Initial import of upstream libjpeg-turbo (svn revision 290)

218 files changed, 75013 insertions, 0 deletions

diff --git a/BUILDING.txt b/BUILDING.txt
new file mode 100644
index 0000000..aa1d1cf
--- /dev/null
+++ b/BUILDING.txt
@@ -0,0 +1,584 @@
+*******************************************************************************
+**     Building on Unix Platforms (including Cygwin)
+*******************************************************************************
+
+
+==================
+Build Requirements
+==================
+
+-- autoconf 2.56 or later
+-- automake 1.7 or later
+-- libtool 1.4 or later
+
+-- NASM
+   * 0.98 or later is required for a 32-bit build
+   * NASM 2.05 or later is required for a 64-bit build
+   * NASM 2.07 or later is required for a 64-bit build on OS X.  This can be
+     obtained from MacPorts (http://www.macports.org/).
+
+   The NASM 2.05 RPMs do not work on older Linux systems, such as Red Hat
+   Enterprise Linux 4.  On such systems, you can easily build and install NASM
+   2.05 from the source RPM by executing the following as root:
+
+     ARCH=`uname -m`
+     wget http://www.nasm.us/pub/nasm/releasebuilds/2.05.01/nasm-2.05.01-1.src.rpm
+     rpmbuild --rebuild nasm-2.05.01-1.src.rpm
+     rpm -Uvh /usr/src/redhat/RPMS/$ARCH/nasm-2.05.01-1.$ARCH.rpm
+
+   NOTE: NASM build will fail if texinfo is not installed.
+
+-- GCC v4.1 or later recommended for best performance
+
+
+==================
+Out-of-Tree Builds
+==================
+
+Binary objects, libraries, and executables are generated in the same directory
+from which configure was executed (the "binary directory"), and this directory
+need not necessarily be the same as the libjpeg-turbo source directory.  You
+can create multiple independent binary directories, in which different versions
+of libjpeg-turbo can be built from the same source tree using different
+compilers or settings.  In the sections below, {build_directory} refers to the
+binary directory, whereas {source_directory} refers to the libjpeg-turbo source
+directory.  For in-tree builds, these directories are the same.
+
+
+======================
+Building libjpeg-turbo
+======================
+
+The following procedure will build libjpeg-turbo on Linux, FreeBSD, 32-bit
+OS X, Cygwin, and Solaris/x86 systems (on Solaris, this generates a 32-bit
+library.  See below for 64-bit build instructions.)
+
+  cd {source_directory}
+  autoreconf -fiv
+  cd {build_directory}
+  sh {source_directory}/configure [additional configure flags]
+  make
+
+NOTE: Running autoreconf in the source directory is only necessary if building
+libjpeg-turbo from the SVN repository.
+
+This will generate the following files under .libs/
+
+  libjpeg.a
+      Static link library for libjpeg-turbo
+
+  libjpeg.so.{version} (Linux, Solaris)
+  libjpeg.{version}.dylib (OS X)
+  cygjpeg-{version}.dll (Cygwin)
+      Shared library for libjpeg-turbo
+
+  libjpeg.so (Linux, Solaris)
+  libjpeg.dylib (OS X)
+  libjpeg.dll.a (Cygwin)
+      Development stub for libjpeg-turbo shared library
+
+  libturbojpeg.a
+      Static link library for TurboJPEG/OSS
+
+  libturbojpeg.so (Linux, Solaris)
+  libturbojpeg.dylib (OS X)
+      Shared library and development stub for TurboJPEG/OSS
+
+  cygturbojpeg.dll (Cygwin)
+      Shared library for TurboJPEG/OSS
+
+  libturbojpeg.dll.a (Cygwin)
+      Development stub for TurboJPEG/OSS shared library
+
+{version} is 62.0.0, 7.0.0, or 8.0.2, depending on whether libjpeg v6b
+(default), v7, or v8b emulation is enabled.  If using Cygwin, {version} is
+62, 7, or 8.
+
+
+libjpeg v7 or v8b Emulation
+---------------------------
+
+Add --with-jpeg7 to the configure command line to build a version of
+libjpeg-turbo that is compatible with libjpeg v7.  Add --with-jpeg8 to the
+configure command to build a version of libjpeg-turbo that is compatible with
+libjpeg v8b.  See README-turbo.txt for more information on libjpeg v7 and v8b
+emulation.
+
+
+
+========================
+Installing libjpeg-turbo
+========================
+
+If you intend to install these libraries and the associated header files, then
+replace 'make' in the instructions above with
+
+  make install prefix={base dir} libdir={library directory}
+
+For example,
+
+  make install prefix=/usr/local libdir=/usr/local/lib64
+
+will install the header files in /usr/local/include and the library files in
+/usr/local/lib64.  If 'prefix' and 'libdir' are not specified, then the default
+is to install the header files in /opt/libjpeg-turbo/include and the library
+files in /opt/libjpeg-turbo/lib.
+
+NOTE: You can specify a prefix of /usr and a libdir of, for instance,
+/usr/lib64 to overwrite the system's version of libjpeg.  If you do this,
+however, then be sure to BACK UP YOUR SYSTEM'S INSTALLATION OF LIBJPEG before
+overwriting it.  It is recommended that you instead install libjpeg-turbo into
+a non-system directory and manipulate the LD_LIBRARY_PATH or create sym links
+to force applications to use libjpeg-turbo instead of libjpeg.  See
+README-turbo.txt for more information.
+
+
+=============
+Build Recipes
+=============
+
+
+32-bit Library Build on 64-bit Linux
+------------------------------------
+
+Add
+
+  --host i686-pc-linux-gnu CFLAGS='-O3 -m32' CXXFLAGS='-O3 -m32' LDFLAGS=-m32
+
+to the configure command line.
+
+
+64-bit Library Build on 64-bit OS X
+-----------------------------------
+
+Add
+
+  --host x86_64-apple-darwin NASM=/opt/local/bin/nasm
+
+to the configure command line.  NASM 2.07 or later from MacPorts must be
+installed.
+
+
+32-bit Library Build on 64-bit OS X
+-----------------------------------
+
+Add
+
+  CFLAGS='-O3 -m32' CXXFLAGS='-O3 -m32' LDFLAGS=-m32
+
+to the configure command line.
+
+
+64-bit Backward-Compatible Library Build on 64-bit OS X
+-------------------------------------------------------
+
+Add
+
+  --host x86_64-apple-darwin NASM=/opt/local/bin/nasm \
+  CFLAGS='-isysroot /Developer/SDKs/MacOSX10.4u.sdk \
+    -mmacosx-version-min=10.4 -O3' \
+    CXXFLAGS='-isysroot /Developer/SDKs/MacOSX10.4u.sdk \
+    -mmacosx-version-min=10.4 -O3' \
+    LDFLAGS='-isysroot /Developer/SDKs/MacOSX10.4u.sdk \
+    -mmacosx-version-min=10.4'
+
+to the configure command line.  The OS X 10.4 SDK, and NASM 2.07 or later from
+MacPorts, must be installed.
+
+
+32-bit Backward-Compatible Library Build on OS X
+------------------------------------------------
+
+Add
+
+  CFLAGS='-isysroot /Developer/SDKs/MacOSX10.4u.sdk \
+    -mmacosx-version-min=10.4 -O3 -m32' \
+    CXXFLAGS='-isysroot /Developer/SDKs/MacOSX10.4u.sdk \
+    -mmacosx-version-min=10.4 -O3 -m32' \
+    LDFLAGS='-isysroot /Developer/SDKs/MacOSX10.4u.sdk \
+    -mmacosx-version-min=10.4 -m32'
+
+to the configure command line.  The OS X 10.4 SDK must be installed.
+
+
+64-bit Library Build on 64-bit Solaris
+--------------------------------------
+
+Add
+
+  --host x86_64-pc-solaris CFLAGS='-O3 -m64' CXXFLAGS='-O3 -m64' LDFLAGS=-m64
+
+to the configure command line.
+
+
+32-bit Library Build on 64-bit FreeBSD
+--------------------------------------
+
+Add
+
+  --host i386-unknown-freebsd CC='gcc -B /usr/lib32' CXX='g++ -B/usr/lib32' \
+    CFLAGS='-O3 -m32' CXXFLAGS='-O3 -m32' LDFLAGS='-B/usr/lib32'
+
+to the configure command line.  NASM 2.07 or later from FreeBSD ports must be
+installed.
+
+
+Sun Studio
+----------
+
+Add
+
+  CC=cc CXX=CC
+
+to the configure command line.  libjpeg-turbo will automatically be built with
+the maximum optimization level (-xO5) unless you override CFLAGS and CXXFLAGS.
+
+To build a 64-bit version of libjpeg-turbo using Sun Studio, add
+
+  --host x86_64-pc-solaris CC=cc CXX=CC CFLAGS='-xO5 -m64' \
+    CXXFLAGS='-xO5 -m64' LDFLAGS=-m64
+
+to the configure command line.
+
+
+MinGW Build on Cygwin
+---------------------
+
+Use CMake (see recipes below)
+
+
+
+*******************************************************************************
+**     Building on Windows (Visual C++ or MinGW)
+*******************************************************************************
+
+
+==================
+Build Requirements
+==================
+
+-- CMake (http://www.cmake.org) v2.6 or later
+
+-- Microsoft Visual C++ 2005 or later
+
+   If you don't already have Visual C++, then the easiest way to get it is by
+   installing the Windows SDK:
+
+   http://msdn.microsoft.com/en-us/windows/bb980924.aspx
+
+   The Windows SDK includes both 32-bit and 64-bit Visual C++ compilers and
+   everything necessary to build libjpeg-turbo.
+
+   * For 32-bit builds, you can also use Microsoft Visual C++ Express
+     Edition.  Visual C++ Express Edition is a free download.
+   * If you intend to build libjpeg-turbo from the command line, then add the
+     appropriate compiler and SDK directories to the INCLUDE, LIB, and PATH
+     environment variables.  This is generally accomplished by executing
+     vcvars32.bat or vcvars64.bat and SetEnv.cmd.  vcvars32.bat and
+     vcvars64.bat are part of Visual C++ and are located in the same directory
+     as the compiler.  SetEnv.cmd is part of the Windows SDK.  You can pass
+     optional arguments to SetEnv.cmd to specify a 32-bit or 64-bit build
+     environment.
+
+... OR ...
+
+-- MinGW
+
+   GCC v4.1 or later recommended for best performance
+
+-- NASM (http://www.nasm.us/) 0.98 or later (NASM 2.05 or later is required for
+   a 64-bit build)
+
+
+==================
+Out-of-Tree Builds
+==================
+
+Binary objects, libraries, and executables are generated in the same directory
+from which cmake was executed (the "binary directory"), and this directory need
+not necessarily be the same as the libjpeg-turbo source directory.  You can
+create multiple independent binary directories, in which different versions of
+libjpeg-turbo can be built from the same source tree using different compilers
+or settings.  In the sections below, {build_directory} refers to the binary
+directory, whereas {source_directory} refers to the libjpeg-turbo source
+directory.  For in-tree builds, these directories are the same.
+
+
+======================
+Building libjpeg-turbo
+======================
+
+
+Visual C++ (Command Line)
+-------------------------
+
+  cd {build_directory}
+  cmake -G "NMake Makefiles" -DCMAKE_BUILD_TYPE=Release {source_directory}
+  nmake
+
+This will build either a 32-bit or a 64-bit version of libjpeg-turbo, depending
+on which version of cl.exe is in the PATH.
+
+The following files will be generated under {build_directory}:
+
+  jpeg-static.lib
+      Static link library for libjpeg-turbo
+  sharedlib/jpeg{version}.dll
+      DLL for libjpeg-turbo
+  sharedlib/jpeg.lib
+      Import library for libjpeg-turbo DLL
+  turbojpeg-static.lib
+      Static link library for TurboJPEG/OSS
+  turbojpeg.dll
+      DLL for TurboJPEG/OSS
+  turbojpeg.lib
+      Import library for TurboJPEG/OSS DLL
+
+{version} is 62, 7, or 8, depending on whether libjpeg v6b (default), v7, or
+v8b emulation is enabled.
+
+
+Visual C++ (IDE)
+----------------
+
+Choose the appropriate CMake generator option for your version of Visual Studio
+(run "cmake" with no arguments for a list of available generators.)  For
+instance:
+
+  cd {build_directory}
+  cmake -G "Visual Studio 9 2008" {source_directory}
+
+You can then open ALL_BUILD.vcproj in Visual Studio and build one of the
+configurations in that project ("Debug", "Release", etc.) to generate a full
+build of libjpeg-turbo.
+
+This will generate the following files under {build_directory}:
+
+  {configuration}/jpeg-static.lib
+      Static link library for libjpeg-turbo
+  sharedlib/{configuration}/jpeg{version}.dll
+      DLL for libjpeg-turbo
+  sharedlib/{configuration}/jpeg.lib
+      Import library for libjpeg-turbo DLL
+  {configuration}/turbojpeg-static.lib
+      Static link library for TurboJPEG/OSS
+  {configuration}/turbojpeg.dll
+      DLL for TurboJPEG/OSS
+  {configuration}/turbojpeg.lib
+      Import library for TurboJPEG/OSS DLL
+
+{configuration} is Debug, Release, RelWithDebInfo, or MinSizeRel, depending on
+the configuration you built in the IDE, and {version} is 62, 7, or 8,
+depending on whether libjpeg v6b (default), v7, or v8b emulation is enabled.
+
+
+MinGW
+-----
+
+  cd {build_directory}
+  cmake -G "MSYS Makefiles" {source_directory}
+  make
+
+This will generate the following files under {build_directory}
+
+  libjpeg.a
+      Static link library for libjpeg-turbo
+  sharedlib/libjpeg-{version}.dll
+      DLL for libjpeg-turbo
+  sharedlib/libjpeg.dll.a
+      Import library for libjpeg-turbo DLL
+  libturbojpeg.a
+      Static link library for TurboJPEG/OSS
+  libturbojpeg.dll
+      DLL for TurboJPEG/OSS
+  libturbojpeg.dll.a
+      Import library for TurboJPEG/OSS DLL
+
+{version} is 62, 7, or 8, depending on whether libjpeg v6b (default), v7, or
+v8b emulation is enabled.
+
+
+Debug Build
+-----------
+
+Add "-DCMAKE_BUILD_TYPE=Debug" to the cmake command line.  Or, if building with
+NMake, remove "-DCMAKE_BUILD_TYPE=Release" (Debug builds are the default with
+NMake.)
+
+
+libjpeg v7 or v8b Emulation
+---------------------------
+
+Add "-DWITH_JPEG7=1" to the cmake command line to build a version of
+libjpeg-turbo that is compatible with libjpeg v7.  Add "-DWITH_JPEG8=1" to the
+cmake command to build a version of libjpeg-turbo that is compatible with
+libjpeg v8b.  See README-turbo.txt for more information on libjpeg v7 and v8b
+emulation.
+
+
+========================
+Installing libjpeg-turbo
+========================
+
+You can use the build system to install libjpeg-turbo into a directory of your
+choosing (as opposed to creating an installer.)  To do this, add:
+
+  -DCMAKE_INSTALL_PREFIX={install_directory}
+
+to the cmake command line.
+
+For example,
+
+  cmake -G "NMake Makefiles" -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_INSTALL_PREFIX=c:\libjpeg-turbo {source_directory}
+  nmake install
+
+will install the header files in c:\libjpeg-turbo\include, the library files
+in c:\libjpeg-turbo\lib, the DLL's in c:\libjpeg-turbo\bin, and the
+documentation in c:\libjpeg-turbo\doc.
+
+
+=============
+Build Recipes
+=============
+
+
+64-bit MinGW Build on Cygwin
+----------------------------
+
+  cd {build_directory}
+  CC=/usr/bin/x86_64-w64-mingw32-gcc CXX=/usr/bin/x86_64-w64-mingw32-g++ \
+    cmake -G "Unix Makefiles" -DCMAKE_SYSTEM_NAME=Windows \
+    -DCMAKE_AR=/usr/bin/x86_64-w64-mingw32-ar \
+    -DCMAKE_RANLIB=/usr/bin/x86_64-w64-mingw32-ranlib {source_directory}
+  make
+
+This produces a 64-bit build of libjpeg-turbo that does not depend on
+cygwin1.dll or other Cygwin DLL's.  The mingw64-x86_64-gcc-core and
+mingw64-x86_64-gcc-g++ packages (and their dependencies) must be installed.
+
+
+32-bit MinGW Build on Cygwin
+----------------------------
+
+  cd {build_directory}
+  CC=/usr/bin/i686-w64-mingw32-gcc CXX=/usr/bin/i686-w64-mingw32-g++ \
+    cmake -G "Unix Makefiles" -DCMAKE_SYSTEM_NAME=Windows \
+    -DDCMAKE_AR=/usr/bin/i686-w64-mingw32-ar \
+    -DCMAKE_RANLIB=/usr/bin/i686-w64-mingw32-ranlib {source_directory}
+  make
+
+This produces a 32-bit build of libjpeg-turbo that does not depend on
+cygwin1.dll or other Cygwin DLL's.  The mingw64-i686-gcc-core and
+mingw64-i686-gcc-g++ packages (and their dependencies) must be installed.
+
+
+MinGW-w64 Build on Windows
+--------------------------
+
+This produces a 64-bit build of libjpeg-turbo using the "native" MinGW-w64
+toolchain (which is faster than the Cygwin version):
+
+  cd {build_directory}
+  CC={mingw-w64_binary_path}/x86_64-w64-mingw32-gcc \
+    CXX={mingw-w64_binary_path}/x86_64-w64-mingw32-g++ \
+    cmake -G "MSYS Makefiles" \
+    -DCMAKE_AR={mingw-w64_binary_path}/x86_64-w64-mingw32-ar \
+    -DCMAKE_RANLIB={mingw-w64_binary_path}/x86_64-w64-mingw32-ranlib \
+    {source_directory}
+  make
+
+
+MinGW Build on Linux
+--------------------
+
+  cd {build_directory}
+  CC={mingw_binary_path}/i386-mingw32-gcc \
+    CXX={mingw_binary_path}/i386-mingw32-g++ \
+    cmake -G "Unix Makefiles" -DCMAKE_SYSTEM_NAME=Windows \
+    -DCMAKE_AR={mingw_binary_path}/i386-mingw32-ar \
+    -DCMAKE_RANLIB={mingw_binary_path}/i386-mingw32-ranlib \
+    {source_directory}
+  make
+
+
+*******************************************************************************
+**     Creating Release Packages
+*******************************************************************************
+
+The following commands can be used to create various types of release packages:
+
+
+Unix
+----
+
+make rpm
+
+  Create Red Hat-style binary RPM package.  Requires RPM v4 or later.
+
+make srpm
+
+  This runs 'make dist' to create a pristine source tarball, then creates a
+  Red Hat-style source RPM package from the tarball.  Requires RPM v4 or later.
+
+make deb
+
+  Create Debian-style binary package.  Requires dpkg.
+
+make dmg
+
+  Create Macintosh package/disk image.  This requires the PackageMaker
+  application, which must be installed in /Developer/Applications/Utilities.
+
+make udmg [BUILDDIR32={32-bit build directory}]
+
+  On 64-bit OS X systems, this creates a version of the Macintosh package and
+  disk image which contains universal i386/x86-64 binaries.  You should first
+  configure a 32-bit out-of-tree build of libjpeg-turbo, then configure a
+  64-bit out-of-tree build, then run 'make udmg' from the 64-bit build
+  directory.  The build system will look for the 32-bit build under
+  {source_directory}/osxx86 by default, but you can override this by setting
+  the BUILDDIR32 variable on the make command line as shown above.
+
+make sunpkg
+
+  Build a Solaris package.  This requires pkgmk, pkgtrans, and bzip2.
+
+make csunpkg [BUILDDIR32={32-bit build directory}]
+
+  On 64-bit Solaris systems, this creates a combined package which contains
+  both 32-bit and 64-bit libraries.  You should first configure a 32-bit
+  out-of-tree build of libjpeg-turbo, then configure a 64-bit out-of-tree
+  build, then run 'make csunpkg' from the 64-bit build directory.  The build
+  system will look for the 32-bit build under {source_directory}/solx86 by
+  default, but you can override this by setting the BUILDDIR32 variable on the
+  make command line as shown above.
+
+make cygwinpkg
+
+  Build a Cygwin binary package.
+
+
+Windows
+-------
+
+If using NMake:
+
+  cd {build_directory}
+  nmake installer
+
+If using MinGW:
+
+  cd {build_directory}
+  make installer
+
+If using the Visual Studio IDE, build the "installer" project.
+
+The installer package (libjpeg-turbo[-gcc][64].exe) will be located under
+{build_directory}.  If building using the Visual Studio IDE, then the installer
+package will be located in a subdirectory with the same name as the
+configuration you built (such as {build_directory}\Debug\ or
+{build_directory}\Release\).
+
+Building a Windows installer requires the Nullsoft Install System
+(http://nsis.sourceforge.net/.)  makensis.exe should be in your PATH.
diff --git a/CMakeLists.txt b/CMakeLists.txt
new file mode 100644
index 0000000..0a07001
--- /dev/null
+++ b/CMakeLists.txt
@@ -0,0 +1,284 @@
+#
+# Setup
+#
+
+cmake_minimum_required(VERSION 2.6)
+
+project(libjpeg-turbo)
+set(VERSION 1.0.90)
+
+if(MINGW OR CYGWIN)
+  execute_process(COMMAND "date" "+%Y%m%d" OUTPUT_VARIABLE BUILD)
+  string(REGEX REPLACE "\n" "" BUILD ${BUILD})
+elseif(WIN32)
+  execute_process(COMMAND "${CMAKE_SOURCE_DIR}/cmakescripts/getdate.bat"
+    OUTPUT_VARIABLE BUILD)
+  string(REGEX REPLACE "\n" "" BUILD ${BUILD})
+else()
+  message(FATAL_ERROR "Platform not supported by this build system.  Use autotools instead.")
+endif()
+
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE Release)
+endif()
+
+message(STATUS "CMAKE_BUILD_TYPE = ${CMAKE_BUILD_TYPE}")
+
+# This only works if building from the command line.  There is currently no way
+# to set a variable's value based on the build type when using the MSVC IDE.
+if(CMAKE_BUILD_TYPE STREQUAL "Debug")
+  set(BUILD "${BUILD}d")
+endif()
+
+message(STATUS "VERSION = ${VERSION}, BUILD = ${BUILD}")
+
+if(NOT DEFINED WITH_SIMD)
+  set(WITH_SIMD 1)
+endif()
+
+set(JPEG_LIB_VERSION 62)
+set(DLL_VERSION ${JPEG_LIB_VERSION})
+set(FULLVERSION ${DLL_VERSION}.0.0)
+if(WITH_JPEG8)
+  set(JPEG_LIB_VERSION 80)
+  set(DLL_VERSION 8)
+  set(FULLVERSION ${DLL_VERSION}.0.2)
+  message(STATUS "Emulating libjpeg v8b API/ABI")
+elseif(WITH_JPEG7)
+  set(JPEG_LIB_VERSION 70)
+  set(DLL_VERSION 7)
+  set(FULLVERSION ${DLL_VERSION}.0.0)
+  message(STATUS "Emulating libjpeg v7 API/ABI")
+endif(WITH_JPEG8)
+
+if(MSVC)
+  # Use the static C library for all build types
+  foreach(var CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
+    CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO
+    CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+    CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+    if(${var} MATCHES "/MD")
+      string(REGEX REPLACE "/MD" "/MT" ${var} "${${var}}")
+    endif()
+  endforeach()
+
+  add_definitions(-W3 -wd4996)
+endif()
+
+# Detect whether compiler is 64-bit
+if(MSVC AND CMAKE_CL_64)
+  set(SIMD_X86_64 1)
+  set(64BIT 1)
+elseif(CMAKE_SIZEOF_VOID_P MATCHES 8)
+  set(SIMD_X86_64 1)
+  set(64BIT 1)
+endif()
+
+if(64BIT)
+  message(STATUS "64-bit build")
+else()
+  message(STATUS "32-bit build")
+endif()
+
+configure_file(win/jconfig.h.in jconfig.h)
+configure_file(win/config.h.in config.h)
+
+include_directories(${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_SOURCE_DIR})
+
+
+#
+# Targets
+#
+
+set(JPEG_SOURCES jcapimin.c jcapistd.c jccoefct.c jccolor.c jcdctmgr.c jchuff.c
+  jcinit.c jcmainct.c jcmarker.c jcmaster.c jcomapi.c jcparam.c jcphuff.c
+  jcprepct.c jcsample.c jctrans.c jdapimin.c jdapistd.c jdatadst.c jdatasrc.c
+  jdcoefct.c jdcolor.c jddctmgr.c jdhuff.c jdinput.c jdmainct.c jdmarker.c
+  jdmaster.c jdmerge.c jdphuff.c jdpostct.c jdsample.c jdtrans.c jerror.c
+  jfdctflt.c jfdctfst.c jfdctint.c jidctflt.c jidctfst.c jidctint.c jidctred.c
+  jquant1.c jquant2.c jutils.c jmemmgr.c jmemnobs.c)
+
+if(WITH_SIMD)
+  add_definitions(-DWITH_SIMD)
+  add_subdirectory(simd)
+  if(SIMD_X86_64)
+    set(JPEG_SOURCES ${JPEG_SOURCES} simd/jsimd_x86_64.c)
+  else()
+    set(JPEG_SOURCES ${JPEG_SOURCES} simd/jsimd_i386.c)
+  endif()
+  # This tells CMake that the "source" files haven't been generated yet
+  set_source_files_properties(${SIMD_OBJS} PROPERTIES GENERATED 1)
+else()
+  set(JPEG_SOURCES ${JPEG_SOURCES} jsimd_none.c)
+  message(STATUS "Not using SIMD acceleration")
+endif()
+
+add_subdirectory(sharedlib)
+
+add_library(jpeg-static STATIC ${JPEG_SOURCES} ${SIMD_OBJS})
+if(NOT MSVC)
+  set_target_properties(jpeg-static PROPERTIES OUTPUT_NAME jpeg)
+endif()
+if(WITH_SIMD)
+  add_dependencies(jpeg-static simd)
+endif()
+
+add_library(turbojpeg SHARED turbojpegl.c)
+set_target_properties(turbojpeg PROPERTIES DEFINE_SYMBOL DLLDEFINE)
+target_link_libraries(turbojpeg jpeg-static)
+set_target_properties(turbojpeg PROPERTIES LINK_INTERFACE_LIBRARIES "")
+
+add_library(turbojpeg-static STATIC ${JPEG_SOURCES} ${SIMD_OBJS}
+  turbojpegl.c)
+if(NOT MSVC)
+  set_target_properties(turbojpeg-static PROPERTIES OUTPUT_NAME turbojpeg)
+endif()
+if(WITH_SIMD)
+  add_dependencies(turbojpeg-static simd)
+endif()
+
+add_executable(jpegut jpegut.c)
+target_link_libraries(jpegut turbojpeg)
+
+add_executable(jpegut-static jpegut.c)
+target_link_libraries(jpegut-static turbojpeg-static)
+
+add_executable(jpgtest jpgtest.cxx bmp.c)
+target_link_libraries(jpgtest turbojpeg)
+
+add_executable(jpgtest-static jpgtest.cxx bmp.c)
+target_link_libraries(jpgtest-static turbojpeg-static)
+
+add_executable(cjpeg-static cjpeg.c cdjpeg.c rdbmp.c rdgif.c rdppm.c rdswitch.c
+  rdtarga.c)
+set_property(TARGET cjpeg-static PROPERTY COMPILE_FLAGS
+  "-DBMP_SUPPORTED -DGIF_SUPPORTED -DPPM_SUPPORTED -DTARGA_SUPPORTED")
+target_link_libraries(cjpeg-static jpeg-static)
+
+add_executable(djpeg-static djpeg.c cdjpeg.c rdcolmap.c rdswitch.c wrbmp.c wrgif.c
+  wrppm.c wrtarga.c)
+set_property(TARGET djpeg-static PROPERTY COMPILE_FLAGS
+  "-DBMP_SUPPORTED -DGIF_SUPPORTED -DPPM_SUPPORTED -DTARGA_SUPPORTED")
+target_link_libraries(djpeg-static jpeg-static)
+
+add_executable(jpegtran-static jpegtran.c cdjpeg.c rdswitch.c transupp.c)
+target_link_libraries(jpegtran-static jpeg-static)
+
+add_executable(rdjpgcom rdjpgcom.c)
+
+add_executable(wrjpgcom rdjpgcom.c)
+
+
+#
+# Tests
+#
+
+enable_testing()
+add_test(jpegut jpegut)
+add_test(cjpeg-int sharedlib/cjpeg -dct int -outfile testoutint.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+add_test(cjpeg-int-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgint.jpg testoutint.jpg)
+add_test(cjpeg-fast sharedlib/cjpeg -dct fast -opt -outfile testoutfst.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+add_test(cjpeg-fast-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgfst.jpg testoutfst.jpg)
+add_test(cjpeg-float sharedlib/cjpeg -dct float -outfile testoutflt.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+if(WITH_SIMD)
+add_test(cjpeg-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgflt.jpg testoutflt.jpg)
+else()
+add_test(cjpeg-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgflt-nosimd.jpg testoutflt.jpg)
+endif()
+add_test(djpeg-int sharedlib/djpeg -dct int -fast -ppm -outfile testoutint.ppm ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(djpeg-int-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgint.ppm testoutint.ppm)
+add_test(djpeg-fast sharedlib/djpeg -dct fast -ppm -outfile testoutfst.ppm ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(djpeg-fast-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgfst.ppm testoutfst.ppm)
+add_test(djpeg-float sharedlib/djpeg -dct float -ppm -outfile testoutflt.ppm ${CMAKE_SOURCE_DIR}/testorig.jpg)
+if(WITH_SIMD)
+add_test(djpeg-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgflt.ppm testoutflt.ppm)
+else()
+add_test(djpeg-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testorig.ppm testoutflt.ppm)
+endif()
+add_test(djpeg-256 sharedlib/djpeg -dct int -bmp -colors 256 -outfile testout.bmp  ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(djpeg-256-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimg.bmp testout.bmp)
+add_test(cjpeg-prog sharedlib/cjpeg -dct int -progressive -outfile testoutp.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+add_test(cjpeg-prog-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgp.jpg testoutp.jpg)
+add_test(jpegtran-prog sharedlib/jpegtran -outfile testoutt.jpg testoutp.jpg)
+add_test(jpegtran-prog-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgint.jpg testoutt.jpg)
+add_test(jpegtran-crop sharedlib/jpegtran -crop 120x90+20+50 -transpose -perfect -outfile testoutcrop.jpg ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(jpegtran-crop-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgcrop.jpg testoutcrop.jpg)
+
+add_test(jpegut-static jpegut-static)
+add_test(cjpeg-static-int cjpeg-static -dct int -outfile testoutint.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+add_test(cjpeg-static-int-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgint.jpg testoutint.jpg)
+add_test(cjpeg-static-fast cjpeg-static -dct fast -opt -outfile testoutfst.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+add_test(cjpeg-static-fast-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgfst.jpg testoutfst.jpg)
+add_test(cjpeg-static-float cjpeg-static -dct float -outfile testoutflt.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+if(WITH_SIMD)
+add_test(cjpeg-static-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgflt.jpg testoutflt.jpg)
+else()
+add_test(cjpeg-static-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgflt-nosimd.jpg testoutflt.jpg)
+endif()
+add_test(djpeg-static-int djpeg-static -dct int -fast -ppm -outfile testoutint.ppm ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(djpeg-static-int-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgint.ppm testoutint.ppm)
+add_test(djpeg-static-fast djpeg-static -dct fast -ppm -outfile testoutfst.ppm ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(djpeg-static-fast-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgfst.ppm testoutfst.ppm)
+add_test(djpeg-static-float djpeg-static -dct float -ppm -outfile testoutflt.ppm ${CMAKE_SOURCE_DIR}/testorig.jpg)
+if(WITH_SIMD)
+add_test(djpeg-static-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgflt.ppm testoutflt.ppm)
+else()
+add_test(djpeg-static-float-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testorig.ppm testoutflt.ppm)
+endif()
+add_test(djpeg-static-256 djpeg-static -dct int -bmp -colors 256 -outfile testout.bmp  ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(djpeg-static-256-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimg.bmp testout.bmp)
+add_test(cjpeg-static-prog cjpeg-static -dct int -progressive -outfile testoutp.jpg ${CMAKE_SOURCE_DIR}/testorig.ppm)
+add_test(cjpeg-static-prog-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgp.jpg testoutp.jpg)
+add_test(jpegtran-static-prog jpegtran-static -outfile testoutt.jpg testoutp.jpg)
+add_test(jpegtran-static-prog-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgint.jpg testoutt.jpg)
+add_test(jpegtran-static-crop jpegtran-static -crop 120x90+20+50 -transpose -perfect -outfile testoutcrop.jpg ${CMAKE_SOURCE_DIR}/testorig.jpg)
+add_test(jpegtran-static-crop-cmp ${CMAKE_COMMAND} -E compare_files ${CMAKE_SOURCE_DIR}/testimgcrop.jpg testoutcrop.jpg)
+
+
+#
+# Installer
+#
+
+set(INST_NAME ${CMAKE_PROJECT_NAME})
+
+if(MSVC)
+  set(INST_PLATFORM "Visual C++")
+elseif(MINGW)
+  set(INST_PLATFORM GCC)
+  set(INST_NAME ${INST_NAME}-gcc)
+  set(INST_DEFS -DGCC)
+endif()
+
+if(64BIT)
+  set(INST_PLATFORM "${INST_PLATFORM} 64-bit")
+  set(INST_NAME ${INST_NAME}64)
+  set(INST_DEFS ${INST_DEFS} -DWIN64)
+endif()
+
+if(MSVC_IDE)
+  set(INST_DEFS ${INST_DEFS} "-DBUILDDIR=$(OutDir)\\")
+else()
+  set(INST_DEFS ${INST_DEFS} "-DBUILDDIR=")
+endif()
+
+configure_file(release/libjpeg-turbo.nsi.in libjpeg-turbo.nsi @ONLY)
+
+add_custom_target(installer
+  makensis -nocd ${INST_DEFS} libjpeg-turbo.nsi
+  DEPENDS jpeg jpeg-static turbojpeg turbojpeg-static
+  SOURCES libjpeg-turbo.nsi)
+
+install(TARGETS jpeg-static turbojpeg turbojpeg-static rdjpgcom wrjpgcom
+  ARCHIVE DESTINATION lib
+  LIBRARY DESTINATION lib
+  RUNTIME DESTINATION bin
+)
+
+install(FILES ${CMAKE_SOURCE_DIR}/LGPL.txt ${CMAKE_SOURCE_DIR}/LICENSE.txt
+  ${CMAKE_SOURCE_DIR}/README ${CMAKE_SOURCE_DIR}/README-turbo.txt
+  ${CMAKE_SOURCE_DIR}/libjpeg.txt ${CMAKE_SOURCE_DIR}/usage.txt
+  DESTINATION doc)
+
+install(FILES ${CMAKE_BINARY_DIR}/jconfig.h ${CMAKE_SOURCE_DIR}/jerror.h
+  ${CMAKE_SOURCE_DIR}/jmorecfg.h ${CMAKE_SOURCE_DIR}/jpeglib.h
+  ${CMAKE_SOURCE_DIR}/turbojpeg.h DESTINATION include)
diff --git a/ChangeLog.txt b/ChangeLog.txt
new file mode 100644
index 0000000..1cabc2c
--- /dev/null
+++ b/ChangeLog.txt
@@ -0,0 +1,65 @@
+Significant changes since 1.0.1
+===============================
+
+[1] Added emulation of the libjpeg v7 and v8b APIs and ABIs.  See
+README-turbo.txt for more details.  This feature was sponsored by CamTrace SAS.
+
+[2] Created a new CMake-based build system for the Visual C++ and MinGW builds.
+
+[3] TurboJPEG/OSS can now compress from/decompress to grayscale bitmaps.
+
+
+Significant changes since 1.0.0
+===============================
+
+[1] The Huffman decoder will now handle erroneous Huffman codes (for instance,
+from a corrupt JPEG image.)  Previously, these would cause libjpeg-turbo to
+crash under certain circumstances.
+
+[2] Fixed typo in SIMD dispatch routines which was causing 4:2:2 upsampling to
+be used instead of 4:2:0 when decompressing JPEG images using SSE2 code.
+
+[3] configure script will now automatically determine whether the
+INCOMPLETE_TYPES_BROKEN macro should be defined.
+
+
+Significant changes since 0.0.93
+================================
+
+[1] 2983700: Further FreeBSD build tweaks (no longer necessary to specify
+--host when configuring on a 64-bit system)
+
+[2] Created sym. links in the Unix/Linux packages so that the TurboJPEG
+include file can always be found in /opt/libjpeg-turbo/include, the 32-bit
+static libraries can always be found in /opt/libjpeg-turbo/lib32, and the
+64-bit static libraries can always be found in /opt/libjpeg-turbo/lib64.
+
+[3] The Unix/Linux distribution packages now include the libjpeg run-time
+programs (cjpeg, etc.) and man pages.
+
+[4] Created a 32-bit supplementary package for amd64 Debian systems which
+contains just the 32-bit libjpeg-turbo libraries.
+
+[5] Moved the libraries from */lib32 to */lib in the i386 Debian package.
+
+[6] Include distribution package for Cygwin
+
+[7] No longer necessary to specify --without-simd on non-x86 architectures, and
+unit tests now work on those architectures.
+
+
+Significant changes since 0.0.91
+================================
+
+[1] 2982659, Fixed x86-64 build on FreeBSD systems
+
+[2] 2988188: Added support for Windows 64-bit systems
+
+
+Significant changes since 0.0.90
+================================
+
+[1] Added documentation to .deb packages
+
+[2] 2968313: Fixed data corruption issues when decompressing large JPEG images
+and/or using buffered I/O with the libjpeg-turbo decompressor
diff --git a/LGPL.txt b/LGPL.txt
new file mode 100644
index 0000000..b1e3f5a
--- /dev/null
+++ b/LGPL.txt
@@ -0,0 +1,504 @@
+		  GNU LESSER GENERAL PUBLIC LICENSE
+		       Version 2.1, February 1999
+
+ Copyright (C) 1991, 1999 Free Software Foundation, Inc.
+     59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+[This is the first released version of the Lesser GPL.  It also counts
+ as the successor of the GNU Library Public License, version 2, hence
+ the version number 2.1.]
+
+			    Preamble
+
+  The licenses for most software are designed to take away your
+freedom to share and change it.  By contrast, the GNU General Public
+Licenses are intended to guarantee your freedom to share and change
+free software--to make sure the software is free for all its users.
+
+  This license, the Lesser General Public License, applies to some
+specially designated software packages--typically libraries--of the
+Free Software Foundation and other authors who decide to use it.  You
+can use it too, but we suggest you first think carefully about whether
+this license or the ordinary General Public License is the better
+strategy to use in any particular case, based on the explanations below.
+
+  When we speak of free software, we are referring to freedom of use,
+not price.  Our General Public Licenses are designed to make sure that
+you have the freedom to distribute copies of free software (and charge
+for this service if you wish); that you receive source code or can get
+it if you want it; that you can change the software and use pieces of
+it in new free programs; and that you are informed that you can do
+these things.
+
+  To protect your rights, we need to make restrictions that forbid
+distributors to deny you these rights or to ask you to surrender these
+rights.  These restrictions translate to certain responsibilities for
+you if you distribute copies of the library or if you modify it.
+
+  For example, if you distribute copies of the library, whether gratis
+or for a fee, you must give the recipients all the rights that we gave
+you.  You must make sure that they, too, receive or can get the source
+code.  If you link other code with the library, you must provide
+complete object files to the recipients, so that they can relink them
+with the library after making changes to the library and recompiling
+it.  And you must show them these terms so they know their rights.
+
+  We protect your rights with a two-step method: (1) we copyright the
+library, and (2) we offer you this license, which gives you legal
+permission to copy, distribute and/or modify the library.
+
+  To protect each distributor, we want to make it very clear that
+there is no warranty for the free library.  Also, if the library is
+modified by someone else and passed on, the recipients should know
+that what they have is not the original version, so that the original
+author's reputation will not be affected by problems that might be
+introduced by others.
+
+  Finally, software patents pose a constant threat to the existence of
+any free program.  We wish to make sure that a company cannot
+effectively restrict the users of a free program by obtaining a
+restrictive license from a patent holder.  Therefore, we insist that
+any patent license obtained for a version of the library must be
+consistent with the full freedom of use specified in this license.
+
+  Most GNU software, including some libraries, is covered by the
+ordinary GNU General Public License.  This license, the GNU Lesser
+General Public License, applies to certain designated libraries, and
+is quite different from the ordinary General Public License.  We use
+this license for certain libraries in order to permit linking those
+libraries into non-free programs.
+
+  When a program is linked with a library, whether statically or using
+a shared library, the combination of the two is legally speaking a
+combined work, a derivative of the original library.  The ordinary
+General Public License therefore permits such linking only if the
+entire combination fits its criteria of freedom.  The Lesser General
+Public License permits more lax criteria for linking other code with
+the library.
+
+  We call this license the "Lesser" General Public License because it
+does Less to protect the user's freedom than the ordinary General
+Public License.  It also provides other free software developers Less
+of an advantage over competing non-free programs.  These disadvantages
+are the reason we use the ordinary General Public License for many
+libraries.  However, the Lesser license provides advantages in certain
+special circumstances.
+
+  For example, on rare occasions, there may be a special need to
+encourage the widest possible use of a certain library, so that it becomes
+a de-facto standard.  To achieve this, non-free programs must be
+allowed to use the library.  A more frequent case is that a free
+library does the same job as widely used non-free libraries.  In this
+case, there is little to gain by limiting the free library to free
+software only, so we use the Lesser General Public License.
+
+  In other cases, permission to use a particular library in non-free
+programs enables a greater number of people to use a large body of
+free software.  For example, permission to use the GNU C Library in
+non-free programs enables many more people to use the whole GNU
+operating system, as well as its variant, the GNU/Linux operating
+system.
+
+  Although the Lesser General Public License is Less protective of the
+users' freedom, it does ensure that the user of a program that is
+linked with the Library has the freedom and the wherewithal to run
+that program using a modified version of the Library.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.  Pay close attention to the difference between a
+"work based on the library" and a "work that uses the library".  The
+former contains code derived from the library, whereas the latter must
+be combined with the library in order to run.
+
+		  GNU LESSER GENERAL PUBLIC LICENSE
+   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+
+  0. This License Agreement applies to any software library or other
+program which contains a notice placed by the copyright holder or
+other authorized party saying it may be distributed under the terms of
+this Lesser General Public License (also called "this License").
+Each licensee is addressed as "you".
+
+  A "library" means a collection of software functions and/or data
+prepared so as to be conveniently linked with application programs
+(which use some of those functions and data) to form executables.
+
+  The "Library", below, refers to any such software library or work
+which has been distributed under these terms.  A "work based on the
+Library" means either the Library or any derivative work under
+copyright law: that is to say, a work containing the Library or a
+portion of it, either verbatim or with modifications and/or translated
+straightforwardly into another language.  (Hereinafter, translation is
+included without limitation in the term "modification".)
+
+  "Source code" for a work means the preferred form of the work for
+making modifications to it.  For a library, complete source code means
+all the source code for all modules it contains, plus any associated
+interface definition files, plus the scripts used to control compilation
+and installation of the library.
+
+  Activities other than copying, distribution and modification are not
+covered by this License; they are outside its scope.  The act of
+running a program using the Library is not restricted, and output from
+such a program is covered only if its contents constitute a work based
+on the Library (independent of the use of the Library in a tool for
+writing it).  Whether that is true depends on what the Library does
+and what the program that uses the Library does.
+  
+  1. You may copy and distribute verbatim copies of the Library's
+complete source code as you receive it, in any medium, provided that
+you conspicuously and appropriately publish on each copy an
+appropriate copyright notice and disclaimer of warranty; keep intact
+all the notices that refer to this License and to the absence of any
+warranty; and distribute a copy of this License along with the
+Library.
+
+  You may charge a fee for the physical act of transferring a copy,
+and you may at your option offer warranty protection in exchange for a
+fee.
+
+  2. You may modify your copy or copies of the Library or any portion
+of it, thus forming a work based on the Library, and copy and
+distribute such modifications or work under the terms of Section 1
+above, provided that you also meet all of these conditions:
+
+    a) The modified work must itself be a software library.
+
+    b) You must cause the files modified to carry prominent notices
+    stating that you changed the files and the date of any change.
+
+    c) You must cause the whole of the work to be licensed at no
+    charge to all third parties under the terms of this License.
+
+    d) If a facility in the modified Library refers to a function or a
+    table of data to be supplied by an application program that uses
+    the facility, other than as an argument passed when the facility
+    is invoked, then you must make a good faith effort to ensure that,
+    in the event an application does not supply such function or
+    table, the facility still operates, and performs whatever part of
+    its purpose remains meaningful.
+
+    (For example, a function in a library to compute square roots has
+    a purpose that is entirely well-defined independent of the
+    application.  Therefore, Subsection 2d requires that any
+    application-supplied function or table used by this function must
+    be optional: if the application does not supply it, the square
+    root function must still compute square roots.)
+
+These requirements apply to the modified work as a whole.  If
+identifiable sections of that work are not derived from the Library,
+and can be reasonably considered independent and separate works in
+themselves, then this License, and its terms, do not apply to those
+sections when you distribute them as separate works.  But when you
+distribute the same sections as part of a whole which is a work based
+on the Library, the distribution of the whole must be on the terms of
+this License, whose permissions for other licensees extend to the
+entire whole, and thus to each and every part regardless of who wrote
+it.
+
+Thus, it is not the intent of this section to claim rights or contest
+your rights to work written entirely by you; rather, the intent is to
+exercise the right to control the distribution of derivative or
+collective works based on the Library.
+
+In addition, mere aggregation of another work not based on the Library
+with the Library (or with a work based on the Library) on a volume of
+a storage or distribution medium does not bring the other work under
+the scope of this License.
+
+  3. You may opt to apply the terms of the ordinary GNU General Public
+License instead of this License to a given copy of the Library.  To do
+this, you must alter all the notices that refer to this License, so
+that they refer to the ordinary GNU General Public License, version 2,
+instead of to this License.  (If a newer version than version 2 of the
+ordinary GNU General Public License has appeared, then you can specify
+that version instead if you wish.)  Do not make any other change in
+these notices.
+
+  Once this change is made in a given copy, it is irreversible for
+that copy, so the ordinary GNU General Public License applies to all
+subsequent copies and derivative works made from that copy.
+
+  This option is useful when you wish to copy part of the code of
+the Library into a program that is not a library.
+
+  4. You may copy and distribute the Library (or a portion or
+derivative of it, under Section 2) in object code or executable form
+under the terms of Sections 1 and 2 above provided that you accompany
+it with the complete corresponding machine-readable source code, which
+must be distributed under the terms of Sections 1 and 2 above on a
+medium customarily used for software interchange.
+
+  If distribution of object code is made by offering access to copy
+from a designated place, then offering equivalent access to copy the
+source code from the same place satisfies the requirement to
+distribute the source code, even though third parties are not
+compelled to copy the source along with the object code.
+
+  5. A program that contains no derivative of any portion of the
+Library, but is designed to work with the Library by being compiled or
+linked with it, is called a "work that uses the Library".  Such a
+work, in isolation, is not a derivative work of the Library, and
+therefore falls outside the scope of this License.
+
+  However, linking a "work that uses the Library" with the Library
+creates an executable that is a derivative of the Library (because it
+contains portions of the Library), rather than a "work that uses the
+library".  The executable is therefore covered by this License.
+Section 6 states terms for distribution of such executables.
+
+  When a "work that uses the Library" uses material from a header file
+that is part of the Library, the object code for the work may be a
+derivative work of the Library even though the source code is not.
+Whether this is true is especially significant if the work can be
+linked without the Library, or if the work is itself a library.  The
+threshold for this to be true is not precisely defined by law.
+
+  If such an object file uses only numerical parameters, data
+structure layouts and accessors, and small macros and small inline
+functions (ten lines or less in length), then the use of the object
+file is unrestricted, regardless of whether it is legally a derivative
+work.  (Executables containing this object code plus portions of the
+Library will still fall under Section 6.)
+
+  Otherwise, if the work is a derivative of the Library, you may
+distribute the object code for the work under the terms of Section 6.
+Any executables containing that work also fall under Section 6,
+whether or not they are linked directly with the Library itself.
+
+  6. As an exception to the Sections above, you may also combine or
+link a "work that uses the Library" with the Library to produce a
+work containing portions of the Library, and distribute that work
+under terms of your choice, provided that the terms permit
+modification of the work for the customer's own use and reverse
+engineering for debugging such modifications.
+
+  You must give prominent notice with each copy of the work that the
+Library is used in it and that the Library and its use are covered by
+this License.  You must supply a copy of this License.  If the work
+during execution displays copyright notices, you must include the
+copyright notice for the Library among them, as well as a reference
+directing the user to the copy of this License.  Also, you must do one
+of these things:
+
+    a) Accompany the work with the complete corresponding
+    machine-readable source code for the Library including whatever
+    changes were used in the work (which must be distributed under
+    Sections 1 and 2 above); and, if the work is an executable linked
+    with the Library, with the complete machine-readable "work that
+    uses the Library", as object code and/or source code, so that the
+    user can modify the Library and then relink to produce a modified
+    executable containing the modified Library.  (It is understood
+    that the user who changes the contents of definitions files in the
+    Library will not necessarily be able to recompile the application
+    to use the modified definitions.)
+
+    b) Use a suitable shared library mechanism for linking with the
+    Library.  A suitable mechanism is one that (1) uses at run time a
+    copy of the library already present on the user's computer system,
+    rather than copying library functions into the executable, and (2)
+    will operate properly with a modified version of the library, if
+    the user installs one, as long as the modified version is
+    interface-compatible with the version that the work was made with.
+
+    c) Accompany the work with a written offer, valid for at
+    least three years, to give the same user the materials
+    specified in Subsection 6a, above, for a charge no more
+    than the cost of performing this distribution.
+
+    d) If distribution of the work is made by offering access to copy
+    from a designated place, offer equivalent access to copy the above
+    specified materials from the same place.
+
+    e) Verify that the user has already received a copy of these
+    materials or that you have already sent this user a copy.
+
+  For an executable, the required form of the "work that uses the
+Library" must include any data and utility programs needed for
+reproducing the executable from it.  However, as a special exception,
+the materials to be distributed need not include anything that is
+normally distributed (in either source or binary form) with the major
+components (compiler, kernel, and so on) of the operating system on
+which the executable runs, unless that component itself accompanies
+the executable.
+
+  It may happen that this requirement contradicts the license
+restrictions of other proprietary libraries that do not normally
+accompany the operating system.  Such a contradiction means you cannot
+use both them and the Library together in an executable that you
+distribute.
+
+  7. You may place library facilities that are a work based on the
+Library side-by-side in a single library together with other library
+facilities not covered by this License, and distribute such a combined
+library, provided that the separate distribution of the work based on
+the Library and of the other library facilities is otherwise
+permitted, and provided that you do these two things:
+
+    a) Accompany the combined library with a copy of the same work
+    based on the Library, uncombined with any other library
+    facilities.  This must be distributed under the terms of the
+    Sections above.
+
+    b) Give prominent notice with the combined library of the fact
+    that part of it is a work based on the Library, and explaining
+    where to find the accompanying uncombined form of the same work.
+
+  8. You may not copy, modify, sublicense, link with, or distribute
+the Library except as expressly provided under this License.  Any
+attempt otherwise to copy, modify, sublicense, link with, or
+distribute the Library is void, and will automatically terminate your
+rights under this License.  However, parties who have received copies,
+or rights, from you under this License will not have their licenses
+terminated so long as such parties remain in full compliance.
+
+  9. You are not required to accept this License, since you have not
+signed it.  However, nothing else grants you permission to modify or
+distribute the Library or its derivative works.  These actions are
+prohibited by law if you do not accept this License.  Therefore, by
+modifying or distributing the Library (or any work based on the
+Library), you indicate your acceptance of this License to do so, and
+all its terms and conditions for copying, distributing or modifying
+the Library or works based on it.
+
+  10. Each time you redistribute the Library (or any work based on the
+Library), the recipient automatically receives a license from the
+original licensor to copy, distribute, link with or modify the Library
+subject to these terms and conditions.  You may not impose any further
+restrictions on the recipients' exercise of the rights granted herein.
+You are not responsible for enforcing compliance by third parties with
+this License.
+
+  11. If, as a consequence of a court judgment or allegation of patent
+infringement or for any other reason (not limited to patent issues),
+conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot
+distribute so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you
+may not distribute the Library at all.  For example, if a patent
+license would not permit royalty-free redistribution of the Library by
+all those who receive copies directly or indirectly through you, then
+the only way you could satisfy both it and this License would be to
+refrain entirely from distribution of the Library.
+
+If any portion of this section is held invalid or unenforceable under any
+particular circumstance, the balance of the section is intended to apply,
+and the section as a whole is intended to apply in other circumstances.
+
+It is not the purpose of this section to induce you to infringe any
+patents or other property right claims or to contest validity of any
+such claims; this section has the sole purpose of protecting the
+integrity of the free software distribution system which is
+implemented by public license practices.  Many people have made
+generous contributions to the wide range of software distributed
+through that system in reliance on consistent application of that
+system; it is up to the author/donor to decide if he or she is willing
+to distribute software through any other system and a licensee cannot
+impose that choice.
+
+This section is intended to make thoroughly clear what is believed to
+be a consequence of the rest of this License.
+
+  12. If the distribution and/or use of the Library is restricted in
+certain countries either by patents or by copyrighted interfaces, the
+original copyright holder who places the Library under this License may add
+an explicit geographical distribution limitation excluding those countries,
+so that distribution is permitted only in or among countries not thus
+excluded.  In such case, this License incorporates the limitation as if
+written in the body of this License.
+
+  13. The Free Software Foundation may publish revised and/or new
+versions of the Lesser General Public License from time to time.
+Such new versions will be similar in spirit to the present version,
+but may differ in detail to address new problems or concerns.
+
+Each version is given a distinguishing version number.  If the Library
+specifies a version number of this License which applies to it and
+"any later version", you have the option of following the terms and
+conditions either of that version or of any later version published by
+the Free Software Foundation.  If the Library does not specify a
+license version number, you may choose any version ever published by
+the Free Software Foundation.
+
+  14. If you wish to incorporate parts of the Library into other free
+programs whose distribution conditions are incompatible with these,
+write to the author to ask for permission.  For software which is
+copyrighted by the Free Software Foundation, write to the Free
+Software Foundation; we sometimes make exceptions for this.  Our
+decision will be guided by the two goals of preserving the free status
+of all derivatives of our free software and of promoting the sharing
+and reuse of software generally.
+
+			    NO WARRANTY
+
+  15. BECAUSE THE LIBRARY IS LICENSED FREE OF CHARGE, THERE IS NO
+WARRANTY FOR THE LIBRARY, TO THE EXTENT PERMITTED BY APPLICABLE LAW.
+EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR
+OTHER PARTIES PROVIDE THE LIBRARY "AS IS" WITHOUT WARRANTY OF ANY
+KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE
+LIBRARY IS WITH YOU.  SHOULD THE LIBRARY PROVE DEFECTIVE, YOU ASSUME
+THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN
+WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY
+AND/OR REDISTRIBUTE THE LIBRARY AS PERMITTED ABOVE, BE LIABLE TO YOU
+FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR
+CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE
+LIBRARY (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING
+RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A
+FAILURE OF THE LIBRARY TO OPERATE WITH ANY OTHER SOFTWARE), EVEN IF
+SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
+DAMAGES.
+
+		     END OF TERMS AND CONDITIONS
+
+           How to Apply These Terms to Your New Libraries
+
+  If you develop a new library, and you want it to be of the greatest
+possible use to the public, we recommend making it free software that
+everyone can redistribute and change.  You can do so by permitting
+redistribution under these terms (or, alternatively, under the terms of the
+ordinary General Public License).
+
+  To apply these terms, attach the following notices to the library.  It is
+safest to attach them to the start of each source file to most effectively
+convey the exclusion of warranty; and each file should have at least the
+"copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the library's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This library is free software; you can redistribute it and/or
+    modify it under the terms of the GNU Lesser General Public
+    License as published by the Free Software Foundation; either
+    version 2.1 of the License, or (at your option) any later version.
+
+    This library is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+    Lesser General Public License for more details.
+
+    You should have received a copy of the GNU Lesser General Public
+    License along with this library; if not, write to the Free Software
+    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+
+Also add information on how to contact you by electronic and paper mail.
+
+You should also get your employer (if you work as a programmer) or your
+school, if any, to sign a "copyright disclaimer" for the library, if
+necessary.  Here is a sample; alter the names:
+
+  Yoyodyne, Inc., hereby disclaims all copyright interest in the
+  library `Frob' (a library for tweaking knobs) written by James Random Hacker.
+
+  <signature of Ty Coon>, 1 April 1990
+  Ty Coon, President of Vice
+
+That's all there is to it!
+
+
diff --git a/LICENSE.txt b/LICENSE.txt
new file mode 100644
index 0000000..fa803b1
--- /dev/null
+++ b/LICENSE.txt
@@ -0,0 +1,53 @@
+                wxWindows Library Licence, Version 3.1
+                ======================================
+
+  Copyright (C) 1998-2005 Julian Smart, Robert Roebling et al
+
+  Everyone is permitted to copy and distribute verbatim copies
+  of this licence document, but changing it is not allowed.
+
+                       WXWINDOWS LIBRARY LICENCE
+     TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+  
+  This library is free software; you can redistribute it and/or modify it
+  under the terms of the GNU Library General Public Licence as published by
+  the Free Software Foundation; either version 2 of the Licence, or (at
+  your option) any later version.
+  
+  This library is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Library
+  General Public Licence for more details.
+
+  You should have received a copy of the GNU Library General Public Licence
+  along with this software, usually in a file named COPYING.LIB.  If not,
+  write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330,
+  Boston, MA 02111-1307 USA.
+
+  EXCEPTION NOTICE
+
+  1. As a special exception, the copyright holders of this library give
+  permission for additional uses of the text contained in this release of
+  the library as licenced under the wxWindows Library Licence, applying
+  either version 3.1 of the Licence, or (at your option) any later version of
+  the Licence as published by the copyright holders of version
+  3.1 of the Licence document.
+
+  2. The exception is that you may use, copy, link, modify and distribute
+  under your own terms, binary object code versions of works based
+  on the Library.
+
+  3. If you copy code from files distributed under the terms of the GNU
+  General Public Licence or the GNU Library General Public Licence into a
+  copy of this library, as this licence permits, the exception does not
+  apply to the code that you add in this way.  To avoid misleading anyone as
+  to the status of such modified files, you must delete this exception
+  notice from such code and/or adjust the licensing conditions notice
+  accordingly.
+
+  4. If you write modifications of your own for this library, it is your
+  choice whether to permit this exception to apply to your modifications. 
+  If you do not wish that, you must delete the exception notice from such
+  code and/or adjust the licensing conditions notice accordingly.
+
+
diff --git a/Makefile.am b/Makefile.am
new file mode 100644
index 0000000..adc4cb6
--- /dev/null
+++ b/Makefile.am
@@ -0,0 +1,225 @@
+lib_LTLIBRARIES = libjpeg.la libturbojpeg.la
+libjpeg_la_LDFLAGS = -version-info ${SO_MAJOR_VERSION}:${SO_MINOR_VERSION} -no-undefined
+libturbojpeg_la_LDFLAGS = -avoid-version -no-undefined
+include_HEADERS = jerror.h jmorecfg.h jpeglib.h turbojpeg.h
+nodist_include_HEADERS = jconfig.h
+
+HDRS = jchuff.h jdct.h jdhuff.h jerror.h jinclude.h jmemsys.h jmorecfg.h \
+	jpegint.h jpeglib.h jversion.h jsimd.h jsimddct.h jpegcomp.h
+
+libjpeg_la_SOURCES = $(HDRS) jcapimin.c jcapistd.c jccoefct.c jccolor.c \
+	jcdctmgr.c jchuff.c jcinit.c jcmainct.c jcmarker.c jcmaster.c \
+	jcomapi.c jcparam.c jcphuff.c jcprepct.c jcsample.c jctrans.c \
+	jdapimin.c jdapistd.c jdatadst.c jdatasrc.c jdcoefct.c jdcolor.c \
+	jddctmgr.c jdhuff.c jdinput.c jdmainct.c jdmarker.c jdmaster.c \
+	jdmerge.c jdphuff.c jdpostct.c jdsample.c jdtrans.c jerror.c \
+	jfdctflt.c jfdctfst.c jfdctint.c jidctflt.c jidctfst.c jidctint.c \
+	jidctred.c jquant1.c jquant2.c jutils.c jmemmgr.c jmemnobs.c
+
+libturbojpeg_la_SOURCES = $(libjpeg_la_SOURCES) turbojpegl.c turbojpeg.h \
+	turbojpeg-mapfile
+
+if ANON_VERSION_SCRIPT
+
+libturbojpeg_la_LDFLAGS += $(ANON_VERSION_SCRIPT_FLAG)$(srcdir)/turbojpeg-mapfile
+
+endif
+
+if VERSION_SCRIPT
+
+libjpeg_la_LDFLAGS += $(VERSION_SCRIPT_FLAG)libjpeg.map
+
+endif
+
+if WITH_SIMD
+
+SUBDIRS = simd
+libjpeg_la_LIBADD = simd/libsimd.la
+libturbojpeg_la_LIBADD = simd/libsimd.la
+
+else
+
+libjpeg_la_SOURCES += jsimd_none.c
+
+endif
+
+TSTHDRS = rrutil.h rrtimer.h
+
+bin_PROGRAMS = cjpeg djpeg jpegtran rdjpgcom wrjpgcom
+noinst_PROGRAMS = jpgtest jpegut
+
+jpgtest_SOURCES = $(TSTHDRS) jpgtest.cxx bmp.h bmp.c
+
+jpgtest_LDADD = libturbojpeg.la
+
+jpegut_SOURCES = $(TSTHDRS) jpegut.c bmp.h bmp.c
+
+jpegut_LDADD = libturbojpeg.la
+
+cjpeg_SOURCES = cdjpeg.h cderror.h cdjpeg.c cjpeg.c rdbmp.c rdgif.c \
+	rdppm.c rdswitch.c rdtarga.c 
+
+cjpeg_LDADD = libjpeg.la
+
+cjpeg_CFLAGS = -DBMP_SUPPORTED -DGIF_SUPPORTED -DPPM_SUPPORTED \
+	-DTARGA_SUPPORTED
+
+djpeg_SOURCES = cdjpeg.h cderror.h cdjpeg.c djpeg.c rdcolmap.c rdswitch.c \
+	wrbmp.c wrgif.c wrppm.c wrtarga.c
+
+djpeg_LDADD = libjpeg.la
+
+djpeg_CFLAGS = -DBMP_SUPPORTED -DGIF_SUPPORTED -DPPM_SUPPORTED \
+	-DTARGA_SUPPORTED
+
+jpegtran_SOURCES = jpegtran.c rdswitch.c cdjpeg.c transupp.c transupp.h
+
+jpegtran_LDADD = libjpeg.la
+
+rdjpgcom_SOURCES = rdjpgcom.c
+
+rdjpgcom_LDADD = libjpeg.la
+
+wrjpgcom_SOURCES = wrjpgcom.c
+
+wrjpgcom_LDADD = libjpeg.la
+
+
+dist_man1_MANS = cjpeg.1 djpeg.1 jpegtran.1 rdjpgcom.1 wrjpgcom.1
+
+DOCS= README install.txt usage.txt wizard.txt example.c libjpeg.txt \
+	structure.txt coderules.txt filelist.txt jconfig.txt change.log \
+	README-turbo.txt rdrle.c wrrle.c LICENSE.txt LGPL.txt BUILDING.txt \
+	ChangeLog.txt
+
+TESTFILES= testorig.jpg testorig.ppm testimg.bmp testimgflt.jpg \
+	testimgfst.jpg testimgint.jpg testimgp.jpg testimgflt.ppm testimgfst.ppm \
+	testimgint.ppm testimgflt-nosimd.jpg testimgcrop.jpg
+
+EXTRA_DIST = win release $(DOCS) $(TESTFILES) CMakeLists.txt \
+	sharedlib/CMakeLists.txt cmakescripts libjpeg.map.in
+
+dist-hook:
+	rm -rf `find $(distdir) -name .svn`
+
+
+if WITH_SIMD
+
+test: testclean all
+	./jpegut
+	./cjpeg -dct int -outfile testoutint.jpg $(srcdir)/testorig.ppm
+	./cjpeg -dct fast -opt -outfile testoutfst.jpg $(srcdir)/testorig.ppm
+	./cjpeg -dct float -outfile testoutflt.jpg $(srcdir)/testorig.ppm
+	cmp $(srcdir)/testimgint.jpg testoutint.jpg
+	cmp $(srcdir)/testimgfst.jpg testoutfst.jpg
+	cmp $(srcdir)/testimgflt.jpg testoutflt.jpg
+	./djpeg -dct int -fast -ppm -outfile testoutint.ppm $(srcdir)/testorig.jpg
+	./djpeg -dct fast -ppm -outfile testoutfst.ppm $(srcdir)/testorig.jpg
+	./djpeg -dct float -ppm -outfile testoutflt.ppm $(srcdir)/testorig.jpg
+	cmp $(srcdir)/testimgint.ppm testoutint.ppm
+	cmp $(srcdir)/testimgfst.ppm testoutfst.ppm
+	cmp $(srcdir)/testimgflt.ppm testoutflt.ppm
+	./djpeg -dct int -bmp -colors 256 -outfile testout.bmp  $(srcdir)/testorig.jpg
+	cmp $(srcdir)/testimg.bmp testout.bmp
+	./cjpeg -dct int -progressive -outfile testoutp.jpg $(srcdir)/testorig.ppm
+	cmp $(srcdir)/testimgp.jpg testoutp.jpg
+	./jpegtran -outfile testoutt.jpg testoutp.jpg
+	cmp $(srcdir)/testimgint.jpg testoutt.jpg
+	./jpegtran -crop 120x90+20+50 -transpose -perfect -outfile testoutcrop.jpg $(srcdir)/testorig.jpg
+	cmp $(srcdir)/testimgcrop.jpg testoutcrop.jpg
+
+else
+
+test: testclean all
+	./jpegut
+	./cjpeg -dct int -outfile testoutint.jpg $(srcdir)/testorig.ppm
+	./cjpeg -dct fast -opt -outfile testoutfst.jpg $(srcdir)/testorig.ppm
+	./cjpeg -dct float -outfile testoutflt.jpg $(srcdir)/testorig.ppm
+	cmp $(srcdir)/testimgint.jpg testoutint.jpg
+	cmp $(srcdir)/testimgfst.jpg testoutfst.jpg
+	cmp $(srcdir)/testimgflt-nosimd.jpg testoutflt.jpg
+	./djpeg -dct int -fast -ppm -outfile testoutint.ppm $(srcdir)/testorig.jpg
+	./djpeg -dct fast -ppm -outfile testoutfst.ppm $(srcdir)/testorig.jpg
+	./djpeg -dct float -ppm -outfile testoutflt.ppm $(srcdir)/testorig.jpg
+	cmp $(srcdir)/testimgint.ppm testoutint.ppm
+	cmp $(srcdir)/testimgfst.ppm testoutfst.ppm
+	cmp $(srcdir)/testorig.ppm testoutflt.ppm
+	./djpeg -dct int -bmp -colors 256 -outfile testout.bmp  $(srcdir)/testorig.jpg
+	cmp $(srcdir)/testimg.bmp testout.bmp
+	./cjpeg -dct int -progressive -outfile testoutp.jpg $(srcdir)/testorig.ppm
+	cmp $(srcdir)/testimgp.jpg testoutp.jpg
+	./jpegtran -outfile testoutt.jpg testoutp.jpg
+	cmp $(srcdir)/testimgint.jpg testoutt.jpg
+	./jpegtran -crop 120x90+20+50 -transpose -perfect -outfile testoutcrop.jpg $(srcdir)/testorig.jpg
+	cmp $(srcdir)/testimgcrop.jpg testoutcrop.jpg
+
+endif
+
+testclean:
+	rm -f testout*
+	rm -f *_GRAYQ[0-9]*.bmp
+	rm -f *_GRAYQ[0-9]*.ppm
+	rm -f *_GRAYQ[0-9]*.jpg
+	rm -f *_420Q[0-9]*.bmp
+	rm -f *_420Q[0-9]*.ppm
+	rm -f *_420Q[0-9]*.jpg
+	rm -f *_422Q[0-9]*.bmp
+	rm -f *_422Q[0-9]*.ppm
+	rm -f *_422Q[0-9]*.jpg
+	rm -f *_444Q[0-9]*.bmp
+	rm -f *_444Q[0-9]*.ppm
+	rm -f *_444Q[0-9]*.jpg
+
+rpm: all
+	TMPDIR=`mktemp -d /tmp/${PACKAGE_NAME}-build.XXXXXX`; \
+	mkdir -p $$TMPDIR/RPMS; \
+	ln -fs `pwd` $$TMPDIR/BUILD; \
+	rm -f ${PACKAGE_NAME}.${RPMARCH}.rpm; \
+	rpmbuild -bb --define "_blddir $$TMPDIR/buildroot"  \
+		--define "_topdir $$TMPDIR" --define "_srcdir ${srcdir}" \
+		--target ${RPMARCH} libjpeg-turbo.spec; \
+	cp $$TMPDIR/RPMS/${RPMARCH}/${PACKAGE_NAME}-${VERSION}-${BUILD}.${RPMARCH}.rpm ${PACKAGE_NAME}.${RPMARCH}.rpm; \
+	rm -rf $$TMPDIR
+
+srpm: dist-gzip
+	TMPDIR=`mktemp -d /tmp/${PACKAGE_NAME}-build.XXXXXX`; \
+	mkdir -p $$TMPDIR/RPMS; \
+	mkdir -p $$TMPDIR/SRPMS; \
+	mkdir -p $$TMPDIR/BUILD; \
+	mkdir -p $$TMPDIR/SOURCES; \
+	mkdir -p $$TMPDIR/SPECS; \
+	rm -f ${PACKAGE_NAME}.src.rpm; \
+	cp ${PACKAGE_NAME}-${VERSION}.tar.gz $$TMPDIR/SOURCES; \
+	cat libjpeg-turbo.spec | sed s/%{_blddir}/%{_tmppath}/g \
+		| sed s@%{_srcdir}/@@g | sed s/#--\>//g \
+		>$$TMPDIR/SPECS/libjpeg-turbo.spec; \
+	rpmbuild -bs --define "_topdir $$TMPDIR" $$TMPDIR/SPECS/libjpeg-turbo.spec; \
+	cp $$TMPDIR/SRPMS/${PACKAGE_NAME}-${VERSION}-${BUILD}.src.rpm ${PACKAGE_NAME}.src.rpm; \
+	rm -rf $$TMPDIR
+
+deb: all
+	sh $(srcdir)/release/makedpkg ${PACKAGE_NAME} ${VERSION} ${BUILD} \
+		${DEBARCH} ${srcdir}
+
+if X86_64
+
+udmg: all
+	sh makemacpkg universal ${BUILDDIR32}
+
+endif
+
+dmg: all
+	sh makemacpkg
+
+if X86_64
+
+csunpkg: all
+	sh makesunpkg combined ${BUILDDIR32}
+
+endif
+
+sunpkg: all
+	sh makesunpkg
+
+cygwinpkg: all
+	sh $(srcdir)/release/makecygwinpkg ${PACKAGE_NAME} ${VERSION} ${srcdir}
diff --git a/README b/README
new file mode 100644
index 0000000..921095d
--- /dev/null
+++ b/README
@@ -0,0 +1,303 @@
+libjpeg-turbo note:  This file is mostly taken from the libjpeg v8b README
+file, and it is included only for reference.  Some parts of it may not apply to
+libjpeg-turbo.  Please see README-turbo.txt for information specific to the
+turbo version.
+
+
+The Independent JPEG Group's JPEG software
+==========================================
+
+This distribution contains a release of the Independent JPEG Group's free JPEG
+software.  You are welcome to redistribute this software and to use it for any
+purpose, subject to the conditions under LEGAL ISSUES, below.
+
+This software is the work of Tom Lane, Guido Vollbeding, Philip Gladstone,
+Bill Allombert, Jim Boucher, Lee Crocker, Bob Friesenhahn, Ben Jackson,
+Julian Minguillon, Luis Ortiz, George Phillips, Davide Rossi, Ge' Weijers,
+and other members of the Independent JPEG Group.
+
+IJG is not affiliated with the official ISO JPEG standards committee.
+
+
+DOCUMENTATION ROADMAP
+=====================
+
+This file contains the following sections:
+
+OVERVIEW            General description of JPEG and the IJG software.
+LEGAL ISSUES        Copyright, lack of warranty, terms of distribution.
+REFERENCES          Where to learn more about JPEG.
+ARCHIVE LOCATIONS   Where to find newer versions of this software.
+FILE FORMAT WARS    Software *not* to get.
+TO DO               Plans for future IJG releases.
+
+Other documentation files in the distribution are:
+
+User documentation:
+  install.txt       How to configure and install the IJG software.
+  usage.txt         Usage instructions for cjpeg, djpeg, jpegtran,
+                    rdjpgcom, and wrjpgcom.
+  *.1               Unix-style man pages for programs (same info as usage.txt).
+  wizard.txt        Advanced usage instructions for JPEG wizards only.
+  change.log        Version-to-version change highlights.
+Programmer and internal documentation:
+  libjpeg.txt       How to use the JPEG library in your own programs.
+  example.c         Sample code for calling the JPEG library.
+  structure.txt     Overview of the JPEG library's internal structure.
+  filelist.txt      Road map of IJG files.
+  coderules.txt     Coding style rules --- please read if you contribute code.
+
+Please read at least the files install.txt and usage.txt.  Some information
+can also be found in the JPEG FAQ (Frequently Asked Questions) article.  See
+ARCHIVE LOCATIONS below to find out where to obtain the FAQ article.
+
+If you want to understand how the JPEG code works, we suggest reading one or
+more of the REFERENCES, then looking at the documentation files (in roughly
+the order listed) before diving into the code.
+
+
+OVERVIEW
+========
+
+This package contains C software to implement JPEG image encoding, decoding,
+and transcoding.  JPEG (pronounced "jay-peg") is a standardized compression
+method for full-color and gray-scale images.  JPEG's strong suit is compressing
+photographic images or other types of images which have smooth color and
+brightness transitions between neighboring pixels.  Images with sharp lines or
+other abrupt features may not compress well with JPEG, and a higher JPEG
+quality may have to be used to avoid visible compression artifacts with such
+images.
+
+JPEG is lossy, meaning that the output pixels are not necessarily identical to
+the input pixels.  However, on photographic content and other "smooth" images,
+very good compression ratios can be obtained with no visible compression
+artifacts, and extremely high compression ratios are possible if you are
+willing to sacrifice image quality (by reducing the "quality" setting in the
+compressor.)
+
+This software implements JPEG baseline, extended-sequential, and progressive
+compression processes.  Provision is made for supporting all variants of these
+processes, although some uncommon parameter settings aren't implemented yet.
+For legal reasons, we are not distributing code for the arithmetic-coding
+variants of JPEG; see LEGAL ISSUES.  We have made no provision for supporting
+the hierarchical or lossless processes defined in the standard.
+
+We provide a set of library routines for reading and writing JPEG image files,
+plus two sample applications "cjpeg" and "djpeg", which use the library to
+perform conversion between JPEG and some other popular image file formats.
+The library is intended to be reused in other applications.
+
+In order to support file conversion and viewing software, we have included
+considerable functionality beyond the bare JPEG coding/decoding capability;
+for example, the color quantization modules are not strictly part of JPEG
+decoding, but they are essential for output to colormapped file formats or
+colormapped displays.  These extra functions can be compiled out of the
+library if not required for a particular application.
+
+We have also included "jpegtran", a utility for lossless transcoding between
+different JPEG processes, and "rdjpgcom" and "wrjpgcom", two simple
+applications for inserting and extracting textual comments in JFIF files.
+
+The emphasis in designing this software has been on achieving portability and
+flexibility, while also making it fast enough to be useful.  In particular,
+the software is not intended to be read as a tutorial on JPEG.  (See the
+REFERENCES section for introductory material.)  Rather, it is intended to
+be reliable, portable, industrial-strength code.  We do not claim to have
+achieved that goal in every aspect of the software, but we strive for it.
+
+We welcome the use of this software as a component of commercial products.
+No royalty is required, but we do ask for an acknowledgement in product
+documentation, as described under LEGAL ISSUES.
+
+
+LEGAL ISSUES
+============
+
+In plain English:
+
+1. We don't promise that this software works.  (But if you find any bugs,
+   please let us know!)
+2. You can use this software for whatever you want.  You don't have to pay us.
+3. You may not pretend that you wrote this software.  If you use it in a
+   program, you must acknowledge somewhere in your documentation that
+   you've used the IJG code.
+
+In legalese:
+
+The authors make NO WARRANTY or representation, either express or implied,
+with respect to this software, its quality, accuracy, merchantability, or
+fitness for a particular purpose.  This software is provided "AS IS", and you,
+its user, assume the entire risk as to its quality and accuracy.
+
+This software is copyright (C) 1991-2010, Thomas G. Lane, Guido Vollbeding.
+All Rights Reserved except as specified below.
+
+Permission is hereby granted to use, copy, modify, and distribute this
+software (or portions thereof) for any purpose, without fee, subject to these
+conditions:
+(1) If any part of the source code for this software is distributed, then this
+README file must be included, with this copyright and no-warranty notice
+unaltered; and any additions, deletions, or changes to the original files
+must be clearly indicated in accompanying documentation.
+(2) If only executable code is distributed, then the accompanying
+documentation must state that "this software is based in part on the work of
+the Independent JPEG Group".
+(3) Permission for use of this software is granted only if the user accepts
+full responsibility for any undesirable consequences; the authors accept
+NO LIABILITY for damages of any kind.
+
+These conditions apply to any software derived from or based on the IJG code,
+not just to the unmodified library.  If you use our work, you ought to
+acknowledge us.
+
+Permission is NOT granted for the use of any IJG author's name or company name
+in advertising or publicity relating to this software or products derived from
+it.  This software may be referred to only as "the Independent JPEG Group's
+software".
+
+We specifically permit and encourage the use of this software as the basis of
+commercial products, provided that all warranty or liability claims are
+assumed by the product vendor.
+
+
+ansi2knr.c is included in this distribution by permission of L. Peter Deutsch,
+sole proprietor of its copyright holder, Aladdin Enterprises of Menlo Park, CA.
+ansi2knr.c is NOT covered by the above copyright and conditions, but instead
+by the usual distribution terms of the Free Software Foundation; principally,
+that you must include source code if you redistribute it.  (See the file
+ansi2knr.c for full details.)  However, since ansi2knr.c is not needed as part
+of any program generated from the IJG code, this does not limit you more than
+the foregoing paragraphs do.
+
+The Unix configuration script "configure" was produced with GNU Autoconf.
+It is copyright by the Free Software Foundation but is freely distributable.
+The same holds for its supporting scripts (config.guess, config.sub,
+ltmain.sh).  Another support script, install-sh, is copyright by X Consortium
+but is also freely distributable.
+
+It appears that the arithmetic coding option of the JPEG spec is covered by
+patents owned by IBM, AT&T, and Mitsubishi.  Hence arithmetic coding cannot
+legally be used without obtaining one or more licenses.  For this reason,
+support for arithmetic coding has been removed from the free JPEG software.
+(Since arithmetic coding provides only a marginal gain over the unpatented
+Huffman mode, it is unlikely that very many implementations will support it.)
+So far as we are aware, there are no patent restrictions on the remaining
+code.
+
+The IJG distribution formerly included code to read and write GIF files.
+To avoid entanglement with the Unisys LZW patent, GIF reading support has
+been removed altogether, and the GIF writer has been simplified to produce
+"uncompressed GIFs".  This technique does not use the LZW algorithm; the
+resulting GIF files are larger than usual, but are readable by all standard
+GIF decoders.
+
+We are required to state that
+    "The Graphics Interchange Format(c) is the Copyright property of
+    CompuServe Incorporated.  GIF(sm) is a Service Mark property of
+    CompuServe Incorporated."
+
+
+REFERENCES
+==========
+
+We recommend reading one or more of these references before trying to
+understand the innards of the JPEG software.
+
+The best short technical introduction to the JPEG compression algorithm is
+	Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+	Communications of the ACM, April 1991 (vol. 34 no. 4), pp. 30-44.
+(Adjacent articles in that issue discuss MPEG motion picture compression,
+applications of JPEG, and related topics.)  If you don't have the CACM issue
+handy, a PostScript file containing a revised version of Wallace's article is
+available at http://www.ijg.org/files/wallace.ps.gz.  The file (actually
+a preprint for an article that appeared in IEEE Trans. Consumer Electronics)
+omits the sample images that appeared in CACM, but it includes corrections
+and some added material.  Note: the Wallace article is copyright ACM and IEEE,
+and it may not be used for commercial purposes.
+
+A somewhat less technical, more leisurely introduction to JPEG can be found in
+"The Data Compression Book" by Mark Nelson and Jean-loup Gailly, published by
+M&T Books (New York), 2nd ed. 1996, ISBN 1-55851-434-1.  This book provides
+good explanations and example C code for a multitude of compression methods
+including JPEG.  It is an excellent source if you are comfortable reading C
+code but don't know much about data compression in general.  The book's JPEG
+sample code is far from industrial-strength, but when you are ready to look
+at a full implementation, you've got one here...
+
+The best currently available description of JPEG is the textbook "JPEG Still
+Image Data Compression Standard" by William B. Pennebaker and Joan L.
+Mitchell, published by Van Nostrand Reinhold, 1993, ISBN 0-442-01272-1.
+Price US$59.95, 638 pp.  The book includes the complete text of the ISO JPEG
+standards (DIS 10918-1 and draft DIS 10918-2).
+Although this is by far the most detailed and comprehensive exposition of
+JPEG publicly available, we point out that it is still missing an explanation
+of the most essential properties and algorithms of the underlying DCT
+technology.
+
+The original JPEG standard is divided into two parts, Part 1 being the actual
+specification, while Part 2 covers compliance testing methods.  Part 1 is
+titled "Digital Compression and Coding of Continuous-tone Still Images,
+Part 1: Requirements and guidelines" and has document numbers ISO/IEC IS
+10918-1, ITU-T T.81.  Part 2 is titled "Digital Compression and Coding of
+Continuous-tone Still Images, Part 2: Compliance testing" and has document
+numbers ISO/IEC IS 10918-2, ITU-T T.83.
+
+The JPEG standard does not specify all details of an interchangeable file
+format.  For the omitted details we follow the "JFIF" conventions, revision
+1.02.  JFIF 1.02 has been adopted as an Ecma International Technical Report
+and thus received a formal publication status.  It is available as a free
+download in PDF format from
+http://www.ecma-international.org/publications/techreports/E-TR-098.htm.
+A PostScript version of the JFIF document is available at
+http://www.ijg.org/files/jfif.ps.gz.  There is also a plain text version at
+http://www.ijg.org/files/jfif.txt.gz, but it is missing the figures.
+
+The TIFF 6.0 file format specification can be obtained by FTP from
+ftp://ftp.sgi.com/graphics/tiff/TIFF6.ps.gz.  The JPEG incorporation scheme
+found in the TIFF 6.0 spec of 3-June-92 has a number of serious problems.
+IJG does not recommend use of the TIFF 6.0 design (TIFF Compression tag 6).
+Instead, we recommend the JPEG design proposed by TIFF Technical Note #2
+(Compression tag 7).  Copies of this Note can be obtained from
+http://www.ijg.org/files/.  It is expected that the next revision
+of the TIFF spec will replace the 6.0 JPEG design with the Note's design.
+Although IJG's own code does not support TIFF/JPEG, the free libtiff library
+uses our library to implement TIFF/JPEG per the Note.
+
+
+ARCHIVE LOCATIONS
+=================
+
+The "official" archive site for this software is www.ijg.org.
+The most recent released version can always be found there in
+directory "files".  This particular version will be archived as
+http://www.ijg.org/files/jpegsrc.v8b.tar.gz, and in Windows-compatible
+"zip" archive format as http://www.ijg.org/files/jpegsr8b.zip.
+
+The JPEG FAQ (Frequently Asked Questions) article is a source of some
+general information about JPEG.
+It is available on the World Wide Web at http://www.faqs.org/faqs/jpeg-faq/
+and other news.answers archive sites, including the official news.answers
+archive at rtfm.mit.edu: ftp://rtfm.mit.edu/pub/usenet/news.answers/jpeg-faq/.
+If you don't have Web or FTP access, send e-mail to mail-server@rtfm.mit.edu
+with body
+	send usenet/news.answers/jpeg-faq/part1
+	send usenet/news.answers/jpeg-faq/part2
+
+
+FILE FORMAT WARS
+================
+
+The ISO JPEG standards committee actually promotes different formats like
+"JPEG 2000" or "JPEG XR" which are incompatible with original DCT-based
+JPEG.  IJG therefore does not support these formats (see REFERENCES).  Indeed,
+one of the original reasons for developing this free software was to help
+force convergence on common, interoperable format standards for JPEG files.
+Don't use an incompatible file format!
+(In any case, our decoder will remain capable of reading existing JPEG
+image files indefinitely.)
+
+
+TO DO
+=====
+
+Please send bug reports, offers of help, etc. to jpeg-info@uc.ag.
diff --git a/README-turbo.txt b/README-turbo.txt
new file mode 100755
index 0000000..1dc58c6
--- /dev/null
+++ b/README-turbo.txt
@@ -0,0 +1,277 @@
+*******************************************************************************
+**     Background
+*******************************************************************************
+
+libjpeg-turbo is a high-speed version of libjpeg for x86 and x86-64 processors
+which uses SIMD instructions (MMX, SSE2, etc.) to accelerate baseline JPEG
+compression and decompression.  libjpeg-turbo is generally 2-4x as fast
+as the unmodified version of libjpeg, all else being equal.
+
+libjpeg-turbo was originally based on libjpeg/SIMD by Miyasaka Masaru, but
+the TigerVNC and VirtualGL projects made numerous enhancements to the codec,
+including improved support for Mac OS X, 64-bit support, support for 32-bit
+and big endian pixel formats, accelerated Huffman encoding/decoding, and
+various bug fixes.  The goal was to produce a fully open source codec that
+could replace the partially closed source TurboJPEG/IPP codec used by VirtualGL
+and TurboVNC.  libjpeg-turbo generally performs in the range of 80-120% of
+TurboJPEG/IPP.  It is faster in some areas but slower in others.
+
+It was decided to split libjpeg-turbo into a separate SDK so that other
+projects could take advantage of this technology.  The libjpeg-turbo shared
+libraries can be used as drop-in replacements for libjpeg on most systems.
+
+
+*******************************************************************************
+**     License
+*******************************************************************************
+
+Some of the optimizations to the Huffman encoder (jchuff.c) and decoder
+(jdhuff.c) were borrowed from VirtualGL, and thus any distribution of
+libjpeg-turbo which includes those optimizations must, as a whole, be subject
+to the terms of the wxWindows Library Licence, Version 3.1.  A copy of this
+license can be found in this directory under LICENSE.txt.  The wxWindows
+Library License is based on the LGPL but includes provisions which allow the
+Library to be statically linked into proprietary libraries and applications
+without requiring the resulting binaries to be distributed under the terms of
+the LGPL.
+
+The rest of the source code, apart from the Huffman codec optimizations, falls
+under a less restrictive, BSD-style license (see README.)  You can choose to
+distribute libjpeg-turbo, as a whole, under this BSD-style license by simply
+replacing the optimized jchuff.c and jdhuff.c with their unoptimized
+counterparts from the libjpeg v6b source.
+
+
+*******************************************************************************
+**     Using libjpeg-turbo
+*******************************************************************************
+
+=============================
+Replacing libjpeg at Run Time
+=============================
+
+If a Unix application is dynamically linked with libjpeg, then you can replace
+libjpeg with libjpeg-turbo at run time by manipulating the LD_LIBRARY_PATH.
+For instance:
+
+  [Using libjpeg]
+  > time cjpeg <vgl_5674_0098.ppm >vgl_5674_0098.jpg
+  real  0m0.392s
+  user  0m0.074s
+  sys   0m0.020s
+
+  [Using libjpeg-turbo]
+  > export LD_LIBRARY_PATH=/opt/libjpeg-turbo/{lib}:$LD_LIBRARY_PATH
+  > time cjpeg <vgl_5674_0098.ppm >vgl_5674_0098.jpg
+  real  0m0.109s
+  user  0m0.029s
+  sys   0m0.010s
+
+NOTE: {lib} can be lib, lib32, lib64, or lib/64, depending on the O/S and
+architecture.
+
+System administrators can also replace the libjpeg sym links in /usr/{lib} with
+links to the libjpeg dynamic library located in /opt/libjpeg-turbo/{lib}.  This
+will effectively accelerate every dynamically linked libjpeg application on the
+system.
+
+The Windows distribution of the libjpeg-turbo SDK installs the libjpeg-turbo
+DLL (jpeg62.dll, jpeg7.dll, or jpeg8.dll, depending on whether libjpeg v6b,
+v7, or v8 emulation is enabled) into c:\libjpeg-turbo[64]\bin, and the PATH
+environment variable can be modified such that this directory is searched
+before any others that might contain a libjpeg DLL.  However, if a libjpeg
+DLL exists in an application's install directory, then Windows will load this
+DLL first whenever the application is launched.  Thus, if an application ships
+with jpeg62.dll, jpeg7.dll, or jpeg8.dll, then back up the application's
+version of this DLL and copy c:\libjpeg-turbo\bin\jpeg*.dll into the
+application's install directory to accelerate it.
+
+The version of the libjpeg-turbo DLL distributed with the "official"
+libjpeg-turbo SDK requires the Visual C++ 2008 C run time DLL (msvcr90.dll).
+msvcr90.dll ships with more recent versions of Windows, but users of older
+Windows releases can obtain it from the Visual C++ 2008 Redistributable
+Package, which is available as a free download from Microsoft's web site.
+
+NOTE:  Features of libjpeg which require passing a C run time structure, such
+as a file handle, from an application to libjpeg will probably not work with
+the "official" version of the libjpeg-turbo DLL unless the application is also
+built to use the Visual C++ 2008 C run time DLL.  In particular, this affects
+jpeg_stdio_dest() and jpeg_stdio_src().
+
+Mac applications typically embed their own copies of the libjpeg dylib inside
+the (hidden) application bundle, so it is not possible to globally replace
+libjpeg on OS X systems.  If an application uses a shared library version of
+libjpeg, then it may be possible to replace the application's version of it.
+This would generally involve copying libjpeg.*.dylib from libjpeg-turbo into
+the appropriate place in the application bundle and using install_name_tool to
+repoint the dylib to the new directory.  This requires an advanced knowledge of
+OS X and would not survive an upgrade or a re-install of the application.
+Thus, it is not recommended for most users.
+
+=======================
+Replacing TurboJPEG/IPP
+=======================
+
+libjpeg-turbo is a drop-in replacement for the TurboJPEG/IPP SDK used by
+VirtualGL 2.1.x and TurboVNC 0.6 (and prior.)  libjpeg-turbo contains a wrapper
+library (TurboJPEG/OSS) that emulates the TurboJPEG API using libjpeg-turbo
+instead of the closed source Intel Performance Primitives.  You can replace the
+TurboJPEG/IPP package on Linux systems with the libjpeg-turbo package in order
+to make existing releases of VirtualGL 2.1.x and TurboVNC use the new codec at
+run time.  Note that the 64-bit libjpeg-turbo packages contain only 64-bit
+binaries, whereas the TurboJPEG/IPP 64-bit packages contained both 64-bit and
+32-bit binaries.  Thus, to replace a TurboJPEG/IPP 64-bit package, install
+both the 64-bit and 32-bit versions of libjpeg-turbo.
+
+You can also build the VirtualGL 2.1.x and TurboVNC 0.6 source code with
+the libjpeg-turbo SDK instead of TurboJPEG/IPP.  It should work identically.
+libjpeg-turbo also includes static library versions of TurboJPEG/OSS, which
+are used to build TurboVNC 1.0 and later.
+
+========================================
+Using libjpeg-turbo in Your Own Programs
+========================================
+
+For the most part, libjpeg-turbo should work identically to libjpeg, so in
+most cases, an application can be built against libjpeg and then run against
+libjpeg-turbo.  On Unix systems (including Cygwin), you can build against
+libjpeg-turbo instead of libjpeg by setting
+
+  CPATH=/opt/libjpeg-turbo/include
+  and
+  LIBRARY_PATH=/opt/libjpeg-turbo/{lib}
+
+({lib} = lib32 or lib64, depending on whether you are building a 32-bit or a
+64-bit application.)
+
+If using MinGW, then set
+
+  CPATH=/c/libjpeg-turbo-gcc[64]/include
+  and
+  LIBRARY_PATH=/c/libjpeg-turbo-gcc[64]/lib
+
+Building against libjpeg-turbo is useful, for instance, if you want to build an
+application that leverages the libjpeg-turbo colorspace extensions (see below.)
+On Linux and Solaris systems, you would still need to manipulate the
+LD_LIBRARY_PATH or sym links appropriately to use libjpeg-turbo at run time.
+On such systems, you can pass -R /opt/libjpeg-turbo/{lib} to the linker to
+force the use of libjpeg-turbo at run time rather than libjpeg (also useful if
+you want to leverage the colorspace extensions), or you can link against the
+libjpeg-turbo static library.
+
+To force a Linux, Solaris, or MinGW application to link against the static
+version of libjpeg-turbo, you can use the following linker options:
+
+  -Wl,-Bstatic -ljpeg -Wl,-Bdynamic
+
+On OS X, simply add /opt/libjpeg-turbo/lib/libjpeg.a to the linker command
+line (this also works on Linux and Solaris.)
+
+To build Visual C++ applications using libjpeg-turbo, add
+c:\libjpeg-turbo[64]\include to your system or user INCLUDE environment
+variable and c:\libjpeg-turbo[64]\lib to your system or user LIB environment
+variable, and then link against either jpeg.lib (to use the DLL version of
+libjpeg-turbo) or jpeg-static.lib (to use the static version of libjpeg-turbo.)
+
+=====================
+Colorspace Extensions
+=====================
+
+libjpeg-turbo includes extensions which allow JPEG images to be compressed
+directly from (and decompressed directly to) buffers which use BGR, BGRA,
+RGBA, ABGR, and ARGB pixel ordering.  This is implemented with six new
+colorspace constants:
+
+  JCS_EXT_RGB   /* red/green/blue */
+  JCS_EXT_RGBX  /* red/green/blue/x */
+  JCS_EXT_BGR   /* blue/green/red */
+  JCS_EXT_BGRX  /* blue/green/red/x */
+  JCS_EXT_XBGR  /* x/blue/green/red */
+  JCS_EXT_XRGB  /* x/red/green/blue */
+
+Setting cinfo.in_color_space (compression) or cinfo.out_color_space
+(decompression) to one of these values will cause libjpeg-turbo to read the
+red, green, and blue values from (or write them to) the appropriate position in
+the pixel when YUV conversion is performed.
+
+Your application can check for the existence of these extensions at compile
+time with:
+
+  #ifdef JCS_EXTENSIONS
+
+At run time, attempting to use these extensions with a version of libjpeg
+that doesn't support them will result in a "Bogus input colorspace" error.
+
+=================================
+libjpeg v7 and v8 API/ABI support
+=================================
+
+libjpeg v7 and v8 added new features to the API/ABI, and, unfortunately, the
+compression and decompression structures were extended in a backward-
+incompatible manner to accommodate these features.  Thus, programs which are
+built to use libjpeg v7 or v8 did not work with libjpeg-turbo, since it is
+based on the libjpeg v6b code base.  Although libjpeg v7 and v8 are still not
+as widely used as v6b, enough programs (including a few Linux distros) have
+made the switch that it was desirable to provide support for the libjpeg v7/v8
+API/ABI in libjpeg-turbo.
+
+Some of the libjpeg v7 and v8 features -- DCT scaling, to name one -- involve
+deep modifications to the code which cannot be accommodated by libjpeg-turbo
+without either breaking compatibility with libjpeg v6b or producing an
+unsupportable mess.  In order to fully support libjpeg v8 with all of its
+features, we would have to essentially port the SIMD extensions to the libjpeg
+v8 code base and maintain two separate code trees.  We are hesitant to do this
+until/unless the newer libjpeg code bases garner more community support and
+involvement and until/unless we have some notion of whether future libjpeg
+releases will also be backward-incompatible.
+
+By passing an argument of --with-jpeg7 or --with-jpeg8 to configure, or an
+argument of -DWITH_JPEG7=1 or -DWITH_JPEG8=1 to cmake, you can build a version
+of libjpeg-turbo which emulates the libjpeg v7 or v8b API/ABI, so that programs
+which are built against libjpeg v7 or v8 can be run with libjpeg-turbo.  The
+following section describes which libjpeg v7+ features are supported and which
+aren't.
+
+libjpeg v7 and v8 Features:
+---------------------------
+
+Fully supported:
+
+-- cjpeg: Separate quality settings for luminance and chrominance
+   Note that the libpjeg v7+ API was extended to accommodate this feature only
+   for convenience purposes.  It has always been possible to implement this
+   feature with libjpeg v6b (see rdswitch.c for an example.)
+
+-- cjpeg: 32-bit BMP support
+
+-- jpegtran: lossless cropping
+
+-- jpegtran: -perfect option
+
+-- rdjpgcom: -raw option
+
+-- rdjpgcom: locale awareness
+
+
+Fully supported when using libjpeg v7/v8 emulation:
+
+-- libjpeg: In-memory source and destination managers
+
+
+Not supported:
+
+-- libjpeg: DCT scaling in compressor
+   cinfo.scale_num and cinfo.scale_denom are silently ignored.
+
+-- libjpeg: IDCT scaling extensions in decompressor
+   libjpeg-turbo still supports IDCT scaling with scaling factors of 1/2, 1/4,
+   and 1/8 (same as libjpeg v6b.)
+
+-- libjpeg: Fancy downsampling in compressor
+   cinfo.do_fancy_downsampling is silently ignored.
+
+-- libjpeg: Arithmetic coding/decoding
+   Not supported due to patent concerns.
+
+-- jpegtran: Scaling
+   Seems to depend on the DCT scaling feature, which isn't supported.
diff --git a/acinclude.m4 b/acinclude.m4
new file mode 100644
index 0000000..cc10a21
--- /dev/null
+++ b/acinclude.m4
@@ -0,0 +1,136 @@
+# AC_PROG_NASM
+# --------------------------
+# Check that NASM exists and determine flags
+AC_DEFUN([AC_PROG_NASM],[
+
+AC_CHECK_PROGS(NASM, [nasm nasmw])
+test -z "$NASM" && AC_MSG_ERROR([no nasm (Netwide Assembler) found])
+
+AC_MSG_CHECKING([for object file format of host system])
+case "$host_os" in
+  cygwin* | mingw* | pw32* | interix*)
+    case "$host_cpu" in
+      x86_64)
+        objfmt='Win64-COFF'
+        ;;
+      *)
+        objfmt='Win32-COFF'
+        ;;
+    esac
+  ;;
+  msdosdjgpp* | go32*)
+    objfmt='COFF'
+  ;;
+  os2-emx*)			# not tested
+    objfmt='MSOMF'		# obj
+  ;;
+  linux*coff* | linux*oldld*)
+    objfmt='COFF'		# ???
+  ;;
+  linux*aout*)
+    objfmt='a.out'
+  ;;
+  linux*)
+    case "$host_cpu" in
+      x86_64)
+        objfmt='ELF64'
+        ;;
+      *)
+        objfmt='ELF'
+        ;;
+    esac
+  ;;
+  freebsd* | netbsd* | openbsd*)
+    if echo __ELF__ | $CC -E - | grep __ELF__ > /dev/null; then
+      objfmt='BSD-a.out'
+    else
+      case "$host_cpu" in
+        x86_64 | amd64)
+          objfmt='ELF64'
+          ;;
+        *)
+          objfmt='ELF'
+          ;;
+      esac
+    fi
+  ;;
+  solaris* | sunos* | sysv* | sco*)
+    case "$host_cpu" in
+      x86_64)
+        objfmt='ELF64'
+        ;;
+      *)
+        objfmt='ELF'
+        ;;
+    esac
+  ;;
+  darwin* | rhapsody* | nextstep* | openstep* | macos*)
+    case "$host_cpu" in
+      x86_64)
+        objfmt='Mach-O64'
+        ;;
+      *)
+        objfmt='Mach-O'
+        ;;
+    esac
+  ;;
+  *)
+    objfmt='ELF ?'
+  ;;
+esac
+
+AC_MSG_RESULT([$objfmt])
+if test "$objfmt" = 'ELF ?'; then
+  objfmt='ELF'
+  AC_MSG_WARN([unexpected host system. assumed that the format is $objfmt.])
+fi
+
+AC_MSG_CHECKING([for object file format specifier (NAFLAGS) ])
+case "$objfmt" in
+  MSOMF)      NAFLAGS='-fobj -DOBJ32';;
+  Win32-COFF) NAFLAGS='-fwin32 -DWIN32';;
+  Win64-COFF) NAFLAGS='-fwin64 -DWIN64 -D__x86_64__';;
+  COFF)       NAFLAGS='-fcoff -DCOFF';;
+  a.out)      NAFLAGS='-faout -DAOUT';;
+  BSD-a.out)  NAFLAGS='-faoutb -DAOUT';;
+  ELF)        NAFLAGS='-felf -DELF';;
+  ELF64)      NAFLAGS='-felf64 -DELF -D__x86_64__';;
+  RDF)        NAFLAGS='-frdf -DRDF';;
+  Mach-O)     NAFLAGS='-fmacho -DMACHO';;
+  Mach-O64)   NAFLAGS='-fmacho64 -DMACHO -D__x86_64__';;
+esac
+AC_MSG_RESULT([$NAFLAGS])
+AC_SUBST([NAFLAGS])
+
+AC_MSG_CHECKING([whether the assembler ($NASM $NAFLAGS) works])
+cat > conftest.asm <<EOF
+[%line __oline__ "configure"
+        section .text
+        global  _main,main
+_main:
+main:   xor     eax,eax
+        ret
+]EOF
+try_nasm='$NASM $NAFLAGS -o conftest.o conftest.asm'
+if AC_TRY_EVAL(try_nasm) && test -s conftest.o; then
+  AC_MSG_RESULT(yes)
+else
+  echo "configure: failed program was:" >&AC_FD_CC
+  cat conftest.asm >&AC_FD_CC
+  rm -rf conftest*
+  AC_MSG_RESULT(no)
+  AC_MSG_ERROR([installation or configuration problem: assembler cannot create object files.])
+fi
+
+AC_MSG_CHECKING([whether the linker accepts assembler output])
+try_nasm='${CC-cc} -o conftest${ac_exeext} $LDFLAGS conftest.o $LIBS 1>&AC_FD_CC'
+if AC_TRY_EVAL(try_nasm) && test -s conftest${ac_exeext}; then
+  rm -rf conftest*
+  AC_MSG_RESULT(yes)
+else
+  rm -rf conftest*
+  AC_MSG_RESULT(no)
+  AC_MSG_ERROR([configuration problem: maybe object file format mismatch.])
+fi
+
+])
diff --git a/bmp.c b/bmp.c
new file mode 100644
index 0000000..3ccc877
--- /dev/null
+++ b/bmp.c
@@ -0,0 +1,370 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005 Sun Microsystems, Inc.
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+*/
+
+#include <fcntl.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <errno.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#ifdef _WIN32
+ #include <io.h>
+#else
+ #include <unistd.h>
+#endif
+#include "./rrutil.h"
+#include "./bmp.h"
+
+#ifndef BI_BITFIELDS
+#define BI_BITFIELDS 3L
+#endif
+#ifndef BI_RGB
+#define BI_RGB 0L
+#endif
+
+#define BMPHDRSIZE 54
+typedef struct _bmphdr
+{
+	unsigned short bfType;
+	unsigned int bfSize;
+	unsigned short bfReserved1, bfReserved2;
+	unsigned int bfOffBits;
+
+	unsigned int biSize;
+	int biWidth, biHeight;
+	unsigned short biPlanes, biBitCount;
+	unsigned int biCompression, biSizeImage;
+	int biXPelsPerMeter, biYPelsPerMeter;
+	unsigned int biClrUsed, biClrImportant;
+} bmphdr;
+
+static const char *__bmperr="No error";
+
+static const int ps[BMPPIXELFORMATS]={3, 4, 3, 4, 4, 4};
+static const int roffset[BMPPIXELFORMATS]={0, 0, 2, 2, 3, 1};
+static const int goffset[BMPPIXELFORMATS]={1, 1, 1, 1, 2, 2};
+static const int boffset[BMPPIXELFORMATS]={2, 2, 0, 0, 1, 3};
+
+#define _throw(m) {__bmperr=m;  retcode=-1;  goto finally;}
+#define _unix(f) {if((f)==-1) _throw(strerror(errno));}
+#define _catch(f) {if((f)==-1) {retcode=-1;  goto finally;}}
+
+#define readme(fd, addr, size) \
+	if((bytesread=read(fd, addr, (size)))==-1) _throw(strerror(errno)); \
+	if(bytesread!=(size)) _throw("Read error");
+
+void pixelconvert(unsigned char *srcbuf, enum BMPPIXELFORMAT srcformat,
+	int srcpitch, unsigned char *dstbuf, enum BMPPIXELFORMAT dstformat, int dstpitch,
+	int w, int h, int flip)
+{
+	unsigned char *srcptr, *srcptr0, *dstptr, *dstptr0;
+	int i, j;
+
+	srcptr=flip? &srcbuf[srcpitch*(h-1)]:srcbuf;
+	for(j=0, dstptr=dstbuf; j<h; j++,
+		srcptr+=flip? -srcpitch:srcpitch, dstptr+=dstpitch)
+	{
+		for(i=0, srcptr0=srcptr, dstptr0=dstptr; i<w; i++,
+			srcptr0+=ps[srcformat], dstptr0+=ps[dstformat])
+		{
+			dstptr0[roffset[dstformat]]=srcptr0[roffset[srcformat]];
+			dstptr0[goffset[dstformat]]=srcptr0[goffset[srcformat]];
+			dstptr0[boffset[dstformat]]=srcptr0[boffset[srcformat]];
+		}
+	}
+}
+
+int loadppm(int *fd, unsigned char **buf, int *w, int *h,
+	enum BMPPIXELFORMAT f, int align, int dstbottomup, int ascii)
+{
+	FILE *fs=NULL;  int retcode=0, scalefactor, dstpitch;
+	unsigned char *tempbuf=NULL;  char temps[255], temps2[255];
+	int numread=0, totalread=0, pixel[3], i, j;
+
+	if((fs=fdopen(*fd, "r"))==NULL) _throw(strerror(errno));
+
+	do
+	{
+		if(!fgets(temps, 255, fs)) _throw("Read error");
+		if(strlen(temps)==0 || temps[0]=='\n') continue;
+		if(sscanf(temps, "%s", temps2)==1 && temps2[1]=='#') continue;
+		switch(totalread)
+		{
+			case 0:
+				if((numread=sscanf(temps, "%d %d %d", w, h, &scalefactor))==EOF)
+					_throw("Read error");
+				break;
+			case 1:
+				if((numread=sscanf(temps, "%d %d", h, &scalefactor))==EOF)
+					_throw("Read error");
+				break;
+			case 2:
+				if((numread=sscanf(temps, "%d", &scalefactor))==EOF)
+					_throw("Read error");
+				break;
+		}
+		totalread+=numread;
+	} while(totalread<3);
+	if((*w)<1 || (*h)<1 || scalefactor<1) _throw("Corrupt PPM header");
+
+	dstpitch=(((*w)*ps[f])+(align-1))&(~(align-1));
+	if((*buf=(unsigned char *)malloc(dstpitch*(*h)))==NULL)
+		_throw("Memory allocation error");
+	if(ascii)
+	{
+		for(j=0; j<*h; j++)
+		{
+			for(i=0; i<*w; i++)
+			{
+				if(fscanf(fs, "%d%d%d", &pixel[0], &pixel[1], &pixel[2])!=3)
+					_throw("Read error");
+				(*buf)[j*dstpitch+i*ps[f]+roffset[f]]=(unsigned char)(pixel[0]*255/scalefactor);
+				(*buf)[j*dstpitch+i*ps[f]+goffset[f]]=(unsigned char)(pixel[1]*255/scalefactor);
+				(*buf)[j*dstpitch+i*ps[f]+boffset[f]]=(unsigned char)(pixel[2]*255/scalefactor);
+			}
+		}
+	}
+	else
+	{
+		if(scalefactor!=255)
+			_throw("Binary PPMs must have 8-bit components");
+		if((tempbuf=(unsigned char *)malloc((*w)*(*h)*3))==NULL)
+			_throw("Memory allocation error");
+		if(fread(tempbuf, (*w)*(*h)*3, 1, fs)!=1) _throw("Read error");
+		pixelconvert(tempbuf, BMP_RGB, (*w)*3, *buf, f, dstpitch, *w, *h, dstbottomup);
+	}
+
+	finally:
+	if(fs) {fclose(fs);  *fd=-1;}
+	if(tempbuf) free(tempbuf);
+	return retcode;
+}
+
+
+int loadbmp(char *filename, unsigned char **buf, int *w, int *h, 
+	enum BMPPIXELFORMAT f, int align, int dstbottomup)
+{
+	int fd=-1, bytesread, srcpitch, srcbottomup=1, srcps, dstpitch,
+		retcode=0;
+	unsigned char *tempbuf=NULL;
+	bmphdr bh;  int flags=O_RDONLY;
+
+	dstbottomup=dstbottomup? 1:0;
+	#ifdef _WIN32
+	flags|=O_BINARY;
+	#endif
+	if(!filename || !buf || !w || !h || f<0 || f>BMPPIXELFORMATS-1 || align<1)
+		_throw("invalid argument to loadbmp()");
+	if((align&(align-1))!=0)
+		_throw("Alignment must be a power of 2");
+	_unix(fd=open(filename, flags));
+
+	readme(fd, &bh.bfType, sizeof(unsigned short));
+	if(!littleendian())	bh.bfType=byteswap16(bh.bfType);
+
+	if(bh.bfType==0x3650)
+	{
+		_catch(loadppm(&fd, buf, w, h, f, align, dstbottomup, 0));
+		goto finally;
+	}
+	if(bh.bfType==0x3350)
+	{
+		_catch(loadppm(&fd, buf, w, h, f, align, dstbottomup, 1));
+		goto finally;
+	}
+
+	readme(fd, &bh.bfSize, sizeof(unsigned int));
+	readme(fd, &bh.bfReserved1, sizeof(unsigned short));
+	readme(fd, &bh.bfReserved2, sizeof(unsigned short));
+	readme(fd, &bh.bfOffBits, sizeof(unsigned int));
+	readme(fd, &bh.biSize, sizeof(unsigned int));
+	readme(fd, &bh.biWidth, sizeof(int));
+	readme(fd, &bh.biHeight, sizeof(int));
+	readme(fd, &bh.biPlanes, sizeof(unsigned short));
+	readme(fd, &bh.biBitCount, sizeof(unsigned short));
+	readme(fd, &bh.biCompression, sizeof(unsigned int));
+	readme(fd, &bh.biSizeImage, sizeof(unsigned int));
+	readme(fd, &bh.biXPelsPerMeter, sizeof(int));
+	readme(fd, &bh.biYPelsPerMeter, sizeof(int));
+	readme(fd, &bh.biClrUsed, sizeof(unsigned int));
+	readme(fd, &bh.biClrImportant, sizeof(unsigned int));
+
+	if(!littleendian())
+	{
+		bh.bfSize=byteswap(bh.bfSize);
+		bh.bfOffBits=byteswap(bh.bfOffBits);
+		bh.biSize=byteswap(bh.biSize);
+		bh.biWidth=byteswap(bh.biWidth);
+		bh.biHeight=byteswap(bh.biHeight);
+		bh.biPlanes=byteswap16(bh.biPlanes);
+		bh.biBitCount=byteswap16(bh.biBitCount);
+		bh.biCompression=byteswap(bh.biCompression);
+		bh.biSizeImage=byteswap(bh.biSizeImage);
+		bh.biXPelsPerMeter=byteswap(bh.biXPelsPerMeter);
+		bh.biYPelsPerMeter=byteswap(bh.biYPelsPerMeter);
+		bh.biClrUsed=byteswap(bh.biClrUsed);
+		bh.biClrImportant=byteswap(bh.biClrImportant);
+	}
+
+	if(bh.bfType!=0x4d42 || bh.bfOffBits<BMPHDRSIZE
+	|| bh.biWidth<1 || bh.biHeight==0)
+		_throw("Corrupt bitmap header");
+	if((bh.biBitCount!=24 && bh.biBitCount!=32) || bh.biCompression!=BI_RGB)
+		_throw("Only uncompessed RGB bitmaps are supported");
+
+	*w=bh.biWidth;  *h=bh.biHeight;  srcps=bh.biBitCount/8;
+	if(*h<0) {*h=-(*h);  srcbottomup=0;}
+	srcpitch=(((*w)*srcps)+3)&(~3);
+	dstpitch=(((*w)*ps[f])+(align-1))&(~(align-1));
+
+	if(srcpitch*(*h)+bh.bfOffBits!=bh.bfSize) _throw("Corrupt bitmap header");
+	if((tempbuf=(unsigned char *)malloc(srcpitch*(*h)))==NULL
+	|| (*buf=(unsigned char *)malloc(dstpitch*(*h)))==NULL)
+		_throw("Memory allocation error");
+	if(lseek(fd, (long)bh.bfOffBits, SEEK_SET)!=(long)bh.bfOffBits)
+		_throw(strerror(errno));
+	_unix(bytesread=read(fd, tempbuf, srcpitch*(*h)));
+	if(bytesread!=srcpitch*(*h)) _throw("Read error");
+
+	pixelconvert(tempbuf, BMP_BGR, srcpitch, *buf, f, dstpitch, *w, *h, 
+		srcbottomup!=dstbottomup);
+
+	finally:
+	if(tempbuf) free(tempbuf);
+	if(fd!=-1) close(fd);
+	return retcode;
+}
+
+#define writeme(fd, addr, size) \
+	if((byteswritten=write(fd, addr, (size)))==-1) _throw(strerror(errno)); \
+	if(byteswritten!=(size)) _throw("Write error");
+
+int saveppm(char *filename, unsigned char *buf, int w, int h,
+	enum BMPPIXELFORMAT f, int srcpitch, int srcbottomup)
+{
+	FILE *fs=NULL;  int retcode=0;
+	unsigned char *tempbuf=NULL;
+
+	if((fs=fopen(filename, "wb"))==NULL) _throw(strerror(errno));
+	if(fprintf(fs, "P6\n")<1) _throw("Write error");
+	if(fprintf(fs, "%d %d\n", w, h)<1) _throw("Write error");
+	if(fprintf(fs, "255\n")<1) _throw("Write error");
+
+	if((tempbuf=(unsigned char *)malloc(w*h*3))==NULL)
+		_throw("Memory allocation error");
+
+	pixelconvert(buf, f, srcpitch, tempbuf, BMP_RGB, w*3, w, h, 
+		srcbottomup);
+
+	if((fwrite(tempbuf, w*h*3, 1, fs))!=1) _throw("Write error");
+
+	finally:
+	if(tempbuf) free(tempbuf);
+	if(fs) fclose(fs);
+	return retcode;
+}
+
+int savebmp(char *filename, unsigned char *buf, int w, int h,
+	enum BMPPIXELFORMAT f, int srcpitch, int srcbottomup)
+{
+	int fd=-1, byteswritten, dstpitch, retcode=0;
+	int flags=O_RDWR|O_CREAT|O_TRUNC;
+	unsigned char *tempbuf=NULL;  char *temp;
+	bmphdr bh;  int mode;
+
+	#ifdef _WIN32
+	flags|=O_BINARY;  mode=_S_IREAD|_S_IWRITE;
+	#else
+	mode=S_IRUSR|S_IWUSR|S_IRGRP|S_IWGRP|S_IROTH|S_IWOTH;
+	#endif
+	if(!filename || !buf || w<1 || h<1 || f<0 || f>BMPPIXELFORMATS-1 || srcpitch<0)
+		_throw("bad argument to savebmp()");
+
+	if(srcpitch==0) srcpitch=w*ps[f];
+
+	if((temp=strrchr(filename, '.'))!=NULL)
+	{
+		if(!stricmp(temp, ".ppm"))
+			return saveppm(filename, buf, w, h, f, srcpitch, srcbottomup);
+	}
+
+	_unix(fd=open(filename, flags, mode));
+	dstpitch=((w*3)+3)&(~3);
+
+	bh.bfType=0x4d42;
+	bh.bfSize=BMPHDRSIZE+dstpitch*h;
+	bh.bfReserved1=0;  bh.bfReserved2=0;
+	bh.bfOffBits=BMPHDRSIZE;
+	bh.biSize=40;
+	bh.biWidth=w;  bh.biHeight=h;
+	bh.biPlanes=0;  bh.biBitCount=24;
+	bh.biCompression=BI_RGB;  bh.biSizeImage=0;
+	bh.biXPelsPerMeter=0;  bh.biYPelsPerMeter=0;
+	bh.biClrUsed=0;  bh.biClrImportant=0;
+
+	if(!littleendian())
+	{
+		bh.bfType=byteswap16(bh.bfType);
+		bh.bfSize=byteswap(bh.bfSize);
+		bh.bfOffBits=byteswap(bh.bfOffBits);
+		bh.biSize=byteswap(bh.biSize);
+		bh.biWidth=byteswap(bh.biWidth);
+		bh.biHeight=byteswap(bh.biHeight);
+		bh.biPlanes=byteswap16(bh.biPlanes);
+		bh.biBitCount=byteswap16(bh.biBitCount);
+		bh.biCompression=byteswap(bh.biCompression);
+		bh.biSizeImage=byteswap(bh.biSizeImage);
+		bh.biXPelsPerMeter=byteswap(bh.biXPelsPerMeter);
+		bh.biYPelsPerMeter=byteswap(bh.biYPelsPerMeter);
+		bh.biClrUsed=byteswap(bh.biClrUsed);
+		bh.biClrImportant=byteswap(bh.biClrImportant);
+	}
+
+	writeme(fd, &bh.bfType, sizeof(unsigned short));
+	writeme(fd, &bh.bfSize, sizeof(unsigned int));
+	writeme(fd, &bh.bfReserved1, sizeof(unsigned short));
+	writeme(fd, &bh.bfReserved2, sizeof(unsigned short));
+	writeme(fd, &bh.bfOffBits, sizeof(unsigned int));
+	writeme(fd, &bh.biSize, sizeof(unsigned int));
+	writeme(fd, &bh.biWidth, sizeof(int));
+	writeme(fd, &bh.biHeight, sizeof(int));
+	writeme(fd, &bh.biPlanes, sizeof(unsigned short));
+	writeme(fd, &bh.biBitCount, sizeof(unsigned short));
+	writeme(fd, &bh.biCompression, sizeof(unsigned int));
+	writeme(fd, &bh.biSizeImage, sizeof(unsigned int));
+	writeme(fd, &bh.biXPelsPerMeter, sizeof(int));
+	writeme(fd, &bh.biYPelsPerMeter, sizeof(int));
+	writeme(fd, &bh.biClrUsed, sizeof(unsigned int));
+	writeme(fd, &bh.biClrImportant, sizeof(unsigned int));
+
+	if((tempbuf=(unsigned char *)malloc(dstpitch*h))==NULL)
+		_throw("Memory allocation error");
+
+	pixelconvert(buf, f, srcpitch, tempbuf, BMP_BGR, dstpitch, w, h, 
+		!srcbottomup);
+
+	if((byteswritten=write(fd, tempbuf, dstpitch*h))!=dstpitch*h)
+		_throw(strerror(errno));
+
+	finally:
+	if(tempbuf) free(tempbuf);
+	if(fd!=-1) close(fd);
+	return retcode;
+}
+
+const char *bmpgeterr(void)
+{
+	return __bmperr;
+}
diff --git a/bmp.h b/bmp.h
new file mode 100644
index 0000000..437d327
--- /dev/null
+++ b/bmp.h
@@ -0,0 +1,48 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005 Sun Microsystems, Inc.
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+*/
+
+// This provides rudimentary facilities for loading and saving true color
+// BMP and PPM files
+
+#ifndef __BMP_H__
+#define __BMP_H__
+
+#define BMPPIXELFORMATS 6
+enum BMPPIXELFORMAT {BMP_RGB=0, BMP_RGBA, BMP_BGR, BMP_BGRA, BMP_ABGR, BMP_ARGB};
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// This will load a Windows bitmap from a file and return a buffer with the
+// specified pixel format, scanline alignment, and orientation.  The width and
+// height are returned in w and h.
+
+int loadbmp(char *filename, unsigned char **buf, int *w, int *h,
+	enum BMPPIXELFORMAT f, int align, int dstbottomup);
+
+// This will save a buffer with the specified pixel format, pitch, orientation,
+// width, and height as a 24-bit Windows bitmap or PPM (the filename determines
+// which format to use)
+
+int savebmp(char *filename, unsigned char *buf, int w, int h,
+	enum BMPPIXELFORMAT f, int srcpitch, int srcbottomup);
+
+const char *bmpgeterr(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/cderror.h b/cderror.h
new file mode 100644
index 0000000..e19c475
--- /dev/null
+++ b/cderror.h
@@ -0,0 +1,134 @@
+/*
+ * cderror.h
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file defines the error and message codes for the cjpeg/djpeg
+ * applications.  These strings are not needed as part of the JPEG library
+ * proper.
+ * Edit this file to add new codes, or to translate the message strings to
+ * some other language.
+ */
+
+/*
+ * To define the enum list of message codes, include this file without
+ * defining macro JMESSAGE.  To create a message string table, include it
+ * again with a suitable JMESSAGE definition (see jerror.c for an example).
+ */
+#ifndef JMESSAGE
+#ifndef CDERROR_H
+#define CDERROR_H
+/* First time through, define the enum list */
+#define JMAKE_ENUM_LIST
+#else
+/* Repeated inclusions of this file are no-ops unless JMESSAGE is defined */
+#define JMESSAGE(code,string)
+#endif /* CDERROR_H */
+#endif /* JMESSAGE */
+
+#ifdef JMAKE_ENUM_LIST
+
+typedef enum {
+
+#define JMESSAGE(code,string)	code ,
+
+#endif /* JMAKE_ENUM_LIST */
+
+JMESSAGE(JMSG_FIRSTADDONCODE=1000, NULL) /* Must be first entry! */
+
+#ifdef BMP_SUPPORTED
+JMESSAGE(JERR_BMP_BADCMAP, "Unsupported BMP colormap format")
+JMESSAGE(JERR_BMP_BADDEPTH, "Only 8- and 24-bit BMP files are supported")
+JMESSAGE(JERR_BMP_BADHEADER, "Invalid BMP file: bad header length")
+JMESSAGE(JERR_BMP_BADPLANES, "Invalid BMP file: biPlanes not equal to 1")
+JMESSAGE(JERR_BMP_COLORSPACE, "BMP output must be grayscale or RGB")
+JMESSAGE(JERR_BMP_COMPRESSED, "Sorry, compressed BMPs not yet supported")
+JMESSAGE(JERR_BMP_EMPTY, "Empty BMP image")
+JMESSAGE(JERR_BMP_NOT, "Not a BMP file - does not start with BM")
+JMESSAGE(JTRC_BMP, "%ux%u 24-bit BMP image")
+JMESSAGE(JTRC_BMP_MAPPED, "%ux%u 8-bit colormapped BMP image")
+JMESSAGE(JTRC_BMP_OS2, "%ux%u 24-bit OS2 BMP image")
+JMESSAGE(JTRC_BMP_OS2_MAPPED, "%ux%u 8-bit colormapped OS2 BMP image")
+#endif /* BMP_SUPPORTED */
+
+#ifdef GIF_SUPPORTED
+JMESSAGE(JERR_GIF_BUG, "GIF output got confused")
+JMESSAGE(JERR_GIF_CODESIZE, "Bogus GIF codesize %d")
+JMESSAGE(JERR_GIF_COLORSPACE, "GIF output must be grayscale or RGB")
+JMESSAGE(JERR_GIF_IMAGENOTFOUND, "Too few images in GIF file")
+JMESSAGE(JERR_GIF_NOT, "Not a GIF file")
+JMESSAGE(JTRC_GIF, "%ux%ux%d GIF image")
+JMESSAGE(JTRC_GIF_BADVERSION,
+	 "Warning: unexpected GIF version number '%c%c%c'")
+JMESSAGE(JTRC_GIF_EXTENSION, "Ignoring GIF extension block of type 0x%02x")
+JMESSAGE(JTRC_GIF_NONSQUARE, "Caution: nonsquare pixels in input")
+JMESSAGE(JWRN_GIF_BADDATA, "Corrupt data in GIF file")
+JMESSAGE(JWRN_GIF_CHAR, "Bogus char 0x%02x in GIF file, ignoring")
+JMESSAGE(JWRN_GIF_ENDCODE, "Premature end of GIF image")
+JMESSAGE(JWRN_GIF_NOMOREDATA, "Ran out of GIF bits")
+#endif /* GIF_SUPPORTED */
+
+#ifdef PPM_SUPPORTED
+JMESSAGE(JERR_PPM_COLORSPACE, "PPM output must be grayscale or RGB")
+JMESSAGE(JERR_PPM_NONNUMERIC, "Nonnumeric data in PPM file")
+JMESSAGE(JERR_PPM_NOT, "Not a PPM/PGM file")
+JMESSAGE(JTRC_PGM, "%ux%u PGM image")
+JMESSAGE(JTRC_PGM_TEXT, "%ux%u text PGM image")
+JMESSAGE(JTRC_PPM, "%ux%u PPM image")
+JMESSAGE(JTRC_PPM_TEXT, "%ux%u text PPM image")
+#endif /* PPM_SUPPORTED */
+
+#ifdef RLE_SUPPORTED
+JMESSAGE(JERR_RLE_BADERROR, "Bogus error code from RLE library")
+JMESSAGE(JERR_RLE_COLORSPACE, "RLE output must be grayscale or RGB")
+JMESSAGE(JERR_RLE_DIMENSIONS, "Image dimensions (%ux%u) too large for RLE")
+JMESSAGE(JERR_RLE_EMPTY, "Empty RLE file")
+JMESSAGE(JERR_RLE_EOF, "Premature EOF in RLE header")
+JMESSAGE(JERR_RLE_MEM, "Insufficient memory for RLE header")
+JMESSAGE(JERR_RLE_NOT, "Not an RLE file")
+JMESSAGE(JERR_RLE_TOOMANYCHANNELS, "Cannot handle %d output channels for RLE")
+JMESSAGE(JERR_RLE_UNSUPPORTED, "Cannot handle this RLE setup")
+JMESSAGE(JTRC_RLE, "%ux%u full-color RLE file")
+JMESSAGE(JTRC_RLE_FULLMAP, "%ux%u full-color RLE file with map of length %d")
+JMESSAGE(JTRC_RLE_GRAY, "%ux%u grayscale RLE file")
+JMESSAGE(JTRC_RLE_MAPGRAY, "%ux%u grayscale RLE file with map of length %d")
+JMESSAGE(JTRC_RLE_MAPPED, "%ux%u colormapped RLE file with map of length %d")
+#endif /* RLE_SUPPORTED */
+
+#ifdef TARGA_SUPPORTED
+JMESSAGE(JERR_TGA_BADCMAP, "Unsupported Targa colormap format")
+JMESSAGE(JERR_TGA_BADPARMS, "Invalid or unsupported Targa file")
+JMESSAGE(JERR_TGA_COLORSPACE, "Targa output must be grayscale or RGB")
+JMESSAGE(JTRC_TGA, "%ux%u RGB Targa image")
+JMESSAGE(JTRC_TGA_GRAY, "%ux%u grayscale Targa image")
+JMESSAGE(JTRC_TGA_MAPPED, "%ux%u colormapped Targa image")
+#else
+JMESSAGE(JERR_TGA_NOTCOMP, "Targa support was not compiled")
+#endif /* TARGA_SUPPORTED */
+
+JMESSAGE(JERR_BAD_CMAP_FILE,
+	 "Color map file is invalid or of unsupported format")
+JMESSAGE(JERR_TOO_MANY_COLORS,
+	 "Output file format cannot handle %d colormap entries")
+JMESSAGE(JERR_UNGETC_FAILED, "ungetc failed")
+#ifdef TARGA_SUPPORTED
+JMESSAGE(JERR_UNKNOWN_FORMAT,
+	 "Unrecognized input file format --- perhaps you need -targa")
+#else
+JMESSAGE(JERR_UNKNOWN_FORMAT, "Unrecognized input file format")
+#endif
+JMESSAGE(JERR_UNSUPPORTED_FORMAT, "Unsupported output file format")
+
+#ifdef JMAKE_ENUM_LIST
+
+  JMSG_LASTADDONCODE
+} ADDON_MESSAGE_CODE;
+
+#undef JMAKE_ENUM_LIST
+#endif /* JMAKE_ENUM_LIST */
+
+/* Zap JMESSAGE macro so that future re-inclusions do nothing by default */
+#undef JMESSAGE
diff --git a/cdjpeg.c b/cdjpeg.c
new file mode 100644
index 0000000..b6250ff
--- /dev/null
+++ b/cdjpeg.c
@@ -0,0 +1,181 @@
+/*
+ * cdjpeg.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains common support routines used by the IJG application
+ * programs (cjpeg, djpeg, jpegtran).
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+#include <ctype.h>		/* to declare isupper(), tolower() */
+#ifdef NEED_SIGNAL_CATCHER
+#include <signal.h>		/* to declare signal() */
+#endif
+#ifdef USE_SETMODE
+#include <fcntl.h>		/* to declare setmode()'s parameter macros */
+/* If you have setmode() but not <io.h>, just delete this line: */
+#include <io.h>			/* to declare setmode() */
+#endif
+
+
+/*
+ * Signal catcher to ensure that temporary files are removed before aborting.
+ * NB: for Amiga Manx C this is actually a global routine named _abort();
+ * we put "#define signal_catcher _abort" in jconfig.h.  Talk about bogus...
+ */
+
+#ifdef NEED_SIGNAL_CATCHER
+
+static j_common_ptr sig_cinfo;
+
+void				/* must be global for Manx C */
+signal_catcher (int signum)
+{
+  if (sig_cinfo != NULL) {
+    if (sig_cinfo->err != NULL) /* turn off trace output */
+      sig_cinfo->err->trace_level = 0;
+    jpeg_destroy(sig_cinfo);	/* clean up memory allocation & temp files */
+  }
+  exit(EXIT_FAILURE);
+}
+
+
+GLOBAL(void)
+enable_signal_catcher (j_common_ptr cinfo)
+{
+  sig_cinfo = cinfo;
+#ifdef SIGINT			/* not all systems have SIGINT */
+  signal(SIGINT, signal_catcher);
+#endif
+#ifdef SIGTERM			/* not all systems have SIGTERM */
+  signal(SIGTERM, signal_catcher);
+#endif
+}
+
+#endif
+
+
+/*
+ * Optional progress monitor: display a percent-done figure on stderr.
+ */
+
+#ifdef PROGRESS_REPORT
+
+METHODDEF(void)
+progress_monitor (j_common_ptr cinfo)
+{
+  cd_progress_ptr prog = (cd_progress_ptr) cinfo->progress;
+  int total_passes = prog->pub.total_passes + prog->total_extra_passes;
+  int percent_done = (int) (prog->pub.pass_counter*100L/prog->pub.pass_limit);
+
+  if (percent_done != prog->percent_done) {
+    prog->percent_done = percent_done;
+    if (total_passes > 1) {
+      fprintf(stderr, "\rPass %d/%d: %3d%% ",
+	      prog->pub.completed_passes + prog->completed_extra_passes + 1,
+	      total_passes, percent_done);
+    } else {
+      fprintf(stderr, "\r %3d%% ", percent_done);
+    }
+    fflush(stderr);
+  }
+}
+
+
+GLOBAL(void)
+start_progress_monitor (j_common_ptr cinfo, cd_progress_ptr progress)
+{
+  /* Enable progress display, unless trace output is on */
+  if (cinfo->err->trace_level == 0) {
+    progress->pub.progress_monitor = progress_monitor;
+    progress->completed_extra_passes = 0;
+    progress->total_extra_passes = 0;
+    progress->percent_done = -1;
+    cinfo->progress = &progress->pub;
+  }
+}
+
+
+GLOBAL(void)
+end_progress_monitor (j_common_ptr cinfo)
+{
+  /* Clear away progress display */
+  if (cinfo->err->trace_level == 0) {
+    fprintf(stderr, "\r                \r");
+    fflush(stderr);
+  }
+}
+
+#endif
+
+
+/*
+ * Case-insensitive matching of possibly-abbreviated keyword switches.
+ * keyword is the constant keyword (must be lower case already),
+ * minchars is length of minimum legal abbreviation.
+ */
+
+GLOBAL(boolean)
+keymatch (char * arg, const char * keyword, int minchars)
+{
+  register int ca, ck;
+  register int nmatched = 0;
+
+  while ((ca = *arg++) != '\0') {
+    if ((ck = *keyword++) == '\0')
+      return FALSE;		/* arg longer than keyword, no good */
+    if (isupper(ca))		/* force arg to lcase (assume ck is already) */
+      ca = tolower(ca);
+    if (ca != ck)
+      return FALSE;		/* no good */
+    nmatched++;			/* count matched characters */
+  }
+  /* reached end of argument; fail if it's too short for unique abbrev */
+  if (nmatched < minchars)
+    return FALSE;
+  return TRUE;			/* A-OK */
+}
+
+
+/*
+ * Routines to establish binary I/O mode for stdin and stdout.
+ * Non-Unix systems often require some hacking to get out of text mode.
+ */
+
+GLOBAL(FILE *)
+read_stdin (void)
+{
+  FILE * input_file = stdin;
+
+#ifdef USE_SETMODE		/* need to hack file mode? */
+  setmode(fileno(stdin), O_BINARY);
+#endif
+#ifdef USE_FDOPEN		/* need to re-open in binary mode? */
+  if ((input_file = fdopen(fileno(stdin), READ_BINARY)) == NULL) {
+    fprintf(stderr, "Cannot reopen stdin\n");
+    exit(EXIT_FAILURE);
+  }
+#endif
+  return input_file;
+}
+
+
+GLOBAL(FILE *)
+write_stdout (void)
+{
+  FILE * output_file = stdout;
+
+#ifdef USE_SETMODE		/* need to hack file mode? */
+  setmode(fileno(stdout), O_BINARY);
+#endif
+#ifdef USE_FDOPEN		/* need to re-open in binary mode? */
+  if ((output_file = fdopen(fileno(stdout), WRITE_BINARY)) == NULL) {
+    fprintf(stderr, "Cannot reopen stdout\n");
+    exit(EXIT_FAILURE);
+  }
+#endif
+  return output_file;
+}
diff --git a/cdjpeg.h b/cdjpeg.h
new file mode 100644
index 0000000..ed024ac
--- /dev/null
+++ b/cdjpeg.h
@@ -0,0 +1,187 @@
+/*
+ * cdjpeg.h
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains common declarations for the sample applications
+ * cjpeg and djpeg.  It is NOT used by the core JPEG library.
+ */
+
+#define JPEG_CJPEG_DJPEG	/* define proper options in jconfig.h */
+#define JPEG_INTERNAL_OPTIONS	/* cjpeg.c,djpeg.c need to see xxx_SUPPORTED */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"		/* get library error codes too */
+#include "cderror.h"		/* get application-specific error codes */
+
+
+/*
+ * Object interface for cjpeg's source file decoding modules
+ */
+
+typedef struct cjpeg_source_struct * cjpeg_source_ptr;
+
+struct cjpeg_source_struct {
+  JMETHOD(void, start_input, (j_compress_ptr cinfo,
+			      cjpeg_source_ptr sinfo));
+  JMETHOD(JDIMENSION, get_pixel_rows, (j_compress_ptr cinfo,
+				       cjpeg_source_ptr sinfo));
+  JMETHOD(void, finish_input, (j_compress_ptr cinfo,
+			       cjpeg_source_ptr sinfo));
+
+  FILE *input_file;
+
+  JSAMPARRAY buffer;
+  JDIMENSION buffer_height;
+};
+
+
+/*
+ * Object interface for djpeg's output file encoding modules
+ */
+
+typedef struct djpeg_dest_struct * djpeg_dest_ptr;
+
+struct djpeg_dest_struct {
+  /* start_output is called after jpeg_start_decompress finishes.
+   * The color map will be ready at this time, if one is needed.
+   */
+  JMETHOD(void, start_output, (j_decompress_ptr cinfo,
+			       djpeg_dest_ptr dinfo));
+  /* Emit the specified number of pixel rows from the buffer. */
+  JMETHOD(void, put_pixel_rows, (j_decompress_ptr cinfo,
+				 djpeg_dest_ptr dinfo,
+				 JDIMENSION rows_supplied));
+  /* Finish up at the end of the image. */
+  JMETHOD(void, finish_output, (j_decompress_ptr cinfo,
+				djpeg_dest_ptr dinfo));
+
+  /* Target file spec; filled in by djpeg.c after object is created. */
+  FILE * output_file;
+
+  /* Output pixel-row buffer.  Created by module init or start_output.
+   * Width is cinfo->output_width * cinfo->output_components;
+   * height is buffer_height.
+   */
+  JSAMPARRAY buffer;
+  JDIMENSION buffer_height;
+};
+
+
+/*
+ * cjpeg/djpeg may need to perform extra passes to convert to or from
+ * the source/destination file format.  The JPEG library does not know
+ * about these passes, but we'd like them to be counted by the progress
+ * monitor.  We use an expanded progress monitor object to hold the
+ * additional pass count.
+ */
+
+struct cdjpeg_progress_mgr {
+  struct jpeg_progress_mgr pub;	/* fields known to JPEG library */
+  int completed_extra_passes;	/* extra passes completed */
+  int total_extra_passes;	/* total extra */
+  /* last printed percentage stored here to avoid multiple printouts */
+  int percent_done;
+};
+
+typedef struct cdjpeg_progress_mgr * cd_progress_ptr;
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jinit_read_bmp		jIRdBMP
+#define jinit_write_bmp		jIWrBMP
+#define jinit_read_gif		jIRdGIF
+#define jinit_write_gif		jIWrGIF
+#define jinit_read_ppm		jIRdPPM
+#define jinit_write_ppm		jIWrPPM
+#define jinit_read_rle		jIRdRLE
+#define jinit_write_rle		jIWrRLE
+#define jinit_read_targa	jIRdTarga
+#define jinit_write_targa	jIWrTarga
+#define read_quant_tables	RdQTables
+#define read_scan_script	RdScnScript
+#define set_quality_ratings     SetQRates
+#define set_quant_slots		SetQSlots
+#define set_sample_factors	SetSFacts
+#define read_color_map		RdCMap
+#define enable_signal_catcher	EnSigCatcher
+#define start_progress_monitor	StProgMon
+#define end_progress_monitor	EnProgMon
+#define read_stdin		RdStdin
+#define write_stdout		WrStdout
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* Module selection routines for I/O modules. */
+
+EXTERN(cjpeg_source_ptr) jinit_read_bmp JPP((j_compress_ptr cinfo));
+EXTERN(djpeg_dest_ptr) jinit_write_bmp JPP((j_decompress_ptr cinfo,
+					    boolean is_os2));
+EXTERN(cjpeg_source_ptr) jinit_read_gif JPP((j_compress_ptr cinfo));
+EXTERN(djpeg_dest_ptr) jinit_write_gif JPP((j_decompress_ptr cinfo));
+EXTERN(cjpeg_source_ptr) jinit_read_ppm JPP((j_compress_ptr cinfo));
+EXTERN(djpeg_dest_ptr) jinit_write_ppm JPP((j_decompress_ptr cinfo));
+EXTERN(cjpeg_source_ptr) jinit_read_rle JPP((j_compress_ptr cinfo));
+EXTERN(djpeg_dest_ptr) jinit_write_rle JPP((j_decompress_ptr cinfo));
+EXTERN(cjpeg_source_ptr) jinit_read_targa JPP((j_compress_ptr cinfo));
+EXTERN(djpeg_dest_ptr) jinit_write_targa JPP((j_decompress_ptr cinfo));
+
+/* cjpeg support routines (in rdswitch.c) */
+
+EXTERN(boolean) read_quant_tables JPP((j_compress_ptr cinfo, char * filename,
+				       boolean force_baseline));
+EXTERN(boolean) read_scan_script JPP((j_compress_ptr cinfo, char * filename));
+EXTERN(boolean) set_quality_ratings JPP((j_compress_ptr cinfo, char *arg,
+					 boolean force_baseline));
+EXTERN(boolean) set_quant_slots JPP((j_compress_ptr cinfo, char *arg));
+EXTERN(boolean) set_sample_factors JPP((j_compress_ptr cinfo, char *arg));
+
+/* djpeg support routines (in rdcolmap.c) */
+
+EXTERN(void) read_color_map JPP((j_decompress_ptr cinfo, FILE * infile));
+
+/* common support routines (in cdjpeg.c) */
+
+EXTERN(void) enable_signal_catcher JPP((j_common_ptr cinfo));
+EXTERN(void) start_progress_monitor JPP((j_common_ptr cinfo,
+					 cd_progress_ptr progress));
+EXTERN(void) end_progress_monitor JPP((j_common_ptr cinfo));
+EXTERN(boolean) keymatch JPP((char * arg, const char * keyword, int minchars));
+EXTERN(FILE *) read_stdin JPP((void));
+EXTERN(FILE *) write_stdout JPP((void));
+
+/* miscellaneous useful macros */
+
+#ifdef DONT_USE_B_MODE		/* define mode parameters for fopen() */
+#define READ_BINARY	"r"
+#define WRITE_BINARY	"w"
+#else
+#ifdef VMS			/* VMS is very nonstandard */
+#define READ_BINARY	"rb", "ctx=stm"
+#define WRITE_BINARY	"wb", "ctx=stm"
+#else				/* standard ANSI-compliant case */
+#define READ_BINARY	"rb"
+#define WRITE_BINARY	"wb"
+#endif
+#endif
+
+#ifndef EXIT_FAILURE		/* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+#ifndef EXIT_SUCCESS
+#ifdef VMS
+#define EXIT_SUCCESS  1		/* VMS is very nonstandard */
+#else
+#define EXIT_SUCCESS  0
+#endif
+#endif
+#ifndef EXIT_WARNING
+#ifdef VMS
+#define EXIT_WARNING  1		/* VMS is very nonstandard */
+#else
+#define EXIT_WARNING  2
+#endif
+#endif
diff --git a/change.log b/change.log
new file mode 100644
index 0000000..c6e9399
--- /dev/null
+++ b/change.log
@@ -0,0 +1,275 @@
+CHANGE LOG for Independent JPEG Group's JPEG software
+
+
+Version 8b  16-May-2010
+-----------------------
+
+Repair problem in new memory source manager with corrupt JPEG data.
+Thank to Ted Campbell and Samuel Chun for the report.
+
+
+Version 8a  28-Feb-2010
+-----------------------
+
+Writing tables-only datastreams via jpeg_write_tables works again.
+
+Support 32-bit BMPs (RGB image with Alpha channel) for read in cjpeg.
+Thank to Brett Blackham for the suggestion.
+
+
+Version 8  10-Jan-2010
+----------------------
+
+Add sanity check in BMP reader module to avoid cjpeg crash for empty input
+image (thank to Isaev Ildar of ISP RAS, Moscow, RU for reporting this error).
+
+Add data source and destination managers for read from and write to
+memory buffers.  New API functions jpeg_mem_src and jpeg_mem_dest.
+Thank to Roberto Boni from Italy for the suggestion.
+
+
+Version 7  27-Jun-2009
+----------------------
+
+cjpeg -quality option has been extended for support of separate quality
+settings for luminance and chrominance (or in general, for every provided
+quantization table slot).
+New API function jpeg_default_qtables() and q_scale_factor array in library.
+
+jpegtran has a new "lossless" cropping feature.
+
+Implement -perfect option in jpegtran, new API function
+jtransform_perfect_transform() in transupp. (DP 204_perfect.dpatch)
+
+Better error messages for jpegtran fopen failure.
+(DP 203_jpegtran_errmsg.dpatch)
+
+Fix byte order issue with 16bit PPM/PGM files in rdppm.c/wrppm.c:
+according to Netpbm, the de facto standard implementation of the PNM formats,
+the most significant byte is first. (DP 203_rdppm.dpatch)
+
+Add -raw option to rdjpgcom not to mangle the output.
+(DP 205_rdjpgcom_raw.dpatch)
+
+Make rdjpgcom locale aware. (DP 201_rdjpgcom_locale.dpatch)
+
+Add extern "C" to jpeglib.h.
+This avoids the need to put extern "C" { ... } around #include "jpeglib.h"
+in your C++ application.  Defining the symbol DONT_USE_EXTERN_C in the
+configuration prevents this. (DP 202_jpeglib.h_c++.dpatch)
+
+
+Version 6b  27-Mar-1998
+-----------------------
+
+jpegtran has new features for lossless image transformations (rotation
+and flipping) as well as "lossless" reduction to grayscale.
+
+jpegtran now copies comments by default; it has a -copy switch to enable
+copying all APPn blocks as well, or to suppress comments.  (Formerly it
+always suppressed comments and APPn blocks.)  jpegtran now also preserves
+JFIF version and resolution information.
+
+New decompressor library feature: COM and APPn markers found in the input
+file can be saved in memory for later use by the application.  (Before,
+you had to code this up yourself with a custom marker processor.)
+
+There is an unused field "void * client_data" now in compress and decompress
+parameter structs; this may be useful in some applications.
+
+JFIF version number information is now saved by the decoder and accepted by
+the encoder.  jpegtran uses this to copy the source file's version number,
+to ensure "jpegtran -copy all" won't create bogus files that contain JFXX
+extensions but claim to be version 1.01.  Applications that generate their
+own JFXX extension markers also (finally) have a supported way to cause the
+encoder to emit JFIF version number 1.02.
+
+djpeg's trace mode reports JFIF 1.02 thumbnail images as such, rather
+than as unknown APP0 markers.
+
+In -verbose mode, djpeg and rdjpgcom will try to print the contents of
+APP12 markers as text.  Some digital cameras store useful text information
+in APP12 markers.
+
+Handling of truncated data streams is more robust: blocks beyond the one in
+which the error occurs will be output as uniform gray, or left unchanged
+if decoding a progressive JPEG.  The appearance no longer depends on the
+Huffman tables being used.
+
+Huffman tables are checked for validity much more carefully than before.
+
+To avoid the Unisys LZW patent, djpeg's GIF output capability has been
+changed to produce "uncompressed GIFs", and cjpeg's GIF input capability
+has been removed altogether.  We're not happy about it either, but there
+seems to be no good alternative.
+
+The configure script now supports building libjpeg as a shared library
+on many flavors of Unix (all the ones that GNU libtool knows how to
+build shared libraries for).  Use "./configure --enable-shared" to
+try this out.
+
+New jconfig file and makefiles for Microsoft Visual C++ and Developer Studio.
+Also, a jconfig file and a build script for Metrowerks CodeWarrior
+on Apple Macintosh.  makefile.dj has been updated for DJGPP v2, and there
+are miscellaneous other minor improvements in the makefiles.
+
+jmemmac.c now knows how to create temporary files following Mac System 7
+conventions.
+
+djpeg's -map switch is now able to read raw-format PPM files reliably.
+
+cjpeg -progressive -restart no longer generates any unnecessary DRI markers.
+
+Multiple calls to jpeg_simple_progression for a single JPEG object
+no longer leak memory.
+
+
+Version 6a  7-Feb-96
+--------------------
+
+Library initialization sequence modified to detect version mismatches
+and struct field packing mismatches between library and calling application.
+This change requires applications to be recompiled, but does not require
+any application source code change.
+
+All routine declarations changed to the style "GLOBAL(type) name ...",
+that is, GLOBAL, LOCAL, METHODDEF, EXTERN are now macros taking the
+routine's return type as an argument.  This makes it possible to add
+Microsoft-style linkage keywords to all the routines by changing just
+these macros.  Note that any application code that was using these macros
+will have to be changed.
+
+DCT coefficient quantization tables are now stored in normal array order
+rather than zigzag order.  Application code that calls jpeg_add_quant_table,
+or otherwise manipulates quantization tables directly, will need to be
+changed.  If you need to make such code work with either older or newer
+versions of the library, a test like "#if JPEG_LIB_VERSION >= 61" is
+recommended.
+
+djpeg's trace capability now dumps DQT tables in natural order, not zigzag
+order.  This allows the trace output to be made into a "-qtables" file
+more easily.
+
+New system-dependent memory manager module for use on Apple Macintosh.
+
+Fix bug in cjpeg's -smooth option: last one or two scanlines would be
+duplicates of the prior line unless the image height mod 16 was 1 or 2.
+
+Repair minor problems in VMS, BCC, MC6 makefiles.
+
+New configure script based on latest GNU Autoconf.
+
+Correct the list of include files needed by MetroWerks C for ccommand().
+
+Numerous small documentation updates.
+
+
+Version 6  2-Aug-95
+-------------------
+
+Progressive JPEG support: library can read and write full progressive JPEG
+files.  A "buffered image" mode supports incremental decoding for on-the-fly
+display of progressive images.  Simply recompiling an existing IJG-v5-based
+decoder with v6 should allow it to read progressive files, though of course
+without any special progressive display.
+
+New "jpegtran" application performs lossless transcoding between different
+JPEG formats; primarily, it can be used to convert baseline to progressive
+JPEG and vice versa.  In support of jpegtran, the library now allows lossless
+reading and writing of JPEG files as DCT coefficient arrays.  This ability
+may be of use in other applications.
+
+Notes for programmers:
+* We changed jpeg_start_decompress() to be able to suspend; this makes all
+decoding modes available to suspending-input applications.  However,
+existing applications that use suspending input will need to be changed
+to check the return value from jpeg_start_decompress().  You don't need to
+do anything if you don't use a suspending data source.
+* We changed the interface to the virtual array routines: access_virt_array
+routines now take a count of the number of rows to access this time.  The
+last parameter to request_virt_array routines is now interpreted as the
+maximum number of rows that may be accessed at once, but not necessarily
+the height of every access.
+
+
+Version 5b  15-Mar-95
+---------------------
+
+Correct bugs with grayscale images having v_samp_factor > 1.
+
+jpeg_write_raw_data() now supports output suspension.
+
+Correct bugs in "configure" script for case of compiling in
+a directory other than the one containing the source files.
+
+Repair bug in jquant1.c: sometimes didn't use as many colors as it could.
+
+Borland C makefile and jconfig file work under either MS-DOS or OS/2.
+
+Miscellaneous improvements to documentation.
+
+
+Version 5a  7-Dec-94
+--------------------
+
+Changed color conversion roundoff behavior so that grayscale values are
+represented exactly.  (This causes test image files to change.)
+
+Make ordered dither use 16x16 instead of 4x4 pattern for a small quality
+improvement.
+
+New configure script based on latest GNU Autoconf.
+Fix configure script to handle CFLAGS correctly.
+Rename *.auto files to *.cfg, so that configure script still works if
+file names have been truncated for DOS.
+
+Fix bug in rdbmp.c: didn't allow for extra data between header and image.
+
+Modify rdppm.c/wrppm.c to handle 2-byte raw PPM/PGM formats for 12-bit data.
+
+Fix several bugs in rdrle.c.
+
+NEED_SHORT_EXTERNAL_NAMES option was broken.
+
+Revise jerror.h/jerror.c for more flexibility in message table.
+
+Repair oversight in jmemname.c NO_MKTEMP case: file could be there
+but unreadable.
+
+
+Version 5  24-Sep-94
+--------------------
+
+Version 5 represents a nearly complete redesign and rewrite of the IJG
+software.  Major user-visible changes include:
+  * Automatic configuration simplifies installation for most Unix systems.
+  * A range of speed vs. image quality tradeoffs are supported.
+    This includes resizing of an image during decompression: scaling down
+    by a factor of 1/2, 1/4, or 1/8 is handled very efficiently.
+  * New programs rdjpgcom and wrjpgcom allow insertion and extraction
+    of text comments in a JPEG file.
+
+The application programmer's interface to the library has changed completely.
+Notable improvements include:
+  * We have eliminated the use of callback routines for handling the
+    uncompressed image data.  The application now sees the library as a
+    set of routines that it calls to read or write image data on a
+    scanline-by-scanline basis.
+  * The application image data is represented in a conventional interleaved-
+    pixel format, rather than as a separate array for each color channel.
+    This can save a copying step in many programs.
+  * The handling of compressed data has been cleaned up: the application can
+    supply routines to source or sink the compressed data.  It is possible to
+    suspend processing on source/sink buffer overrun, although this is not
+    supported in all operating modes.
+  * All static state has been eliminated from the library, so that multiple
+    instances of compression or decompression can be active concurrently.
+  * JPEG abbreviated datastream formats are supported, ie, quantization and
+    Huffman tables can be stored separately from the image data.
+  * And not only that, but the documentation of the library has improved
+    considerably!
+
+
+The last widely used release before the version 5 rewrite was version 4A of
+18-Feb-93.  Change logs before that point have been discarded, since they
+are not of much interest after the rewrite.
diff --git a/cjpeg.1 b/cjpeg.1
new file mode 100644
index 0000000..63a4e39
--- /dev/null
+++ b/cjpeg.1
@@ -0,0 +1,318 @@
+.TH CJPEG 1 "11 October 2010"
+.SH NAME
+cjpeg \- compress an image file to a JPEG file
+.SH SYNOPSIS
+.B cjpeg
+[
+.I options
+]
+[
+.I filename
+]
+.LP
+.SH DESCRIPTION
+.LP
+.B cjpeg
+compresses the named image file, or the standard input if no file is
+named, and produces a JPEG/JFIF file on the standard output.
+The currently supported input file formats are: PPM (PBMPLUS color
+format), PGM (PBMPLUS gray-scale format), BMP, Targa, and RLE (Utah Raster
+Toolkit format).  (RLE is supported only if the URT library is available.)
+.SH OPTIONS
+All switch names may be abbreviated; for example,
+.B \-grayscale
+may be written
+.B \-gray
+or
+.BR \-gr .
+Most of the "basic" switches can be abbreviated to as little as one letter.
+Upper and lower case are equivalent (thus
+.B \-BMP
+is the same as
+.BR \-bmp ).
+British spellings are also accepted (e.g.,
+.BR \-greyscale ),
+though for brevity these are not mentioned below.
+.PP
+The basic switches are:
+.TP
+.BI \-quality " N[,...]"
+Scale quantization tables to adjust image quality.  Quality is 0 (worst) to
+100 (best); default is 75.  (See below for more info.)
+.TP
+.B \-grayscale
+Create monochrome JPEG file from color input.  Be sure to use this switch when
+compressing a grayscale BMP file, because
+.B cjpeg
+isn't bright enough to notice whether a BMP file uses only shades of gray.
+By saying
+.BR \-grayscale ,
+you'll get a smaller JPEG file that takes less time to process.
+.TP
+.B \-optimize
+Perform optimization of entropy encoding parameters.  Without this, default
+encoding parameters are used.
+.B \-optimize
+usually makes the JPEG file a little smaller, but
+.B cjpeg
+runs somewhat slower and needs much more memory.  Image quality and speed of
+decompression are unaffected by
+.BR \-optimize .
+.TP
+.B \-progressive
+Create progressive JPEG file (see below).
+.TP
+.B \-targa
+Input file is Targa format.  Targa files that contain an "identification"
+field will not be automatically recognized by
+.BR cjpeg ;
+for such files you must specify
+.B \-targa
+to make
+.B cjpeg
+treat the input as Targa format.
+For most Targa files, you won't need this switch.
+.PP
+The
+.B \-quality
+switch lets you trade off compressed file size against quality of the
+reconstructed image: the higher the quality setting, the larger the JPEG file,
+and the closer the output image will be to the original input.  Normally you
+want to use the lowest quality setting (smallest file) that decompresses into
+something visually indistinguishable from the original image.  For this
+purpose the quality setting should be between 50 and 95; the default of 75 is
+often about right.  If you see defects at
+.B \-quality
+75, then go up 5 or 10 counts at a time until you are happy with the output
+image.  (The optimal setting will vary from one image to another.)
+.PP
+.B \-quality
+100 will generate a quantization table of all 1's, minimizing loss in the
+quantization step (but there is still information loss in subsampling, as well
+as roundoff error).  This setting is mainly of interest for experimental
+purposes.  Quality values above about 95 are
+.B not
+recommended for normal use; the compressed file size goes up dramatically for
+hardly any gain in output image quality.
+.PP
+In the other direction, quality values below 50 will produce very small files
+of low image quality.  Settings around 5 to 10 might be useful in preparing an
+index of a large image library, for example.  Try
+.B \-quality
+2 (or so) for some amusing Cubist effects.  (Note: quality
+values below about 25 generate 2-byte quantization tables, which are
+considered optional in the JPEG standard.
+.B cjpeg
+emits a warning message when you give such a quality value, because some
+other JPEG programs may be unable to decode the resulting file.  Use
+.B \-baseline
+if you need to ensure compatibility at low quality values.)
+.PP
+The \fB-quality\fR option has been extended in this version of \fBcjpeg\fR to
+support separate quality settings for luminance and chrominance (or, in
+general, separate settings for every quantization table slot.)  The principle
+is the same as chrominance subsampling:  since the human eye is more sensitive
+to spatial changes in brightness than spatial changes in color, the chrominance
+components can be quantized more than the luminance components without
+incurring any visible image quality loss.  However, unlike subsampling, this
+feature reduces data in the frequency domain instead of the spatial domain,
+which allows for more fine-grained control.  This option is useful in
+quality-sensitive applications, for which the artifacts generated by
+subsampling may be unacceptable.
+.PP
+The \fB-quality\fR option accepts a comma-separated list of parameters, which
+respectively refer to the quality levels which should be assigned to the
+quantization table slots.  If there are more q-table slots than parameters,
+then the last parameter is replicated.  Thus, if only one quality parameter is
+given, this is used for both luminance and chrominance (slots 0 and 1,
+respectively), preserving the legacy behavior of cjpeg v6b and prior.
+More (or customized) quantization tables can be set with the \fB-qtables\fR
+option and assigned to components with the \fB-qslots\fR option (see the
+"wizard" switches below.)
+.PP
+JPEG files generated with separate luminance and chrominance quality are fully
+compliant with standard JPEG decoders.
+.PP
+.BR CAUTION:
+For this setting to be useful, be sure to pass an argument of \fB-sample 1x1\fR
+to \fBcjpeg\fR to disable chrominance subsampling.  Otherwise, the default
+subsampling level (2x2, AKA "4:2:0") will be used.
+.PP
+The
+.B \-progressive
+switch creates a "progressive JPEG" file.  In this type of JPEG file, the data
+is stored in multiple scans of increasing quality.  If the file is being
+transmitted over a slow communications link, the decoder can use the first
+scan to display a low-quality image very quickly, and can then improve the
+display with each subsequent scan.  The final image is exactly equivalent to a
+standard JPEG file of the same quality setting, and the total file size is
+about the same --- often a little smaller.
+.PP
+Switches for advanced users:
+.TP
+.B \-dct int
+Use integer DCT method (default).
+.TP
+.B \-dct fast
+Use fast integer DCT (less accurate).
+.TP
+.B \-dct float
+Use floating-point DCT method.
+The float method is very slightly more accurate than the int method, but is
+much slower unless your machine has very fast floating-point hardware.  Also
+note that results of the floating-point method may vary slightly across
+machines, while the integer methods should give the same results everywhere.
+The fast integer method is much less accurate than the other two.
+.TP
+.BI \-restart " N"
+Emit a JPEG restart marker every N MCU rows, or every N MCU blocks if "B" is
+attached to the number.
+.B \-restart 0
+(the default) means no restart markers.
+.TP
+.BI \-smooth " N"
+Smooth the input image to eliminate dithering noise.  N, ranging from 1 to
+100, indicates the strength of smoothing.  0 (the default) means no smoothing.
+.TP
+.BI \-maxmemory " N"
+Set limit for amount of memory to use in processing large images.  Value is
+in thousands of bytes, or millions of bytes if "M" is attached to the
+number.  For example,
+.B \-max 4m
+selects 4000000 bytes.  If more space is needed, temporary files will be used.
+.TP
+.BI \-outfile " name"
+Send output image to the named file, not to standard output.
+.TP
+.B \-verbose
+Enable debug printout.  More
+.BR \-v 's
+give more output.  Also, version information is printed at startup.
+.TP
+.B \-debug
+Same as
+.BR \-verbose .
+.PP
+The
+.B \-restart
+option inserts extra markers that allow a JPEG decoder to resynchronize after
+a transmission error.  Without restart markers, any damage to a compressed
+file will usually ruin the image from the point of the error to the end of the
+image; with restart markers, the damage is usually confined to the portion of
+the image up to the next restart marker.  Of course, the restart markers
+occupy extra space.  We recommend
+.B \-restart 1
+for images that will be transmitted across unreliable networks such as Usenet.
+.PP
+The
+.B \-smooth
+option filters the input to eliminate fine-scale noise.  This is often useful
+when converting dithered images to JPEG: a moderate smoothing factor of 10 to
+50 gets rid of dithering patterns in the input file, resulting in a smaller
+JPEG file and a better-looking image.  Too large a smoothing factor will
+visibly blur the image, however.
+.PP
+Switches for wizards:
+.TP
+.B \-baseline
+Force baseline-compatible quantization tables to be generated.  This clamps
+quantization values to 8 bits even at low quality settings.  (This switch is
+poorly named, since it does not ensure that the output is actually baseline
+JPEG.  For example, you can use
+.B \-baseline
+and
+.B \-progressive
+together.)
+.TP
+.BI \-qtables " file"
+Use the quantization tables given in the specified text file.
+.TP
+.BI \-qslots " N[,...]"
+Select which quantization table to use for each color component.
+.TP
+.BI \-sample " HxV[,...]"
+Set JPEG sampling factors for each color component.
+.TP
+.BI \-scans " file"
+Use the scan script given in the specified text file.
+.PP
+The "wizard" switches are intended for experimentation with JPEG.  If you
+don't know what you are doing, \fBdon't use them\fR.  These switches are
+documented further in the file wizard.txt.
+.SH EXAMPLES
+.LP
+This example compresses the PPM file foo.ppm with a quality factor of
+60 and saves the output as foo.jpg:
+.IP
+.B cjpeg \-quality
+.I 60 foo.ppm
+.B >
+.I foo.jpg
+.SH HINTS
+Color GIF files are not the ideal input for JPEG; JPEG is really intended for
+compressing full-color (24-bit) images.  In particular, don't try to convert
+cartoons, line drawings, and other images that have only a few distinct
+colors.  GIF works great on these, JPEG does not.  If you want to convert a
+GIF to JPEG, you should experiment with
+.BR cjpeg 's
+.B \-quality
+and
+.B \-smooth
+options to get a satisfactory conversion.
+.B \-smooth 10
+or so is often helpful.
+.PP
+Avoid running an image through a series of JPEG compression/decompression
+cycles.  Image quality loss will accumulate; after ten or so cycles the image
+may be noticeably worse than it was after one cycle.  It's best to use a
+lossless format while manipulating an image, then convert to JPEG format when
+you are ready to file the image away.
+.PP
+The
+.B \-optimize
+option to
+.B cjpeg
+is worth using when you are making a "final" version for posting or archiving.
+It's also a win when you are using low quality settings to make very small
+JPEG files; the percentage improvement is often a lot more than it is on
+larger files.  (At present,
+.B \-optimize
+mode is always selected when generating progressive JPEG files.)
+.SH ENVIRONMENT
+.TP
+.B JPEGMEM
+If this environment variable is set, its value is the default memory limit.
+The value is specified as described for the
+.B \-maxmemory
+switch.
+.B JPEGMEM
+overrides the default value specified when the program was compiled, and
+itself is overridden by an explicit
+.BR \-maxmemory .
+.SH SEE ALSO
+.BR djpeg (1),
+.BR jpegtran (1),
+.BR rdjpgcom (1),
+.BR wrjpgcom (1)
+.br
+.BR ppm (5),
+.BR pgm (5)
+.br
+Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+Communications of the ACM, April 1991 (vol. 34, no. 4), pp. 30-44.
+.SH AUTHOR
+Independent JPEG Group
+.SH BUGS
+Arithmetic coding is not supported for legal reasons.
+.PP
+Support for GIF input files was removed in cjpeg v6b due to concerns over
+the Unisys LZW patent.  Although this patent expired in 2006, cjpeg still
+lacks GIF support, for these historical reasons.  (Conversion of GIF files to
+JPEG is usually a bad idea anyway.)
+.PP
+Not all variants of BMP and Targa file formats are supported.
+.PP
+The
+.B \-targa
+switch is not a bug, it's a feature.  (It would be a bug if the Targa format
+designers had not been clueless.)
diff --git a/cjpeg.c b/cjpeg.c
new file mode 100644
index 0000000..3daa259
--- /dev/null
+++ b/cjpeg.c
@@ -0,0 +1,605 @@
+/*
+ * cjpeg.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2003-2008 by Guido Vollbeding.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a command-line user interface for the JPEG compressor.
+ * It should work on any system with Unix- or MS-DOS-style command lines.
+ *
+ * Two different command line styles are permitted, depending on the
+ * compile-time switch TWO_FILE_COMMANDLINE:
+ *	cjpeg [options]  inputfile outputfile
+ *	cjpeg [options]  [inputfile]
+ * In the second style, output is always to standard output, which you'd
+ * normally redirect to a file or pipe to some other program.  Input is
+ * either from a named file or from standard input (typically redirected).
+ * The second style is convenient on Unix but is unhelpful on systems that
+ * don't support pipes.  Also, you MUST use the first style if your system
+ * doesn't do binary I/O to stdin/stdout.
+ * To simplify script writing, the "-outfile" switch is provided.  The syntax
+ *	cjpeg [options]  -outfile outputfile  inputfile
+ * works regardless of which command line style is used.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+#include "jversion.h"		/* for version message */
+#include "config.h"
+
+#ifdef USE_CCOMMAND		/* command-line reader for Macintosh */
+#ifdef __MWERKS__
+#include <SIOUX.h>              /* Metrowerks needs this */
+#include <console.h>		/* ... and this */
+#endif
+#ifdef THINK_C
+#include <console.h>		/* Think declares it here */
+#endif
+#endif
+
+
+/* Create the add-on message string table. */
+
+#define JMESSAGE(code,string)	string ,
+
+static const char * const cdjpeg_message_table[] = {
+#include "cderror.h"
+  NULL
+};
+
+
+/*
+ * This routine determines what format the input file is,
+ * and selects the appropriate input-reading module.
+ *
+ * To determine which family of input formats the file belongs to,
+ * we may look only at the first byte of the file, since C does not
+ * guarantee that more than one character can be pushed back with ungetc.
+ * Looking at additional bytes would require one of these approaches:
+ *     1) assume we can fseek() the input file (fails for piped input);
+ *     2) assume we can push back more than one character (works in
+ *        some C implementations, but unportable);
+ *     3) provide our own buffering (breaks input readers that want to use
+ *        stdio directly, such as the RLE library);
+ * or  4) don't put back the data, and modify the input_init methods to assume
+ *        they start reading after the start of file (also breaks RLE library).
+ * #1 is attractive for MS-DOS but is untenable on Unix.
+ *
+ * The most portable solution for file types that can't be identified by their
+ * first byte is to make the user tell us what they are.  This is also the
+ * only approach for "raw" file types that contain only arbitrary values.
+ * We presently apply this method for Targa files.  Most of the time Targa
+ * files start with 0x00, so we recognize that case.  Potentially, however,
+ * a Targa file could start with any byte value (byte 0 is the length of the
+ * seldom-used ID field), so we provide a switch to force Targa input mode.
+ */
+
+static boolean is_targa;	/* records user -targa switch */
+
+
+LOCAL(cjpeg_source_ptr)
+select_file_type (j_compress_ptr cinfo, FILE * infile)
+{
+  int c;
+
+  if (is_targa) {
+#ifdef TARGA_SUPPORTED
+    return jinit_read_targa(cinfo);
+#else
+    ERREXIT(cinfo, JERR_TGA_NOTCOMP);
+#endif
+  }
+
+  if ((c = getc(infile)) == EOF)
+    ERREXIT(cinfo, JERR_INPUT_EMPTY);
+  if (ungetc(c, infile) == EOF)
+    ERREXIT(cinfo, JERR_UNGETC_FAILED);
+
+  switch (c) {
+#ifdef BMP_SUPPORTED
+  case 'B':
+    return jinit_read_bmp(cinfo);
+#endif
+#ifdef GIF_SUPPORTED
+  case 'G':
+    return jinit_read_gif(cinfo);
+#endif
+#ifdef PPM_SUPPORTED
+  case 'P':
+    return jinit_read_ppm(cinfo);
+#endif
+#ifdef RLE_SUPPORTED
+  case 'R':
+    return jinit_read_rle(cinfo);
+#endif
+#ifdef TARGA_SUPPORTED
+  case 0x00:
+    return jinit_read_targa(cinfo);
+#endif
+  default:
+    ERREXIT(cinfo, JERR_UNKNOWN_FORMAT);
+    break;
+  }
+
+  return NULL;			/* suppress compiler warnings */
+}
+
+
+/*
+ * Argument-parsing code.
+ * The switch parser is designed to be useful with DOS-style command line
+ * syntax, ie, intermixed switches and file names, where only the switches
+ * to the left of a given file name affect processing of that file.
+ * The main program in this file doesn't actually use this capability...
+ */
+
+
+static const char * progname;	/* program name for error messages */
+static char * outfilename;	/* for -outfile switch */
+
+
+LOCAL(void)
+usage (void)
+/* complain about bad command line */
+{
+  fprintf(stderr, "usage: %s [switches] ", progname);
+#ifdef TWO_FILE_COMMANDLINE
+  fprintf(stderr, "inputfile outputfile\n");
+#else
+  fprintf(stderr, "[inputfile]\n");
+#endif
+
+  fprintf(stderr, "Switches (names may be abbreviated):\n");
+  fprintf(stderr, "  -quality N[,...]   Compression quality (0..100; 5-95 is useful range)\n");
+  fprintf(stderr, "  -grayscale     Create monochrome JPEG file\n");
+#ifdef ENTROPY_OPT_SUPPORTED
+  fprintf(stderr, "  -optimize      Optimize Huffman table (smaller file, but slow compression)\n");
+#endif
+#ifdef C_PROGRESSIVE_SUPPORTED
+  fprintf(stderr, "  -progressive   Create progressive JPEG file\n");
+#endif
+#ifdef TARGA_SUPPORTED
+  fprintf(stderr, "  -targa         Input file is Targa format (usually not needed)\n");
+#endif
+  fprintf(stderr, "Switches for advanced users:\n");
+#ifdef DCT_ISLOW_SUPPORTED
+  fprintf(stderr, "  -dct int       Use integer DCT method%s\n",
+	  (JDCT_DEFAULT == JDCT_ISLOW ? " (default)" : ""));
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  fprintf(stderr, "  -dct fast      Use fast integer DCT (less accurate)%s\n",
+	  (JDCT_DEFAULT == JDCT_IFAST ? " (default)" : ""));
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  fprintf(stderr, "  -dct float     Use floating-point DCT method%s\n",
+	  (JDCT_DEFAULT == JDCT_FLOAT ? " (default)" : ""));
+#endif
+  fprintf(stderr, "  -restart N     Set restart interval in rows, or in blocks with B\n");
+#ifdef INPUT_SMOOTHING_SUPPORTED
+  fprintf(stderr, "  -smooth N      Smooth dithered input (N=1..100 is strength)\n");
+#endif
+  fprintf(stderr, "  -maxmemory N   Maximum memory to use (in kbytes)\n");
+  fprintf(stderr, "  -outfile name  Specify name for output file\n");
+  fprintf(stderr, "  -verbose  or  -debug   Emit debug output\n");
+  fprintf(stderr, "Switches for wizards:\n");
+#ifdef C_ARITH_CODING_SUPPORTED
+  fprintf(stderr, "  -arithmetic    Use arithmetic coding\n");
+#endif
+  fprintf(stderr, "  -baseline      Force baseline quantization tables\n");
+  fprintf(stderr, "  -qtables file  Use quantization tables given in file\n");
+  fprintf(stderr, "  -qslots N[,...]    Set component quantization tables\n");
+  fprintf(stderr, "  -sample HxV[,...]  Set component sampling factors\n");
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+  fprintf(stderr, "  -scans file    Create multi-scan JPEG per script file\n");
+#endif
+  exit(EXIT_FAILURE);
+}
+
+
+LOCAL(int)
+parse_switches (j_compress_ptr cinfo, int argc, char **argv,
+		int last_file_arg_seen, boolean for_real)
+/* Parse optional switches.
+ * Returns argv[] index of first file-name argument (== argc if none).
+ * Any file names with indexes <= last_file_arg_seen are ignored;
+ * they have presumably been processed in a previous iteration.
+ * (Pass 0 for last_file_arg_seen on the first or only iteration.)
+ * for_real is FALSE on the first (dummy) pass; we may skip any expensive
+ * processing.
+ */
+{
+  int argn;
+  char * arg;
+  boolean force_baseline;
+  boolean simple_progressive;
+  char * qualityarg = NULL;	/* saves -quality parm if any */
+  char * qtablefile = NULL;	/* saves -qtables filename if any */
+  char * qslotsarg = NULL;	/* saves -qslots parm if any */
+  char * samplearg = NULL;	/* saves -sample parm if any */
+  char * scansarg = NULL;	/* saves -scans parm if any */
+
+  /* Set up default JPEG parameters. */
+
+  force_baseline = FALSE;	/* by default, allow 16-bit quantizers */
+  simple_progressive = FALSE;
+  is_targa = FALSE;
+  outfilename = NULL;
+  cinfo->err->trace_level = 0;
+
+  /* Scan command line options, adjust parameters */
+
+  for (argn = 1; argn < argc; argn++) {
+    arg = argv[argn];
+    if (*arg != '-') {
+      /* Not a switch, must be a file name argument */
+      if (argn <= last_file_arg_seen) {
+	outfilename = NULL;	/* -outfile applies to just one input file */
+	continue;		/* ignore this name if previously processed */
+      }
+      break;			/* else done parsing switches */
+    }
+    arg++;			/* advance past switch marker character */
+
+    if (keymatch(arg, "arithmetic", 1)) {
+      /* Use arithmetic coding. */
+#ifdef C_ARITH_CODING_SUPPORTED
+      cinfo->arith_code = TRUE;
+#else
+      fprintf(stderr, "%s: sorry, arithmetic coding not supported\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "baseline", 1)) {
+      /* Force baseline-compatible output (8-bit quantizer values). */
+      force_baseline = TRUE;
+
+    } else if (keymatch(arg, "dct", 2)) {
+      /* Select DCT algorithm. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (keymatch(argv[argn], "int", 1)) {
+	cinfo->dct_method = JDCT_ISLOW;
+      } else if (keymatch(argv[argn], "fast", 2)) {
+	cinfo->dct_method = JDCT_IFAST;
+      } else if (keymatch(argv[argn], "float", 2)) {
+	cinfo->dct_method = JDCT_FLOAT;
+      } else
+	usage();
+
+    } else if (keymatch(arg, "debug", 1) || keymatch(arg, "verbose", 1)) {
+      /* Enable debug printouts. */
+      /* On first -d, print version identification */
+      static boolean printed_version = FALSE;
+
+      if (! printed_version) {
+	fprintf(stderr, "%s version %s (build %s)\n",
+		PACKAGE_NAME, VERSION, BUILD);
+	fprintf(stderr, "%s\n\n", LJTCOPYRIGHT);
+	fprintf(stderr, "Based on Independent JPEG Group's libjpeg, version %s\n%s\n\n",
+		JVERSION, JCOPYRIGHT);
+	printed_version = TRUE;
+      }
+      cinfo->err->trace_level++;
+
+    } else if (keymatch(arg, "grayscale", 2) || keymatch(arg, "greyscale",2)) {
+      /* Force a monochrome JPEG file to be generated. */
+      jpeg_set_colorspace(cinfo, JCS_GRAYSCALE);
+
+    } else if (keymatch(arg, "maxmemory", 3)) {
+      /* Maximum memory in Kb (or Mb with 'm'). */
+      long lval;
+      char ch = 'x';
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%ld%c", &lval, &ch) < 1)
+	usage();
+      if (ch == 'm' || ch == 'M')
+	lval *= 1000L;
+      cinfo->mem->max_memory_to_use = lval * 1000L;
+
+    } else if (keymatch(arg, "optimize", 1) || keymatch(arg, "optimise", 1)) {
+      /* Enable entropy parm optimization. */
+#ifdef ENTROPY_OPT_SUPPORTED
+      cinfo->optimize_coding = TRUE;
+#else
+      fprintf(stderr, "%s: sorry, entropy optimization was not compiled\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "outfile", 4)) {
+      /* Set output file name. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      outfilename = argv[argn];	/* save it away for later use */
+
+    } else if (keymatch(arg, "progressive", 1)) {
+      /* Select simple progressive mode. */
+#ifdef C_PROGRESSIVE_SUPPORTED
+      simple_progressive = TRUE;
+      /* We must postpone execution until num_components is known. */
+#else
+      fprintf(stderr, "%s: sorry, progressive output was not compiled\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "quality", 1)) {
+      /* Quality ratings (quantization table scaling factors). */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      qualityarg = argv[argn];
+
+    } else if (keymatch(arg, "qslots", 2)) {
+      /* Quantization table slot numbers. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      qslotsarg = argv[argn];
+      /* Must delay setting qslots until after we have processed any
+       * colorspace-determining switches, since jpeg_set_colorspace sets
+       * default quant table numbers.
+       */
+
+    } else if (keymatch(arg, "qtables", 2)) {
+      /* Quantization tables fetched from file. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      qtablefile = argv[argn];
+      /* We postpone actually reading the file in case -quality comes later. */
+
+    } else if (keymatch(arg, "restart", 1)) {
+      /* Restart interval in MCU rows (or in MCUs with 'b'). */
+      long lval;
+      char ch = 'x';
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%ld%c", &lval, &ch) < 1)
+	usage();
+      if (lval < 0 || lval > 65535L)
+	usage();
+      if (ch == 'b' || ch == 'B') {
+	cinfo->restart_interval = (unsigned int) lval;
+	cinfo->restart_in_rows = 0; /* else prior '-restart n' overrides me */
+      } else {
+	cinfo->restart_in_rows = (int) lval;
+	/* restart_interval will be computed during startup */
+      }
+
+    } else if (keymatch(arg, "sample", 2)) {
+      /* Set sampling factors. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      samplearg = argv[argn];
+      /* Must delay setting sample factors until after we have processed any
+       * colorspace-determining switches, since jpeg_set_colorspace sets
+       * default sampling factors.
+       */
+
+    } else if (keymatch(arg, "scans", 4)) {
+      /* Set scan script. */
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      scansarg = argv[argn];
+      /* We must postpone reading the file in case -progressive appears. */
+#else
+      fprintf(stderr, "%s: sorry, multi-scan output was not compiled\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "smooth", 2)) {
+      /* Set input smoothing factor. */
+      int val;
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%d", &val) != 1)
+	usage();
+      if (val < 0 || val > 100)
+	usage();
+      cinfo->smoothing_factor = val;
+
+    } else if (keymatch(arg, "targa", 1)) {
+      /* Input file is Targa format. */
+      is_targa = TRUE;
+
+    } else {
+      usage();			/* bogus switch */
+    }
+  }
+
+  /* Post-switch-scanning cleanup */
+
+  if (for_real) {
+
+    /* Set quantization tables for selected quality. */
+    /* Some or all may be overridden if -qtables is present. */
+    if (qualityarg != NULL)	/* process -quality if it was present */
+      if (! set_quality_ratings(cinfo, qualityarg, force_baseline))
+	usage();
+
+    if (qtablefile != NULL)	/* process -qtables if it was present */
+      if (! read_quant_tables(cinfo, qtablefile, force_baseline))
+	usage();
+
+    if (qslotsarg != NULL)	/* process -qslots if it was present */
+      if (! set_quant_slots(cinfo, qslotsarg))
+	usage();
+
+    if (samplearg != NULL)	/* process -sample if it was present */
+      if (! set_sample_factors(cinfo, samplearg))
+	usage();
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+    if (simple_progressive)	/* process -progressive; -scans can override */
+      jpeg_simple_progression(cinfo);
+#endif
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+    if (scansarg != NULL)	/* process -scans if it was present */
+      if (! read_scan_script(cinfo, scansarg))
+	usage();
+#endif
+  }
+
+  return argn;			/* return index of next arg (file name) */
+}
+
+
+/*
+ * The main program.
+ */
+
+int
+main (int argc, char **argv)
+{
+  struct jpeg_compress_struct cinfo;
+  struct jpeg_error_mgr jerr;
+#ifdef PROGRESS_REPORT
+  struct cdjpeg_progress_mgr progress;
+#endif
+  int file_index;
+  cjpeg_source_ptr src_mgr;
+  FILE * input_file;
+  FILE * output_file;
+  JDIMENSION num_scanlines;
+
+  /* On Mac, fetch a command line. */
+#ifdef USE_CCOMMAND
+  argc = ccommand(&argv);
+#endif
+
+  progname = argv[0];
+  if (progname == NULL || progname[0] == 0)
+    progname = "cjpeg";		/* in case C library doesn't provide it */
+
+  /* Initialize the JPEG compression object with default error handling. */
+  cinfo.err = jpeg_std_error(&jerr);
+  jpeg_create_compress(&cinfo);
+  /* Add some application-specific error messages (from cderror.h) */
+  jerr.addon_message_table = cdjpeg_message_table;
+  jerr.first_addon_message = JMSG_FIRSTADDONCODE;
+  jerr.last_addon_message = JMSG_LASTADDONCODE;
+
+  /* Now safe to enable signal catcher. */
+#ifdef NEED_SIGNAL_CATCHER
+  enable_signal_catcher((j_common_ptr) &cinfo);
+#endif
+
+  /* Initialize JPEG parameters.
+   * Much of this may be overridden later.
+   * In particular, we don't yet know the input file's color space,
+   * but we need to provide some value for jpeg_set_defaults() to work.
+   */
+
+  cinfo.in_color_space = JCS_RGB; /* arbitrary guess */
+  jpeg_set_defaults(&cinfo);
+
+  /* Scan command line to find file names.
+   * It is convenient to use just one switch-parsing routine, but the switch
+   * values read here are ignored; we will rescan the switches after opening
+   * the input file.
+   */
+
+  file_index = parse_switches(&cinfo, argc, argv, 0, FALSE);
+
+#ifdef TWO_FILE_COMMANDLINE
+  /* Must have either -outfile switch or explicit output file name */
+  if (outfilename == NULL) {
+    if (file_index != argc-2) {
+      fprintf(stderr, "%s: must name one input and one output file\n",
+	      progname);
+      usage();
+    }
+    outfilename = argv[file_index+1];
+  } else {
+    if (file_index != argc-1) {
+      fprintf(stderr, "%s: must name one input and one output file\n",
+	      progname);
+      usage();
+    }
+  }
+#else
+  /* Unix style: expect zero or one file name */
+  if (file_index < argc-1) {
+    fprintf(stderr, "%s: only one input file\n", progname);
+    usage();
+  }
+#endif /* TWO_FILE_COMMANDLINE */
+
+  /* Open the input file. */
+  if (file_index < argc) {
+    if ((input_file = fopen(argv[file_index], READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s\n", progname, argv[file_index]);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default input file is stdin */
+    input_file = read_stdin();
+  }
+
+  /* Open the output file. */
+  if (outfilename != NULL) {
+    if ((output_file = fopen(outfilename, WRITE_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s\n", progname, outfilename);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default output file is stdout */
+    output_file = write_stdout();
+  }
+
+#ifdef PROGRESS_REPORT
+  start_progress_monitor((j_common_ptr) &cinfo, &progress);
+#endif
+
+  /* Figure out the input file format, and set up to read it. */
+  src_mgr = select_file_type(&cinfo, input_file);
+  src_mgr->input_file = input_file;
+
+  /* Read the input file header to obtain file size & colorspace. */
+  (*src_mgr->start_input) (&cinfo, src_mgr);
+
+  /* Now that we know input colorspace, fix colorspace-dependent defaults */
+  jpeg_default_colorspace(&cinfo);
+
+  /* Adjust default compression parameters by re-parsing the options */
+  file_index = parse_switches(&cinfo, argc, argv, 0, TRUE);
+
+  /* Specify data destination for compression */
+  jpeg_stdio_dest(&cinfo, output_file);
+
+  /* Start compressor */
+  jpeg_start_compress(&cinfo, TRUE);
+
+  /* Process data */
+  while (cinfo.next_scanline < cinfo.image_height) {
+    num_scanlines = (*src_mgr->get_pixel_rows) (&cinfo, src_mgr);
+    (void) jpeg_write_scanlines(&cinfo, src_mgr->buffer, num_scanlines);
+  }
+
+  /* Finish compression and release memory */
+  (*src_mgr->finish_input) (&cinfo, src_mgr);
+  jpeg_finish_compress(&cinfo);
+  jpeg_destroy_compress(&cinfo);
+
+  /* Close files, if we opened them */
+  if (input_file != stdin)
+    fclose(input_file);
+  if (output_file != stdout)
+    fclose(output_file);
+
+#ifdef PROGRESS_REPORT
+  end_progress_monitor((j_common_ptr) &cinfo);
+#endif
+
+  /* All done. */
+  exit(jerr.num_warnings ? EXIT_WARNING : EXIT_SUCCESS);
+  return 0;			/* suppress no-return-value warnings */
+}
diff --git a/cmakescripts/getdate.bat b/cmakescripts/getdate.bat
new file mode 100644
index 0000000..b4251bb
--- /dev/null
+++ b/cmakescripts/getdate.bat
@@ -0,0 +1,3 @@
+@echo off
+for /f "tokens=1-4 eol=/ DELIMS=/ " %%i in ('date /t') do set BUILD=%%l%%j%%k
+echo %BUILD%
diff --git a/coderules.txt b/coderules.txt
new file mode 100644
index 0000000..357929f
--- /dev/null
+++ b/coderules.txt
@@ -0,0 +1,118 @@
+IJG JPEG LIBRARY:  CODING RULES
+
+Copyright (C) 1991-1996, Thomas G. Lane.
+This file is part of the Independent JPEG Group's software.
+For conditions of distribution and use, see the accompanying README file.
+
+
+Since numerous people will be contributing code and bug fixes, it's important
+to establish a common coding style.  The goal of using similar coding styles
+is much more important than the details of just what that style is.
+
+In general we follow the recommendations of "Recommended C Style and Coding
+Standards" revision 6.1 (Cannon et al. as modified by Spencer, Keppel and
+Brader).  This document is available in the IJG FTP archive (see
+jpeg/doc/cstyle.ms.tbl.Z, or cstyle.txt.Z for those without nroff/tbl).
+
+Block comments should be laid out thusly:
+
+/*
+ *  Block comments in this style.
+ */
+
+We indent statements in K&R style, e.g.,
+	if (test) {
+	  then-part;
+	} else {
+	  else-part;
+	}
+with two spaces per indentation level.  (This indentation convention is
+handled automatically by GNU Emacs and many other text editors.)
+
+Multi-word names should be written in lower case with underscores, e.g.,
+multi_word_name (not multiWordName).  Preprocessor symbols and enum constants
+are similar but upper case (MULTI_WORD_NAME).  Names should be unique within
+the first fifteen characters.  (On some older systems, global names must be
+unique within six characters.  We accommodate this without cluttering the
+source code by using macros to substitute shorter names.)
+
+We use function prototypes everywhere; we rely on automatic source code
+transformation to feed prototype-less C compilers.  Transformation is done
+by the simple and portable tool 'ansi2knr.c' (courtesy of Ghostscript).
+ansi2knr is not very bright, so it imposes a format requirement on function
+declarations: the function name MUST BEGIN IN COLUMN 1.  Thus all functions
+should be written in the following style:
+
+LOCAL(int *)
+function_name (int a, char *b)
+{
+    code...
+}
+
+Note that each function definition must begin with GLOBAL(type), LOCAL(type),
+or METHODDEF(type).  These macros expand to "static type" or just "type" as
+appropriate.  They provide a readable indication of the routine's usage and
+can readily be changed for special needs.  (For instance, special linkage
+keywords can be inserted for use in Windows DLLs.)
+
+ansi2knr does not transform method declarations (function pointers in
+structs).  We handle these with a macro JMETHOD, defined as
+	#ifdef HAVE_PROTOTYPES
+	#define JMETHOD(type,methodname,arglist)  type (*methodname) arglist
+	#else
+	#define JMETHOD(type,methodname,arglist)  type (*methodname) ()
+	#endif
+which is used like this:
+	struct function_pointers {
+	  JMETHOD(void, init_entropy_encoder, (int somearg, jparms *jp));
+	  JMETHOD(void, term_entropy_encoder, (void));
+	};
+Note the set of parentheses surrounding the parameter list.
+
+A similar solution is used for forward and external function declarations
+(see the EXTERN and JPP macros).
+
+If the code is to work on non-ANSI compilers, we cannot rely on a prototype
+declaration to coerce actual parameters into the right types.  Therefore, use
+explicit casts on actual parameters whenever the actual parameter type is not
+identical to the formal parameter.  Beware of implicit conversions to "int".
+
+It seems there are some non-ANSI compilers in which the sizeof() operator
+is defined to return int, yet size_t is defined as long.  Needless to say,
+this is brain-damaged.  Always use the SIZEOF() macro in place of sizeof(),
+so that the result is guaranteed to be of type size_t.
+
+
+The JPEG library is intended to be used within larger programs.  Furthermore,
+we want it to be reentrant so that it can be used by applications that process
+multiple images concurrently.  The following rules support these requirements:
+
+1. Avoid direct use of file I/O, "malloc", error report printouts, etc;
+pass these through the common routines provided.
+
+2. Minimize global namespace pollution.  Functions should be declared static
+wherever possible.  (Note that our method-based calling conventions help this
+a lot: in many modules only the initialization function will ever need to be
+called directly, so only that function need be externally visible.)  All
+global function names should begin with "jpeg_", and should have an
+abbreviated name (unique in the first six characters) substituted by macro
+when NEED_SHORT_EXTERNAL_NAMES is set.
+
+3. Don't use global variables; anything that must be used in another module
+should be in the common data structures.
+
+4. Don't use static variables except for read-only constant tables.  Variables
+that should be private to a module can be placed into private structures (see
+the system architecture document, structure.txt).
+
+5. Source file names should begin with "j" for files that are part of the
+library proper; source files that are not part of the library, such as cjpeg.c
+and djpeg.c, do not begin with "j".  Keep source file names to eight
+characters (plus ".c" or ".h", etc) to make life easy for MS-DOSers.  Keep
+compression and decompression code in separate source files --- some
+applications may want only one half of the library.
+
+Note: these rules (particularly #4) are not followed religiously in the
+modules that are used in cjpeg/djpeg but are not part of the JPEG library
+proper.  Those modules are not really intended to be used in other
+applications.
diff --git a/configure.ac b/configure.ac
new file mode 100644
index 0000000..fb022f2
--- /dev/null
+++ b/configure.ac
@@ -0,0 +1,271 @@
+#                                               -*- Autoconf -*-
+# Process this file with autoconf to produce a configure script.
+
+AC_PREREQ([2.56])
+AC_INIT([libjpeg-turbo], [1.0.90])
+BUILD=`date +%Y%m%d`
+
+AM_INIT_AUTOMAKE([-Wall foreign dist-bzip2])
+AC_PREFIX_DEFAULT(/opt/libjpeg-turbo)
+
+# Always build with prototypes
+AC_DEFINE([HAVE_PROTOTYPES], 1, [Define if your compiler supports prototypes])
+
+# Checks for programs.
+SAVED_CFLAGS=${CFLAGS}
+SAVED_CXXFLAGS=${CXXFLAGS}
+AC_PROG_CPP
+AC_PROG_CC
+AC_PROG_CXX
+AC_PROG_INSTALL
+AC_PROG_LIBTOOL
+AC_PROG_LN_S
+
+# Check whether compiler supports pointers to undefined structures
+AC_MSG_CHECKING(whether compiler supports pointers to undefined structures)
+AC_TRY_COMPILE([ typedef struct undefined_structure * undef_struct_ptr; ], ,
+AC_MSG_RESULT(yes),
+[AC_MSG_RESULT(no)
+AC_DEFINE([INCOMPLETE_TYPES_BROKEN],[1],[Compiler does not support pointers to undefined structures.])])
+
+if test "x${GCC}" = "xyes"; then
+  if test "x${SAVED_CFLAGS}" = "x"; then
+    CFLAGS=-O3
+  fi
+  if test "x${SAVED_CXXFLAGS}" = "x"; then
+    CXXFLAGS=-O3
+  fi
+fi
+
+AC_CHECK_DECL([__SUNPRO_C], [SUNCC="yes"], [SUNCC="no"])
+if test "x${SUNCC}" = "xyes"; then
+  if test "x${SAVED_CFLAGS}" = "x"; then
+    CFLAGS=-xO5
+  fi
+  if test "x${SAVED_CXXFLAGS}" = "x"; then
+    CXXFLAGS=-xO5
+  fi
+fi
+
+# Checks for libraries.
+
+# Checks for header files.
+AC_HEADER_STDC
+AC_CHECK_HEADERS([stddef.h stdlib.h string.h])
+AC_CHECK_HEADER([sys/types.h], AC_DEFINE([NEED_SYS_TYPES_H], 1, [Define if you have sys/types.h]))
+
+# Checks for typedefs, structures, and compiler characteristics.
+AC_C_CONST
+AC_C_CHAR_UNSIGNED
+AC_C_INLINE
+AC_TYPE_SIZE_T
+AC_CHECK_TYPES([unsigned char, unsigned short])
+
+AC_MSG_CHECKING([if right shift is signed])
+AC_TRY_RUN(
+	[#include <stdio.h>
+	 int is_shifting_signed (long arg) {
+	 long res = arg >> 4;
+
+	 if (res == -0x7F7E80CL)
+		return 1; /* right shift is signed */
+
+	 /* see if unsigned-shift hack will fix it. */
+	 /* we can't just test exact value since it depends on width of long... */
+	 res |= (~0L) << (32-4);
+	 if (res == -0x7F7E80CL)
+		return 0; /* right shift is unsigned */
+
+	 printf("Right shift isn't acting as I expect it to.\n");
+	 printf("I fear the JPEG software will not work at all.\n\n");
+	 return 0; /* try it with unsigned anyway */
+	 }
+	 int main (void) {
+		exit(is_shifting_signed(-0x7F7E80B1L));
+	 }],
+	[AC_MSG_RESULT(no)
+	 AC_DEFINE([RIGHT_SHIFT_IS_UNSIGNED], 1, [Define if shift is unsigned])],
+	[AC_MSG_RESULT(yes)],
+	[AC_MSG_RESULT(Assuming that right shift is signed on target machine.)])
+
+# test whether global names are unique to at least 15 chars
+AC_MSG_CHECKING([for short external names])
+AC_TRY_LINK(
+	[int possibly_duplicate_function () { return 0; }
+	 int possibly_dupli_function () { return 1; }], [ ],
+	[AC_MSG_RESULT(ok)],
+	[AC_MSG_RESULT(short)
+	 AC_DEFINE([NEED_SHORT_EXTERNAL_NAMES], 1, [Define if you need short function names])])
+
+# Checks for library functions.
+AC_CHECK_FUNCS([memset memcpy], [],
+	[AC_DEFINE([NEED_BSD_STRINGS], 1,
+		   [Define if you have BSD-like bzero and bcopy])])
+
+# Set flags to indicate platform
+case "$host_os" in
+  cygwin* | mingw* | pw32* | interix*)
+    is_win32=1
+  ;;
+esac
+AM_CONDITIONAL([IS_WIN32], [test "x$is_win32" = "x1"])
+
+AC_MSG_CHECKING([libjpeg API version])
+AC_ARG_VAR(JPEG_LIB_VERSION, [libjpeg API version (62, 70, or 80)])
+if test "x$JPEG_LIB_VERSION" = "x"; then
+    AC_ARG_WITH([jpeg7],
+        AC_HELP_STRING([--with-jpeg7], [Emulate libjpeg v7 API/ABI (this makes libjpeg-turbo backward incompatible with libjpeg v6b.)]))
+    AC_ARG_WITH([jpeg8],
+        AC_HELP_STRING([--with-jpeg8], [Emulate libjpeg v8b API/ABI (this makes libjpeg-turbo backward incompatible with libjpeg v6b.)]))
+    if test "x${with_jpeg8}" = "xyes"; then
+        JPEG_LIB_VERSION=80
+    else
+        if test "x${with_jpeg7}" = "xyes"; then
+            JPEG_LIB_VERSION=70
+        else
+            JPEG_LIB_VERSION=62
+        fi
+    fi
+fi
+JPEG_LIB_VERSION_DECIMAL=`expr $JPEG_LIB_VERSION / 10`.`expr $JPEG_LIB_VERSION % 10`
+AC_SUBST(JPEG_LIB_VERSION_DECIMAL)
+AC_MSG_RESULT([$JPEG_LIB_VERSION_DECIMAL])
+AC_DEFINE_UNQUOTED(JPEG_LIB_VERSION, [$JPEG_LIB_VERSION], [libjpeg API version])
+
+AC_MSG_CHECKING([libjpeg shared library version])
+AC_ARG_VAR(SO_MAJOR_VERSION, [Major version of the libjpeg-turbo shared library (default is determined by the API version)])
+AC_ARG_VAR(SO_MINOR_VERSION, [Minor version of the libjpeg-turbo shared library (default is determined by the API version)])
+if test "x$SO_MAJOR_VERSION" = "x"; then
+    case "$JPEG_LIB_VERSION" in
+        62)  SO_MAJOR_VERSION=$JPEG_LIB_VERSION ;;
+        *)   SO_MAJOR_VERSION=`expr $JPEG_LIB_VERSION / 10` ;;
+    esac
+fi
+if test "x$SO_MINOR_VERSION" = "x"; then
+    case "$JPEG_LIB_VERSION" in
+        80)  SO_MINOR_VERSION=2 ;;
+        *)   SO_MINOR_VERSION=0 ;;
+    esac
+fi
+AC_MSG_RESULT([$SO_MAJOR_VERSION:$SO_MINOR_VERSION])
+AC_SUBST(SO_MAJOR_VERSION)
+AC_SUBST(SO_MINOR_VERSION)
+
+VERSION_SCRIPT=yes
+AC_ARG_ENABLE([ld-version-script],
+  AS_HELP_STRING([--disable-ld-version-script],
+    [Disable linker version script for libjpeg-turbo (default is to use linker version script if the linker supports it)]),
+    [VERSION_SCRIPT=$enableval], [])
+
+AC_MSG_CHECKING([whether the linker supports version scripts])
+SAVED_LDFLAGS="$LDFLAGS"
+LDFLAGS="$LDFLAGS -Wl,--version-script,conftest.map"
+cat > conftest.map <<EOF
+VERS_1 {
+  global: *;
+};
+EOF
+AC_LINK_IFELSE(AC_LANG_PROGRAM([], []),
+  [VERSION_SCRIPT_FLAG=-Wl,--version-script,; AC_MSG_RESULT([yes (GNU style)])], [])
+if test "x$VERSION_SCRIPT_FLAG" = "x"; then
+  LDFLAGS="$SAVED_LDFLAGS -Wl,-M,conftest.map"
+  AC_LINK_IFELSE(AC_LANG_PROGRAM([], []),
+    [VERSION_SCRIPT_FLAG=-Wl,-M,; AC_MSG_RESULT([yes (Sun style)])], [])
+fi
+if test "x$VERSION_SCRIPT_FLAG" = "x"; then
+  VERSION_SCRIPT=no
+  AC_MSG_RESULT(no)
+fi
+
+ANON_VERSION_SCRIPT=yes
+AC_MSG_CHECKING([whether the linker supports anonymous version scripts])
+LDFLAGS="$SAVED_LDFLAGS -Wl,--version-script,conftest.map"
+cat > conftest.map <<EOF
+{
+  global: main;
+  local: *;
+};
+EOF
+AC_LINK_IFELSE(AC_LANG_PROGRAM([], []),
+  [ANON_VERSION_SCRIPT_FLAG=-Wl,--version-script,; AC_MSG_RESULT([yes (GNU style)])], [])
+if test "x$ANON_VERSION_SCRIPT_FLAG" = "x"; then
+  LDFLAGS="$SAVED_LDFLAGS -Wl,-M,conftest.map"
+  AC_LINK_IFELSE(AC_LANG_PROGRAM([], []),
+    [ANON_VERSION_SCRIPT_FLAG=-Wl,-M,; AC_MSG_RESULT([yes (Sun style)])], [])
+fi
+if test "x$ANON_VERSION_SCRIPT_FLAG" = "x"; then
+  ANON_VERSION_SCRIPT=no
+  AC_MSG_RESULT(no)
+fi
+LDFLAGS="$SAVED_LDFLAGS"
+
+AC_MSG_CHECKING([whether to use version script when building libjpeg])
+AC_MSG_RESULT($VERSION_SCRIPT)
+
+AM_CONDITIONAL(VERSION_SCRIPT, test "x$VERSION_SCRIPT" = "xyes")
+AM_CONDITIONAL(ANON_VERSION_SCRIPT, test "x$ANON_VERSION_SCRIPT" = "xyes")
+AC_SUBST(VERSION_SCRIPT_FLAG)
+AC_SUBST(ANON_VERSION_SCRIPT_FLAG)
+
+# SIMD is optional
+AC_ARG_WITH([simd],
+    AC_HELP_STRING([--without-simd],[Omit accelerated SIMD routines.]))
+if test "x${with_simd}" != "xno"; then
+  # Check if we're on a supported CPU
+  AC_MSG_CHECKING([if we have SIMD optimisations for cpu type])
+  case "$host_cpu" in
+    x86_64 | amd64)
+      AC_MSG_RESULT([yes (x86_64)])
+      AC_PROG_NASM
+      simd_arch=x86_64
+    ;;
+    i*86 | x86 | ia32)
+      AC_MSG_RESULT([yes (i386)])
+      AC_PROG_NASM
+      simd_arch=i386
+    ;;
+    *)
+      AC_MSG_RESULT([no ("$host_cpu")])
+      AC_MSG_WARN([SIMD support not available for this CPU.  Performance will suffer.])
+      with_simd=no;
+    ;;
+  esac
+
+  if test "x${with_simd}" != "xno"; then
+    AC_DEFINE([WITH_SIMD], [1], [Use accelerated SIMD routines.])
+  fi
+fi
+
+AM_CONDITIONAL([WITH_SIMD], [test "x$with_simd" != "xno"])
+AM_CONDITIONAL([SIMD_I386], [test "x$simd_arch" = "xi386"])
+AM_CONDITIONAL([SIMD_X86_64], [test "x$simd_arch" = "xx86_64"])
+AM_CONDITIONAL([X86_64], [test "x$host_cpu" = "xx86_64" -o "x$host_cpu" = "xamd64"])
+
+case "$host_cpu" in
+  x86_64)
+    RPMARCH=x86_64
+    DEBARCH=amd64
+    ;;
+  i*86 | x86 | ia32)
+    RPMARCH=i386
+    DEBARCH=i386
+    ;;
+esac
+
+AC_SUBST(RPMARCH)
+AC_SUBST(DEBARCH)
+AC_SUBST(BUILD)
+AC_DEFINE_UNQUOTED([BUILD], "$BUILD", [Build number])
+
+# jconfig.h is the file we use, but we have another before that to
+# fool autoheader. the reason is that we include this header in our
+# API headers, which can screw things up for users of the lib.
+# jconfig.h is a minimal version that allows this package to be built
+AC_CONFIG_HEADERS([config.h])
+AC_CONFIG_HEADERS([jconfig.h])
+AC_CONFIG_FILES([libjpeg-turbo.spec:release/libjpeg-turbo.spec.in])
+AC_CONFIG_FILES([makemacpkg:release/makemacpkg.in])
+AC_CONFIG_FILES([makesunpkg:release/makesunpkg.in])
+AC_CONFIG_FILES([libjpeg.map])
+AC_CONFIG_FILES([Makefile simd/Makefile])
+AC_OUTPUT
diff --git a/djpeg.1 b/djpeg.1
new file mode 100644
index 0000000..8d98c66
--- /dev/null
+++ b/djpeg.1
@@ -0,0 +1,251 @@
+.TH DJPEG 1 "11 October 2010"
+.SH NAME
+djpeg \- decompress a JPEG file to an image file
+.SH SYNOPSIS
+.B djpeg
+[
+.I options
+]
+[
+.I filename
+]
+.LP
+.SH DESCRIPTION
+.LP
+.B djpeg
+decompresses the named JPEG file, or the standard input if no file is named,
+and produces an image file on the standard output.  PBMPLUS (PPM/PGM), BMP,
+GIF, Targa, or RLE (Utah Raster Toolkit) output format can be selected.
+(RLE is supported only if the URT library is available.)
+.SH OPTIONS
+All switch names may be abbreviated; for example,
+.B \-grayscale
+may be written
+.B \-gray
+or
+.BR \-gr .
+Most of the "basic" switches can be abbreviated to as little as one letter.
+Upper and lower case are equivalent (thus
+.B \-BMP
+is the same as
+.BR \-bmp ).
+British spellings are also accepted (e.g.,
+.BR \-greyscale ),
+though for brevity these are not mentioned below.
+.PP
+The basic switches are:
+.TP
+.BI \-colors " N"
+Reduce image to at most N colors.  This reduces the number of colors used in
+the output image, so that it can be displayed on a colormapped display or
+stored in a colormapped file format.  For example, if you have an 8-bit
+display, you'd need to reduce to 256 or fewer colors.
+.TP
+.BI \-quantize " N"
+Same as
+.BR \-colors .
+.B \-colors
+is the recommended name,
+.B \-quantize
+is provided only for backwards compatibility.
+.TP
+.B \-fast
+Select recommended processing options for fast, low quality output.  (The
+default options are chosen for highest quality output.)  Currently, this is
+equivalent to \fB\-dct fast \-nosmooth \-onepass \-dither ordered\fR.
+.TP
+.B \-grayscale
+Force gray-scale output even if JPEG file is color.  Useful for viewing on
+monochrome displays; also,
+.B djpeg
+runs noticeably faster in this mode.
+.TP
+.BI \-scale " M/N"
+Scale the output image by a factor M/N.  Currently the scale factor must be
+1/1, 1/2, 1/4, or 1/8.  Scaling is handy if the image is larger than your
+screen; also,
+.B djpeg
+runs much faster when scaling down the output.
+.TP
+.B \-bmp
+Select BMP output format (Windows flavor).  8-bit colormapped format is
+emitted if
+.B \-colors
+or
+.B \-grayscale
+is specified, or if the JPEG file is gray-scale; otherwise, 24-bit full-color
+format is emitted.
+.TP
+.B \-gif
+Select GIF output format.  Since GIF does not support more than 256 colors,
+.B \-colors 256
+is assumed (unless you specify a smaller number of colors).
+.TP
+.B \-os2
+Select BMP output format (OS/2 1.x flavor).  8-bit colormapped format is
+emitted if
+.B \-colors
+or
+.B \-grayscale
+is specified, or if the JPEG file is gray-scale; otherwise, 24-bit full-color
+format is emitted.
+.TP
+.B \-pnm
+Select PBMPLUS (PPM/PGM) output format (this is the default format).
+PGM is emitted if the JPEG file is gray-scale or if
+.B \-grayscale
+is specified; otherwise PPM is emitted.
+.TP
+.B \-rle
+Select RLE output format.  (Requires URT library.)
+.TP
+.B \-targa
+Select Targa output format.  Gray-scale format is emitted if the JPEG file is
+gray-scale or if
+.B \-grayscale
+is specified; otherwise, colormapped format is emitted if
+.B \-colors
+is specified; otherwise, 24-bit full-color format is emitted.
+.PP
+Switches for advanced users:
+.TP
+.B \-dct int
+Use integer DCT method (default).
+.TP
+.B \-dct fast
+Use fast integer DCT (less accurate).
+.TP
+.B \-dct float
+Use floating-point DCT method.
+The float method is very slightly more accurate than the int method, but is
+much slower unless your machine has very fast floating-point hardware.  Also
+note that results of the floating-point method may vary slightly across
+machines, while the integer methods should give the same results everywhere.
+The fast integer method is much less accurate than the other two.
+.TP
+.B \-dither fs
+Use Floyd-Steinberg dithering in color quantization.
+.TP
+.B \-dither ordered
+Use ordered dithering in color quantization.
+.TP
+.B \-dither none
+Do not use dithering in color quantization.
+By default, Floyd-Steinberg dithering is applied when quantizing colors; this
+is slow but usually produces the best results.  Ordered dither is a compromise
+between speed and quality; no dithering is fast but usually looks awful.  Note
+that these switches have no effect unless color quantization is being done.
+Ordered dither is only available in
+.B \-onepass
+mode.
+.TP
+.BI \-map " file"
+Quantize to the colors used in the specified image file.  This is useful for
+producing multiple files with identical color maps, or for forcing a
+predefined set of colors to be used.  The
+.I file
+must be a GIF or PPM file. This option overrides
+.B \-colors
+and
+.BR \-onepass .
+.TP
+.B \-nosmooth
+Use a faster, lower-quality upsampling routine.
+.TP
+.B \-onepass
+Use one-pass instead of two-pass color quantization.  The one-pass method is
+faster and needs less memory, but it produces a lower-quality image.
+.B \-onepass
+is ignored unless you also say
+.B \-colors
+.IR N .
+Also, the one-pass method is always used for gray-scale output (the two-pass
+method is no improvement then).
+.TP
+.BI \-maxmemory " N"
+Set limit for amount of memory to use in processing large images.  Value is
+in thousands of bytes, or millions of bytes if "M" is attached to the
+number.  For example,
+.B \-max 4m
+selects 4000000 bytes.  If more space is needed, temporary files will be used.
+.TP
+.BI \-outfile " name"
+Send output image to the named file, not to standard output.
+.TP
+.B \-verbose
+Enable debug printout.  More
+.BR \-v 's
+give more output.  Also, version information is printed at startup.
+.TP
+.B \-debug
+Same as
+.BR \-verbose .
+.SH EXAMPLES
+.LP
+This example decompresses the JPEG file foo.jpg, quantizes it to
+256 colors, and saves the output in 8-bit BMP format in foo.bmp:
+.IP
+.B djpeg \-colors 256 \-bmp
+.I foo.jpg
+.B >
+.I foo.bmp
+.SH HINTS
+To get a quick preview of an image, use the
+.B \-grayscale
+and/or
+.B \-scale
+switches.
+.B \-grayscale \-scale 1/8
+is the fastest case.
+.PP
+Several options are available that trade off image quality to gain speed.
+.B \-fast
+turns on the recommended settings.
+.PP
+.B \-dct fast
+and/or
+.B \-nosmooth
+gain speed at a small sacrifice in quality.
+When producing a color-quantized image,
+.B \-onepass \-dither ordered
+is fast but much lower quality than the default behavior.
+.B \-dither none
+may give acceptable results in two-pass mode, but is seldom tolerable in
+one-pass mode.
+.PP
+If you are fortunate enough to have very fast floating point hardware,
+\fB\-dct float\fR may be even faster than \fB\-dct fast\fR.  But on most
+machines \fB\-dct float\fR is slower than \fB\-dct int\fR; in this case it is
+not worth using, because its theoretical accuracy advantage is too small to be
+significant in practice.
+.SH ENVIRONMENT
+.TP
+.B JPEGMEM
+If this environment variable is set, its value is the default memory limit.
+The value is specified as described for the
+.B \-maxmemory
+switch.
+.B JPEGMEM
+overrides the default value specified when the program was compiled, and
+itself is overridden by an explicit
+.BR \-maxmemory .
+.SH SEE ALSO
+.BR cjpeg (1),
+.BR jpegtran (1),
+.BR rdjpgcom (1),
+.BR wrjpgcom (1)
+.br
+.BR ppm (5),
+.BR pgm (5)
+.br
+Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+Communications of the ACM, April 1991 (vol. 34, no. 4), pp. 30-44.
+.SH AUTHOR
+Independent JPEG Group
+.SH BUGS
+Arithmetic coding is not supported for legal reasons.
+.PP
+To avoid the Unisys LZW patent,
+.B djpeg
+produces uncompressed GIF files.  These are larger than they should be, but
+are readable by standard GIF decoders.
diff --git a/djpeg.c b/djpeg.c
new file mode 100644
index 0000000..466973f
--- /dev/null
+++ b/djpeg.c
@@ -0,0 +1,621 @@
+/*
+ * djpeg.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a command-line user interface for the JPEG decompressor.
+ * It should work on any system with Unix- or MS-DOS-style command lines.
+ *
+ * Two different command line styles are permitted, depending on the
+ * compile-time switch TWO_FILE_COMMANDLINE:
+ *	djpeg [options]  inputfile outputfile
+ *	djpeg [options]  [inputfile]
+ * In the second style, output is always to standard output, which you'd
+ * normally redirect to a file or pipe to some other program.  Input is
+ * either from a named file or from standard input (typically redirected).
+ * The second style is convenient on Unix but is unhelpful on systems that
+ * don't support pipes.  Also, you MUST use the first style if your system
+ * doesn't do binary I/O to stdin/stdout.
+ * To simplify script writing, the "-outfile" switch is provided.  The syntax
+ *	djpeg [options]  -outfile outputfile  inputfile
+ * works regardless of which command line style is used.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+#include "jversion.h"		/* for version message */
+#include "config.h"
+
+#include <ctype.h>		/* to declare isprint() */
+
+#ifdef USE_CCOMMAND		/* command-line reader for Macintosh */
+#ifdef __MWERKS__
+#include <SIOUX.h>              /* Metrowerks needs this */
+#include <console.h>		/* ... and this */
+#endif
+#ifdef THINK_C
+#include <console.h>		/* Think declares it here */
+#endif
+#endif
+
+
+/* Create the add-on message string table. */
+
+#define JMESSAGE(code,string)	string ,
+
+static const char * const cdjpeg_message_table[] = {
+#include "cderror.h"
+  NULL
+};
+
+
+/*
+ * This list defines the known output image formats
+ * (not all of which need be supported by a given version).
+ * You can change the default output format by defining DEFAULT_FMT;
+ * indeed, you had better do so if you undefine PPM_SUPPORTED.
+ */
+
+typedef enum {
+	FMT_BMP,		/* BMP format (Windows flavor) */
+	FMT_GIF,		/* GIF format */
+	FMT_OS2,		/* BMP format (OS/2 flavor) */
+	FMT_PPM,		/* PPM/PGM (PBMPLUS formats) */
+	FMT_RLE,		/* RLE format */
+	FMT_TARGA,		/* Targa format */
+	FMT_TIFF		/* TIFF format */
+} IMAGE_FORMATS;
+
+#ifndef DEFAULT_FMT		/* so can override from CFLAGS in Makefile */
+#define DEFAULT_FMT	FMT_PPM
+#endif
+
+static IMAGE_FORMATS requested_fmt;
+
+
+/*
+ * Argument-parsing code.
+ * The switch parser is designed to be useful with DOS-style command line
+ * syntax, ie, intermixed switches and file names, where only the switches
+ * to the left of a given file name affect processing of that file.
+ * The main program in this file doesn't actually use this capability...
+ */
+
+
+static const char * progname;	/* program name for error messages */
+static char * outfilename;	/* for -outfile switch */
+
+
+LOCAL(void)
+usage (void)
+/* complain about bad command line */
+{
+  fprintf(stderr, "usage: %s [switches] ", progname);
+#ifdef TWO_FILE_COMMANDLINE
+  fprintf(stderr, "inputfile outputfile\n");
+#else
+  fprintf(stderr, "[inputfile]\n");
+#endif
+
+  fprintf(stderr, "Switches (names may be abbreviated):\n");
+  fprintf(stderr, "  -colors N      Reduce image to no more than N colors\n");
+  fprintf(stderr, "  -fast          Fast, low-quality processing\n");
+  fprintf(stderr, "  -grayscale     Force grayscale output\n");
+#ifdef IDCT_SCALING_SUPPORTED
+  fprintf(stderr, "  -scale M/N     Scale output image by fraction M/N, eg, 1/8\n");
+#endif
+#ifdef BMP_SUPPORTED
+  fprintf(stderr, "  -bmp           Select BMP output format (Windows style)%s\n",
+	  (DEFAULT_FMT == FMT_BMP ? " (default)" : ""));
+#endif
+#ifdef GIF_SUPPORTED
+  fprintf(stderr, "  -gif           Select GIF output format%s\n",
+	  (DEFAULT_FMT == FMT_GIF ? " (default)" : ""));
+#endif
+#ifdef BMP_SUPPORTED
+  fprintf(stderr, "  -os2           Select BMP output format (OS/2 style)%s\n",
+	  (DEFAULT_FMT == FMT_OS2 ? " (default)" : ""));
+#endif
+#ifdef PPM_SUPPORTED
+  fprintf(stderr, "  -pnm           Select PBMPLUS (PPM/PGM) output format%s\n",
+	  (DEFAULT_FMT == FMT_PPM ? " (default)" : ""));
+#endif
+#ifdef RLE_SUPPORTED
+  fprintf(stderr, "  -rle           Select Utah RLE output format%s\n",
+	  (DEFAULT_FMT == FMT_RLE ? " (default)" : ""));
+#endif
+#ifdef TARGA_SUPPORTED
+  fprintf(stderr, "  -targa         Select Targa output format%s\n",
+	  (DEFAULT_FMT == FMT_TARGA ? " (default)" : ""));
+#endif
+  fprintf(stderr, "Switches for advanced users:\n");
+#ifdef DCT_ISLOW_SUPPORTED
+  fprintf(stderr, "  -dct int       Use integer DCT method%s\n",
+	  (JDCT_DEFAULT == JDCT_ISLOW ? " (default)" : ""));
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  fprintf(stderr, "  -dct fast      Use fast integer DCT (less accurate)%s\n",
+	  (JDCT_DEFAULT == JDCT_IFAST ? " (default)" : ""));
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  fprintf(stderr, "  -dct float     Use floating-point DCT method%s\n",
+	  (JDCT_DEFAULT == JDCT_FLOAT ? " (default)" : ""));
+#endif
+  fprintf(stderr, "  -dither fs     Use F-S dithering (default)\n");
+  fprintf(stderr, "  -dither none   Don't use dithering in quantization\n");
+  fprintf(stderr, "  -dither ordered  Use ordered dither (medium speed, quality)\n");
+#ifdef QUANT_2PASS_SUPPORTED
+  fprintf(stderr, "  -map FILE      Map to colors used in named image file\n");
+#endif
+  fprintf(stderr, "  -nosmooth      Don't use high-quality upsampling\n");
+#ifdef QUANT_1PASS_SUPPORTED
+  fprintf(stderr, "  -onepass       Use 1-pass quantization (fast, low quality)\n");
+#endif
+  fprintf(stderr, "  -maxmemory N   Maximum memory to use (in kbytes)\n");
+  fprintf(stderr, "  -outfile name  Specify name for output file\n");
+  fprintf(stderr, "  -verbose  or  -debug   Emit debug output\n");
+  exit(EXIT_FAILURE);
+}
+
+
+LOCAL(int)
+parse_switches (j_decompress_ptr cinfo, int argc, char **argv,
+		int last_file_arg_seen, boolean for_real)
+/* Parse optional switches.
+ * Returns argv[] index of first file-name argument (== argc if none).
+ * Any file names with indexes <= last_file_arg_seen are ignored;
+ * they have presumably been processed in a previous iteration.
+ * (Pass 0 for last_file_arg_seen on the first or only iteration.)
+ * for_real is FALSE on the first (dummy) pass; we may skip any expensive
+ * processing.
+ */
+{
+  int argn;
+  char * arg;
+
+  /* Set up default JPEG parameters. */
+  requested_fmt = DEFAULT_FMT;	/* set default output file format */
+  outfilename = NULL;
+  cinfo->err->trace_level = 0;
+
+  /* Scan command line options, adjust parameters */
+
+  for (argn = 1; argn < argc; argn++) {
+    arg = argv[argn];
+    if (*arg != '-') {
+      /* Not a switch, must be a file name argument */
+      if (argn <= last_file_arg_seen) {
+	outfilename = NULL;	/* -outfile applies to just one input file */
+	continue;		/* ignore this name if previously processed */
+      }
+      break;			/* else done parsing switches */
+    }
+    arg++;			/* advance past switch marker character */
+
+    if (keymatch(arg, "bmp", 1)) {
+      /* BMP output format. */
+      requested_fmt = FMT_BMP;
+
+    } else if (keymatch(arg, "colors", 1) || keymatch(arg, "colours", 1) ||
+	       keymatch(arg, "quantize", 1) || keymatch(arg, "quantise", 1)) {
+      /* Do color quantization. */
+      int val;
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%d", &val) != 1)
+	usage();
+      cinfo->desired_number_of_colors = val;
+      cinfo->quantize_colors = TRUE;
+
+    } else if (keymatch(arg, "dct", 2)) {
+      /* Select IDCT algorithm. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (keymatch(argv[argn], "int", 1)) {
+	cinfo->dct_method = JDCT_ISLOW;
+      } else if (keymatch(argv[argn], "fast", 2)) {
+	cinfo->dct_method = JDCT_IFAST;
+      } else if (keymatch(argv[argn], "float", 2)) {
+	cinfo->dct_method = JDCT_FLOAT;
+      } else
+	usage();
+
+    } else if (keymatch(arg, "dither", 2)) {
+      /* Select dithering algorithm. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (keymatch(argv[argn], "fs", 2)) {
+	cinfo->dither_mode = JDITHER_FS;
+      } else if (keymatch(argv[argn], "none", 2)) {
+	cinfo->dither_mode = JDITHER_NONE;
+      } else if (keymatch(argv[argn], "ordered", 2)) {
+	cinfo->dither_mode = JDITHER_ORDERED;
+      } else
+	usage();
+
+    } else if (keymatch(arg, "debug", 1) || keymatch(arg, "verbose", 1)) {
+      /* Enable debug printouts. */
+      /* On first -d, print version identification */
+      static boolean printed_version = FALSE;
+
+      if (! printed_version) {
+	fprintf(stderr, "%s version %s (build %s)\n",
+		PACKAGE_NAME, VERSION, BUILD);
+	fprintf(stderr, "%s\n\n", LJTCOPYRIGHT);
+	fprintf(stderr, "Based on Independent JPEG Group's libjpeg, version %s\n%s\n\n",
+		JVERSION, JCOPYRIGHT);
+	printed_version = TRUE;
+      }
+      cinfo->err->trace_level++;
+
+    } else if (keymatch(arg, "fast", 1)) {
+      /* Select recommended processing options for quick-and-dirty output. */
+      cinfo->two_pass_quantize = FALSE;
+      cinfo->dither_mode = JDITHER_ORDERED;
+      if (! cinfo->quantize_colors) /* don't override an earlier -colors */
+	cinfo->desired_number_of_colors = 216;
+      cinfo->dct_method = JDCT_FASTEST;
+      cinfo->do_fancy_upsampling = FALSE;
+
+    } else if (keymatch(arg, "gif", 1)) {
+      /* GIF output format. */
+      requested_fmt = FMT_GIF;
+
+    } else if (keymatch(arg, "grayscale", 2) || keymatch(arg, "greyscale",2)) {
+      /* Force monochrome output. */
+      cinfo->out_color_space = JCS_GRAYSCALE;
+
+    } else if (keymatch(arg, "map", 3)) {
+      /* Quantize to a color map taken from an input file. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (for_real) {		/* too expensive to do twice! */
+#ifdef QUANT_2PASS_SUPPORTED	/* otherwise can't quantize to supplied map */
+	FILE * mapfile;
+
+	if ((mapfile = fopen(argv[argn], READ_BINARY)) == NULL) {
+	  fprintf(stderr, "%s: can't open %s\n", progname, argv[argn]);
+	  exit(EXIT_FAILURE);
+	}
+	read_color_map(cinfo, mapfile);
+	fclose(mapfile);
+	cinfo->quantize_colors = TRUE;
+#else
+	ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+      }
+
+    } else if (keymatch(arg, "maxmemory", 3)) {
+      /* Maximum memory in Kb (or Mb with 'm'). */
+      long lval;
+      char ch = 'x';
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%ld%c", &lval, &ch) < 1)
+	usage();
+      if (ch == 'm' || ch == 'M')
+	lval *= 1000L;
+      cinfo->mem->max_memory_to_use = lval * 1000L;
+
+    } else if (keymatch(arg, "nosmooth", 3)) {
+      /* Suppress fancy upsampling */
+      cinfo->do_fancy_upsampling = FALSE;
+
+    } else if (keymatch(arg, "onepass", 3)) {
+      /* Use fast one-pass quantization. */
+      cinfo->two_pass_quantize = FALSE;
+
+    } else if (keymatch(arg, "os2", 3)) {
+      /* BMP output format (OS/2 flavor). */
+      requested_fmt = FMT_OS2;
+
+    } else if (keymatch(arg, "outfile", 4)) {
+      /* Set output file name. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      outfilename = argv[argn];	/* save it away for later use */
+
+    } else if (keymatch(arg, "pnm", 1) || keymatch(arg, "ppm", 1)) {
+      /* PPM/PGM output format. */
+      requested_fmt = FMT_PPM;
+
+    } else if (keymatch(arg, "rle", 1)) {
+      /* RLE output format. */
+      requested_fmt = FMT_RLE;
+
+    } else if (keymatch(arg, "scale", 1)) {
+      /* Scale the output image by a fraction M/N. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%d/%d",
+		 &cinfo->scale_num, &cinfo->scale_denom) != 2)
+	usage();
+
+    } else if (keymatch(arg, "targa", 1)) {
+      /* Targa output format. */
+      requested_fmt = FMT_TARGA;
+
+    } else {
+      usage();			/* bogus switch */
+    }
+  }
+
+  return argn;			/* return index of next arg (file name) */
+}
+
+
+/*
+ * Marker processor for COM and interesting APPn markers.
+ * This replaces the library's built-in processor, which just skips the marker.
+ * We want to print out the marker as text, to the extent possible.
+ * Note this code relies on a non-suspending data source.
+ */
+
+LOCAL(unsigned int)
+jpeg_getc (j_decompress_ptr cinfo)
+/* Read next byte */
+{
+  struct jpeg_source_mgr * datasrc = cinfo->src;
+
+  if (datasrc->bytes_in_buffer == 0) {
+    if (! (*datasrc->fill_input_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  }
+  datasrc->bytes_in_buffer--;
+  return GETJOCTET(*datasrc->next_input_byte++);
+}
+
+
+METHODDEF(boolean)
+print_text_marker (j_decompress_ptr cinfo)
+{
+  boolean traceit = (cinfo->err->trace_level >= 1);
+  INT32 length;
+  unsigned int ch;
+  unsigned int lastch = 0;
+
+  length = jpeg_getc(cinfo) << 8;
+  length += jpeg_getc(cinfo);
+  length -= 2;			/* discount the length word itself */
+
+  if (traceit) {
+    if (cinfo->unread_marker == JPEG_COM)
+      fprintf(stderr, "Comment, length %ld:\n", (long) length);
+    else			/* assume it is an APPn otherwise */
+      fprintf(stderr, "APP%d, length %ld:\n",
+	      cinfo->unread_marker - JPEG_APP0, (long) length);
+  }
+
+  while (--length >= 0) {
+    ch = jpeg_getc(cinfo);
+    if (traceit) {
+      /* Emit the character in a readable form.
+       * Nonprintables are converted to \nnn form,
+       * while \ is converted to \\.
+       * Newlines in CR, CR/LF, or LF form will be printed as one newline.
+       */
+      if (ch == '\r') {
+	fprintf(stderr, "\n");
+      } else if (ch == '\n') {
+	if (lastch != '\r')
+	  fprintf(stderr, "\n");
+      } else if (ch == '\\') {
+	fprintf(stderr, "\\\\");
+      } else if (isprint(ch)) {
+	putc(ch, stderr);
+      } else {
+	fprintf(stderr, "\\%03o", ch);
+      }
+      lastch = ch;
+    }
+  }
+
+  if (traceit)
+    fprintf(stderr, "\n");
+
+  return TRUE;
+}
+
+
+/*
+ * The main program.
+ */
+
+int
+main (int argc, char **argv)
+{
+  struct jpeg_decompress_struct cinfo;
+  struct jpeg_error_mgr jerr;
+#ifdef PROGRESS_REPORT
+  struct cdjpeg_progress_mgr progress;
+#endif
+  int file_index;
+  djpeg_dest_ptr dest_mgr = NULL;
+  FILE * input_file;
+  FILE * output_file;
+  JDIMENSION num_scanlines;
+
+  /* On Mac, fetch a command line. */
+#ifdef USE_CCOMMAND
+  argc = ccommand(&argv);
+#endif
+
+  progname = argv[0];
+  if (progname == NULL || progname[0] == 0)
+    progname = "djpeg";		/* in case C library doesn't provide it */
+
+  /* Initialize the JPEG decompression object with default error handling. */
+  cinfo.err = jpeg_std_error(&jerr);
+  jpeg_create_decompress(&cinfo);
+  /* Add some application-specific error messages (from cderror.h) */
+  jerr.addon_message_table = cdjpeg_message_table;
+  jerr.first_addon_message = JMSG_FIRSTADDONCODE;
+  jerr.last_addon_message = JMSG_LASTADDONCODE;
+
+  /* Insert custom marker processor for COM and APP12.
+   * APP12 is used by some digital camera makers for textual info,
+   * so we provide the ability to display it as text.
+   * If you like, additional APPn marker types can be selected for display,
+   * but don't try to override APP0 or APP14 this way (see libjpeg.txt).
+   */
+  jpeg_set_marker_processor(&cinfo, JPEG_COM, print_text_marker);
+  jpeg_set_marker_processor(&cinfo, JPEG_APP0+12, print_text_marker);
+
+  /* Now safe to enable signal catcher. */
+#ifdef NEED_SIGNAL_CATCHER
+  enable_signal_catcher((j_common_ptr) &cinfo);
+#endif
+
+  /* Scan command line to find file names. */
+  /* It is convenient to use just one switch-parsing routine, but the switch
+   * values read here are ignored; we will rescan the switches after opening
+   * the input file.
+   * (Exception: tracing level set here controls verbosity for COM markers
+   * found during jpeg_read_header...)
+   */
+
+  file_index = parse_switches(&cinfo, argc, argv, 0, FALSE);
+
+#ifdef TWO_FILE_COMMANDLINE
+  /* Must have either -outfile switch or explicit output file name */
+  if (outfilename == NULL) {
+    if (file_index != argc-2) {
+      fprintf(stderr, "%s: must name one input and one output file\n",
+	      progname);
+      usage();
+    }
+    outfilename = argv[file_index+1];
+  } else {
+    if (file_index != argc-1) {
+      fprintf(stderr, "%s: must name one input and one output file\n",
+	      progname);
+      usage();
+    }
+  }
+#else
+  /* Unix style: expect zero or one file name */
+  if (file_index < argc-1) {
+    fprintf(stderr, "%s: only one input file\n", progname);
+    usage();
+  }
+#endif /* TWO_FILE_COMMANDLINE */
+
+  /* Open the input file. */
+  if (file_index < argc) {
+    if ((input_file = fopen(argv[file_index], READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s\n", progname, argv[file_index]);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default input file is stdin */
+    input_file = read_stdin();
+  }
+
+  /* Open the output file. */
+  if (outfilename != NULL) {
+    if ((output_file = fopen(outfilename, WRITE_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s\n", progname, outfilename);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default output file is stdout */
+    output_file = write_stdout();
+  }
+
+#ifdef PROGRESS_REPORT
+  start_progress_monitor((j_common_ptr) &cinfo, &progress);
+#endif
+
+  /* Specify data source for decompression */
+  jpeg_stdio_src(&cinfo, input_file);
+
+  /* Read file header, set default decompression parameters */
+  (void) jpeg_read_header(&cinfo, TRUE);
+
+  /* Adjust default decompression parameters by re-parsing the options */
+  file_index = parse_switches(&cinfo, argc, argv, 0, TRUE);
+
+  /* Initialize the output module now to let it override any crucial
+   * option settings (for instance, GIF wants to force color quantization).
+   */
+  switch (requested_fmt) {
+#ifdef BMP_SUPPORTED
+  case FMT_BMP:
+    dest_mgr = jinit_write_bmp(&cinfo, FALSE);
+    break;
+  case FMT_OS2:
+    dest_mgr = jinit_write_bmp(&cinfo, TRUE);
+    break;
+#endif
+#ifdef GIF_SUPPORTED
+  case FMT_GIF:
+    dest_mgr = jinit_write_gif(&cinfo);
+    break;
+#endif
+#ifdef PPM_SUPPORTED
+  case FMT_PPM:
+    dest_mgr = jinit_write_ppm(&cinfo);
+    break;
+#endif
+#ifdef RLE_SUPPORTED
+  case FMT_RLE:
+    dest_mgr = jinit_write_rle(&cinfo);
+    break;
+#endif
+#ifdef TARGA_SUPPORTED
+  case FMT_TARGA:
+    dest_mgr = jinit_write_targa(&cinfo);
+    break;
+#endif
+  default:
+    ERREXIT(&cinfo, JERR_UNSUPPORTED_FORMAT);
+    break;
+  }
+  dest_mgr->output_file = output_file;
+
+  /* Start decompressor */
+  (void) jpeg_start_decompress(&cinfo);
+
+  /* Write output file header */
+  (*dest_mgr->start_output) (&cinfo, dest_mgr);
+
+  /* Process data */
+  while (cinfo.output_scanline < cinfo.output_height) {
+    num_scanlines = jpeg_read_scanlines(&cinfo, dest_mgr->buffer,
+					dest_mgr->buffer_height);
+    (*dest_mgr->put_pixel_rows) (&cinfo, dest_mgr, num_scanlines);
+  }
+
+#ifdef PROGRESS_REPORT
+  /* Hack: count final pass as done in case finish_output does an extra pass.
+   * The library won't have updated completed_passes.
+   */
+  progress.pub.completed_passes = progress.pub.total_passes;
+#endif
+
+  /* Finish decompression and release memory.
+   * I must do it in this order because output module has allocated memory
+   * of lifespan JPOOL_IMAGE; it needs to finish before releasing memory.
+   */
+  (*dest_mgr->finish_output) (&cinfo, dest_mgr);
+  (void) jpeg_finish_decompress(&cinfo);
+  jpeg_destroy_decompress(&cinfo);
+
+  /* Close files, if we opened them */
+  if (input_file != stdin)
+    fclose(input_file);
+  if (output_file != stdout)
+    fclose(output_file);
+
+#ifdef PROGRESS_REPORT
+  end_progress_monitor((j_common_ptr) &cinfo);
+#endif
+
+  /* All done. */
+  exit(jerr.num_warnings ? EXIT_WARNING : EXIT_SUCCESS);
+  return 0;			/* suppress no-return-value warnings */
+}
diff --git a/example.c b/example.c
new file mode 100644
index 0000000..1d6f6cc
--- /dev/null
+++ b/example.c
@@ -0,0 +1,433 @@
+/*
+ * example.c
+ *
+ * This file illustrates how to use the IJG code as a subroutine library
+ * to read or write JPEG image files.  You should look at this code in
+ * conjunction with the documentation file libjpeg.txt.
+ *
+ * This code will not do anything useful as-is, but it may be helpful as a
+ * skeleton for constructing routines that call the JPEG library.  
+ *
+ * We present these routines in the same coding style used in the JPEG code
+ * (ANSI function definitions, etc); but you are of course free to code your
+ * routines in a different style if you prefer.
+ */
+
+#include <stdio.h>
+
+/*
+ * Include file for users of JPEG library.
+ * You will need to have included system headers that define at least
+ * the typedefs FILE and size_t before you can include jpeglib.h.
+ * (stdio.h is sufficient on ANSI-conforming systems.)
+ * You may also wish to include "jerror.h".
+ */
+
+#include "jpeglib.h"
+
+/*
+ * <setjmp.h> is used for the optional error recovery mechanism shown in
+ * the second part of the example.
+ */
+
+#include <setjmp.h>
+
+
+
+/******************** JPEG COMPRESSION SAMPLE INTERFACE *******************/
+
+/* This half of the example shows how to feed data into the JPEG compressor.
+ * We present a minimal version that does not worry about refinements such
+ * as error recovery (the JPEG code will just exit() if it gets an error).
+ */
+
+
+/*
+ * IMAGE DATA FORMATS:
+ *
+ * The standard input image format is a rectangular array of pixels, with
+ * each pixel having the same number of "component" values (color channels).
+ * Each pixel row is an array of JSAMPLEs (which typically are unsigned chars).
+ * If you are working with color data, then the color values for each pixel
+ * must be adjacent in the row; for example, R,G,B,R,G,B,R,G,B,... for 24-bit
+ * RGB color.
+ *
+ * For this example, we'll assume that this data structure matches the way
+ * our application has stored the image in memory, so we can just pass a
+ * pointer to our image buffer.  In particular, let's say that the image is
+ * RGB color and is described by:
+ */
+
+extern JSAMPLE * image_buffer;	/* Points to large array of R,G,B-order data */
+extern int image_height;	/* Number of rows in image */
+extern int image_width;		/* Number of columns in image */
+
+
+/*
+ * Sample routine for JPEG compression.  We assume that the target file name
+ * and a compression quality factor are passed in.
+ */
+
+GLOBAL(void)
+write_JPEG_file (char * filename, int quality)
+{
+  /* This struct contains the JPEG compression parameters and pointers to
+   * working space (which is allocated as needed by the JPEG library).
+   * It is possible to have several such structures, representing multiple
+   * compression/decompression processes, in existence at once.  We refer
+   * to any one struct (and its associated working data) as a "JPEG object".
+   */
+  struct jpeg_compress_struct cinfo;
+  /* This struct represents a JPEG error handler.  It is declared separately
+   * because applications often want to supply a specialized error handler
+   * (see the second half of this file for an example).  But here we just
+   * take the easy way out and use the standard error handler, which will
+   * print a message on stderr and call exit() if compression fails.
+   * Note that this struct must live as long as the main JPEG parameter
+   * struct, to avoid dangling-pointer problems.
+   */
+  struct jpeg_error_mgr jerr;
+  /* More stuff */
+  FILE * outfile;		/* target file */
+  JSAMPROW row_pointer[1];	/* pointer to JSAMPLE row[s] */
+  int row_stride;		/* physical row width in image buffer */
+
+  /* Step 1: allocate and initialize JPEG compression object */
+
+  /* We have to set up the error handler first, in case the initialization
+   * step fails.  (Unlikely, but it could happen if you are out of memory.)
+   * This routine fills in the contents of struct jerr, and returns jerr's
+   * address which we place into the link field in cinfo.
+   */
+  cinfo.err = jpeg_std_error(&jerr);
+  /* Now we can initialize the JPEG compression object. */
+  jpeg_create_compress(&cinfo);
+
+  /* Step 2: specify data destination (eg, a file) */
+  /* Note: steps 2 and 3 can be done in either order. */
+
+  /* Here we use the library-supplied code to send compressed data to a
+   * stdio stream.  You can also write your own code to do something else.
+   * VERY IMPORTANT: use "b" option to fopen() if you are on a machine that
+   * requires it in order to write binary files.
+   */
+  if ((outfile = fopen(filename, "wb")) == NULL) {
+    fprintf(stderr, "can't open %s\n", filename);
+    exit(1);
+  }
+  jpeg_stdio_dest(&cinfo, outfile);
+
+  /* Step 3: set parameters for compression */
+
+  /* First we supply a description of the input image.
+   * Four fields of the cinfo struct must be filled in:
+   */
+  cinfo.image_width = image_width; 	/* image width and height, in pixels */
+  cinfo.image_height = image_height;
+  cinfo.input_components = 3;		/* # of color components per pixel */
+  cinfo.in_color_space = JCS_RGB; 	/* colorspace of input image */
+  /* Now use the library's routine to set default compression parameters.
+   * (You must set at least cinfo.in_color_space before calling this,
+   * since the defaults depend on the source color space.)
+   */
+  jpeg_set_defaults(&cinfo);
+  /* Now you can set any non-default parameters you wish to.
+   * Here we just illustrate the use of quality (quantization table) scaling:
+   */
+  jpeg_set_quality(&cinfo, quality, TRUE /* limit to baseline-JPEG values */);
+
+  /* Step 4: Start compressor */
+
+  /* TRUE ensures that we will write a complete interchange-JPEG file.
+   * Pass TRUE unless you are very sure of what you're doing.
+   */
+  jpeg_start_compress(&cinfo, TRUE);
+
+  /* Step 5: while (scan lines remain to be written) */
+  /*           jpeg_write_scanlines(...); */
+
+  /* Here we use the library's state variable cinfo.next_scanline as the
+   * loop counter, so that we don't have to keep track ourselves.
+   * To keep things simple, we pass one scanline per call; you can pass
+   * more if you wish, though.
+   */
+  row_stride = image_width * 3;	/* JSAMPLEs per row in image_buffer */
+
+  while (cinfo.next_scanline < cinfo.image_height) {
+    /* jpeg_write_scanlines expects an array of pointers to scanlines.
+     * Here the array is only one element long, but you could pass
+     * more than one scanline at a time if that's more convenient.
+     */
+    row_pointer[0] = & image_buffer[cinfo.next_scanline * row_stride];
+    (void) jpeg_write_scanlines(&cinfo, row_pointer, 1);
+  }
+
+  /* Step 6: Finish compression */
+
+  jpeg_finish_compress(&cinfo);
+  /* After finish_compress, we can close the output file. */
+  fclose(outfile);
+
+  /* Step 7: release JPEG compression object */
+
+  /* This is an important step since it will release a good deal of memory. */
+  jpeg_destroy_compress(&cinfo);
+
+  /* And we're done! */
+}
+
+
+/*
+ * SOME FINE POINTS:
+ *
+ * In the above loop, we ignored the return value of jpeg_write_scanlines,
+ * which is the number of scanlines actually written.  We could get away
+ * with this because we were only relying on the value of cinfo.next_scanline,
+ * which will be incremented correctly.  If you maintain additional loop
+ * variables then you should be careful to increment them properly.
+ * Actually, for output to a stdio stream you needn't worry, because
+ * then jpeg_write_scanlines will write all the lines passed (or else exit
+ * with a fatal error).  Partial writes can only occur if you use a data
+ * destination module that can demand suspension of the compressor.
+ * (If you don't know what that's for, you don't need it.)
+ *
+ * If the compressor requires full-image buffers (for entropy-coding
+ * optimization or a multi-scan JPEG file), it will create temporary
+ * files for anything that doesn't fit within the maximum-memory setting.
+ * (Note that temp files are NOT needed if you use the default parameters.)
+ * On some systems you may need to set up a signal handler to ensure that
+ * temporary files are deleted if the program is interrupted.  See libjpeg.txt.
+ *
+ * Scanlines MUST be supplied in top-to-bottom order if you want your JPEG
+ * files to be compatible with everyone else's.  If you cannot readily read
+ * your data in that order, you'll need an intermediate array to hold the
+ * image.  See rdtarga.c or rdbmp.c for examples of handling bottom-to-top
+ * source data using the JPEG code's internal virtual-array mechanisms.
+ */
+
+
+
+/******************** JPEG DECOMPRESSION SAMPLE INTERFACE *******************/
+
+/* This half of the example shows how to read data from the JPEG decompressor.
+ * It's a bit more refined than the above, in that we show:
+ *   (a) how to modify the JPEG library's standard error-reporting behavior;
+ *   (b) how to allocate workspace using the library's memory manager.
+ *
+ * Just to make this example a little different from the first one, we'll
+ * assume that we do not intend to put the whole image into an in-memory
+ * buffer, but to send it line-by-line someplace else.  We need a one-
+ * scanline-high JSAMPLE array as a work buffer, and we will let the JPEG
+ * memory manager allocate it for us.  This approach is actually quite useful
+ * because we don't need to remember to deallocate the buffer separately: it
+ * will go away automatically when the JPEG object is cleaned up.
+ */
+
+
+/*
+ * ERROR HANDLING:
+ *
+ * The JPEG library's standard error handler (jerror.c) is divided into
+ * several "methods" which you can override individually.  This lets you
+ * adjust the behavior without duplicating a lot of code, which you might
+ * have to update with each future release.
+ *
+ * Our example here shows how to override the "error_exit" method so that
+ * control is returned to the library's caller when a fatal error occurs,
+ * rather than calling exit() as the standard error_exit method does.
+ *
+ * We use C's setjmp/longjmp facility to return control.  This means that the
+ * routine which calls the JPEG library must first execute a setjmp() call to
+ * establish the return point.  We want the replacement error_exit to do a
+ * longjmp().  But we need to make the setjmp buffer accessible to the
+ * error_exit routine.  To do this, we make a private extension of the
+ * standard JPEG error handler object.  (If we were using C++, we'd say we
+ * were making a subclass of the regular error handler.)
+ *
+ * Here's the extended error handler struct:
+ */
+
+struct my_error_mgr {
+  struct jpeg_error_mgr pub;	/* "public" fields */
+
+  jmp_buf setjmp_buffer;	/* for return to caller */
+};
+
+typedef struct my_error_mgr * my_error_ptr;
+
+/*
+ * Here's the routine that will replace the standard error_exit method:
+ */
+
+METHODDEF(void)
+my_error_exit (j_common_ptr cinfo)
+{
+  /* cinfo->err really points to a my_error_mgr struct, so coerce pointer */
+  my_error_ptr myerr = (my_error_ptr) cinfo->err;
+
+  /* Always display the message. */
+  /* We could postpone this until after returning, if we chose. */
+  (*cinfo->err->output_message) (cinfo);
+
+  /* Return control to the setjmp point */
+  longjmp(myerr->setjmp_buffer, 1);
+}
+
+
+/*
+ * Sample routine for JPEG decompression.  We assume that the source file name
+ * is passed in.  We want to return 1 on success, 0 on error.
+ */
+
+
+GLOBAL(int)
+read_JPEG_file (char * filename)
+{
+  /* This struct contains the JPEG decompression parameters and pointers to
+   * working space (which is allocated as needed by the JPEG library).
+   */
+  struct jpeg_decompress_struct cinfo;
+  /* We use our private extension JPEG error handler.
+   * Note that this struct must live as long as the main JPEG parameter
+   * struct, to avoid dangling-pointer problems.
+   */
+  struct my_error_mgr jerr;
+  /* More stuff */
+  FILE * infile;		/* source file */
+  JSAMPARRAY buffer;		/* Output row buffer */
+  int row_stride;		/* physical row width in output buffer */
+
+  /* In this example we want to open the input file before doing anything else,
+   * so that the setjmp() error recovery below can assume the file is open.
+   * VERY IMPORTANT: use "b" option to fopen() if you are on a machine that
+   * requires it in order to read binary files.
+   */
+
+  if ((infile = fopen(filename, "rb")) == NULL) {
+    fprintf(stderr, "can't open %s\n", filename);
+    return 0;
+  }
+
+  /* Step 1: allocate and initialize JPEG decompression object */
+
+  /* We set up the normal JPEG error routines, then override error_exit. */
+  cinfo.err = jpeg_std_error(&jerr.pub);
+  jerr.pub.error_exit = my_error_exit;
+  /* Establish the setjmp return context for my_error_exit to use. */
+  if (setjmp(jerr.setjmp_buffer)) {
+    /* If we get here, the JPEG code has signaled an error.
+     * We need to clean up the JPEG object, close the input file, and return.
+     */
+    jpeg_destroy_decompress(&cinfo);
+    fclose(infile);
+    return 0;
+  }
+  /* Now we can initialize the JPEG decompression object. */
+  jpeg_create_decompress(&cinfo);
+
+  /* Step 2: specify data source (eg, a file) */
+
+  jpeg_stdio_src(&cinfo, infile);
+
+  /* Step 3: read file parameters with jpeg_read_header() */
+
+  (void) jpeg_read_header(&cinfo, TRUE);
+  /* We can ignore the return value from jpeg_read_header since
+   *   (a) suspension is not possible with the stdio data source, and
+   *   (b) we passed TRUE to reject a tables-only JPEG file as an error.
+   * See libjpeg.txt for more info.
+   */
+
+  /* Step 4: set parameters for decompression */
+
+  /* In this example, we don't need to change any of the defaults set by
+   * jpeg_read_header(), so we do nothing here.
+   */
+
+  /* Step 5: Start decompressor */
+
+  (void) jpeg_start_decompress(&cinfo);
+  /* We can ignore the return value since suspension is not possible
+   * with the stdio data source.
+   */
+
+  /* We may need to do some setup of our own at this point before reading
+   * the data.  After jpeg_start_decompress() we have the correct scaled
+   * output image dimensions available, as well as the output colormap
+   * if we asked for color quantization.
+   * In this example, we need to make an output work buffer of the right size.
+   */ 
+  /* JSAMPLEs per row in output buffer */
+  row_stride = cinfo.output_width * cinfo.output_components;
+  /* Make a one-row-high sample array that will go away when done with image */
+  buffer = (*cinfo.mem->alloc_sarray)
+		((j_common_ptr) &cinfo, JPOOL_IMAGE, row_stride, 1);
+
+  /* Step 6: while (scan lines remain to be read) */
+  /*           jpeg_read_scanlines(...); */
+
+  /* Here we use the library's state variable cinfo.output_scanline as the
+   * loop counter, so that we don't have to keep track ourselves.
+   */
+  while (cinfo.output_scanline < cinfo.output_height) {
+    /* jpeg_read_scanlines expects an array of pointers to scanlines.
+     * Here the array is only one element long, but you could ask for
+     * more than one scanline at a time if that's more convenient.
+     */
+    (void) jpeg_read_scanlines(&cinfo, buffer, 1);
+    /* Assume put_scanline_someplace wants a pointer and sample count. */
+    put_scanline_someplace(buffer[0], row_stride);
+  }
+
+  /* Step 7: Finish decompression */
+
+  (void) jpeg_finish_decompress(&cinfo);
+  /* We can ignore the return value since suspension is not possible
+   * with the stdio data source.
+   */
+
+  /* Step 8: Release JPEG decompression object */
+
+  /* This is an important step since it will release a good deal of memory. */
+  jpeg_destroy_decompress(&cinfo);
+
+  /* After finish_decompress, we can close the input file.
+   * Here we postpone it until after no more JPEG errors are possible,
+   * so as to simplify the setjmp error logic above.  (Actually, I don't
+   * think that jpeg_destroy can do an error exit, but why assume anything...)
+   */
+  fclose(infile);
+
+  /* At this point you may want to check to see whether any corrupt-data
+   * warnings occurred (test whether jerr.pub.num_warnings is nonzero).
+   */
+
+  /* And we're done! */
+  return 1;
+}
+
+
+/*
+ * SOME FINE POINTS:
+ *
+ * In the above code, we ignored the return value of jpeg_read_scanlines,
+ * which is the number of scanlines actually read.  We could get away with
+ * this because we asked for only one line at a time and we weren't using
+ * a suspending data source.  See libjpeg.txt for more info.
+ *
+ * We cheated a bit by calling alloc_sarray() after jpeg_start_decompress();
+ * we should have done it beforehand to ensure that the space would be
+ * counted against the JPEG max_memory setting.  In some systems the above
+ * code would risk an out-of-memory error.  However, in general we don't
+ * know the output image dimensions before jpeg_start_decompress(), unless we
+ * call jpeg_calc_output_dimensions().  See libjpeg.txt for more about this.
+ *
+ * Scanlines are returned in the same order as they appear in the JPEG file,
+ * which is standardly top-to-bottom.  If you must emit data bottom-to-top,
+ * you can use one of the virtual arrays provided by the JPEG memory manager
+ * to invert the data.  See wrbmp.c for an example.
+ *
+ * As with compression, some operating modes may require temporary files.
+ * On some systems you may need to set up a signal handler to ensure that
+ * temporary files are deleted if the program is interrupted.  See libjpeg.txt.
+ */
diff --git a/filelist.txt b/filelist.txt
new file mode 100644
index 0000000..a26ffcd
--- /dev/null
+++ b/filelist.txt
@@ -0,0 +1,210 @@
+IJG JPEG LIBRARY:  FILE LIST
+
+Copyright (C) 1994-2010, Thomas G. Lane, Guido Vollbeding, D. R. Commander.
+This file is part of the Independent JPEG Group's software.
+For conditions of distribution and use, see the accompanying README file.
+
+
+Here is a road map to the files in the IJG JPEG distribution.  The
+distribution includes the JPEG library proper, plus two application
+programs ("cjpeg" and "djpeg") which use the library to convert JPEG
+files to and from some other popular image formats.  A third application
+"jpegtran" uses the library to do lossless conversion between different
+variants of JPEG.  There are also two stand-alone applications,
+"rdjpgcom" and "wrjpgcom".
+
+
+THE JPEG LIBRARY
+================
+
+Include files:
+
+jpeglib.h	JPEG library's exported data and function declarations.
+jconfig.h	Configuration declarations.  Note: this file is not present
+		in the distribution; it is generated during installation.
+jmorecfg.h	Additional configuration declarations; need not be changed
+		for a standard installation.
+jerror.h	Declares JPEG library's error and trace message codes.
+jinclude.h	Central include file used by all IJG .c files to reference
+		system include files.
+jpegint.h	JPEG library's internal data structures.
+jchuff.h	Private declarations for Huffman encoder modules.
+jdhuff.h	Private declarations for Huffman decoder modules.
+jdct.h		Private declarations for forward & reverse DCT subsystems.
+jmemsys.h	Private declarations for memory management subsystem.
+jversion.h	Version information.
+
+Applications using the library should include jpeglib.h (which in turn
+includes jconfig.h and jmorecfg.h).  Optionally, jerror.h may be included
+if the application needs to reference individual JPEG error codes.  The
+other include files are intended for internal use and would not normally
+be included by an application program.  (cjpeg/djpeg/etc do use jinclude.h,
+since its function is to improve portability of the whole IJG distribution.
+Most other applications will directly include the system include files they
+want, and hence won't need jinclude.h.)
+
+
+C source code files:
+
+These files contain most of the functions intended to be called directly by
+an application program:
+
+jcapimin.c	Application program interface: core routines for compression.
+jcapistd.c	Application program interface: standard compression.
+jdapimin.c	Application program interface: core routines for decompression.
+jdapistd.c	Application program interface: standard decompression.
+jcomapi.c	Application program interface routines common to compression
+		and decompression.
+jcparam.c	Compression parameter setting helper routines.
+jctrans.c	API and library routines for transcoding compression.
+jdtrans.c	API and library routines for transcoding decompression.
+
+Compression side of the library:
+
+jcinit.c	Initialization: determines which other modules to use.
+jcmaster.c	Master control: setup and inter-pass sequencing logic.
+jcmainct.c	Main buffer controller (preprocessor => JPEG compressor).
+jcprepct.c	Preprocessor buffer controller.
+jccoefct.c	Buffer controller for DCT coefficient buffer.
+jccolor.c	Color space conversion.
+jcsample.c	Downsampling.
+jcdctmgr.c	DCT manager (DCT implementation selection & control).
+jfdctint.c	Forward DCT using slow-but-accurate integer method.
+jfdctfst.c	Forward DCT using faster, less accurate integer method.
+jfdctflt.c	Forward DCT using floating-point arithmetic.
+jchuff.c	Huffman entropy coding for sequential JPEG.
+jcphuff.c	Huffman entropy coding for progressive JPEG.
+jcmarker.c	JPEG marker writing.
+jdatadst.c	Data destination managers for memory and stdio output.
+
+Decompression side of the library:
+
+jdmaster.c	Master control: determines which other modules to use.
+jdinput.c	Input controller: controls input processing modules.
+jdmainct.c	Main buffer controller (JPEG decompressor => postprocessor).
+jdcoefct.c	Buffer controller for DCT coefficient buffer.
+jdpostct.c	Postprocessor buffer controller.
+jdmarker.c	JPEG marker reading.
+jdhuff.c	Huffman entropy decoding for sequential JPEG.
+jdphuff.c	Huffman entropy decoding for progressive JPEG.
+jddctmgr.c	IDCT manager (IDCT implementation selection & control).
+jidctint.c	Inverse DCT using slow-but-accurate integer method.
+jidctfst.c	Inverse DCT using faster, less accurate integer method.
+jidctflt.c	Inverse DCT using floating-point arithmetic.
+jidctred.c	Inverse DCTs with reduced-size outputs.
+jdsample.c	Upsampling.
+jdcolor.c	Color space conversion.
+jdmerge.c	Merged upsampling/color conversion (faster, lower quality).
+jquant1.c	One-pass color quantization using a fixed-spacing colormap.
+jquant2.c	Two-pass color quantization using a custom-generated colormap.
+		Also handles one-pass quantization to an externally given map.
+jdatasrc.c	Data source managers for memory and stdio input.
+
+Support files for both compression and decompression:
+
+jerror.c	Standard error handling routines (application replaceable).
+jmemmgr.c	System-independent (more or less) memory management code.
+jutils.c	Miscellaneous utility routines.
+
+jmemmgr.c relies on a system-dependent memory management module.  The IJG
+distribution includes the following implementations of the system-dependent
+module:
+
+jmemnobs.c	"No backing store": assumes adequate virtual memory exists.
+jmemansi.c	Makes temporary files with ANSI-standard routine tmpfile().
+jmemname.c	Makes temporary files with program-generated file names.
+jmemdos.c	Custom implementation for MS-DOS (16-bit environment only):
+		can use extended and expanded memory as well as temp files.
+jmemmac.c	Custom implementation for Apple Macintosh.
+
+Exactly one of the system-dependent modules should be configured into an
+installed JPEG library (see install.txt for hints about which one to use).
+On unusual systems you may find it worthwhile to make a special
+system-dependent memory manager.
+
+
+Non-C source code files:
+
+jmemdosa.asm	80x86 assembly code support for jmemdos.c; used only in
+		MS-DOS-specific configurations of the JPEG library.
+
+
+CJPEG/DJPEG/JPEGTRAN
+====================
+
+Include files:
+
+cdjpeg.h	Declarations shared by cjpeg/djpeg/jpegtran modules.
+cderror.h	Additional error and trace message codes for cjpeg et al.
+transupp.h	Declarations for jpegtran support routines in transupp.c.
+
+C source code files:
+
+cjpeg.c		Main program for cjpeg.
+djpeg.c		Main program for djpeg.
+jpegtran.c	Main program for jpegtran.
+cdjpeg.c	Utility routines used by all three programs.
+rdcolmap.c	Code to read a colormap file for djpeg's "-map" switch.
+rdswitch.c	Code to process some of cjpeg's more complex switches.
+		Also used by jpegtran.
+transupp.c	Support code for jpegtran: lossless image manipulations.
+
+Image file reader modules for cjpeg:
+
+rdbmp.c		BMP file input.
+rdgif.c		GIF file input (now just a stub).
+rdppm.c		PPM/PGM file input.
+rdrle.c		Utah RLE file input.
+rdtarga.c	Targa file input.
+
+Image file writer modules for djpeg:
+
+wrbmp.c		BMP file output.
+wrgif.c		GIF file output (a mere shadow of its former self).
+wrppm.c		PPM/PGM file output.
+wrrle.c		Utah RLE file output.
+wrtarga.c	Targa file output.
+
+
+RDJPGCOM/WRJPGCOM
+=================
+
+C source code files:
+
+rdjpgcom.c	Stand-alone rdjpgcom application.
+wrjpgcom.c	Stand-alone wrjpgcom application.
+
+These programs do not depend on the IJG library.  They do use
+jconfig.h and jinclude.h, only to improve portability.
+
+
+ADDITIONAL FILES
+================
+
+Documentation (see README for a guide to the documentation files):
+
+README		Master documentation file.
+*.txt		Other documentation files.
+*.1		Documentation in Unix man page format.
+change.log	Version-to-version change highlights.
+example.c	Sample code for calling JPEG library.
+
+Configuration/installation files and programs (see install.txt for more info):
+
+configure	Unix shell script to perform automatic configuration.
+configure.ac	Source file for use with Autoconf to generate configure.
+ltmain.sh	Support scripts for configure (from GNU libtool).
+config.guess
+config.sub
+depcomp
+missing
+install-sh	Install shell script for those Unix systems lacking one.
+Makefile.in	Makefile input for configure.
+Makefile.am	Source file for use with Automake to generate Makefile.in.
+jconfig.txt	Template for making jconfig.h by hand.
+aclocal.m4	M4 macro definitions for use with Autoconf.
+
+Test files (see install.txt for test procedure):
+
+test*.*		Source and comparison files for confidence test.
+		These are binary image files, NOT text files.
diff --git a/install.txt b/install.txt
new file mode 100644
index 0000000..2ee86ad
--- /dev/null
+++ b/install.txt
@@ -0,0 +1,1096 @@
+INSTALLATION INSTRUCTIONS for the Independent JPEG Group's JPEG software
+
+Copyright (C) 1991-2010, Thomas G. Lane, Guido Vollbeding.
+This file is part of the Independent JPEG Group's software.
+For conditions of distribution and use, see the accompanying README file.
+
+
+This file explains how to configure and install the IJG software.  We have
+tried to make this software extremely portable and flexible, so that it can be
+adapted to almost any environment.  The downside of this decision is that the
+installation process is complicated.  We have provided shortcuts to simplify
+the task on common systems.  But in any case, you will need at least a little
+familiarity with C programming and program build procedures for your system.
+
+If you are only using this software as part of a larger program, the larger
+program's installation procedure may take care of configuring the IJG code.
+For example, Ghostscript's installation script will configure the IJG code.
+You don't need to read this file if you just want to compile Ghostscript.
+
+If you are on a Unix machine, you may not need to read this file at all.
+Try doing
+	./configure
+	make
+	make test
+If that doesn't complain, do
+	make install
+(better do "make -n install" first to see if the makefile will put the files
+where you want them).  Read further if you run into snags or want to customize
+the code for your system.
+
+
+TABLE OF CONTENTS
+-----------------
+
+Before you start
+Configuring the software:
+	using the automatic "configure" script
+	using one of the supplied jconfig and makefile files
+	by hand
+Building the software
+Testing the software
+Installing the software
+Optional stuff
+Optimization
+Hints for specific systems
+
+
+BEFORE YOU START
+================
+
+Before installing the software you must unpack the distributed source code.
+Since you are reading this file, you have probably already succeeded in this
+task.  However, there is a potential for error if you needed to convert the
+files to the local standard text file format (for example, if you are on
+MS-DOS you may have converted LF end-of-line to CR/LF).  You must apply
+such conversion to all the files EXCEPT those whose names begin with "test".
+The test files contain binary data; if you change them in any way then the
+self-test will give bad results.
+
+Please check the last section of this file to see if there are hints for the
+specific machine or compiler you are using.
+
+
+CONFIGURING THE SOFTWARE
+========================
+
+To configure the IJG code for your system, you need to create two files:
+  * jconfig.h: contains values for system-dependent #define symbols.
+  * Makefile: controls the compilation process.
+(On a non-Unix machine, you may create "project files" or some other
+substitute for a Makefile.  jconfig.h is needed in any environment.)
+
+We provide three different ways to generate these files:
+  * On a Unix system, you can just run the "configure" script.
+  * We provide sample jconfig files and makefiles for popular machines;
+    if your machine matches one of the samples, just copy the right sample
+    files to jconfig.h and Makefile.
+  * If all else fails, read the instructions below and make your own files.
+
+
+Configuring the software using the automatic "configure" script
+---------------------------------------------------------------
+
+If you are on a Unix machine, you can just type
+	./configure
+and let the configure script construct appropriate configuration files.
+If you're using "csh" on an old version of System V, you might need to type
+	sh configure
+instead to prevent csh from trying to execute configure itself.
+Expect configure to run for a few minutes, particularly on slower machines;
+it works by compiling a series of test programs.
+
+Configure was created with GNU Autoconf and it follows the usual conventions
+for GNU configure scripts.  It makes a few assumptions that you may want to
+override.  You can do this by providing optional switches to configure:
+
+* Configure will build both static and shared libraries, if possible.
+If you want to build libjpeg only as a static library, say
+	./configure --disable-shared
+If you want to build libjpeg only as a shared library, say
+	./configure --disable-static
+Configure uses GNU libtool to take care of system-dependent shared library
+building methods.
+
+* Configure will use gcc (GNU C compiler) if it's available, otherwise cc.
+To force a particular compiler to be selected, use the CC option, for example
+	./configure CC='cc'
+The same method can be used to include any unusual compiler switches.
+For example, on HP-UX you probably want to say
+	./configure CC='cc -Aa'
+to get HP's compiler to run in ANSI mode.
+
+* The default CFLAGS setting is "-g" for non-gcc compilers, "-g -O2" for gcc.
+You can override this by saying, for example,
+	./configure CFLAGS='-O2'
+if you want to compile without debugging support.
+
+* Configure will set up the makefile so that "make install" will install files
+into /usr/local/bin, /usr/local/man, etc.  You can specify an installation
+prefix other than "/usr/local" by giving configure the option "--prefix=PATH".
+
+* If you don't have a lot of swap space, you may need to enable the IJG
+software's internal virtual memory mechanism.  To do this, give the option
+"--enable-maxmem=N" where N is the default maxmemory limit in megabytes.
+This is discussed in more detail under "Selecting a memory manager", below.
+You probably don't need to worry about this on reasonably-sized Unix machines,
+unless you plan to process very large images.
+
+Configure has some other features that are useful if you are cross-compiling
+or working in a network of multiple machine types; but if you need those
+features, you probably already know how to use them.
+
+
+Configuring the software using one of the supplied jconfig and makefile files
+-----------------------------------------------------------------------------
+
+If you have one of these systems, you can just use the provided configuration
+files:
+
+Makefile	jconfig file	System and/or compiler
+
+makefile.manx	jconfig.manx	Amiga, Manx Aztec C
+makefile.sas	jconfig.sas	Amiga, SAS C
+makeproj.mac	jconfig.mac	Apple Macintosh, Metrowerks CodeWarrior
+mak*jpeg.st	jconfig.st	Atari ST/STE/TT, Pure C or Turbo C
+makefile.bcc	jconfig.bcc	MS-DOS or OS/2, Borland C
+makefile.dj	jconfig.dj	MS-DOS, DJGPP (Delorie's port of GNU C)
+makefile.mc6	jconfig.mc6	MS-DOS, Microsoft C (16-bit only)
+makefile.wat	jconfig.wat	MS-DOS, OS/2, or Windows NT, Watcom C
+makefile.vc	jconfig.vc	Windows NT/95, MS Visual C++
+make*.vc6	jconfig.vc	Windows NT/95, MS Visual C++ 6
+make*.v10	jconfig.vc	Windows NT/95, MS Visual C++ 2010 (v10)
+makefile.mms	jconfig.vms	Digital VMS, with MMS software
+makefile.vms	jconfig.vms	Digital VMS, without MMS software
+
+Copy the proper jconfig file to jconfig.h and the makefile to Makefile (or
+whatever your system uses as the standard makefile name).  For more info see
+the appropriate system-specific hints section near the end of this file.
+
+
+Configuring the software by hand
+--------------------------------
+
+First, generate a jconfig.h file.  If you are moderately familiar with C,
+the comments in jconfig.txt should be enough information to do this; just
+copy jconfig.txt to jconfig.h and edit it appropriately.  Otherwise, you may
+prefer to use the ckconfig.c program.  You will need to compile and execute
+ckconfig.c by hand --- we hope you know at least enough to do that.
+ckconfig.c may not compile the first try (in fact, the whole idea is for it
+to fail if anything is going to).  If you get compile errors, fix them by
+editing ckconfig.c according to the directions given in ckconfig.c.  Once
+you get it to run, it will write a suitable jconfig.h file, and will also
+print out some advice about which makefile to use.
+
+You may also want to look at the canned jconfig files, if there is one for a
+system similar to yours.
+
+Second, select a makefile and copy it to Makefile (or whatever your system
+uses as the standard makefile name).  The most generic makefiles we provide
+are
+	makefile.ansi:	if your C compiler supports function prototypes
+	makefile.unix:	if not.
+(You have function prototypes if ckconfig.c put "#define HAVE_PROTOTYPES"
+in jconfig.h.)  You may want to start from one of the other makefiles if
+there is one for a system similar to yours.
+
+Look over the selected Makefile and adjust options as needed.  In particular
+you may want to change the CC and CFLAGS definitions.  For instance, if you
+are using GCC, set CC=gcc.  If you had to use any compiler switches to get
+ckconfig.c to work, make sure the same switches are in CFLAGS.
+
+If you are on a system that doesn't use makefiles, you'll need to set up
+project files (or whatever you do use) to compile all the source files and
+link them into executable files cjpeg, djpeg, jpegtran, rdjpgcom, and wrjpgcom.
+See the file lists in any of the makefiles to find out which files go into
+each program.  Note that the provided makefiles all make a "library" file
+libjpeg first, but you don't have to do that if you don't want to; the file
+lists identify which source files are actually needed for compression,
+decompression, or both.  As a last resort, you can make a batch script that
+just compiles everything and links it all together; makefile.vms is an example
+of this (it's for VMS systems that have no make-like utility).
+
+Here are comments about some specific configuration decisions you'll
+need to make:
+
+Command line style
+------------------
+
+These programs can use a Unix-like command line style which supports
+redirection and piping, like this:
+	cjpeg inputfile >outputfile
+	cjpeg <inputfile >outputfile
+	source program | cjpeg >outputfile
+The simpler "two file" command line style is just
+	cjpeg inputfile outputfile
+You may prefer the two-file style, particularly if you don't have pipes.
+
+You MUST use two-file style on any system that doesn't cope well with binary
+data fed through stdin/stdout; this is true for some MS-DOS compilers, for
+example.  If you're not on a Unix system, it's safest to assume you need
+two-file style.  (But if your compiler provides either the Posix-standard
+fdopen() library routine or a Microsoft-compatible setmode() routine, you
+can safely use the Unix command line style, by defining USE_FDOPEN or
+USE_SETMODE respectively.)
+
+To use the two-file style, make jconfig.h say "#define TWO_FILE_COMMANDLINE".
+
+Selecting a memory manager
+--------------------------
+
+The IJG code is capable of working on images that are too big to fit in main
+memory; data is swapped out to temporary files as necessary.  However, the
+code to do this is rather system-dependent.  We provide five different
+memory managers:
+
+* jmemansi.c	This version uses the ANSI-standard library routine tmpfile(),
+		which not all non-ANSI systems have.  On some systems
+		tmpfile() may put the temporary file in a non-optimal
+		location; if you don't like what it does, use jmemname.c.
+
+* jmemname.c	This version creates named temporary files.  For anything
+		except a Unix machine, you'll need to configure the
+		select_file_name() routine appropriately; see the comments
+		near the head of jmemname.c.  If you use this version, define
+		NEED_SIGNAL_CATCHER in jconfig.h to make sure the temp files
+		are removed if the program is aborted.
+
+* jmemnobs.c	(That stands for No Backing Store :-).)  This will compile on
+		almost any system, but it assumes you have enough main memory
+		or virtual memory to hold the biggest images you work with.
+
+* jmemdos.c	This should be used with most 16-bit MS-DOS compilers.
+		See the system-specific notes about MS-DOS for more info.
+		IMPORTANT: if you use this, define USE_MSDOS_MEMMGR in
+		jconfig.h, and include the assembly file jmemdosa.asm in the
+		programs.  The supplied makefiles and jconfig files for
+		16-bit MS-DOS compilers already do both.
+
+* jmemmac.c	Custom version for Apple Macintosh; see the system-specific
+		notes for Macintosh for more info.
+
+To use a particular memory manager, change the SYSDEPMEM variable in your
+makefile to equal the corresponding object file name (for example, jmemansi.o
+or jmemansi.obj for jmemansi.c).
+
+If you have plenty of (real or virtual) main memory, just use jmemnobs.c.
+"Plenty" means about ten bytes for every pixel in the largest images
+you plan to process, so a lot of systems don't meet this criterion.
+If yours doesn't, try jmemansi.c first.  If that doesn't compile, you'll have
+to use jmemname.c; be sure to adjust select_file_name() for local conditions.
+You may also need to change unlink() to remove() in close_backing_store().
+
+Except with jmemnobs.c or jmemmac.c, you need to adjust the DEFAULT_MAX_MEM
+setting to a reasonable value for your system (either by adding a #define for
+DEFAULT_MAX_MEM to jconfig.h, or by adding a -D switch to the Makefile).
+This value limits the amount of data space the program will attempt to
+allocate.  Code and static data space isn't counted, so the actual memory
+needs for cjpeg or djpeg are typically 100 to 150Kb more than the max-memory
+setting.  Larger max-memory settings reduce the amount of I/O needed to
+process a large image, but too large a value can result in "insufficient
+memory" failures.  On most Unix machines (and other systems with virtual
+memory), just set DEFAULT_MAX_MEM to several million and forget it.  At the
+other end of the spectrum, for MS-DOS machines you probably can't go much
+above 300K to 400K.  (On MS-DOS the value refers to conventional memory only.
+Extended/expanded memory is handled separately by jmemdos.c.)
+
+
+BUILDING THE SOFTWARE
+=====================
+
+Now you should be able to compile the software.  Just say "make" (or
+whatever's necessary to start the compilation).  Have a cup of coffee.
+
+Here are some things that could go wrong:
+
+If your compiler complains about undefined structures, you should be able to
+shut it up by putting "#define INCOMPLETE_TYPES_BROKEN" in jconfig.h.
+
+If you have trouble with missing system include files or inclusion of the
+wrong ones, read jinclude.h.  This shouldn't happen if you used configure
+or ckconfig.c to set up jconfig.h.
+
+There are a fair number of routines that do not use all of their parameters;
+some compilers will issue warnings about this, which you can ignore.  There
+are also a few configuration checks that may give "unreachable code" warnings.
+Any other warning deserves investigation.
+
+If you don't have a getenv() library routine, define NO_GETENV.
+
+Also see the system-specific hints, below.
+
+
+TESTING THE SOFTWARE
+====================
+
+As a quick test of functionality we've included a small sample image in
+several forms:
+	testorig.jpg	Starting point for the djpeg tests.
+	testimg.ppm	The output of djpeg testorig.jpg
+	testimg.bmp	The output of djpeg -bmp -colors 256 testorig.jpg
+	testimg.jpg	The output of cjpeg testimg.ppm
+	testprog.jpg	Progressive-mode equivalent of testorig.jpg.
+	testimgp.jpg	The output of cjpeg -progressive -optimize testimg.ppm
+(The first- and second-generation .jpg files aren't identical since JPEG is
+lossy.)  If you can generate duplicates of the testimg* files then you
+probably have working programs.
+
+With most of the makefiles, "make test" will perform the necessary
+comparisons.
+
+If you're using a makefile that doesn't provide the test option, run djpeg
+and cjpeg by hand and compare the output files to testimg* with whatever
+binary file comparison tool you have.  The files should be bit-for-bit
+identical.
+
+If the programs complain "MAX_ALLOC_CHUNK is wrong, please fix", then you
+need to reduce MAX_ALLOC_CHUNK to a value that fits in type size_t.
+Try adding "#define MAX_ALLOC_CHUNK 65520L" to jconfig.h.  A less likely
+configuration error is "ALIGN_TYPE is wrong, please fix": defining ALIGN_TYPE
+as long should take care of that one.
+
+If the cjpeg test run fails with "Missing Huffman code table entry", it's a
+good bet that you needed to define RIGHT_SHIFT_IS_UNSIGNED.  Go back to the
+configuration step and run ckconfig.c.  (This is a good plan for any other
+test failure, too.)
+
+If you are using Unix (one-file) command line style on a non-Unix system,
+it's a good idea to check that binary I/O through stdin/stdout actually
+works.  You should get the same results from "djpeg <testorig.jpg >out.ppm"
+as from "djpeg -outfile out.ppm testorig.jpg".  Note that the makefiles all
+use the latter style and therefore do not exercise stdin/stdout!  If this
+check fails, try recompiling with USE_SETMODE or USE_FDOPEN defined.
+If it still doesn't work, better use two-file style.
+
+If you chose a memory manager other than jmemnobs.c, you should test that
+temporary-file usage works.  Try "djpeg -bmp -colors 256 -max 0 testorig.jpg"
+and make sure its output matches testimg.bmp.  If you have any really large
+images handy, try compressing them with -optimize and/or decompressing with
+-colors 256 to make sure your DEFAULT_MAX_MEM setting is not too large.
+
+NOTE: this is far from an exhaustive test of the JPEG software; some modules,
+such as 1-pass color quantization, are not exercised at all.  It's just a
+quick test to give you some confidence that you haven't missed something
+major.
+
+
+INSTALLING THE SOFTWARE
+=======================
+
+Once you're done with the above steps, you can install the software by
+copying the executable files (cjpeg, djpeg, jpegtran, rdjpgcom, and wrjpgcom)
+to wherever you normally install programs.  On Unix systems, you'll also want
+to put the man pages (cjpeg.1, djpeg.1, jpegtran.1, rdjpgcom.1, wrjpgcom.1)
+in the man-page directory.  The pre-fab makefiles don't support this step
+since there's such a wide variety of installation procedures on different
+systems.
+
+If you generated a Makefile with the "configure" script, you can just say
+	make install
+to install the programs and their man pages into the standard places.
+(You'll probably need to be root to do this.)  We recommend first saying
+	make -n install
+to see where configure thought the files should go.  You may need to edit
+the Makefile, particularly if your system's conventions for man page
+filenames don't match what configure expects.
+
+If you want to install the IJG library itself, for use in compiling other
+programs besides ours, then you need to put the four include files
+	jpeglib.h jerror.h jconfig.h jmorecfg.h
+into your include-file directory, and put the library file libjpeg.a
+(extension may vary depending on system) wherever library files go.
+If you generated a Makefile with "configure", it will do what it thinks
+is the right thing if you say
+	make install-lib
+
+
+OPTIONAL STUFF
+==============
+
+Progress monitor:
+
+If you like, you can #define PROGRESS_REPORT (in jconfig.h) to enable display
+of percent-done progress reports.  The routine provided in cdjpeg.c merely
+prints percentages to stderr, but you can customize it to do something
+fancier.
+
+Utah RLE file format support:
+
+We distribute the software with support for RLE image files (Utah Raster
+Toolkit format) disabled, because the RLE support won't compile without the
+Utah library.  If you have URT version 3.1 or later, you can enable RLE
+support as follows:
+	1.  #define RLE_SUPPORTED in jconfig.h.
+	2.  Add a -I option to CFLAGS in the Makefile for the directory
+	    containing the URT .h files (typically the "include"
+	    subdirectory of the URT distribution).
+	3.  Add -L... -lrle to LDLIBS in the Makefile, where ... specifies
+	    the directory containing the URT "librle.a" file (typically the
+	    "lib" subdirectory of the URT distribution).
+
+Support for 12-bit-deep pixel data:
+
+The JPEG standard allows either 8-bit or 12-bit data precision.  (For color,
+this means 8 or 12 bits per channel, of course.)  If you need to work with
+deeper than 8-bit data, you can compile the IJG code for 12-bit operation.
+To do so:
+  1. In jmorecfg.h, define BITS_IN_JSAMPLE as 12 rather than 8.
+  2. In jconfig.h, undefine BMP_SUPPORTED, RLE_SUPPORTED, and TARGA_SUPPORTED,
+     because the code for those formats doesn't handle 12-bit data and won't
+     even compile.  (The PPM code does work, as explained below.  The GIF
+     code works too; it scales 8-bit GIF data to and from 12-bit depth
+     automatically.)
+  3. Compile.  Don't expect "make test" to pass, since the supplied test
+     files are for 8-bit data.
+
+Currently, 12-bit support does not work on 16-bit-int machines.
+
+Note that a 12-bit version will not read 8-bit JPEG files, nor vice versa;
+so you'll want to keep around a regular 8-bit compilation as well.
+(Run-time selection of data depth, to allow a single copy that does both,
+is possible but would probably slow things down considerably; it's very low
+on our to-do list.)
+
+The PPM reader (rdppm.c) can read 12-bit data from either text-format or
+binary-format PPM and PGM files.  Binary-format PPM/PGM files which have a
+maxval greater than 255 are assumed to use 2 bytes per sample, MSB first
+(big-endian order).  As of early 1995, 2-byte binary format is not
+officially supported by the PBMPLUS library, but it is expected that a
+future release of PBMPLUS will support it.  Note that the PPM reader will
+read files of any maxval regardless of the BITS_IN_JSAMPLE setting; incoming
+data is automatically rescaled to either maxval=255 or maxval=4095 as
+appropriate for the cjpeg bit depth.
+
+The PPM writer (wrppm.c) will normally write 2-byte binary PPM or PGM
+format, maxval 4095, when compiled with BITS_IN_JSAMPLE=12.  Since this
+format is not yet widely supported, you can disable it by compiling wrppm.c
+with PPM_NORAWWORD defined; then the data is scaled down to 8 bits to make a
+standard 1-byte/sample PPM or PGM file.  (Yes, this means still another copy
+of djpeg to keep around.  But hopefully you won't need it for very long.
+Poskanzer's supposed to get that new PBMPLUS release out Real Soon Now.)
+
+Of course, if you are working with 12-bit data, you probably have it stored
+in some other, nonstandard format.  In that case you'll probably want to
+write your own I/O modules to read and write your format.
+
+Note that a 12-bit version of cjpeg always runs in "-optimize" mode, in
+order to generate valid Huffman tables.  This is necessary because our
+default Huffman tables only cover 8-bit data.
+
+Removing code:
+
+If you need to make a smaller version of the JPEG software, some optional
+functions can be removed at compile time.  See the xxx_SUPPORTED #defines in
+jconfig.h and jmorecfg.h.  If at all possible, we recommend that you leave in
+decoder support for all valid JPEG files, to ensure that you can read anyone's
+output.  Taking out support for image file formats that you don't use is the
+most painless way to make the programs smaller.  Another possibility is to
+remove some of the DCT methods: in particular, the "IFAST" method may not be
+enough faster than the others to be worth keeping on your machine.  (If you
+do remove ISLOW or IFAST, be sure to redefine JDCT_DEFAULT or JDCT_FASTEST
+to a supported method, by adding a #define in jconfig.h.)
+
+
+OPTIMIZATION
+============
+
+Unless you own a Cray, you'll probably be interested in making the JPEG
+software go as fast as possible.  This section covers some machine-dependent
+optimizations you may want to try.  We suggest that before trying any of
+this, you first get the basic installation to pass the self-test step.
+Repeat the self-test after any optimization to make sure that you haven't
+broken anything.
+
+The integer DCT routines perform a lot of multiplications.  These
+multiplications must yield 32-bit results, but none of their input values
+are more than 16 bits wide.  On many machines, notably the 680x0 and 80x86
+CPUs, a 16x16=>32 bit multiply instruction is faster than a full 32x32=>32
+bit multiply.  Unfortunately there is no portable way to specify such a
+multiplication in C, but some compilers can generate one when you use the
+right combination of casts.  See the MULTIPLYxxx macro definitions in
+jdct.h.  If your compiler makes "int" be 32 bits and "short" be 16 bits,
+defining SHORTxSHORT_32 is fairly likely to work.  When experimenting with
+alternate definitions, be sure to test not only whether the code still works
+(use the self-test), but also whether it is actually faster --- on some
+compilers, alternate definitions may compute the right answer, yet be slower
+than the default.  Timing cjpeg on a large PGM (grayscale) input file is the
+best way to check this, as the DCT will be the largest fraction of the runtime
+in that mode.  (Note: some of the distributed compiler-specific jconfig files
+already contain #define switches to select appropriate MULTIPLYxxx
+definitions.)
+
+If your machine has sufficiently fast floating point hardware, you may find
+that the float DCT method is faster than the integer DCT methods, even
+after tweaking the integer multiply macros.  In that case you may want to
+make the float DCT be the default method.  (The only objection to this is
+that float DCT results may vary slightly across machines.)  To do that, add
+"#define JDCT_DEFAULT JDCT_FLOAT" to jconfig.h.  Even if you don't change
+the default, you should redefine JDCT_FASTEST, which is the method selected
+by djpeg's -fast switch.  Don't forget to update the documentation files
+(usage.txt and/or cjpeg.1, djpeg.1) to agree with what you've done.
+
+If access to "short" arrays is slow on your machine, it may be a win to
+define type JCOEF as int rather than short.  This will cost a good deal of
+memory though, particularly in some multi-pass modes, so don't do it unless
+you have memory to burn and short is REALLY slow.
+
+If your compiler can compile function calls in-line, make sure the INLINE
+macro in jmorecfg.h is defined as the keyword that marks a function
+inline-able.  Some compilers have a switch that tells the compiler to inline
+any function it thinks is profitable (e.g., -finline-functions for gcc).
+Enabling such a switch is likely to make the compiled code bigger but faster.
+
+In general, it's worth trying the maximum optimization level of your compiler,
+and experimenting with any optional optimizations such as loop unrolling.
+(Unfortunately, far too many compilers have optimizer bugs ... be prepared to
+back off if the code fails self-test.)  If you do any experimentation along
+these lines, please report the optimal settings to jpeg-info@uc.ag so we
+can mention them in future releases.  Be sure to specify your machine
+and compiler version.
+
+
+HINTS FOR SPECIFIC SYSTEMS
+==========================
+
+We welcome reports on changes needed for systems not mentioned here.  Submit
+'em to jpeg-info@uc.ag.  Also, if configure or ckconfig.c is wrong about how
+to configure the JPEG software for your system, please let us know.
+
+
+Acorn RISC OS:
+
+(Thanks to Simon Middleton for these hints on compiling with Desktop C.)
+After renaming the files according to Acorn conventions, take a copy of
+makefile.ansi, change all occurrences of 'libjpeg.a' to 'libjpeg.o' and
+change these definitions as indicated:
+
+CFLAGS= -throwback -IC: -Wn
+LDLIBS=C:o.Stubs
+SYSDEPMEM=jmemansi.o
+LN=Link
+AR=LibFile -c -o
+
+Also add a new line '.c.o:; $(cc) $< $(cflags) -c -o $@'.  Remove the
+lines '$(RM) libjpeg.o' and '$(AR2) libjpeg.o' and the 'jconfig.h'
+dependency section.
+
+Copy jconfig.txt to jconfig.h.  Edit jconfig.h to define TWO_FILE_COMMANDLINE
+and CHAR_IS_UNSIGNED.
+
+Run the makefile using !AMU not !Make.  If you want to use the 'clean' and
+'test' makefile entries then you will have to fiddle with the syntax a bit
+and rename the test files.
+
+
+Amiga:
+
+SAS C 6.50 reportedly is too buggy to compile the IJG code properly.
+A patch to update to 6.51 is available from SAS or AmiNet FTP sites.
+
+The supplied config files are set up to use jmemname.c as the memory
+manager, with temporary files being created on the device named by
+"JPEGTMP:".
+
+
+Atari ST/STE/TT:
+
+Copy the project files makcjpeg.st, makdjpeg.st, maktjpeg.st, and makljpeg.st
+to cjpeg.prj, djpeg.prj, jpegtran.prj, and libjpeg.prj respectively.  The
+project files should work as-is with Pure C.  For Turbo C, change library
+filenames "pc..." to "tc..." in each project file.  Note that libjpeg.prj
+selects jmemansi.c as the recommended memory manager.  You'll probably want to
+adjust the DEFAULT_MAX_MEM setting --- you want it to be a couple hundred K
+less than your normal free memory.  Put "#define DEFAULT_MAX_MEM nnnn" into
+jconfig.h to do this.
+
+To use the 68881/68882 coprocessor for the floating point DCT, add the
+compiler option "-8" to the project files and replace pcfltlib.lib with
+pc881lib.lib in cjpeg.prj and djpeg.prj.  Or if you don't have a
+coprocessor, you may prefer to remove the float DCT code by undefining
+DCT_FLOAT_SUPPORTED in jmorecfg.h (since without a coprocessor, the float
+code will be too slow to be useful).  In that case, you can delete
+pcfltlib.lib from the project files.
+
+Note that you must make libjpeg.lib before making cjpeg.ttp, djpeg.ttp,
+or jpegtran.ttp.  You'll have to perform the self-test by hand.
+
+We haven't bothered to include project files for rdjpgcom and wrjpgcom.
+Those source files should just be compiled by themselves; they don't
+depend on the JPEG library.  You can use the default.prj project file
+of the Pure C distribution to make the programs.
+
+There is a bug in some older versions of the Turbo C library which causes the
+space used by temporary files created with "tmpfile()" not to be freed after
+an abnormal program exit.  If you check your disk afterwards, you will find
+cluster chains that are allocated but not used by a file.  This should not
+happen in cjpeg/djpeg/jpegtran, since we enable a signal catcher to explicitly
+close temp files before exiting.  But if you use the JPEG library with your
+own code, be sure to supply a signal catcher, or else use a different
+system-dependent memory manager.
+
+
+Cray:
+
+Should you be so fortunate as to be running JPEG on a Cray YMP, there is a
+compiler bug in old versions of Cray's Standard C (prior to 3.1).  If you
+still have an old compiler, you'll need to insert a line reading
+"#pragma novector" just before the loop	
+    for (i = 1; i <= (int) htbl->bits[l]; i++)
+      huffsize[p++] = (char) l;
+in fix_huff_tbl (in V5beta1, line 204 of jchuff.c and line 176 of jdhuff.c).
+[This bug may or may not still occur with the current IJG code, but it's
+probably a dead issue anyway...]
+
+
+HP-UX:
+
+If you have HP-UX 7.05 or later with the "software development" C compiler,
+you should run the compiler in ANSI mode.  If using the configure script,
+say
+	./configure CC='cc -Aa'
+(or -Ae if you prefer).  If configuring by hand, use makefile.ansi and add
+"-Aa" to the CFLAGS line in the makefile.
+
+If you have a pre-7.05 system, or if you are using the non-ANSI C compiler
+delivered with a minimum HP-UX system, then you must use makefile.unix
+(and do NOT add -Aa); or just run configure without the CC option.
+
+On HP 9000 series 800 machines, the HP C compiler is buggy in revisions prior
+to A.08.07.  If you get complaints about "not a typedef name", you'll have to
+use makefile.unix, or run configure without the CC option.
+
+
+Macintosh, generic comments:
+
+The supplied user-interface files (cjpeg.c, djpeg.c, etc) are set up to
+provide a Unix-style command line interface.  You can use this interface on
+the Mac by means of the ccommand() library routine provided by Metrowerks
+CodeWarrior or Think C.  This is only appropriate for testing the library,
+however; to make a user-friendly equivalent of cjpeg/djpeg you'd really want
+to develop a Mac-style user interface.  There isn't a complete example
+available at the moment, but there are some helpful starting points:
+1. Sam Bushell's free "To JPEG" applet provides drag-and-drop conversion to
+JPEG under System 7 and later.  This only illustrates how to use the
+compression half of the library, but it does a very nice job of that part.
+The CodeWarrior source code is available from http://www.pobox.com/~jsam.
+2. Jim Brunner prepared a Mac-style user interface for both compression and
+decompression.  Unfortunately, it hasn't been updated since IJG v4, and
+the library's API has changed considerably since then.  Still it may be of
+some help, particularly as a guide to compiling the IJG code under Think C.
+Jim's code is available from the Info-Mac archives, at sumex-aim.stanford.edu
+or mirrors thereof; see file /info-mac/dev/src/jpeg-convert-c.hqx.
+
+jmemmac.c is the recommended memory manager back end for Macintosh.  It uses
+NewPtr/DisposePtr instead of malloc/free, and has a Mac-specific
+implementation of jpeg_mem_available().  It also creates temporary files that
+follow Mac conventions.  (That part of the code relies on System-7-or-later OS
+functions.  See the comments in jmemmac.c if you need to run it on System 6.)
+NOTE that USE_MAC_MEMMGR must be defined in jconfig.h to use jmemmac.c.
+
+You can also use jmemnobs.c, if you don't care about handling images larger
+than available memory.  If you use any memory manager back end other than
+jmemmac.c, we recommend replacing "malloc" and "free" by "NewPtr" and
+"DisposePtr", because Mac C libraries often have peculiar implementations of
+malloc/free.  (For instance, free() may not return the freed space to the
+Mac Memory Manager.  This is undesirable for the IJG code because jmemmgr.c
+already clumps space requests.)
+
+
+Macintosh, Metrowerks CodeWarrior:
+
+The Unix-command-line-style interface can be used by defining USE_CCOMMAND.
+You'll also need to define TWO_FILE_COMMANDLINE to avoid stdin/stdout.
+This means that when using the cjpeg/djpeg programs, you'll have to type the
+input and output file names in the "Arguments" text-edit box, rather than
+using the file radio buttons.  (Perhaps USE_FDOPEN or USE_SETMODE would
+eliminate the problem, but I haven't heard from anyone who's tried it.)
+
+On 680x0 Macs, Metrowerks defines type "double" as a 10-byte IEEE extended
+float.  jmemmgr.c won't like this: it wants sizeof(ALIGN_TYPE) to be a power
+of 2.  Add "#define ALIGN_TYPE long" to jconfig.h to eliminate the complaint.
+
+The supplied configuration file jconfig.mac can be used for your jconfig.h;
+it includes all the recommended symbol definitions.  If you have AppleScript
+installed, you can run the supplied script makeproj.mac to create CodeWarrior
+project files for the library and the testbed applications, then build the
+library and applications.  (Thanks to Dan Sears and Don Agro for this nifty
+hack, which saves us from trying to maintain CodeWarrior project files as part
+of the IJG distribution...)
+
+
+Macintosh, Think C:
+
+The documentation in Jim Brunner's "JPEG Convert" source code (see above)
+includes detailed build instructions for Think C; it's probably somewhat
+out of date for the current release, but may be helpful.
+
+If you want to build the minimal command line version, proceed as follows.
+You'll have to prepare project files for the programs; we don't include any
+in the distribution since they are not text files.  Use the file lists in
+any of the supplied makefiles as a guide.  Also add the ANSI and Unix C
+libraries in a separate segment.  You may need to divide the JPEG files into
+more than one segment; we recommend dividing compression and decompression
+modules.  Define USE_CCOMMAND in jconfig.h so that the ccommand() routine is
+called.  You must also define TWO_FILE_COMMANDLINE because stdin/stdout
+don't handle binary data correctly.
+
+On 680x0 Macs, Think C defines type "double" as a 12-byte IEEE extended float.
+jmemmgr.c won't like this: it wants sizeof(ALIGN_TYPE) to be a power of 2.
+Add "#define ALIGN_TYPE long" to jconfig.h to eliminate the complaint.
+
+jconfig.mac should work as a jconfig.h configuration file for Think C,
+but the makeproj.mac AppleScript script is specific to CodeWarrior.  Sorry.
+
+
+MIPS R3000:
+
+MIPS's cc version 1.31 has a rather nasty optimization bug.  Don't use -O
+if you have that compiler version.  (Use "cc -V" to check the version.)
+Note that the R3000 chip is found in workstations from DEC and others.
+
+
+MS-DOS, generic comments for 16-bit compilers:
+
+The IJG code is designed to work well in 80x86 "small" or "medium" memory
+models (i.e., data pointers are 16 bits unless explicitly declared "far";
+code pointers can be either size).  You may be able to use small model to
+compile cjpeg or djpeg by itself, but you will probably have to use medium
+model for any larger application.  This won't make much difference in
+performance.  You *will* take a noticeable performance hit if you use a
+large-data memory model, and you should avoid "huge" model if at all
+possible.  Be sure that NEED_FAR_POINTERS is defined in jconfig.h if you use
+a small-data memory model; be sure it is NOT defined if you use a large-data
+model.  (The supplied makefiles and jconfig files for Borland and Microsoft C
+compile in medium model and define NEED_FAR_POINTERS.)
+
+The DOS-specific memory manager, jmemdos.c, should be used if possible.
+It needs some assembly-code routines which are in jmemdosa.asm; make sure
+your makefile assembles that file and includes it in the library.  If you
+don't have a suitable assembler, you can get pre-assembled object files for
+jmemdosa by FTP from ftp.uu.net:/graphics/jpeg/jdosaobj.zip.  (DOS-oriented
+distributions of the IJG source code often include these object files.)
+
+When using jmemdos.c, jconfig.h must define USE_MSDOS_MEMMGR and must set
+MAX_ALLOC_CHUNK to less than 64K (65520L is a typical value).  If your
+C library's far-heap malloc() can't allocate blocks that large, reduce
+MAX_ALLOC_CHUNK to whatever it can handle.
+
+If you can't use jmemdos.c for some reason --- for example, because you
+don't have an assembler to assemble jmemdosa.asm --- you'll have to fall
+back to jmemansi.c or jmemname.c.  You'll probably still need to set
+MAX_ALLOC_CHUNK in jconfig.h, because most DOS C libraries won't malloc()
+more than 64K at a time.  IMPORTANT: if you use jmemansi.c or jmemname.c,
+you will have to compile in a large-data memory model in order to get the
+right stdio library.  Too bad.
+
+wrjpgcom needs to be compiled in large model, because it malloc()s a 64KB
+work area to hold the comment text.  If your C library's malloc can't
+handle that, reduce MAX_COM_LENGTH as necessary in wrjpgcom.c.
+
+Most MS-DOS compilers treat stdin/stdout as text files, so you must use
+two-file command line style.  But if your compiler has either fdopen() or
+setmode(), you can use one-file style if you like.  To do this, define
+USE_SETMODE or USE_FDOPEN so that stdin/stdout will be set to binary mode.
+(USE_SETMODE seems to work with more DOS compilers than USE_FDOPEN.)  You
+should test that I/O through stdin/stdout produces the same results as I/O
+to explicitly named files... the "make test" procedures in the supplied
+makefiles do NOT use stdin/stdout.
+
+
+MS-DOS, generic comments for 32-bit compilers:
+
+None of the above comments about memory models apply if you are using a
+32-bit flat-memory-space environment, such as DJGPP or Watcom C.  (And you
+should use one if you have it, as performance will be much better than
+8086-compatible code!)  For flat-memory-space compilers, do NOT define
+NEED_FAR_POINTERS, and do NOT use jmemdos.c.  Use jmemnobs.c if the
+environment supplies adequate virtual memory, otherwise use jmemansi.c or
+jmemname.c.
+
+You'll still need to be careful about binary I/O through stdin/stdout.
+See the last paragraph of the previous section.
+
+
+MS-DOS, Borland C:
+
+Be sure to convert all the source files to DOS text format (CR/LF newlines).
+Although Borland C will often work OK with unmodified Unix (LF newlines)
+source files, sometimes it will give bogus compile errors.
+"Illegal character '#'" is the most common such error.  (This is true with
+Borland C 3.1, but perhaps is fixed in newer releases.)
+
+If you want one-file command line style, just undefine TWO_FILE_COMMANDLINE.
+jconfig.bcc already includes #define USE_SETMODE to make this work.
+(fdopen does not work correctly.)
+
+
+MS-DOS, Microsoft C:
+
+makefile.mc6 works with Microsoft C, DOS Visual C++, etc.  It should only
+be used if you want to build a 16-bit (small or medium memory model) program.
+
+If you want one-file command line style, just undefine TWO_FILE_COMMANDLINE.
+jconfig.mc6 already includes #define USE_SETMODE to make this work.
+(fdopen does not work correctly.)
+
+Note that this makefile assumes that the working copy of itself is called
+"makefile".  If you want to call it something else, say "makefile.mak",
+be sure to adjust the dependency line that reads "$(RFILE) : makefile".
+Otherwise the make will fail because it doesn't know how to create "makefile".
+Worse, some releases of Microsoft's make utilities give an incorrect error
+message in this situation.
+
+Old versions of MS C fail with an "out of macro expansion space" error
+because they can't cope with the macro TRACEMS8 (defined in jerror.h).
+If this happens to you, the easiest solution is to change TRACEMS8 to
+expand to nothing.  You'll lose the ability to dump out JPEG coefficient
+tables with djpeg -debug -debug, but at least you can compile.
+
+Original MS C 6.0 is very buggy; it compiles incorrect code unless you turn
+off optimization entirely (remove -O from CFLAGS).  6.00A is better, but it
+still generates bad code if you enable loop optimizations (-Ol or -Ox).
+
+MS C 8.0 crashes when compiling jquant1.c with optimization switch /Oo ...
+which is on by default.  To work around this bug, compile that one file
+with /Oo-.
+
+
+Microsoft Windows (all versions), generic comments:
+
+Some Windows system include files define typedef boolean as "unsigned char".
+The IJG code also defines typedef boolean, but we make it "int" by default.
+This doesn't affect the IJG programs because we don't import those Windows
+include files.  But if you use the JPEG library in your own program, and some
+of your program's files import one definition of boolean while some import the
+other, you can get all sorts of mysterious problems.  A good preventive step
+is to make the IJG library use "unsigned char" for boolean.  To do that,
+add something like this to your jconfig.h file:
+	/* Define "boolean" as unsigned char, not int, per Windows custom */
+	#ifndef __RPCNDR_H__	/* don't conflict if rpcndr.h already read */
+	typedef unsigned char boolean;
+	#endif
+	#define HAVE_BOOLEAN	/* prevent jmorecfg.h from redefining it */
+(This is already in jconfig.vc, by the way.)
+
+windef.h contains the declarations
+	#define far
+	#define FAR far
+Since jmorecfg.h tries to define FAR as empty, you may get a compiler
+warning if you include both jpeglib.h and windef.h (which windows.h
+includes).  To suppress the warning, you can put "#ifndef FAR"/"#endif"
+around the line "#define FAR" in jmorecfg.h.
+(Something like this is already in jmorecfg.h, by the way.)
+
+When using the library in a Windows application, you will almost certainly
+want to modify or replace the error handler module jerror.c, since our
+default error handler does a couple of inappropriate things:
+  1. it tries to write error and warning messages on stderr;
+  2. in event of a fatal error, it exits by calling exit().
+
+A simple stopgap solution for problem 1 is to replace the line
+	fprintf(stderr, "%s\n", buffer);
+(in output_message in jerror.c) with
+	MessageBox(GetActiveWindow(),buffer,"JPEG Error",MB_OK|MB_ICONERROR);
+It's highly recommended that you at least do that much, since otherwise
+error messages will disappear into nowhere.  (Beginning with IJG v6b, this
+code is already present in jerror.c; just define USE_WINDOWS_MESSAGEBOX in
+jconfig.h to enable it.)
+
+The proper solution for problem 2 is to return control to your calling
+application after a library error.  This can be done with the setjmp/longjmp
+technique discussed in libjpeg.txt and illustrated in example.c.  (NOTE:
+some older Windows C compilers provide versions of setjmp/longjmp that
+don't actually work under Windows.  You may need to use the Windows system
+functions Catch and Throw instead.)
+
+The recommended memory manager under Windows is jmemnobs.c; in other words,
+let Windows do any virtual memory management needed.  You should NOT use
+jmemdos.c nor jmemdosa.asm under Windows.
+
+For Windows 3.1, we recommend compiling in medium or large memory model;
+for newer Windows versions, use a 32-bit flat memory model.  (See the MS-DOS
+sections above for more info about memory models.)  In the 16-bit memory
+models only, you'll need to put
+	#define MAX_ALLOC_CHUNK 65520L	/* Maximum request to malloc() */
+into jconfig.h to limit allocation chunks to 64Kb.  (Without that, you'd
+have to use huge memory model, which slows things down unnecessarily.)
+jmemnobs.c works without modification in large or flat memory models, but to
+use medium model, you need to modify its jpeg_get_large and jpeg_free_large
+routines to allocate far memory.  In any case, you might like to replace
+its calls to malloc and free with direct calls on Windows memory allocation
+functions.
+
+You may also want to modify jdatasrc.c and jdatadst.c to use Windows file
+operations rather than fread/fwrite.  This is only necessary if your C
+compiler doesn't provide a competent implementation of C stdio functions.
+
+You might want to tweak the RGB_xxx macros in jmorecfg.h so that the library
+will accept or deliver color pixels in BGR sample order, not RGB; BGR order
+is usually more convenient under Windows.  Note that this change will break
+the sample applications cjpeg/djpeg, but the library itself works fine.
+
+
+Many people want to convert the IJG library into a DLL.  This is reasonably
+straightforward, but watch out for the following:
+
+  1. Don't try to compile as a DLL in small or medium memory model; use
+large model, or even better, 32-bit flat model.  Many places in the IJG code
+assume the address of a local variable is an ordinary (not FAR) pointer;
+that isn't true in a medium-model DLL.
+
+  2. Microsoft C cannot pass file pointers between applications and DLLs.
+(See Microsoft Knowledge Base, PSS ID Number Q50336.)  So jdatasrc.c and
+jdatadst.c don't work if you open a file in your application and then pass
+the pointer to the DLL.  One workaround is to make jdatasrc.c/jdatadst.c
+part of your main application rather than part of the DLL.
+
+  3. You'll probably need to modify the macros GLOBAL() and EXTERN() to
+attach suitable linkage keywords to the exported routine names.  Similarly,
+you'll want to modify METHODDEF() and JMETHOD() to ensure function pointers
+are declared in a way that lets application routines be called back through
+the function pointers.  These macros are in jmorecfg.h.  Typical definitions
+for a 16-bit DLL are:
+	#define GLOBAL(type)		type _far _pascal _loadds _export
+	#define EXTERN(type)		extern type _far _pascal _loadds
+	#define METHODDEF(type)		static type _far _pascal
+	#define JMETHOD(type,methodname,arglist)  \
+		type (_far _pascal *methodname) arglist
+For a 32-bit DLL you may want something like
+	#define GLOBAL(type)		__declspec(dllexport) type
+	#define EXTERN(type)		extern __declspec(dllexport) type
+Although not all the GLOBAL routines are actually intended to be called by
+the application, the performance cost of making them all DLL entry points is
+negligible.
+
+The unmodified IJG library presents a very C-specific application interface,
+so the resulting DLL is only usable from C or C++ applications.  There has
+been some talk of writing wrapper code that would present a simpler interface
+usable from other languages, such as Visual Basic.  This is on our to-do list
+but hasn't been very high priority --- any volunteers out there?
+
+
+Microsoft Windows, Borland C:
+
+The provided jconfig.bcc should work OK in a 32-bit Windows environment,
+but you'll need to tweak it in a 16-bit environment (you'd need to define
+NEED_FAR_POINTERS and MAX_ALLOC_CHUNK).  Beware that makefile.bcc will need
+alteration if you want to use it for Windows --- in particular, you should
+use jmemnobs.c not jmemdos.c under Windows.
+
+Borland C++ 4.5 fails with an internal compiler error when trying to compile
+jdmerge.c in 32-bit mode.  If enough people complain, perhaps Borland will fix
+it.  In the meantime, the simplest known workaround is to add a redundant
+definition of the variable range_limit in h2v1_merged_upsample(), at the head
+of the block that handles odd image width (about line 268 in v6 jdmerge.c):
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    register JSAMPLE * range_limit = cinfo->sample_range_limit; /* ADD THIS */
+    cb = GETJSAMPLE(*inptr1);
+Pretty bizarre, especially since the very similar routine h2v2_merged_upsample
+doesn't trigger the bug.
+Recent reports suggest that this bug does not occur with "bcc32a" (the
+Pentium-optimized version of the compiler).
+
+Another report from a user of Borland C 4.5 was that incorrect code (leading
+to a color shift in processed images) was produced if any of the following
+optimization switch combinations were used: 
+	-Ot -Og
+	-Ot -Op
+	-Ot -Om
+So try backing off on optimization if you see such a problem.  (Are there
+several different releases all numbered "4.5"??)
+
+
+Microsoft Windows, Microsoft Visual C++:
+
+jconfig.vc should work OK with any Microsoft compiler for a 32-bit memory
+model.  makefile.vc is intended for command-line use.  (If you are using
+the Developer Studio environment, you may prefer the DevStudio project
+files; see below.)
+
+IJG JPEG 7 adds extern "C" to jpeglib.h.  This avoids the need to put
+extern "C" { ... } around #include "jpeglib.h" in your C++ application.
+You can also force VC++ to treat the library as C++ code by renaming
+all the *.c files to *.cpp (and adjusting the makefile to match).
+In this case you also need to define the symbol DONT_USE_EXTERN_C in
+the configuration to prevent jpeglib.h from using extern "C".
+
+
+Microsoft Windows, Microsoft Visual C++ 6 Developer Studio:
+
+We include makefiles that should work as project files in DevStudio 6.0 or
+later.  There is a library makefile that builds the IJG library as a static
+Win32 library, and application makefiles that build the sample applications
+as Win32 console applications.  (Even if you only want the library, we
+recommend building the applications so that you can run the self-test.)
+
+To use:
+1. Open the command prompt, change to the main directory and execute the
+   command line
+	NMAKE /f makefile.vc  setup-vc6
+   This will move jconfig.vc to jconfig.h and makefiles to project files.
+   (Note that the renaming is critical!)
+2. Open the workspace file jpeg.dsw, build the library project.
+   (If you are using DevStudio more recent than 6.0, you'll probably
+   get a message saying that the project files are being updated.)
+3. Open the workspace file apps.dsw, build the application projects.
+4. To perform the self-test, execute the command line
+	NMAKE /f makefile.vc  test-build
+5. Move the application .exe files from `app`\Release to an
+   appropriate location on your path.
+
+
+Microsoft Windows, Microsoft Visual C++ 2010 Developer Studio (v10):
+
+We include makefiles that should work as project files in Visual Studio
+2010 or later.  There is a library makefile that builds the IJG library
+as a static Win32 library, and application makefiles that build the sample
+applications as Win32 console applications.  (Even if you only want the
+library, we recommend building the applications so that you can run the
+self-test.)
+
+To use:
+1. Open the command prompt, change to the main directory and execute the
+   command line
+	NMAKE /f makefile.vc  setup-v10
+   This will move jconfig.vc to jconfig.h and makefiles to project files.
+   (Note that the renaming is critical!)
+2. Open the solution file jpeg.sln, build the library project.
+   (If you are using Visual Studio more recent than 2010 (v10), you'll
+   probably get a message saying that the project files are being updated.)
+3. Open the solution file apps.sln, build the application projects.
+4. To perform the self-test, execute the command line
+	NMAKE /f makefile.vc  test-build
+5. Move the application .exe files from `app`\Release to an
+   appropriate location on your path.
+
+Note:
+There seems to be an optimization bug in the compiler which causes the
+self-test to fail with the color quantization option.
+We have disabled optimization for the file jquant2.c in the library
+project file which causes the self-test to pass properly.
+
+
+OS/2, Borland C++:
+
+Watch out for optimization bugs in older Borland compilers; you may need
+to back off the optimization switch settings.  See the comments in
+makefile.bcc.
+
+
+SGI:
+
+On some SGI systems, you may need to set "AR2= ar -ts" in the Makefile.
+If you are using configure, you can do this by saying
+	./configure RANLIB='ar -ts'
+This change is not needed on all SGIs.  Use it only if the make fails at the
+stage of linking the completed programs.
+
+On the MIPS R4000 architecture (Indy, etc.), the compiler option "-mips2"
+reportedly speeds up the float DCT method substantially, enough to make it
+faster than the default int method (but still slower than the fast int
+method).  If you use -mips2, you may want to alter the default DCT method to
+be float.  To do this, put "#define JDCT_DEFAULT JDCT_FLOAT" in jconfig.h.
+
+
+VMS:
+
+On an Alpha/VMS system with MMS, be sure to use the "/Marco=Alpha=1"
+qualifier with MMS when building the JPEG package.
+
+VAX/VMS v5.5-1 may have problems with the test step of the build procedure
+reporting differences when it compares the original and test images.  If the
+error points to the last block of the files, it is most likely bogus and may
+be safely ignored.  It seems to be because the files are Stream_LF and
+Backup/Compare has difficulty with the (presumably) null padded files.
+This problem was not observed on VAX/VMS v6.1 or AXP/VMS v6.1.
diff --git a/jcapimin.c b/jcapimin.c
new file mode 100644
index 0000000..20ba9e9
--- /dev/null
+++ b/jcapimin.c
@@ -0,0 +1,292 @@
+/*
+ * jcapimin.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-compression case or the transcoding-only
+ * case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jcapistd.c.  But also see jcparam.c for
+ * parameter-setup helper routines, jcomapi.c for routines shared by
+ * compression and decompression, and jctrans.c for the transcoding case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Initialization of a JPEG compression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateCompress (j_compress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;		/* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != SIZEOF(struct jpeg_compress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE, 
+	     (int) SIZEOF(struct jpeg_compress_struct), (int) structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr * err = cinfo->err;
+    void * client_data = cinfo->client_data; /* ignore Purify complaint here */
+    MEMZERO(cinfo, SIZEOF(struct jpeg_compress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = FALSE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr) cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->dest = NULL;
+
+  cinfo->comp_info = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    cinfo->quant_tbl_ptrs[i] = NULL;
+#if JPEG_LIB_VERSION >= 70
+    cinfo->q_scale_factor[i] = 100;
+#endif
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+#if JPEG_LIB_VERSION >= 80
+  /* Must do it here for emit_dqt in case jpeg_write_tables is used */
+  cinfo->block_size = DCTSIZE;
+  cinfo->natural_order = jpeg_natural_order;
+  cinfo->lim_Se = DCTSIZE2-1;
+#endif
+
+  cinfo->script_space = NULL;
+
+  cinfo->input_gamma = 1.0;	/* in case application forgets */
+
+  /* OK, I'm ready */
+  cinfo->global_state = CSTATE_START;
+}
+
+
+/*
+ * Destruction of a JPEG compression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_compress (j_compress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG compression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_compress (j_compress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Forcibly suppress or un-suppress all quantization and Huffman tables.
+ * Marks all currently defined tables as already written (if suppress)
+ * or not written (if !suppress).  This will control whether they get emitted
+ * by a subsequent jpeg_start_compress call.
+ *
+ * This routine is exported for use by applications that want to produce
+ * abbreviated JPEG datastreams.  It logically belongs in jcparam.c, but
+ * since it is called by jpeg_start_compress, we put it here --- otherwise
+ * jcparam.o would be linked whether the application used it or not.
+ */
+
+GLOBAL(void)
+jpeg_suppress_tables (j_compress_ptr cinfo, boolean suppress)
+{
+  int i;
+  JQUANT_TBL * qtbl;
+  JHUFF_TBL * htbl;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if ((qtbl = cinfo->quant_tbl_ptrs[i]) != NULL)
+      qtbl->sent_table = suppress;
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    if ((htbl = cinfo->dc_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+    if ((htbl = cinfo->ac_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+  }
+}
+
+
+/*
+ * Finish JPEG compression.
+ *
+ * If a multipass operating mode was selected, this may do a great deal of
+ * work including most of the actual output.
+ */
+
+GLOBAL(void)
+jpeg_finish_compress (j_compress_ptr cinfo)
+{
+  JDIMENSION iMCU_row;
+
+  if (cinfo->global_state == CSTATE_SCANNING ||
+      cinfo->global_state == CSTATE_RAW_OK) {
+    /* Terminate first pass */
+    if (cinfo->next_scanline < cinfo->image_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_pass) (cinfo);
+  } else if (cinfo->global_state != CSTATE_WRCOEFS)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any remaining passes */
+  while (! cinfo->master->is_last_pass) {
+    (*cinfo->master->prepare_for_pass) (cinfo);
+    for (iMCU_row = 0; iMCU_row < cinfo->total_iMCU_rows; iMCU_row++) {
+      if (cinfo->progress != NULL) {
+	cinfo->progress->pass_counter = (long) iMCU_row;
+	cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows;
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      }
+      /* We bypass the main controller and invoke coef controller directly;
+       * all work is being done from the coefficient buffer.
+       */
+      if (! (*cinfo->coef->compress_data) (cinfo, (JSAMPIMAGE) NULL))
+	ERREXIT(cinfo, JERR_CANT_SUSPEND);
+    }
+    (*cinfo->master->finish_pass) (cinfo);
+  }
+  /* Write EOI, do final cleanup */
+  (*cinfo->marker->write_file_trailer) (cinfo);
+  (*cinfo->dest->term_destination) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr) cinfo);
+}
+
+
+/*
+ * Write a special marker.
+ * This is only recommended for writing COM or APPn markers.
+ * Must be called after jpeg_start_compress() and before
+ * first call to jpeg_write_scanlines() or jpeg_write_raw_data().
+ */
+
+GLOBAL(void)
+jpeg_write_marker (j_compress_ptr cinfo, int marker,
+		   const JOCTET *dataptr, unsigned int datalen)
+{
+  JMETHOD(void, write_marker_byte, (j_compress_ptr info, int val));
+
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+  write_marker_byte = cinfo->marker->write_marker_byte;	/* copy for speed */
+  while (datalen--) {
+    (*write_marker_byte) (cinfo, *dataptr);
+    dataptr++;
+  }
+}
+
+/* Same, but piecemeal. */
+
+GLOBAL(void)
+jpeg_write_m_header (j_compress_ptr cinfo, int marker, unsigned int datalen)
+{
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+}
+
+GLOBAL(void)
+jpeg_write_m_byte (j_compress_ptr cinfo, int val)
+{
+  (*cinfo->marker->write_marker_byte) (cinfo, val);
+}
+
+
+/*
+ * Alternate compression function: just write an abbreviated table file.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * To produce a pair of files containing abbreviated tables and abbreviated
+ * image data, one would proceed as follows:
+ *
+ *		initialize JPEG object
+ *		set JPEG parameters
+ *		set destination to table file
+ *		jpeg_write_tables(cinfo);
+ *		set destination to image file
+ *		jpeg_start_compress(cinfo, FALSE);
+ *		write data...
+ *		jpeg_finish_compress(cinfo);
+ *
+ * jpeg_write_tables has the side effect of marking all tables written
+ * (same as jpeg_suppress_tables(..., TRUE)).  Thus a subsequent start_compress
+ * will not re-emit the tables unless it is passed write_all_tables=TRUE.
+ */
+
+GLOBAL(void)
+jpeg_write_tables (j_compress_ptr cinfo)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Initialize the marker writer ... bit of a crock to do it here. */
+  jinit_marker_writer(cinfo);
+  /* Write them tables! */
+  (*cinfo->marker->write_tables_only) (cinfo);
+  /* And clean up. */
+  (*cinfo->dest->term_destination) (cinfo);
+  /*
+   * In library releases up through v6a, we called jpeg_abort() here to free
+   * any working memory allocated by the destination manager and marker
+   * writer.  Some applications had a problem with that: they allocated space
+   * of their own from the library memory manager, and didn't want it to go
+   * away during write_tables.  So now we do nothing.  This will cause a
+   * memory leak if an app calls write_tables repeatedly without doing a full
+   * compression cycle or otherwise resetting the JPEG object.  However, that
+   * seems less bad than unexpectedly freeing memory in the normal case.
+   * An app that prefers the old behavior can call jpeg_abort for itself after
+   * each call to jpeg_write_tables().
+   */
+}
diff --git a/jcapistd.c b/jcapistd.c
new file mode 100644
index 0000000..c0320b1
--- /dev/null
+++ b/jcapistd.c
@@ -0,0 +1,161 @@
+/*
+ * jcapistd.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-compression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_compress, it will end up linking in the entire compressor.
+ * We thus must separate this file from jcapimin.c to avoid linking the
+ * whole compression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Compression initialization.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * We require a write_all_tables parameter as a failsafe check when writing
+ * multiple datastreams from the same compression object.  Since prior runs
+ * will have left all the tables marked sent_table=TRUE, a subsequent run
+ * would emit an abbreviated stream (no tables) by default.  This may be what
+ * is wanted, but for safety's sake it should not be the default behavior:
+ * programmers should have to make a deliberate choice to emit abbreviated
+ * images.  Therefore the documentation and examples should encourage people
+ * to pass write_all_tables=TRUE; then it will take active thought to do the
+ * wrong thing.
+ */
+
+GLOBAL(void)
+jpeg_start_compress (j_compress_ptr cinfo, boolean write_all_tables)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (write_all_tables)
+    jpeg_suppress_tables(cinfo, FALSE);	/* mark all tables to be written */
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  jinit_compress_master(cinfo);
+  /* Set up for the first pass */
+  (*cinfo->master->prepare_for_pass) (cinfo);
+  /* Ready for application to drive first pass through jpeg_write_scanlines
+   * or jpeg_write_raw_data.
+   */
+  cinfo->next_scanline = 0;
+  cinfo->global_state = (cinfo->raw_data_in ? CSTATE_RAW_OK : CSTATE_SCANNING);
+}
+
+
+/*
+ * Write some scanlines of data to the JPEG compressor.
+ *
+ * The return value will be the number of lines actually written.
+ * This should be less than the supplied num_lines only in case that
+ * the data destination module has requested suspension of the compressor,
+ * or if more than image_height scanlines are passed in.
+ *
+ * Note: we warn about excess calls to jpeg_write_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * excess scanlines passed in the last valid call are *silently* ignored,
+ * so that the application need not adjust num_lines for end-of-image
+ * when using a multiple-scanline buffer.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_write_scanlines (j_compress_ptr cinfo, JSAMPARRAY scanlines,
+		      JDIMENSION num_lines)
+{
+  JDIMENSION row_ctr, rows_left;
+
+  if (cinfo->global_state != CSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height)
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * jpeg_write_scanlines.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and jpeg_write_scanlines.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Ignore any extra scanlines at bottom of image. */
+  rows_left = cinfo->image_height - cinfo->next_scanline;
+  if (num_lines > rows_left)
+    num_lines = rows_left;
+
+  row_ctr = 0;
+  (*cinfo->main->process_data) (cinfo, scanlines, &row_ctr, num_lines);
+  cinfo->next_scanline += row_ctr;
+  return row_ctr;
+}
+
+
+/*
+ * Alternate entry point to write raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_write_raw_data (j_compress_ptr cinfo, JSAMPIMAGE data,
+		     JDIMENSION num_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->global_state != CSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * jpeg_write_raw_data.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and jpeg_write_raw_data.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Verify that at least one iMCU row has been passed. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * DCTSIZE;
+  if (num_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Directly compress the row. */
+  if (! (*cinfo->coef->compress_data) (cinfo, data)) {
+    /* If compressor did not consume the whole row, suspend processing. */
+    return 0;
+  }
+
+  /* OK, we processed one iMCU row. */
+  cinfo->next_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
diff --git a/jccoefct.c b/jccoefct.c
new file mode 100644
index 0000000..1963ddb
--- /dev/null
+++ b/jccoefct.c
@@ -0,0 +1,449 @@
+/*
+ * jccoefct.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the coefficient buffer controller for compression.
+ * This controller is the top level of the JPEG compressor proper.
+ * The coefficient buffer lies between forward-DCT and entropy encoding steps.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* We use a full-image coefficient buffer when doing Huffman optimization,
+ * and also for writing multiple-scan JPEG files.  In all cases, the DCT
+ * step is run during the first pass, and subsequent passes need only read
+ * the buffered coefficients.
+ */
+#ifdef ENTROPY_OPT_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#else
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#endif
+#endif
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;	/* iMCU row # within image */
+  JDIMENSION mcu_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* For single-pass compression, it's sufficient to buffer just one MCU
+   * (although this may prove a bit slow in practice).  We allocate a
+   * workspace of C_MAX_BLOCKS_IN_MCU coefficient blocks, and reuse it for each
+   * MCU constructed and sent.  (On 80x86, the workspace is FAR even though
+   * it's not really very big; this is to keep the module interfaces unchanged
+   * when a large coefficient buffer is necessary.)
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays.
+   */
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+
+/* Forward declarations */
+METHODDEF(boolean) compress_data
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+METHODDEF(boolean) compress_first_pass
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+METHODDEF(boolean) compress_output
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+#endif
+
+
+LOCAL(void)
+start_iMCU_row (j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (coef->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_data;
+    break;
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_first_pass;
+    break;
+  case JBUF_CRANK_DEST:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_output;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data in the single-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(boolean)
+compress_data (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, bi, ci, yindex, yoffset, blockcnt;
+  JDIMENSION ypos, xpos;
+  jpeg_component_info *compptr;
+
+  /* Loop to write as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num <= last_MCU_col;
+	 MCU_col_num++) {
+      /* Determine where data comes from in input_buf and do the DCT thing.
+       * Each call on forward_DCT processes a horizontal row of DCT blocks
+       * as wide as an MCU; we rely on having allocated the MCU_buffer[] blocks
+       * sequentially.  Dummy blocks at the right or bottom edge are filled in
+       * specially.  The data in them does not matter for image reconstruction,
+       * so we fill them with values that will encode to the smallest amount of
+       * data, viz: all zeroes in the AC entries, DC entries equal to previous
+       * block's DC value.  (Thanks to Thomas Kinsman for this idea.)
+       */
+      blkn = 0;
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						: compptr->last_col_width;
+	xpos = MCU_col_num * compptr->MCU_sample_width;
+	ypos = yoffset * DCTSIZE; /* ypos == (yoffset+yindex) * DCTSIZE */
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (coef->iMCU_row_num < last_iMCU_row ||
+	      yoffset+yindex < compptr->last_row_height) {
+	    (*cinfo->fdct->forward_DCT) (cinfo, compptr,
+					 input_buf[compptr->component_index],
+					 coef->MCU_buffer[blkn],
+					 ypos, xpos, (JDIMENSION) blockcnt);
+	    if (blockcnt < compptr->MCU_width) {
+	      /* Create some dummy blocks at the right edge of the image. */
+	      jzero_far((void FAR *) coef->MCU_buffer[blkn + blockcnt],
+			(compptr->MCU_width - blockcnt) * SIZEOF(JBLOCK));
+	      for (bi = blockcnt; bi < compptr->MCU_width; bi++) {
+		coef->MCU_buffer[blkn+bi][0][0] = coef->MCU_buffer[blkn+bi-1][0][0];
+	      }
+	    }
+	  } else {
+	    /* Create a row of dummy blocks at the bottom of the image. */
+	    jzero_far((void FAR *) coef->MCU_buffer[blkn],
+		      compptr->MCU_width * SIZEOF(JBLOCK));
+	    for (bi = 0; bi < compptr->MCU_width; bi++) {
+	      coef->MCU_buffer[blkn+bi][0][0] = coef->MCU_buffer[blkn-1][0][0];
+	    }
+	  }
+	  blkn += compptr->MCU_width;
+	  ypos += DCTSIZE;
+	}
+      }
+      /* Try to write the MCU.  In event of a suspension failure, we will
+       * re-DCT the MCU on restart (a bit inefficient, could be fixed...)
+       */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+
+/*
+ * Process some data in the first pass of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * This amount of data is read from the source buffer, DCT'd and quantized,
+ * and saved into the virtual arrays.  We also generate suitable dummy blocks
+ * as needed at the right and lower edges.  (The dummy blocks are constructed
+ * in the virtual arrays, which have been padded appropriately.)  This makes
+ * it possible for subsequent passes not to worry about real vs. dummy blocks.
+ *
+ * We must also emit the data to the entropy encoder.  This is conveniently
+ * done by calling compress_output() after we've loaded the current strip
+ * of the virtual arrays.
+ *
+ * NB: input_buf contains a plane for each component in image.  All
+ * components are DCT'd and loaded into the virtual arrays in this pass.
+ * However, it may be that only a subset of the components are emitted to
+ * the entropy encoder during this first pass; be careful about looking
+ * at the scan-dependent variables (MCU dimensions, etc).
+ */
+
+METHODDEF(boolean)
+compress_first_pass (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION blocks_across, MCUs_across, MCUindex;
+  int bi, ci, h_samp_factor, block_row, block_rows, ndummy;
+  JCOEF lastDC;
+  jpeg_component_info *compptr;
+  JBLOCKARRAY buffer;
+  JBLOCKROW thisblockrow, lastblockrow;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[ci],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, TRUE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (coef->iMCU_row_num < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here, since may not be set! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    blocks_across = compptr->width_in_blocks;
+    h_samp_factor = compptr->h_samp_factor;
+    /* Count number of dummy blocks to be added at the right margin. */
+    ndummy = (int) (blocks_across % h_samp_factor);
+    if (ndummy > 0)
+      ndummy = h_samp_factor - ndummy;
+    /* Perform DCT for all non-dummy blocks in this iMCU row.  Each call
+     * on forward_DCT processes a complete horizontal row of DCT blocks.
+     */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      thisblockrow = buffer[block_row];
+      (*cinfo->fdct->forward_DCT) (cinfo, compptr,
+				   input_buf[ci], thisblockrow,
+				   (JDIMENSION) (block_row * DCTSIZE),
+				   (JDIMENSION) 0, blocks_across);
+      if (ndummy > 0) {
+	/* Create dummy blocks at the right edge of the image. */
+	thisblockrow += blocks_across; /* => first dummy block */
+	jzero_far((void FAR *) thisblockrow, ndummy * SIZEOF(JBLOCK));
+	lastDC = thisblockrow[-1][0];
+	for (bi = 0; bi < ndummy; bi++) {
+	  thisblockrow[bi][0] = lastDC;
+	}
+      }
+    }
+    /* If at end of image, create dummy block rows as needed.
+     * The tricky part here is that within each MCU, we want the DC values
+     * of the dummy blocks to match the last real block's DC value.
+     * This squeezes a few more bytes out of the resulting file...
+     */
+    if (coef->iMCU_row_num == last_iMCU_row) {
+      blocks_across += ndummy;	/* include lower right corner */
+      MCUs_across = blocks_across / h_samp_factor;
+      for (block_row = block_rows; block_row < compptr->v_samp_factor;
+	   block_row++) {
+	thisblockrow = buffer[block_row];
+	lastblockrow = buffer[block_row-1];
+	jzero_far((void FAR *) thisblockrow,
+		  (size_t) (blocks_across * SIZEOF(JBLOCK)));
+	for (MCUindex = 0; MCUindex < MCUs_across; MCUindex++) {
+	  lastDC = lastblockrow[h_samp_factor-1][0];
+	  for (bi = 0; bi < h_samp_factor; bi++) {
+	    thisblockrow[bi][0] = lastDC;
+	  }
+	  thisblockrow += h_samp_factor; /* advance to next MCU in row */
+	  lastblockrow += h_samp_factor;
+	}
+      }
+    }
+  }
+  /* NB: compress_output will increment iMCU_row_num if successful.
+   * A suspension return will result in redoing all the work above next time.
+   */
+
+  /* Emit data to the entropy encoder, sharing code with subsequent passes */
+  return compress_output(cinfo, input_buf);
+}
+
+
+/*
+ * Process some data in subsequent passes of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(boolean)
+compress_output (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan.
+   * NB: during first pass, this is safe only because the buffers will
+   * already be aligned properly, so jmemmgr.c won't need to do any I/O.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+	    coef->MCU_buffer[blkn++] = buffer_ptr++;
+	  }
+	}
+      }
+      /* Try to write the MCU. */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+#endif /* FULL_COEF_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+jinit_c_coef_controller (j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *) coef;
+  coef->pub.start_pass = start_pass_coef;
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    int ci;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
+				(long) compptr->h_samp_factor),
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor),
+	 (JDIMENSION) compptr->v_samp_factor);
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+    for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->whole_image[0] = NULL; /* flag for no virtual arrays */
+  }
+}
diff --git a/jccolor.c b/jccolor.c
new file mode 100644
index 0000000..2e2bfd2
--- /dev/null
+++ b/jccolor.c
@@ -0,0 +1,567 @@
+/*
+ * jccolor.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright 2009 D. R. Commander
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains input colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_converter pub; /* public fields */
+
+  /* Private state for RGB->YCC conversion */
+  INT32 * rgb_ycc_tab;		/* => table for RGB to YCbCr conversion */
+} my_color_converter;
+
+typedef my_color_converter * my_cconvert_ptr;
+
+
+/**************** RGB -> YCbCr conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *	Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *	Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + CENTERJSAMPLE
+ *	Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + CENTERJSAMPLE
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ * Note: older versions of the IJG code used a zero offset of MAXJSAMPLE/2,
+ * rather than CENTERJSAMPLE, for Cb and Cr.  This gave equal positive and
+ * negative swings for Cb/Cr, but meant that grayscale values (Cb=Cr=0)
+ * were not represented exactly.  Now we sacrifice exact representation of
+ * maximum red and maximum blue in order to get exact grayscales.
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times R,G,B for all possible values.
+ * For 8-bit JSAMPLEs this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The CENTERJSAMPLE offsets and the rounding fudge-factor of 0.5 are included
+ * in the tables to save adding them separately in the inner loop.
+ */
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define CBCR_OFFSET	((INT32) CENTERJSAMPLE << SCALEBITS)
+#define ONE_HALF	((INT32) 1 << (SCALEBITS-1))
+#define FIX(x)		((INT32) ((x) * (1L<<SCALEBITS) + 0.5))
+
+/* We allocate one big table and divide it up into eight parts, instead of
+ * doing eight alloc_small requests.  This lets us use a single table base
+ * address, which can be held in a register in the inner loops on many
+ * machines (more than can hold all eight addresses, anyway).
+ */
+
+#define R_Y_OFF		0			/* offset to R => Y section */
+#define G_Y_OFF		(1*(MAXJSAMPLE+1))	/* offset to G => Y section */
+#define B_Y_OFF		(2*(MAXJSAMPLE+1))	/* etc. */
+#define R_CB_OFF	(3*(MAXJSAMPLE+1))
+#define G_CB_OFF	(4*(MAXJSAMPLE+1))
+#define B_CB_OFF	(5*(MAXJSAMPLE+1))
+#define R_CR_OFF	B_CB_OFF		/* B=>Cb, R=>Cr are the same */
+#define G_CR_OFF	(6*(MAXJSAMPLE+1))
+#define B_CR_OFF	(7*(MAXJSAMPLE+1))
+#define TABLE_SIZE	(8*(MAXJSAMPLE+1))
+
+
+#if BITS_IN_JSAMPLE == 8
+
+const unsigned char red_lut[256] = {
+  0 , 0 , 1 , 1 , 1 , 1 , 2 , 2 , 2 , 3 , 3 , 3 , 4 , 4 , 4 , 4 ,
+  5 , 5 , 5 , 6 , 6 , 6 , 7 , 7 , 7 , 7 , 8 , 8 , 8 , 9 , 9 , 9 ,
+  10, 10, 10, 10, 11, 11, 11, 12, 12, 12, 13, 13, 13, 13, 14, 14,
+  14, 15, 15, 15, 16, 16, 16, 16, 17, 17, 17, 18, 18, 18, 19, 19,
+  19, 19, 20, 20, 20, 21, 21, 21, 22, 22, 22, 22, 23, 23, 23, 24,
+  24, 24, 25, 25, 25, 25, 26, 26, 26, 27, 27, 27, 28, 28, 28, 28,
+  29, 29, 29, 30, 30, 30, 30, 31, 31, 31, 32, 32, 32, 33, 33, 33,
+  33, 34, 34, 34, 35, 35, 35, 36, 36, 36, 36, 37, 37, 37, 38, 38,
+  38, 39, 39, 39, 39, 40, 40, 40, 41, 41, 41, 42, 42, 42, 42, 43,
+  43, 43, 44, 44, 44, 45, 45, 45, 45, 46, 46, 46, 47, 47, 47, 48,
+  48, 48, 48, 49, 49, 49, 50, 50, 50, 51, 51, 51, 51, 52, 52, 52,
+  53, 53, 53, 54, 54, 54, 54, 55, 55, 55, 56, 56, 56, 57, 57, 57,
+  57, 58, 58, 58, 59, 59, 59, 60, 60, 60, 60, 61, 61, 61, 62, 62,
+  62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 67,
+  67, 67, 68, 68, 68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 71,
+  72, 72, 72, 73, 73, 73, 74, 74, 74, 74, 75, 75, 75, 76, 76, 76
+};
+
+const unsigned char green_lut[256] = {
+  0  , 1  , 1  , 2  , 2  , 3  , 4  , 4  , 5  , 5  , 6  , 6  ,
+  7  , 8  , 8  , 9  , 9  , 10 , 11 , 11 , 12 , 12 , 13 , 14 ,
+  14 , 15 , 15 , 16 , 16 , 17 , 18 , 18 , 19 , 19 , 20 , 21 ,
+  21 , 22 , 22 , 23 , 23 , 24 , 25 , 25 , 26 , 26 , 27 , 28 ,
+  28 , 29 , 29 , 30 , 31 , 31 , 32 , 32 , 33 , 33 , 34 , 35 ,
+  35 , 36 , 36 , 37 , 38 , 38 , 39 , 39 , 40 , 41 , 41 , 42 ,
+  42 , 43 , 43 , 44 , 45 , 45 , 46 , 46 , 47 , 48 , 48 , 49 ,
+  49 , 50 , 50 , 51 , 52 , 52 , 53 , 53 , 54 , 55 , 55 , 56 ,
+  56 , 57 , 58 , 58 , 59 , 59 , 60 , 60 , 61 , 62 , 62 , 63 ,
+  63 , 64 , 65 , 65 , 66 , 66 , 67 , 68 , 68 , 69 , 69 , 70 ,
+  70 , 71 , 72 , 72 , 73 , 73 , 74 , 75 , 75 , 76 , 76 , 77 ,
+  77 , 78 , 79 , 79 , 80 , 80 , 81 , 82 , 82 , 83 , 83 , 84 ,
+  85 , 85 , 86 , 86 , 87 , 87 , 88 , 89 , 89 , 90 , 90 , 91 ,
+  92 , 92 , 93 , 93 , 94 , 95 , 95 , 96 , 96 , 97 , 97 , 98 ,
+  99 , 99 , 100, 100, 101, 102, 102, 103, 103, 104, 104, 105,
+  106, 106, 107, 107, 108, 109, 109, 110, 110, 111, 112, 112,
+  113, 113, 114, 114, 115, 116, 116, 117, 117, 118, 119, 119,
+  120, 120, 121, 122, 122, 123, 123, 124, 124, 125, 126, 126,
+  127, 127, 128, 129, 129, 130, 130, 131, 131, 132, 133, 133,
+  34, 134, 135, 136, 136, 137, 137, 138, 139, 139, 140, 140,
+  141, 141, 142, 143, 143, 144, 144, 145, 146, 146, 147, 147,
+  148, 149, 149, 150
+};
+
+const unsigned char blue_lut[256] = {
+  0 , 0 , 0 , 0 , 0 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 2 ,
+  2 , 2 , 2 , 2 , 2 , 2 , 3 , 3 , 3 , 3 , 3 , 3 , 3 , 3 , 3 , 4 ,
+  4 , 4 , 4 , 4 , 4 , 4 , 4 , 4 , 5 , 5 , 5 , 5 , 5 , 5 , 5 , 5 ,
+  5 , 6 , 6 , 6 , 6 , 6 , 6 , 6 , 6 , 6 , 7 , 7 , 7 , 7 , 7 , 7 ,
+  7 , 7 , 8 , 8 , 8 , 8 , 8 , 8 , 8 , 8 , 8 , 9 , 9 , 9 , 9 , 9 ,
+  9 , 9 , 9 , 9 , 10, 10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11,
+  11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 12, 13, 13,
+  13, 13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+  15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16,
+  16, 17, 17, 17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18, 18, 18,
+  18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19, 19, 20, 20, 20, 20,
+  20, 20, 20, 20, 21, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22, 22,
+  22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23, 23, 24,
+  24, 24, 24, 24, 24, 24, 24, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+  26, 26, 26, 26, 26, 26, 26, 26, 26, 27, 27, 27, 27, 27, 27, 27,
+  27, 27, 28, 28, 28, 28, 28, 28, 28, 28, 29, 29, 29, 29, 29, 29
+};
+
+#endif
+
+
+/*
+ * Initialize for RGB->YCC colorspace conversion.
+ */
+
+METHODDEF(void)
+rgb_ycc_start (j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  INT32 * rgb_ycc_tab;
+  INT32 i;
+
+  /* Allocate and fill in the conversion tables. */
+  cconvert->rgb_ycc_tab = rgb_ycc_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(TABLE_SIZE * SIZEOF(INT32)));
+
+  for (i = 0; i <= MAXJSAMPLE; i++) {
+    rgb_ycc_tab[i+R_Y_OFF] = FIX(0.29900) * i;
+    rgb_ycc_tab[i+G_Y_OFF] = FIX(0.58700) * i;
+    rgb_ycc_tab[i+B_Y_OFF] = FIX(0.11400) * i     + ONE_HALF;
+    rgb_ycc_tab[i+R_CB_OFF] = (-FIX(0.16874)) * i;
+    rgb_ycc_tab[i+G_CB_OFF] = (-FIX(0.33126)) * i;
+    /* We use a rounding fudge-factor of 0.5-epsilon for Cb and Cr.
+     * This ensures that the maximum output will round to MAXJSAMPLE
+     * not MAXJSAMPLE+1, and thus that we don't have to range-limit.
+     */
+    rgb_ycc_tab[i+B_CB_OFF] = FIX(0.50000) * i    + CBCR_OFFSET + ONE_HALF-1;
+/*  B=>Cb and R=>Cr tables are the same
+    rgb_ycc_tab[i+R_CR_OFF] = FIX(0.50000) * i    + CBCR_OFFSET + ONE_HALF-1;
+*/
+    rgb_ycc_tab[i+G_CR_OFF] = (-FIX(0.41869)) * i;
+    rgb_ycc_tab[i+B_CR_OFF] = (-FIX(0.08131)) * i;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ *
+ * Note that we change from the application's interleaved-pixel format
+ * to our internal noninterleaved, one-plane-per-component format.
+ * The input buffer is therefore three times as wide as the output buffer.
+ *
+ * A starting row offset is provided only for the output buffer.  The caller
+ * can easily adjust the passed input_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+METHODDEF(void)
+rgb_ycc_convert (j_compress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		 JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int r, g, b;
+  register INT32 * ctab = cconvert->rgb_ycc_tab;
+  register JSAMPROW inptr;
+  register JSAMPROW outptr0, outptr1, outptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = GETJSAMPLE(inptr[rgb_red[cinfo->in_color_space]]);
+      g = GETJSAMPLE(inptr[rgb_green[cinfo->in_color_space]]);
+      b = GETJSAMPLE(inptr[rgb_blue[cinfo->in_color_space]]);
+      inptr += rgb_pixelsize[cinfo->in_color_space];
+      /* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (JSAMPLE)
+		((ctab[r+R_CB_OFF] + ctab[g+G_CB_OFF] + ctab[b+B_CB_OFF])
+		 >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (JSAMPLE)
+		((ctab[r+R_CR_OFF] + ctab[g+G_CR_OFF] + ctab[b+B_CR_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/**************** Cases other than RGB -> YCbCr **************/
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles RGB->grayscale conversion, which is the same
+ * as the RGB->Y portion of RGB->YCbCr.
+ * We assume rgb_ycc_start has been called (we only use the Y tables).
+ */
+
+METHODDEF(void)
+rgb_gray_convert (j_compress_ptr cinfo,
+		  JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		  JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  #if BITS_IN_JSAMPLE != 8
+  register INT32 * ctab = cconvert->rgb_ycc_tab;
+  #endif
+  register JSAMPROW inptr;
+  register JSAMPROW outptr;
+  JSAMPLE *maxoutptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+  int rindex = rgb_red[cinfo->in_color_space];
+  int gindex = rgb_green[cinfo->in_color_space];
+  int bindex = rgb_blue[cinfo->in_color_space];
+  int rgbstride = rgb_pixelsize[cinfo->in_color_space];
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    maxoutptr = &outptr[num_cols];
+    output_row++;
+    for (; outptr < maxoutptr; outptr++, inptr += rgbstride) {
+      /* Y */
+      #if BITS_IN_JSAMPLE == 8
+      *outptr = red_lut[inptr[rindex]] + green_lut[inptr[gindex]]
+	    + blue_lut[inptr[bindex]];
+      #else
+      *outptr = (JSAMPLE)
+	    ((ctab[GETJSAMPLE(inptr[rindex])+R_Y_OFF]
+	     + ctab[GETJSAMPLE(inptr[gindex])+G_Y_OFF]
+	     + ctab[GETJSAMPLE(inptr[bindex])+B_Y_OFF])
+	     >> SCALEBITS);
+      #endif
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles Adobe-style CMYK->YCCK conversion,
+ * where we convert R=1-C, G=1-M, and B=1-Y to YCbCr using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume rgb_ycc_start has been called.
+ */
+
+METHODDEF(void)
+cmyk_ycck_convert (j_compress_ptr cinfo,
+		   JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		   JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int r, g, b;
+  register INT32 * ctab = cconvert->rgb_ycc_tab;
+  register JSAMPROW inptr;
+  register JSAMPROW outptr0, outptr1, outptr2, outptr3;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    outptr3 = output_buf[3][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = MAXJSAMPLE - GETJSAMPLE(inptr[0]);
+      g = MAXJSAMPLE - GETJSAMPLE(inptr[1]);
+      b = MAXJSAMPLE - GETJSAMPLE(inptr[2]);
+      /* K passes through as-is */
+      outptr3[col] = inptr[3];	/* don't need GETJSAMPLE here */
+      inptr += 4;
+      /* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (JSAMPLE)
+		((ctab[r+R_CB_OFF] + ctab[g+G_CB_OFF] + ctab[b+B_CB_OFF])
+		 >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (JSAMPLE)
+		((ctab[r+R_CR_OFF] + ctab[g+G_CR_OFF] + ctab[b+B_CR_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles grayscale output with no conversion.
+ * The source can be either plain grayscale or YCbCr (since Y == gray).
+ */
+
+METHODDEF(void)
+grayscale_convert (j_compress_ptr cinfo,
+		   JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		   JDIMENSION output_row, int num_rows)
+{
+  register JSAMPROW inptr;
+  register JSAMPROW outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+  int instride = cinfo->input_components;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      outptr[col] = inptr[0];	/* don't need GETJSAMPLE() here */
+      inptr += instride;
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles multi-component colorspaces without conversion.
+ * We assume input_components == num_components.
+ */
+
+METHODDEF(void)
+null_convert (j_compress_ptr cinfo,
+	      JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+	      JDIMENSION output_row, int num_rows)
+{
+  register JSAMPROW inptr;
+  register JSAMPROW outptr;
+  register JDIMENSION col;
+  register int ci;
+  int nc = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    /* It seems fastest to make a separate pass for each component. */
+    for (ci = 0; ci < nc; ci++) {
+      inptr = *input_buf;
+      outptr = output_buf[ci][output_row];
+      for (col = 0; col < num_cols; col++) {
+	outptr[col] = inptr[ci]; /* don't need GETJSAMPLE() here */
+	inptr += nc;
+      }
+    }
+    input_buf++;
+    output_row++;
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+null_method (j_compress_ptr cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for input colorspace conversion.
+ */
+
+GLOBAL(void)
+jinit_color_converter (j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_color_converter));
+  cinfo->cconvert = (struct jpeg_color_converter *) cconvert;
+  /* set start_pass to null method until we find out differently */
+  cconvert->pub.start_pass = null_method;
+
+  /* Make sure input_components agrees with in_color_space */
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->input_components != 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+    if (cinfo->input_components != rgb_pixelsize[cinfo->in_color_space])
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_YCbCr:
+    if (cinfo->input_components != 3)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->input_components != 4)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  default:			/* JCS_UNKNOWN can be anything */
+    if (cinfo->input_components < 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+  }
+
+  /* Check num_components, set conversion method based on requested space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_GRAYSCALE)
+      cconvert->pub.color_convert = grayscale_convert;
+    else if (cinfo->in_color_space == JCS_RGB ||
+             cinfo->in_color_space == JCS_EXT_RGB ||
+             cinfo->in_color_space == JCS_EXT_RGBX ||
+             cinfo->in_color_space == JCS_EXT_BGR ||
+             cinfo->in_color_space == JCS_EXT_BGRX ||
+             cinfo->in_color_space == JCS_EXT_XBGR ||
+             cinfo->in_color_space == JCS_EXT_XRGB) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = rgb_gray_convert;
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub.color_convert = grayscale_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == cinfo->jpeg_color_space &&
+      rgb_pixelsize[cinfo->in_color_space] == 3)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_RGB ||
+        cinfo->in_color_space == JCS_EXT_RGB ||
+        cinfo->in_color_space == JCS_EXT_RGBX ||
+        cinfo->in_color_space == JCS_EXT_BGR ||
+        cinfo->in_color_space == JCS_EXT_BGRX ||
+        cinfo->in_color_space == JCS_EXT_XBGR ||
+        cinfo->in_color_space == JCS_EXT_XRGB) {
+      if (jsimd_can_rgb_ycc())
+        cconvert->pub.color_convert = jsimd_rgb_ycc_convert;
+      else {
+        cconvert->pub.start_pass = rgb_ycc_start;
+        cconvert->pub.color_convert = rgb_ycc_convert;
+      }
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = cmyk_ycck_convert;
+    } else if (cinfo->in_color_space == JCS_YCCK)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:			/* allow null conversion of JCS_UNKNOWN */
+    if (cinfo->jpeg_color_space != cinfo->in_color_space ||
+	cinfo->num_components != cinfo->input_components)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    cconvert->pub.color_convert = null_convert;
+    break;
+  }
+}
diff --git a/jcdctmgr.c b/jcdctmgr.c
new file mode 100644
index 0000000..156957a
--- /dev/null
+++ b/jcdctmgr.c
@@ -0,0 +1,632 @@
+/*
+ * jcdctmgr.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the forward-DCT management logic.
+ * This code selects a particular DCT implementation to be used,
+ * and it performs related housekeeping chores including coefficient
+ * quantization.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+#include "jsimddct.h"
+
+
+/* Private subobject for this module */
+
+typedef JMETHOD(void, forward_DCT_method_ptr, (DCTELEM * data));
+typedef JMETHOD(void, float_DCT_method_ptr, (FAST_FLOAT * data));
+
+typedef JMETHOD(void, convsamp_method_ptr,
+                (JSAMPARRAY sample_data, JDIMENSION start_col,
+                 DCTELEM * workspace));
+typedef JMETHOD(void, float_convsamp_method_ptr,
+                (JSAMPARRAY sample_data, JDIMENSION start_col,
+                 FAST_FLOAT *workspace));
+
+typedef JMETHOD(void, quantize_method_ptr,
+                (JCOEFPTR coef_block, DCTELEM * divisors,
+                 DCTELEM * workspace));
+typedef JMETHOD(void, float_quantize_method_ptr,
+                (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+                 FAST_FLOAT * workspace));
+
+typedef struct {
+  struct jpeg_forward_dct pub;	/* public fields */
+
+  /* Pointer to the DCT routine actually in use */
+  forward_DCT_method_ptr dct;
+  convsamp_method_ptr convsamp;
+  quantize_method_ptr quantize;
+
+  /* The actual post-DCT divisors --- not identical to the quant table
+   * entries, because of scaling (especially for an unnormalized DCT).
+   * Each table is given in normal array order.
+   */
+  DCTELEM * divisors[NUM_QUANT_TBLS];
+
+  /* work area for FDCT subroutine */
+  DCTELEM * workspace;
+
+#ifdef DCT_FLOAT_SUPPORTED
+  /* Same as above for the floating-point case. */
+  float_DCT_method_ptr float_dct;
+  float_convsamp_method_ptr float_convsamp;
+  float_quantize_method_ptr float_quantize;
+  FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
+  FAST_FLOAT * float_workspace;
+#endif
+} my_fdct_controller;
+
+typedef my_fdct_controller * my_fdct_ptr;
+
+
+/*
+ * Find the highest bit in an integer through binary search.
+ */
+LOCAL(int)
+flss (UINT16 val)
+{
+  int bit;
+
+  bit = 16;
+
+  if (!val)
+    return 0;
+
+  if (!(val & 0xff00)) {
+    bit -= 8;
+    val <<= 8;
+  }
+  if (!(val & 0xf000)) {
+    bit -= 4;
+    val <<= 4;
+  }
+  if (!(val & 0xc000)) {
+    bit -= 2;
+    val <<= 2;
+  }
+  if (!(val & 0x8000)) {
+    bit -= 1;
+    val <<= 1;
+  }
+
+  return bit;
+}
+
+/*
+ * Compute values to do a division using reciprocal.
+ *
+ * This implementation is based on an algorithm described in
+ *   "How to optimize for the Pentium family of microprocessors"
+ *   (http://www.agner.org/assem/).
+ * More information about the basic algorithm can be found in
+ * the paper "Integer Division Using Reciprocals" by Robert Alverson.
+ *
+ * The basic idea is to replace x/d by x * d^-1. In order to store
+ * d^-1 with enough precision we shift it left a few places. It turns
+ * out that this algoright gives just enough precision, and also fits
+ * into DCTELEM:
+ *
+ *   b = (the number of significant bits in divisor) - 1
+ *   r = (word size) + b
+ *   f = 2^r / divisor
+ *
+ * f will not be an integer for most cases, so we need to compensate
+ * for the rounding error introduced:
+ *
+ *   no fractional part:
+ *
+ *       result = input >> r
+ *
+ *   fractional part of f < 0.5:
+ *
+ *       round f down to nearest integer
+ *       result = ((input + 1) * f) >> r
+ *
+ *   fractional part of f > 0.5:
+ *
+ *       round f up to nearest integer
+ *       result = (input * f) >> r
+ *
+ * This is the original algorithm that gives truncated results. But we
+ * want properly rounded results, so we replace "input" with
+ * "input + divisor/2".
+ *
+ * In order to allow SIMD implementations we also tweak the values to
+ * allow the same calculation to be made at all times:
+ * 
+ *   dctbl[0] = f rounded to nearest integer
+ *   dctbl[1] = divisor / 2 (+ 1 if fractional part of f < 0.5)
+ *   dctbl[2] = 1 << ((word size) * 2 - r)
+ *   dctbl[3] = r - (word size)
+ *
+ * dctbl[2] is for stupid instruction sets where the shift operation
+ * isn't member wise (e.g. MMX).
+ *
+ * The reason dctbl[2] and dctbl[3] reduce the shift with (word size)
+ * is that most SIMD implementations have a "multiply and store top
+ * half" operation.
+ *
+ * Lastly, we store each of the values in their own table instead
+ * of in a consecutive manner, yet again in order to allow SIMD
+ * routines.
+ */
+LOCAL(void)
+compute_reciprocal (UINT16 divisor, DCTELEM * dtbl)
+{
+  UDCTELEM2 fq, fr;
+  UDCTELEM c;
+  int b, r;
+
+  b = flss(divisor) - 1;
+  r  = sizeof(DCTELEM) * 8 + b;
+
+  fq = ((UDCTELEM2)1 << r) / divisor;
+  fr = ((UDCTELEM2)1 << r) % divisor;
+
+  c = divisor / 2; /* for rounding */
+
+  if (fr == 0) { /* divisor is power of two */
+    /* fq will be one bit too large to fit in DCTELEM, so adjust */
+    fq >>= 1;
+    r--;
+  } else if (fr <= (divisor / 2)) { /* fractional part is < 0.5 */
+    c++;
+  } else { /* fractional part is > 0.5 */
+    fq++;
+  }
+
+  dtbl[DCTSIZE2 * 0] = (DCTELEM) fq;      /* reciprocal */
+  dtbl[DCTSIZE2 * 1] = (DCTELEM) c;       /* correction + roundfactor */
+  dtbl[DCTSIZE2 * 2] = (DCTELEM) (1 << (sizeof(DCTELEM)*8*2 - r));  /* scale */
+  dtbl[DCTSIZE2 * 3] = (DCTELEM) r - sizeof(DCTELEM)*8; /* shift */
+}
+
+/*
+ * Initialize for a processing pass.
+ * Verify that all referenced Q-tables are present, and set up
+ * the divisor table for each one.
+ * In the current implementation, DCT of all components is done during
+ * the first pass, even if only some components will be output in the
+ * first scan.  Hence all components should be examined here.
+ */
+
+METHODDEF(void)
+start_pass_fdctmgr (j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  int ci, qtblno, i;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtbl;
+  DCTELEM * dtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    qtblno = compptr->quant_tbl_no;
+    /* Make sure specified quantization table is present */
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+	cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    qtbl = cinfo->quant_tbl_ptrs[qtblno];
+    /* Compute divisors for this quant table */
+    /* We may do this more than once for same table, but it's not a big deal */
+    switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+    case JDCT_ISLOW:
+      /* For LL&M IDCT method, divisors are equal to raw quantization
+       * coefficients multiplied by 8 (to counteract scaling).
+       */
+      if (fdct->divisors[qtblno] == NULL) {
+	fdct->divisors[qtblno] = (DCTELEM *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      (DCTSIZE2 * 4) * SIZEOF(DCTELEM));
+      }
+      dtbl = fdct->divisors[qtblno];
+      for (i = 0; i < DCTSIZE2; i++) {
+	compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i]);
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+	/* For AA&N IDCT method, divisors are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 8.
+	 */
+#define CONST_BITS 14
+	static const INT16 aanscales[DCTSIZE2] = {
+	  /* precomputed values scaled up by 14 bits */
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+	};
+	SHIFT_TEMPS
+
+	if (fdct->divisors[qtblno] == NULL) {
+	  fdct->divisors[qtblno] = (DCTELEM *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					(DCTSIZE2 * 4) * SIZEOF(DCTELEM));
+	}
+	dtbl = fdct->divisors[qtblno];
+	for (i = 0; i < DCTSIZE2; i++) {
+	  compute_reciprocal(
+	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
+				  (INT32) aanscales[i]),
+		    CONST_BITS-3), &dtbl[i]);
+	}
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+	/* For float AA&N IDCT method, divisors are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 8.
+	 * What's actually stored is 1/divisor so that the inner loop can
+	 * use a multiplication rather than a division.
+	 */
+	FAST_FLOAT * fdtbl;
+	int row, col;
+	static const double aanscalefactor[DCTSIZE] = {
+	  1.0, 1.387039845, 1.306562965, 1.175875602,
+	  1.0, 0.785694958, 0.541196100, 0.275899379
+	};
+
+	if (fdct->float_divisors[qtblno] == NULL) {
+	  fdct->float_divisors[qtblno] = (FAST_FLOAT *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					DCTSIZE2 * SIZEOF(FAST_FLOAT));
+	}
+	fdtbl = fdct->float_divisors[qtblno];
+	i = 0;
+	for (row = 0; row < DCTSIZE; row++) {
+	  for (col = 0; col < DCTSIZE; col++) {
+	    fdtbl[i] = (FAST_FLOAT)
+	      (1.0 / (((double) qtbl->quantval[i] *
+		       aanscalefactor[row] * aanscalefactor[col] * 8.0)));
+	    i++;
+	  }
+	}
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Load data into workspace, applying unsigned->signed conversion.
+ */
+
+METHODDEF(void)
+convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, DCTELEM * workspace)
+{
+  register DCTELEM *workspaceptr;
+  register JSAMPROW elemptr;
+  register int elemr;
+
+  workspaceptr = workspace;
+  for (elemr = 0; elemr < DCTSIZE; elemr++) {
+    elemptr = sample_data[elemr] + start_col;
+
+#if DCTSIZE == 8		/* unroll the inner loop */
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+#else
+    {
+      register int elemc;
+      for (elemc = DCTSIZE; elemc > 0; elemc--)
+        *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+    }
+#endif
+  }
+}
+
+
+/*
+ * Quantize/descale the coefficients, and store into coef_blocks[].
+ */
+
+METHODDEF(void)
+quantize (JCOEFPTR coef_block, DCTELEM * divisors, DCTELEM * workspace)
+{
+  int i;
+  DCTELEM temp;
+  UDCTELEM recip, corr, shift;
+  UDCTELEM2 product;
+  JCOEFPTR output_ptr = coef_block;
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    temp = workspace[i];
+    recip = divisors[i + DCTSIZE2 * 0];
+    corr =  divisors[i + DCTSIZE2 * 1];
+    shift = divisors[i + DCTSIZE2 * 3];
+
+    if (temp < 0) {
+      temp = -temp;
+      product = (UDCTELEM2)(temp + corr) * recip;
+      product >>= shift + sizeof(DCTELEM)*8;
+      temp = product;
+      temp = -temp;
+    } else {
+      product = (UDCTELEM2)(temp + corr) * recip;
+      product >>= shift + sizeof(DCTELEM)*8;
+      temp = product;
+    }
+
+    output_ptr[i] = (JCOEF) temp;
+  }
+}
+
+
+/*
+ * Perform forward DCT on one or more blocks of a component.
+ *
+ * The input samples are taken from the sample_data[] array starting at
+ * position start_row/start_col, and moving to the right for any additional
+ * blocks. The quantized coefficients are returned in coef_blocks[].
+ */
+
+METHODDEF(void)
+forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
+	     JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+	     JDIMENSION start_row, JDIMENSION start_col,
+	     JDIMENSION num_blocks)
+/* This version is used for integer DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
+  DCTELEM * workspace;
+  JDIMENSION bi;
+
+  /* Make sure the compiler doesn't look up these every pass */
+  forward_DCT_method_ptr do_dct = fdct->dct;
+  convsamp_method_ptr do_convsamp = fdct->convsamp;
+  quantize_method_ptr do_quantize = fdct->quantize;
+  workspace = fdct->workspace;
+
+  sample_data += start_row;	/* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
+    /* Load data into workspace, applying unsigned->signed conversion */
+    (*do_convsamp) (sample_data, start_col, workspace);
+
+    /* Perform the DCT */
+    (*do_dct) (workspace);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    (*do_quantize) (coef_blocks[bi], divisors, workspace);
+  }
+}
+
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+METHODDEF(void)
+convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, FAST_FLOAT * workspace)
+{
+  register FAST_FLOAT *workspaceptr;
+  register JSAMPROW elemptr;
+  register int elemr;
+
+  workspaceptr = workspace;
+  for (elemr = 0; elemr < DCTSIZE; elemr++) {
+    elemptr = sample_data[elemr] + start_col;
+#if DCTSIZE == 8		/* unroll the inner loop */
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+#else
+    {
+      register int elemc;
+      for (elemc = DCTSIZE; elemc > 0; elemc--)
+        *workspaceptr++ = (FAST_FLOAT)
+                          (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+    }
+#endif
+  }
+}
+
+
+METHODDEF(void)
+quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors, FAST_FLOAT * workspace)
+{
+  register FAST_FLOAT temp;
+  register int i;
+  register JCOEFPTR output_ptr = coef_block;
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    /* Apply the quantization and scaling factor */
+    temp = workspace[i] * divisors[i];
+
+    /* Round to nearest integer.
+     * Since C does not specify the direction of rounding for negative
+     * quotients, we have to force the dividend positive for portability.
+     * The maximum coefficient size is +-16K (for 12-bit data), so this
+     * code should work for either 16-bit or 32-bit ints.
+     */
+    output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
+  }
+}
+
+
+METHODDEF(void)
+forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+		   JDIMENSION start_row, JDIMENSION start_col,
+		   JDIMENSION num_blocks)
+/* This version is used for floating-point DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
+  FAST_FLOAT * workspace;
+  JDIMENSION bi;
+
+
+  /* Make sure the compiler doesn't look up these every pass */
+  float_DCT_method_ptr do_dct = fdct->float_dct;
+  float_convsamp_method_ptr do_convsamp = fdct->float_convsamp;
+  float_quantize_method_ptr do_quantize = fdct->float_quantize;
+  workspace = fdct->float_workspace;
+
+  sample_data += start_row;	/* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
+    /* Load data into workspace, applying unsigned->signed conversion */
+    (*do_convsamp) (sample_data, start_col, workspace);
+
+    /* Perform the DCT */
+    (*do_dct) (workspace);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    (*do_quantize) (coef_blocks[bi], divisors, workspace);
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
+
+
+/*
+ * Initialize FDCT manager.
+ */
+
+GLOBAL(void)
+jinit_forward_dct (j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct;
+  int i;
+
+  fdct = (my_fdct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_fdct_controller));
+  cinfo->fdct = (struct jpeg_forward_dct *) fdct;
+  fdct->pub.start_pass = start_pass_fdctmgr;
+
+  /* First determine the DCT... */
+  switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+  case JDCT_ISLOW:
+    fdct->pub.forward_DCT = forward_DCT;
+    if (jsimd_can_fdct_islow())
+      fdct->dct = jsimd_fdct_islow;
+    else
+      fdct->dct = jpeg_fdct_islow;
+    break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  case JDCT_IFAST:
+    fdct->pub.forward_DCT = forward_DCT;
+    if (jsimd_can_fdct_ifast())
+      fdct->dct = jsimd_fdct_ifast;
+    else
+      fdct->dct = jpeg_fdct_ifast;
+    break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  case JDCT_FLOAT:
+    fdct->pub.forward_DCT = forward_DCT_float;
+    if (jsimd_can_fdct_float())
+      fdct->float_dct = jsimd_fdct_float;
+    else
+      fdct->float_dct = jpeg_fdct_float;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+
+  /* ...then the supporting stages. */
+  switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+  case JDCT_ISLOW:
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  case JDCT_IFAST:
+#endif
+#if defined(DCT_ISLOW_SUPPORTED) || defined(DCT_IFAST_SUPPORTED)
+    if (jsimd_can_convsamp())
+      fdct->convsamp = jsimd_convsamp;
+    else
+      fdct->convsamp = convsamp;
+    if (jsimd_can_quantize())
+      fdct->quantize = jsimd_quantize;
+    else
+      fdct->quantize = quantize;
+    break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  case JDCT_FLOAT:
+    if (jsimd_can_convsamp_float())
+      fdct->float_convsamp = jsimd_convsamp_float;
+    else
+      fdct->float_convsamp = convsamp_float;
+    if (jsimd_can_quantize_float())
+      fdct->float_quantize = jsimd_quantize_float;
+    else
+      fdct->float_quantize = quantize_float;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+
+  /* Allocate workspace memory */
+#ifdef DCT_FLOAT_SUPPORTED
+  if (cinfo->dct_method == JDCT_FLOAT)
+    fdct->float_workspace = (FAST_FLOAT *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(FAST_FLOAT) * DCTSIZE2);
+  else
+#endif
+    fdct->workspace = (DCTELEM *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Mark divisor tables unallocated */
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    fdct->divisors[i] = NULL;
+#ifdef DCT_FLOAT_SUPPORTED
+    fdct->float_divisors[i] = NULL;
+#endif
+  }
+}
diff --git a/jchuff.c b/jchuff.c
new file mode 100644
index 0000000..b05c8e7
--- /dev/null
+++ b/jchuff.c
@@ -0,0 +1,1046 @@
+/*
+ * jchuff.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy encoding routines.
+ *
+ * Much of the complexity here has to do with supporting output suspension.
+ * If the data destination module demands suspension, we want to be able to
+ * back up to the start of the current MCU.  To do this, we copy state
+ * variables into local working storage, and update them back to the
+ * permanent JPEG objects only upon successful completion of an MCU.
+ */
+
+/* Modifications:
+ * Copyright (C)2007 Sun Microsystems, Inc.
+ * Copyright (C)2009 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jchuff.h"		/* Declarations shared with jcphuff.c */
+#include <limits.h>
+
+static unsigned char jpeg_first_bit_table[65536];
+int jpeg_first_bit_table_init=0;
+
+#ifndef min
+ #define min(a,b) ((a)<(b)?(a):(b))
+#endif
+
+/* Expanded entropy encoder object for Huffman encoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  size_t put_buffer;		/* current bit-accumulation buffer */
+  int put_bits;			/* # of bits now in it */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).put_buffer = (src).put_buffer, \
+	 (dest).put_bits = (src).put_bits, \
+	 (dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  savable_state saved;		/* Bit buffer & DC state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
+  c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
+
+#ifdef ENTROPY_OPT_SUPPORTED	/* Statistics tables for optimization */
+  long * dc_count_ptrs[NUM_HUFF_TBLS];
+  long * ac_count_ptrs[NUM_HUFF_TBLS];
+#endif
+} huff_entropy_encoder;
+
+typedef huff_entropy_encoder * huff_entropy_ptr;
+
+/* Working state while writing an MCU.
+ * This struct contains all the fields that are needed by subroutines.
+ */
+
+typedef struct {
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+  savable_state cur;		/* Current bit buffer & DC state */
+  j_compress_ptr cinfo;		/* dump_buffer needs access to this */
+} working_state;
+
+
+/* Forward declarations */
+METHODDEF(boolean) encode_mcu_huff JPP((j_compress_ptr cinfo,
+					JBLOCKROW *MCU_data));
+METHODDEF(void) finish_pass_huff JPP((j_compress_ptr cinfo));
+#ifdef ENTROPY_OPT_SUPPORTED
+METHODDEF(boolean) encode_mcu_gather JPP((j_compress_ptr cinfo,
+					  JBLOCKROW *MCU_data));
+METHODDEF(void) finish_pass_gather JPP((j_compress_ptr cinfo));
+#endif
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ * If gather_statistics is TRUE, we do not output anything during the scan,
+ * just count the Huffman symbols used and generate Huffman code tables.
+ */
+
+METHODDEF(void)
+start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, dctbl, actbl;
+  jpeg_component_info * compptr;
+
+  if (gather_statistics) {
+#ifdef ENTROPY_OPT_SUPPORTED
+    entropy->pub.encode_mcu = encode_mcu_gather;
+    entropy->pub.finish_pass = finish_pass_gather;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    entropy->pub.encode_mcu = encode_mcu_huff;
+    entropy->pub.finish_pass = finish_pass_huff;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    if (gather_statistics) {
+#ifdef ENTROPY_OPT_SUPPORTED
+      /* Check for invalid table indexes */
+      /* (make_c_derived_tbl does this in the other path) */
+      if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
+	ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
+      if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
+	ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
+      /* Allocate and zero the statistics tables */
+      /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+      if (entropy->dc_count_ptrs[dctbl] == NULL)
+	entropy->dc_count_ptrs[dctbl] = (long *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      257 * SIZEOF(long));
+      MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * SIZEOF(long));
+      if (entropy->ac_count_ptrs[actbl] == NULL)
+	entropy->ac_count_ptrs[actbl] = (long *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      257 * SIZEOF(long));
+      MEMZERO(entropy->ac_count_ptrs[actbl], 257 * SIZEOF(long));
+#endif
+    } else {
+      /* Compute derived values for Huffman tables */
+      /* We may do this more than once for a table, but it's not expensive */
+      jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
+			      & entropy->dc_derived_tbls[dctbl]);
+      jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
+			      & entropy->ac_derived_tbls[actbl]);
+    }
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Initialize bit buffer to empty */
+
+  entropy->saved.put_buffer = 0;
+  entropy->saved.put_bits = 0;
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ *
+ * Note this is also used by jcphuff.c.
+ */
+
+GLOBAL(void)
+jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
+			 c_derived_tbl ** pdtbl)
+{
+  JHUFF_TBL *htbl;
+  c_derived_tbl *dtbl;
+  int p, i, l, lastp, si, maxsymbol;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (c_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(c_derived_tbl));
+  dtbl = *pdtbl;
+  
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int) htbl->bits[l];
+    if (i < 0 || p + i > 256)	/* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char) l;
+  }
+  huffsize[p] = 0;
+  lastp = p;
+  
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int) huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((INT32) code) >= (((INT32) 1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+  
+  /* Figure C.3: generate encoding tables */
+  /* These are code and size indexed by symbol value */
+
+  /* Set all codeless symbols to have code length 0;
+   * this lets us detect duplicate VAL entries here, and later
+   * allows emit_bits to detect any attempt to emit such symbols.
+   */
+  MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
+
+  /* This is also a convenient place to check for out-of-range
+   * and duplicated VAL entries.  We allow 0..255 for AC symbols
+   * but only 0..15 for DC.  (We could constrain them further
+   * based on data depth and mode, but this seems enough.)
+   */
+  maxsymbol = isDC ? 15 : 255;
+
+  for (p = 0; p < lastp; p++) {
+    i = htbl->huffval[p];
+    if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    dtbl->ehufco[i] = huffcode[p];
+    dtbl->ehufsi[i] = huffsize[p];
+  }
+
+  if(!jpeg_first_bit_table_init) {
+    for(i = 0; i < 65536; i++) {
+      int bit = 0, val = i;
+      while (val) {val >>= 1;  bit++;}
+      jpeg_first_bit_table[i] = bit;
+    }
+    jpeg_first_bit_table_init = 1;
+  }
+}
+
+
+/* Outputting bytes to the file */
+
+/* Emit a byte, taking 'action' if must suspend. */
+#define emit_byte(state,val,action)  \
+	{ *(state)->next_output_byte++ = (JOCTET) (val);  \
+	  if (--(state)->free_in_buffer == 0)  \
+	    if (! dump_buffer(state))  \
+	      { action; } }
+
+
+LOCAL(boolean)
+dump_buffer (working_state * state)
+/* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
+{
+  struct jpeg_destination_mgr * dest = state->cinfo->dest;
+
+  dest->free_in_buffer = state->free_in_buffer;
+
+  if (! (*dest->empty_output_buffer) (state->cinfo))
+    return FALSE;
+  /* After a successful buffer dump, must reset buffer pointers */
+  state->next_output_byte = dest->next_output_byte;
+  state->free_in_buffer = dest->free_in_buffer;
+  return TRUE;
+}
+
+
+/* Outputting bits to the file */
+
+/* Only the right 24 bits of put_buffer are used; the valid bits are
+ * left-justified in this part.  At most 16 bits can be passed to emit_bits
+ * in one call, and we never retain more than 7 bits in put_buffer
+ * between calls, so 24 bits are sufficient.
+ */
+
+/***************************************************************/
+
+#define EMIT_BYTE() {                                           \
+  if (0xFF == (*buffer++ =  (unsigned char)(put_buffer >> (put_bits -= 8))))  \
+    *buffer++ = 0;                                              \
+ }
+
+/***************************************************************/
+
+#define DUMP_BITS_(code, size) {                                \
+  put_bits += size;                                             \
+  put_buffer = (put_buffer << size) | code;                     \
+  if (put_bits > 7)                                             \
+    while(put_bits > 7)                                         \
+      EMIT_BYTE()                                               \
+ }
+
+/***************************************************************/
+
+#define CHECKBUF15() {                                          \
+  if (put_bits > 15) {                                          \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+  }                                                             \
+}
+
+#define CHECKBUF47() {                                          \
+  if (put_bits > 47) {                                          \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+  }                                                             \
+}
+
+#define CHECKBUF31() {                                          \
+  if (put_bits > 31) {                                          \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+    EMIT_BYTE()                                                 \
+  }                                                             \
+}
+
+/***************************************************************/
+
+#define DUMP_BITS_NOCHECK(code, size) {                         \
+  put_bits += size;                                             \
+  put_buffer = (put_buffer << size) | code;                     \
+ }
+
+#if __WORDSIZE==64 || defined(_WIN64)
+
+#define DUMP_BITS(code, size) {                                 \
+  CHECKBUF47()                                                  \
+  put_bits += size;                                             \
+  put_buffer = (put_buffer << size) | code;                     \
+ }
+
+#else
+
+#define DUMP_BITS(code, size) {                                 \
+  put_bits += size;                                             \
+  put_buffer = (put_buffer << size) | code;                     \
+  CHECKBUF15()                                                  \
+ }
+
+#endif
+
+/***************************************************************/
+
+#define DUMP_SINGLE_VALUE(ht, codevalue) { \
+  size = ht->ehufsi[codevalue];            \
+  code = ht->ehufco[codevalue];            \
+                                           \
+  DUMP_BITS(code, size)                    \
+ }
+
+/***************************************************************/
+
+#define DUMP_VALUE_SLOW(ht, codevalue, t, nbits) { \
+  size = ht->ehufsi[codevalue];               \
+  code = ht->ehufco[codevalue];               \
+  t &= ~(-1 << nbits);                        \
+  DUMP_BITS_NOCHECK(code, size)               \
+  CHECKBUF15()                                \
+  DUMP_BITS_NOCHECK(t, nbits)                 \
+  CHECKBUF15()                                \
+ }
+
+int _max=0;
+
+#if __WORDSIZE==64 || defined(_WIN64)
+
+#define DUMP_VALUE(ht, codevalue, t, nbits) { \
+  size = ht->ehufsi[codevalue];               \
+  code = ht->ehufco[codevalue];               \
+  t &= ~(-1 << nbits);                        \
+  CHECKBUF31()                                \
+  DUMP_BITS_NOCHECK(code, size)               \
+  DUMP_BITS_NOCHECK(t, nbits)                 \
+ }
+
+#else
+
+#define DUMP_VALUE(ht, codevalue, t, nbits) { \
+  size = ht->ehufsi[codevalue];               \
+  code = ht->ehufco[codevalue];               \
+  t &= ~(-1 << nbits);                        \
+  DUMP_BITS_NOCHECK(code, size)               \
+  CHECKBUF15()                                \
+  DUMP_BITS_NOCHECK(t, nbits)                 \
+  CHECKBUF15()                                \
+ }
+
+#endif
+
+/***************************************************************/
+
+#define BUFSIZE (DCTSIZE2 * 2)
+
+#define LOAD_BUFFER() {                                           \
+  if (state->free_in_buffer < BUFSIZE) {                          \
+    localbuf = 1;                                                 \
+    buffer = _buffer;                                             \
+  }                                                               \
+  else buffer = state->next_output_byte;                          \
+ }
+
+#define STORE_BUFFER() {                                          \
+  if (localbuf) {                                                 \
+    bytes = buffer - _buffer;                                     \
+    buffer = _buffer;                                             \
+    while (bytes > 0) {                                           \
+      bytestocopy = min(bytes, state->free_in_buffer);            \
+      MEMCOPY(state->next_output_byte, buffer, bytestocopy);      \
+      state->next_output_byte += bytestocopy;                     \
+      buffer += bytestocopy;                                      \
+      state->free_in_buffer -= bytestocopy;                       \
+      if (state->free_in_buffer == 0)                             \
+        if (! dump_buffer(state)) return FALSE;                   \
+      bytes -= bytestocopy;                                       \
+    }                                                             \
+  }                                                               \
+  else {                                                          \
+    state->free_in_buffer -= (buffer - state->next_output_byte);  \
+    state->next_output_byte = buffer;                             \
+  }                                                               \
+ }
+
+/***************************************************************/
+
+LOCAL(boolean)
+flush_bits (working_state * state)
+{
+  unsigned char _buffer[BUFSIZE], *buffer;
+  size_t put_buffer;  int put_bits;
+  size_t bytes, bytestocopy;  int localbuf = 0;
+
+  put_buffer = state->cur.put_buffer;
+  put_bits = state->cur.put_bits;
+  LOAD_BUFFER()
+
+  DUMP_BITS_(0x7F, 7)
+
+  state->cur.put_buffer = 0;	/* and reset bit-buffer to empty */
+  state->cur.put_bits = 0;
+  STORE_BUFFER()
+
+  return TRUE;
+}
+
+/* Encode a single block's worth of coefficients */
+
+LOCAL(boolean)
+encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
+		  c_derived_tbl *dctbl, c_derived_tbl *actbl)
+{
+  int temp, temp2;
+  int nbits;
+  int r, sflag, size, code;
+  unsigned char _buffer[BUFSIZE], *buffer;
+  size_t put_buffer;  int put_bits;
+  int code_0xf0 = actbl->ehufco[0xf0], size_0xf0 = actbl->ehufsi[0xf0];
+  size_t bytes, bytestocopy;  int localbuf = 0;
+
+  put_buffer = state->cur.put_buffer;
+  put_bits = state->cur.put_bits;
+  LOAD_BUFFER()
+
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+  
+  temp = temp2 = block[0] - last_dc_val;
+
+  sflag = temp >> 31;
+  temp -= ((temp + temp) & sflag);
+  temp2 += sflag;
+  nbits = jpeg_first_bit_table[temp];
+  DUMP_VALUE_SLOW(dctbl, nbits, temp2, nbits)
+
+  /* Encode the AC coefficients per section F.1.2.2 */
+  
+  r = 0;			/* r = run length of zeros */
+
+#define innerloop(order) {  \
+  temp2  = *(JCOEF*)((unsigned char*)block + order);  \
+  if(temp2 == 0) r++;  \
+  else {  \
+    temp = (JCOEF)temp2;  \
+    sflag = temp >> 31;  \
+    temp = (temp ^ sflag) - sflag;  \
+    temp2 += sflag;  \
+    nbits = jpeg_first_bit_table[temp];  \
+    for(; r > 15; r -= 16) DUMP_BITS(code_0xf0, size_0xf0)  \
+    sflag = (r << 4) + nbits;  \
+    DUMP_VALUE(actbl, sflag, temp2, nbits)  \
+    r = 0;  \
+  }}
+
+  innerloop(2*1);   innerloop(2*8);   innerloop(2*16);  innerloop(2*9);
+  innerloop(2*2);   innerloop(2*3);   innerloop(2*10);  innerloop(2*17);
+  innerloop(2*24);  innerloop(2*32);  innerloop(2*25);  innerloop(2*18);
+  innerloop(2*11);  innerloop(2*4);   innerloop(2*5);   innerloop(2*12);
+  innerloop(2*19);  innerloop(2*26);  innerloop(2*33);  innerloop(2*40);
+  innerloop(2*48);  innerloop(2*41);  innerloop(2*34);  innerloop(2*27);
+  innerloop(2*20);  innerloop(2*13);  innerloop(2*6);   innerloop(2*7);
+  innerloop(2*14);  innerloop(2*21);  innerloop(2*28);  innerloop(2*35);
+  innerloop(2*42);  innerloop(2*49);  innerloop(2*56);  innerloop(2*57);
+  innerloop(2*50);  innerloop(2*43);  innerloop(2*36);  innerloop(2*29);
+  innerloop(2*22);  innerloop(2*15);  innerloop(2*23);  innerloop(2*30);
+  innerloop(2*37);  innerloop(2*44);  innerloop(2*51);  innerloop(2*58);
+  innerloop(2*59);  innerloop(2*52);  innerloop(2*45);  innerloop(2*38);
+  innerloop(2*31);  innerloop(2*39);  innerloop(2*46);  innerloop(2*53);
+  innerloop(2*60);  innerloop(2*61);  innerloop(2*54);  innerloop(2*47);
+  innerloop(2*55);  innerloop(2*62);  innerloop(2*63);
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0) DUMP_SINGLE_VALUE(actbl, 0x0)
+
+  state->cur.put_buffer = put_buffer;
+  state->cur.put_bits = put_bits;
+  STORE_BUFFER()
+
+  return TRUE;
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(boolean)
+emit_restart (working_state * state, int restart_num)
+{
+  int ci;
+
+  if (! flush_bits(state))
+    return FALSE;
+
+  emit_byte(state, 0xFF, return FALSE);
+  emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
+    state->cur.last_dc_val[ci] = 0;
+
+  /* The restart counter is not updated until we successfully write the MCU. */
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of Huffman-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  working_state state;
+  int blkn, ci;
+  jpeg_component_info * compptr;
+
+  /* Load up working state */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  ASSIGN_STATE(state.cur, entropy->saved);
+  state.cinfo = cinfo;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! emit_restart(&state, entropy->next_restart_num))
+	return FALSE;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    if (! encode_one_block(&state,
+			   MCU_data[blkn][0], state.cur.last_dc_val[ci],
+			   entropy->dc_derived_tbls[compptr->dc_tbl_no],
+			   entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+      return FALSE;
+    /* Update last_dc_val */
+    state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  /* Completed MCU, so update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  ASSIGN_STATE(entropy->saved, state.cur);
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass_huff (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  working_state state;
+
+  /* Load up working state ... flush_bits needs it */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  ASSIGN_STATE(state.cur, entropy->saved);
+  state.cinfo = cinfo;
+
+  /* Flush out the last data */
+  if (! flush_bits(&state))
+    ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+  /* Update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  ASSIGN_STATE(entropy->saved, state.cur);
+}
+
+
+/*
+ * Huffman coding optimization.
+ *
+ * We first scan the supplied data and count the number of uses of each symbol
+ * that is to be Huffman-coded. (This process MUST agree with the code above.)
+ * Then we build a Huffman coding tree for the observed counts.
+ * Symbols which are not needed at all for the particular image are not
+ * assigned any code, which saves space in the DHT marker as well as in
+ * the compressed data.
+ */
+
+#ifdef ENTROPY_OPT_SUPPORTED
+
+
+/* Process a single block's worth of coefficients */
+
+LOCAL(void)
+htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
+		 long dc_counts[], long ac_counts[])
+{
+  register int temp;
+  register int nbits;
+  register int k, r;
+  
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+  
+  temp = block[0] - last_dc_val;
+  if (temp < 0)
+    temp = -temp;
+  
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = 0;
+  while (temp) {
+    nbits++;
+    temp >>= 1;
+  }
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > MAX_COEF_BITS+1)
+    ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+  /* Count the Huffman symbol for the number of bits */
+  dc_counts[nbits]++;
+  
+  /* Encode the AC coefficients per section F.1.2.2 */
+  
+  r = 0;			/* r = run length of zeros */
+  
+  for (k = 1; k < DCTSIZE2; k++) {
+    if ((temp = block[jpeg_natural_order[k]]) == 0) {
+      r++;
+    } else {
+      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+      while (r > 15) {
+	ac_counts[0xF0]++;
+	r -= 16;
+      }
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      if (temp < 0)
+	temp = -temp;
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      nbits = 1;		/* there must be at least one 1 bit */
+      while ((temp >>= 1))
+	nbits++;
+      /* Check for out-of-range coefficient values */
+      if (nbits > MAX_COEF_BITS)
+	ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+      
+      /* Count Huffman symbol for run length / number of bits */
+      ac_counts[(r << 4) + nbits]++;
+      
+      r = 0;
+    }
+  }
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0)
+    ac_counts[0]++;
+}
+
+
+/*
+ * Trial-encode one MCU's worth of Huffman-compressed coefficients.
+ * No data is actually output, so no suspension return is possible.
+ */
+
+METHODDEF(boolean)
+encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int blkn, ci;
+  jpeg_component_info * compptr;
+
+  /* Take care of restart intervals if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      /* Re-initialize DC predictions to 0 */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+	entropy->saved.last_dc_val[ci] = 0;
+      /* Update restart state */
+      entropy->restarts_to_go = cinfo->restart_interval;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
+		    entropy->dc_count_ptrs[compptr->dc_tbl_no],
+		    entropy->ac_count_ptrs[compptr->ac_tbl_no]);
+    entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Generate the best Huffman code table for the given counts, fill htbl.
+ * Note this is also used by jcphuff.c.
+ *
+ * The JPEG standard requires that no symbol be assigned a codeword of all
+ * one bits (so that padding bits added at the end of a compressed segment
+ * can't look like a valid code).  Because of the canonical ordering of
+ * codewords, this just means that there must be an unused slot in the
+ * longest codeword length category.  Section K.2 of the JPEG spec suggests
+ * reserving such a slot by pretending that symbol 256 is a valid symbol
+ * with count 1.  In theory that's not optimal; giving it count zero but
+ * including it in the symbol set anyway should give a better Huffman code.
+ * But the theoretically better code actually seems to come out worse in
+ * practice, because it produces more all-ones bytes (which incur stuffed
+ * zero bytes in the final file).  In any case the difference is tiny.
+ *
+ * The JPEG standard requires Huffman codes to be no more than 16 bits long.
+ * If some symbols have a very small but nonzero probability, the Huffman tree
+ * must be adjusted to meet the code length restriction.  We currently use
+ * the adjustment method suggested in JPEG section K.2.  This method is *not*
+ * optimal; it may not choose the best possible limited-length code.  But
+ * typically only very-low-frequency symbols will be given less-than-optimal
+ * lengths, so the code is almost optimal.  Experimental comparisons against
+ * an optimal limited-length-code algorithm indicate that the difference is
+ * microscopic --- usually less than a hundredth of a percent of total size.
+ * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
+ */
+
+GLOBAL(void)
+jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
+{
+#define MAX_CLEN 32		/* assumed maximum initial code length */
+  UINT8 bits[MAX_CLEN+1];	/* bits[k] = # of symbols with code length k */
+  int codesize[257];		/* codesize[k] = code length of symbol k */
+  int others[257];		/* next symbol in current branch of tree */
+  int c1, c2;
+  int p, i, j;
+  long v;
+
+  /* This algorithm is explained in section K.2 of the JPEG standard */
+
+  MEMZERO(bits, SIZEOF(bits));
+  MEMZERO(codesize, SIZEOF(codesize));
+  for (i = 0; i < 257; i++)
+    others[i] = -1;		/* init links to empty */
+  
+  freq[256] = 1;		/* make sure 256 has a nonzero count */
+  /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
+   * that no real symbol is given code-value of all ones, because 256
+   * will be placed last in the largest codeword category.
+   */
+
+  /* Huffman's basic algorithm to assign optimal code lengths to symbols */
+
+  for (;;) {
+    /* Find the smallest nonzero frequency, set c1 = its symbol */
+    /* In case of ties, take the larger symbol number */
+    c1 = -1;
+    v = 1000000000L;
+    for (i = 0; i <= 256; i++) {
+      if (freq[i] && freq[i] <= v) {
+	v = freq[i];
+	c1 = i;
+      }
+    }
+
+    /* Find the next smallest nonzero frequency, set c2 = its symbol */
+    /* In case of ties, take the larger symbol number */
+    c2 = -1;
+    v = 1000000000L;
+    for (i = 0; i <= 256; i++) {
+      if (freq[i] && freq[i] <= v && i != c1) {
+	v = freq[i];
+	c2 = i;
+      }
+    }
+
+    /* Done if we've merged everything into one frequency */
+    if (c2 < 0)
+      break;
+    
+    /* Else merge the two counts/trees */
+    freq[c1] += freq[c2];
+    freq[c2] = 0;
+
+    /* Increment the codesize of everything in c1's tree branch */
+    codesize[c1]++;
+    while (others[c1] >= 0) {
+      c1 = others[c1];
+      codesize[c1]++;
+    }
+    
+    others[c1] = c2;		/* chain c2 onto c1's tree branch */
+    
+    /* Increment the codesize of everything in c2's tree branch */
+    codesize[c2]++;
+    while (others[c2] >= 0) {
+      c2 = others[c2];
+      codesize[c2]++;
+    }
+  }
+
+  /* Now count the number of symbols of each code length */
+  for (i = 0; i <= 256; i++) {
+    if (codesize[i]) {
+      /* The JPEG standard seems to think that this can't happen, */
+      /* but I'm paranoid... */
+      if (codesize[i] > MAX_CLEN)
+	ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
+
+      bits[codesize[i]]++;
+    }
+  }
+
+  /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
+   * Huffman procedure assigned any such lengths, we must adjust the coding.
+   * Here is what the JPEG spec says about how this next bit works:
+   * Since symbols are paired for the longest Huffman code, the symbols are
+   * removed from this length category two at a time.  The prefix for the pair
+   * (which is one bit shorter) is allocated to one of the pair; then,
+   * skipping the BITS entry for that prefix length, a code word from the next
+   * shortest nonzero BITS entry is converted into a prefix for two code words
+   * one bit longer.
+   */
+  
+  for (i = MAX_CLEN; i > 16; i--) {
+    while (bits[i] > 0) {
+      j = i - 2;		/* find length of new prefix to be used */
+      while (bits[j] == 0)
+	j--;
+      
+      bits[i] -= 2;		/* remove two symbols */
+      bits[i-1]++;		/* one goes in this length */
+      bits[j+1] += 2;		/* two new symbols in this length */
+      bits[j]--;		/* symbol of this length is now a prefix */
+    }
+  }
+
+  /* Remove the count for the pseudo-symbol 256 from the largest codelength */
+  while (bits[i] == 0)		/* find largest codelength still in use */
+    i--;
+  bits[i]--;
+  
+  /* Return final symbol counts (only for lengths 0..16) */
+  MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
+  
+  /* Return a list of the symbols sorted by code length */
+  /* It's not real clear to me why we don't need to consider the codelength
+   * changes made above, but the JPEG spec seems to think this works.
+   */
+  p = 0;
+  for (i = 1; i <= MAX_CLEN; i++) {
+    for (j = 0; j <= 255; j++) {
+      if (codesize[j] == i) {
+	htbl->huffval[p] = (UINT8) j;
+	p++;
+      }
+    }
+  }
+
+  /* Set sent_table FALSE so updated table will be written to JPEG file. */
+  htbl->sent_table = FALSE;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, dctbl, actbl;
+  jpeg_component_info * compptr;
+  JHUFF_TBL **htblptr;
+  boolean did_dc[NUM_HUFF_TBLS];
+  boolean did_ac[NUM_HUFF_TBLS];
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  MEMZERO(did_dc, SIZEOF(did_dc));
+  MEMZERO(did_ac, SIZEOF(did_ac));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    if (! did_dc[dctbl]) {
+      htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
+      if (*htblptr == NULL)
+	*htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
+      did_dc[dctbl] = TRUE;
+    }
+    if (! did_ac[actbl]) {
+      htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
+      if (*htblptr == NULL)
+	*htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
+      did_ac[actbl] = TRUE;
+    }
+  }
+}
+
+
+#endif /* ENTROPY_OPT_SUPPORTED */
+
+
+/*
+ * Module initialization routine for Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_huff_encoder (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(huff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass_huff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+#ifdef ENTROPY_OPT_SUPPORTED
+    entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
+#endif
+  }
+}
diff --git a/jchuff.h b/jchuff.h
new file mode 100644
index 0000000..a9599fc
--- /dev/null
+++ b/jchuff.h
@@ -0,0 +1,47 @@
+/*
+ * jchuff.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains declarations for Huffman entropy encoding routines
+ * that are shared between the sequential encoder (jchuff.c) and the
+ * progressive encoder (jcphuff.c).  No other modules need to see these.
+ */
+
+/* The legal range of a DCT coefficient is
+ *  -1024 .. +1023  for 8-bit data;
+ * -16384 .. +16383 for 12-bit data.
+ * Hence the magnitude should always fit in 10 or 14 bits respectively.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MAX_COEF_BITS 10
+#else
+#define MAX_COEF_BITS 14
+#endif
+
+/* Derived data constructed for each Huffman table */
+
+typedef struct {
+  unsigned int ehufco[256];	/* code for each symbol */
+  char ehufsi[256];		/* length of code for each symbol */
+  /* If no code has been allocated for a symbol S, ehufsi[S] contains 0 */
+} c_derived_tbl;
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_make_c_derived_tbl	jMkCDerived
+#define jpeg_gen_optimal_table	jGenOptTbl
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* Expand a Huffman table definition into the derived format */
+EXTERN(void) jpeg_make_c_derived_tbl
+	JPP((j_compress_ptr cinfo, boolean isDC, int tblno,
+	     c_derived_tbl ** pdtbl));
+
+/* Generate an optimal table definition given the specified counts */
+EXTERN(void) jpeg_gen_optimal_table
+	JPP((j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[]));
diff --git a/jcinit.c b/jcinit.c
new file mode 100644
index 0000000..5efffe3
--- /dev/null
+++ b/jcinit.c
@@ -0,0 +1,72 @@
+/*
+ * jcinit.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains initialization logic for the JPEG compressor.
+ * This routine is in charge of selecting the modules to be executed and
+ * making an initialization call to each one.
+ *
+ * Logically, this code belongs in jcmaster.c.  It's split out because
+ * linking this routine implies linking the entire compression library.
+ * For a transcoding-only application, we want to be able to use jcmaster.c
+ * without linking in the whole library.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Master selection of compression modules.
+ * This is done once at the start of processing an image.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ */
+
+GLOBAL(void)
+jinit_compress_master (j_compress_ptr cinfo)
+{
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, FALSE /* full compression */);
+
+  /* Preprocessing */
+  if (! cinfo->raw_data_in) {
+    jinit_color_converter(cinfo);
+    jinit_downsampler(cinfo);
+    jinit_c_prep_controller(cinfo, FALSE /* never need full buffer here */);
+  }
+  /* Forward DCT */
+  jinit_forward_dct(cinfo);
+  /* Entropy encoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      jinit_phuff_encoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_encoder(cinfo);
+  }
+
+  /* Need a full-image coefficient buffer in any multi-pass mode. */
+  jinit_c_coef_controller(cinfo,
+		(boolean) (cinfo->num_scans > 1 || cinfo->optimize_coding));
+  jinit_c_main_controller(cinfo, FALSE /* never need full buffer here */);
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Write the datastream header (SOI) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}
diff --git a/jcmainct.c b/jcmainct.c
new file mode 100644
index 0000000..e0279a7
--- /dev/null
+++ b/jcmainct.c
@@ -0,0 +1,293 @@
+/*
+ * jcmainct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the main buffer controller for compression.
+ * The main buffer lies between the pre-processor and the JPEG
+ * compressor proper; it holds downsampled data in the JPEG colorspace.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Note: currently, there is no operating mode in which a full-image buffer
+ * is needed at this step.  If there were, that mode could not be used with
+ * "raw data" input, since this module is bypassed in that case.  However,
+ * we've left the code here for possible use in special applications.
+ */
+#undef FULL_MAIN_BUFFER_SUPPORTED
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_main_controller pub; /* public fields */
+
+  JDIMENSION cur_iMCU_row;	/* number of current iMCU row */
+  JDIMENSION rowgroup_ctr;	/* counts row groups received in iMCU row */
+  boolean suspended;		/* remember if we suspended output */
+  J_BUF_MODE pass_mode;		/* current operating mode */
+
+  /* If using just a strip buffer, this points to the entire set of buffers
+   * (we allocate one for each component).  In the full-image case, this
+   * points to the currently accessible strips of the virtual arrays.
+   */
+  JSAMPARRAY buffer[MAX_COMPONENTS];
+
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+  /* If using full-image storage, this array holds pointers to virtual-array
+   * control blocks for each component.  Unused if not full-image storage.
+   */
+  jvirt_sarray_ptr whole_image[MAX_COMPONENTS];
+#endif
+} my_main_controller;
+
+typedef my_main_controller * my_main_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main
+	JPP((j_compress_ptr cinfo, JSAMPARRAY input_buf,
+	     JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail));
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+METHODDEF(void) process_data_buffer_main
+	JPP((j_compress_ptr cinfo, JSAMPARRAY input_buf,
+	     JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail));
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  /* Do nothing in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  main->cur_iMCU_row = 0;	/* initialize counters */
+  main->rowgroup_ctr = 0;
+  main->suspended = FALSE;
+  main->pass_mode = pass_mode;	/* save mode for use by process_data */
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    if (main->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+    main->pub.process_data = process_data_simple_main;
+    break;
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+  case JBUF_SAVE_SOURCE:
+  case JBUF_CRANK_DEST:
+  case JBUF_SAVE_AND_PASS:
+    if (main->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    main->pub.process_data = process_data_buffer_main;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This routine handles the simple pass-through mode,
+ * where we have only a strip buffer.
+ */
+
+METHODDEF(void)
+process_data_simple_main (j_compress_ptr cinfo,
+			  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			  JDIMENSION in_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  while (main->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Read input data if we haven't filled the main buffer yet */
+    if (main->rowgroup_ctr < DCTSIZE)
+      (*cinfo->prep->pre_process_data) (cinfo,
+					input_buf, in_row_ctr, in_rows_avail,
+					main->buffer, &main->rowgroup_ctr,
+					(JDIMENSION) DCTSIZE);
+
+    /* If we don't have a full iMCU row buffered, return to application for
+     * more data.  Note that preprocessor will always pad to fill the iMCU row
+     * at the bottom of the image.
+     */
+    if (main->rowgroup_ctr != DCTSIZE)
+      return;
+
+    /* Send the completed row to the compressor */
+    if (! (*cinfo->coef->compress_data) (cinfo, main->buffer)) {
+      /* If compressor did not consume the whole row, then we must need to
+       * suspend processing and return to the application.  In this situation
+       * we pretend we didn't yet consume the last input row; otherwise, if
+       * it happened to be the last row of the image, the application would
+       * think we were done.
+       */
+      if (! main->suspended) {
+	(*in_row_ctr)--;
+	main->suspended = TRUE;
+      }
+      return;
+    }
+    /* We did finish the row.  Undo our little suspension hack if a previous
+     * call suspended; then mark the main buffer empty.
+     */
+    if (main->suspended) {
+      (*in_row_ctr)++;
+      main->suspended = FALSE;
+    }
+    main->rowgroup_ctr = 0;
+    main->cur_iMCU_row++;
+  }
+}
+
+
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+
+/*
+ * Process some data.
+ * This routine handles all of the modes that use a full-size buffer.
+ */
+
+METHODDEF(void)
+process_data_buffer_main (j_compress_ptr cinfo,
+			  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			  JDIMENSION in_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci;
+  jpeg_component_info *compptr;
+  boolean writing = (main->pass_mode != JBUF_CRANK_DEST);
+
+  while (main->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Realign the virtual buffers if at the start of an iMCU row. */
+    if (main->rowgroup_ctr == 0) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	   ci++, compptr++) {
+	main->buffer[ci] = (*cinfo->mem->access_virt_sarray)
+	  ((j_common_ptr) cinfo, main->whole_image[ci],
+	   main->cur_iMCU_row * (compptr->v_samp_factor * DCTSIZE),
+	   (JDIMENSION) (compptr->v_samp_factor * DCTSIZE), writing);
+      }
+      /* In a read pass, pretend we just read some source data. */
+      if (! writing) {
+	*in_row_ctr += cinfo->max_v_samp_factor * DCTSIZE;
+	main->rowgroup_ctr = DCTSIZE;
+      }
+    }
+
+    /* If a write pass, read input data until the current iMCU row is full. */
+    /* Note: preprocessor will pad if necessary to fill the last iMCU row. */
+    if (writing) {
+      (*cinfo->prep->pre_process_data) (cinfo,
+					input_buf, in_row_ctr, in_rows_avail,
+					main->buffer, &main->rowgroup_ctr,
+					(JDIMENSION) DCTSIZE);
+      /* Return to application if we need more data to fill the iMCU row. */
+      if (main->rowgroup_ctr < DCTSIZE)
+	return;
+    }
+
+    /* Emit data, unless this is a sink-only pass. */
+    if (main->pass_mode != JBUF_SAVE_SOURCE) {
+      if (! (*cinfo->coef->compress_data) (cinfo, main->buffer)) {
+	/* If compressor did not consume the whole row, then we must need to
+	 * suspend processing and return to the application.  In this situation
+	 * we pretend we didn't yet consume the last input row; otherwise, if
+	 * it happened to be the last row of the image, the application would
+	 * think we were done.
+	 */
+	if (! main->suspended) {
+	  (*in_row_ctr)--;
+	  main->suspended = TRUE;
+	}
+	return;
+      }
+      /* We did finish the row.  Undo our little suspension hack if a previous
+       * call suspended; then mark the main buffer empty.
+       */
+      if (main->suspended) {
+	(*in_row_ctr)++;
+	main->suspended = FALSE;
+      }
+    }
+
+    /* If get here, we are done with this iMCU row.  Mark buffer empty. */
+    main->rowgroup_ctr = 0;
+    main->cur_iMCU_row++;
+  }
+}
+
+#endif /* FULL_MAIN_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+jinit_c_main_controller (j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_main_ptr main;
+  int ci;
+  jpeg_component_info *compptr;
+
+  main = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_main_controller));
+  cinfo->main = (struct jpeg_c_main_controller *) main;
+  main->pub.start_pass = start_pass_main;
+
+  /* We don't need to create a buffer in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  /* Create the buffer.  It holds downsampled data, so each component
+   * may be of a different size.
+   */
+  if (need_full_buffer) {
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component */
+    /* Note we pad the bottom to a multiple of the iMCU height */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      main->whole_image[ci] = (*cinfo->mem->request_virt_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 compptr->width_in_blocks * DCTSIZE,
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor) * DCTSIZE,
+	 (JDIMENSION) (compptr->v_samp_factor * DCTSIZE));
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    main->whole_image[0] = NULL; /* flag for no virtual arrays */
+#endif
+    /* Allocate a strip buffer for each component */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      main->buffer[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 compptr->width_in_blocks * DCTSIZE,
+	 (JDIMENSION) (compptr->v_samp_factor * DCTSIZE));
+    }
+  }
+}
diff --git a/jcmarker.c b/jcmarker.c
new file mode 100644
index 0000000..b1c1e45
--- /dev/null
+++ b/jcmarker.c
@@ -0,0 +1,666 @@
+/*
+ * jcmarker.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write JPEG datastream markers.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+
+typedef enum {			/* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+  
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+  
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+  
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+  
+  M_DHT   = 0xc4,
+  
+  M_DAC   = 0xcc,
+  
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+  
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+  
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+  
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+  
+  M_TEM   = 0x01,
+  
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_writer pub; /* public fields */
+
+  unsigned int last_restart_interval; /* last DRI value emitted; 0 after SOI */
+} my_marker_writer;
+
+typedef my_marker_writer * my_marker_ptr;
+
+
+/*
+ * Basic output routines.
+ *
+ * Note that we do not support suspension while writing a marker.
+ * Therefore, an application using suspension must ensure that there is
+ * enough buffer space for the initial markers (typ. 600-700 bytes) before
+ * calling jpeg_start_compress, and enough space to write the trailing EOI
+ * (a few bytes) before calling jpeg_finish_compress.  Multipass compression
+ * modes are not supported at all with suspension, so those two are the only
+ * points where markers will be written.
+ */
+
+LOCAL(void)
+emit_byte (j_compress_ptr cinfo, int val)
+/* Emit a byte */
+{
+  struct jpeg_destination_mgr * dest = cinfo->dest;
+
+  *(dest->next_output_byte)++ = (JOCTET) val;
+  if (--dest->free_in_buffer == 0) {
+    if (! (*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  }
+}
+
+
+LOCAL(void)
+emit_marker (j_compress_ptr cinfo, JPEG_MARKER mark)
+/* Emit a marker code */
+{
+  emit_byte(cinfo, 0xFF);
+  emit_byte(cinfo, (int) mark);
+}
+
+
+LOCAL(void)
+emit_2bytes (j_compress_ptr cinfo, int value)
+/* Emit a 2-byte integer; these are always MSB first in JPEG files */
+{
+  emit_byte(cinfo, (value >> 8) & 0xFF);
+  emit_byte(cinfo, value & 0xFF);
+}
+
+
+/*
+ * Routines to write specific marker types.
+ */
+
+LOCAL(int)
+emit_dqt (j_compress_ptr cinfo, int index)
+/* Emit a DQT marker */
+/* Returns the precision used (0 = 8bits, 1 = 16bits) for baseline checking */
+{
+  JQUANT_TBL * qtbl = cinfo->quant_tbl_ptrs[index];
+  int prec;
+  int i;
+
+  if (qtbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, index);
+
+  prec = 0;
+  for (i = 0; i < DCTSIZE2; i++) {
+    if (qtbl->quantval[i] > 255)
+      prec = 1;
+  }
+
+  if (! qtbl->sent_table) {
+    emit_marker(cinfo, M_DQT);
+
+    emit_2bytes(cinfo, prec ? DCTSIZE2*2 + 1 + 2 : DCTSIZE2 + 1 + 2);
+
+    emit_byte(cinfo, index + (prec<<4));
+
+    for (i = 0; i < DCTSIZE2; i++) {
+      /* The table entries must be emitted in zigzag order. */
+      unsigned int qval = qtbl->quantval[jpeg_natural_order[i]];
+      if (prec)
+	emit_byte(cinfo, (int) (qval >> 8));
+      emit_byte(cinfo, (int) (qval & 0xFF));
+    }
+
+    qtbl->sent_table = TRUE;
+  }
+
+  return prec;
+}
+
+
+LOCAL(void)
+emit_dht (j_compress_ptr cinfo, int index, boolean is_ac)
+/* Emit a DHT marker */
+{
+  JHUFF_TBL * htbl;
+  int length, i;
+  
+  if (is_ac) {
+    htbl = cinfo->ac_huff_tbl_ptrs[index];
+    index += 0x10;		/* output index has AC bit set */
+  } else {
+    htbl = cinfo->dc_huff_tbl_ptrs[index];
+  }
+
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, index);
+  
+  if (! htbl->sent_table) {
+    emit_marker(cinfo, M_DHT);
+    
+    length = 0;
+    for (i = 1; i <= 16; i++)
+      length += htbl->bits[i];
+    
+    emit_2bytes(cinfo, length + 2 + 1 + 16);
+    emit_byte(cinfo, index);
+    
+    for (i = 1; i <= 16; i++)
+      emit_byte(cinfo, htbl->bits[i]);
+    
+    for (i = 0; i < length; i++)
+      emit_byte(cinfo, htbl->huffval[i]);
+    
+    htbl->sent_table = TRUE;
+  }
+}
+
+
+LOCAL(void)
+emit_dac (j_compress_ptr cinfo)
+/* Emit a DAC marker */
+/* Since the useful info is so small, we want to emit all the tables in */
+/* one DAC marker.  Therefore this routine does its own scan of the table. */
+{
+#ifdef C_ARITH_CODING_SUPPORTED
+  char dc_in_use[NUM_ARITH_TBLS];
+  char ac_in_use[NUM_ARITH_TBLS];
+  int length, i;
+  jpeg_component_info *compptr;
+  
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    dc_in_use[i] = ac_in_use[i] = 0;
+  
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    dc_in_use[compptr->dc_tbl_no] = 1;
+    ac_in_use[compptr->ac_tbl_no] = 1;
+  }
+  
+  length = 0;
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    length += dc_in_use[i] + ac_in_use[i];
+  
+  emit_marker(cinfo, M_DAC);
+  
+  emit_2bytes(cinfo, length*2 + 2);
+  
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    if (dc_in_use[i]) {
+      emit_byte(cinfo, i);
+      emit_byte(cinfo, cinfo->arith_dc_L[i] + (cinfo->arith_dc_U[i]<<4));
+    }
+    if (ac_in_use[i]) {
+      emit_byte(cinfo, i + 0x10);
+      emit_byte(cinfo, cinfo->arith_ac_K[i]);
+    }
+  }
+#endif /* C_ARITH_CODING_SUPPORTED */
+}
+
+
+LOCAL(void)
+emit_dri (j_compress_ptr cinfo)
+/* Emit a DRI marker */
+{
+  emit_marker(cinfo, M_DRI);
+  
+  emit_2bytes(cinfo, 4);	/* fixed length */
+
+  emit_2bytes(cinfo, (int) cinfo->restart_interval);
+}
+
+
+LOCAL(void)
+emit_sof (j_compress_ptr cinfo, JPEG_MARKER code)
+/* Emit a SOF marker */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  
+  emit_marker(cinfo, code);
+  
+  emit_2bytes(cinfo, 3 * cinfo->num_components + 2 + 5 + 1); /* length */
+
+  /* Make sure image isn't bigger than SOF field can handle */
+  if ((long) cinfo->_jpeg_height > 65535L ||
+      (long) cinfo->_jpeg_width > 65535L)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) 65535);
+
+  emit_byte(cinfo, cinfo->data_precision);
+  emit_2bytes(cinfo, (int) cinfo->_jpeg_height);
+  emit_2bytes(cinfo, (int) cinfo->_jpeg_width);
+
+  emit_byte(cinfo, cinfo->num_components);
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    emit_byte(cinfo, compptr->component_id);
+    emit_byte(cinfo, (compptr->h_samp_factor << 4) + compptr->v_samp_factor);
+    emit_byte(cinfo, compptr->quant_tbl_no);
+  }
+}
+
+
+LOCAL(void)
+emit_sos (j_compress_ptr cinfo)
+/* Emit a SOS marker */
+{
+  int i, td, ta;
+  jpeg_component_info *compptr;
+  
+  emit_marker(cinfo, M_SOS);
+  
+  emit_2bytes(cinfo, 2 * cinfo->comps_in_scan + 2 + 1 + 3); /* length */
+  
+  emit_byte(cinfo, cinfo->comps_in_scan);
+  
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    emit_byte(cinfo, compptr->component_id);
+    td = compptr->dc_tbl_no;
+    ta = compptr->ac_tbl_no;
+    if (cinfo->progressive_mode) {
+      /* Progressive mode: only DC or only AC tables are used in one scan;
+       * furthermore, Huffman coding of DC refinement uses no table at all.
+       * We emit 0 for unused field(s); this is recommended by the P&M text
+       * but does not seem to be specified in the standard.
+       */
+      if (cinfo->Ss == 0) {
+	ta = 0;			/* DC scan */
+	if (cinfo->Ah != 0 && !cinfo->arith_code)
+	  td = 0;		/* no DC table either */
+      } else {
+	td = 0;			/* AC scan */
+      }
+    }
+    emit_byte(cinfo, (td << 4) + ta);
+  }
+
+  emit_byte(cinfo, cinfo->Ss);
+  emit_byte(cinfo, cinfo->Se);
+  emit_byte(cinfo, (cinfo->Ah << 4) + cinfo->Al);
+}
+
+
+LOCAL(void)
+emit_jfif_app0 (j_compress_ptr cinfo)
+/* Emit a JFIF-compliant APP0 marker */
+{
+  /*
+   * Length of APP0 block	(2 bytes)
+   * Block ID			(4 bytes - ASCII "JFIF")
+   * Zero byte			(1 byte to terminate the ID string)
+   * Version Major, Minor	(2 bytes - major first)
+   * Units			(1 byte - 0x00 = none, 0x01 = inch, 0x02 = cm)
+   * Xdpu			(2 bytes - dots per unit horizontal)
+   * Ydpu			(2 bytes - dots per unit vertical)
+   * Thumbnail X size		(1 byte)
+   * Thumbnail Y size		(1 byte)
+   */
+  
+  emit_marker(cinfo, M_APP0);
+  
+  emit_2bytes(cinfo, 2 + 4 + 1 + 2 + 1 + 2 + 2 + 1 + 1); /* length */
+
+  emit_byte(cinfo, 0x4A);	/* Identifier: ASCII "JFIF" */
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0x49);
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0);
+  emit_byte(cinfo, cinfo->JFIF_major_version); /* Version fields */
+  emit_byte(cinfo, cinfo->JFIF_minor_version);
+  emit_byte(cinfo, cinfo->density_unit); /* Pixel size information */
+  emit_2bytes(cinfo, (int) cinfo->X_density);
+  emit_2bytes(cinfo, (int) cinfo->Y_density);
+  emit_byte(cinfo, 0);		/* No thumbnail image */
+  emit_byte(cinfo, 0);
+}
+
+
+LOCAL(void)
+emit_adobe_app14 (j_compress_ptr cinfo)
+/* Emit an Adobe APP14 marker */
+{
+  /*
+   * Length of APP14 block	(2 bytes)
+   * Block ID			(5 bytes - ASCII "Adobe")
+   * Version Number		(2 bytes - currently 100)
+   * Flags0			(2 bytes - currently 0)
+   * Flags1			(2 bytes - currently 0)
+   * Color transform		(1 byte)
+   *
+   * Although Adobe TN 5116 mentions Version = 101, all the Adobe files
+   * now in circulation seem to use Version = 100, so that's what we write.
+   *
+   * We write the color transform byte as 1 if the JPEG color space is
+   * YCbCr, 2 if it's YCCK, 0 otherwise.  Adobe's definition has to do with
+   * whether the encoder performed a transformation, which is pretty useless.
+   */
+  
+  emit_marker(cinfo, M_APP14);
+  
+  emit_2bytes(cinfo, 2 + 5 + 2 + 2 + 2 + 1); /* length */
+
+  emit_byte(cinfo, 0x41);	/* Identifier: ASCII "Adobe" */
+  emit_byte(cinfo, 0x64);
+  emit_byte(cinfo, 0x6F);
+  emit_byte(cinfo, 0x62);
+  emit_byte(cinfo, 0x65);
+  emit_2bytes(cinfo, 100);	/* Version */
+  emit_2bytes(cinfo, 0);	/* Flags0 */
+  emit_2bytes(cinfo, 0);	/* Flags1 */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_YCbCr:
+    emit_byte(cinfo, 1);	/* Color transform = 1 */
+    break;
+  case JCS_YCCK:
+    emit_byte(cinfo, 2);	/* Color transform = 2 */
+    break;
+  default:
+    emit_byte(cinfo, 0);	/* Color transform = 0 */
+    break;
+  }
+}
+
+
+/*
+ * These routines allow writing an arbitrary marker with parameters.
+ * The only intended use is to emit COM or APPn markers after calling
+ * write_file_header and before calling write_frame_header.
+ * Other uses are not guaranteed to produce desirable results.
+ * Counting the parameter bytes properly is the caller's responsibility.
+ */
+
+METHODDEF(void)
+write_marker_header (j_compress_ptr cinfo, int marker, unsigned int datalen)
+/* Emit an arbitrary marker header */
+{
+  if (datalen > (unsigned int) 65533)		/* safety check */
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  emit_marker(cinfo, (JPEG_MARKER) marker);
+
+  emit_2bytes(cinfo, (int) (datalen + 2));	/* total length */
+}
+
+METHODDEF(void)
+write_marker_byte (j_compress_ptr cinfo, int val)
+/* Emit one byte of marker parameters following write_marker_header */
+{
+  emit_byte(cinfo, val);
+}
+
+
+/*
+ * Write datastream header.
+ * This consists of an SOI and optional APPn markers.
+ * We recommend use of the JFIF marker, but not the Adobe marker,
+ * when using YCbCr or grayscale data.  The JFIF marker should NOT
+ * be used for any other JPEG colorspace.  The Adobe marker is helpful
+ * to distinguish RGB, CMYK, and YCCK colorspaces.
+ * Note that an application can write additional header markers after
+ * jpeg_start_compress returns.
+ */
+
+METHODDEF(void)
+write_file_header (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  emit_marker(cinfo, M_SOI);	/* first the SOI */
+
+  /* SOI is defined to reset restart interval to 0 */
+  marker->last_restart_interval = 0;
+
+  if (cinfo->write_JFIF_header)	/* next an optional JFIF APP0 */
+    emit_jfif_app0(cinfo);
+  if (cinfo->write_Adobe_marker) /* next an optional Adobe APP14 */
+    emit_adobe_app14(cinfo);
+}
+
+
+/*
+ * Write frame header.
+ * This consists of DQT and SOFn markers.
+ * Note that we do not emit the SOF until we have emitted the DQT(s).
+ * This avoids compatibility problems with incorrect implementations that
+ * try to error-check the quant table numbers as soon as they see the SOF.
+ */
+
+METHODDEF(void)
+write_frame_header (j_compress_ptr cinfo)
+{
+  int ci, prec;
+  boolean is_baseline;
+  jpeg_component_info *compptr;
+  
+  /* Emit DQT for each quantization table.
+   * Note that emit_dqt() suppresses any duplicate tables.
+   */
+  prec = 0;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    prec += emit_dqt(cinfo, compptr->quant_tbl_no);
+  }
+  /* now prec is nonzero iff there are any 16-bit quant tables. */
+
+  /* Check for a non-baseline specification.
+   * Note we assume that Huffman table numbers won't be changed later.
+   */
+  if (cinfo->arith_code || cinfo->progressive_mode ||
+      cinfo->data_precision != 8) {
+    is_baseline = FALSE;
+  } else {
+    is_baseline = TRUE;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      if (compptr->dc_tbl_no > 1 || compptr->ac_tbl_no > 1)
+	is_baseline = FALSE;
+    }
+    if (prec && is_baseline) {
+      is_baseline = FALSE;
+      /* If it's baseline except for quantizer size, warn the user */
+      TRACEMS(cinfo, 0, JTRC_16BIT_TABLES);
+    }
+  }
+
+  /* Emit the proper SOF marker */
+  if (cinfo->arith_code) {
+    emit_sof(cinfo, M_SOF9);	/* SOF code for arithmetic coding */
+  } else {
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF2);	/* SOF code for progressive Huffman */
+    else if (is_baseline)
+      emit_sof(cinfo, M_SOF0);	/* SOF code for baseline implementation */
+    else
+      emit_sof(cinfo, M_SOF1);	/* SOF code for non-baseline Huffman file */
+  }
+}
+
+
+/*
+ * Write scan header.
+ * This consists of DHT or DAC markers, optional DRI, and SOS.
+ * Compressed data will be written following the SOS.
+ */
+
+METHODDEF(void)
+write_scan_header (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  int i;
+  jpeg_component_info *compptr;
+
+  if (cinfo->arith_code) {
+    /* Emit arith conditioning info.  We may have some duplication
+     * if the file has multiple scans, but it's so small it's hardly
+     * worth worrying about.
+     */
+    emit_dac(cinfo);
+  } else {
+    /* Emit Huffman tables.
+     * Note that emit_dht() suppresses any duplicate tables.
+     */
+    for (i = 0; i < cinfo->comps_in_scan; i++) {
+      compptr = cinfo->cur_comp_info[i];
+      if (cinfo->progressive_mode) {
+	/* Progressive mode: only DC or only AC tables are used in one scan */
+	if (cinfo->Ss == 0) {
+	  if (cinfo->Ah == 0)	/* DC needs no table for refinement scan */
+	    emit_dht(cinfo, compptr->dc_tbl_no, FALSE);
+	} else {
+	  emit_dht(cinfo, compptr->ac_tbl_no, TRUE);
+	}
+      } else {
+	/* Sequential mode: need both DC and AC tables */
+	emit_dht(cinfo, compptr->dc_tbl_no, FALSE);
+	emit_dht(cinfo, compptr->ac_tbl_no, TRUE);
+      }
+    }
+  }
+
+  /* Emit DRI if required --- note that DRI value could change for each scan.
+   * We avoid wasting space with unnecessary DRIs, however.
+   */
+  if (cinfo->restart_interval != marker->last_restart_interval) {
+    emit_dri(cinfo);
+    marker->last_restart_interval = cinfo->restart_interval;
+  }
+
+  emit_sos(cinfo);
+}
+
+
+/*
+ * Write datastream trailer.
+ */
+
+METHODDEF(void)
+write_file_trailer (j_compress_ptr cinfo)
+{
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Write an abbreviated table-specification datastream.
+ * This consists of SOI, DQT and DHT tables, and EOI.
+ * Any table that is defined and not marked sent_table = TRUE will be
+ * emitted.  Note that all tables will be marked sent_table = TRUE at exit.
+ */
+
+METHODDEF(void)
+write_tables_only (j_compress_ptr cinfo)
+{
+  int i;
+
+  emit_marker(cinfo, M_SOI);
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if (cinfo->quant_tbl_ptrs[i] != NULL)
+      (void) emit_dqt(cinfo, i);
+  }
+
+  if (! cinfo->arith_code) {
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      if (cinfo->dc_huff_tbl_ptrs[i] != NULL)
+	emit_dht(cinfo, i, FALSE);
+      if (cinfo->ac_huff_tbl_ptrs[i] != NULL)
+	emit_dht(cinfo, i, TRUE);
+    }
+  }
+
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Initialize the marker writer module.
+ */
+
+GLOBAL(void)
+jinit_marker_writer (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker;
+
+  /* Create the subobject */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_marker_writer));
+  cinfo->marker = (struct jpeg_marker_writer *) marker;
+  /* Initialize method pointers */
+  marker->pub.write_file_header = write_file_header;
+  marker->pub.write_frame_header = write_frame_header;
+  marker->pub.write_scan_header = write_scan_header;
+  marker->pub.write_file_trailer = write_file_trailer;
+  marker->pub.write_tables_only = write_tables_only;
+  marker->pub.write_marker_header = write_marker_header;
+  marker->pub.write_marker_byte = write_marker_byte;
+  /* Initialize private state */
+  marker->last_restart_interval = 0;
+}
diff --git a/jcmaster.c b/jcmaster.c
new file mode 100644
index 0000000..74df555
--- /dev/null
+++ b/jcmaster.c
@@ -0,0 +1,622 @@
+/*
+ * jcmaster.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains master control logic for the JPEG compressor.
+ * These routines are concerned with parameter validation, initial setup,
+ * and inter-pass control (determining the number of passes and the work 
+ * to be done in each pass).
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+
+/* Private state */
+
+typedef enum {
+	main_pass,		/* input data, also do first output step */
+	huff_opt_pass,		/* Huffman code optimization pass */
+	output_pass		/* data output pass */
+} c_pass_type;
+
+typedef struct {
+  struct jpeg_comp_master pub;	/* public fields */
+
+  c_pass_type pass_type;	/* the type of the current pass */
+
+  int pass_number;		/* # of passes completed */
+  int total_passes;		/* total # of passes needed */
+
+  int scan_number;		/* current index in scan_info[] */
+} my_comp_master;
+
+typedef my_comp_master * my_master_ptr;
+
+
+/*
+ * Support routines that do various essential calculations.
+ */
+
+#if JPEG_LIB_VERSION >= 70
+/*
+ * Compute JPEG image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ */
+
+GLOBAL(void)
+jpeg_calc_jpeg_dimensions (j_compress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+  /* Hardwire it to "no scaling" */
+  cinfo->jpeg_width = cinfo->image_width;
+  cinfo->jpeg_height = cinfo->image_height;
+  cinfo->min_DCT_h_scaled_size = DCTSIZE;
+  cinfo->min_DCT_v_scaled_size = DCTSIZE;
+}
+#endif
+
+
+LOCAL(void)
+initial_setup (j_compress_ptr cinfo, boolean transcode_only)
+/* Do computations that are needed before master selection phase */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+#if JPEG_LIB_VERSION >= 70
+  if (!transcode_only)
+    jpeg_calc_jpeg_dimensions(cinfo);
+#endif
+
+  /* Sanity check on image dimensions */
+  if (cinfo->_jpeg_height <= 0 || cinfo->_jpeg_width <= 0
+      || cinfo->num_components <= 0 || cinfo->input_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long) cinfo->_jpeg_height > (long) JPEG_MAX_DIMENSION ||
+      (long) cinfo->_jpeg_width > (long) JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* Width of an input scanline must be representable as JDIMENSION. */
+  samplesperrow = (long) cinfo->image_width * (long) cinfo->input_components;
+  jd_samplesperrow = (JDIMENSION) samplesperrow;
+  if ((long) jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* For now, precision must match compiled-in value... */
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	     MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor<=0 || compptr->h_samp_factor>MAX_SAMP_FACTOR ||
+	compptr->v_samp_factor<=0 || compptr->v_samp_factor>MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+				   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+				   compptr->v_samp_factor);
+  }
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Fill in the correct component_index value; don't rely on application */
+    compptr->component_index = ci;
+    /* For compression, we never do DCT scaling. */
+#if JPEG_LIB_VERSION >= 70
+    compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size = DCTSIZE;
+#else
+    compptr->DCT_scaled_size = DCTSIZE;
+#endif
+    /* Size in DCT blocks */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->_jpeg_width * (long) compptr->h_samp_factor,
+		    (long) (cinfo->max_h_samp_factor * DCTSIZE));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->_jpeg_height * (long) compptr->v_samp_factor,
+		    (long) (cinfo->max_v_samp_factor * DCTSIZE));
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->_jpeg_width * (long) compptr->h_samp_factor,
+		    (long) cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->_jpeg_height * (long) compptr->v_samp_factor,
+		    (long) cinfo->max_v_samp_factor);
+    /* Mark component needed (this flag isn't actually used for compression) */
+    compptr->component_needed = TRUE;
+  }
+
+  /* Compute number of fully interleaved MCU rows (number of times that
+   * main controller will call coefficient controller).
+   */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long) cinfo->_jpeg_height,
+		  (long) (cinfo->max_v_samp_factor*DCTSIZE));
+}
+
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+
+LOCAL(void)
+validate_script (j_compress_ptr cinfo)
+/* Verify that the scan script in cinfo->scan_info[] is valid; also
+ * determine whether it uses progressive JPEG, and set cinfo->progressive_mode.
+ */
+{
+  const jpeg_scan_info * scanptr;
+  int scanno, ncomps, ci, coefi, thisi;
+  int Ss, Se, Ah, Al;
+  boolean component_sent[MAX_COMPONENTS];
+#ifdef C_PROGRESSIVE_SUPPORTED
+  int * last_bitpos_ptr;
+  int last_bitpos[MAX_COMPONENTS][DCTSIZE2];
+  /* -1 until that coefficient has been seen; then last Al for it */
+#endif
+
+  if (cinfo->num_scans <= 0)
+    ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, 0);
+
+  /* For sequential JPEG, all scans must have Ss=0, Se=DCTSIZE2-1;
+   * for progressive JPEG, no scan can have this.
+   */
+  scanptr = cinfo->scan_info;
+  if (scanptr->Ss != 0 || scanptr->Se != DCTSIZE2-1) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    cinfo->progressive_mode = TRUE;
+    last_bitpos_ptr = & last_bitpos[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      for (coefi = 0; coefi < DCTSIZE2; coefi++)
+	*last_bitpos_ptr++ = -1;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      component_sent[ci] = FALSE;
+  }
+
+  for (scanno = 1; scanno <= cinfo->num_scans; scanptr++, scanno++) {
+    /* Validate component indexes */
+    ncomps = scanptr->comps_in_scan;
+    if (ncomps <= 0 || ncomps > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, ncomps, MAX_COMPS_IN_SCAN);
+    for (ci = 0; ci < ncomps; ci++) {
+      thisi = scanptr->component_index[ci];
+      if (thisi < 0 || thisi >= cinfo->num_components)
+	ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+      /* Components must appear in SOF order within each scan */
+      if (ci > 0 && thisi <= scanptr->component_index[ci-1])
+	ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+    }
+    /* Validate progression parameters */
+    Ss = scanptr->Ss;
+    Se = scanptr->Se;
+    Ah = scanptr->Ah;
+    Al = scanptr->Al;
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      /* The JPEG spec simply gives the ranges 0..13 for Ah and Al, but that
+       * seems wrong: the upper bound ought to depend on data precision.
+       * Perhaps they really meant 0..N+1 for N-bit precision.
+       * Here we allow 0..10 for 8-bit data; Al larger than 10 results in
+       * out-of-range reconstructed DC values during the first DC scan,
+       * which might cause problems for some decoders.
+       */
+#if BITS_IN_JSAMPLE == 8
+#define MAX_AH_AL 10
+#else
+#define MAX_AH_AL 13
+#endif
+      if (Ss < 0 || Ss >= DCTSIZE2 || Se < Ss || Se >= DCTSIZE2 ||
+	  Ah < 0 || Ah > MAX_AH_AL || Al < 0 || Al > MAX_AH_AL)
+	ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      if (Ss == 0) {
+	if (Se != 0)		/* DC and AC together not OK */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      } else {
+	if (ncomps != 1)	/* AC scans must be for only one component */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      }
+      for (ci = 0; ci < ncomps; ci++) {
+	last_bitpos_ptr = & last_bitpos[scanptr->component_index[ci]][0];
+	if (Ss != 0 && last_bitpos_ptr[0] < 0) /* AC without prior DC scan */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	for (coefi = Ss; coefi <= Se; coefi++) {
+	  if (last_bitpos_ptr[coefi] < 0) {
+	    /* first scan of this coefficient */
+	    if (Ah != 0)
+	      ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	  } else {
+	    /* not first scan */
+	    if (Ah != last_bitpos_ptr[coefi] || Al != Ah-1)
+	      ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	  }
+	  last_bitpos_ptr[coefi] = Al;
+	}
+      }
+#endif
+    } else {
+      /* For sequential JPEG, all progression parameters must be these: */
+      if (Ss != 0 || Se != DCTSIZE2-1 || Ah != 0 || Al != 0)
+	ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      /* Make sure components are not sent twice */
+      for (ci = 0; ci < ncomps; ci++) {
+	thisi = scanptr->component_index[ci];
+	if (component_sent[thisi])
+	  ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+	component_sent[thisi] = TRUE;
+      }
+    }
+  }
+
+  /* Now verify that everything got sent. */
+  if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    /* For progressive mode, we only check that at least some DC data
+     * got sent for each component; the spec does not require that all bits
+     * of all coefficients be transmitted.  Would it be wiser to enforce
+     * transmission of all coefficient bits??
+     */
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (last_bitpos[ci][0] < 0)
+	ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+#endif
+  } else {
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (! component_sent[ci])
+	ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+  }
+}
+
+#endif /* C_MULTISCAN_FILES_SUPPORTED */
+
+
+LOCAL(void)
+select_scan_parameters (j_compress_ptr cinfo)
+/* Set up the scan parameters for the current scan */
+{
+  int ci;
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+  if (cinfo->scan_info != NULL) {
+    /* Prepare for current scan --- the script is already validated */
+    my_master_ptr master = (my_master_ptr) cinfo->master;
+    const jpeg_scan_info * scanptr = cinfo->scan_info + master->scan_number;
+
+    cinfo->comps_in_scan = scanptr->comps_in_scan;
+    for (ci = 0; ci < scanptr->comps_in_scan; ci++) {
+      cinfo->cur_comp_info[ci] =
+	&cinfo->comp_info[scanptr->component_index[ci]];
+    }
+    cinfo->Ss = scanptr->Ss;
+    cinfo->Se = scanptr->Se;
+    cinfo->Ah = scanptr->Ah;
+    cinfo->Al = scanptr->Al;
+  }
+  else
+#endif
+  {
+    /* Prepare for single sequential-JPEG scan containing all components */
+    if (cinfo->num_components > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	       MAX_COMPS_IN_SCAN);
+    cinfo->comps_in_scan = cinfo->num_components;
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      cinfo->cur_comp_info[ci] = &cinfo->comp_info[ci];
+    }
+    cinfo->Ss = 0;
+    cinfo->Se = DCTSIZE2-1;
+    cinfo->Ah = 0;
+    cinfo->Al = 0;
+  }
+}
+
+
+LOCAL(void)
+per_scan_setup (j_compress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] are already set */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  
+  if (cinfo->comps_in_scan == 1) {
+    
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+    
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = DCTSIZE;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+    
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+    
+  } else {
+    
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+	       MAX_COMPS_IN_SCAN);
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long) cinfo->_jpeg_width,
+		    (long) (cinfo->max_h_samp_factor*DCTSIZE));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long) cinfo->_jpeg_height,
+		    (long) (cinfo->max_v_samp_factor*DCTSIZE));
+    
+    cinfo->blocks_in_MCU = 0;
+    
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * DCTSIZE;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int) (compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int) (compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > C_MAX_BLOCKS_IN_MCU)
+	ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+	cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+    
+  }
+
+  /* Convert restart specified in rows to actual MCU count. */
+  /* Note that count must fit in 16 bits, so we provide limiting. */
+  if (cinfo->restart_in_rows > 0) {
+    long nominal = (long) cinfo->restart_in_rows * (long) cinfo->MCUs_per_row;
+    cinfo->restart_interval = (unsigned int) MIN(nominal, 65535L);
+  }
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each pass.  We determine which modules
+ * will be active during this pass and give them appropriate start_pass calls.
+ * We also set is_last_pass to indicate whether any more passes will be
+ * required.
+ */
+
+METHODDEF(void)
+prepare_for_pass (j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  switch (master->pass_type) {
+  case main_pass:
+    /* Initial pass: will collect input data, and do either Huffman
+     * optimization or data output for the first scan.
+     */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (! cinfo->raw_data_in) {
+      (*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->downsample->start_pass) (cinfo);
+      (*cinfo->prep->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+    (*cinfo->fdct->start_pass) (cinfo);
+    (*cinfo->entropy->start_pass) (cinfo, cinfo->optimize_coding);
+    (*cinfo->coef->start_pass) (cinfo,
+				(master->total_passes > 1 ?
+				 JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+    (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    if (cinfo->optimize_coding) {
+      /* No immediate data output; postpone writing frame/scan headers */
+      master->pub.call_pass_startup = FALSE;
+    } else {
+      /* Will write frame/scan headers at first jpeg_write_scanlines call */
+      master->pub.call_pass_startup = TRUE;
+    }
+    break;
+#ifdef ENTROPY_OPT_SUPPORTED
+  case huff_opt_pass:
+    /* Do Huffman optimization for a scan after the first one. */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (cinfo->Ss != 0 || cinfo->Ah == 0 || cinfo->arith_code) {
+      (*cinfo->entropy->start_pass) (cinfo, TRUE);
+      (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+      master->pub.call_pass_startup = FALSE;
+      break;
+    }
+    /* Special case: Huffman DC refinement scans need no Huffman table
+     * and therefore we can skip the optimization pass for them.
+     */
+    master->pass_type = output_pass;
+    master->pass_number++;
+    /*FALLTHROUGH*/
+#endif
+  case output_pass:
+    /* Do a data-output pass. */
+    /* We need not repeat per-scan setup if prior optimization pass did it. */
+    if (! cinfo->optimize_coding) {
+      select_scan_parameters(cinfo);
+      per_scan_setup(cinfo);
+    }
+    (*cinfo->entropy->start_pass) (cinfo, FALSE);
+    (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+    /* We emit frame/scan headers now */
+    if (master->scan_number == 0)
+      (*cinfo->marker->write_frame_header) (cinfo);
+    (*cinfo->marker->write_scan_header) (cinfo);
+    master->pub.call_pass_startup = FALSE;
+    break;
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+  }
+
+  master->pub.is_last_pass = (master->pass_number == master->total_passes-1);
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->total_passes;
+  }
+}
+
+
+/*
+ * Special start-of-pass hook.
+ * This is called by jpeg_write_scanlines if call_pass_startup is TRUE.
+ * In single-pass processing, we need this hook because we don't want to
+ * write frame/scan headers during jpeg_start_compress; we want to let the
+ * application write COM markers etc. between jpeg_start_compress and the
+ * jpeg_write_scanlines loop.
+ * In multi-pass processing, this routine is not used.
+ */
+
+METHODDEF(void)
+pass_startup (j_compress_ptr cinfo)
+{
+  cinfo->master->call_pass_startup = FALSE; /* reset flag so call only once */
+
+  (*cinfo->marker->write_frame_header) (cinfo);
+  (*cinfo->marker->write_scan_header) (cinfo);
+}
+
+
+/*
+ * Finish up at end of pass.
+ */
+
+METHODDEF(void)
+finish_pass_master (j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  /* The entropy coder always needs an end-of-pass call,
+   * either to analyze statistics or to flush its output buffer.
+   */
+  (*cinfo->entropy->finish_pass) (cinfo);
+
+  /* Update state for next pass */
+  switch (master->pass_type) {
+  case main_pass:
+    /* next pass is either output of scan 0 (after optimization)
+     * or output of scan 1 (if no optimization).
+     */
+    master->pass_type = output_pass;
+    if (! cinfo->optimize_coding)
+      master->scan_number++;
+    break;
+  case huff_opt_pass:
+    /* next pass is always output of current scan */
+    master->pass_type = output_pass;
+    break;
+  case output_pass:
+    /* next pass is either optimization or output of next scan */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    master->scan_number++;
+    break;
+  }
+
+  master->pass_number++;
+}
+
+
+/*
+ * Initialize master compression control.
+ */
+
+GLOBAL(void)
+jinit_c_master_control (j_compress_ptr cinfo, boolean transcode_only)
+{
+  my_master_ptr master;
+
+  master = (my_master_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(my_comp_master));
+  cinfo->master = (struct jpeg_comp_master *) master;
+  master->pub.prepare_for_pass = prepare_for_pass;
+  master->pub.pass_startup = pass_startup;
+  master->pub.finish_pass = finish_pass_master;
+  master->pub.is_last_pass = FALSE;
+
+  /* Validate parameters, determine derived values */
+  initial_setup(cinfo, transcode_only);
+
+  if (cinfo->scan_info != NULL) {
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+    validate_script(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    cinfo->num_scans = 1;
+  }
+
+  if (cinfo->progressive_mode)	/*  TEMPORARY HACK ??? */
+    cinfo->optimize_coding = TRUE; /* assume default tables no good for progressive mode */
+
+  /* Initialize my private state */
+  if (transcode_only) {
+    /* no main pass in transcoding */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    else
+      master->pass_type = output_pass;
+  } else {
+    /* for normal compression, first pass is always this type: */
+    master->pass_type = main_pass;
+  }
+  master->scan_number = 0;
+  master->pass_number = 0;
+  if (cinfo->optimize_coding)
+    master->total_passes = cinfo->num_scans * 2;
+  else
+    master->total_passes = cinfo->num_scans;
+}
diff --git a/jcomapi.c b/jcomapi.c
new file mode 100644
index 0000000..9b1fa75
--- /dev/null
+++ b/jcomapi.c
@@ -0,0 +1,106 @@
+/*
+ * jcomapi.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface routines that are used for both
+ * compression and decompression.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Abort processing of a JPEG compression or decompression operation,
+ * but don't destroy the object itself.
+ *
+ * For this, we merely clean up all the nonpermanent memory pools.
+ * Note that temp files (virtual arrays) are not allowed to belong to
+ * the permanent pool, so we will be able to close all temp files here.
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_abort (j_common_ptr cinfo)
+{
+  int pool;
+
+  /* Do nothing if called on a not-initialized or destroyed JPEG object. */
+  if (cinfo->mem == NULL)
+    return;
+
+  /* Releasing pools in reverse order might help avoid fragmentation
+   * with some (brain-damaged) malloc libraries.
+   */
+  for (pool = JPOOL_NUMPOOLS-1; pool > JPOOL_PERMANENT; pool--) {
+    (*cinfo->mem->free_pool) (cinfo, pool);
+  }
+
+  /* Reset overall state for possible reuse of object */
+  if (cinfo->is_decompressor) {
+    cinfo->global_state = DSTATE_START;
+    /* Try to keep application from accessing now-deleted marker list.
+     * A bit kludgy to do it here, but this is the most central place.
+     */
+    ((j_decompress_ptr) cinfo)->marker_list = NULL;
+  } else {
+    cinfo->global_state = CSTATE_START;
+  }
+}
+
+
+/*
+ * Destruction of a JPEG object.
+ *
+ * Everything gets deallocated except the master jpeg_compress_struct itself
+ * and the error manager struct.  Both of these are supplied by the application
+ * and must be freed, if necessary, by the application.  (Often they are on
+ * the stack and so don't need to be freed anyway.)
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_destroy (j_common_ptr cinfo)
+{
+  /* We need only tell the memory manager to release everything. */
+  /* NB: mem pointer is NULL if memory mgr failed to initialize. */
+  if (cinfo->mem != NULL)
+    (*cinfo->mem->self_destruct) (cinfo);
+  cinfo->mem = NULL;		/* be safe if jpeg_destroy is called twice */
+  cinfo->global_state = 0;	/* mark it destroyed */
+}
+
+
+/*
+ * Convenience routines for allocating quantization and Huffman tables.
+ * (Would jutils.c be a more reasonable place to put these?)
+ */
+
+GLOBAL(JQUANT_TBL *)
+jpeg_alloc_quant_table (j_common_ptr cinfo)
+{
+  JQUANT_TBL *tbl;
+
+  tbl = (JQUANT_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, SIZEOF(JQUANT_TBL));
+  tbl->sent_table = FALSE;	/* make sure this is false in any new table */
+  return tbl;
+}
+
+
+GLOBAL(JHUFF_TBL *)
+jpeg_alloc_huff_table (j_common_ptr cinfo)
+{
+  JHUFF_TBL *tbl;
+
+  tbl = (JHUFF_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, SIZEOF(JHUFF_TBL));
+  tbl->sent_table = FALSE;	/* make sure this is false in any new table */
+  return tbl;
+}
diff --git a/jconfig.h.in b/jconfig.h.in
new file mode 100644
index 0000000..c8d0dd7
--- /dev/null
+++ b/jconfig.h.in
@@ -0,0 +1,54 @@
+/* Version ID for the JPEG library.
+ * Might be useful for tests like "#if JPEG_LIB_VERSION >= 60".
+ */
+#define JPEG_LIB_VERSION  62	/* Version 6b */
+
+/* Define if your compiler supports prototypes */
+#undef HAVE_PROTOTYPES
+
+/* Define to 1 if you have the <stddef.h> header file. */
+#undef HAVE_STDDEF_H
+
+/* Define to 1 if you have the <stdlib.h> header file. */
+#undef HAVE_STDLIB_H
+
+/* Define to 1 if the system has the type `unsigned char'. */
+#undef HAVE_UNSIGNED_CHAR
+
+/* Define to 1 if the system has the type `unsigned short'. */
+#undef HAVE_UNSIGNED_SHORT
+
+/* Define if you want use complete types */
+#undef INCOMPLETE_TYPES_BROKEN
+
+/* Define if you have BSD-like bzero and bcopy */
+#undef NEED_BSD_STRINGS
+
+/* Define if you need short function names */
+#undef NEED_SHORT_EXTERNAL_NAMES
+
+/* Define if you have sys/types.h */
+#undef NEED_SYS_TYPES_H
+
+/* Define if shift is unsigned */
+#undef RIGHT_SHIFT_IS_UNSIGNED
+
+/* Use accelerated SIMD routines. */
+#undef WITH_SIMD
+
+/* Define to 1 if type `char' is unsigned and you are not using gcc.  */
+#ifndef __CHAR_UNSIGNED__
+# undef __CHAR_UNSIGNED__
+#endif
+
+/* Define to empty if `const' does not conform to ANSI C. */
+#undef const
+
+/* Define to `__inline__' or `__inline' if that's what the C compiler
+   calls it, or to nothing if 'inline' is not supported under any name.  */
+#ifndef __cplusplus
+#undef inline
+#endif
+
+/* Define to `unsigned int' if <sys/types.h> does not define. */
+#undef size_t
diff --git a/jconfig.txt b/jconfig.txt
new file mode 100644
index 0000000..b96d312
--- /dev/null
+++ b/jconfig.txt
@@ -0,0 +1,164 @@
+/*
+ * jconfig.txt
+ *
+ * Copyright (C) 1991-1994, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file documents the configuration options that are required to
+ * customize the JPEG software for a particular system.
+ *
+ * The actual configuration options for a particular installation are stored
+ * in jconfig.h.  On many machines, jconfig.h can be generated automatically
+ * or copied from one of the "canned" jconfig files that we supply.  But if
+ * you need to generate a jconfig.h file by hand, this file tells you how.
+ *
+ * DO NOT EDIT THIS FILE --- IT WON'T ACCOMPLISH ANYTHING.
+ * EDIT A COPY NAMED JCONFIG.H.
+ */
+
+
+/*
+ * These symbols indicate the properties of your machine or compiler.
+ * #define the symbol if yes, #undef it if no.
+ */
+
+/* Does your compiler support function prototypes?
+ * (If not, you also need to use ansi2knr, see install.txt)
+ */
+#define HAVE_PROTOTYPES
+
+/* Does your compiler support the declaration "unsigned char" ?
+ * How about "unsigned short" ?
+ */
+#define HAVE_UNSIGNED_CHAR
+#define HAVE_UNSIGNED_SHORT
+
+/* Define "void" as "char" if your compiler doesn't know about type void.
+ * NOTE: be sure to define void such that "void *" represents the most general
+ * pointer type, e.g., that returned by malloc().
+ */
+/* #define void char */
+
+/* Define "const" as empty if your compiler doesn't know the "const" keyword.
+ */
+/* #define const */
+
+/* Define this if an ordinary "char" type is unsigned.
+ * If you're not sure, leaving it undefined will work at some cost in speed.
+ * If you defined HAVE_UNSIGNED_CHAR then the speed difference is minimal.
+ */
+#undef CHAR_IS_UNSIGNED
+
+/* Define this if your system has an ANSI-conforming <stddef.h> file.
+ */
+#define HAVE_STDDEF_H
+
+/* Define this if your system has an ANSI-conforming <stdlib.h> file.
+ */
+#define HAVE_STDLIB_H
+
+/* Define this if your system does not have an ANSI/SysV <string.h>,
+ * but does have a BSD-style <strings.h>.
+ */
+#undef NEED_BSD_STRINGS
+
+/* Define this if your system does not provide typedef size_t in any of the
+ * ANSI-standard places (stddef.h, stdlib.h, or stdio.h), but places it in
+ * <sys/types.h> instead.
+ */
+#undef NEED_SYS_TYPES_H
+
+/* For 80x86 machines, you need to define NEED_FAR_POINTERS,
+ * unless you are using a large-data memory model or 80386 flat-memory mode.
+ * On less brain-damaged CPUs this symbol must not be defined.
+ * (Defining this symbol causes large data structures to be referenced through
+ * "far" pointers and to be allocated with a special version of malloc.)
+ */
+#undef NEED_FAR_POINTERS
+
+/* Define this if your linker needs global names to be unique in less
+ * than the first 15 characters.
+ */
+#undef NEED_SHORT_EXTERNAL_NAMES
+
+/* Although a real ANSI C compiler can deal perfectly well with pointers to
+ * unspecified structures (see "incomplete types" in the spec), a few pre-ANSI
+ * and pseudo-ANSI compilers get confused.  To keep one of these bozos happy,
+ * define INCOMPLETE_TYPES_BROKEN.  This is not recommended unless you
+ * actually get "missing structure definition" warnings or errors while
+ * compiling the JPEG code.
+ */
+#undef INCOMPLETE_TYPES_BROKEN
+
+/* Define "boolean" as unsigned char, not int, on Windows systems.
+ */
+#ifdef _WIN32
+#ifndef __RPCNDR_H__		/* don't conflict if rpcndr.h already read */
+typedef unsigned char boolean;
+#endif
+#define HAVE_BOOLEAN		/* prevent jmorecfg.h from redefining it */
+#endif
+
+
+/*
+ * The following options affect code selection within the JPEG library,
+ * but they don't need to be visible to applications using the library.
+ * To minimize application namespace pollution, the symbols won't be
+ * defined unless JPEG_INTERNALS has been defined.
+ */
+
+#ifdef JPEG_INTERNALS
+
+/* Define this if your compiler implements ">>" on signed values as a logical
+ * (unsigned) shift; leave it undefined if ">>" is a signed (arithmetic) shift,
+ * which is the normal and rational definition.
+ */
+#undef RIGHT_SHIFT_IS_UNSIGNED
+
+
+#endif /* JPEG_INTERNALS */
+
+
+/*
+ * The remaining options do not affect the JPEG library proper,
+ * but only the sample applications cjpeg/djpeg (see cjpeg.c, djpeg.c).
+ * Other applications can ignore these.
+ */
+
+#ifdef JPEG_CJPEG_DJPEG
+
+/* These defines indicate which image (non-JPEG) file formats are allowed. */
+
+#define BMP_SUPPORTED		/* BMP image file format */
+#define GIF_SUPPORTED		/* GIF image file format */
+#define PPM_SUPPORTED		/* PBMPLUS PPM/PGM image file format */
+#undef RLE_SUPPORTED		/* Utah RLE image file format */
+#define TARGA_SUPPORTED		/* Targa image file format */
+
+/* Define this if you want to name both input and output files on the command
+ * line, rather than using stdout and optionally stdin.  You MUST do this if
+ * your system can't cope with binary I/O to stdin/stdout.  See comments at
+ * head of cjpeg.c or djpeg.c.
+ */
+#undef TWO_FILE_COMMANDLINE
+
+/* Define this if your system needs explicit cleanup of temporary files.
+ * This is crucial under MS-DOS, where the temporary "files" may be areas
+ * of extended memory; on most other systems it's not as important.
+ */
+#undef NEED_SIGNAL_CATCHER
+
+/* By default, we open image files with fopen(...,"rb") or fopen(...,"wb").
+ * This is necessary on systems that distinguish text files from binary files,
+ * and is harmless on most systems that don't.  If you have one of the rare
+ * systems that complains about the "b" spec, define this symbol.
+ */
+#undef DONT_USE_B_MODE
+
+/* Define this if you want percent-done progress reports from cjpeg/djpeg.
+ */
+#undef PROGRESS_REPORT
+
+
+#endif /* JPEG_CJPEG_DJPEG */
diff --git a/jcparam.c b/jcparam.c
new file mode 100644
index 0000000..27b5a03
--- /dev/null
+++ b/jcparam.c
@@ -0,0 +1,645 @@
+/*
+ * jcparam.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2003-2008 by Guido Vollbeding.
+ * Copyright (C) 2009-2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains optional default-setting code for the JPEG compressor.
+ * Applications do not have to use this file, but those that don't use it
+ * must know a lot more about the innards of the JPEG code.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Quantization table setup routines
+ */
+
+GLOBAL(void)
+jpeg_add_quant_table (j_compress_ptr cinfo, int which_tbl,
+		      const unsigned int *basic_table,
+		      int scale_factor, boolean force_baseline)
+/* Define a quantization table equal to the basic_table times
+ * a scale factor (given as a percentage).
+ * If force_baseline is TRUE, the computed quantization table entries
+ * are limited to 1..255 for JPEG baseline compatibility.
+ */
+{
+  JQUANT_TBL ** qtblptr;
+  int i;
+  long temp;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (which_tbl < 0 || which_tbl >= NUM_QUANT_TBLS)
+    ERREXIT1(cinfo, JERR_DQT_INDEX, which_tbl);
+
+  qtblptr = & cinfo->quant_tbl_ptrs[which_tbl];
+
+  if (*qtblptr == NULL)
+    *qtblptr = jpeg_alloc_quant_table((j_common_ptr) cinfo);
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    temp = ((long) basic_table[i] * scale_factor + 50L) / 100L;
+    /* limit the values to the valid range */
+    if (temp <= 0L) temp = 1L;
+    if (temp > 32767L) temp = 32767L; /* max quantizer needed for 12 bits */
+    if (force_baseline && temp > 255L)
+      temp = 255L;		/* limit to baseline range if requested */
+    (*qtblptr)->quantval[i] = (UINT16) temp;
+  }
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*qtblptr)->sent_table = FALSE;
+}
+
+
+/* These are the sample quantization tables given in JPEG spec section K.1.
+ * The spec says that the values given produce "good" quality, and
+ * when divided by 2, "very good" quality.
+ */
+static const unsigned int std_luminance_quant_tbl[DCTSIZE2] = {
+  16,  11,  10,  16,  24,  40,  51,  61,
+  12,  12,  14,  19,  26,  58,  60,  55,
+  14,  13,  16,  24,  40,  57,  69,  56,
+  14,  17,  22,  29,  51,  87,  80,  62,
+  18,  22,  37,  56,  68, 109, 103,  77,
+  24,  35,  55,  64,  81, 104, 113,  92,
+  49,  64,  78,  87, 103, 121, 120, 101,
+  72,  92,  95,  98, 112, 100, 103,  99
+};
+static const unsigned int std_chrominance_quant_tbl[DCTSIZE2] = {
+  17,  18,  24,  47,  99,  99,  99,  99,
+  18,  21,  26,  66,  99,  99,  99,  99,
+  24,  26,  56,  99,  99,  99,  99,  99,
+  47,  66,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99
+};
+
+
+#if JPEG_LIB_VERSION >= 70
+GLOBAL(void)
+jpeg_default_qtables (j_compress_ptr cinfo, boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and straight percentage-scaling quality scales.
+ * This entry point allows different scalings for luminance and chrominance.
+ */
+{
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+		       cinfo->q_scale_factor[0], force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+		       cinfo->q_scale_factor[1], force_baseline);
+}
+#endif
+
+
+GLOBAL(void)
+jpeg_set_linear_quality (j_compress_ptr cinfo, int scale_factor,
+			 boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and a straight percentage-scaling quality scale.  In most cases it's better
+ * to use jpeg_set_quality (below); this entry point is provided for
+ * applications that insist on a linear percentage scaling.
+ */
+{
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+		       scale_factor, force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+		       scale_factor, force_baseline);
+}
+
+
+GLOBAL(int)
+jpeg_quality_scaling (int quality)
+/* Convert a user-specified quality rating to a percentage scaling factor
+ * for an underlying quantization table, using our recommended scaling curve.
+ * The input 'quality' factor should be 0 (terrible) to 100 (very good).
+ */
+{
+  /* Safety limit on quality factor.  Convert 0 to 1 to avoid zero divide. */
+  if (quality <= 0) quality = 1;
+  if (quality > 100) quality = 100;
+
+  /* The basic table is used as-is (scaling 100) for a quality of 50.
+   * Qualities 50..100 are converted to scaling percentage 200 - 2*Q;
+   * note that at Q=100 the scaling is 0, which will cause jpeg_add_quant_table
+   * to make all the table entries 1 (hence, minimum quantization loss).
+   * Qualities 1..50 are converted to scaling percentage 5000/Q.
+   */
+  if (quality < 50)
+    quality = 5000 / quality;
+  else
+    quality = 200 - quality*2;
+
+  return quality;
+}
+
+
+GLOBAL(void)
+jpeg_set_quality (j_compress_ptr cinfo, int quality, boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables.
+ * This is the standard quality-adjusting entry point for typical user
+ * interfaces; only those who want detailed control over quantization tables
+ * would use the preceding three routines directly.
+ */
+{
+  /* Convert user 0-100 rating to percentage scaling */
+  quality = jpeg_quality_scaling(quality);
+
+  /* Set up standard quality tables */
+  jpeg_set_linear_quality(cinfo, quality, force_baseline);
+}
+
+
+/*
+ * Huffman table setup routines
+ */
+
+LOCAL(void)
+add_huff_table (j_compress_ptr cinfo,
+		JHUFF_TBL **htblptr, const UINT8 *bits, const UINT8 *val)
+/* Define a Huffman table */
+{
+  int nsymbols, len;
+
+  if (*htblptr == NULL)
+    *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+
+  /* Copy the number-of-symbols-of-each-code-length counts */
+  MEMCOPY((*htblptr)->bits, bits, SIZEOF((*htblptr)->bits));
+
+  /* Validate the counts.  We do this here mainly so we can copy the right
+   * number of symbols from the val[] array, without risking marching off
+   * the end of memory.  jchuff.c will do a more thorough test later.
+   */
+  nsymbols = 0;
+  for (len = 1; len <= 16; len++)
+    nsymbols += bits[len];
+  if (nsymbols < 1 || nsymbols > 256)
+    ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+  MEMCOPY((*htblptr)->huffval, val, nsymbols * SIZEOF(UINT8));
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*htblptr)->sent_table = FALSE;
+}
+
+
+LOCAL(void)
+std_huff_tables (j_compress_ptr cinfo)
+/* Set up the standard Huffman tables (cf. JPEG standard section K.3) */
+/* IMPORTANT: these are only valid for 8-bit data precision! */
+{
+  static const UINT8 bits_dc_luminance[17] =
+    { /* 0-base */ 0, 0, 1, 5, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0 };
+  static const UINT8 val_dc_luminance[] =
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 };
+  
+  static const UINT8 bits_dc_chrominance[17] =
+    { /* 0-base */ 0, 0, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
+  static const UINT8 val_dc_chrominance[] =
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 };
+  
+  static const UINT8 bits_ac_luminance[17] =
+    { /* 0-base */ 0, 0, 2, 1, 3, 3, 2, 4, 3, 5, 5, 4, 4, 0, 0, 1, 0x7d };
+  static const UINT8 val_ac_luminance[] =
+    { 0x01, 0x02, 0x03, 0x00, 0x04, 0x11, 0x05, 0x12,
+      0x21, 0x31, 0x41, 0x06, 0x13, 0x51, 0x61, 0x07,
+      0x22, 0x71, 0x14, 0x32, 0x81, 0x91, 0xa1, 0x08,
+      0x23, 0x42, 0xb1, 0xc1, 0x15, 0x52, 0xd1, 0xf0,
+      0x24, 0x33, 0x62, 0x72, 0x82, 0x09, 0x0a, 0x16,
+      0x17, 0x18, 0x19, 0x1a, 0x25, 0x26, 0x27, 0x28,
+      0x29, 0x2a, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39,
+      0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48, 0x49,
+      0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59,
+      0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69,
+      0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79,
+      0x7a, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89,
+      0x8a, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98,
+      0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7,
+      0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6,
+      0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3, 0xc4, 0xc5,
+      0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2, 0xd3, 0xd4,
+      0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xe1, 0xe2,
+      0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea,
+      0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+      0xf9, 0xfa };
+  
+  static const UINT8 bits_ac_chrominance[17] =
+    { /* 0-base */ 0, 0, 2, 1, 2, 4, 4, 3, 4, 7, 5, 4, 4, 0, 1, 2, 0x77 };
+  static const UINT8 val_ac_chrominance[] =
+    { 0x00, 0x01, 0x02, 0x03, 0x11, 0x04, 0x05, 0x21,
+      0x31, 0x06, 0x12, 0x41, 0x51, 0x07, 0x61, 0x71,
+      0x13, 0x22, 0x32, 0x81, 0x08, 0x14, 0x42, 0x91,
+      0xa1, 0xb1, 0xc1, 0x09, 0x23, 0x33, 0x52, 0xf0,
+      0x15, 0x62, 0x72, 0xd1, 0x0a, 0x16, 0x24, 0x34,
+      0xe1, 0x25, 0xf1, 0x17, 0x18, 0x19, 0x1a, 0x26,
+      0x27, 0x28, 0x29, 0x2a, 0x35, 0x36, 0x37, 0x38,
+      0x39, 0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48,
+      0x49, 0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58,
+      0x59, 0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
+      0x69, 0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78,
+      0x79, 0x7a, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+      0x88, 0x89, 0x8a, 0x92, 0x93, 0x94, 0x95, 0x96,
+      0x97, 0x98, 0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5,
+      0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4,
+      0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3,
+      0xc4, 0xc5, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2,
+      0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda,
+      0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9,
+      0xea, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+      0xf9, 0xfa };
+  
+  add_huff_table(cinfo, &cinfo->dc_huff_tbl_ptrs[0],
+		 bits_dc_luminance, val_dc_luminance);
+  add_huff_table(cinfo, &cinfo->ac_huff_tbl_ptrs[0],
+		 bits_ac_luminance, val_ac_luminance);
+  add_huff_table(cinfo, &cinfo->dc_huff_tbl_ptrs[1],
+		 bits_dc_chrominance, val_dc_chrominance);
+  add_huff_table(cinfo, &cinfo->ac_huff_tbl_ptrs[1],
+		 bits_ac_chrominance, val_ac_chrominance);
+}
+
+
+/*
+ * Default parameter setup for compression.
+ *
+ * Applications that don't choose to use this routine must do their
+ * own setup of all these parameters.  Alternately, you can call this
+ * to establish defaults and then alter parameters selectively.  This
+ * is the recommended approach since, if we add any new parameters,
+ * your code will still work (they'll be set to reasonable defaults).
+ */
+
+GLOBAL(void)
+jpeg_set_defaults (j_compress_ptr cinfo)
+{
+  int i;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Allocate comp_info array large enough for maximum component count.
+   * Array is made permanent in case application wants to compress
+   * multiple images at same param settings.
+   */
+  if (cinfo->comp_info == NULL)
+    cinfo->comp_info = (jpeg_component_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  MAX_COMPONENTS * SIZEOF(jpeg_component_info));
+
+  /* Initialize everything not dependent on the color space */
+
+#if JPEG_LIB_VERSION >= 70
+  cinfo->scale_num = 1;		/* 1:1 scaling */
+  cinfo->scale_denom = 1;
+#endif
+  cinfo->data_precision = BITS_IN_JSAMPLE;
+  /* Set up two quantization tables using default quality of 75 */
+  jpeg_set_quality(cinfo, 75, TRUE);
+  /* Set up two Huffman tables */
+  std_huff_tables(cinfo);
+
+  /* Initialize default arithmetic coding conditioning */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+
+  /* Default is no multiple-scan output */
+  cinfo->scan_info = NULL;
+  cinfo->num_scans = 0;
+
+  /* Expect normal source image, not raw downsampled data */
+  cinfo->raw_data_in = FALSE;
+
+  /* Use Huffman coding, not arithmetic coding, by default */
+  cinfo->arith_code = FALSE;
+
+  /* By default, don't do extra passes to optimize entropy coding */
+  cinfo->optimize_coding = FALSE;
+  /* The standard Huffman tables are only valid for 8-bit data precision.
+   * If the precision is higher, force optimization on so that usable
+   * tables will be computed.  This test can be removed if default tables
+   * are supplied that are valid for the desired precision.
+   */
+  if (cinfo->data_precision > 8)
+    cinfo->optimize_coding = TRUE;
+
+  /* By default, use the simpler non-cosited sampling alignment */
+  cinfo->CCIR601_sampling = FALSE;
+
+#if JPEG_LIB_VERSION >= 70
+  /* By default, apply fancy downsampling */
+  cinfo->do_fancy_downsampling = TRUE;
+#endif
+
+  /* No input smoothing */
+  cinfo->smoothing_factor = 0;
+
+  /* DCT algorithm preference */
+  cinfo->dct_method = JDCT_DEFAULT;
+
+  /* No restart markers */
+  cinfo->restart_interval = 0;
+  cinfo->restart_in_rows = 0;
+
+  /* Fill in default JFIF marker parameters.  Note that whether the marker
+   * will actually be written is determined by jpeg_set_colorspace.
+   *
+   * By default, the library emits JFIF version code 1.01.
+   * An application that wants to emit JFIF 1.02 extension markers should set
+   * JFIF_minor_version to 2.  We could probably get away with just defaulting
+   * to 1.02, but there may still be some decoders in use that will complain
+   * about that; saying 1.01 should minimize compatibility problems.
+   */
+  cinfo->JFIF_major_version = 1; /* Default JFIF version = 1.01 */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;	/* Pixel size is unknown by default */
+  cinfo->X_density = 1;		/* Pixel aspect ratio is square by default */
+  cinfo->Y_density = 1;
+
+  /* Choose JPEG colorspace based on input space, set defaults accordingly */
+
+  jpeg_default_colorspace(cinfo);
+}
+
+
+/*
+ * Select an appropriate JPEG colorspace for in_color_space.
+ */
+
+GLOBAL(void)
+jpeg_default_colorspace (j_compress_ptr cinfo)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    jpeg_set_colorspace(cinfo, JCS_GRAYSCALE);
+    break;
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_YCbCr:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_CMYK:
+    jpeg_set_colorspace(cinfo, JCS_CMYK); /* By default, no translation */
+    break;
+  case JCS_YCCK:
+    jpeg_set_colorspace(cinfo, JCS_YCCK);
+    break;
+  case JCS_UNKNOWN:
+    jpeg_set_colorspace(cinfo, JCS_UNKNOWN);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+  }
+}
+
+
+/*
+ * Set the JPEG colorspace, and choose colorspace-dependent default values.
+ */
+
+GLOBAL(void)
+jpeg_set_colorspace (j_compress_ptr cinfo, J_COLOR_SPACE colorspace)
+{
+  jpeg_component_info * compptr;
+  int ci;
+
+#define SET_COMP(index,id,hsamp,vsamp,quant,dctbl,actbl)  \
+  (compptr = &cinfo->comp_info[index], \
+   compptr->component_id = (id), \
+   compptr->h_samp_factor = (hsamp), \
+   compptr->v_samp_factor = (vsamp), \
+   compptr->quant_tbl_no = (quant), \
+   compptr->dc_tbl_no = (dctbl), \
+   compptr->ac_tbl_no = (actbl) )
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* For all colorspaces, we use Q and Huff tables 0 for luminance components,
+   * tables 1 for chrominance components.
+   */
+
+  cinfo->jpeg_color_space = colorspace;
+
+  cinfo->write_JFIF_header = FALSE; /* No marker for non-JFIF colorspaces */
+  cinfo->write_Adobe_marker = FALSE; /* write no Adobe marker by default */
+
+  switch (colorspace) {
+  case JCS_GRAYSCALE:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 1;
+    /* JFIF specifies component ID 1 */
+    SET_COMP(0, 1, 1,1, 0, 0,0);
+    break;
+  case JCS_RGB:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag RGB */
+    cinfo->num_components = 3;
+    SET_COMP(0, 0x52 /* 'R' */, 1,1, 0, 0,0);
+    SET_COMP(1, 0x47 /* 'G' */, 1,1, 0, 0,0);
+    SET_COMP(2, 0x42 /* 'B' */, 1,1, 0, 0,0);
+    break;
+  case JCS_YCbCr:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 3;
+    /* JFIF specifies component IDs 1,2,3 */
+    /* We default to 2x2 subsamples of chrominance */
+    SET_COMP(0, 1, 2,2, 0, 0,0);
+    SET_COMP(1, 2, 1,1, 1, 1,1);
+    SET_COMP(2, 3, 1,1, 1, 1,1);
+    break;
+  case JCS_CMYK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag CMYK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 0x43 /* 'C' */, 1,1, 0, 0,0);
+    SET_COMP(1, 0x4D /* 'M' */, 1,1, 0, 0,0);
+    SET_COMP(2, 0x59 /* 'Y' */, 1,1, 0, 0,0);
+    SET_COMP(3, 0x4B /* 'K' */, 1,1, 0, 0,0);
+    break;
+  case JCS_YCCK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag YCCK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 1, 2,2, 0, 0,0);
+    SET_COMP(1, 2, 1,1, 1, 1,1);
+    SET_COMP(2, 3, 1,1, 1, 1,1);
+    SET_COMP(3, 4, 2,2, 0, 0,0);
+    break;
+  case JCS_UNKNOWN:
+    cinfo->num_components = cinfo->input_components;
+    if (cinfo->num_components < 1 || cinfo->num_components > MAX_COMPONENTS)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	       MAX_COMPONENTS);
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      SET_COMP(ci, ci, 1,1, 0, 0,0);
+    }
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+  }
+}
+
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+LOCAL(jpeg_scan_info *)
+fill_a_scan (jpeg_scan_info * scanptr, int ci,
+	     int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for specified component */
+{
+  scanptr->comps_in_scan = 1;
+  scanptr->component_index[0] = ci;
+  scanptr->Ss = Ss;
+  scanptr->Se = Se;
+  scanptr->Ah = Ah;
+  scanptr->Al = Al;
+  scanptr++;
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_scans (jpeg_scan_info * scanptr, int ncomps,
+	    int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for each component */
+{
+  int ci;
+
+  for (ci = 0; ci < ncomps; ci++) {
+    scanptr->comps_in_scan = 1;
+    scanptr->component_index[0] = ci;
+    scanptr->Ss = Ss;
+    scanptr->Se = Se;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  }
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_dc_scans (jpeg_scan_info * scanptr, int ncomps, int Ah, int Al)
+/* Support routine: generate interleaved DC scan if possible, else N scans */
+{
+  int ci;
+
+  if (ncomps <= MAX_COMPS_IN_SCAN) {
+    /* Single interleaved DC scan */
+    scanptr->comps_in_scan = ncomps;
+    for (ci = 0; ci < ncomps; ci++)
+      scanptr->component_index[ci] = ci;
+    scanptr->Ss = scanptr->Se = 0;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  } else {
+    /* Noninterleaved DC scan for each component */
+    scanptr = fill_scans(scanptr, ncomps, 0, 0, Ah, Al);
+  }
+  return scanptr;
+}
+
+
+/*
+ * Create a recommended progressive-JPEG script.
+ * cinfo->num_components and cinfo->jpeg_color_space must be correct.
+ */
+
+GLOBAL(void)
+jpeg_simple_progression (j_compress_ptr cinfo)
+{
+  int ncomps = cinfo->num_components;
+  int nscans;
+  jpeg_scan_info * scanptr;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Figure space needed for script.  Calculation must match code below! */
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    nscans = 10;
+  } else {
+    /* All-purpose script for other color spaces. */
+    if (ncomps > MAX_COMPS_IN_SCAN)
+      nscans = 6 * ncomps;	/* 2 DC + 4 AC scans per component */
+    else
+      nscans = 2 + 4 * ncomps;	/* 2 DC scans; 4 AC scans per component */
+  }
+
+  /* Allocate space for script.
+   * We need to put it in the permanent pool in case the application performs
+   * multiple compressions without changing the settings.  To avoid a memory
+   * leak if jpeg_simple_progression is called repeatedly for the same JPEG
+   * object, we try to re-use previously allocated space, and we allocate
+   * enough space to handle YCbCr even if initially asked for grayscale.
+   */
+  if (cinfo->script_space == NULL || cinfo->script_space_size < nscans) {
+    cinfo->script_space_size = MAX(nscans, 10);
+    cinfo->script_space = (jpeg_scan_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+			cinfo->script_space_size * SIZEOF(jpeg_scan_info));
+  }
+  scanptr = cinfo->script_space;
+  cinfo->scan_info = scanptr;
+  cinfo->num_scans = nscans;
+
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    /* Initial DC scan */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    /* Initial AC scan: get some luma data out in a hurry */
+    scanptr = fill_a_scan(scanptr, 0, 1, 5, 0, 2);
+    /* Chroma data is too small to be worth expending many scans on */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 0, 1);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 0, 1);
+    /* Complete spectral selection for luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 6, 63, 0, 2);
+    /* Refine next bit of luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 2, 1);
+    /* Finish DC successive approximation */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    /* Finish AC successive approximation */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 1, 0);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 1, 0);
+    /* Luma bottom bit comes last since it's usually largest scan */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 1, 0);
+  } else {
+    /* All-purpose script for other color spaces. */
+    /* Successive approximation first pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    scanptr = fill_scans(scanptr, ncomps, 1, 5, 0, 2);
+    scanptr = fill_scans(scanptr, ncomps, 6, 63, 0, 2);
+    /* Successive approximation second pass */
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 2, 1);
+    /* Successive approximation final pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 1, 0);
+  }
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */
diff --git a/jcphuff.c b/jcphuff.c
new file mode 100644
index 0000000..3102871
--- /dev/null
+++ b/jcphuff.c
@@ -0,0 +1,831 @@
+/*
+ * jcphuff.c
+ *
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy encoding routines for progressive JPEG.
+ *
+ * We do not support output suspension in this module, since the library
+ * currently does not allow multiple-scan files to be written with output
+ * suspension.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jchuff.h"		/* Declarations shared with jchuff.c */
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+/* Expanded entropy encoder object for progressive Huffman encoding. */
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  /* Mode flag: TRUE for optimization, FALSE for actual data output */
+  boolean gather_statistics;
+
+  /* Bit-level coding status.
+   * next_output_byte/free_in_buffer are local copies of cinfo->dest fields.
+   */
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+  INT32 put_buffer;		/* current bit-accumulation buffer */
+  int put_bits;			/* # of bits now in it */
+  j_compress_ptr cinfo;		/* link to cinfo (needed for dump_buffer) */
+
+  /* Coding status for DC components */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+
+  /* Coding status for AC components */
+  int ac_tbl_no;		/* the table number of the single component */
+  unsigned int EOBRUN;		/* run length of EOBs */
+  unsigned int BE;		/* # of buffered correction bits before MCU */
+  char * bit_buffer;		/* buffer for correction bits (1 per char) */
+  /* packing correction bits tightly would save some space but cost time... */
+
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan).
+   * Since any one scan codes only DC or only AC, we only need one set
+   * of tables, not one for DC and one for AC.
+   */
+  c_derived_tbl * derived_tbls[NUM_HUFF_TBLS];
+
+  /* Statistics tables for optimization; again, one set is enough */
+  long * count_ptrs[NUM_HUFF_TBLS];
+} phuff_entropy_encoder;
+
+typedef phuff_entropy_encoder * phuff_entropy_ptr;
+
+/* MAX_CORR_BITS is the number of bits the AC refinement correction-bit
+ * buffer can hold.  Larger sizes may slightly improve compression, but
+ * 1000 is already well into the realm of overkill.
+ * The minimum safe size is 64 bits.
+ */
+
+#define MAX_CORR_BITS  1000	/* Max # of correction bits I can buffer */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than INT32.
+ * We assume that int right shift is unsigned if INT32 right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	int ishift_temp;
+#define IRIGHT_SHIFT(x,shft)  \
+	((ishift_temp = (x)) < 0 ? \
+	 (ishift_temp >> (shft)) | ((~0) << (16-(shft))) : \
+	 (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+/* Forward declarations */
+METHODDEF(boolean) encode_mcu_DC_first JPP((j_compress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(boolean) encode_mcu_AC_first JPP((j_compress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(boolean) encode_mcu_DC_refine JPP((j_compress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+METHODDEF(boolean) encode_mcu_AC_refine JPP((j_compress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+METHODDEF(void) finish_pass_phuff JPP((j_compress_ptr cinfo));
+METHODDEF(void) finish_pass_gather_phuff JPP((j_compress_ptr cinfo));
+
+
+/*
+ * Initialize for a Huffman-compressed scan using progressive JPEG.
+ */
+
+METHODDEF(void)
+start_pass_phuff (j_compress_ptr cinfo, boolean gather_statistics)
+{  
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  boolean is_DC_band;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  entropy->cinfo = cinfo;
+  entropy->gather_statistics = gather_statistics;
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* We assume jcmaster.c already validated the scan parameters. */
+
+  /* Select execution routines */
+  if (cinfo->Ah == 0) {
+    if (is_DC_band)
+      entropy->pub.encode_mcu = encode_mcu_DC_first;
+    else
+      entropy->pub.encode_mcu = encode_mcu_AC_first;
+  } else {
+    if (is_DC_band)
+      entropy->pub.encode_mcu = encode_mcu_DC_refine;
+    else {
+      entropy->pub.encode_mcu = encode_mcu_AC_refine;
+      /* AC refinement needs a correction bit buffer */
+      if (entropy->bit_buffer == NULL)
+	entropy->bit_buffer = (char *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      MAX_CORR_BITS * SIZEOF(char));
+    }
+  }
+  if (gather_statistics)
+    entropy->pub.finish_pass = finish_pass_gather_phuff;
+  else
+    entropy->pub.finish_pass = finish_pass_phuff;
+
+  /* Only DC coefficients may be interleaved, so cinfo->comps_in_scan = 1
+   * for AC coefficients.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Initialize DC predictions to 0 */
+    entropy->last_dc_val[ci] = 0;
+    /* Get table index */
+    if (is_DC_band) {
+      if (cinfo->Ah != 0)	/* DC refinement needs no table */
+	continue;
+      tbl = compptr->dc_tbl_no;
+    } else {
+      entropy->ac_tbl_no = tbl = compptr->ac_tbl_no;
+    }
+    if (gather_statistics) {
+      /* Check for invalid table index */
+      /* (make_c_derived_tbl does this in the other path) */
+      if (tbl < 0 || tbl >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tbl);
+      /* Allocate and zero the statistics tables */
+      /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+      if (entropy->count_ptrs[tbl] == NULL)
+	entropy->count_ptrs[tbl] = (long *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      257 * SIZEOF(long));
+      MEMZERO(entropy->count_ptrs[tbl], 257 * SIZEOF(long));
+    } else {
+      /* Compute derived values for Huffman table */
+      /* We may do this more than once for a table, but it's not expensive */
+      jpeg_make_c_derived_tbl(cinfo, is_DC_band, tbl,
+			      & entropy->derived_tbls[tbl]);
+    }
+  }
+
+  /* Initialize AC stuff */
+  entropy->EOBRUN = 0;
+  entropy->BE = 0;
+
+  /* Initialize bit buffer to empty */
+  entropy->put_buffer = 0;
+  entropy->put_bits = 0;
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/* Outputting bytes to the file.
+ * NB: these must be called only when actually outputting,
+ * that is, entropy->gather_statistics == FALSE.
+ */
+
+/* Emit a byte */
+#define emit_byte(entropy,val)  \
+	{ *(entropy)->next_output_byte++ = (JOCTET) (val);  \
+	  if (--(entropy)->free_in_buffer == 0)  \
+	    dump_buffer(entropy); }
+
+
+LOCAL(void)
+dump_buffer (phuff_entropy_ptr entropy)
+/* Empty the output buffer; we do not support suspension in this module. */
+{
+  struct jpeg_destination_mgr * dest = entropy->cinfo->dest;
+
+  if (! (*dest->empty_output_buffer) (entropy->cinfo))
+    ERREXIT(entropy->cinfo, JERR_CANT_SUSPEND);
+  /* After a successful buffer dump, must reset buffer pointers */
+  entropy->next_output_byte = dest->next_output_byte;
+  entropy->free_in_buffer = dest->free_in_buffer;
+}
+
+
+/* Outputting bits to the file */
+
+/* Only the right 24 bits of put_buffer are used; the valid bits are
+ * left-justified in this part.  At most 16 bits can be passed to emit_bits
+ * in one call, and we never retain more than 7 bits in put_buffer
+ * between calls, so 24 bits are sufficient.
+ */
+
+LOCAL(void)
+emit_bits (phuff_entropy_ptr entropy, unsigned int code, int size)
+/* Emit some bits, unless we are in gather mode */
+{
+  /* This routine is heavily used, so it's worth coding tightly. */
+  register INT32 put_buffer = (INT32) code;
+  register int put_bits = entropy->put_bits;
+
+  /* if size is 0, caller used an invalid Huffman table entry */
+  if (size == 0)
+    ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+  if (entropy->gather_statistics)
+    return;			/* do nothing if we're only getting stats */
+
+  put_buffer &= (((INT32) 1)<<size) - 1; /* mask off any extra bits in code */
+  
+  put_bits += size;		/* new number of bits in buffer */
+  
+  put_buffer <<= 24 - put_bits; /* align incoming bits */
+
+  put_buffer |= entropy->put_buffer; /* and merge with old buffer contents */
+
+  while (put_bits >= 8) {
+    int c = (int) ((put_buffer >> 16) & 0xFF);
+    
+    emit_byte(entropy, c);
+    if (c == 0xFF) {		/* need to stuff a zero byte? */
+      emit_byte(entropy, 0);
+    }
+    put_buffer <<= 8;
+    put_bits -= 8;
+  }
+
+  entropy->put_buffer = put_buffer; /* update variables */
+  entropy->put_bits = put_bits;
+}
+
+
+LOCAL(void)
+flush_bits (phuff_entropy_ptr entropy)
+{
+  emit_bits(entropy, 0x7F, 7); /* fill any partial byte with ones */
+  entropy->put_buffer = 0;     /* and reset bit-buffer to empty */
+  entropy->put_bits = 0;
+}
+
+
+/*
+ * Emit (or just count) a Huffman symbol.
+ */
+
+LOCAL(void)
+emit_symbol (phuff_entropy_ptr entropy, int tbl_no, int symbol)
+{
+  if (entropy->gather_statistics)
+    entropy->count_ptrs[tbl_no][symbol]++;
+  else {
+    c_derived_tbl * tbl = entropy->derived_tbls[tbl_no];
+    emit_bits(entropy, tbl->ehufco[symbol], tbl->ehufsi[symbol]);
+  }
+}
+
+
+/*
+ * Emit bits from a correction bit buffer.
+ */
+
+LOCAL(void)
+emit_buffered_bits (phuff_entropy_ptr entropy, char * bufstart,
+		    unsigned int nbits)
+{
+  if (entropy->gather_statistics)
+    return;			/* no real work */
+
+  while (nbits > 0) {
+    emit_bits(entropy, (unsigned int) (*bufstart), 1);
+    bufstart++;
+    nbits--;
+  }
+}
+
+
+/*
+ * Emit any pending EOBRUN symbol.
+ */
+
+LOCAL(void)
+emit_eobrun (phuff_entropy_ptr entropy)
+{
+  register int temp, nbits;
+
+  if (entropy->EOBRUN > 0) {	/* if there is any pending EOBRUN */
+    temp = entropy->EOBRUN;
+    nbits = 0;
+    while ((temp >>= 1))
+      nbits++;
+    /* safety check: shouldn't happen given limited correction-bit buffer */
+    if (nbits > 14)
+      ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+    emit_symbol(entropy, entropy->ac_tbl_no, nbits << 4);
+    if (nbits)
+      emit_bits(entropy, entropy->EOBRUN, nbits);
+
+    entropy->EOBRUN = 0;
+
+    /* Emit any buffered correction bits */
+    emit_buffered_bits(entropy, entropy->bit_buffer, entropy->BE);
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(void)
+emit_restart (phuff_entropy_ptr entropy, int restart_num)
+{
+  int ci;
+
+  emit_eobrun(entropy);
+
+  if (! entropy->gather_statistics) {
+    flush_bits(entropy);
+    emit_byte(entropy, 0xFF);
+    emit_byte(entropy, JPEG_RST0 + restart_num);
+  }
+
+  if (entropy->cinfo->Ss == 0) {
+    /* Re-initialize DC predictions to 0 */
+    for (ci = 0; ci < entropy->cinfo->comps_in_scan; ci++)
+      entropy->last_dc_val[ci] = 0;
+  } else {
+    /* Re-initialize all AC-related fields to 0 */
+    entropy->EOBRUN = 0;
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  register int temp, temp2;
+  register int nbits;
+  int blkn, ci;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+  jpeg_component_info * compptr;
+  ISHIFT_TEMPS
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    temp2 = IRIGHT_SHIFT((int) ((*block)[0]), Al);
+
+    /* DC differences are figured on the point-transformed values. */
+    temp = temp2 - entropy->last_dc_val[ci];
+    entropy->last_dc_val[ci] = temp2;
+
+    /* Encode the DC coefficient difference per section G.1.2.1 */
+    temp2 = temp;
+    if (temp < 0) {
+      temp = -temp;		/* temp is abs value of input */
+      /* For a negative input, want temp2 = bitwise complement of abs(input) */
+      /* This code assumes we are on a two's complement machine */
+      temp2--;
+    }
+    
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = 0;
+    while (temp) {
+      nbits++;
+      temp >>= 1;
+    }
+    /* Check for out-of-range coefficient values.
+     * Since we're encoding a difference, the range limit is twice as much.
+     */
+    if (nbits > MAX_COEF_BITS+1)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+    
+    /* Count/emit the Huffman-coded symbol for the number of bits */
+    emit_symbol(entropy, compptr->dc_tbl_no, nbits);
+    
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    if (nbits)			/* emit_bits rejects calls with size 0 */
+      emit_bits(entropy, (unsigned int) temp2, nbits);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  register int temp, temp2;
+  register int nbits;
+  register int r, k;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+
+  /* Encode the AC coefficients per section G.1.2.2, fig. G.3 */
+  
+  r = 0;			/* r = run length of zeros */
+   
+  for (k = cinfo->Ss; k <= Se; k++) {
+    if ((temp = (*block)[jpeg_natural_order[k]]) == 0) {
+      r++;
+      continue;
+    }
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value; so the code is
+     * interwoven with finding the abs value (temp) and output bits (temp2).
+     */
+    if (temp < 0) {
+      temp = -temp;		/* temp is abs value of input */
+      temp >>= Al;		/* apply the point transform */
+      /* For a negative coef, want temp2 = bitwise complement of abs(coef) */
+      temp2 = ~temp;
+    } else {
+      temp >>= Al;		/* apply the point transform */
+      temp2 = temp;
+    }
+    /* Watch out for case that nonzero coef is zero after point transform */
+    if (temp == 0) {
+      r++;
+      continue;
+    }
+
+    /* Emit any pending EOBRUN */
+    if (entropy->EOBRUN > 0)
+      emit_eobrun(entropy);
+    /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+    while (r > 15) {
+      emit_symbol(entropy, entropy->ac_tbl_no, 0xF0);
+      r -= 16;
+    }
+
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = 1;			/* there must be at least one 1 bit */
+    while ((temp >>= 1))
+      nbits++;
+    /* Check for out-of-range coefficient values */
+    if (nbits > MAX_COEF_BITS)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+    /* Count/emit Huffman symbol for run length / number of bits */
+    emit_symbol(entropy, entropy->ac_tbl_no, (r << 4) + nbits);
+
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    emit_bits(entropy, (unsigned int) temp2, nbits);
+
+    r = 0;			/* reset zero run length */
+  }
+
+  if (r > 0) {			/* If there are trailing zeroes, */
+    entropy->EOBRUN++;		/* count an EOB */
+    if (entropy->EOBRUN == 0x7FFF)
+      emit_eobrun(entropy);	/* force it out to avoid overflow */
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  register int temp;
+  int blkn;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    temp = (*block)[0];
+    emit_bits(entropy, (unsigned int) (temp >> Al), 1);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  register int temp;
+  register int r, k;
+  int EOB;
+  char *BR_buffer;
+  unsigned int BR;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+  int absvalues[DCTSIZE2];
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+
+  /* It is convenient to make a pre-pass to determine the transformed
+   * coefficients' absolute values and the EOB position.
+   */
+  EOB = 0;
+  for (k = cinfo->Ss; k <= Se; k++) {
+    temp = (*block)[jpeg_natural_order[k]];
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if (temp < 0)
+      temp = -temp;		/* temp is abs value of input */
+    temp >>= Al;		/* apply the point transform */
+    absvalues[k] = temp;	/* save abs value for main pass */
+    if (temp == 1)
+      EOB = k;			/* EOB = index of last newly-nonzero coef */
+  }
+
+  /* Encode the AC coefficients per section G.1.2.3, fig. G.7 */
+  
+  r = 0;			/* r = run length of zeros */
+  BR = 0;			/* BR = count of buffered bits added now */
+  BR_buffer = entropy->bit_buffer + entropy->BE; /* Append bits to buffer */
+
+  for (k = cinfo->Ss; k <= Se; k++) {
+    if ((temp = absvalues[k]) == 0) {
+      r++;
+      continue;
+    }
+
+    /* Emit any required ZRLs, but not if they can be folded into EOB */
+    while (r > 15 && k <= EOB) {
+      /* emit any pending EOBRUN and the BE correction bits */
+      emit_eobrun(entropy);
+      /* Emit ZRL */
+      emit_symbol(entropy, entropy->ac_tbl_no, 0xF0);
+      r -= 16;
+      /* Emit buffered correction bits that must be associated with ZRL */
+      emit_buffered_bits(entropy, BR_buffer, BR);
+      BR_buffer = entropy->bit_buffer; /* BE bits are gone now */
+      BR = 0;
+    }
+
+    /* If the coef was previously nonzero, it only needs a correction bit.
+     * NOTE: a straight translation of the spec's figure G.7 would suggest
+     * that we also need to test r > 15.  But if r > 15, we can only get here
+     * if k > EOB, which implies that this coefficient is not 1.
+     */
+    if (temp > 1) {
+      /* The correction bit is the next bit of the absolute value. */
+      BR_buffer[BR++] = (char) (temp & 1);
+      continue;
+    }
+
+    /* Emit any pending EOBRUN and the BE correction bits */
+    emit_eobrun(entropy);
+
+    /* Count/emit Huffman symbol for run length / number of bits */
+    emit_symbol(entropy, entropy->ac_tbl_no, (r << 4) + 1);
+
+    /* Emit output bit for newly-nonzero coef */
+    temp = ((*block)[jpeg_natural_order[k]] < 0) ? 0 : 1;
+    emit_bits(entropy, (unsigned int) temp, 1);
+
+    /* Emit buffered correction bits that must be associated with this code */
+    emit_buffered_bits(entropy, BR_buffer, BR);
+    BR_buffer = entropy->bit_buffer; /* BE bits are gone now */
+    BR = 0;
+    r = 0;			/* reset zero run length */
+  }
+
+  if (r > 0 || BR > 0) {	/* If there are trailing zeroes, */
+    entropy->EOBRUN++;		/* count an EOB */
+    entropy->BE += BR;		/* concat my correction bits to older ones */
+    /* We force out the EOB if we risk either:
+     * 1. overflow of the EOB counter;
+     * 2. overflow of the correction bit buffer during the next MCU.
+     */
+    if (entropy->EOBRUN == 0x7FFF || entropy->BE > (MAX_CORR_BITS-DCTSIZE2+1))
+      emit_eobrun(entropy);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed progressive scan.
+ */
+
+METHODDEF(void)
+finish_pass_phuff (j_compress_ptr cinfo)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Flush out any buffered data */
+  emit_eobrun(entropy);
+  flush_bits(entropy);
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather_phuff (j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  boolean is_DC_band;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+  JHUFF_TBL **htblptr;
+  boolean did[NUM_HUFF_TBLS];
+
+  /* Flush out buffered data (all we care about is counting the EOB symbol) */
+  emit_eobrun(entropy);
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  MEMZERO(did, SIZEOF(did));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (is_DC_band) {
+      if (cinfo->Ah != 0)	/* DC refinement needs no table */
+	continue;
+      tbl = compptr->dc_tbl_no;
+    } else {
+      tbl = compptr->ac_tbl_no;
+    }
+    if (! did[tbl]) {
+      if (is_DC_band)
+        htblptr = & cinfo->dc_huff_tbl_ptrs[tbl];
+      else
+        htblptr = & cinfo->ac_huff_tbl_ptrs[tbl];
+      if (*htblptr == NULL)
+        *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->count_ptrs[tbl]);
+      did[tbl] = TRUE;
+    }
+  }
+}
+
+
+/*
+ * Module initialization routine for progressive Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_phuff_encoder (j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy;
+  int i;
+
+  entropy = (phuff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(phuff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass_phuff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->derived_tbls[i] = NULL;
+    entropy->count_ptrs[i] = NULL;
+  }
+  entropy->bit_buffer = NULL;	/* needed only in AC refinement scan */
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */
diff --git a/jcprepct.c b/jcprepct.c
new file mode 100644
index 0000000..fa93333
--- /dev/null
+++ b/jcprepct.c
@@ -0,0 +1,354 @@
+/*
+ * jcprepct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the compression preprocessing controller.
+ * This controller manages the color conversion, downsampling,
+ * and edge expansion steps.
+ *
+ * Most of the complexity here is associated with buffering input rows
+ * as required by the downsampler.  See the comments at the head of
+ * jcsample.c for the downsampler's needs.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* At present, jcsample.c can request context rows only for smoothing.
+ * In the future, we might also need context rows for CCIR601 sampling
+ * or other more-complex downsampling procedures.  The code to support
+ * context rows should be compiled only if needed.
+ */
+#ifdef INPUT_SMOOTHING_SUPPORTED
+#define CONTEXT_ROWS_SUPPORTED
+#endif
+
+
+/*
+ * For the simple (no-context-row) case, we just need to buffer one
+ * row group's worth of pixels for the downsampling step.  At the bottom of
+ * the image, we pad to a full row group by replicating the last pixel row.
+ * The downsampler's last output row is then replicated if needed to pad
+ * out to a full iMCU row.
+ *
+ * When providing context rows, we must buffer three row groups' worth of
+ * pixels.  Three row groups are physically allocated, but the row pointer
+ * arrays are made five row groups high, with the extra pointers above and
+ * below "wrapping around" to point to the last and first real row groups.
+ * This allows the downsampler to access the proper context rows.
+ * At the top and bottom of the image, we create dummy context rows by
+ * copying the first or last real pixel row.  This copying could be avoided
+ * by pointer hacking as is done in jdmainct.c, but it doesn't seem worth the
+ * trouble on the compression side.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_prep_controller pub; /* public fields */
+
+  /* Downsampling input buffer.  This buffer holds color-converted data
+   * until we have enough to do a downsample step.
+   */
+  JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  JDIMENSION rows_to_go;	/* counts rows remaining in source image */
+  int next_buf_row;		/* index of next row to store in color_buf */
+
+#ifdef CONTEXT_ROWS_SUPPORTED	/* only needed for context case */
+  int this_row_group;		/* starting row index of group to process */
+  int next_buf_stop;		/* downsample when we reach this index */
+#endif
+} my_prep_controller;
+
+typedef my_prep_controller * my_prep_ptr;
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_prep (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+
+  if (pass_mode != JBUF_PASS_THRU)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Initialize total-height counter for detecting bottom of image */
+  prep->rows_to_go = cinfo->image_height;
+  /* Mark the conversion buffer empty */
+  prep->next_buf_row = 0;
+#ifdef CONTEXT_ROWS_SUPPORTED
+  /* Preset additional state variables for context mode.
+   * These aren't used in non-context mode, so we needn't test which mode.
+   */
+  prep->this_row_group = 0;
+  /* Set next_buf_stop to stop after two row groups have been read in. */
+  prep->next_buf_stop = 2 * cinfo->max_v_samp_factor;
+#endif
+}
+
+
+/*
+ * Expand an image vertically from height input_rows to height output_rows,
+ * by duplicating the bottom row.
+ */
+
+LOCAL(void)
+expand_bottom_edge (JSAMPARRAY image_data, JDIMENSION num_cols,
+		    int input_rows, int output_rows)
+{
+  register int row;
+
+  for (row = input_rows; row < output_rows; row++) {
+    jcopy_sample_rows(image_data, input_rows-1, image_data, row,
+		      1, num_cols);
+  }
+}
+
+
+/*
+ * Process some data in the simple no-context case.
+ *
+ * Preprocessor output data is counted in "row groups".  A row group
+ * is defined to be v_samp_factor sample rows of each component.
+ * Downsampling will produce this much data from each max_v_samp_factor
+ * input rows.
+ */
+
+METHODDEF(void)
+pre_process_data (j_compress_ptr cinfo,
+		  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+		  JDIMENSION in_rows_avail,
+		  JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+		  JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int numrows, ci;
+  JDIMENSION inrows;
+  jpeg_component_info * compptr;
+
+  while (*in_row_ctr < in_rows_avail &&
+	 *out_row_group_ctr < out_row_groups_avail) {
+    /* Do color conversion to fill the conversion buffer. */
+    inrows = in_rows_avail - *in_row_ctr;
+    numrows = cinfo->max_v_samp_factor - prep->next_buf_row;
+    numrows = (int) MIN((JDIMENSION) numrows, inrows);
+    (*cinfo->cconvert->color_convert) (cinfo, input_buf + *in_row_ctr,
+				       prep->color_buf,
+				       (JDIMENSION) prep->next_buf_row,
+				       numrows);
+    *in_row_ctr += numrows;
+    prep->next_buf_row += numrows;
+    prep->rows_to_go -= numrows;
+    /* If at bottom of image, pad to fill the conversion buffer. */
+    if (prep->rows_to_go == 0 &&
+	prep->next_buf_row < cinfo->max_v_samp_factor) {
+      for (ci = 0; ci < cinfo->num_components; ci++) {
+	expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+			   prep->next_buf_row, cinfo->max_v_samp_factor);
+      }
+      prep->next_buf_row = cinfo->max_v_samp_factor;
+    }
+    /* If we've filled the conversion buffer, empty it. */
+    if (prep->next_buf_row == cinfo->max_v_samp_factor) {
+      (*cinfo->downsample->downsample) (cinfo,
+					prep->color_buf, (JDIMENSION) 0,
+					output_buf, *out_row_group_ctr);
+      prep->next_buf_row = 0;
+      (*out_row_group_ctr)++;
+    }
+    /* If at bottom of image, pad the output to a full iMCU height.
+     * Note we assume the caller is providing a one-iMCU-height output buffer!
+     */
+    if (prep->rows_to_go == 0 &&
+	*out_row_group_ctr < out_row_groups_avail) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	   ci++, compptr++) {
+	expand_bottom_edge(output_buf[ci],
+			   compptr->width_in_blocks * DCTSIZE,
+			   (int) (*out_row_group_ctr * compptr->v_samp_factor),
+			   (int) (out_row_groups_avail * compptr->v_samp_factor));
+      }
+      *out_row_group_ctr = out_row_groups_avail;
+      break;			/* can exit outer loop without test */
+    }
+  }
+}
+
+
+#ifdef CONTEXT_ROWS_SUPPORTED
+
+/*
+ * Process some data in the context case.
+ */
+
+METHODDEF(void)
+pre_process_context (j_compress_ptr cinfo,
+		     JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+		     JDIMENSION in_rows_avail,
+		     JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+		     JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int numrows, ci;
+  int buf_height = cinfo->max_v_samp_factor * 3;
+  JDIMENSION inrows;
+
+  while (*out_row_group_ctr < out_row_groups_avail) {
+    if (*in_row_ctr < in_rows_avail) {
+      /* Do color conversion to fill the conversion buffer. */
+      inrows = in_rows_avail - *in_row_ctr;
+      numrows = prep->next_buf_stop - prep->next_buf_row;
+      numrows = (int) MIN((JDIMENSION) numrows, inrows);
+      (*cinfo->cconvert->color_convert) (cinfo, input_buf + *in_row_ctr,
+					 prep->color_buf,
+					 (JDIMENSION) prep->next_buf_row,
+					 numrows);
+      /* Pad at top of image, if first time through */
+      if (prep->rows_to_go == cinfo->image_height) {
+	for (ci = 0; ci < cinfo->num_components; ci++) {
+	  int row;
+	  for (row = 1; row <= cinfo->max_v_samp_factor; row++) {
+	    jcopy_sample_rows(prep->color_buf[ci], 0,
+			      prep->color_buf[ci], -row,
+			      1, cinfo->image_width);
+	  }
+	}
+      }
+      *in_row_ctr += numrows;
+      prep->next_buf_row += numrows;
+      prep->rows_to_go -= numrows;
+    } else {
+      /* Return for more data, unless we are at the bottom of the image. */
+      if (prep->rows_to_go != 0)
+	break;
+      /* When at bottom of image, pad to fill the conversion buffer. */
+      if (prep->next_buf_row < prep->next_buf_stop) {
+	for (ci = 0; ci < cinfo->num_components; ci++) {
+	  expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+			     prep->next_buf_row, prep->next_buf_stop);
+	}
+	prep->next_buf_row = prep->next_buf_stop;
+      }
+    }
+    /* If we've gotten enough data, downsample a row group. */
+    if (prep->next_buf_row == prep->next_buf_stop) {
+      (*cinfo->downsample->downsample) (cinfo,
+					prep->color_buf,
+					(JDIMENSION) prep->this_row_group,
+					output_buf, *out_row_group_ctr);
+      (*out_row_group_ctr)++;
+      /* Advance pointers with wraparound as necessary. */
+      prep->this_row_group += cinfo->max_v_samp_factor;
+      if (prep->this_row_group >= buf_height)
+	prep->this_row_group = 0;
+      if (prep->next_buf_row >= buf_height)
+	prep->next_buf_row = 0;
+      prep->next_buf_stop = prep->next_buf_row + cinfo->max_v_samp_factor;
+    }
+  }
+}
+
+
+/*
+ * Create the wrapped-around downsampling input buffer needed for context mode.
+ */
+
+LOCAL(void)
+create_context_buffer (j_compress_ptr cinfo)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int rgroup_height = cinfo->max_v_samp_factor;
+  int ci, i;
+  jpeg_component_info * compptr;
+  JSAMPARRAY true_buffer, fake_buffer;
+
+  /* Grab enough space for fake row pointers for all the components;
+   * we need five row groups' worth of pointers for each component.
+   */
+  fake_buffer = (JSAMPARRAY)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(cinfo->num_components * 5 * rgroup_height) *
+				SIZEOF(JSAMPROW));
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate the actual buffer space (3 row groups) for this component.
+     * We make the buffer wide enough to allow the downsampler to edge-expand
+     * horizontally within the buffer, if it so chooses.
+     */
+    true_buffer = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) (((long) compptr->width_in_blocks * DCTSIZE *
+		      cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+       (JDIMENSION) (3 * rgroup_height));
+    /* Copy true buffer row pointers into the middle of the fake row array */
+    MEMCOPY(fake_buffer + rgroup_height, true_buffer,
+	    3 * rgroup_height * SIZEOF(JSAMPROW));
+    /* Fill in the above and below wraparound pointers */
+    for (i = 0; i < rgroup_height; i++) {
+      fake_buffer[i] = true_buffer[2 * rgroup_height + i];
+      fake_buffer[4 * rgroup_height + i] = true_buffer[i];
+    }
+    prep->color_buf[ci] = fake_buffer + rgroup_height;
+    fake_buffer += 5 * rgroup_height; /* point to space for next component */
+  }
+}
+
+#endif /* CONTEXT_ROWS_SUPPORTED */
+
+
+/*
+ * Initialize preprocessing controller.
+ */
+
+GLOBAL(void)
+jinit_c_prep_controller (j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_prep_ptr prep;
+  int ci;
+  jpeg_component_info * compptr;
+
+  if (need_full_buffer)		/* safety check */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  prep = (my_prep_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_prep_controller));
+  cinfo->prep = (struct jpeg_c_prep_controller *) prep;
+  prep->pub.start_pass = start_pass_prep;
+
+  /* Allocate the color conversion buffer.
+   * We make the buffer wide enough to allow the downsampler to edge-expand
+   * horizontally within the buffer, if it so chooses.
+   */
+  if (cinfo->downsample->need_context_rows) {
+    /* Set up to provide context rows */
+#ifdef CONTEXT_ROWS_SUPPORTED
+    prep->pub.pre_process_data = pre_process_context;
+    create_context_buffer(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* No context, just make it tall enough for one row group */
+    prep->pub.pre_process_data = pre_process_data;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      prep->color_buf[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 (JDIMENSION) (((long) compptr->width_in_blocks * DCTSIZE *
+			cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+	 (JDIMENSION) cinfo->max_v_samp_factor);
+    }
+  }
+}
diff --git a/jcsample.c b/jcsample.c
new file mode 100644
index 0000000..eea376f
--- /dev/null
+++ b/jcsample.c
@@ -0,0 +1,527 @@
+/*
+ * jcsample.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains downsampling routines.
+ *
+ * Downsampling input data is counted in "row groups".  A row group
+ * is defined to be max_v_samp_factor pixel rows of each component,
+ * from which the downsampler produces v_samp_factor sample rows.
+ * A single row group is processed in each call to the downsampler module.
+ *
+ * The downsampler is responsible for edge-expansion of its output data
+ * to fill an integral number of DCT blocks horizontally.  The source buffer
+ * may be modified if it is helpful for this purpose (the source buffer is
+ * allocated wide enough to correspond to the desired output width).
+ * The caller (the prep controller) is responsible for vertical padding.
+ *
+ * The downsampler may request "context rows" by setting need_context_rows
+ * during startup.  In this case, the input arrays will contain at least
+ * one row group's worth of pixels above and below the passed-in data;
+ * the caller will create dummy rows at image top and bottom by replicating
+ * the first or last real pixel row.
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ *
+ * The downsampling algorithm used here is a simple average of the source
+ * pixels covered by the output pixel.  The hi-falutin sampling literature
+ * refers to this as a "box filter".  In general the characteristics of a box
+ * filter are not very good, but for the specific cases we normally use (1:1
+ * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
+ * nearly so bad.  If you intend to use other sampling ratios, you'd be well
+ * advised to improve this code.
+ *
+ * A simple input-smoothing capability is provided.  This is mainly intended
+ * for cleaning up color-dithered GIF input files (if you find it inadequate,
+ * we suggest using an external filtering program such as pnmconvol).  When
+ * enabled, each input pixel P is replaced by a weighted sum of itself and its
+ * eight neighbors.  P's weight is 1-8*SF and each neighbor's weight is SF,
+ * where SF = (smoothing_factor / 1024).
+ * Currently, smoothing is only supported for 2h2v sampling factors.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+
+
+/* Pointer to routine to downsample a single component */
+typedef JMETHOD(void, downsample1_ptr,
+		(j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data));
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_downsampler pub;	/* public fields */
+
+  /* Downsampling method pointers, one per component */
+  downsample1_ptr methods[MAX_COMPONENTS];
+} my_downsampler;
+
+typedef my_downsampler * my_downsample_ptr;
+
+
+/*
+ * Initialize for a downsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_downsample (j_compress_ptr cinfo)
+{
+  /* no work for now */
+}
+
+
+/*
+ * Expand a component horizontally from width input_cols to width output_cols,
+ * by duplicating the rightmost samples.
+ */
+
+LOCAL(void)
+expand_right_edge (JSAMPARRAY image_data, int num_rows,
+		   JDIMENSION input_cols, JDIMENSION output_cols)
+{
+  register JSAMPROW ptr;
+  register JSAMPLE pixval;
+  register int count;
+  int row;
+  int numcols = (int) (output_cols - input_cols);
+
+  if (numcols > 0) {
+    for (row = 0; row < num_rows; row++) {
+      ptr = image_data[row] + input_cols;
+      pixval = ptr[-1];		/* don't need GETJSAMPLE() here */
+      for (count = numcols; count > 0; count--)
+	*ptr++ = pixval;
+    }
+  }
+}
+
+
+/*
+ * Do downsampling for a whole row group (all components).
+ *
+ * In this version we simply downsample each component independently.
+ */
+
+METHODDEF(void)
+sep_downsample (j_compress_ptr cinfo,
+		JSAMPIMAGE input_buf, JDIMENSION in_row_index,
+		JSAMPIMAGE output_buf, JDIMENSION out_row_group_index)
+{
+  my_downsample_ptr downsample = (my_downsample_ptr) cinfo->downsample;
+  int ci;
+  jpeg_component_info * compptr;
+  JSAMPARRAY in_ptr, out_ptr;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    in_ptr = input_buf[ci] + in_row_index;
+    out_ptr = output_buf[ci] + (out_row_group_index * compptr->v_samp_factor);
+    (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr);
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * One row group is processed per call.
+ * This version handles arbitrary integral sampling ratios, without smoothing.
+ * Note that this version is not actually used for customary sampling ratios.
+ */
+
+METHODDEF(void)
+int_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v;
+  JDIMENSION outcol, outcol_h;	/* outcol_h == outcol*h_expand */
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  JSAMPROW inptr, outptr;
+  INT32 outvalue;
+
+  h_expand = cinfo->max_h_samp_factor / compptr->h_samp_factor;
+  v_expand = cinfo->max_v_samp_factor / compptr->v_samp_factor;
+  numpix = h_expand * v_expand;
+  numpix2 = numpix/2;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * h_expand);
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    for (outcol = 0, outcol_h = 0; outcol < output_cols;
+	 outcol++, outcol_h += h_expand) {
+      outvalue = 0;
+      for (v = 0; v < v_expand; v++) {
+	inptr = input_data[inrow+v] + outcol_h;
+	for (h = 0; h < h_expand; h++) {
+	  outvalue += (INT32) GETJSAMPLE(*inptr++);
+	}
+      }
+      *outptr++ = (JSAMPLE) ((outvalue + numpix2) / numpix);
+    }
+    inrow += v_expand;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+fullsize_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  /* Copy the data */
+  jcopy_sample_rows(input_data, 0, output_data, 0,
+		    cinfo->max_v_samp_factor, cinfo->image_width);
+  /* Edge-expand */
+  expand_right_edge(output_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, compptr->width_in_blocks * DCTSIZE);
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the common case of 2:1 horizontal and 1:1 vertical,
+ * without smoothing.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int outrow;
+  JDIMENSION outcol;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  register JSAMPROW inptr, outptr;
+  register int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * 2);
+
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+    bias = 0;			/* bias = 0,1,0,1,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr) + GETJSAMPLE(inptr[1])
+			      + bias) >> 1);
+      bias ^= 1;		/* 0=>1, 1=>0 */
+      inptr += 2;
+    }
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION outcol;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  register JSAMPROW inptr0, inptr1, outptr;
+  register int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * 2);
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow+1];
+    bias = 1;			/* bias = 1,2,1,2,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+			      GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1])
+			      + bias) >> 2);
+      bias ^= 3;		/* 1=>2, 2=>1 */
+      inptr0 += 2; inptr1 += 2;
+    }
+    inrow += 2;
+  }
+}
+
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+h2v2_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+			JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION colctr;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  register JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
+  INT32 membersum, neighsum, memberscale, neighscale;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+		    cinfo->image_width, output_cols * 2);
+
+  /* We don't bother to form the individual "smoothed" input pixel values;
+   * we can directly compute the output which is the average of the four
+   * smoothed values.  Each of the four member pixels contributes a fraction
+   * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
+   * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
+   * output.  The four corner-adjacent neighbor pixels contribute a fraction
+   * SF to just one smoothed pixel, or SF/4 to the final output; while the
+   * eight edge-adjacent neighbors contribute SF to each of two smoothed
+   * pixels, or SF/2 overall.  In order to use integer arithmetic, these
+   * factors are scaled by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
+  neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow+1];
+    above_ptr = input_data[inrow-1];
+    below_ptr = input_data[inrow+2];
+
+    /* Special case for first column: pretend column -1 is same as column 0 */
+    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+	       GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) +
+	       GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]);
+    neighsum += neighsum;
+    neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) +
+		GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]);
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+    inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      /* sum of pixels directly mapped to this output element */
+      membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		  GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+      /* sum of edge-neighbor pixels */
+      neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+		 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+		 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) +
+		 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]);
+      /* The edge-neighbors count twice as much as corner-neighbors */
+      neighsum += neighsum;
+      /* Add in the corner-neighbors */
+      neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) +
+		  GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]);
+      /* form final output scaled up by 2^16 */
+      membersum = membersum * memberscale + neighsum * neighscale;
+      /* round, descale and output it */
+      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+      inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
+    }
+
+    /* Special case for last column */
+    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+	       GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) +
+	       GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]);
+    neighsum += neighsum;
+    neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) +
+		GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]);
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
+
+    inrow += 2;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+fullsize_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
+			    JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int outrow;
+  JDIMENSION colctr;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  register JSAMPROW inptr, above_ptr, below_ptr, outptr;
+  INT32 membersum, neighsum, memberscale, neighscale;
+  int colsum, lastcolsum, nextcolsum;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+		    cinfo->image_width, output_cols);
+
+  /* Each of the eight neighbor pixels contributes a fraction SF to the
+   * smoothed pixel, while the main pixel contributes (1-8*SF).  In order
+   * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
+  neighscale = cinfo->smoothing_factor * 64; /* scaled SF */
+
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+    above_ptr = input_data[outrow-1];
+    below_ptr = input_data[outrow+1];
+
+    /* Special case for first column */
+    colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) +
+	     GETJSAMPLE(*inptr);
+    membersum = GETJSAMPLE(*inptr++);
+    nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
+		 GETJSAMPLE(*inptr);
+    neighsum = colsum + (colsum - membersum) + nextcolsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+    lastcolsum = colsum; colsum = nextcolsum;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      membersum = GETJSAMPLE(*inptr++);
+      above_ptr++; below_ptr++;
+      nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
+		   GETJSAMPLE(*inptr);
+      neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
+      membersum = membersum * memberscale + neighsum * neighscale;
+      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+      lastcolsum = colsum; colsum = nextcolsum;
+    }
+
+    /* Special case for last column */
+    membersum = GETJSAMPLE(*inptr);
+    neighsum = lastcolsum + (colsum - membersum) + colsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
+
+  }
+}
+
+#endif /* INPUT_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Module initialization routine for downsampling.
+ * Note that we must select a routine for each component.
+ */
+
+GLOBAL(void)
+jinit_downsampler (j_compress_ptr cinfo)
+{
+  my_downsample_ptr downsample;
+  int ci;
+  jpeg_component_info * compptr;
+  boolean smoothok = TRUE;
+
+  downsample = (my_downsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_downsampler));
+  cinfo->downsample = (struct jpeg_downsampler *) downsample;
+  downsample->pub.start_pass = start_pass_downsample;
+  downsample->pub.downsample = sep_downsample;
+  downsample->pub.need_context_rows = FALSE;
+
+  if (cinfo->CCIR601_sampling)
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* Verify we can handle the sampling factors, and set up method pointers */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor == cinfo->max_h_samp_factor &&
+	compptr->v_samp_factor == cinfo->max_v_samp_factor) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+	downsample->methods[ci] = fullsize_smooth_downsample;
+	downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+	downsample->methods[ci] = fullsize_downsample;
+    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
+	       compptr->v_samp_factor == cinfo->max_v_samp_factor) {
+      smoothok = FALSE;
+      if (jsimd_can_h2v1_downsample())
+        downsample->methods[ci] = jsimd_h2v1_downsample;
+      else
+        downsample->methods[ci] = h2v1_downsample;
+    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
+	       compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+	downsample->methods[ci] = h2v2_smooth_downsample;
+	downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+	if (jsimd_can_h2v2_downsample())
+	  downsample->methods[ci] = jsimd_h2v2_downsample;
+	else
+	  downsample->methods[ci] = h2v2_downsample;
+    } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 &&
+	       (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) {
+      smoothok = FALSE;
+      downsample->methods[ci] = int_downsample;
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+  }
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+  if (cinfo->smoothing_factor && !smoothok)
+    TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
+#endif
+}
diff --git a/jctrans.c b/jctrans.c
new file mode 100644
index 0000000..c7956b5
--- /dev/null
+++ b/jctrans.c
@@ -0,0 +1,395 @@
+/*
+ * jctrans.c
+ *
+ * Copyright (C) 1995-1998, Thomas G. Lane.
+ * Modified 2000-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains library routines for transcoding compression,
+ * that is, writing raw DCT coefficient arrays to an output JPEG file.
+ * The routines in jcapimin.c will also be needed by a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Forward declarations */
+LOCAL(void) transencode_master_selection
+	JPP((j_compress_ptr cinfo, jvirt_barray_ptr * coef_arrays));
+LOCAL(void) transencode_coef_controller
+	JPP((j_compress_ptr cinfo, jvirt_barray_ptr * coef_arrays));
+
+
+/*
+ * Compression initialization for writing raw-coefficient data.
+ * Before calling this, all parameters and a data destination must be set up.
+ * Call jpeg_finish_compress() to actually write the data.
+ *
+ * The number of passed virtual arrays must match cinfo->num_components.
+ * Note that the virtual arrays need not be filled or even realized at
+ * the time write_coefficients is called; indeed, if the virtual arrays
+ * were requested from this compression object's memory manager, they
+ * typically will be realized during this routine and filled afterwards.
+ */
+
+GLOBAL(void)
+jpeg_write_coefficients (j_compress_ptr cinfo, jvirt_barray_ptr * coef_arrays)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Mark all tables to be written */
+  jpeg_suppress_tables(cinfo, FALSE);
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  transencode_master_selection(cinfo, coef_arrays);
+  /* Wait for jpeg_finish_compress() call */
+  cinfo->next_scanline = 0;	/* so jpeg_write_marker works */
+  cinfo->global_state = CSTATE_WRCOEFS;
+}
+
+
+/*
+ * Initialize the compression object with default parameters,
+ * then copy from the source object all parameters needed for lossless
+ * transcoding.  Parameters that can be varied without loss (such as
+ * scan script and Huffman optimization) are left in their default states.
+ */
+
+GLOBAL(void)
+jpeg_copy_critical_parameters (j_decompress_ptr srcinfo,
+			       j_compress_ptr dstinfo)
+{
+  JQUANT_TBL ** qtblptr;
+  jpeg_component_info *incomp, *outcomp;
+  JQUANT_TBL *c_quant, *slot_quant;
+  int tblno, ci, coefi;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (dstinfo->global_state != CSTATE_START)
+    ERREXIT1(dstinfo, JERR_BAD_STATE, dstinfo->global_state);
+  /* Copy fundamental image dimensions */
+  dstinfo->image_width = srcinfo->image_width;
+  dstinfo->image_height = srcinfo->image_height;
+  dstinfo->input_components = srcinfo->num_components;
+  dstinfo->in_color_space = srcinfo->jpeg_color_space;
+#if JPEG_LIB_VERSION >= 70
+  dstinfo->jpeg_width = srcinfo->output_width;
+  dstinfo->jpeg_height = srcinfo->output_height;
+  dstinfo->min_DCT_h_scaled_size = srcinfo->min_DCT_h_scaled_size;
+  dstinfo->min_DCT_v_scaled_size = srcinfo->min_DCT_v_scaled_size;
+#endif
+  /* Initialize all parameters to default values */
+  jpeg_set_defaults(dstinfo);
+  /* jpeg_set_defaults may choose wrong colorspace, eg YCbCr if input is RGB.
+   * Fix it to get the right header markers for the image colorspace.
+   */
+  jpeg_set_colorspace(dstinfo, srcinfo->jpeg_color_space);
+  dstinfo->data_precision = srcinfo->data_precision;
+  dstinfo->CCIR601_sampling = srcinfo->CCIR601_sampling;
+  /* Copy the source's quantization tables. */
+  for (tblno = 0; tblno < NUM_QUANT_TBLS; tblno++) {
+    if (srcinfo->quant_tbl_ptrs[tblno] != NULL) {
+      qtblptr = & dstinfo->quant_tbl_ptrs[tblno];
+      if (*qtblptr == NULL)
+	*qtblptr = jpeg_alloc_quant_table((j_common_ptr) dstinfo);
+      MEMCOPY((*qtblptr)->quantval,
+	      srcinfo->quant_tbl_ptrs[tblno]->quantval,
+	      SIZEOF((*qtblptr)->quantval));
+      (*qtblptr)->sent_table = FALSE;
+    }
+  }
+  /* Copy the source's per-component info.
+   * Note we assume jpeg_set_defaults has allocated the dest comp_info array.
+   */
+  dstinfo->num_components = srcinfo->num_components;
+  if (dstinfo->num_components < 1 || dstinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(dstinfo, JERR_COMPONENT_COUNT, dstinfo->num_components,
+	     MAX_COMPONENTS);
+  for (ci = 0, incomp = srcinfo->comp_info, outcomp = dstinfo->comp_info;
+       ci < dstinfo->num_components; ci++, incomp++, outcomp++) {
+    outcomp->component_id = incomp->component_id;
+    outcomp->h_samp_factor = incomp->h_samp_factor;
+    outcomp->v_samp_factor = incomp->v_samp_factor;
+    outcomp->quant_tbl_no = incomp->quant_tbl_no;
+    /* Make sure saved quantization table for component matches the qtable
+     * slot.  If not, the input file re-used this qtable slot.
+     * IJG encoder currently cannot duplicate this.
+     */
+    tblno = outcomp->quant_tbl_no;
+    if (tblno < 0 || tblno >= NUM_QUANT_TBLS ||
+	srcinfo->quant_tbl_ptrs[tblno] == NULL)
+      ERREXIT1(dstinfo, JERR_NO_QUANT_TABLE, tblno);
+    slot_quant = srcinfo->quant_tbl_ptrs[tblno];
+    c_quant = incomp->quant_table;
+    if (c_quant != NULL) {
+      for (coefi = 0; coefi < DCTSIZE2; coefi++) {
+	if (c_quant->quantval[coefi] != slot_quant->quantval[coefi])
+	  ERREXIT1(dstinfo, JERR_MISMATCHED_QUANT_TABLE, tblno);
+      }
+    }
+    /* Note: we do not copy the source's Huffman table assignments;
+     * instead we rely on jpeg_set_colorspace to have made a suitable choice.
+     */
+  }
+  /* Also copy JFIF version and resolution information, if available.
+   * Strictly speaking this isn't "critical" info, but it's nearly
+   * always appropriate to copy it if available.  In particular,
+   * if the application chooses to copy JFIF 1.02 extension markers from
+   * the source file, we need to copy the version to make sure we don't
+   * emit a file that has 1.02 extensions but a claimed version of 1.01.
+   * We will *not*, however, copy version info from mislabeled "2.01" files.
+   */
+  if (srcinfo->saw_JFIF_marker) {
+    if (srcinfo->JFIF_major_version == 1) {
+      dstinfo->JFIF_major_version = srcinfo->JFIF_major_version;
+      dstinfo->JFIF_minor_version = srcinfo->JFIF_minor_version;
+    }
+    dstinfo->density_unit = srcinfo->density_unit;
+    dstinfo->X_density = srcinfo->X_density;
+    dstinfo->Y_density = srcinfo->Y_density;
+  }
+}
+
+
+/*
+ * Master selection of compression modules for transcoding.
+ * This substitutes for jcinit.c's initialization of the full compressor.
+ */
+
+LOCAL(void)
+transencode_master_selection (j_compress_ptr cinfo,
+			      jvirt_barray_ptr * coef_arrays)
+{
+  /* Although we don't actually use input_components for transcoding,
+   * jcmaster.c's initial_setup will complain if input_components is 0.
+   */
+  cinfo->input_components = 1;
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, TRUE /* transcode only */);
+
+  /* Entropy encoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      jinit_phuff_encoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_encoder(cinfo);
+  }
+
+  /* We need a special coefficient buffer controller. */
+  transencode_coef_controller(cinfo, coef_arrays);
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Write the datastream header (SOI, JFIF) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}
+
+
+/*
+ * The rest of this file is a special implementation of the coefficient
+ * buffer controller.  This is similar to jccoefct.c, but it handles only
+ * output from presupplied virtual arrays.  Furthermore, we generate any
+ * dummy padding blocks on-the-fly rather than expecting them to be present
+ * in the arrays.
+ */
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;	/* iMCU row # within image */
+  JDIMENSION mcu_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* Virtual block array for each component. */
+  jvirt_barray_ptr * whole_image;
+
+  /* Workspace for constructing dummy blocks at right/bottom edges. */
+  JBLOCKROW dummy_buffer[C_MAX_BLOCKS_IN_MCU];
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+
+LOCAL(void)
+start_iMCU_row (j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  if (pass_mode != JBUF_CRANK_DEST)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Process some data.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(boolean)
+compress_output (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, blockcnt;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						: compptr->last_col_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (coef->iMCU_row_num < last_iMCU_row ||
+	      yindex+yoffset < compptr->last_row_height) {
+	    /* Fill in pointers to real blocks in this row */
+	    buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	    for (xindex = 0; xindex < blockcnt; xindex++)
+	      MCU_buffer[blkn++] = buffer_ptr++;
+	  } else {
+	    /* At bottom of image, need a whole row of dummy blocks */
+	    xindex = 0;
+	  }
+	  /* Fill in any dummy blocks needed in this row.
+	   * Dummy blocks are filled in the same way as in jccoefct.c:
+	   * all zeroes in the AC entries, DC entries equal to previous
+	   * block's DC value.  The init routine has already zeroed the
+	   * AC entries, so we need only set the DC entries correctly.
+	   */
+	  for (; xindex < compptr->MCU_width; xindex++) {
+	    MCU_buffer[blkn] = coef->dummy_buffer[blkn];
+	    MCU_buffer[blkn][0][0] = MCU_buffer[blkn-1][0][0];
+	    blkn++;
+	  }
+	}
+      }
+      /* Try to write the MCU. */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Initialize coefficient buffer controller.
+ *
+ * Each passed coefficient array must be the right size for that
+ * coefficient: width_in_blocks wide and height_in_blocks high,
+ * with unitheight at least v_samp_factor.
+ */
+
+LOCAL(void)
+transencode_coef_controller (j_compress_ptr cinfo,
+			     jvirt_barray_ptr * coef_arrays)
+{
+  my_coef_ptr coef;
+  JBLOCKROW buffer;
+  int i;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *) coef;
+  coef->pub.start_pass = start_pass_coef;
+  coef->pub.compress_data = compress_output;
+
+  /* Save pointer to virtual arrays */
+  coef->whole_image = coef_arrays;
+
+  /* Allocate and pre-zero space for dummy DCT blocks. */
+  buffer = (JBLOCKROW)
+    (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+  jzero_far((void FAR *) buffer, C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+  for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+    coef->dummy_buffer[i] = buffer + i;
+  }
+}
diff --git a/jdapimin.c b/jdapimin.c
new file mode 100644
index 0000000..cadb59f
--- /dev/null
+++ b/jdapimin.c
@@ -0,0 +1,395 @@
+/*
+ * jdapimin.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-decompression case or the
+ * transcoding-only case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jdapistd.c.  But also see jcomapi.c for routines
+ * shared by compression and decompression, and jdtrans.c for the transcoding
+ * case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Initialization of a JPEG decompression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateDecompress (j_decompress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;		/* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != SIZEOF(struct jpeg_decompress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE, 
+	     (int) SIZEOF(struct jpeg_decompress_struct), (int) structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr * err = cinfo->err;
+    void * client_data = cinfo->client_data; /* ignore Purify complaint here */
+    MEMZERO(cinfo, SIZEOF(struct jpeg_decompress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = TRUE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr) cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->src = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++)
+    cinfo->quant_tbl_ptrs[i] = NULL;
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+  /* Initialize marker processor so application can override methods
+   * for COM, APPn markers before calling jpeg_read_header.
+   */
+  cinfo->marker_list = NULL;
+  jinit_marker_reader(cinfo);
+
+  /* And initialize the overall input controller. */
+  jinit_input_controller(cinfo);
+
+  /* OK, I'm ready */
+  cinfo->global_state = DSTATE_START;
+}
+
+
+/*
+ * Destruction of a JPEG decompression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_decompress (j_decompress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG decompression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_decompress (j_decompress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Set default decompression parameters.
+ */
+
+LOCAL(void)
+default_decompress_parms (j_decompress_ptr cinfo)
+{
+  /* Guess the input colorspace, and set output colorspace accordingly. */
+  /* (Wish JPEG committee had provided a real way to specify this...) */
+  /* Note application may override our guesses. */
+  switch (cinfo->num_components) {
+  case 1:
+    cinfo->jpeg_color_space = JCS_GRAYSCALE;
+    cinfo->out_color_space = JCS_GRAYSCALE;
+    break;
+    
+  case 3:
+    if (cinfo->saw_JFIF_marker) {
+      cinfo->jpeg_color_space = JCS_YCbCr; /* JFIF implies YCbCr */
+    } else if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+	cinfo->jpeg_color_space = JCS_RGB;
+	break;
+      case 1:
+	cinfo->jpeg_color_space = JCS_YCbCr;
+	break;
+      default:
+	WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+	break;
+      }
+    } else {
+      /* Saw no special markers, try to guess from the component IDs */
+      int cid0 = cinfo->comp_info[0].component_id;
+      int cid1 = cinfo->comp_info[1].component_id;
+      int cid2 = cinfo->comp_info[2].component_id;
+
+      if (cid0 == 1 && cid1 == 2 && cid2 == 3)
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume JFIF w/out marker */
+      else if (cid0 == 82 && cid1 == 71 && cid2 == 66)
+	cinfo->jpeg_color_space = JCS_RGB; /* ASCII 'R', 'G', 'B' */
+      else {
+	TRACEMS3(cinfo, 1, JTRC_UNKNOWN_IDS, cid0, cid1, cid2);
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+      }
+    }
+    /* Always guess RGB is proper output colorspace. */
+    cinfo->out_color_space = JCS_RGB;
+    break;
+    
+  case 4:
+    if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+	cinfo->jpeg_color_space = JCS_CMYK;
+	break;
+      case 2:
+	cinfo->jpeg_color_space = JCS_YCCK;
+	break;
+      default:
+	WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+	cinfo->jpeg_color_space = JCS_YCCK; /* assume it's YCCK */
+	break;
+      }
+    } else {
+      /* No special markers, assume straight CMYK. */
+      cinfo->jpeg_color_space = JCS_CMYK;
+    }
+    cinfo->out_color_space = JCS_CMYK;
+    break;
+    
+  default:
+    cinfo->jpeg_color_space = JCS_UNKNOWN;
+    cinfo->out_color_space = JCS_UNKNOWN;
+    break;
+  }
+
+  /* Set defaults for other decompression parameters. */
+  cinfo->scale_num = 1;		/* 1:1 scaling */
+  cinfo->scale_denom = 1;
+  cinfo->output_gamma = 1.0;
+  cinfo->buffered_image = FALSE;
+  cinfo->raw_data_out = FALSE;
+  cinfo->dct_method = JDCT_DEFAULT;
+  cinfo->do_fancy_upsampling = TRUE;
+  cinfo->do_block_smoothing = TRUE;
+  cinfo->quantize_colors = FALSE;
+  /* We set these in case application only sets quantize_colors. */
+  cinfo->dither_mode = JDITHER_FS;
+#ifdef QUANT_2PASS_SUPPORTED
+  cinfo->two_pass_quantize = TRUE;
+#else
+  cinfo->two_pass_quantize = FALSE;
+#endif
+  cinfo->desired_number_of_colors = 256;
+  cinfo->colormap = NULL;
+  /* Initialize for no mode change in buffered-image mode. */
+  cinfo->enable_1pass_quant = FALSE;
+  cinfo->enable_external_quant = FALSE;
+  cinfo->enable_2pass_quant = FALSE;
+}
+
+
+/*
+ * Decompression startup: read start of JPEG datastream to see what's there.
+ * Need only initialize JPEG object and supply a data source before calling.
+ *
+ * This routine will read as far as the first SOS marker (ie, actual start of
+ * compressed data), and will save all tables and parameters in the JPEG
+ * object.  It will also initialize the decompression parameters to default
+ * values, and finally return JPEG_HEADER_OK.  On return, the application may
+ * adjust the decompression parameters and then call jpeg_start_decompress.
+ * (Or, if the application only wanted to determine the image parameters,
+ * the data need not be decompressed.  In that case, call jpeg_abort or
+ * jpeg_destroy to release any temporary space.)
+ * If an abbreviated (tables only) datastream is presented, the routine will
+ * return JPEG_HEADER_TABLES_ONLY upon reaching EOI.  The application may then
+ * re-use the JPEG object to read the abbreviated image datastream(s).
+ * It is unnecessary (but OK) to call jpeg_abort in this case.
+ * The JPEG_SUSPENDED return code only occurs if the data source module
+ * requests suspension of the decompressor.  In this case the application
+ * should load more source data and then re-call jpeg_read_header to resume
+ * processing.
+ * If a non-suspending data source is used and require_image is TRUE, then the
+ * return code need not be inspected since only JPEG_HEADER_OK is possible.
+ *
+ * This routine is now just a front end to jpeg_consume_input, with some
+ * extra error checking.
+ */
+
+GLOBAL(int)
+jpeg_read_header (j_decompress_ptr cinfo, boolean require_image)
+{
+  int retcode;
+
+  if (cinfo->global_state != DSTATE_START &&
+      cinfo->global_state != DSTATE_INHEADER)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  retcode = jpeg_consume_input(cinfo);
+
+  switch (retcode) {
+  case JPEG_REACHED_SOS:
+    retcode = JPEG_HEADER_OK;
+    break;
+  case JPEG_REACHED_EOI:
+    if (require_image)		/* Complain if application wanted an image */
+      ERREXIT(cinfo, JERR_NO_IMAGE);
+    /* Reset to start state; it would be safer to require the application to
+     * call jpeg_abort, but we can't change it now for compatibility reasons.
+     * A side effect is to free any temporary memory (there shouldn't be any).
+     */
+    jpeg_abort((j_common_ptr) cinfo); /* sets state = DSTATE_START */
+    retcode = JPEG_HEADER_TABLES_ONLY;
+    break;
+  case JPEG_SUSPENDED:
+    /* no work */
+    break;
+  }
+
+  return retcode;
+}
+
+
+/*
+ * Consume data in advance of what the decompressor requires.
+ * This can be called at any time once the decompressor object has
+ * been created and a data source has been set up.
+ *
+ * This routine is essentially a state machine that handles a couple
+ * of critical state-transition actions, namely initial setup and
+ * transition from header scanning to ready-for-start_decompress.
+ * All the actual input is done via the input controller's consume_input
+ * method.
+ */
+
+GLOBAL(int)
+jpeg_consume_input (j_decompress_ptr cinfo)
+{
+  int retcode = JPEG_SUSPENDED;
+
+  /* NB: every possible DSTATE value should be listed in this switch */
+  switch (cinfo->global_state) {
+  case DSTATE_START:
+    /* Start-of-datastream actions: reset appropriate modules */
+    (*cinfo->inputctl->reset_input_controller) (cinfo);
+    /* Initialize application's data source module */
+    (*cinfo->src->init_source) (cinfo);
+    cinfo->global_state = DSTATE_INHEADER;
+    /*FALLTHROUGH*/
+  case DSTATE_INHEADER:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    if (retcode == JPEG_REACHED_SOS) { /* Found SOS, prepare to decompress */
+      /* Set up default parameters based on header data */
+      default_decompress_parms(cinfo);
+      /* Set global state: ready for start_decompress */
+      cinfo->global_state = DSTATE_READY;
+    }
+    break;
+  case DSTATE_READY:
+    /* Can't advance past first SOS until start_decompress is called */
+    retcode = JPEG_REACHED_SOS;
+    break;
+  case DSTATE_PRELOAD:
+  case DSTATE_PRESCAN:
+  case DSTATE_SCANNING:
+  case DSTATE_RAW_OK:
+  case DSTATE_BUFIMAGE:
+  case DSTATE_BUFPOST:
+  case DSTATE_STOPPING:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    break;
+  default:
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  return retcode;
+}
+
+
+/*
+ * Have we finished reading the input file?
+ */
+
+GLOBAL(boolean)
+jpeg_input_complete (j_decompress_ptr cinfo)
+{
+  /* Check for valid jpeg object */
+  if (cinfo->global_state < DSTATE_START ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->eoi_reached;
+}
+
+
+/*
+ * Is there more than one scan?
+ */
+
+GLOBAL(boolean)
+jpeg_has_multiple_scans (j_decompress_ptr cinfo)
+{
+  /* Only valid after jpeg_read_header completes */
+  if (cinfo->global_state < DSTATE_READY ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->has_multiple_scans;
+}
+
+
+/*
+ * Finish JPEG decompression.
+ *
+ * This will normally just verify the file trailer and release temp storage.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_finish_decompress (j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && ! cinfo->buffered_image) {
+    /* Terminate final pass of non-buffered mode */
+    if (cinfo->output_scanline < cinfo->output_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state == DSTATE_BUFIMAGE) {
+    /* Finishing after a buffered-image operation */
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state != DSTATE_STOPPING) {
+    /* STOPPING = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read until EOI */
+  while (! cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;		/* Suspend, come back later */
+  }
+  /* Do final cleanup */
+  (*cinfo->src->term_source) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr) cinfo);
+  return TRUE;
+}
diff --git a/jdapistd.c b/jdapistd.c
new file mode 100644
index 0000000..2343da5
--- /dev/null
+++ b/jdapistd.c
@@ -0,0 +1,277 @@
+/*
+ * jdapistd.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-decompression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_decompress, it will end up linking in the entire decompressor.
+ * We thus must separate this file from jdapimin.c to avoid linking the
+ * whole decompression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+
+/* Forward declarations */
+LOCAL(boolean) output_pass_setup JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Decompression initialization.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * If a multipass operating mode was selected, this will do all but the
+ * last pass, and thus may take a great deal of time.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_start_decompress (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize master control, select active modules */
+    jinit_master_decompress(cinfo);
+    if (cinfo->buffered_image) {
+      /* No more work here; expecting jpeg_start_output next */
+      cinfo->global_state = DSTATE_BUFIMAGE;
+      return TRUE;
+    }
+    cinfo->global_state = DSTATE_PRELOAD;
+  }
+  if (cinfo->global_state == DSTATE_PRELOAD) {
+    /* If file has multiple scans, absorb them all into the coef buffer */
+    if (cinfo->inputctl->has_multiple_scans) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+      for (;;) {
+	int retcode;
+	/* Call progress monitor hook if present */
+	if (cinfo->progress != NULL)
+	  (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+	/* Absorb some more input */
+	retcode = (*cinfo->inputctl->consume_input) (cinfo);
+	if (retcode == JPEG_SUSPENDED)
+	  return FALSE;
+	if (retcode == JPEG_REACHED_EOI)
+	  break;
+	/* Advance progress counter if appropriate */
+	if (cinfo->progress != NULL &&
+	    (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+	  if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+	    /* jdmaster underestimated number of scans; ratchet up one scan */
+	    cinfo->progress->pass_limit += (long) cinfo->total_iMCU_rows;
+	  }
+	}
+      }
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+    }
+    cinfo->output_scan_number = cinfo->input_scan_number;
+  } else if (cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any dummy output passes, and set up for the final pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Set up for an output pass, and perform any dummy pass(es) needed.
+ * Common subroutine for jpeg_start_decompress and jpeg_start_output.
+ * Entry: global_state = DSTATE_PRESCAN only if previously suspended.
+ * Exit: If done, returns TRUE and sets global_state for proper output mode.
+ *       If suspended, returns FALSE and sets global_state = DSTATE_PRESCAN.
+ */
+
+LOCAL(boolean)
+output_pass_setup (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state != DSTATE_PRESCAN) {
+    /* First call: do pass setup */
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+    cinfo->global_state = DSTATE_PRESCAN;
+  }
+  /* Loop over any required dummy passes */
+  while (cinfo->master->is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Crank through the dummy pass */
+    while (cinfo->output_scanline < cinfo->output_height) {
+      JDIMENSION last_scanline;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL) {
+	cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+	cinfo->progress->pass_limit = (long) cinfo->output_height;
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      }
+      /* Process some data */
+      last_scanline = cinfo->output_scanline;
+      (*cinfo->main->process_data) (cinfo, (JSAMPARRAY) NULL,
+				    &cinfo->output_scanline, (JDIMENSION) 0);
+      if (cinfo->output_scanline == last_scanline)
+	return FALSE;		/* No progress made, must suspend */
+    }
+    /* Finish up dummy pass, and set up for another one */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  }
+  /* Ready for application to drive output pass through
+   * jpeg_read_scanlines or jpeg_read_raw_data.
+   */
+  cinfo->global_state = cinfo->raw_data_out ? DSTATE_RAW_OK : DSTATE_SCANNING;
+  return TRUE;
+}
+
+
+/*
+ * Read some scanlines of data from the JPEG decompressor.
+ *
+ * The return value will be the number of lines actually read.
+ * This may be less than the number requested in several cases,
+ * including bottom of image, data source suspension, and operating
+ * modes that emit multiple scanlines at a time.
+ *
+ * Note: we warn about excess calls to jpeg_read_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * an oversize buffer (max_lines > scanlines remaining) is not an error.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_read_scanlines (j_decompress_ptr cinfo, JSAMPARRAY scanlines,
+		     JDIMENSION max_lines)
+{
+  JDIMENSION row_ctr;
+
+  if (cinfo->global_state != DSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Process some data */
+  row_ctr = 0;
+  (*cinfo->main->process_data) (cinfo, scanlines, &row_ctr, max_lines);
+  cinfo->output_scanline += row_ctr;
+  return row_ctr;
+}
+
+
+/*
+ * Alternate entry point to read raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_read_raw_data (j_decompress_ptr cinfo, JSAMPIMAGE data,
+		    JDIMENSION max_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->global_state != DSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Verify that at least one iMCU row can be returned. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * cinfo->_min_DCT_scaled_size;
+  if (max_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Decompress directly into user's buffer. */
+  if (! (*cinfo->coef->decompress_data) (cinfo, data))
+    return 0;			/* suspension forced, can do nothing more */
+
+  /* OK, we processed one iMCU row. */
+  cinfo->output_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
+
+
+/* Additional entry points for buffered-image mode. */
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Initialize for an output pass in buffered-image mode.
+ */
+
+GLOBAL(boolean)
+jpeg_start_output (j_decompress_ptr cinfo, int scan_number)
+{
+  if (cinfo->global_state != DSTATE_BUFIMAGE &&
+      cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Limit scan number to valid range */
+  if (scan_number <= 0)
+    scan_number = 1;
+  if (cinfo->inputctl->eoi_reached &&
+      scan_number > cinfo->input_scan_number)
+    scan_number = cinfo->input_scan_number;
+  cinfo->output_scan_number = scan_number;
+  /* Perform any dummy output passes, and set up for the real pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Finish up after an output pass in buffered-image mode.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_finish_output (j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && cinfo->buffered_image) {
+    /* Terminate this pass. */
+    /* We do not require the whole pass to have been completed. */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_BUFPOST;
+  } else if (cinfo->global_state != DSTATE_BUFPOST) {
+    /* BUFPOST = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read markers looking for SOS or EOI */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+	 ! cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;		/* Suspend, come back later */
+  }
+  cinfo->global_state = DSTATE_BUFIMAGE;
+  return TRUE;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
diff --git a/jdatadst.c b/jdatadst.c
new file mode 100644
index 0000000..2f48869
--- /dev/null
+++ b/jdatadst.c
@@ -0,0 +1,277 @@
+/*
+ * jdatadst.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains compression data destination routines for the case of
+ * emitting JPEG data to memory or to a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different destination manager.
+ * IMPORTANT: we assume that fwrite() will correctly transcribe an array of
+ * JOCTETs into 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare malloc(),free() */
+extern void * malloc JPP((size_t size));
+extern void free JPP((void *ptr));
+#endif
+
+
+/* Expanded data destination object for stdio output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  FILE * outfile;		/* target stream */
+  JOCTET * buffer;		/* start of buffer */
+} my_destination_mgr;
+
+typedef my_destination_mgr * my_dest_ptr;
+
+#define OUTPUT_BUF_SIZE  4096	/* choose an efficiently fwrite'able size */
+
+
+#if JPEG_LIB_VERSION >= 80
+/* Expanded data destination object for memory output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  unsigned char ** outbuffer;	/* target buffer */
+  unsigned long * outsize;
+  unsigned char * newbuffer;	/* newly allocated buffer */
+  JOCTET * buffer;		/* start of buffer */
+  size_t bufsize;
+} my_mem_destination_mgr;
+
+typedef my_mem_destination_mgr * my_mem_dest_ptr;
+#endif
+
+
+/*
+ * Initialize destination --- called by jpeg_start_compress
+ * before any data is actually written.
+ */
+
+METHODDEF(void)
+init_destination (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+
+  /* Allocate the output buffer --- it will be released when done with image */
+  dest->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  OUTPUT_BUF_SIZE * SIZEOF(JOCTET));
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+}
+
+#if JPEG_LIB_VERSION >= 80
+METHODDEF(void)
+init_mem_destination (j_compress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+#endif
+
+
+/*
+ * Empty the output buffer --- called whenever buffer fills up.
+ *
+ * In typical applications, this should write the entire output buffer
+ * (ignoring the current state of next_output_byte & free_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been dumped.
+ *
+ * In applications that need to be able to suspend compression due to output
+ * overrun, a FALSE return indicates that the buffer cannot be emptied now.
+ * In this situation, the compressor will return to its caller (possibly with
+ * an indication that it has not accepted all the supplied scanlines).  The
+ * application should resume compression after it has made more room in the
+ * output buffer.  Note that there are substantial restrictions on the use of
+ * suspension --- see the documentation.
+ *
+ * When suspending, the compressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_output_byte & free_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point will be regenerated after resumption, so do not
+ * write it out when emptying the buffer externally.
+ */
+
+METHODDEF(boolean)
+empty_output_buffer (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+
+  if (JFWRITE(dest->outfile, dest->buffer, OUTPUT_BUF_SIZE) !=
+      (size_t) OUTPUT_BUF_SIZE)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+
+  return TRUE;
+}
+
+#if JPEG_LIB_VERSION >= 80
+METHODDEF(boolean)
+empty_mem_output_buffer (j_compress_ptr cinfo)
+{
+  size_t nextsize;
+  JOCTET * nextbuffer;
+  my_mem_dest_ptr dest = (my_mem_dest_ptr) cinfo->dest;
+
+  /* Try to allocate new buffer with double size */
+  nextsize = dest->bufsize * 2;
+  nextbuffer = malloc(nextsize);
+
+  if (nextbuffer == NULL)
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+
+  MEMCOPY(nextbuffer, dest->buffer, dest->bufsize);
+
+  if (dest->newbuffer != NULL)
+    free(dest->newbuffer);
+
+  dest->newbuffer = nextbuffer;
+
+  dest->pub.next_output_byte = nextbuffer + dest->bufsize;
+  dest->pub.free_in_buffer = dest->bufsize;
+
+  dest->buffer = nextbuffer;
+  dest->bufsize = nextsize;
+
+  return TRUE;
+}
+#endif
+
+
+/*
+ * Terminate destination --- called by jpeg_finish_compress
+ * after all data has been written.  Usually needs to flush buffer.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_destination (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+  size_t datacount = OUTPUT_BUF_SIZE - dest->pub.free_in_buffer;
+
+  /* Write any data remaining in the buffer */
+  if (datacount > 0) {
+    if (JFWRITE(dest->outfile, dest->buffer, datacount) != datacount)
+      ERREXIT(cinfo, JERR_FILE_WRITE);
+  }
+  fflush(dest->outfile);
+  /* Make sure we wrote the output file OK */
+  if (ferror(dest->outfile))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+#if JPEG_LIB_VERSION >= 80
+METHODDEF(void)
+term_mem_destination (j_compress_ptr cinfo)
+{
+  my_mem_dest_ptr dest = (my_mem_dest_ptr) cinfo->dest;
+
+  *dest->outbuffer = dest->buffer;
+  *dest->outsize = dest->bufsize - dest->pub.free_in_buffer;
+}
+#endif
+
+
+/*
+ * Prepare for output to a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing compression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_dest (j_compress_ptr cinfo, FILE * outfile)
+{
+  my_dest_ptr dest;
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same file without re-executing jpeg_stdio_dest.
+   * This makes it dangerous to use this manager and a different destination
+   * manager serially with the same JPEG object, because their private object
+   * sizes may be different.  Caveat programmer.
+   */
+  if (cinfo->dest == NULL) {	/* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_destination_mgr));
+  }
+
+  dest = (my_dest_ptr) cinfo->dest;
+  dest->pub.init_destination = init_destination;
+  dest->pub.empty_output_buffer = empty_output_buffer;
+  dest->pub.term_destination = term_destination;
+  dest->outfile = outfile;
+}
+
+
+#if JPEG_LIB_VERSION >= 80
+/*
+ * Prepare for output to a memory buffer.
+ * The caller may supply an own initial buffer with appropriate size.
+ * Otherwise, or when the actual data output exceeds the given size,
+ * the library adapts the buffer size as necessary.
+ * The standard library functions malloc/free are used for allocating
+ * larger memory, so the buffer is available to the application after
+ * finishing compression, and then the application is responsible for
+ * freeing the requested memory.
+ */
+
+GLOBAL(void)
+jpeg_mem_dest (j_compress_ptr cinfo,
+	       unsigned char ** outbuffer, unsigned long * outsize)
+{
+  my_mem_dest_ptr dest;
+
+  if (outbuffer == NULL || outsize == NULL)	/* sanity check */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same buffer without re-executing jpeg_mem_dest.
+   */
+  if (cinfo->dest == NULL) {	/* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_mem_destination_mgr));
+  }
+
+  dest = (my_mem_dest_ptr) cinfo->dest;
+  dest->pub.init_destination = init_mem_destination;
+  dest->pub.empty_output_buffer = empty_mem_output_buffer;
+  dest->pub.term_destination = term_mem_destination;
+  dest->outbuffer = outbuffer;
+  dest->outsize = outsize;
+  dest->newbuffer = NULL;
+
+  if (*outbuffer == NULL || *outsize == 0) {
+    /* Allocate initial buffer */
+    dest->newbuffer = *outbuffer = malloc(OUTPUT_BUF_SIZE);
+    if (dest->newbuffer == NULL)
+      ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+    *outsize = OUTPUT_BUF_SIZE;
+  }
+
+  dest->pub.next_output_byte = dest->buffer = *outbuffer;
+  dest->pub.free_in_buffer = dest->bufsize = *outsize;
+}
+#endif
diff --git a/jdatasrc.c b/jdatasrc.c
new file mode 100644
index 0000000..7609f76
--- /dev/null
+++ b/jdatasrc.c
@@ -0,0 +1,280 @@
+/*
+ * jdatasrc.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains decompression data source routines for the case of
+ * reading JPEG data from memory or from a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different source manager.
+ * IMPORTANT: we assume that fread() will correctly transcribe an array of
+ * JOCTETs from 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/* Expanded data source object for stdio input */
+
+typedef struct {
+  struct jpeg_source_mgr pub;	/* public fields */
+
+  FILE * infile;		/* source stream */
+  JOCTET * buffer;		/* start of buffer */
+  boolean start_of_file;	/* have we gotten any data yet? */
+} my_source_mgr;
+
+typedef my_source_mgr * my_src_ptr;
+
+#define INPUT_BUF_SIZE  4096	/* choose an efficiently fread'able size */
+
+
+/*
+ * Initialize source --- called by jpeg_read_header
+ * before any data is actually read.
+ */
+
+METHODDEF(void)
+init_source (j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+
+  /* We reset the empty-input-file flag for each image,
+   * but we don't clear the input buffer.
+   * This is correct behavior for reading a series of images from one source.
+   */
+  src->start_of_file = TRUE;
+}
+
+#if JPEG_LIB_VERSION >= 80
+METHODDEF(void)
+init_mem_source (j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+#endif
+
+
+/*
+ * Fill the input buffer --- called whenever buffer is emptied.
+ *
+ * In typical applications, this should read fresh data into the buffer
+ * (ignoring the current state of next_input_byte & bytes_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been reloaded.  It is not necessary to
+ * fill the buffer entirely, only to obtain at least one more byte.
+ *
+ * There is no such thing as an EOF return.  If the end of the file has been
+ * reached, the routine has a choice of ERREXIT() or inserting fake data into
+ * the buffer.  In most cases, generating a warning message and inserting a
+ * fake EOI marker is the best course of action --- this will allow the
+ * decompressor to output however much of the image is there.  However,
+ * the resulting error message is misleading if the real problem is an empty
+ * input file, so we handle that case specially.
+ *
+ * In applications that need to be able to suspend compression due to input
+ * not being available yet, a FALSE return indicates that no more data can be
+ * obtained right now, but more may be forthcoming later.  In this situation,
+ * the decompressor will return to its caller (with an indication of the
+ * number of scanlines it has read, if any).  The application should resume
+ * decompression after it has loaded more data into the input buffer.  Note
+ * that there are substantial restrictions on the use of suspension --- see
+ * the documentation.
+ *
+ * When suspending, the decompressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_input_byte & bytes_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point must be rescanned after resumption, so move it to
+ * the front of the buffer rather than discarding it.
+ */
+
+METHODDEF(boolean)
+fill_input_buffer (j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+  size_t nbytes;
+
+  nbytes = JFREAD(src->infile, src->buffer, INPUT_BUF_SIZE);
+
+  if (nbytes <= 0) {
+    if (src->start_of_file)	/* Treat empty input file as fatal error */
+      ERREXIT(cinfo, JERR_INPUT_EMPTY);
+    WARNMS(cinfo, JWRN_JPEG_EOF);
+    /* Insert a fake EOI marker */
+    src->buffer[0] = (JOCTET) 0xFF;
+    src->buffer[1] = (JOCTET) JPEG_EOI;
+    nbytes = 2;
+  }
+
+  src->pub.next_input_byte = src->buffer;
+  src->pub.bytes_in_buffer = nbytes;
+  src->start_of_file = FALSE;
+
+  return TRUE;
+}
+
+#if JPEG_LIB_VERSION >= 80
+METHODDEF(boolean)
+fill_mem_input_buffer (j_decompress_ptr cinfo)
+{
+  static JOCTET mybuffer[4];
+
+  /* The whole JPEG data is expected to reside in the supplied memory
+   * buffer, so any request for more data beyond the given buffer size
+   * is treated as an error.
+   */
+  WARNMS(cinfo, JWRN_JPEG_EOF);
+  /* Insert a fake EOI marker */
+  mybuffer[0] = (JOCTET) 0xFF;
+  mybuffer[1] = (JOCTET) JPEG_EOI;
+
+  cinfo->src->next_input_byte = mybuffer;
+  cinfo->src->bytes_in_buffer = 2;
+
+  return TRUE;
+}
+#endif
+
+
+/*
+ * Skip data --- used to skip over a potentially large amount of
+ * uninteresting data (such as an APPn marker).
+ *
+ * Writers of suspendable-input applications must note that skip_input_data
+ * is not granted the right to give a suspension return.  If the skip extends
+ * beyond the data currently in the buffer, the buffer can be marked empty so
+ * that the next read will cause a fill_input_buffer call that can suspend.
+ * Arranging for additional bytes to be discarded before reloading the input
+ * buffer is the application writer's problem.
+ */
+
+METHODDEF(void)
+skip_input_data (j_decompress_ptr cinfo, long num_bytes)
+{
+  struct jpeg_source_mgr * src = cinfo->src;
+
+  /* Just a dumb implementation for now.  Could use fseek() except
+   * it doesn't work on pipes.  Not clear that being smart is worth
+   * any trouble anyway --- large skips are infrequent.
+   */
+  if (num_bytes > 0) {
+    while (num_bytes > (long) src->bytes_in_buffer) {
+      num_bytes -= (long) src->bytes_in_buffer;
+      (void) (*src->fill_input_buffer) (cinfo);
+      /* note we assume that fill_input_buffer will never return FALSE,
+       * so suspension need not be handled.
+       */
+    }
+    src->next_input_byte += (size_t) num_bytes;
+    src->bytes_in_buffer -= (size_t) num_bytes;
+  }
+}
+
+
+/*
+ * An additional method that can be provided by data source modules is the
+ * resync_to_restart method for error recovery in the presence of RST markers.
+ * For the moment, this source module just uses the default resync method
+ * provided by the JPEG library.  That method assumes that no backtracking
+ * is possible.
+ */
+
+
+/*
+ * Terminate source --- called by jpeg_finish_decompress
+ * after all data has been read.  Often a no-op.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_source (j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Prepare for input from a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing decompression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_src (j_decompress_ptr cinfo, FILE * infile)
+{
+  my_src_ptr src;
+
+  /* The source object and input buffer are made permanent so that a series
+   * of JPEG images can be read from the same file by calling jpeg_stdio_src
+   * only before the first one.  (If we discarded the buffer at the end of
+   * one image, we'd likely lose the start of the next one.)
+   * This makes it unsafe to use this manager and a different source
+   * manager serially with the same JPEG object.  Caveat programmer.
+   */
+  if (cinfo->src == NULL) {	/* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_source_mgr));
+    src = (my_src_ptr) cinfo->src;
+    src->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  INPUT_BUF_SIZE * SIZEOF(JOCTET));
+  }
+
+  src = (my_src_ptr) cinfo->src;
+  src->pub.init_source = init_source;
+  src->pub.fill_input_buffer = fill_input_buffer;
+  src->pub.skip_input_data = skip_input_data;
+  src->pub.resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->pub.term_source = term_source;
+  src->infile = infile;
+  src->pub.bytes_in_buffer = 0; /* forces fill_input_buffer on first read */
+  src->pub.next_input_byte = NULL; /* until buffer loaded */
+}
+
+
+#if JPEG_LIB_VERSION >= 80
+/*
+ * Prepare for input from a supplied memory buffer.
+ * The buffer must contain the whole JPEG data.
+ */
+
+GLOBAL(void)
+jpeg_mem_src (j_decompress_ptr cinfo,
+	      unsigned char * inbuffer, unsigned long insize)
+{
+  struct jpeg_source_mgr * src;
+
+  if (inbuffer == NULL || insize == 0)	/* Treat empty input as fatal error */
+    ERREXIT(cinfo, JERR_INPUT_EMPTY);
+
+  /* The source object is made permanent so that a series of JPEG images
+   * can be read from the same buffer by calling jpeg_mem_src only before
+   * the first one.
+   */
+  if (cinfo->src == NULL) {	/* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(struct jpeg_source_mgr));
+  }
+
+  src = cinfo->src;
+  src->init_source = init_mem_source;
+  src->fill_input_buffer = fill_mem_input_buffer;
+  src->skip_input_data = skip_input_data;
+  src->resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->term_source = term_source;
+  src->bytes_in_buffer = (size_t) insize;
+  src->next_input_byte = (JOCTET *) inbuffer;
+}
+#endif
diff --git a/jdcoefct.c b/jdcoefct.c
new file mode 100644
index 0000000..48a9fc6
--- /dev/null
+++ b/jdcoefct.c
@@ -0,0 +1,749 @@
+/*
+ * jdcoefct.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the coefficient buffer controller for decompression.
+ * This controller is the top level of the JPEG decompressor proper.
+ * The coefficient buffer lies between entropy decoding and inverse-DCT steps.
+ *
+ * In buffered-image mode, this controller is the interface between
+ * input-oriented processing and output-oriented processing.
+ * Also, the input side (only) is used when reading a file for transcoding.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+/* Block smoothing is only applicable for progressive JPEG, so: */
+#ifndef D_PROGRESSIVE_SUPPORTED
+#undef BLOCK_SMOOTHING_SUPPORTED
+#endif
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_coef_controller pub; /* public fields */
+
+  /* These variables keep track of the current location of the input side. */
+  /* cinfo->input_iMCU_row is also used for this. */
+  JDIMENSION MCU_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* The output side's location is represented by cinfo->output_iMCU_row. */
+
+  /* In single-pass modes, it's sufficient to buffer just one MCU.
+   * We allocate a workspace of D_MAX_BLOCKS_IN_MCU coefficient blocks,
+   * and let the entropy decoder write into that workspace each time.
+   * (On 80x86, the workspace is FAR even though it's not really very big;
+   * this is to keep the module interfaces unchanged when a large coefficient
+   * buffer is necessary.)
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays; it is used only by the input side.
+   */
+  JBLOCKROW MCU_buffer[D_MAX_BLOCKS_IN_MCU];
+
+  /* Temporary workspace for one MCU */
+  JCOEF * workspace;
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+#endif
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  /* When doing block smoothing, we latch coefficient Al values here */
+  int * coef_bits_latch;
+#define SAVED_COEFS  6		/* we save coef_bits[0..5] */
+#endif
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+/* Forward declarations */
+METHODDEF(int) decompress_onepass
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+METHODDEF(int) decompress_data
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#endif
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+LOCAL(boolean) smoothing_ok JPP((j_decompress_ptr cinfo));
+METHODDEF(int) decompress_smooth_data
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#endif
+
+
+LOCAL(void)
+start_iMCU_row (j_decompress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row (input side) */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (cinfo->input_iMCU_row < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->MCU_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for an input processing pass.
+ */
+
+METHODDEF(void)
+start_input_pass (j_decompress_ptr cinfo)
+{
+  cinfo->input_iMCU_row = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Initialize for an output processing pass.
+ */
+
+METHODDEF(void)
+start_output_pass (j_decompress_ptr cinfo)
+{
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* If multipass, check to see whether to use block smoothing on this pass */
+  if (coef->pub.coef_arrays != NULL) {
+    if (cinfo->do_block_smoothing && smoothing_ok(cinfo))
+      coef->pub.decompress_data = decompress_smooth_data;
+    else
+      coef->pub.decompress_data = decompress_data;
+  }
+#endif
+  cinfo->output_iMCU_row = 0;
+}
+
+
+/*
+ * Decompress and return some data in the single-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Input and output must run in lockstep since we have only a one-MCU buffer.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(int)
+decompress_onepass (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, useful_width;
+  JSAMPARRAY output_ptr;
+  JDIMENSION start_col, output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+
+  /* Loop to process as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num <= last_MCU_col;
+	 MCU_col_num++) {
+      /* Try to fetch an MCU.  Entropy decoder expects buffer to be zeroed. */
+      jzero_far((void FAR *) coef->MCU_buffer[0],
+		(size_t) (cinfo->blocks_in_MCU * SIZEOF(JBLOCK)));
+      if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->MCU_ctr = MCU_col_num;
+	return JPEG_SUSPENDED;
+      }
+      /* Determine where data should go in output_buf and do the IDCT thing.
+       * We skip dummy blocks at the right and bottom edges (but blkn gets
+       * incremented past them!).  Note the inner loop relies on having
+       * allocated the MCU_buffer[] blocks sequentially.
+       */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	/* Don't bother to IDCT an uninteresting component. */
+	if (! compptr->component_needed) {
+	  blkn += compptr->MCU_blocks;
+	  continue;
+	}
+	inverse_DCT = cinfo->idct->inverse_DCT[compptr->component_index];
+	useful_width = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						    : compptr->last_col_width;
+	output_ptr = output_buf[compptr->component_index] +
+	  yoffset * compptr->_DCT_scaled_size;
+	start_col = MCU_col_num * compptr->MCU_sample_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (cinfo->input_iMCU_row < last_iMCU_row ||
+	      yoffset+yindex < compptr->last_row_height) {
+	    output_col = start_col;
+	    for (xindex = 0; xindex < useful_width; xindex++) {
+	      (*inverse_DCT) (cinfo, compptr,
+			      (JCOEFPTR) coef->MCU_buffer[blkn+xindex],
+			      output_ptr, output_col);
+	      output_col += compptr->_DCT_scaled_size;
+	    }
+	  }
+	  blkn += compptr->MCU_width;
+	  output_ptr += compptr->_DCT_scaled_size;
+	}
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  cinfo->output_iMCU_row++;
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Dummy consume-input routine for single-pass operation.
+ */
+
+METHODDEF(int)
+dummy_consume_data (j_decompress_ptr cinfo)
+{
+  return JPEG_SUSPENDED;	/* Always indicate nothing was done */
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Consume input data and store it in the full-image coefficient buffer.
+ * We read as much as one fully interleaved MCU row ("iMCU" row) per call,
+ * ie, v_samp_factor block rows for each component in the scan.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ */
+
+METHODDEF(int)
+consume_data (j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       cinfo->input_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, TRUE);
+    /* Note: entropy decoder expects buffer to be zeroed,
+     * but this is handled automatically by the memory manager
+     * because we requested a pre-zeroed array.
+     */
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+	    coef->MCU_buffer[blkn++] = buffer_ptr++;
+	  }
+	}
+      }
+      /* Try to fetch the MCU. */
+      if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->MCU_ctr = MCU_col_num;
+	return JPEG_SUSPENDED;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Decompress and return some data in the multi-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image.
+ */
+
+METHODDEF(int)
+decompress_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num;
+  int ci, block_row, block_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr;
+  JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number < cinfo->output_scan_number ||
+	 (cinfo->input_scan_number == cinfo->output_scan_number &&
+	  cinfo->input_iMCU_row <= cinfo->output_iMCU_row)) {
+    if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (! compptr->component_needed)
+      continue;
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[ci],
+       cinfo->output_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    inverse_DCT = cinfo->idct->inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row];
+      output_col = 0;
+      for (block_num = 0; block_num < compptr->width_in_blocks; block_num++) {
+	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) buffer_ptr,
+			output_ptr, output_col);
+	buffer_ptr++;
+	output_col += compptr->_DCT_scaled_size;
+      }
+      output_ptr += compptr->_DCT_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+
+/*
+ * This code applies interblock smoothing as described by section K.8
+ * of the JPEG standard: the first 5 AC coefficients are estimated from
+ * the DC values of a DCT block and its 8 neighboring blocks.
+ * We apply smoothing only for progressive JPEG decoding, and only if
+ * the coefficients it can estimate are not yet known to full precision.
+ */
+
+/* Natural-order array positions of the first 5 zigzag-order coefficients */
+#define Q01_POS  1
+#define Q10_POS  8
+#define Q20_POS  16
+#define Q11_POS  9
+#define Q02_POS  2
+
+/*
+ * Determine whether block smoothing is applicable and safe.
+ * We also latch the current states of the coef_bits[] entries for the
+ * AC coefficients; otherwise, if the input side of the decompressor
+ * advances into a new scan, we might think the coefficients are known
+ * more accurately than they really are.
+ */
+
+LOCAL(boolean)
+smoothing_ok (j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  boolean smoothing_useful = FALSE;
+  int ci, coefi;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtable;
+  int * coef_bits;
+  int * coef_bits_latch;
+
+  if (! cinfo->progressive_mode || cinfo->coef_bits == NULL)
+    return FALSE;
+
+  /* Allocate latch area if not already done */
+  if (coef->coef_bits_latch == NULL)
+    coef->coef_bits_latch = (int *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  cinfo->num_components *
+				  (SAVED_COEFS * SIZEOF(int)));
+  coef_bits_latch = coef->coef_bits_latch;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* All components' quantization values must already be latched. */
+    if ((qtable = compptr->quant_table) == NULL)
+      return FALSE;
+    /* Verify DC & first 5 AC quantizers are nonzero to avoid zero-divide. */
+    if (qtable->quantval[0] == 0 ||
+	qtable->quantval[Q01_POS] == 0 ||
+	qtable->quantval[Q10_POS] == 0 ||
+	qtable->quantval[Q20_POS] == 0 ||
+	qtable->quantval[Q11_POS] == 0 ||
+	qtable->quantval[Q02_POS] == 0)
+      return FALSE;
+    /* DC values must be at least partly known for all components. */
+    coef_bits = cinfo->coef_bits[ci];
+    if (coef_bits[0] < 0)
+      return FALSE;
+    /* Block smoothing is helpful if some AC coefficients remain inaccurate. */
+    for (coefi = 1; coefi <= 5; coefi++) {
+      coef_bits_latch[coefi] = coef_bits[coefi];
+      if (coef_bits[coefi] != 0)
+	smoothing_useful = TRUE;
+    }
+    coef_bits_latch += SAVED_COEFS;
+  }
+
+  return smoothing_useful;
+}
+
+
+/*
+ * Variant of decompress_data for use when doing block smoothing.
+ */
+
+METHODDEF(int)
+decompress_smooth_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num, last_block_column;
+  int ci, block_row, block_rows, access_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr, prev_block_row, next_block_row;
+  JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+  boolean first_row, last_row;
+  JCOEF * workspace;
+  int *coef_bits;
+  JQUANT_TBL *quanttbl;
+  INT32 Q00,Q01,Q02,Q10,Q11,Q20, num;
+  int DC1,DC2,DC3,DC4,DC5,DC6,DC7,DC8,DC9;
+  int Al, pred;
+
+  /* Keep a local variable to avoid looking it up more than once */
+  workspace = coef->workspace;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+	 ! cinfo->inputctl->eoi_reached) {
+    if (cinfo->input_scan_number == cinfo->output_scan_number) {
+      /* If input is working on current scan, we ordinarily want it to
+       * have completed the current row.  But if input scan is DC,
+       * we want it to keep one row ahead so that next block row's DC
+       * values are up to date.
+       */
+      JDIMENSION delta = (cinfo->Ss == 0) ? 1 : 0;
+      if (cinfo->input_iMCU_row > cinfo->output_iMCU_row+delta)
+	break;
+    }
+    if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (! compptr->component_needed)
+      continue;
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row) {
+      block_rows = compptr->v_samp_factor;
+      access_rows = block_rows * 2; /* this and next iMCU row */
+      last_row = FALSE;
+    } else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+      access_rows = block_rows; /* this iMCU row only */
+      last_row = TRUE;
+    }
+    /* Align the virtual buffer for this component. */
+    if (cinfo->output_iMCU_row > 0) {
+      access_rows += compptr->v_samp_factor; /* prior iMCU row too */
+      buffer = (*cinfo->mem->access_virt_barray)
+	((j_common_ptr) cinfo, coef->whole_image[ci],
+	 (cinfo->output_iMCU_row - 1) * compptr->v_samp_factor,
+	 (JDIMENSION) access_rows, FALSE);
+      buffer += compptr->v_samp_factor;	/* point to current iMCU row */
+      first_row = FALSE;
+    } else {
+      buffer = (*cinfo->mem->access_virt_barray)
+	((j_common_ptr) cinfo, coef->whole_image[ci],
+	 (JDIMENSION) 0, (JDIMENSION) access_rows, FALSE);
+      first_row = TRUE;
+    }
+    /* Fetch component-dependent info */
+    coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
+    quanttbl = compptr->quant_table;
+    Q00 = quanttbl->quantval[0];
+    Q01 = quanttbl->quantval[Q01_POS];
+    Q10 = quanttbl->quantval[Q10_POS];
+    Q20 = quanttbl->quantval[Q20_POS];
+    Q11 = quanttbl->quantval[Q11_POS];
+    Q02 = quanttbl->quantval[Q02_POS];
+    inverse_DCT = cinfo->idct->inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row];
+      if (first_row && block_row == 0)
+	prev_block_row = buffer_ptr;
+      else
+	prev_block_row = buffer[block_row-1];
+      if (last_row && block_row == block_rows-1)
+	next_block_row = buffer_ptr;
+      else
+	next_block_row = buffer[block_row+1];
+      /* We fetch the surrounding DC values using a sliding-register approach.
+       * Initialize all nine here so as to do the right thing on narrow pics.
+       */
+      DC1 = DC2 = DC3 = (int) prev_block_row[0][0];
+      DC4 = DC5 = DC6 = (int) buffer_ptr[0][0];
+      DC7 = DC8 = DC9 = (int) next_block_row[0][0];
+      output_col = 0;
+      last_block_column = compptr->width_in_blocks - 1;
+      for (block_num = 0; block_num <= last_block_column; block_num++) {
+	/* Fetch current DCT block into workspace so we can modify it. */
+	jcopy_block_row(buffer_ptr, (JBLOCKROW) workspace, (JDIMENSION) 1);
+	/* Update DC values */
+	if (block_num < last_block_column) {
+	  DC3 = (int) prev_block_row[1][0];
+	  DC6 = (int) buffer_ptr[1][0];
+	  DC9 = (int) next_block_row[1][0];
+	}
+	/* Compute coefficient estimates per K.8.
+	 * An estimate is applied only if coefficient is still zero,
+	 * and is not known to be fully accurate.
+	 */
+	/* AC01 */
+	if ((Al=coef_bits[1]) != 0 && workspace[1] == 0) {
+	  num = 36 * Q00 * (DC4 - DC6);
+	  if (num >= 0) {
+	    pred = (int) (((Q01<<7) + num) / (Q01<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q01<<7) - num) / (Q01<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[1] = (JCOEF) pred;
+	}
+	/* AC10 */
+	if ((Al=coef_bits[2]) != 0 && workspace[8] == 0) {
+	  num = 36 * Q00 * (DC2 - DC8);
+	  if (num >= 0) {
+	    pred = (int) (((Q10<<7) + num) / (Q10<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q10<<7) - num) / (Q10<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[8] = (JCOEF) pred;
+	}
+	/* AC20 */
+	if ((Al=coef_bits[3]) != 0 && workspace[16] == 0) {
+	  num = 9 * Q00 * (DC2 + DC8 - 2*DC5);
+	  if (num >= 0) {
+	    pred = (int) (((Q20<<7) + num) / (Q20<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q20<<7) - num) / (Q20<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[16] = (JCOEF) pred;
+	}
+	/* AC11 */
+	if ((Al=coef_bits[4]) != 0 && workspace[9] == 0) {
+	  num = 5 * Q00 * (DC1 - DC3 - DC7 + DC9);
+	  if (num >= 0) {
+	    pred = (int) (((Q11<<7) + num) / (Q11<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q11<<7) - num) / (Q11<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[9] = (JCOEF) pred;
+	}
+	/* AC02 */
+	if ((Al=coef_bits[5]) != 0 && workspace[2] == 0) {
+	  num = 9 * Q00 * (DC4 + DC6 - 2*DC5);
+	  if (num >= 0) {
+	    pred = (int) (((Q02<<7) + num) / (Q02<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q02<<7) - num) / (Q02<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[2] = (JCOEF) pred;
+	}
+	/* OK, do the IDCT */
+	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) workspace,
+			output_ptr, output_col);
+	/* Advance for next column */
+	DC1 = DC2; DC2 = DC3;
+	DC4 = DC5; DC5 = DC6;
+	DC7 = DC8; DC8 = DC9;
+	buffer_ptr++, prev_block_row++, next_block_row++;
+	output_col += compptr->_DCT_scaled_size;
+      }
+      output_ptr += compptr->_DCT_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* BLOCK_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+jinit_d_coef_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_d_coef_controller *) coef;
+  coef->pub.start_input_pass = start_input_pass;
+  coef->pub.start_output_pass = start_output_pass;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  coef->coef_bits_latch = NULL;
+#endif
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    /* Note we ask for a pre-zeroed array. */
+    int ci, access_rows;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      access_rows = compptr->v_samp_factor;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+      /* If block smoothing could be used, need a bigger window */
+      if (cinfo->progressive_mode)
+	access_rows *= 3;
+#endif
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, TRUE,
+	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
+				(long) compptr->h_samp_factor),
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor),
+	 (JDIMENSION) access_rows);
+    }
+    coef->pub.consume_data = consume_data;
+    coef->pub.decompress_data = decompress_data;
+    coef->pub.coef_arrays = coef->whole_image; /* link to virtual arrays */
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  D_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+    for (i = 0; i < D_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->pub.consume_data = dummy_consume_data;
+    coef->pub.decompress_data = decompress_onepass;
+    coef->pub.coef_arrays = NULL; /* flag for no virtual arrays */
+  }
+
+  /* Allocate the workspace buffer */
+  coef->workspace = (JCOEF *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+                                SIZEOF(JCOEF) * DCTSIZE2);
+}
diff --git a/jdcolor.c b/jdcolor.c
new file mode 100644
index 0000000..e02ea4f
--- /dev/null
+++ b/jdcolor.c
@@ -0,0 +1,415 @@
+/*
+ * jdcolor.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2009, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains output colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_deconverter pub; /* public fields */
+
+  /* Private state for YCC->RGB conversion */
+  int * Cr_r_tab;		/* => table for Cr to R conversion */
+  int * Cb_b_tab;		/* => table for Cb to B conversion */
+  INT32 * Cr_g_tab;		/* => table for Cr to G conversion */
+  INT32 * Cb_g_tab;		/* => table for Cb to G conversion */
+} my_color_deconverter;
+
+typedef my_color_deconverter * my_cconvert_ptr;
+
+
+/**************** YCbCr -> RGB conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *	R = Y                + 1.40200 * Cr
+ *	G = Y - 0.34414 * Cb - 0.71414 * Cr
+ *	B = Y + 1.77200 * Cb
+ * where Cb and Cr represent the incoming values less CENTERJSAMPLE.
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ * Notice that Y, being an integral input, does not contribute any fraction
+ * so it need not participate in the rounding.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times Cb and Cr for all possible values.
+ * For 8-bit JSAMPLEs this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The Cr=>R and Cb=>B values can be rounded to integers in advance; the
+ * values for the G calculation are left scaled up, since we must add them
+ * together before rounding.
+ */
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define ONE_HALF	((INT32) 1 << (SCALEBITS-1))
+#define FIX(x)		((INT32) ((x) * (1L<<SCALEBITS) + 0.5))
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table (j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int i;
+  INT32 x;
+  SHIFT_TEMPS
+
+  cconvert->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  cconvert->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  cconvert->Cr_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+  cconvert->Cb_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+
+  for (i = 0, x = -CENTERJSAMPLE; i <= MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0..MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    cconvert->Cr_r_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    cconvert->Cb_b_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    cconvert->Cr_g_tab[i] = (- FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    cconvert->Cb_g_tab[i] = (- FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the output colorspace.
+ *
+ * Note that we change from noninterleaved, one-plane-per-component format
+ * to interleaved-pixel format.  The output buffer is therefore three times
+ * as wide as the input buffer.
+ * A starting row offset is provided only for the input buffer.  The caller
+ * can easily adjust the passed output_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+METHODDEF(void)
+ycc_rgb_convert (j_decompress_ptr cinfo,
+		 JSAMPIMAGE input_buf, JDIMENSION input_row,
+		 JSAMPARRAY output_buf, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int y, cb, cr;
+  register JSAMPROW outptr;
+  register JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  register int * Crrtab = cconvert->Cr_r_tab;
+  register int * Cbbtab = cconvert->Cb_b_tab;
+  register INT32 * Crgtab = cconvert->Cr_g_tab;
+  register INT32 * Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = GETJSAMPLE(inptr0[col]);
+      cb = GETJSAMPLE(inptr1[col]);
+      cr = GETJSAMPLE(inptr2[col]);
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[rgb_red[cinfo->out_color_space]] =   range_limit[y + Crrtab[cr]];
+      outptr[rgb_green[cinfo->out_color_space]] = range_limit[y +
+			      ((int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+						 SCALEBITS))];
+      outptr[rgb_blue[cinfo->out_color_space]] =  range_limit[y + Cbbtab[cb]];
+      outptr += rgb_pixelsize[cinfo->out_color_space];
+    }
+  }
+}
+
+
+/**************** Cases other than YCbCr -> RGB **************/
+
+
+/*
+ * Color conversion for no colorspace change: just copy the data,
+ * converting from separate-planes to interleaved representation.
+ */
+
+METHODDEF(void)
+null_convert (j_decompress_ptr cinfo,
+	      JSAMPIMAGE input_buf, JDIMENSION input_row,
+	      JSAMPARRAY output_buf, int num_rows)
+{
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION count;
+  register int num_components = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->output_width;
+  int ci;
+
+  while (--num_rows >= 0) {
+    for (ci = 0; ci < num_components; ci++) {
+      inptr = input_buf[ci][input_row];
+      outptr = output_buf[0] + ci;
+      for (count = num_cols; count > 0; count--) {
+	*outptr = *inptr++;	/* needn't bother with GETJSAMPLE() here */
+	outptr += num_components;
+      }
+    }
+    input_row++;
+    output_buf++;
+  }
+}
+
+
+/*
+ * Color conversion for grayscale: just copy the data.
+ * This also works for YCbCr -> grayscale conversion, in which
+ * we just copy the Y (luminance) component and ignore chrominance.
+ */
+
+METHODDEF(void)
+grayscale_convert (j_decompress_ptr cinfo,
+		   JSAMPIMAGE input_buf, JDIMENSION input_row,
+		   JSAMPARRAY output_buf, int num_rows)
+{
+  jcopy_sample_rows(input_buf[0], (int) input_row, output_buf, 0,
+		    num_rows, cinfo->output_width);
+}
+
+
+/*
+ * Convert grayscale to RGB: just duplicate the graylevel three times.
+ * This is provided to support applications that don't want to cope
+ * with grayscale as a separate case.
+ */
+
+METHODDEF(void)
+gray_rgb_convert (j_decompress_ptr cinfo,
+		  JSAMPIMAGE input_buf, JDIMENSION input_row,
+		  JSAMPARRAY output_buf, int num_rows)
+{
+  register JSAMPROW inptr, outptr;
+  JSAMPLE *maxinptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  int rindex = rgb_red[cinfo->out_color_space];
+  int gindex = rgb_green[cinfo->out_color_space];
+  int bindex = rgb_blue[cinfo->out_color_space];
+  int rgbstride = rgb_pixelsize[cinfo->out_color_space];
+
+  while (--num_rows >= 0) {
+    inptr = input_buf[0][input_row++];
+    maxinptr = &inptr[num_cols];
+    outptr = *output_buf++;
+    for (; inptr < maxinptr; inptr++, outptr += rgbstride) {
+      /* We can dispense with GETJSAMPLE() here */
+      outptr[rindex] = outptr[gindex] = outptr[bindex] = *inptr;
+    }
+  }
+}
+
+
+/*
+ * Adobe-style YCCK->CMYK conversion.
+ * We convert YCbCr to R=1-C, G=1-M, and B=1-Y using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume build_ycc_rgb_table has been called.
+ */
+
+METHODDEF(void)
+ycck_cmyk_convert (j_decompress_ptr cinfo,
+		   JSAMPIMAGE input_buf, JDIMENSION input_row,
+		   JSAMPARRAY output_buf, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int y, cb, cr;
+  register JSAMPROW outptr;
+  register JSAMPROW inptr0, inptr1, inptr2, inptr3;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  register int * Crrtab = cconvert->Cr_r_tab;
+  register int * Cbbtab = cconvert->Cb_b_tab;
+  register INT32 * Crgtab = cconvert->Cr_g_tab;
+  register INT32 * Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    inptr3 = input_buf[3][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = GETJSAMPLE(inptr0[col]);
+      cb = GETJSAMPLE(inptr1[col]);
+      cr = GETJSAMPLE(inptr2[col]);
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[0] = range_limit[MAXJSAMPLE - (y + Crrtab[cr])];	/* red */
+      outptr[1] = range_limit[MAXJSAMPLE - (y +			/* green */
+			      ((int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+						 SCALEBITS)))];
+      outptr[2] = range_limit[MAXJSAMPLE - (y + Cbbtab[cb])];	/* blue */
+      /* K passes through unchanged */
+      outptr[3] = inptr3[col];	/* don't need GETJSAMPLE here */
+      outptr += 4;
+    }
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+start_pass_dcolor (j_decompress_ptr cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for output colorspace conversion.
+ */
+
+GLOBAL(void)
+jinit_color_deconverter (j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+  int ci;
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_color_deconverter));
+  cinfo->cconvert = (struct jpeg_color_deconverter *) cconvert;
+  cconvert->pub.start_pass = start_pass_dcolor;
+
+  /* Make sure num_components agrees with jpeg_color_space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  default:			/* JCS_UNKNOWN can be anything */
+    if (cinfo->num_components < 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+  }
+
+  /* Set out_color_components and conversion method based on requested space.
+   * Also clear the component_needed flags for any unused components,
+   * so that earlier pipeline stages can avoid useless computation.
+   */
+
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    if (cinfo->jpeg_color_space == JCS_GRAYSCALE ||
+	cinfo->jpeg_color_space == JCS_YCbCr) {
+      cconvert->pub.color_convert = grayscale_convert;
+      /* For color->grayscale conversion, only the Y (0) component is needed */
+      for (ci = 1; ci < cinfo->num_components; ci++)
+	cinfo->comp_info[ci].component_needed = FALSE;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+    cinfo->out_color_components = rgb_pixelsize[cinfo->out_color_space];
+    if (cinfo->jpeg_color_space == JCS_YCbCr) {
+      if (jsimd_can_ycc_rgb())
+        cconvert->pub.color_convert = jsimd_ycc_rgb_convert;
+      else {
+        cconvert->pub.color_convert = ycc_rgb_convert;
+        build_ycc_rgb_table(cinfo);
+      }
+    } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
+      cconvert->pub.color_convert = gray_rgb_convert;
+    } else if (cinfo->jpeg_color_space == cinfo->out_color_space &&
+      rgb_pixelsize[cinfo->out_color_space] == 3) {
+      cconvert->pub.color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+    cinfo->out_color_components = 4;
+    if (cinfo->jpeg_color_space == JCS_YCCK) {
+      cconvert->pub.color_convert = ycck_cmyk_convert;
+      build_ycc_rgb_table(cinfo);
+    } else if (cinfo->jpeg_color_space == JCS_CMYK) {
+      cconvert->pub.color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:
+    /* Permit null conversion to same output space */
+    if (cinfo->out_color_space == cinfo->jpeg_color_space) {
+      cinfo->out_color_components = cinfo->num_components;
+      cconvert->pub.color_convert = null_convert;
+    } else			/* unsupported non-null conversion */
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+  }
+
+  if (cinfo->quantize_colors)
+    cinfo->output_components = 1; /* single colormapped output component */
+  else
+    cinfo->output_components = cinfo->out_color_components;
+}
diff --git a/jdct.h b/jdct.h
new file mode 100644
index 0000000..7b49a97
--- /dev/null
+++ b/jdct.h
@@ -0,0 +1,184 @@
+/*
+ * jdct.h
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This include file contains common declarations for the forward and
+ * inverse DCT modules.  These declarations are private to the DCT managers
+ * (jcdctmgr.c, jddctmgr.c) and the individual DCT algorithms.
+ * The individual DCT algorithms are kept in separate files to ease 
+ * machine-dependent tuning (e.g., assembly coding).
+ */
+
+
+/*
+ * A forward DCT routine is given a pointer to a work area of type DCTELEM[];
+ * the DCT is to be performed in-place in that buffer.  Type DCTELEM is int
+ * for 8-bit samples, INT32 for 12-bit samples.  (NOTE: Floating-point DCT
+ * implementations use an array of type FAST_FLOAT, instead.)
+ * The DCT inputs are expected to be signed (range +-CENTERJSAMPLE).
+ * The DCT outputs are returned scaled up by a factor of 8; they therefore
+ * have a range of +-8K for 8-bit data, +-128K for 12-bit data.  This
+ * convention improves accuracy in integer implementations and saves some
+ * work in floating-point ones.
+ * Quantization of the output coefficients is done by jcdctmgr.c. This
+ * step requires an unsigned type and also one with twice the bits.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#ifndef WITH_SIMD
+typedef int DCTELEM;		/* 16 or 32 bits is fine */
+typedef unsigned int UDCTELEM;
+typedef unsigned long long UDCTELEM2;
+#else
+typedef short DCTELEM;  /* prefer 16 bit with SIMD for parellelism */
+typedef unsigned short UDCTELEM;
+typedef unsigned int UDCTELEM2;
+#endif
+#else
+typedef INT32 DCTELEM;		/* must have 32 bits */
+typedef UINT32 UDCTELEM;
+typedef unsigned long long UDCTELEM2;
+#endif
+
+
+/*
+ * An inverse DCT routine is given a pointer to the input JBLOCK and a pointer
+ * to an output sample array.  The routine must dequantize the input data as
+ * well as perform the IDCT; for dequantization, it uses the multiplier table
+ * pointed to by compptr->dct_table.  The output data is to be placed into the
+ * sample array starting at a specified column.  (Any row offset needed will
+ * be applied to the array pointer before it is passed to the IDCT code.)
+ * Note that the number of samples emitted by the IDCT routine is
+ * DCT_scaled_size * DCT_scaled_size.
+ */
+
+/* typedef inverse_DCT_method_ptr is declared in jpegint.h */
+
+/*
+ * Each IDCT routine has its own ideas about the best dct_table element type.
+ */
+
+typedef MULTIPLIER ISLOW_MULT_TYPE; /* short or int, whichever is faster */
+#if BITS_IN_JSAMPLE == 8
+typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
+#define IFAST_SCALE_BITS  2	/* fractional bits in scale factors */
+#else
+typedef INT32 IFAST_MULT_TYPE;	/* need 32 bits for scaled quantizers */
+#define IFAST_SCALE_BITS  13	/* fractional bits in scale factors */
+#endif
+typedef FAST_FLOAT FLOAT_MULT_TYPE; /* preferred floating type */
+
+
+/*
+ * Each IDCT routine is responsible for range-limiting its results and
+ * converting them to unsigned form (0..MAXJSAMPLE).  The raw outputs could
+ * be quite far out of range if the input data is corrupt, so a bulletproof
+ * range-limiting step is required.  We use a mask-and-table-lookup method
+ * to do the combined operations quickly.  See the comments with
+ * prepare_range_limit_table (in jdmaster.c) for more info.
+ */
+
+#define IDCT_range_limit(cinfo)  ((cinfo)->sample_range_limit + CENTERJSAMPLE)
+
+#define RANGE_MASK  (MAXJSAMPLE * 4 + 3) /* 2 bits wider than legal samples */
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_fdct_islow		jFDislow
+#define jpeg_fdct_ifast		jFDifast
+#define jpeg_fdct_float		jFDfloat
+#define jpeg_idct_islow		jRDislow
+#define jpeg_idct_ifast		jRDifast
+#define jpeg_idct_float		jRDfloat
+#define jpeg_idct_4x4		jRD4x4
+#define jpeg_idct_2x2		jRD2x2
+#define jpeg_idct_1x1		jRD1x1
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* Extern declarations for the forward and inverse DCT routines. */
+
+EXTERN(void) jpeg_fdct_islow JPP((DCTELEM * data));
+EXTERN(void) jpeg_fdct_ifast JPP((DCTELEM * data));
+EXTERN(void) jpeg_fdct_float JPP((FAST_FLOAT * data));
+
+EXTERN(void) jpeg_idct_islow
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_ifast
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_float
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_4x4
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_2x2
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_1x1
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+
+
+/*
+ * Macros for handling fixed-point arithmetic; these are used by many
+ * but not all of the DCT/IDCT modules.
+ *
+ * All values are expected to be of type INT32.
+ * Fractional constants are scaled left by CONST_BITS bits.
+ * CONST_BITS is defined within each module using these macros,
+ * and may differ from one module to the next.
+ */
+
+#define ONE	((INT32) 1)
+#define CONST_SCALE (ONE << CONST_BITS)
+
+/* Convert a positive real constant to an integer scaled by CONST_SCALE.
+ * Caution: some C compilers fail to reduce "FIX(constant)" at compile time,
+ * thus causing a lot of useless floating-point operations at run time.
+ */
+
+#define FIX(x)	((INT32) ((x) * CONST_SCALE + 0.5))
+
+/* Descale and correctly round an INT32 value that's scaled by N bits.
+ * We assume RIGHT_SHIFT rounds towards minus infinity, so adding
+ * the fudge factor is correct for either sign of X.
+ */
+
+#define DESCALE(x,n)  RIGHT_SHIFT((x) + (ONE << ((n)-1)), n)
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * This macro is used only when the two inputs will actually be no more than
+ * 16 bits wide, so that a 16x16->32 bit multiply can be used instead of a
+ * full 32x32 multiply.  This provides a useful speedup on many machines.
+ * Unfortunately there is no way to specify a 16x16->32 multiply portably
+ * in C, but some C compilers will do the right thing if you provide the
+ * correct combination of casts.
+ */
+
+#ifdef SHORTxSHORT_32		/* may work if 'int' is 32 bits */
+#define MULTIPLY16C16(var,const)  (((INT16) (var)) * ((INT16) (const)))
+#endif
+#ifdef SHORTxLCONST_32		/* known to work with Microsoft C 6.0 */
+#define MULTIPLY16C16(var,const)  (((INT16) (var)) * ((INT32) (const)))
+#endif
+
+#ifndef MULTIPLY16C16		/* default definition */
+#define MULTIPLY16C16(var,const)  ((var) * (const))
+#endif
+
+/* Same except both inputs are variables. */
+
+#ifdef SHORTxSHORT_32		/* may work if 'int' is 32 bits */
+#define MULTIPLY16V16(var1,var2)  (((INT16) (var1)) * ((INT16) (var2)))
+#endif
+
+#ifndef MULTIPLY16V16		/* default definition */
+#define MULTIPLY16V16(var1,var2)  ((var1) * (var2))
+#endif
diff --git a/jddctmgr.c b/jddctmgr.c
new file mode 100644
index 0000000..044e469
--- /dev/null
+++ b/jddctmgr.c
@@ -0,0 +1,288 @@
+/*
+ * jddctmgr.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the inverse-DCT management logic.
+ * This code selects a particular IDCT implementation to be used,
+ * and it performs related housekeeping chores.  No code in this file
+ * is executed per IDCT step, only during output pass setup.
+ *
+ * Note that the IDCT routines are responsible for performing coefficient
+ * dequantization as well as the IDCT proper.  This module sets up the
+ * dequantization multiplier table needed by the IDCT routine.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+#include "jsimddct.h"
+#include "jpegcomp.h"
+
+
+/*
+ * The decompressor input side (jdinput.c) saves away the appropriate
+ * quantization table for each component at the start of the first scan
+ * involving that component.  (This is necessary in order to correctly
+ * decode files that reuse Q-table slots.)
+ * When we are ready to make an output pass, the saved Q-table is converted
+ * to a multiplier table that will actually be used by the IDCT routine.
+ * The multiplier table contents are IDCT-method-dependent.  To support
+ * application changes in IDCT method between scans, we can remake the
+ * multiplier tables if necessary.
+ * In buffered-image mode, the first output pass may occur before any data
+ * has been seen for some components, and thus before their Q-tables have
+ * been saved away.  To handle this case, multiplier tables are preset
+ * to zeroes; the result of the IDCT will be a neutral gray level.
+ */
+
+
+/* Private subobject for this module */
+
+typedef struct {
+  struct jpeg_inverse_dct pub;	/* public fields */
+
+  /* This array contains the IDCT method code that each multiplier table
+   * is currently set up for, or -1 if it's not yet set up.
+   * The actual multiplier tables are pointed to by dct_table in the
+   * per-component comp_info structures.
+   */
+  int cur_method[MAX_COMPONENTS];
+} my_idct_controller;
+
+typedef my_idct_controller * my_idct_ptr;
+
+
+/* Allocated multiplier tables: big enough for any supported variant */
+
+typedef union {
+  ISLOW_MULT_TYPE islow_array[DCTSIZE2];
+#ifdef DCT_IFAST_SUPPORTED
+  IFAST_MULT_TYPE ifast_array[DCTSIZE2];
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  FLOAT_MULT_TYPE float_array[DCTSIZE2];
+#endif
+} multiplier_table;
+
+
+/* The current scaled-IDCT routines require ISLOW-style multiplier tables,
+ * so be sure to compile that code if either ISLOW or SCALING is requested.
+ */
+#ifdef DCT_ISLOW_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#else
+#ifdef IDCT_SCALING_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#endif
+#endif
+
+
+/*
+ * Prepare for an output pass.
+ * Here we select the proper IDCT routine for each component and build
+ * a matching multiplier table.
+ */
+
+METHODDEF(void)
+start_pass (j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct = (my_idct_ptr) cinfo->idct;
+  int ci, i;
+  jpeg_component_info *compptr;
+  int method = 0;
+  inverse_DCT_method_ptr method_ptr = NULL;
+  JQUANT_TBL * qtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Select the proper IDCT routine for this component's scaling */
+    switch (compptr->_DCT_scaled_size) {
+#ifdef IDCT_SCALING_SUPPORTED
+    case 1:
+      method_ptr = jpeg_idct_1x1;
+      method = JDCT_ISLOW;	/* jidctred uses islow-style table */
+      break;
+    case 2:
+      if (jsimd_can_idct_2x2())
+        method_ptr = jsimd_idct_2x2;
+      else
+        method_ptr = jpeg_idct_2x2;
+      method = JDCT_ISLOW;	/* jidctred uses islow-style table */
+      break;
+    case 4:
+      if (jsimd_can_idct_4x4())
+        method_ptr = jsimd_idct_4x4;
+      else
+        method_ptr = jpeg_idct_4x4;
+      method = JDCT_ISLOW;	/* jidctred uses islow-style table */
+      break;
+#endif
+    case DCTSIZE:
+      switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+      case JDCT_ISLOW:
+	if (jsimd_can_idct_islow())
+	  method_ptr = jsimd_idct_islow;
+	else
+	  method_ptr = jpeg_idct_islow;
+	method = JDCT_ISLOW;
+	break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+      case JDCT_IFAST:
+	if (jsimd_can_idct_ifast())
+	  method_ptr = jsimd_idct_ifast;
+	else
+	  method_ptr = jpeg_idct_ifast;
+	method = JDCT_IFAST;
+	break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+      case JDCT_FLOAT:
+	if (jsimd_can_idct_float())
+	  method_ptr = jsimd_idct_float;
+	else
+	  method_ptr = jpeg_idct_float;
+	method = JDCT_FLOAT;
+	break;
+#endif
+      default:
+	ERREXIT(cinfo, JERR_NOT_COMPILED);
+	break;
+      }
+      break;
+    default:
+      ERREXIT1(cinfo, JERR_BAD_DCTSIZE, compptr->_DCT_scaled_size);
+      break;
+    }
+    idct->pub.inverse_DCT[ci] = method_ptr;
+    /* Create multiplier table from quant table.
+     * However, we can skip this if the component is uninteresting
+     * or if we already built the table.  Also, if no quant table
+     * has yet been saved for the component, we leave the
+     * multiplier table all-zero; we'll be reading zeroes from the
+     * coefficient controller's buffer anyway.
+     */
+    if (! compptr->component_needed || idct->cur_method[ci] == method)
+      continue;
+    qtbl = compptr->quant_table;
+    if (qtbl == NULL)		/* happens if no data yet for component */
+      continue;
+    idct->cur_method[ci] = method;
+    switch (method) {
+#ifdef PROVIDE_ISLOW_TABLES
+    case JDCT_ISLOW:
+      {
+	/* For LL&M IDCT method, multipliers are equal to raw quantization
+	 * coefficients, but are stored as ints to ensure access efficiency.
+	 */
+	ISLOW_MULT_TYPE * ismtbl = (ISLOW_MULT_TYPE *) compptr->dct_table;
+	for (i = 0; i < DCTSIZE2; i++) {
+	  ismtbl[i] = (ISLOW_MULT_TYPE) qtbl->quantval[i];
+	}
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+	/* For AA&N IDCT method, multipliers are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * For integer operation, the multiplier table is to be scaled by
+	 * IFAST_SCALE_BITS.
+	 */
+	IFAST_MULT_TYPE * ifmtbl = (IFAST_MULT_TYPE *) compptr->dct_table;
+#define CONST_BITS 14
+	static const INT16 aanscales[DCTSIZE2] = {
+	  /* precomputed values scaled up by 14 bits */
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+	};
+	SHIFT_TEMPS
+
+	for (i = 0; i < DCTSIZE2; i++) {
+	  ifmtbl[i] = (IFAST_MULT_TYPE)
+	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
+				  (INT32) aanscales[i]),
+		    CONST_BITS-IFAST_SCALE_BITS);
+	}
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+	/* For float AA&N IDCT method, multipliers are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 */
+	FLOAT_MULT_TYPE * fmtbl = (FLOAT_MULT_TYPE *) compptr->dct_table;
+	int row, col;
+	static const double aanscalefactor[DCTSIZE] = {
+	  1.0, 1.387039845, 1.306562965, 1.175875602,
+	  1.0, 0.785694958, 0.541196100, 0.275899379
+	};
+
+	i = 0;
+	for (row = 0; row < DCTSIZE; row++) {
+	  for (col = 0; col < DCTSIZE; col++) {
+	    fmtbl[i] = (FLOAT_MULT_TYPE)
+	      ((double) qtbl->quantval[i] *
+	       aanscalefactor[row] * aanscalefactor[col]);
+	    i++;
+	  }
+	}
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Initialize IDCT manager.
+ */
+
+GLOBAL(void)
+jinit_inverse_dct (j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct;
+  int ci;
+  jpeg_component_info *compptr;
+
+  idct = (my_idct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_idct_controller));
+  cinfo->idct = (struct jpeg_inverse_dct *) idct;
+  idct->pub.start_pass = start_pass;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate and pre-zero a multiplier table for each component */
+    compptr->dct_table =
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(multiplier_table));
+    MEMZERO(compptr->dct_table, SIZEOF(multiplier_table));
+    /* Mark multiplier table not yet set up for any method */
+    idct->cur_method[ci] = -1;
+  }
+}
diff --git a/jdhuff.c b/jdhuff.c
new file mode 100644
index 0000000..ab8e254
--- /dev/null
+++ b/jdhuff.c
@@ -0,0 +1,826 @@
+/*
+ * jdhuff.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy decoding routines.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ */
+
+/* Performance enhancements:
+ * Copyright (C)2007 Sun Microsystems, Inc.
+ * Copyright (C)2009-2010 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdhuff.h"		/* Declarations shared with jdphuff.c */
+#include "jpegcomp.h"
+
+
+/*
+ * Expanded entropy decoder object for Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;	/* Bit buffer at start of MCU */
+  savable_state saved;		/* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
+  d_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
+
+  /* Precalculated info set up by start_pass for use in decode_mcu: */
+
+  /* Pointers to derived tables to be used for each block within an MCU */
+  d_derived_tbl * dc_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  d_derived_tbl * ac_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  /* Whether we care about the DC and AC coefficient values for each block */
+  boolean dc_needed[D_MAX_BLOCKS_IN_MCU];
+  boolean ac_needed[D_MAX_BLOCKS_IN_MCU];
+} huff_entropy_decoder;
+
+typedef huff_entropy_decoder * huff_entropy_ptr;
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_huff_decoder (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, blkn, dctbl, actbl;
+  jpeg_component_info * compptr;
+
+  /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+   * This ought to be an error condition, but we make it a warning because
+   * there are some baseline files out there with all zeroes in these bytes.
+   */
+  if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2-1 ||
+      cinfo->Ah != 0 || cinfo->Al != 0)
+    WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    /* Compute derived values for Huffman tables */
+    /* We may do this more than once for a table, but it's not expensive */
+    jpeg_make_d_derived_tbl(cinfo, TRUE, dctbl,
+			    & entropy->dc_derived_tbls[dctbl]);
+    jpeg_make_d_derived_tbl(cinfo, FALSE, actbl,
+			    & entropy->ac_derived_tbls[actbl]);
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Precalculate decoding info for each block in an MCU of this scan */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    /* Precalculate which table to use for each block */
+    entropy->dc_cur_tbls[blkn] = entropy->dc_derived_tbls[compptr->dc_tbl_no];
+    entropy->ac_cur_tbls[blkn] = entropy->ac_derived_tbls[compptr->ac_tbl_no];
+    /* Decide whether we really care about the coefficient values */
+    if (compptr->component_needed) {
+      entropy->dc_needed[blkn] = TRUE;
+      /* we don't need the ACs if producing a 1/8th-size image */
+      entropy->ac_needed[blkn] = (compptr->_DCT_scaled_size > 1);
+    } else {
+      entropy->dc_needed[blkn] = entropy->ac_needed[blkn] = FALSE;
+    }
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ *
+ * Note this is also used by jdphuff.c.
+ */
+
+GLOBAL(void)
+jpeg_make_d_derived_tbl (j_decompress_ptr cinfo, boolean isDC, int tblno,
+			 d_derived_tbl ** pdtbl)
+{
+  JHUFF_TBL *htbl;
+  d_derived_tbl *dtbl;
+  int p, i, l, si, numsymbols;
+  int lookbits, ctr;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (d_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(d_derived_tbl));
+  dtbl = *pdtbl;
+  dtbl->pub = htbl;		/* fill in back link */
+  
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int) htbl->bits[l];
+    if (i < 0 || p + i > 256)	/* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char) l;
+  }
+  huffsize[p] = 0;
+  numsymbols = p;
+  
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+  
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int) huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((INT32) code) >= (((INT32) 1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+
+  /* Figure F.15: generate decoding tables for bit-sequential decoding */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    if (htbl->bits[l]) {
+      /* valoffset[l] = huffval[] index of 1st symbol of code length l,
+       * minus the minimum code of length l
+       */
+      dtbl->valoffset[l] = (INT32) p - (INT32) huffcode[p];
+      p += htbl->bits[l];
+      dtbl->maxcode[l] = huffcode[p-1]; /* maximum code of length l */
+    } else {
+      dtbl->maxcode[l] = -1;	/* -1 if no codes of this length */
+    }
+  }
+  dtbl->valoffset[17] = 0;
+  dtbl->maxcode[17] = 0xFFFFFL; /* ensures jpeg_huff_decode terminates */
+
+  /* Compute lookahead tables to speed up decoding.
+   * First we set all the table entries to 0, indicating "too long";
+   * then we iterate through the Huffman codes that are short enough and
+   * fill in all the entries that correspond to bit sequences starting
+   * with that code.
+   */
+
+   for (i = 0; i < (1 << HUFF_LOOKAHEAD); i++)
+     dtbl->lookup[i] = (HUFF_LOOKAHEAD + 1) << HUFF_LOOKAHEAD;
+
+  p = 0;
+  for (l = 1; l <= HUFF_LOOKAHEAD; l++) {
+    for (i = 1; i <= (int) htbl->bits[l]; i++, p++) {
+      /* l = current code's length, p = its index in huffcode[] & huffval[]. */
+      /* Generate left-justified code followed by all possible bit sequences */
+      lookbits = huffcode[p] << (HUFF_LOOKAHEAD-l);
+      for (ctr = 1 << (HUFF_LOOKAHEAD-l); ctr > 0; ctr--) {
+	dtbl->lookup[lookbits] = (l << HUFF_LOOKAHEAD) | htbl->huffval[p];
+	lookbits++;
+      }
+    }
+  }
+
+  /* Validate symbols as being reasonable.
+   * For AC tables, we make no check, but accept all byte values 0..255.
+   * For DC tables, we require the symbols to be in range 0..15.
+   * (Tighter bounds could be applied depending on the data depth and mode,
+   * but this is sufficient to ensure safe decoding.)
+   */
+  if (isDC) {
+    for (i = 0; i < numsymbols; i++) {
+      int sym = htbl->huffval[i];
+      if (sym < 0 || sym > 15)
+	ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    }
+  }
+}
+
+
+/*
+ * Out-of-line code for bit fetching (shared with jdphuff.c).
+ * See jdhuff.h for info about usage.
+ * Note: current values of get_buffer and bits_left are passed as parameters,
+ * but are returned in the corresponding fields of the state struct.
+ *
+ * On most machines MIN_GET_BITS should be 25 to allow the full 32-bit width
+ * of get_buffer to be used.  (On machines with wider words, an even larger
+ * buffer could be used.)  However, on some machines 32-bit shifts are
+ * quite slow and take time proportional to the number of places shifted.
+ * (This is true with most PC compilers, for instance.)  In this case it may
+ * be a win to set MIN_GET_BITS to the minimum value of 15.  This reduces the
+ * average shift distance at the cost of more calls to jpeg_fill_bit_buffer.
+ */
+
+#ifdef SLOW_SHIFT_32
+#define MIN_GET_BITS  15	/* minimum allowable value */
+#else
+#define MIN_GET_BITS  (BIT_BUF_SIZE-7)
+#endif
+
+
+GLOBAL(boolean)
+jpeg_fill_bit_buffer (bitread_working_state * state,
+		      register bit_buf_type get_buffer, register int bits_left,
+		      int nbits)
+/* Load up the bit buffer to a depth of at least nbits */
+{
+  /* Copy heavily used state fields into locals (hopefully registers) */
+  register const JOCTET * next_input_byte = state->next_input_byte;
+  register size_t bytes_in_buffer = state->bytes_in_buffer;
+  j_decompress_ptr cinfo = state->cinfo;
+
+  /* Attempt to load at least MIN_GET_BITS bits into get_buffer. */
+  /* (It is assumed that no request will be for more than that many bits.) */
+  /* We fail to do so only if we hit a marker or are forced to suspend. */
+
+  if (cinfo->unread_marker == 0) {	/* cannot advance past a marker */
+    while (bits_left < MIN_GET_BITS) {
+      register int c;
+
+      /* Attempt to read a byte */
+      if (bytes_in_buffer == 0) {
+	if (! (*cinfo->src->fill_input_buffer) (cinfo))
+	  return FALSE;
+	next_input_byte = cinfo->src->next_input_byte;
+	bytes_in_buffer = cinfo->src->bytes_in_buffer;
+      }
+      bytes_in_buffer--;
+      c = GETJOCTET(*next_input_byte++);
+
+      /* If it's 0xFF, check and discard stuffed zero byte */
+      if (c == 0xFF) {
+	/* Loop here to discard any padding FF's on terminating marker,
+	 * so that we can save a valid unread_marker value.  NOTE: we will
+	 * accept multiple FF's followed by a 0 as meaning a single FF data
+	 * byte.  This data pattern is not valid according to the standard.
+	 */
+	do {
+	  if (bytes_in_buffer == 0) {
+	    if (! (*cinfo->src->fill_input_buffer) (cinfo))
+	      return FALSE;
+	    next_input_byte = cinfo->src->next_input_byte;
+	    bytes_in_buffer = cinfo->src->bytes_in_buffer;
+	  }
+	  bytes_in_buffer--;
+	  c = GETJOCTET(*next_input_byte++);
+	} while (c == 0xFF);
+
+	if (c == 0) {
+	  /* Found FF/00, which represents an FF data byte */
+	  c = 0xFF;
+	} else {
+	  /* Oops, it's actually a marker indicating end of compressed data.
+	   * Save the marker code for later use.
+	   * Fine point: it might appear that we should save the marker into
+	   * bitread working state, not straight into permanent state.  But
+	   * once we have hit a marker, we cannot need to suspend within the
+	   * current MCU, because we will read no more bytes from the data
+	   * source.  So it is OK to update permanent state right away.
+	   */
+	  cinfo->unread_marker = c;
+	  /* See if we need to insert some fake zero bits. */
+	  goto no_more_bytes;
+	}
+      }
+
+      /* OK, load c into get_buffer */
+      get_buffer = (get_buffer << 8) | c;
+      bits_left += 8;
+    } /* end while */
+  } else {
+  no_more_bytes:
+    /* We get here if we've read the marker that terminates the compressed
+     * data segment.  There should be enough bits in the buffer register
+     * to satisfy the request; if so, no problem.
+     */
+    if (nbits > bits_left) {
+      /* Uh-oh.  Report corrupted data to user and stuff zeroes into
+       * the data stream, so that we can produce some kind of image.
+       * We use a nonvolatile flag to ensure that only one warning message
+       * appears per data segment.
+       */
+      if (! cinfo->entropy->insufficient_data) {
+	WARNMS(cinfo, JWRN_HIT_MARKER);
+	cinfo->entropy->insufficient_data = TRUE;
+      }
+      /* Fill the buffer with zero bits */
+      get_buffer <<= MIN_GET_BITS - bits_left;
+      bits_left = MIN_GET_BITS;
+    }
+  }
+
+  /* Unload the local registers */
+  state->next_input_byte = next_input_byte;
+  state->bytes_in_buffer = bytes_in_buffer;
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  return TRUE;
+}
+
+
+/*
+ * Out-of-line code for Huffman code decoding.
+ * See jdhuff.h for info about usage.
+ */
+
+GLOBAL(int)
+jpeg_huff_decode (bitread_working_state * state,
+		  register bit_buf_type get_buffer, register int bits_left,
+		  d_derived_tbl * htbl, int min_bits)
+{
+  register int l = min_bits;
+  register INT32 code;
+
+  /* HUFF_DECODE has determined that the code is at least min_bits */
+  /* bits long, so fetch that many bits in one swoop. */
+
+  CHECK_BIT_BUFFER(*state, l, return -1);
+  code = GET_BITS(l);
+
+  /* Collect the rest of the Huffman code one bit at a time. */
+  /* This is per Figure F.16 in the JPEG spec. */
+
+  while (code > htbl->maxcode[l]) {
+    code <<= 1;
+    CHECK_BIT_BUFFER(*state, 1, return -1);
+    code |= GET_BITS(1);
+    l++;
+  }
+
+  /* Unload the local registers */
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  /* With garbage input we may reach the sentinel value l = 17. */
+
+  if (l > 16) {
+    WARNMS(state->cinfo, JWRN_HUFF_BAD_CODE);
+    return 0;			/* fake a zero as the safest result */
+  }
+
+  return htbl->pub->huffval[ (int) (code + htbl->valoffset[l]) ];
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and add will be faster than a table lookup.
+ */
+
+#define AVOID_TABLES
+#ifdef AVOID_TABLES
+
+#define HUFF_EXTEND(x,s)  ((x) + ((((x) - (1<<((s)-1))) >> 31) & (((-1)<<(s)) + 1)))
+
+#else
+
+#define HUFF_EXTEND(x,s)  ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
+
+static const int extend_test[16] =   /* entry n is 2**(n-1) */
+  { 0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
+    0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000 };
+
+static const int extend_offset[16] = /* entry n is (-1 << n) + 1 */
+  { 0, ((-1)<<1) + 1, ((-1)<<2) + 1, ((-1)<<3) + 1, ((-1)<<4) + 1,
+    ((-1)<<5) + 1, ((-1)<<6) + 1, ((-1)<<7) + 1, ((-1)<<8) + 1,
+    ((-1)<<9) + 1, ((-1)<<10) + 1, ((-1)<<11) + 1, ((-1)<<12) + 1,
+    ((-1)<<13) + 1, ((-1)<<14) + 1, ((-1)<<15) + 1 };
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(boolean)
+process_restart (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->pub.insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+decode_mcu_slow (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  BITREAD_STATE_VARS;
+  int blkn;
+  savable_state state;
+  /* Outer loop handles each block in the MCU */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+  ASSIGN_STATE(state, entropy->saved);
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    JBLOCKROW block = MCU_data[blkn];
+    d_derived_tbl * dctbl = entropy->dc_cur_tbls[blkn];
+    d_derived_tbl * actbl = entropy->ac_cur_tbls[blkn];
+    register int s, k, r;
+
+    /* Decode a single block's worth of coefficients */
+
+    /* Section F.2.2.1: decode the DC coefficient difference */
+    HUFF_DECODE(s, br_state, dctbl, return FALSE, label1);
+    if (s) {
+      CHECK_BIT_BUFFER(br_state, s, return FALSE);
+      r = GET_BITS(s);
+      s = HUFF_EXTEND(r, s);
+    }
+
+    if (entropy->dc_needed[blkn]) {
+      /* Convert DC difference to actual value, update last_dc_val */
+      int ci = cinfo->MCU_membership[blkn];
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      /* Output the DC coefficient (assumes jpeg_natural_order[0] = 0) */
+      (*block)[0] = (JCOEF) s;
+    }
+
+    if (entropy->ac_needed[blkn]) {
+
+      /* Section F.2.2.2: decode the AC coefficients */
+      /* Since zeroes are skipped, output area must be cleared beforehand */
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE(s, br_state, actbl, return FALSE, label2);
+
+        r = s >> 4;
+        s &= 15;
+      
+        if (s) {
+          k += r;
+          CHECK_BIT_BUFFER(br_state, s, return FALSE);
+          r = GET_BITS(s);
+          s = HUFF_EXTEND(r, s);
+          /* Output coefficient in natural (dezigzagged) order.
+           * Note: the extra entries in jpeg_natural_order[] will save us
+           * if k >= DCTSIZE2, which could happen if the data is corrupted.
+           */
+          (*block)[jpeg_natural_order[k]] = (JCOEF) s;
+        } else {
+          if (r != 15)
+            break;
+          k += 15;
+        }
+      }
+
+    } else {
+
+      /* Section F.2.2.2: decode the AC coefficients */
+      /* In this path we just discard the values */
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE(s, br_state, actbl, return FALSE, label3);
+
+        r = s >> 4;
+        s &= 15;
+
+        if (s) {
+          k += r;
+          CHECK_BIT_BUFFER(br_state, s, return FALSE);
+          DROP_BITS(s);
+        } else {
+          if (r != 15)
+            break;
+          k += 15;
+        }
+      }
+    }
+  }
+
+  /* Completed MCU, so update state */
+  BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+  ASSIGN_STATE(entropy->saved, state);
+  return TRUE;
+}
+
+
+/***************************************************************/
+
+#define ADD_BYTE  {                                     \
+  int val0 = *(buffer++);                               \
+  int val1 = *(buffer);                                 \
+                                                        \
+  bits_left += 8;                                       \
+  get_buffer = (get_buffer << 8) | (val0);              \
+  if (val0 == 0xFF) {                                   \
+    buffer++;                                           \
+    if (val1 != 0) {                                    \
+      buffer   -= 2;                                    \
+      get_buffer      &= ~0xFF;                         \
+    }                                                   \
+  }                                                     \
+}
+
+/***************************************************************/
+
+#if __WORDSIZE == 64 || defined(_WIN64)
+
+#define ENSURE_SHORT \
+  if (bits_left < 16) { \
+    ADD_BYTE ADD_BYTE ADD_BYTE ADD_BYTE ADD_BYTE ADD_BYTE \
+  }
+
+#else
+
+#define ENSURE_SHORT  if (bits_left < 16) { ADD_BYTE ADD_BYTE }
+
+#endif
+
+/***************************************************************/
+
+#define HUFF_DECODE_FAST(symbol, size, htbl) { \
+  ENSURE_SHORT \
+  symbol = PEEK_BITS(HUFF_LOOKAHEAD); \
+  symbol = htbl->lookup[symbol]; \
+  size = symbol >> 8; \
+  bits_left -= size; \
+  symbol = symbol & ((1 << HUFF_LOOKAHEAD) - 1); \
+  if (size == HUFF_LOOKAHEAD + 1) { \
+    symbol = (get_buffer >> bits_left) & ((1 << (size)) - 1); \
+    while (symbol > htbl->maxcode[size]) { \
+      symbol <<= 1; \
+      symbol |= GET_BITS(1); \
+      size++; \
+    } \
+    symbol = htbl->pub->huffval[ (int) (symbol + htbl->valoffset[size]) ]; \
+  } \
+}
+
+/***************************************************************/
+
+LOCAL(boolean)
+decode_mcu_fast (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  BITREAD_STATE_VARS;
+  JOCTET *buffer;
+  int blkn;
+  savable_state state;
+  /* Outer loop handles each block in the MCU */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+  buffer = (JOCTET *) br_state.next_input_byte;
+  ASSIGN_STATE(state, entropy->saved);
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    JBLOCKROW block = MCU_data[blkn];
+    d_derived_tbl * dctbl = entropy->dc_cur_tbls[blkn];
+    d_derived_tbl * actbl = entropy->ac_cur_tbls[blkn];
+    register int s, k, r, l;
+
+    HUFF_DECODE_FAST(s, l, dctbl);
+    if (s) {
+      ENSURE_SHORT
+      r = GET_BITS(s);
+      s = HUFF_EXTEND(r, s);
+    }
+
+    if (entropy->dc_needed[blkn]) {
+      int ci = cinfo->MCU_membership[blkn];
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      (*block)[0] = (JCOEF) s;
+    }
+
+    if (entropy->ac_needed[blkn]) {
+
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE_FAST(s, l, actbl);
+        r = s >> 4;
+        s &= 15;
+      
+        if (s) {
+          k += r;
+          ENSURE_SHORT
+          r = GET_BITS(s);
+          s = HUFF_EXTEND(r, s);
+          (*block)[jpeg_natural_order[k]] = (JCOEF) s;
+        } else {
+          if (r != 15) break;
+          k += 15;
+        }
+      }
+
+    } else {
+
+      for (k = 1; k < DCTSIZE2; k++) {
+        HUFF_DECODE_FAST(s, l, actbl);
+        r = s >> 4;
+        s &= 15;
+
+        if (s) {
+          k += r;
+          ENSURE_SHORT
+          DROP_BITS(s);
+        } else {
+          if (r != 15) break;
+          k += 15;
+        }
+      }
+    }
+  }
+
+  br_state.bytes_in_buffer -= (buffer - br_state.next_input_byte);
+  br_state.next_input_byte = buffer;
+  BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+  ASSIGN_STATE(entropy->saved, state);
+  return TRUE;
+}
+
+
+/*
+ * Decode and return one MCU's worth of Huffman-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA HAS BEEN ZEROED BY THE CALLER.
+ * (Wholesale zeroing is usually a little faster than retail...)
+ *
+ * Returns FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * this module, since we'll just re-assign them on the next call.)
+ */
+
+#define BUFSIZE (DCTSIZE2 * 2)
+
+METHODDEF(boolean)
+decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int usefast = 1;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+    usefast = 0;
+  }
+
+  if (cinfo->src->bytes_in_buffer < BUFSIZE * cinfo->blocks_in_MCU)
+    usefast = 0;
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    if (usefast) {
+      if (!decode_mcu_fast(cinfo, MCU_data)) return FALSE;
+    }
+    else {
+      if (!decode_mcu_slow(cinfo, MCU_data)) return FALSE;
+    }
+
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * Module initialization routine for Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_huff_decoder (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(huff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass_huff_decoder;
+  entropy->pub.decode_mcu = decode_mcu;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+  }
+}
diff --git a/jdhuff.h b/jdhuff.h
new file mode 100644
index 0000000..2697488
--- /dev/null
+++ b/jdhuff.h
@@ -0,0 +1,214 @@
+/*
+ * jdhuff.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains declarations for Huffman entropy decoding routines
+ * that are shared between the sequential decoder (jdhuff.c) and the
+ * progressive decoder (jdphuff.c).  No other modules need to see these.
+ */
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_make_d_derived_tbl	jMkDDerived
+#define jpeg_fill_bit_buffer	jFilBitBuf
+#define jpeg_huff_decode	jHufDecode
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/* Derived data constructed for each Huffman table */
+
+#define HUFF_LOOKAHEAD	8	/* # of bits of lookahead */
+
+typedef struct {
+  /* Basic tables: (element [0] of each array is unused) */
+  INT32 maxcode[18];		/* largest code of length k (-1 if none) */
+  /* (maxcode[17] is a sentinel to ensure jpeg_huff_decode terminates) */
+  INT32 valoffset[18];		/* huffval[] offset for codes of length k */
+  /* valoffset[k] = huffval[] index of 1st symbol of code length k, less
+   * the smallest code of length k; so given a code of length k, the
+   * corresponding symbol is huffval[code + valoffset[k]]
+   */
+
+  /* Link to public Huffman table (needed only in jpeg_huff_decode) */
+  JHUFF_TBL *pub;
+
+  /* Lookahead table: indexed by the next HUFF_LOOKAHEAD bits of
+   * the input data stream.  If the next Huffman code is no more
+   * than HUFF_LOOKAHEAD bits long, we can obtain its length and
+   * the corresponding symbol directly from this tables.
+   *
+   * The lower 8 bits of each table entry contain the number of
+   * bits in the corresponding Huffman code, or HUFF_LOOKAHEAD + 1
+   * if too long.  The next 8 bits of each entry contain the
+   * symbol.
+   */
+  int lookup[1<<HUFF_LOOKAHEAD];
+} d_derived_tbl;
+
+/* Expand a Huffman table definition into the derived format */
+EXTERN(void) jpeg_make_d_derived_tbl
+	JPP((j_decompress_ptr cinfo, boolean isDC, int tblno,
+	     d_derived_tbl ** pdtbl));
+
+
+/*
+ * Fetching the next N bits from the input stream is a time-critical operation
+ * for the Huffman decoders.  We implement it with a combination of inline
+ * macros and out-of-line subroutines.  Note that N (the number of bits
+ * demanded at one time) never exceeds 15 for JPEG use.
+ *
+ * We read source bytes into get_buffer and dole out bits as needed.
+ * If get_buffer already contains enough bits, they are fetched in-line
+ * by the macros CHECK_BIT_BUFFER and GET_BITS.  When there aren't enough
+ * bits, jpeg_fill_bit_buffer is called; it will attempt to fill get_buffer
+ * as full as possible (not just to the number of bits needed; this
+ * prefetching reduces the overhead cost of calling jpeg_fill_bit_buffer).
+ * Note that jpeg_fill_bit_buffer may return FALSE to indicate suspension.
+ * On TRUE return, jpeg_fill_bit_buffer guarantees that get_buffer contains
+ * at least the requested number of bits --- dummy zeroes are inserted if
+ * necessary.
+ */
+
+#if __WORDSIZE == 64 || defined(_WIN64)
+
+typedef size_t bit_buf_type;	/* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  64		/* size of buffer in bits */
+
+#else
+
+typedef INT32 bit_buf_type;	/* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  32		/* size of buffer in bits */
+
+#endif
+
+/* If long is > 32 bits on your machine, and shifting/masking longs is
+ * reasonably fast, making bit_buf_type be long and setting BIT_BUF_SIZE
+ * appropriately should be a win.  Unfortunately we can't define the size
+ * with something like  #define BIT_BUF_SIZE (sizeof(bit_buf_type)*8)
+ * because not all machines measure sizeof in 8-bit bytes.
+ */
+
+typedef struct {		/* Bitreading state saved across MCUs */
+  bit_buf_type get_buffer;	/* current bit-extraction buffer */
+  int bits_left;		/* # of unused bits in it */
+} bitread_perm_state;
+
+typedef struct {		/* Bitreading working state within an MCU */
+  /* Current data source location */
+  /* We need a copy, rather than munging the original, in case of suspension */
+  const JOCTET * next_input_byte; /* => next byte to read from source */
+  size_t bytes_in_buffer;	/* # of bytes remaining in source buffer */
+  /* Bit input buffer --- note these values are kept in register variables,
+   * not in this struct, inside the inner loops.
+   */
+  bit_buf_type get_buffer;	/* current bit-extraction buffer */
+  int bits_left;		/* # of unused bits in it */
+  /* Pointer needed by jpeg_fill_bit_buffer. */
+  j_decompress_ptr cinfo;	/* back link to decompress master record */
+} bitread_working_state;
+
+/* Macros to declare and load/save bitread local variables. */
+#define BITREAD_STATE_VARS  \
+	register bit_buf_type get_buffer;  \
+	register int bits_left;  \
+	bitread_working_state br_state
+
+#define BITREAD_LOAD_STATE(cinfop,permstate)  \
+	br_state.cinfo = cinfop; \
+	br_state.next_input_byte = cinfop->src->next_input_byte; \
+	br_state.bytes_in_buffer = cinfop->src->bytes_in_buffer; \
+	get_buffer = permstate.get_buffer; \
+	bits_left = permstate.bits_left;
+
+#define BITREAD_SAVE_STATE(cinfop,permstate)  \
+	cinfop->src->next_input_byte = br_state.next_input_byte; \
+	cinfop->src->bytes_in_buffer = br_state.bytes_in_buffer; \
+	permstate.get_buffer = get_buffer; \
+	permstate.bits_left = bits_left
+
+/*
+ * These macros provide the in-line portion of bit fetching.
+ * Use CHECK_BIT_BUFFER to ensure there are N bits in get_buffer
+ * before using GET_BITS, PEEK_BITS, or DROP_BITS.
+ * The variables get_buffer and bits_left are assumed to be locals,
+ * but the state struct might not be (jpeg_huff_decode needs this).
+ *	CHECK_BIT_BUFFER(state,n,action);
+ *		Ensure there are N bits in get_buffer; if suspend, take action.
+ *      val = GET_BITS(n);
+ *		Fetch next N bits.
+ *      val = PEEK_BITS(n);
+ *		Fetch next N bits without removing them from the buffer.
+ *	DROP_BITS(n);
+ *		Discard next N bits.
+ * The value N should be a simple variable, not an expression, because it
+ * is evaluated multiple times.
+ */
+
+#define CHECK_BIT_BUFFER(state,nbits,action) \
+	{ if (bits_left < (nbits)) {  \
+	    if (! jpeg_fill_bit_buffer(&(state),get_buffer,bits_left,nbits))  \
+	      { action; }  \
+	    get_buffer = (state).get_buffer; bits_left = (state).bits_left; } }
+
+#define GET_BITS(nbits) \
+	(((int) (get_buffer >> (bits_left -= (nbits)))) & ((1<<(nbits))-1))
+
+#define PEEK_BITS(nbits) \
+	(((int) (get_buffer >> (bits_left -  (nbits)))) & ((1<<(nbits))-1))
+
+#define DROP_BITS(nbits) \
+	(bits_left -= (nbits))
+
+/* Load up the bit buffer to a depth of at least nbits */
+EXTERN(boolean) jpeg_fill_bit_buffer
+	JPP((bitread_working_state * state, register bit_buf_type get_buffer,
+	     register int bits_left, int nbits));
+
+
+/*
+ * Code for extracting next Huffman-coded symbol from input bit stream.
+ * Again, this is time-critical and we make the main paths be macros.
+ *
+ * We use a lookahead table to process codes of up to HUFF_LOOKAHEAD bits
+ * without looping.  Usually, more than 95% of the Huffman codes will be 8
+ * or fewer bits long.  The few overlength codes are handled with a loop,
+ * which need not be inline code.
+ *
+ * Notes about the HUFF_DECODE macro:
+ * 1. Near the end of the data segment, we may fail to get enough bits
+ *    for a lookahead.  In that case, we do it the hard way.
+ * 2. If the lookahead table contains no entry, the next code must be
+ *    more than HUFF_LOOKAHEAD bits long.
+ * 3. jpeg_huff_decode returns -1 if forced to suspend.
+ */
+
+#define HUFF_DECODE(result,state,htbl,failaction,slowlabel) \
+{ register int nb, look; \
+  if (bits_left < HUFF_LOOKAHEAD) { \
+    if (! jpeg_fill_bit_buffer(&state,get_buffer,bits_left, 0)) {failaction;} \
+    get_buffer = state.get_buffer; bits_left = state.bits_left; \
+    if (bits_left < HUFF_LOOKAHEAD) { \
+      nb = 1; goto slowlabel; \
+    } \
+  } \
+  look = PEEK_BITS(HUFF_LOOKAHEAD); \
+  if ((nb = (htbl->lookup[look] >> HUFF_LOOKAHEAD)) <= HUFF_LOOKAHEAD) { \
+    DROP_BITS(nb); \
+    result = htbl->lookup[look] & ((1 << HUFF_LOOKAHEAD) - 1); \
+  } else { \
+slowlabel: \
+    if ((result=jpeg_huff_decode(&state,get_buffer,bits_left,htbl,nb)) < 0) \
+	{ failaction; } \
+    get_buffer = state.get_buffer; bits_left = state.bits_left; \
+  } \
+}
+
+/* Out-of-line case for Huffman code fetching */
+EXTERN(int) jpeg_huff_decode
+	JPP((bitread_working_state * state, register bit_buf_type get_buffer,
+	     register int bits_left, d_derived_tbl * htbl, int min_bits));
diff --git a/jdinput.c b/jdinput.c
new file mode 100644
index 0000000..9fcd089
--- /dev/null
+++ b/jdinput.c
@@ -0,0 +1,471 @@
+/*
+ * jdinput.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2002-2009 by Guido Vollbeding.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains input control logic for the JPEG decompressor.
+ * These routines are concerned with controlling the decompressor's input
+ * processing (marker reading and coefficient decoding).  The actual input
+ * reading is done in jdmarker.c, jdhuff.c, and jdphuff.c.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_input_controller pub; /* public fields */
+
+  boolean inheaders;		/* TRUE until first SOS is reached */
+} my_input_controller;
+
+typedef my_input_controller * my_inputctl_ptr;
+
+
+/* Forward declarations */
+METHODDEF(int) consume_markers JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Routines to calculate various quantities related to the size of the image.
+ */
+
+
+#if JPEG_LIB_VERSION >= 80
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ */
+
+GLOBAL(void)
+jpeg_core_output_dimensions (j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase.
+ * This function is used for transcoding and full decompression.
+ */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+
+  /* Compute actual output image dimensions and DCT scaling choices. */
+  if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom) {
+    /* Provide 1/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 1;
+    cinfo->min_DCT_v_scaled_size = 1;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 2) {
+    /* Provide 2/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 2L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 2L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 2;
+    cinfo->min_DCT_v_scaled_size = 2;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 4) {
+    /* Provide 4/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 4L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 4L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 4;
+    cinfo->min_DCT_v_scaled_size = 4;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 8) {
+    /* Provide 8/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 8L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 8L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 8;
+    cinfo->min_DCT_v_scaled_size = 8;
+  }
+  /* Recompute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->DCT_h_scaled_size = cinfo->min_DCT_h_scaled_size;
+    compptr->DCT_v_scaled_size = cinfo->min_DCT_v_scaled_size;
+  }
+
+#else /* !IDCT_SCALING_SUPPORTED */
+
+  /* Hardwire it to "no scaling" */
+  cinfo->output_width = cinfo->image_width;
+  cinfo->output_height = cinfo->image_height;
+  /* jdinput.c has already initialized DCT_scaled_size,
+   * and has computed unscaled downsampled_width and downsampled_height.
+   */
+
+#endif /* IDCT_SCALING_SUPPORTED */
+}
+#endif
+
+
+LOCAL(void)
+initial_setup (j_decompress_ptr cinfo)
+/* Called once, when first SOS marker is reached */
+{
+  int ci;
+  jpeg_component_info *compptr;
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long) cinfo->image_height > (long) JPEG_MAX_DIMENSION ||
+      (long) cinfo->image_width > (long) JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* For now, precision must match compiled-in value... */
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	     MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor<=0 || compptr->h_samp_factor>MAX_SAMP_FACTOR ||
+	compptr->v_samp_factor<=0 || compptr->v_samp_factor>MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+				   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+				   compptr->v_samp_factor);
+  }
+
+#if JPEG_LIB_VERSION >=80
+    cinfo->block_size = DCTSIZE;
+    cinfo->natural_order = jpeg_natural_order;
+    cinfo->lim_Se = DCTSIZE2-1;
+#endif
+
+  /* We initialize DCT_scaled_size and min_DCT_scaled_size to DCTSIZE.
+   * In the full decompressor, this will be overridden by jdmaster.c;
+   * but in the transcoder, jdmaster.c is not used, so we must do it here.
+   */
+#if JPEG_LIB_VERSION >= 70
+  cinfo->min_DCT_h_scaled_size = cinfo->min_DCT_v_scaled_size = DCTSIZE;
+#else
+  cinfo->min_DCT_scaled_size = DCTSIZE;
+#endif
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+#if JPEG_LIB_VERSION >= 70
+    compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size = DCTSIZE;
+#else
+    compptr->DCT_scaled_size = DCTSIZE;
+#endif
+    /* Size in DCT blocks */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) (cinfo->max_h_samp_factor * DCTSIZE));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) (cinfo->max_v_samp_factor * DCTSIZE));
+    /* downsampled_width and downsampled_height will also be overridden by
+     * jdmaster.c if we are doing full decompression.  The transcoder library
+     * doesn't use these values, but the calling application might.
+     */
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) cinfo->max_v_samp_factor);
+    /* Mark component needed, until color conversion says otherwise */
+    compptr->component_needed = TRUE;
+    /* Mark no quantization table yet saved for component */
+    compptr->quant_table = NULL;
+  }
+
+  /* Compute number of fully interleaved MCU rows. */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long) cinfo->image_height,
+		  (long) (cinfo->max_v_samp_factor*DCTSIZE));
+
+  /* Decide whether file contains multiple scans */
+  if (cinfo->comps_in_scan < cinfo->num_components || cinfo->progressive_mode)
+    cinfo->inputctl->has_multiple_scans = TRUE;
+  else
+    cinfo->inputctl->has_multiple_scans = FALSE;
+}
+
+
+LOCAL(void)
+per_scan_setup (j_decompress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] were set from SOS marker */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  
+  if (cinfo->comps_in_scan == 1) {
+    
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+    
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = compptr->_DCT_scaled_size;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+    
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+    
+  } else {
+    
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+	       MAX_COMPS_IN_SCAN);
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width,
+		    (long) (cinfo->max_h_samp_factor*DCTSIZE));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height,
+		    (long) (cinfo->max_v_samp_factor*DCTSIZE));
+    
+    cinfo->blocks_in_MCU = 0;
+    
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * compptr->_DCT_scaled_size;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int) (compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int) (compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > D_MAX_BLOCKS_IN_MCU)
+	ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+	cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+    
+  }
+}
+
+
+/*
+ * Save away a copy of the Q-table referenced by each component present
+ * in the current scan, unless already saved during a prior scan.
+ *
+ * In a multiple-scan JPEG file, the encoder could assign different components
+ * the same Q-table slot number, but change table definitions between scans
+ * so that each component uses a different Q-table.  (The IJG encoder is not
+ * currently capable of doing this, but other encoders might.)  Since we want
+ * to be able to dequantize all the components at the end of the file, this
+ * means that we have to save away the table actually used for each component.
+ * We do this by copying the table at the start of the first scan containing
+ * the component.
+ * The JPEG spec prohibits the encoder from changing the contents of a Q-table
+ * slot between scans of a component using that slot.  If the encoder does so
+ * anyway, this decoder will simply use the Q-table values that were current
+ * at the start of the first scan for the component.
+ *
+ * The decompressor output side looks only at the saved quant tables,
+ * not at the current Q-table slots.
+ */
+
+LOCAL(void)
+latch_quant_tables (j_decompress_ptr cinfo)
+{
+  int ci, qtblno;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtbl;
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* No work if we already saved Q-table for this component */
+    if (compptr->quant_table != NULL)
+      continue;
+    /* Make sure specified quantization table is present */
+    qtblno = compptr->quant_tbl_no;
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+	cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    /* OK, save away the quantization table */
+    qtbl = (JQUANT_TBL *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(JQUANT_TBL));
+    MEMCOPY(qtbl, cinfo->quant_tbl_ptrs[qtblno], SIZEOF(JQUANT_TBL));
+    compptr->quant_table = qtbl;
+  }
+}
+
+
+/*
+ * Initialize the input modules to read a scan of compressed data.
+ * The first call to this is done by jdmaster.c after initializing
+ * the entire decompressor (during jpeg_start_decompress).
+ * Subsequent calls come from consume_markers, below.
+ */
+
+METHODDEF(void)
+start_input_pass (j_decompress_ptr cinfo)
+{
+  per_scan_setup(cinfo);
+  latch_quant_tables(cinfo);
+  (*cinfo->entropy->start_pass) (cinfo);
+  (*cinfo->coef->start_input_pass) (cinfo);
+  cinfo->inputctl->consume_input = cinfo->coef->consume_data;
+}
+
+
+/*
+ * Finish up after inputting a compressed-data scan.
+ * This is called by the coefficient controller after it's read all
+ * the expected data of the scan.
+ */
+
+METHODDEF(void)
+finish_input_pass (j_decompress_ptr cinfo)
+{
+  cinfo->inputctl->consume_input = consume_markers;
+}
+
+
+/*
+ * Read JPEG markers before, between, or after compressed-data scans.
+ * Change state as necessary when a new scan is reached.
+ * Return value is JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ *
+ * The consume_input method pointer points either here or to the
+ * coefficient controller's consume_data routine, depending on whether
+ * we are reading a compressed data segment or inter-segment markers.
+ */
+
+METHODDEF(int)
+consume_markers (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr) cinfo->inputctl;
+  int val;
+
+  if (inputctl->pub.eoi_reached) /* After hitting EOI, read no further */
+    return JPEG_REACHED_EOI;
+
+  val = (*cinfo->marker->read_markers) (cinfo);
+
+  switch (val) {
+  case JPEG_REACHED_SOS:	/* Found SOS */
+    if (inputctl->inheaders) {	/* 1st SOS */
+      initial_setup(cinfo);
+      inputctl->inheaders = FALSE;
+      /* Note: start_input_pass must be called by jdmaster.c
+       * before any more input can be consumed.  jdapimin.c is
+       * responsible for enforcing this sequencing.
+       */
+    } else {			/* 2nd or later SOS marker */
+      if (! inputctl->pub.has_multiple_scans)
+	ERREXIT(cinfo, JERR_EOI_EXPECTED); /* Oops, I wasn't expecting this! */
+      start_input_pass(cinfo);
+    }
+    break;
+  case JPEG_REACHED_EOI:	/* Found EOI */
+    inputctl->pub.eoi_reached = TRUE;
+    if (inputctl->inheaders) {	/* Tables-only datastream, apparently */
+      if (cinfo->marker->saw_SOF)
+	ERREXIT(cinfo, JERR_SOF_NO_SOS);
+    } else {
+      /* Prevent infinite loop in coef ctlr's decompress_data routine
+       * if user set output_scan_number larger than number of scans.
+       */
+      if (cinfo->output_scan_number > cinfo->input_scan_number)
+	cinfo->output_scan_number = cinfo->input_scan_number;
+    }
+    break;
+  case JPEG_SUSPENDED:
+    break;
+  }
+
+  return val;
+}
+
+
+/*
+ * Reset state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_input_controller (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr) cinfo->inputctl;
+
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = TRUE;
+  /* Reset other modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->marker->reset_marker_reader) (cinfo);
+  /* Reset progression state -- would be cleaner if entropy decoder did this */
+  cinfo->coef_bits = NULL;
+}
+
+
+/*
+ * Initialize the input controller module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_input_controller (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl;
+
+  /* Create subobject in permanent pool */
+  inputctl = (my_inputctl_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				SIZEOF(my_input_controller));
+  cinfo->inputctl = (struct jpeg_input_controller *) inputctl;
+  /* Initialize method pointers */
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.reset_input_controller = reset_input_controller;
+  inputctl->pub.start_input_pass = start_input_pass;
+  inputctl->pub.finish_input_pass = finish_input_pass;
+  /* Initialize state: can't use reset_input_controller since we don't
+   * want to try to reset other modules yet.
+   */
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = TRUE;
+}
diff --git a/jdmainct.c b/jdmainct.c
new file mode 100644
index 0000000..67f6215
--- /dev/null
+++ b/jdmainct.c
@@ -0,0 +1,514 @@
+/*
+ * jdmainct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the main buffer controller for decompression.
+ * The main buffer lies between the JPEG decompressor proper and the
+ * post-processor; it holds downsampled data in the JPEG colorspace.
+ *
+ * Note that this code is bypassed in raw-data mode, since the application
+ * supplies the equivalent of the main buffer in that case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+
+/*
+ * In the current system design, the main buffer need never be a full-image
+ * buffer; any full-height buffers will be found inside the coefficient or
+ * postprocessing controllers.  Nonetheless, the main controller is not
+ * trivial.  Its responsibility is to provide context rows for upsampling/
+ * rescaling, and doing this in an efficient fashion is a bit tricky.
+ *
+ * Postprocessor input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  (We require DCT_scaled_size values to be
+ * chosen such that these numbers are integers.  In practice DCT_scaled_size
+ * values will likely be powers of two, so we actually have the stronger
+ * condition that DCT_scaled_size / min_DCT_scaled_size is an integer.)
+ * Upsampling will typically produce max_v_samp_factor pixel rows from each
+ * row group (times any additional scale factor that the upsampler is
+ * applying).
+ *
+ * The coefficient controller will deliver data to us one iMCU row at a time;
+ * each iMCU row contains v_samp_factor * DCT_scaled_size sample rows, or
+ * exactly min_DCT_scaled_size row groups.  (This amount of data corresponds
+ * to one row of MCUs when the image is fully interleaved.)  Note that the
+ * number of sample rows varies across components, but the number of row
+ * groups does not.  Some garbage sample rows may be included in the last iMCU
+ * row at the bottom of the image.
+ *
+ * Depending on the vertical scaling algorithm used, the upsampler may need
+ * access to the sample row(s) above and below its current input row group.
+ * The upsampler is required to set need_context_rows TRUE at global selection
+ * time if so.  When need_context_rows is FALSE, this controller can simply
+ * obtain one iMCU row at a time from the coefficient controller and dole it
+ * out as row groups to the postprocessor.
+ *
+ * When need_context_rows is TRUE, this controller guarantees that the buffer
+ * passed to postprocessing contains at least one row group's worth of samples
+ * above and below the row group(s) being processed.  Note that the context
+ * rows "above" the first passed row group appear at negative row offsets in
+ * the passed buffer.  At the top and bottom of the image, the required
+ * context rows are manufactured by duplicating the first or last real sample
+ * row; this avoids having special cases in the upsampling inner loops.
+ *
+ * The amount of context is fixed at one row group just because that's a
+ * convenient number for this controller to work with.  The existing
+ * upsamplers really only need one sample row of context.  An upsampler
+ * supporting arbitrary output rescaling might wish for more than one row
+ * group of context when shrinking the image; tough, we don't handle that.
+ * (This is justified by the assumption that downsizing will be handled mostly
+ * by adjusting the DCT_scaled_size values, so that the actual scale factor at
+ * the upsample step needn't be much less than one.)
+ *
+ * To provide the desired context, we have to retain the last two row groups
+ * of one iMCU row while reading in the next iMCU row.  (The last row group
+ * can't be processed until we have another row group for its below-context,
+ * and so we have to save the next-to-last group too for its above-context.)
+ * We could do this most simply by copying data around in our buffer, but
+ * that'd be very slow.  We can avoid copying any data by creating a rather
+ * strange pointer structure.  Here's how it works.  We allocate a workspace
+ * consisting of M+2 row groups (where M = min_DCT_scaled_size is the number
+ * of row groups per iMCU row).  We create two sets of redundant pointers to
+ * the workspace.  Labeling the physical row groups 0 to M+1, the synthesized
+ * pointer lists look like this:
+ *                   M+1                          M-1
+ * master pointer --> 0         master pointer --> 0
+ *                    1                            1
+ *                   ...                          ...
+ *                   M-3                          M-3
+ *                   M-2                           M
+ *                   M-1                          M+1
+ *                    M                           M-2
+ *                   M+1                          M-1
+ *                    0                            0
+ * We read alternate iMCU rows using each master pointer; thus the last two
+ * row groups of the previous iMCU row remain un-overwritten in the workspace.
+ * The pointer lists are set up so that the required context rows appear to
+ * be adjacent to the proper places when we pass the pointer lists to the
+ * upsampler.
+ *
+ * The above pictures describe the normal state of the pointer lists.
+ * At top and bottom of the image, we diddle the pointer lists to duplicate
+ * the first or last sample row as necessary (this is cheaper than copying
+ * sample rows around).
+ *
+ * This scheme breaks down if M < 2, ie, min_DCT_scaled_size is 1.  In that
+ * situation each iMCU row provides only one row group so the buffering logic
+ * must be different (eg, we must read two iMCU rows before we can emit the
+ * first row group).  For now, we simply do not support providing context
+ * rows when min_DCT_scaled_size is 1.  That combination seems unlikely to
+ * be worth providing --- if someone wants a 1/8th-size preview, they probably
+ * want it quick and dirty, so a context-free upsampler is sufficient.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_main_controller pub; /* public fields */
+
+  /* Pointer to allocated workspace (M or M+2 row groups). */
+  JSAMPARRAY buffer[MAX_COMPONENTS];
+
+  boolean buffer_full;		/* Have we gotten an iMCU row from decoder? */
+  JDIMENSION rowgroup_ctr;	/* counts row groups output to postprocessor */
+
+  /* Remaining fields are only used in the context case. */
+
+  /* These are the master pointers to the funny-order pointer lists. */
+  JSAMPIMAGE xbuffer[2];	/* pointers to weird pointer lists */
+
+  int whichptr;			/* indicates which pointer set is now in use */
+  int context_state;		/* process_data state machine status */
+  JDIMENSION rowgroups_avail;	/* row groups available to postprocessor */
+  JDIMENSION iMCU_row_ctr;	/* counts iMCU rows to detect image top/bot */
+} my_main_controller;
+
+typedef my_main_controller * my_main_ptr;
+
+/* context_state values: */
+#define CTX_PREPARE_FOR_IMCU	0	/* need to prepare for MCU row */
+#define CTX_PROCESS_IMCU	1	/* feeding iMCU to postprocessor */
+#define CTX_POSTPONED_ROW	2	/* feeding postponed row group */
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+METHODDEF(void) process_data_context_main
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) process_data_crank_post
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+#endif
+
+
+LOCAL(void)
+alloc_funny_pointers (j_decompress_ptr cinfo)
+/* Allocate space for the funny pointer lists.
+ * This is done only once, not once per pass.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, rgroup;
+  int M = cinfo->_min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf;
+
+  /* Get top-level space for component array pointers.
+   * We alloc both arrays with one call to save a few cycles.
+   */
+  main->xbuffer[0] = (JSAMPIMAGE)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				cinfo->num_components * 2 * SIZEOF(JSAMPARRAY));
+  main->xbuffer[1] = main->xbuffer[0] + cinfo->num_components;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    /* Get space for pointer lists --- M+4 row groups in each list.
+     * We alloc both pointer lists with one call to save a few cycles.
+     */
+    xbuf = (JSAMPARRAY)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  2 * (rgroup * (M + 4)) * SIZEOF(JSAMPROW));
+    xbuf += rgroup;		/* want one row group at negative offsets */
+    main->xbuffer[0][ci] = xbuf;
+    xbuf += rgroup * (M + 4);
+    main->xbuffer[1][ci] = xbuf;
+  }
+}
+
+
+LOCAL(void)
+make_funny_pointers (j_decompress_ptr cinfo)
+/* Create the funny pointer lists discussed in the comments above.
+ * The actual workspace is already allocated (in main->buffer),
+ * and the space for the pointer lists is allocated too.
+ * This routine just fills in the curiously ordered lists.
+ * This will be repeated at the beginning of each pass.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->_min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY buf, xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    xbuf0 = main->xbuffer[0][ci];
+    xbuf1 = main->xbuffer[1][ci];
+    /* First copy the workspace pointers as-is */
+    buf = main->buffer[ci];
+    for (i = 0; i < rgroup * (M + 2); i++) {
+      xbuf0[i] = xbuf1[i] = buf[i];
+    }
+    /* In the second list, put the last four row groups in swapped order */
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf1[rgroup*(M-2) + i] = buf[rgroup*M + i];
+      xbuf1[rgroup*M + i] = buf[rgroup*(M-2) + i];
+    }
+    /* The wraparound pointers at top and bottom will be filled later
+     * (see set_wraparound_pointers, below).  Initially we want the "above"
+     * pointers to duplicate the first actual data line.  This only needs
+     * to happen in xbuffer[0].
+     */
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[0];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_wraparound_pointers (j_decompress_ptr cinfo)
+/* Set up the "wraparound" pointers at top and bottom of the pointer lists.
+ * This changes the pointer list state from top-of-image to the normal state.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->_min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    xbuf0 = main->xbuffer[0][ci];
+    xbuf1 = main->xbuffer[1][ci];
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[rgroup*(M+1) + i];
+      xbuf1[i - rgroup] = xbuf1[rgroup*(M+1) + i];
+      xbuf0[rgroup*(M+2) + i] = xbuf0[i];
+      xbuf1[rgroup*(M+2) + i] = xbuf1[i];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_bottom_pointers (j_decompress_ptr cinfo)
+/* Change the pointer lists to duplicate the last sample row at the bottom
+ * of the image.  whichptr indicates which xbuffer holds the final iMCU row.
+ * Also sets rowgroups_avail to indicate number of nondummy row groups in row.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup, iMCUheight, rows_left;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Count sample rows in one iMCU row and in one row group */
+    iMCUheight = compptr->v_samp_factor * compptr->_DCT_scaled_size;
+    rgroup = iMCUheight / cinfo->_min_DCT_scaled_size;
+    /* Count nondummy sample rows remaining for this component */
+    rows_left = (int) (compptr->downsampled_height % (JDIMENSION) iMCUheight);
+    if (rows_left == 0) rows_left = iMCUheight;
+    /* Count nondummy row groups.  Should get same answer for each component,
+     * so we need only do it once.
+     */
+    if (ci == 0) {
+      main->rowgroups_avail = (JDIMENSION) ((rows_left-1) / rgroup + 1);
+    }
+    /* Duplicate the last real sample row rgroup*2 times; this pads out the
+     * last partial rowgroup and ensures at least one full rowgroup of context.
+     */
+    xbuf = main->xbuffer[main->whichptr][ci];
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf[rows_left + i] = xbuf[rows_left-1];
+    }
+  }
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main (j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->upsample->need_context_rows) {
+      main->pub.process_data = process_data_context_main;
+      make_funny_pointers(cinfo); /* Create the xbuffer[] lists */
+      main->whichptr = 0;	/* Read first iMCU row into xbuffer[0] */
+      main->context_state = CTX_PREPARE_FOR_IMCU;
+      main->iMCU_row_ctr = 0;
+    } else {
+      /* Simple case with no context needed */
+      main->pub.process_data = process_data_simple_main;
+    }
+    main->buffer_full = FALSE;	/* Mark buffer empty */
+    main->rowgroup_ctr = 0;
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_CRANK_DEST:
+    /* For last pass of 2-pass quantization, just crank the postprocessor */
+    main->pub.process_data = process_data_crank_post;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the simple case where no context is required.
+ */
+
+METHODDEF(void)
+process_data_simple_main (j_decompress_ptr cinfo,
+			  JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			  JDIMENSION out_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  JDIMENSION rowgroups_avail;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (! main->buffer_full) {
+    if (! (*cinfo->coef->decompress_data) (cinfo, main->buffer))
+      return;			/* suspension forced, can do nothing more */
+    main->buffer_full = TRUE;	/* OK, we have an iMCU row to work with */
+  }
+
+  /* There are always min_DCT_scaled_size row groups in an iMCU row. */
+  rowgroups_avail = (JDIMENSION) cinfo->_min_DCT_scaled_size;
+  /* Note: at the bottom of the image, we may pass extra garbage row groups
+   * to the postprocessor.  The postprocessor has to check for bottom
+   * of image anyway (at row resolution), so no point in us doing it too.
+   */
+
+  /* Feed the postprocessor */
+  (*cinfo->post->post_process_data) (cinfo, main->buffer,
+				     &main->rowgroup_ctr, rowgroups_avail,
+				     output_buf, out_row_ctr, out_rows_avail);
+
+  /* Has postprocessor consumed all the data yet? If so, mark buffer empty */
+  if (main->rowgroup_ctr >= rowgroups_avail) {
+    main->buffer_full = FALSE;
+    main->rowgroup_ctr = 0;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the case where context rows must be provided.
+ */
+
+METHODDEF(void)
+process_data_context_main (j_decompress_ptr cinfo,
+			   JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			   JDIMENSION out_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (! main->buffer_full) {
+    if (! (*cinfo->coef->decompress_data) (cinfo,
+					   main->xbuffer[main->whichptr]))
+      return;			/* suspension forced, can do nothing more */
+    main->buffer_full = TRUE;	/* OK, we have an iMCU row to work with */
+    main->iMCU_row_ctr++;	/* count rows received */
+  }
+
+  /* Postprocessor typically will not swallow all the input data it is handed
+   * in one call (due to filling the output buffer first).  Must be prepared
+   * to exit and restart.  This switch lets us keep track of how far we got.
+   * Note that each case falls through to the next on successful completion.
+   */
+  switch (main->context_state) {
+  case CTX_POSTPONED_ROW:
+    /* Call postprocessor using previously set pointers for postponed row */
+    (*cinfo->post->post_process_data) (cinfo, main->xbuffer[main->whichptr],
+			&main->rowgroup_ctr, main->rowgroups_avail,
+			output_buf, out_row_ctr, out_rows_avail);
+    if (main->rowgroup_ctr < main->rowgroups_avail)
+      return;			/* Need to suspend */
+    main->context_state = CTX_PREPARE_FOR_IMCU;
+    if (*out_row_ctr >= out_rows_avail)
+      return;			/* Postprocessor exactly filled output buf */
+    /*FALLTHROUGH*/
+  case CTX_PREPARE_FOR_IMCU:
+    /* Prepare to process first M-1 row groups of this iMCU row */
+    main->rowgroup_ctr = 0;
+    main->rowgroups_avail = (JDIMENSION) (cinfo->_min_DCT_scaled_size - 1);
+    /* Check for bottom of image: if so, tweak pointers to "duplicate"
+     * the last sample row, and adjust rowgroups_avail to ignore padding rows.
+     */
+    if (main->iMCU_row_ctr == cinfo->total_iMCU_rows)
+      set_bottom_pointers(cinfo);
+    main->context_state = CTX_PROCESS_IMCU;
+    /*FALLTHROUGH*/
+  case CTX_PROCESS_IMCU:
+    /* Call postprocessor using previously set pointers */
+    (*cinfo->post->post_process_data) (cinfo, main->xbuffer[main->whichptr],
+			&main->rowgroup_ctr, main->rowgroups_avail,
+			output_buf, out_row_ctr, out_rows_avail);
+    if (main->rowgroup_ctr < main->rowgroups_avail)
+      return;			/* Need to suspend */
+    /* After the first iMCU, change wraparound pointers to normal state */
+    if (main->iMCU_row_ctr == 1)
+      set_wraparound_pointers(cinfo);
+    /* Prepare to load new iMCU row using other xbuffer list */
+    main->whichptr ^= 1;	/* 0=>1 or 1=>0 */
+    main->buffer_full = FALSE;
+    /* Still need to process last row group of this iMCU row, */
+    /* which is saved at index M+1 of the other xbuffer */
+    main->rowgroup_ctr = (JDIMENSION) (cinfo->_min_DCT_scaled_size + 1);
+    main->rowgroups_avail = (JDIMENSION) (cinfo->_min_DCT_scaled_size + 2);
+    main->context_state = CTX_POSTPONED_ROW;
+  }
+}
+
+
+/*
+ * Process some data.
+ * Final pass of two-pass quantization: just call the postprocessor.
+ * Source data will be the postprocessor controller's internal buffer.
+ */
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+METHODDEF(void)
+process_data_crank_post (j_decompress_ptr cinfo,
+			 JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			 JDIMENSION out_rows_avail)
+{
+  (*cinfo->post->post_process_data) (cinfo, (JSAMPIMAGE) NULL,
+				     (JDIMENSION *) NULL, (JDIMENSION) 0,
+				     output_buf, out_row_ctr, out_rows_avail);
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+jinit_d_main_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_main_ptr main;
+  int ci, rgroup, ngroups;
+  jpeg_component_info *compptr;
+
+  main = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_main_controller));
+  cinfo->main = (struct jpeg_d_main_controller *) main;
+  main->pub.start_pass = start_pass_main;
+
+  if (need_full_buffer)		/* shouldn't happen */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Allocate the workspace.
+   * ngroups is the number of row groups we need.
+   */
+  if (cinfo->upsample->need_context_rows) {
+    if (cinfo->_min_DCT_scaled_size < 2) /* unsupported, see comments above */
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    alloc_funny_pointers(cinfo); /* Alloc space for xbuffer[] lists */
+    ngroups = cinfo->_min_DCT_scaled_size + 2;
+  } else {
+    ngroups = cinfo->_min_DCT_scaled_size;
+  }
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+      cinfo->_min_DCT_scaled_size; /* height of a row group of component */
+    main->buffer[ci] = (*cinfo->mem->alloc_sarray)
+			((j_common_ptr) cinfo, JPOOL_IMAGE,
+			 compptr->width_in_blocks * compptr->_DCT_scaled_size,
+			 (JDIMENSION) (rgroup * ngroups));
+  }
+}
diff --git a/jdmarker.c b/jdmarker.c
new file mode 100644
index 0000000..f4cca8c
--- /dev/null
+++ b/jdmarker.c
@@ -0,0 +1,1360 @@
+/*
+ * jdmarker.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to decode JPEG datastream markers.
+ * Most of the complexity arises from our desire to support input
+ * suspension: if not all of the data for a marker is available,
+ * we must exit back to the application.  On resumption, we reprocess
+ * the marker.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+typedef enum {			/* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+  
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+  
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+  
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+  
+  M_DHT   = 0xc4,
+  
+  M_DAC   = 0xcc,
+  
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+  
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+  
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+  
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+  
+  M_TEM   = 0x01,
+  
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_reader pub; /* public fields */
+
+  /* Application-overridable marker processing methods */
+  jpeg_marker_parser_method process_COM;
+  jpeg_marker_parser_method process_APPn[16];
+
+  /* Limit on marker data length to save for each marker type */
+  unsigned int length_limit_COM;
+  unsigned int length_limit_APPn[16];
+
+  /* Status of COM/APPn marker saving */
+  jpeg_saved_marker_ptr cur_marker;	/* NULL if not processing a marker */
+  unsigned int bytes_read;		/* data bytes read so far in marker */
+  /* Note: cur_marker is not linked into marker_list until it's all read. */
+} my_marker_reader;
+
+typedef my_marker_reader * my_marker_ptr;
+
+
+/*
+ * Macros for fetching data from the data source module.
+ *
+ * At all times, cinfo->src->next_input_byte and ->bytes_in_buffer reflect
+ * the current restart point; we update them only when we have reached a
+ * suitable place to restart if a suspension occurs.
+ */
+
+/* Declare and initialize local copies of input pointer/count */
+#define INPUT_VARS(cinfo)  \
+	struct jpeg_source_mgr * datasrc = (cinfo)->src;  \
+	const JOCTET * next_input_byte = datasrc->next_input_byte;  \
+	size_t bytes_in_buffer = datasrc->bytes_in_buffer
+
+/* Unload the local copies --- do this only at a restart boundary */
+#define INPUT_SYNC(cinfo)  \
+	( datasrc->next_input_byte = next_input_byte,  \
+	  datasrc->bytes_in_buffer = bytes_in_buffer )
+
+/* Reload the local copies --- used only in MAKE_BYTE_AVAIL */
+#define INPUT_RELOAD(cinfo)  \
+	( next_input_byte = datasrc->next_input_byte,  \
+	  bytes_in_buffer = datasrc->bytes_in_buffer )
+
+/* Internal macro for INPUT_BYTE and INPUT_2BYTES: make a byte available.
+ * Note we do *not* do INPUT_SYNC before calling fill_input_buffer,
+ * but we must reload the local copies after a successful fill.
+ */
+#define MAKE_BYTE_AVAIL(cinfo,action)  \
+	if (bytes_in_buffer == 0) {  \
+	  if (! (*datasrc->fill_input_buffer) (cinfo))  \
+	    { action; }  \
+	  INPUT_RELOAD(cinfo);  \
+	}
+
+/* Read a byte into variable V.
+ * If must suspend, take the specified action (typically "return FALSE").
+ */
+#define INPUT_BYTE(cinfo,V,action)  \
+	MAKESTMT( MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V = GETJOCTET(*next_input_byte++); )
+
+/* As above, but read two bytes interpreted as an unsigned 16-bit integer.
+ * V should be declared unsigned int or perhaps INT32.
+ */
+#define INPUT_2BYTES(cinfo,V,action)  \
+	MAKESTMT( MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V = ((unsigned int) GETJOCTET(*next_input_byte++)) << 8; \
+		  MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V += GETJOCTET(*next_input_byte++); )
+
+
+/*
+ * Routines to process JPEG markers.
+ *
+ * Entry condition: JPEG marker itself has been read and its code saved
+ *   in cinfo->unread_marker; input restart point is just after the marker.
+ *
+ * Exit: if return TRUE, have read and processed any parameters, and have
+ *   updated the restart point to point after the parameters.
+ *   If return FALSE, was forced to suspend before reaching end of
+ *   marker parameters; restart point has not been moved.  Same routine
+ *   will be called again after application supplies more input data.
+ *
+ * This approach to suspension assumes that all of a marker's parameters
+ * can fit into a single input bufferload.  This should hold for "normal"
+ * markers.  Some COM/APPn markers might have large parameter segments
+ * that might not fit.  If we are simply dropping such a marker, we use
+ * skip_input_data to get past it, and thereby put the problem on the
+ * source manager's shoulders.  If we are saving the marker's contents
+ * into memory, we use a slightly different convention: when forced to
+ * suspend, the marker processor updates the restart point to the end of
+ * what it's consumed (ie, the end of the buffer) before returning FALSE.
+ * On resumption, cinfo->unread_marker still contains the marker code,
+ * but the data source will point to the next chunk of marker data.
+ * The marker processor must retain internal state to deal with this.
+ *
+ * Note that we don't bother to avoid duplicate trace messages if a
+ * suspension occurs within marker parameters.  Other side effects
+ * require more care.
+ */
+
+
+LOCAL(boolean)
+get_soi (j_decompress_ptr cinfo)
+/* Process an SOI marker */
+{
+  int i;
+  
+  TRACEMS(cinfo, 1, JTRC_SOI);
+
+  if (cinfo->marker->saw_SOI)
+    ERREXIT(cinfo, JERR_SOI_DUPLICATE);
+
+  /* Reset all parameters that are defined to be reset by SOI */
+
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+  cinfo->restart_interval = 0;
+
+  /* Set initial assumptions for colorspace etc */
+
+  cinfo->jpeg_color_space = JCS_UNKNOWN;
+  cinfo->CCIR601_sampling = FALSE; /* Assume non-CCIR sampling??? */
+
+  cinfo->saw_JFIF_marker = FALSE;
+  cinfo->JFIF_major_version = 1; /* set default JFIF APP0 values */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;
+  cinfo->X_density = 1;
+  cinfo->Y_density = 1;
+  cinfo->saw_Adobe_marker = FALSE;
+  cinfo->Adobe_transform = 0;
+
+  cinfo->marker->saw_SOI = TRUE;
+
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_sof (j_decompress_ptr cinfo, boolean is_prog, boolean is_arith)
+/* Process a SOFn marker */
+{
+  INT32 length;
+  int c, ci;
+  jpeg_component_info * compptr;
+  INPUT_VARS(cinfo);
+
+  cinfo->progressive_mode = is_prog;
+  cinfo->arith_code = is_arith;
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, cinfo->data_precision, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_height, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_width, return FALSE);
+  INPUT_BYTE(cinfo, cinfo->num_components, return FALSE);
+
+  length -= 8;
+
+  TRACEMS4(cinfo, 1, JTRC_SOF, cinfo->unread_marker,
+	   (int) cinfo->image_width, (int) cinfo->image_height,
+	   cinfo->num_components);
+
+  if (cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOF_DUPLICATE);
+
+  /* We don't support files in which the image height is initially specified */
+  /* as 0 and is later redefined by DNL.  As long as we have to check that,  */
+  /* might as well have a general sanity check. */
+  if (cinfo->image_height <= 0 || cinfo->image_width <= 0
+      || cinfo->num_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  if (length != (cinfo->num_components * 3))
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  if (cinfo->comp_info == NULL)	/* do only once, even if suspend */
+    cinfo->comp_info = (jpeg_component_info *) (*cinfo->mem->alloc_small)
+			((j_common_ptr) cinfo, JPOOL_IMAGE,
+			 cinfo->num_components * SIZEOF(jpeg_component_info));
+  
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->component_index = ci;
+    INPUT_BYTE(cinfo, compptr->component_id, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    compptr->h_samp_factor = (c >> 4) & 15;
+    compptr->v_samp_factor = (c     ) & 15;
+    INPUT_BYTE(cinfo, compptr->quant_tbl_no, return FALSE);
+
+    TRACEMS4(cinfo, 1, JTRC_SOF_COMPONENT,
+	     compptr->component_id, compptr->h_samp_factor,
+	     compptr->v_samp_factor, compptr->quant_tbl_no);
+  }
+
+  cinfo->marker->saw_SOF = TRUE;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_sos (j_decompress_ptr cinfo)
+/* Process a SOS marker */
+{
+  INT32 length;
+  int i, ci, n, c, cc;
+  jpeg_component_info * compptr;
+  INPUT_VARS(cinfo);
+
+  if (! cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOS_NO_SOF);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, n, return FALSE); /* Number of components */
+
+  TRACEMS1(cinfo, 1, JTRC_SOS, n);
+
+  if (length != (n * 2 + 6) || n < 1 || n > MAX_COMPS_IN_SCAN)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  cinfo->comps_in_scan = n;
+
+  /* Collect the component-spec parameters */
+
+  for (i = 0; i < n; i++) {
+    INPUT_BYTE(cinfo, cc, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      if (cc == compptr->component_id)
+	goto id_found;
+    }
+
+    ERREXIT1(cinfo, JERR_BAD_COMPONENT_ID, cc);
+
+  id_found:
+
+    cinfo->cur_comp_info[i] = compptr;
+    compptr->dc_tbl_no = (c >> 4) & 15;
+    compptr->ac_tbl_no = (c     ) & 15;
+    
+    TRACEMS3(cinfo, 1, JTRC_SOS_COMPONENT, cc,
+	     compptr->dc_tbl_no, compptr->ac_tbl_no);
+  }
+
+  /* Collect the additional scan parameters Ss, Se, Ah/Al. */
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ss = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Se = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ah = (c >> 4) & 15;
+  cinfo->Al = (c     ) & 15;
+
+  TRACEMS4(cinfo, 1, JTRC_SOS_PARAMS, cinfo->Ss, cinfo->Se,
+	   cinfo->Ah, cinfo->Al);
+
+  /* Prepare to scan data & restart markers */
+  cinfo->marker->next_restart_num = 0;
+
+  /* Count another SOS marker */
+  cinfo->input_scan_number++;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+#ifdef D_ARITH_CODING_SUPPORTED
+
+LOCAL(boolean)
+get_dac (j_decompress_ptr cinfo)
+/* Process a DAC marker */
+{
+  INT32 length;
+  int index, val;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  while (length > 0) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+    INPUT_BYTE(cinfo, val, return FALSE);
+
+    length -= 2;
+
+    TRACEMS2(cinfo, 1, JTRC_DAC, index, val);
+
+    if (index < 0 || index >= (2*NUM_ARITH_TBLS))
+      ERREXIT1(cinfo, JERR_DAC_INDEX, index);
+
+    if (index >= NUM_ARITH_TBLS) { /* define AC table */
+      cinfo->arith_ac_K[index-NUM_ARITH_TBLS] = (UINT8) val;
+    } else {			/* define DC table */
+      cinfo->arith_dc_L[index] = (UINT8) (val & 0x0F);
+      cinfo->arith_dc_U[index] = (UINT8) (val >> 4);
+      if (cinfo->arith_dc_L[index] > cinfo->arith_dc_U[index])
+	ERREXIT1(cinfo, JERR_DAC_VALUE, val);
+    }
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+#else /* ! D_ARITH_CODING_SUPPORTED */
+
+#define get_dac(cinfo)  skip_variable(cinfo)
+
+#endif /* D_ARITH_CODING_SUPPORTED */
+
+
+LOCAL(boolean)
+get_dht (j_decompress_ptr cinfo)
+/* Process a DHT marker */
+{
+  INT32 length;
+  UINT8 bits[17];
+  UINT8 huffval[256];
+  int i, index, count;
+  JHUFF_TBL **htblptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  while (length > 16) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+
+    TRACEMS1(cinfo, 1, JTRC_DHT, index);
+      
+    bits[0] = 0;
+    count = 0;
+    for (i = 1; i <= 16; i++) {
+      INPUT_BYTE(cinfo, bits[i], return FALSE);
+      count += bits[i];
+    }
+
+    length -= 1 + 16;
+
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+	     bits[1], bits[2], bits[3], bits[4],
+	     bits[5], bits[6], bits[7], bits[8]);
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+	     bits[9], bits[10], bits[11], bits[12],
+	     bits[13], bits[14], bits[15], bits[16]);
+
+    /* Here we just do minimal validation of the counts to avoid walking
+     * off the end of our table space.  jdhuff.c will check more carefully.
+     */
+    if (count > 256 || ((INT32) count) > length)
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+    for (i = 0; i < count; i++)
+      INPUT_BYTE(cinfo, huffval[i], return FALSE);
+
+    length -= count;
+
+    if (index & 0x10) {		/* AC table definition */
+      index -= 0x10;
+      htblptr = &cinfo->ac_huff_tbl_ptrs[index];
+    } else {			/* DC table definition */
+      htblptr = &cinfo->dc_huff_tbl_ptrs[index];
+    }
+
+    if (index < 0 || index >= NUM_HUFF_TBLS)
+      ERREXIT1(cinfo, JERR_DHT_INDEX, index);
+
+    if (*htblptr == NULL)
+      *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+  
+    MEMCOPY((*htblptr)->bits, bits, SIZEOF((*htblptr)->bits));
+    MEMCOPY((*htblptr)->huffval, huffval, SIZEOF((*htblptr)->huffval));
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_dqt (j_decompress_ptr cinfo)
+/* Process a DQT marker */
+{
+  INT32 length;
+  int n, i, prec;
+  unsigned int tmp;
+  JQUANT_TBL *quant_ptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  while (length > 0) {
+    INPUT_BYTE(cinfo, n, return FALSE);
+    prec = n >> 4;
+    n &= 0x0F;
+
+    TRACEMS2(cinfo, 1, JTRC_DQT, n, prec);
+
+    if (n >= NUM_QUANT_TBLS)
+      ERREXIT1(cinfo, JERR_DQT_INDEX, n);
+      
+    if (cinfo->quant_tbl_ptrs[n] == NULL)
+      cinfo->quant_tbl_ptrs[n] = jpeg_alloc_quant_table((j_common_ptr) cinfo);
+    quant_ptr = cinfo->quant_tbl_ptrs[n];
+
+    for (i = 0; i < DCTSIZE2; i++) {
+      if (prec)
+	INPUT_2BYTES(cinfo, tmp, return FALSE);
+      else
+	INPUT_BYTE(cinfo, tmp, return FALSE);
+      /* We convert the zigzag-order table to natural array order. */
+      quant_ptr->quantval[jpeg_natural_order[i]] = (UINT16) tmp;
+    }
+
+    if (cinfo->err->trace_level >= 2) {
+      for (i = 0; i < DCTSIZE2; i += 8) {
+	TRACEMS8(cinfo, 2, JTRC_QUANTVALS,
+		 quant_ptr->quantval[i],   quant_ptr->quantval[i+1],
+		 quant_ptr->quantval[i+2], quant_ptr->quantval[i+3],
+		 quant_ptr->quantval[i+4], quant_ptr->quantval[i+5],
+		 quant_ptr->quantval[i+6], quant_ptr->quantval[i+7]);
+      }
+    }
+
+    length -= DCTSIZE2+1;
+    if (prec) length -= DCTSIZE2;
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_dri (j_decompress_ptr cinfo)
+/* Process a DRI marker */
+{
+  INT32 length;
+  unsigned int tmp;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  
+  if (length != 4)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_2BYTES(cinfo, tmp, return FALSE);
+
+  TRACEMS1(cinfo, 1, JTRC_DRI, tmp);
+
+  cinfo->restart_interval = tmp;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Routines for processing APPn and COM markers.
+ * These are either saved in memory or discarded, per application request.
+ * APP0 and APP14 are specially checked to see if they are
+ * JFIF and Adobe markers, respectively.
+ */
+
+#define APP0_DATA_LEN	14	/* Length of interesting data in APP0 */
+#define APP14_DATA_LEN	12	/* Length of interesting data in APP14 */
+#define APPN_DATA_LEN	14	/* Must be the largest of the above!! */
+
+
+LOCAL(void)
+examine_app0 (j_decompress_ptr cinfo, JOCTET FAR * data,
+	      unsigned int datalen, INT32 remaining)
+/* Examine first few bytes from an APP0.
+ * Take appropriate action if it is a JFIF marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  INT32 totallen = (INT32) datalen + remaining;
+
+  if (datalen >= APP0_DATA_LEN &&
+      GETJOCTET(data[0]) == 0x4A &&
+      GETJOCTET(data[1]) == 0x46 &&
+      GETJOCTET(data[2]) == 0x49 &&
+      GETJOCTET(data[3]) == 0x46 &&
+      GETJOCTET(data[4]) == 0) {
+    /* Found JFIF APP0 marker: save info */
+    cinfo->saw_JFIF_marker = TRUE;
+    cinfo->JFIF_major_version = GETJOCTET(data[5]);
+    cinfo->JFIF_minor_version = GETJOCTET(data[6]);
+    cinfo->density_unit = GETJOCTET(data[7]);
+    cinfo->X_density = (GETJOCTET(data[8]) << 8) + GETJOCTET(data[9]);
+    cinfo->Y_density = (GETJOCTET(data[10]) << 8) + GETJOCTET(data[11]);
+    /* Check version.
+     * Major version must be 1, anything else signals an incompatible change.
+     * (We used to treat this as an error, but now it's a nonfatal warning,
+     * because some bozo at Hijaak couldn't read the spec.)
+     * Minor version should be 0..2, but process anyway if newer.
+     */
+    if (cinfo->JFIF_major_version != 1)
+      WARNMS2(cinfo, JWRN_JFIF_MAJOR,
+	      cinfo->JFIF_major_version, cinfo->JFIF_minor_version);
+    /* Generate trace messages */
+    TRACEMS5(cinfo, 1, JTRC_JFIF,
+	     cinfo->JFIF_major_version, cinfo->JFIF_minor_version,
+	     cinfo->X_density, cinfo->Y_density, cinfo->density_unit);
+    /* Validate thumbnail dimensions and issue appropriate messages */
+    if (GETJOCTET(data[12]) | GETJOCTET(data[13]))
+      TRACEMS2(cinfo, 1, JTRC_JFIF_THUMBNAIL,
+	       GETJOCTET(data[12]), GETJOCTET(data[13]));
+    totallen -= APP0_DATA_LEN;
+    if (totallen !=
+	((INT32)GETJOCTET(data[12]) * (INT32)GETJOCTET(data[13]) * (INT32) 3))
+      TRACEMS1(cinfo, 1, JTRC_JFIF_BADTHUMBNAILSIZE, (int) totallen);
+  } else if (datalen >= 6 &&
+      GETJOCTET(data[0]) == 0x4A &&
+      GETJOCTET(data[1]) == 0x46 &&
+      GETJOCTET(data[2]) == 0x58 &&
+      GETJOCTET(data[3]) == 0x58 &&
+      GETJOCTET(data[4]) == 0) {
+    /* Found JFIF "JFXX" extension APP0 marker */
+    /* The library doesn't actually do anything with these,
+     * but we try to produce a helpful trace message.
+     */
+    switch (GETJOCTET(data[5])) {
+    case 0x10:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_JPEG, (int) totallen);
+      break;
+    case 0x11:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_PALETTE, (int) totallen);
+      break;
+    case 0x13:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_RGB, (int) totallen);
+      break;
+    default:
+      TRACEMS2(cinfo, 1, JTRC_JFIF_EXTENSION,
+	       GETJOCTET(data[5]), (int) totallen);
+      break;
+    }
+  } else {
+    /* Start of APP0 does not match "JFIF" or "JFXX", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP0, (int) totallen);
+  }
+}
+
+
+LOCAL(void)
+examine_app14 (j_decompress_ptr cinfo, JOCTET FAR * data,
+	       unsigned int datalen, INT32 remaining)
+/* Examine first few bytes from an APP14.
+ * Take appropriate action if it is an Adobe marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  unsigned int version, flags0, flags1, transform;
+
+  if (datalen >= APP14_DATA_LEN &&
+      GETJOCTET(data[0]) == 0x41 &&
+      GETJOCTET(data[1]) == 0x64 &&
+      GETJOCTET(data[2]) == 0x6F &&
+      GETJOCTET(data[3]) == 0x62 &&
+      GETJOCTET(data[4]) == 0x65) {
+    /* Found Adobe APP14 marker */
+    version = (GETJOCTET(data[5]) << 8) + GETJOCTET(data[6]);
+    flags0 = (GETJOCTET(data[7]) << 8) + GETJOCTET(data[8]);
+    flags1 = (GETJOCTET(data[9]) << 8) + GETJOCTET(data[10]);
+    transform = GETJOCTET(data[11]);
+    TRACEMS4(cinfo, 1, JTRC_ADOBE, version, flags0, flags1, transform);
+    cinfo->saw_Adobe_marker = TRUE;
+    cinfo->Adobe_transform = (UINT8) transform;
+  } else {
+    /* Start of APP14 does not match "Adobe", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP14, (int) (datalen + remaining));
+  }
+}
+
+
+METHODDEF(boolean)
+get_interesting_appn (j_decompress_ptr cinfo)
+/* Process an APP0 or APP14 marker without saving it */
+{
+  INT32 length;
+  JOCTET b[APPN_DATA_LEN];
+  unsigned int i, numtoread;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  /* get the interesting part of the marker data */
+  if (length >= APPN_DATA_LEN)
+    numtoread = APPN_DATA_LEN;
+  else if (length > 0)
+    numtoread = (unsigned int) length;
+  else
+    numtoread = 0;
+  for (i = 0; i < numtoread; i++)
+    INPUT_BYTE(cinfo, b[i], return FALSE);
+  length -= numtoread;
+
+  /* process it */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, (JOCTET FAR *) b, numtoread, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, (JOCTET FAR *) b, numtoread, length);
+    break;
+  default:
+    /* can't get here unless jpeg_save_markers chooses wrong processor */
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+METHODDEF(boolean)
+save_marker (j_decompress_ptr cinfo)
+/* Save an APPn or COM marker into the marker list */
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  jpeg_saved_marker_ptr cur_marker = marker->cur_marker;
+  unsigned int bytes_read, data_length;
+  JOCTET FAR * data;
+  INT32 length = 0;
+  INPUT_VARS(cinfo);
+
+  if (cur_marker == NULL) {
+    /* begin reading a marker */
+    INPUT_2BYTES(cinfo, length, return FALSE);
+    length -= 2;
+    if (length >= 0) {		/* watch out for bogus length word */
+      /* figure out how much we want to save */
+      unsigned int limit;
+      if (cinfo->unread_marker == (int) M_COM)
+	limit = marker->length_limit_COM;
+      else
+	limit = marker->length_limit_APPn[cinfo->unread_marker - (int) M_APP0];
+      if ((unsigned int) length < limit)
+	limit = (unsigned int) length;
+      /* allocate and initialize the marker item */
+      cur_marker = (jpeg_saved_marker_ptr)
+	(*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				    SIZEOF(struct jpeg_marker_struct) + limit);
+      cur_marker->next = NULL;
+      cur_marker->marker = (UINT8) cinfo->unread_marker;
+      cur_marker->original_length = (unsigned int) length;
+      cur_marker->data_length = limit;
+      /* data area is just beyond the jpeg_marker_struct */
+      data = cur_marker->data = (JOCTET FAR *) (cur_marker + 1);
+      marker->cur_marker = cur_marker;
+      marker->bytes_read = 0;
+      bytes_read = 0;
+      data_length = limit;
+    } else {
+      /* deal with bogus length word */
+      bytes_read = data_length = 0;
+      data = NULL;
+    }
+  } else {
+    /* resume reading a marker */
+    bytes_read = marker->bytes_read;
+    data_length = cur_marker->data_length;
+    data = cur_marker->data + bytes_read;
+  }
+
+  while (bytes_read < data_length) {
+    INPUT_SYNC(cinfo);		/* move the restart point to here */
+    marker->bytes_read = bytes_read;
+    /* If there's not at least one byte in buffer, suspend */
+    MAKE_BYTE_AVAIL(cinfo, return FALSE);
+    /* Copy bytes with reasonable rapidity */
+    while (bytes_read < data_length && bytes_in_buffer > 0) {
+      *data++ = *next_input_byte++;
+      bytes_in_buffer--;
+      bytes_read++;
+    }
+  }
+
+  /* Done reading what we want to read */
+  if (cur_marker != NULL) {	/* will be NULL if bogus length word */
+    /* Add new marker to end of list */
+    if (cinfo->marker_list == NULL) {
+      cinfo->marker_list = cur_marker;
+    } else {
+      jpeg_saved_marker_ptr prev = cinfo->marker_list;
+      while (prev->next != NULL)
+	prev = prev->next;
+      prev->next = cur_marker;
+    }
+    /* Reset pointer & calc remaining data length */
+    data = cur_marker->data;
+    length = cur_marker->original_length - data_length;
+  }
+  /* Reset to initial state for next marker */
+  marker->cur_marker = NULL;
+
+  /* Process the marker if interesting; else just make a generic trace msg */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, data, data_length, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, data, data_length, length);
+    break;
+  default:
+    TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker,
+	     (int) (data_length + length));
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);		/* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+METHODDEF(boolean)
+skip_variable (j_decompress_ptr cinfo)
+/* Skip over an unknown or uninteresting variable-length marker */
+{
+  INT32 length;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker, (int) length);
+
+  INPUT_SYNC(cinfo);		/* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+
+/*
+ * Find the next JPEG marker, save it in cinfo->unread_marker.
+ * Returns FALSE if had to suspend before reaching a marker;
+ * in that case cinfo->unread_marker is unchanged.
+ *
+ * Note that the result might not be a valid marker code,
+ * but it will never be 0 or FF.
+ */
+
+LOCAL(boolean)
+next_marker (j_decompress_ptr cinfo)
+{
+  int c;
+  INPUT_VARS(cinfo);
+
+  for (;;) {
+    INPUT_BYTE(cinfo, c, return FALSE);
+    /* Skip any non-FF bytes.
+     * This may look a bit inefficient, but it will not occur in a valid file.
+     * We sync after each discarded byte so that a suspending data source
+     * can discard the byte from its buffer.
+     */
+    while (c != 0xFF) {
+      cinfo->marker->discarded_bytes++;
+      INPUT_SYNC(cinfo);
+      INPUT_BYTE(cinfo, c, return FALSE);
+    }
+    /* This loop swallows any duplicate FF bytes.  Extra FFs are legal as
+     * pad bytes, so don't count them in discarded_bytes.  We assume there
+     * will not be so many consecutive FF bytes as to overflow a suspending
+     * data source's input buffer.
+     */
+    do {
+      INPUT_BYTE(cinfo, c, return FALSE);
+    } while (c == 0xFF);
+    if (c != 0)
+      break;			/* found a valid marker, exit loop */
+    /* Reach here if we found a stuffed-zero data sequence (FF/00).
+     * Discard it and loop back to try again.
+     */
+    cinfo->marker->discarded_bytes += 2;
+    INPUT_SYNC(cinfo);
+  }
+
+  if (cinfo->marker->discarded_bytes != 0) {
+    WARNMS2(cinfo, JWRN_EXTRANEOUS_DATA, cinfo->marker->discarded_bytes, c);
+    cinfo->marker->discarded_bytes = 0;
+  }
+
+  cinfo->unread_marker = c;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+first_marker (j_decompress_ptr cinfo)
+/* Like next_marker, but used to obtain the initial SOI marker. */
+/* For this marker, we do not allow preceding garbage or fill; otherwise,
+ * we might well scan an entire input file before realizing it ain't JPEG.
+ * If an application wants to process non-JFIF files, it must seek to the
+ * SOI before calling the JPEG library.
+ */
+{
+  int c, c2;
+  INPUT_VARS(cinfo);
+
+  INPUT_BYTE(cinfo, c, return FALSE);
+  INPUT_BYTE(cinfo, c2, return FALSE);
+  if (c != 0xFF || c2 != (int) M_SOI)
+    ERREXIT2(cinfo, JERR_NO_SOI, c, c2);
+
+  cinfo->unread_marker = c2;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Read markers until SOS or EOI.
+ *
+ * Returns same codes as are defined for jpeg_consume_input:
+ * JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ */
+
+METHODDEF(int)
+read_markers (j_decompress_ptr cinfo)
+{
+  /* Outer loop repeats once for each marker. */
+  for (;;) {
+    /* Collect the marker proper, unless we already did. */
+    /* NB: first_marker() enforces the requirement that SOI appear first. */
+    if (cinfo->unread_marker == 0) {
+      if (! cinfo->marker->saw_SOI) {
+	if (! first_marker(cinfo))
+	  return JPEG_SUSPENDED;
+      } else {
+	if (! next_marker(cinfo))
+	  return JPEG_SUSPENDED;
+      }
+    }
+    /* At this point cinfo->unread_marker contains the marker code and the
+     * input point is just past the marker proper, but before any parameters.
+     * A suspension will cause us to return with this state still true.
+     */
+    switch (cinfo->unread_marker) {
+    case M_SOI:
+      if (! get_soi(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF0:		/* Baseline */
+    case M_SOF1:		/* Extended sequential, Huffman */
+      if (! get_sof(cinfo, FALSE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF2:		/* Progressive, Huffman */
+      if (! get_sof(cinfo, TRUE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF9:		/* Extended sequential, arithmetic */
+      if (! get_sof(cinfo, FALSE, TRUE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF10:		/* Progressive, arithmetic */
+      if (! get_sof(cinfo, TRUE, TRUE))
+	return JPEG_SUSPENDED;
+      break;
+
+    /* Currently unsupported SOFn types */
+    case M_SOF3:		/* Lossless, Huffman */
+    case M_SOF5:		/* Differential sequential, Huffman */
+    case M_SOF6:		/* Differential progressive, Huffman */
+    case M_SOF7:		/* Differential lossless, Huffman */
+    case M_JPG:			/* Reserved for JPEG extensions */
+    case M_SOF11:		/* Lossless, arithmetic */
+    case M_SOF13:		/* Differential sequential, arithmetic */
+    case M_SOF14:		/* Differential progressive, arithmetic */
+    case M_SOF15:		/* Differential lossless, arithmetic */
+      ERREXIT1(cinfo, JERR_SOF_UNSUPPORTED, cinfo->unread_marker);
+      break;
+
+    case M_SOS:
+      if (! get_sos(cinfo))
+	return JPEG_SUSPENDED;
+      cinfo->unread_marker = 0;	/* processed the marker */
+      return JPEG_REACHED_SOS;
+    
+    case M_EOI:
+      TRACEMS(cinfo, 1, JTRC_EOI);
+      cinfo->unread_marker = 0;	/* processed the marker */
+      return JPEG_REACHED_EOI;
+      
+    case M_DAC:
+      if (! get_dac(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DHT:
+      if (! get_dht(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DQT:
+      if (! get_dqt(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DRI:
+      if (! get_dri(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_APP0:
+    case M_APP1:
+    case M_APP2:
+    case M_APP3:
+    case M_APP4:
+    case M_APP5:
+    case M_APP6:
+    case M_APP7:
+    case M_APP8:
+    case M_APP9:
+    case M_APP10:
+    case M_APP11:
+    case M_APP12:
+    case M_APP13:
+    case M_APP14:
+    case M_APP15:
+      if (! (*((my_marker_ptr) cinfo->marker)->process_APPn[
+		cinfo->unread_marker - (int) M_APP0]) (cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_COM:
+      if (! (*((my_marker_ptr) cinfo->marker)->process_COM) (cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_RST0:		/* these are all parameterless */
+    case M_RST1:
+    case M_RST2:
+    case M_RST3:
+    case M_RST4:
+    case M_RST5:
+    case M_RST6:
+    case M_RST7:
+    case M_TEM:
+      TRACEMS1(cinfo, 1, JTRC_PARMLESS_MARKER, cinfo->unread_marker);
+      break;
+
+    case M_DNL:			/* Ignore DNL ... perhaps the wrong thing */
+      if (! skip_variable(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    default:			/* must be DHP, EXP, JPGn, or RESn */
+      /* For now, we treat the reserved markers as fatal errors since they are
+       * likely to be used to signal incompatible JPEG Part 3 extensions.
+       * Once the JPEG 3 version-number marker is well defined, this code
+       * ought to change!
+       */
+      ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+      break;
+    }
+    /* Successfully processed marker, so reset state variable */
+    cinfo->unread_marker = 0;
+  } /* end loop */
+}
+
+
+/*
+ * Read a restart marker, which is expected to appear next in the datastream;
+ * if the marker is not there, take appropriate recovery action.
+ * Returns FALSE if suspension is required.
+ *
+ * This is called by the entropy decoder after it has read an appropriate
+ * number of MCUs.  cinfo->unread_marker may be nonzero if the entropy decoder
+ * has already read a marker from the data source.  Under normal conditions
+ * cinfo->unread_marker will be reset to 0 before returning; if not reset,
+ * it holds a marker which the decoder will be unable to read past.
+ */
+
+METHODDEF(boolean)
+read_restart_marker (j_decompress_ptr cinfo)
+{
+  /* Obtain a marker unless we already did. */
+  /* Note that next_marker will complain if it skips any data. */
+  if (cinfo->unread_marker == 0) {
+    if (! next_marker(cinfo))
+      return FALSE;
+  }
+
+  if (cinfo->unread_marker ==
+      ((int) M_RST0 + cinfo->marker->next_restart_num)) {
+    /* Normal case --- swallow the marker and let entropy decoder continue */
+    TRACEMS1(cinfo, 3, JTRC_RST, cinfo->marker->next_restart_num);
+    cinfo->unread_marker = 0;
+  } else {
+    /* Uh-oh, the restart markers have been messed up. */
+    /* Let the data source manager determine how to resync. */
+    if (! (*cinfo->src->resync_to_restart) (cinfo,
+					    cinfo->marker->next_restart_num))
+      return FALSE;
+  }
+
+  /* Update next-restart state */
+  cinfo->marker->next_restart_num = (cinfo->marker->next_restart_num + 1) & 7;
+
+  return TRUE;
+}
+
+
+/*
+ * This is the default resync_to_restart method for data source managers
+ * to use if they don't have any better approach.  Some data source managers
+ * may be able to back up, or may have additional knowledge about the data
+ * which permits a more intelligent recovery strategy; such managers would
+ * presumably supply their own resync method.
+ *
+ * read_restart_marker calls resync_to_restart if it finds a marker other than
+ * the restart marker it was expecting.  (This code is *not* used unless
+ * a nonzero restart interval has been declared.)  cinfo->unread_marker is
+ * the marker code actually found (might be anything, except 0 or FF).
+ * The desired restart marker number (0..7) is passed as a parameter.
+ * This routine is supposed to apply whatever error recovery strategy seems
+ * appropriate in order to position the input stream to the next data segment.
+ * Note that cinfo->unread_marker is treated as a marker appearing before
+ * the current data-source input point; usually it should be reset to zero
+ * before returning.
+ * Returns FALSE if suspension is required.
+ *
+ * This implementation is substantially constrained by wanting to treat the
+ * input as a data stream; this means we can't back up.  Therefore, we have
+ * only the following actions to work with:
+ *   1. Simply discard the marker and let the entropy decoder resume at next
+ *      byte of file.
+ *   2. Read forward until we find another marker, discarding intervening
+ *      data.  (In theory we could look ahead within the current bufferload,
+ *      without having to discard data if we don't find the desired marker.
+ *      This idea is not implemented here, in part because it makes behavior
+ *      dependent on buffer size and chance buffer-boundary positions.)
+ *   3. Leave the marker unread (by failing to zero cinfo->unread_marker).
+ *      This will cause the entropy decoder to process an empty data segment,
+ *      inserting dummy zeroes, and then we will reprocess the marker.
+ *
+ * #2 is appropriate if we think the desired marker lies ahead, while #3 is
+ * appropriate if the found marker is a future restart marker (indicating
+ * that we have missed the desired restart marker, probably because it got
+ * corrupted).
+ * We apply #2 or #3 if the found marker is a restart marker no more than
+ * two counts behind or ahead of the expected one.  We also apply #2 if the
+ * found marker is not a legal JPEG marker code (it's certainly bogus data).
+ * If the found marker is a restart marker more than 2 counts away, we do #1
+ * (too much risk that the marker is erroneous; with luck we will be able to
+ * resync at some future point).
+ * For any valid non-restart JPEG marker, we apply #3.  This keeps us from
+ * overrunning the end of a scan.  An implementation limited to single-scan
+ * files might find it better to apply #2 for markers other than EOI, since
+ * any other marker would have to be bogus data in that case.
+ */
+
+GLOBAL(boolean)
+jpeg_resync_to_restart (j_decompress_ptr cinfo, int desired)
+{
+  int marker = cinfo->unread_marker;
+  int action = 1;
+  
+  /* Always put up a warning. */
+  WARNMS2(cinfo, JWRN_MUST_RESYNC, marker, desired);
+  
+  /* Outer loop handles repeated decision after scanning forward. */
+  for (;;) {
+    if (marker < (int) M_SOF0)
+      action = 2;		/* invalid marker */
+    else if (marker < (int) M_RST0 || marker > (int) M_RST7)
+      action = 3;		/* valid non-restart marker */
+    else {
+      if (marker == ((int) M_RST0 + ((desired+1) & 7)) ||
+	  marker == ((int) M_RST0 + ((desired+2) & 7)))
+	action = 3;		/* one of the next two expected restarts */
+      else if (marker == ((int) M_RST0 + ((desired-1) & 7)) ||
+	       marker == ((int) M_RST0 + ((desired-2) & 7)))
+	action = 2;		/* a prior restart, so advance */
+      else
+	action = 1;		/* desired restart or too far away */
+    }
+    TRACEMS2(cinfo, 4, JTRC_RECOVERY_ACTION, marker, action);
+    switch (action) {
+    case 1:
+      /* Discard marker and let entropy decoder resume processing. */
+      cinfo->unread_marker = 0;
+      return TRUE;
+    case 2:
+      /* Scan to the next marker, and repeat the decision loop. */
+      if (! next_marker(cinfo))
+	return FALSE;
+      marker = cinfo->unread_marker;
+      break;
+    case 3:
+      /* Return without advancing past this marker. */
+      /* Entropy decoder will be forced to process an empty segment. */
+      return TRUE;
+    }
+  } /* end loop */
+}
+
+
+/*
+ * Reset marker processing state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_marker_reader (j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  cinfo->comp_info = NULL;		/* until allocated by get_sof */
+  cinfo->input_scan_number = 0;		/* no SOS seen yet */
+  cinfo->unread_marker = 0;		/* no pending marker */
+  marker->pub.saw_SOI = FALSE;		/* set internal state too */
+  marker->pub.saw_SOF = FALSE;
+  marker->pub.discarded_bytes = 0;
+  marker->cur_marker = NULL;
+}
+
+
+/*
+ * Initialize the marker reader module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_marker_reader (j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker;
+  int i;
+
+  /* Create subobject in permanent pool */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				SIZEOF(my_marker_reader));
+  cinfo->marker = (struct jpeg_marker_reader *) marker;
+  /* Initialize public method pointers */
+  marker->pub.reset_marker_reader = reset_marker_reader;
+  marker->pub.read_markers = read_markers;
+  marker->pub.read_restart_marker = read_restart_marker;
+  /* Initialize COM/APPn processing.
+   * By default, we examine and then discard APP0 and APP14,
+   * but simply discard COM and all other APPn.
+   */
+  marker->process_COM = skip_variable;
+  marker->length_limit_COM = 0;
+  for (i = 0; i < 16; i++) {
+    marker->process_APPn[i] = skip_variable;
+    marker->length_limit_APPn[i] = 0;
+  }
+  marker->process_APPn[0] = get_interesting_appn;
+  marker->process_APPn[14] = get_interesting_appn;
+  /* Reset marker processing state */
+  reset_marker_reader(cinfo);
+}
+
+
+/*
+ * Control saving of COM and APPn markers into marker_list.
+ */
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+GLOBAL(void)
+jpeg_save_markers (j_decompress_ptr cinfo, int marker_code,
+		   unsigned int length_limit)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  long maxlength;
+  jpeg_marker_parser_method processor;
+
+  /* Length limit mustn't be larger than what we can allocate
+   * (should only be a concern in a 16-bit environment).
+   */
+  maxlength = cinfo->mem->max_alloc_chunk - SIZEOF(struct jpeg_marker_struct);
+  if (((long) length_limit) > maxlength)
+    length_limit = (unsigned int) maxlength;
+
+  /* Choose processor routine to use.
+   * APP0/APP14 have special requirements.
+   */
+  if (length_limit) {
+    processor = save_marker;
+    /* If saving APP0/APP14, save at least enough for our internal use. */
+    if (marker_code == (int) M_APP0 && length_limit < APP0_DATA_LEN)
+      length_limit = APP0_DATA_LEN;
+    else if (marker_code == (int) M_APP14 && length_limit < APP14_DATA_LEN)
+      length_limit = APP14_DATA_LEN;
+  } else {
+    processor = skip_variable;
+    /* If discarding APP0/APP14, use our regular on-the-fly processor. */
+    if (marker_code == (int) M_APP0 || marker_code == (int) M_APP14)
+      processor = get_interesting_appn;
+  }
+
+  if (marker_code == (int) M_COM) {
+    marker->process_COM = processor;
+    marker->length_limit_COM = length_limit;
+  } else if (marker_code >= (int) M_APP0 && marker_code <= (int) M_APP15) {
+    marker->process_APPn[marker_code - (int) M_APP0] = processor;
+    marker->length_limit_APPn[marker_code - (int) M_APP0] = length_limit;
+  } else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+/*
+ * Install a special processing method for COM or APPn markers.
+ */
+
+GLOBAL(void)
+jpeg_set_marker_processor (j_decompress_ptr cinfo, int marker_code,
+			   jpeg_marker_parser_method routine)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  if (marker_code == (int) M_COM)
+    marker->process_COM = routine;
+  else if (marker_code >= (int) M_APP0 && marker_code <= (int) M_APP15)
+    marker->process_APPn[marker_code - (int) M_APP0] = routine;
+  else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}
diff --git a/jdmaster.c b/jdmaster.c
new file mode 100644
index 0000000..38b7dd0
--- /dev/null
+++ b/jdmaster.c
@@ -0,0 +1,589 @@
+/*
+ * jdmaster.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Copyright (C) 2009-2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains master control logic for the JPEG decompressor.
+ * These routines are concerned with selecting the modules to be executed
+ * and with determining the number of passes and the work to be done in each
+ * pass.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jpegcomp.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_decomp_master pub; /* public fields */
+
+  int pass_number;		/* # of passes completed */
+
+  boolean using_merged_upsample; /* TRUE if using merged upsample/cconvert */
+
+  /* Saved references to initialized quantizer modules,
+   * in case we need to switch modes.
+   */
+  struct jpeg_color_quantizer * quantizer_1pass;
+  struct jpeg_color_quantizer * quantizer_2pass;
+} my_decomp_master;
+
+typedef my_decomp_master * my_master_ptr;
+
+
+/*
+ * Determine whether merged upsample/color conversion should be used.
+ * CRUCIAL: this must match the actual capabilities of jdmerge.c!
+ */
+
+LOCAL(boolean)
+use_merged_upsample (j_decompress_ptr cinfo)
+{
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  /* Merging is the equivalent of plain box-filter upsampling */
+  if (cinfo->do_fancy_upsampling || cinfo->CCIR601_sampling)
+    return FALSE;
+  /* jdmerge.c only supports YCC=>RGB color conversion */
+  if (cinfo->jpeg_color_space != JCS_YCbCr || cinfo->num_components != 3 ||
+      (cinfo->out_color_space != JCS_RGB &&
+      cinfo->out_color_space != JCS_EXT_RGB &&
+      cinfo->out_color_space != JCS_EXT_RGBX &&
+      cinfo->out_color_space != JCS_EXT_BGR &&
+      cinfo->out_color_space != JCS_EXT_BGRX &&
+      cinfo->out_color_space != JCS_EXT_XBGR &&
+      cinfo->out_color_space != JCS_EXT_XRGB) ||
+      cinfo->out_color_components != rgb_pixelsize[cinfo->out_color_space])
+    return FALSE;
+  /* and it only handles 2h1v or 2h2v sampling ratios */
+  if (cinfo->comp_info[0].h_samp_factor != 2 ||
+      cinfo->comp_info[1].h_samp_factor != 1 ||
+      cinfo->comp_info[2].h_samp_factor != 1 ||
+      cinfo->comp_info[0].v_samp_factor >  2 ||
+      cinfo->comp_info[1].v_samp_factor != 1 ||
+      cinfo->comp_info[2].v_samp_factor != 1)
+    return FALSE;
+  /* furthermore, it doesn't work if we've scaled the IDCTs differently */
+  if (cinfo->comp_info[0]._DCT_scaled_size != cinfo->_min_DCT_scaled_size ||
+      cinfo->comp_info[1]._DCT_scaled_size != cinfo->_min_DCT_scaled_size ||
+      cinfo->comp_info[2]._DCT_scaled_size != cinfo->_min_DCT_scaled_size)
+    return FALSE;
+  /* ??? also need to test for upsample-time rescaling, when & if supported */
+  return TRUE;			/* by golly, it'll work... */
+#else
+  return FALSE;
+#endif
+}
+
+
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ * Also note that it may be called before the master module is initialized!
+ */
+
+GLOBAL(void)
+jpeg_calc_output_dimensions (j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+#endif
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_READY)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+  /* Compute actual output image dimensions and DCT scaling choices. */
+  if (cinfo->scale_num * 8 <= cinfo->scale_denom) {
+    /* Provide 1/8 scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, 8L);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, 8L);
+#if JPEG_LIB_VERSION >= 70
+    cinfo->min_DCT_h_scaled_size = cinfo->min_DCT_v_scaled_size = 1;
+#else
+    cinfo->min_DCT_scaled_size = 1;
+#endif
+  } else if (cinfo->scale_num * 4 <= cinfo->scale_denom) {
+    /* Provide 1/4 scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, 4L);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, 4L);
+#if JPEG_LIB_VERSION >= 70
+    cinfo->min_DCT_h_scaled_size = cinfo->min_DCT_v_scaled_size = 2;
+#else
+    cinfo->min_DCT_scaled_size = 2;
+#endif
+  } else if (cinfo->scale_num * 2 <= cinfo->scale_denom) {
+    /* Provide 1/2 scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, 2L);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, 2L);
+#if JPEG_LIB_VERSION >= 70
+    cinfo->min_DCT_h_scaled_size = cinfo->min_DCT_v_scaled_size = 4;
+#else
+    cinfo->min_DCT_scaled_size = 4;
+#endif
+  } else {
+    /* Provide 1/1 scaling */
+    cinfo->output_width = cinfo->image_width;
+    cinfo->output_height = cinfo->image_height;
+#if JPEG_LIB_VERSION >= 70
+    cinfo->min_DCT_h_scaled_size = cinfo->min_DCT_v_scaled_size = DCTSIZE;
+#else
+    cinfo->min_DCT_scaled_size = DCTSIZE;
+#endif
+  }
+  /* In selecting the actual DCT scaling for each component, we try to
+   * scale up the chroma components via IDCT scaling rather than upsampling.
+   * This saves time if the upsampler gets to use 1:1 scaling.
+   * Note this code assumes that the supported DCT scalings are powers of 2.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    int ssize = cinfo->_min_DCT_scaled_size;
+    while (ssize < DCTSIZE &&
+	   (compptr->h_samp_factor * ssize * 2 <=
+	    cinfo->max_h_samp_factor * cinfo->_min_DCT_scaled_size) &&
+	   (compptr->v_samp_factor * ssize * 2 <=
+	    cinfo->max_v_samp_factor * cinfo->_min_DCT_scaled_size)) {
+      ssize = ssize * 2;
+    }
+#if JPEG_LIB_VERSION >= 70
+    compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size = ssize;
+#else
+    compptr->DCT_scaled_size = ssize;
+#endif
+  }
+
+  /* Recompute downsampled dimensions of components;
+   * application needs to know these if using raw downsampled data.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Size in samples, after IDCT scaling */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width *
+		    (long) (compptr->h_samp_factor * compptr->_DCT_scaled_size),
+		    (long) (cinfo->max_h_samp_factor * DCTSIZE));
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height *
+		    (long) (compptr->v_samp_factor * compptr->_DCT_scaled_size),
+		    (long) (cinfo->max_v_samp_factor * DCTSIZE));
+  }
+
+#else /* !IDCT_SCALING_SUPPORTED */
+
+  /* Hardwire it to "no scaling" */
+  cinfo->output_width = cinfo->image_width;
+  cinfo->output_height = cinfo->image_height;
+  /* jdinput.c has already initialized DCT_scaled_size to DCTSIZE,
+   * and has computed unscaled downsampled_width and downsampled_height.
+   */
+
+#endif /* IDCT_SCALING_SUPPORTED */
+
+  /* Report number of components in selected colorspace. */
+  /* Probably this should be in the color conversion module... */
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    break;
+  case JCS_RGB:
+  case JCS_EXT_RGB:
+  case JCS_EXT_RGBX:
+  case JCS_EXT_BGR:
+  case JCS_EXT_BGRX:
+  case JCS_EXT_XBGR:
+  case JCS_EXT_XRGB:
+    cinfo->out_color_components = rgb_pixelsize[cinfo->out_color_space];
+    break;
+  case JCS_YCbCr:
+    cinfo->out_color_components = 3;
+    break;
+  case JCS_CMYK:
+  case JCS_YCCK:
+    cinfo->out_color_components = 4;
+    break;
+  default:			/* else must be same colorspace as in file */
+    cinfo->out_color_components = cinfo->num_components;
+    break;
+  }
+  cinfo->output_components = (cinfo->quantize_colors ? 1 :
+			      cinfo->out_color_components);
+
+  /* See if upsampler will want to emit more than one row at a time */
+  if (use_merged_upsample(cinfo))
+    cinfo->rec_outbuf_height = cinfo->max_v_samp_factor;
+  else
+    cinfo->rec_outbuf_height = 1;
+}
+
+
+/*
+ * Several decompression processes need to range-limit values to the range
+ * 0..MAXJSAMPLE; the input value may fall somewhat outside this range
+ * due to noise introduced by quantization, roundoff error, etc.  These
+ * processes are inner loops and need to be as fast as possible.  On most
+ * machines, particularly CPUs with pipelines or instruction prefetch,
+ * a (subscript-check-less) C table lookup
+ *		x = sample_range_limit[x];
+ * is faster than explicit tests
+ *		if (x < 0)  x = 0;
+ *		else if (x > MAXJSAMPLE)  x = MAXJSAMPLE;
+ * These processes all use a common table prepared by the routine below.
+ *
+ * For most steps we can mathematically guarantee that the initial value
+ * of x is within MAXJSAMPLE+1 of the legal range, so a table running from
+ * -(MAXJSAMPLE+1) to 2*MAXJSAMPLE+1 is sufficient.  But for the initial
+ * limiting step (just after the IDCT), a wildly out-of-range value is 
+ * possible if the input data is corrupt.  To avoid any chance of indexing
+ * off the end of memory and getting a bad-pointer trap, we perform the
+ * post-IDCT limiting thus:
+ *		x = range_limit[x & MASK];
+ * where MASK is 2 bits wider than legal sample data, ie 10 bits for 8-bit
+ * samples.  Under normal circumstances this is more than enough range and
+ * a correct output will be generated; with bogus input data the mask will
+ * cause wraparound, and we will safely generate a bogus-but-in-range output.
+ * For the post-IDCT step, we want to convert the data from signed to unsigned
+ * representation by adding CENTERJSAMPLE at the same time that we limit it.
+ * So the post-IDCT limiting table ends up looking like this:
+ *   CENTERJSAMPLE,CENTERJSAMPLE+1,...,MAXJSAMPLE,
+ *   MAXJSAMPLE (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0          (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0,1,...,CENTERJSAMPLE-1
+ * Negative inputs select values from the upper half of the table after
+ * masking.
+ *
+ * We can save some space by overlapping the start of the post-IDCT table
+ * with the simpler range limiting table.  The post-IDCT table begins at
+ * sample_range_limit + CENTERJSAMPLE.
+ *
+ * Note that the table is allocated in near data space on PCs; it's small
+ * enough and used often enough to justify this.
+ */
+
+LOCAL(void)
+prepare_range_limit_table (j_decompress_ptr cinfo)
+/* Allocate and fill in the sample_range_limit table */
+{
+  JSAMPLE * table;
+  int i;
+
+  table = (JSAMPLE *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+		(5 * (MAXJSAMPLE+1) + CENTERJSAMPLE) * SIZEOF(JSAMPLE));
+  table += (MAXJSAMPLE+1);	/* allow negative subscripts of simple table */
+  cinfo->sample_range_limit = table;
+  /* First segment of "simple" table: limit[x] = 0 for x < 0 */
+  MEMZERO(table - (MAXJSAMPLE+1), (MAXJSAMPLE+1) * SIZEOF(JSAMPLE));
+  /* Main part of "simple" table: limit[x] = x */
+  for (i = 0; i <= MAXJSAMPLE; i++)
+    table[i] = (JSAMPLE) i;
+  table += CENTERJSAMPLE;	/* Point to where post-IDCT table starts */
+  /* End of simple table, rest of first half of post-IDCT table */
+  for (i = CENTERJSAMPLE; i < 2*(MAXJSAMPLE+1); i++)
+    table[i] = MAXJSAMPLE;
+  /* Second half of post-IDCT table */
+  MEMZERO(table + (2 * (MAXJSAMPLE+1)),
+	  (2 * (MAXJSAMPLE+1) - CENTERJSAMPLE) * SIZEOF(JSAMPLE));
+  MEMCOPY(table + (4 * (MAXJSAMPLE+1) - CENTERJSAMPLE),
+	  cinfo->sample_range_limit, CENTERJSAMPLE * SIZEOF(JSAMPLE));
+}
+
+
+/*
+ * Master selection of decompression modules.
+ * This is done once at jpeg_start_decompress time.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ * We also initialize the decompressor input side to begin consuming data.
+ *
+ * Since jpeg_read_header has finished, we know what is in the SOF
+ * and (first) SOS markers.  We also have all the application parameter
+ * settings.
+ */
+
+LOCAL(void)
+master_selection (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+  boolean use_c_buffer;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+  /* Initialize dimensions and other stuff */
+  jpeg_calc_output_dimensions(cinfo);
+  prepare_range_limit_table(cinfo);
+
+  /* Width of an output scanline must be representable as JDIMENSION. */
+  samplesperrow = (long) cinfo->output_width * (long) cinfo->out_color_components;
+  jd_samplesperrow = (JDIMENSION) samplesperrow;
+  if ((long) jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* Initialize my private state */
+  master->pass_number = 0;
+  master->using_merged_upsample = use_merged_upsample(cinfo);
+
+  /* Color quantizer selection */
+  master->quantizer_1pass = NULL;
+  master->quantizer_2pass = NULL;
+  /* No mode changes if not using buffered-image mode. */
+  if (! cinfo->quantize_colors || ! cinfo->buffered_image) {
+    cinfo->enable_1pass_quant = FALSE;
+    cinfo->enable_external_quant = FALSE;
+    cinfo->enable_2pass_quant = FALSE;
+  }
+  if (cinfo->quantize_colors) {
+    if (cinfo->raw_data_out)
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    /* 2-pass quantizer only works in 3-component color space. */
+    if (cinfo->out_color_components != 3) {
+      cinfo->enable_1pass_quant = TRUE;
+      cinfo->enable_external_quant = FALSE;
+      cinfo->enable_2pass_quant = FALSE;
+      cinfo->colormap = NULL;
+    } else if (cinfo->colormap != NULL) {
+      cinfo->enable_external_quant = TRUE;
+    } else if (cinfo->two_pass_quantize) {
+      cinfo->enable_2pass_quant = TRUE;
+    } else {
+      cinfo->enable_1pass_quant = TRUE;
+    }
+
+    if (cinfo->enable_1pass_quant) {
+#ifdef QUANT_1PASS_SUPPORTED
+      jinit_1pass_quantizer(cinfo);
+      master->quantizer_1pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+
+    /* We use the 2-pass code to map to external colormaps. */
+    if (cinfo->enable_2pass_quant || cinfo->enable_external_quant) {
+#ifdef QUANT_2PASS_SUPPORTED
+      jinit_2pass_quantizer(cinfo);
+      master->quantizer_2pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+    /* If both quantizers are initialized, the 2-pass one is left active;
+     * this is necessary for starting with quantization to an external map.
+     */
+  }
+
+  /* Post-processing: in particular, color conversion first */
+  if (! cinfo->raw_data_out) {
+    if (master->using_merged_upsample) {
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+      jinit_merged_upsampler(cinfo); /* does color conversion too */
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else {
+      jinit_color_deconverter(cinfo);
+      jinit_upsampler(cinfo);
+    }
+    jinit_d_post_controller(cinfo, cinfo->enable_2pass_quant);
+  }
+  /* Inverse DCT */
+  jinit_inverse_dct(cinfo);
+  /* Entropy decoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef D_PROGRESSIVE_SUPPORTED
+      jinit_phuff_decoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_decoder(cinfo);
+  }
+
+  /* Initialize principal buffer controllers. */
+  use_c_buffer = cinfo->inputctl->has_multiple_scans || cinfo->buffered_image;
+  jinit_d_coef_controller(cinfo, use_c_buffer);
+
+  if (! cinfo->raw_data_out)
+    jinit_d_main_controller(cinfo, FALSE /* never need full buffer here */);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* If jpeg_start_decompress will read the whole file, initialize
+   * progress monitoring appropriately.  The input step is counted
+   * as one pass.
+   */
+  if (cinfo->progress != NULL && ! cinfo->buffered_image &&
+      cinfo->inputctl->has_multiple_scans) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = (cinfo->enable_2pass_quant ? 3 : 2);
+    /* Count the input pass as done */
+    master->pass_number++;
+  }
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each output pass.  We determine which
+ * modules will be active during this pass and give them appropriate
+ * start_pass calls.  We also set is_dummy_pass to indicate whether this
+ * is a "real" output pass or a dummy pass for color quantization.
+ * (In the latter case, jdapistd.c will crank the pass to completion.)
+ */
+
+METHODDEF(void)
+prepare_for_output_pass (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  if (master->pub.is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Final pass of 2-pass quantization */
+    master->pub.is_dummy_pass = FALSE;
+    (*cinfo->cquantize->start_pass) (cinfo, FALSE);
+    (*cinfo->post->start_pass) (cinfo, JBUF_CRANK_DEST);
+    (*cinfo->main->start_pass) (cinfo, JBUF_CRANK_DEST);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  } else {
+    if (cinfo->quantize_colors && cinfo->colormap == NULL) {
+      /* Select new quantization method */
+      if (cinfo->two_pass_quantize && cinfo->enable_2pass_quant) {
+	cinfo->cquantize = master->quantizer_2pass;
+	master->pub.is_dummy_pass = TRUE;
+      } else if (cinfo->enable_1pass_quant) {
+	cinfo->cquantize = master->quantizer_1pass;
+      } else {
+	ERREXIT(cinfo, JERR_MODE_CHANGE);
+      }
+    }
+    (*cinfo->idct->start_pass) (cinfo);
+    (*cinfo->coef->start_output_pass) (cinfo);
+    if (! cinfo->raw_data_out) {
+      if (! master->using_merged_upsample)
+	(*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->upsample->start_pass) (cinfo);
+      if (cinfo->quantize_colors)
+	(*cinfo->cquantize->start_pass) (cinfo, master->pub.is_dummy_pass);
+      (*cinfo->post->start_pass) (cinfo,
+	    (master->pub.is_dummy_pass ? JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+      (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+  }
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->pass_number +
+				    (master->pub.is_dummy_pass ? 2 : 1);
+    /* In buffered-image mode, we assume one more output pass if EOI not
+     * yet reached, but no more passes if EOI has been reached.
+     */
+    if (cinfo->buffered_image && ! cinfo->inputctl->eoi_reached) {
+      cinfo->progress->total_passes += (cinfo->enable_2pass_quant ? 2 : 1);
+    }
+  }
+}
+
+
+/*
+ * Finish up at end of an output pass.
+ */
+
+METHODDEF(void)
+finish_output_pass (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  if (cinfo->quantize_colors)
+    (*cinfo->cquantize->finish_pass) (cinfo);
+  master->pass_number++;
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Switch to a new external colormap between output passes.
+ */
+
+GLOBAL(void)
+jpeg_new_colormap (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_BUFIMAGE)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (cinfo->quantize_colors && cinfo->enable_external_quant &&
+      cinfo->colormap != NULL) {
+    /* Select 2-pass quantizer for external colormap use */
+    cinfo->cquantize = master->quantizer_2pass;
+    /* Notify quantizer of colormap change */
+    (*cinfo->cquantize->new_color_map) (cinfo);
+    master->pub.is_dummy_pass = FALSE; /* just in case */
+  } else
+    ERREXIT(cinfo, JERR_MODE_CHANGE);
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+/*
+ * Initialize master decompression control and select active modules.
+ * This is performed at the start of jpeg_start_decompress.
+ */
+
+GLOBAL(void)
+jinit_master_decompress (j_decompress_ptr cinfo)
+{
+  my_master_ptr master;
+
+  master = (my_master_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(my_decomp_master));
+  cinfo->master = (struct jpeg_decomp_master *) master;
+  master->pub.prepare_for_output_pass = prepare_for_output_pass;
+  master->pub.finish_output_pass = finish_output_pass;
+
+  master->pub.is_dummy_pass = FALSE;
+
+  master_selection(cinfo);
+}
diff --git a/jdmerge.c b/jdmerge.c
new file mode 100644
index 0000000..edf061a
--- /dev/null
+++ b/jdmerge.c
@@ -0,0 +1,409 @@
+/*
+ * jdmerge.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2009, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains code for merged upsampling/color conversion.
+ *
+ * This file combines functions from jdsample.c and jdcolor.c;
+ * read those files first to understand what's going on.
+ *
+ * When the chroma components are to be upsampled by simple replication
+ * (ie, box filtering), we can save some work in color conversion by
+ * calculating all the output pixels corresponding to a pair of chroma
+ * samples at one time.  In the conversion equations
+ *	R = Y           + K1 * Cr
+ *	G = Y + K2 * Cb + K3 * Cr
+ *	B = Y + K4 * Cb
+ * only the Y term varies among the group of pixels corresponding to a pair
+ * of chroma samples, so the rest of the terms can be calculated just once.
+ * At typical sampling ratios, this eliminates half or three-quarters of the
+ * multiplications needed for color conversion.
+ *
+ * This file currently provides implementations for the following cases:
+ *	YCbCr => RGB color conversion only.
+ *	Sampling ratios of 2h1v or 2h2v.
+ *	No scaling needed at upsample time.
+ *	Corner-aligned (non-CCIR601) sampling alignment.
+ * Other special cases could be added, but in most applications these are
+ * the only common cases.  (For uncommon cases we fall back on the more
+ * general code in jdsample.c and jdcolor.c.)
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;	/* public fields */
+
+  /* Pointer to routine to do actual upsampling/conversion of one row group */
+  JMETHOD(void, upmethod, (j_decompress_ptr cinfo,
+			   JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+			   JSAMPARRAY output_buf));
+
+  /* Private state for YCC->RGB conversion */
+  int * Cr_r_tab;		/* => table for Cr to R conversion */
+  int * Cb_b_tab;		/* => table for Cb to B conversion */
+  INT32 * Cr_g_tab;		/* => table for Cr to G conversion */
+  INT32 * Cb_g_tab;		/* => table for Cb to G conversion */
+
+  /* For 2:1 vertical sampling, we produce two output rows at a time.
+   * We need a "spare" row buffer to hold the second output row if the
+   * application provides just a one-row buffer; we also use the spare
+   * to discard the dummy last row if the image height is odd.
+   */
+  JSAMPROW spare_row;
+  boolean spare_full;		/* T if spare buffer is occupied */
+
+  JDIMENSION out_row_width;	/* samples per output row */
+  JDIMENSION rows_to_go;	/* counts rows remaining in image */
+} my_upsampler;
+
+typedef my_upsampler * my_upsample_ptr;
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define ONE_HALF	((INT32) 1 << (SCALEBITS-1))
+#define FIX(x)		((INT32) ((x) * (1L<<SCALEBITS) + 0.5))
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ * This is taken directly from jdcolor.c; see that file for more info.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int i;
+  INT32 x;
+  SHIFT_TEMPS
+
+  upsample->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  upsample->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  upsample->Cr_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+  upsample->Cb_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+
+  for (i = 0, x = -CENTERJSAMPLE; i <= MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0..MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    upsample->Cr_r_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    upsample->Cb_b_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    upsample->Cr_g_tab[i] = (- FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    upsample->Cb_g_tab[i] = (- FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_merged_upsample (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Mark the spare buffer empty */
+  upsample->spare_full = FALSE;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * The control routine just handles the row buffering considerations.
+ */
+
+METHODDEF(void)
+merged_2v_upsample (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+/* 2:1 vertical sampling case: may need a spare row. */
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  JSAMPROW work_ptrs[2];
+  JDIMENSION num_rows;		/* number of rows returned to caller */
+
+  if (upsample->spare_full) {
+    /* If we have a spare row saved from a previous cycle, just return it. */
+    jcopy_sample_rows(& upsample->spare_row, 0, output_buf + *out_row_ctr, 0,
+		      1, upsample->out_row_width);
+    num_rows = 1;
+    upsample->spare_full = FALSE;
+  } else {
+    /* Figure number of rows to return to caller. */
+    num_rows = 2;
+    /* Not more than the distance to the end of the image. */
+    if (num_rows > upsample->rows_to_go)
+      num_rows = upsample->rows_to_go;
+    /* And not more than what the client can accept: */
+    out_rows_avail -= *out_row_ctr;
+    if (num_rows > out_rows_avail)
+      num_rows = out_rows_avail;
+    /* Create output pointer array for upsampler. */
+    work_ptrs[0] = output_buf[*out_row_ctr];
+    if (num_rows > 1) {
+      work_ptrs[1] = output_buf[*out_row_ctr + 1];
+    } else {
+      work_ptrs[1] = upsample->spare_row;
+      upsample->spare_full = TRUE;
+    }
+    /* Now do the upsampling. */
+    (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr, work_ptrs);
+  }
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (! upsample->spare_full)
+    (*in_row_group_ctr)++;
+}
+
+
+METHODDEF(void)
+merged_1v_upsample (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+/* 1:1 vertical sampling case: much easier, never need a spare row. */
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Just do the upsampling. */
+  (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr,
+			 output_buf + *out_row_ctr);
+  /* Adjust counts */
+  (*out_row_ctr)++;
+  (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by the control routines to do
+ * the actual upsampling/conversion.  One row group is processed per call.
+ *
+ * Note: since we may be writing directly into application-supplied buffers,
+ * we have to be honest about the output width; we can't assume the buffer
+ * has been rounded up to an even width.
+ */
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+METHODDEF(void)
+h2v1_merged_upsample (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+		      JSAMPARRAY output_buf)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register JSAMPROW outptr;
+  JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = upsample->Cr_r_tab;
+  int * Cbbtab = upsample->Cb_b_tab;
+  INT32 * Crgtab = upsample->Cr_g_tab;
+  INT32 * Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+  /* Loop for each pair of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = GETJSAMPLE(*inptr1++);
+    cr = GETJSAMPLE(*inptr2++);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 2 Y values and emit 2 pixels */
+    y  = GETJSAMPLE(*inptr0++);
+    outptr[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    outptr += rgb_pixelsize[cinfo->out_color_space];
+    y  = GETJSAMPLE(*inptr0++);
+    outptr[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    outptr += rgb_pixelsize[cinfo->out_color_space];
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = GETJSAMPLE(*inptr1);
+    cr = GETJSAMPLE(*inptr2);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = GETJSAMPLE(*inptr0);
+    outptr[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+  }
+}
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ */
+
+METHODDEF(void)
+h2v2_merged_upsample (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+		      JSAMPARRAY output_buf)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register JSAMPROW outptr0, outptr1;
+  JSAMPROW inptr00, inptr01, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = upsample->Cr_r_tab;
+  int * Cbbtab = upsample->Cb_b_tab;
+  INT32 * Crgtab = upsample->Cr_g_tab;
+  INT32 * Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr00 = input_buf[0][in_row_group_ctr*2];
+  inptr01 = input_buf[0][in_row_group_ctr*2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+  /* Loop for each group of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = GETJSAMPLE(*inptr1++);
+    cr = GETJSAMPLE(*inptr2++);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 4 Y values and emit 4 pixels */
+    y  = GETJSAMPLE(*inptr00++);
+    outptr0[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr0[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr0[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    outptr0 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr00++);
+    outptr0[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr0[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr0[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    outptr0 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr01++);
+    outptr1[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr1[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr1[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    outptr1 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr01++);
+    outptr1[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr1[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr1[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    outptr1 += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = GETJSAMPLE(*inptr1);
+    cr = GETJSAMPLE(*inptr2);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = GETJSAMPLE(*inptr00);
+    outptr0[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr0[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr0[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+    y  = GETJSAMPLE(*inptr01);
+    outptr1[rgb_red[cinfo->out_color_space]] =   range_limit[y + cred];
+    outptr1[rgb_green[cinfo->out_color_space]] = range_limit[y + cgreen];
+    outptr1[rgb_blue[cinfo->out_color_space]] =  range_limit[y + cblue];
+  }
+}
+
+
+/*
+ * Module initialization routine for merged upsampling/color conversion.
+ *
+ * NB: this is called under the conditions determined by use_merged_upsample()
+ * in jdmaster.c.  That routine MUST correspond to the actual capabilities
+ * of this module; no safety checks are made here.
+ */
+
+GLOBAL(void)
+jinit_merged_upsampler (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+
+  upsample = (my_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *) upsample;
+  upsample->pub.start_pass = start_pass_merged_upsample;
+  upsample->pub.need_context_rows = FALSE;
+
+  upsample->out_row_width = cinfo->output_width * cinfo->out_color_components;
+
+  if (cinfo->max_v_samp_factor == 2) {
+    upsample->pub.upsample = merged_2v_upsample;
+    if (jsimd_can_h2v2_merged_upsample())
+      upsample->upmethod = jsimd_h2v2_merged_upsample;
+    else
+      upsample->upmethod = h2v2_merged_upsample;
+    /* Allocate a spare row buffer */
+    upsample->spare_row = (JSAMPROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+		(size_t) (upsample->out_row_width * SIZEOF(JSAMPLE)));
+  } else {
+    upsample->pub.upsample = merged_1v_upsample;
+    if (jsimd_can_h2v1_merged_upsample())
+      upsample->upmethod = jsimd_h2v1_merged_upsample;
+    else
+      upsample->upmethod = h2v1_merged_upsample;
+    /* No spare row needed */
+    upsample->spare_row = NULL;
+  }
+
+  build_ycc_rgb_table(cinfo);
+}
+
+#endif /* UPSAMPLE_MERGING_SUPPORTED */
diff --git a/jdphuff.c b/jdphuff.c
new file mode 100644
index 0000000..2267809
--- /dev/null
+++ b/jdphuff.c
@@ -0,0 +1,668 @@
+/*
+ * jdphuff.c
+ *
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy decoding routines for progressive JPEG.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdhuff.h"		/* Declarations shared with jdhuff.c */
+
+
+#ifdef D_PROGRESSIVE_SUPPORTED
+
+/*
+ * Expanded entropy decoder object for progressive Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  unsigned int EOBRUN;			/* remaining EOBs in EOBRUN */
+  int last_dc_val[MAX_COMPS_IN_SCAN];	/* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).EOBRUN = (src).EOBRUN, \
+	 (dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;	/* Bit buffer at start of MCU */
+  savable_state saved;		/* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl * derived_tbls[NUM_HUFF_TBLS];
+
+  d_derived_tbl * ac_derived_tbl; /* active table during an AC scan */
+} phuff_entropy_decoder;
+
+typedef phuff_entropy_decoder * phuff_entropy_ptr;
+
+/* Forward declarations */
+METHODDEF(boolean) decode_mcu_DC_first JPP((j_decompress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(boolean) decode_mcu_AC_first JPP((j_decompress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(boolean) decode_mcu_DC_refine JPP((j_decompress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+METHODDEF(boolean) decode_mcu_AC_refine JPP((j_decompress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_phuff_decoder (j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  boolean is_DC_band, bad;
+  int ci, coefi, tbl;
+  int *coef_bit_ptr;
+  jpeg_component_info * compptr;
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* Validate scan parameters */
+  bad = FALSE;
+  if (is_DC_band) {
+    if (cinfo->Se != 0)
+      bad = TRUE;
+  } else {
+    /* need not check Ss/Se < 0 since they came from unsigned bytes */
+    if (cinfo->Ss > cinfo->Se || cinfo->Se >= DCTSIZE2)
+      bad = TRUE;
+    /* AC scans may have only one component */
+    if (cinfo->comps_in_scan != 1)
+      bad = TRUE;
+  }
+  if (cinfo->Ah != 0) {
+    /* Successive approximation refinement scan: must have Al = Ah-1. */
+    if (cinfo->Al != cinfo->Ah-1)
+      bad = TRUE;
+  }
+  if (cinfo->Al > 13)		/* need not check for < 0 */
+    bad = TRUE;
+  /* Arguably the maximum Al value should be less than 13 for 8-bit precision,
+   * but the spec doesn't say so, and we try to be liberal about what we
+   * accept.  Note: large Al values could result in out-of-range DC
+   * coefficients during early scans, leading to bizarre displays due to
+   * overflows in the IDCT math.  But we won't crash.
+   */
+  if (bad)
+    ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+	     cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+  /* Update progression status, and verify that scan order is legal.
+   * Note that inter-scan inconsistencies are treated as warnings
+   * not fatal errors ... not clear if this is right way to behave.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    int cindex = cinfo->cur_comp_info[ci]->component_index;
+    coef_bit_ptr = & cinfo->coef_bits[cindex][0];
+    if (!is_DC_band && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+      WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+    for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+      int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+      if (cinfo->Ah != expected)
+	WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+      coef_bit_ptr[coefi] = cinfo->Al;
+    }
+  }
+
+  /* Select MCU decoding routine */
+  if (cinfo->Ah == 0) {
+    if (is_DC_band)
+      entropy->pub.decode_mcu = decode_mcu_DC_first;
+    else
+      entropy->pub.decode_mcu = decode_mcu_AC_first;
+  } else {
+    if (is_DC_band)
+      entropy->pub.decode_mcu = decode_mcu_DC_refine;
+    else
+      entropy->pub.decode_mcu = decode_mcu_AC_refine;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Make sure requested tables are present, and compute derived tables.
+     * We may build same derived table more than once, but it's not expensive.
+     */
+    if (is_DC_band) {
+      if (cinfo->Ah == 0) {	/* DC refinement needs no table */
+	tbl = compptr->dc_tbl_no;
+	jpeg_make_d_derived_tbl(cinfo, TRUE, tbl,
+				& entropy->derived_tbls[tbl]);
+      }
+    } else {
+      tbl = compptr->ac_tbl_no;
+      jpeg_make_d_derived_tbl(cinfo, FALSE, tbl,
+			      & entropy->derived_tbls[tbl]);
+      /* remember the single active table */
+      entropy->ac_derived_tbl = entropy->derived_tbls[tbl];
+    }
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize private state variables */
+  entropy->saved.EOBRUN = 0;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and add will be faster than a table lookup.
+ */
+
+#ifdef AVOID_TABLES
+
+#define HUFF_EXTEND(x,s)  ((x) < (1<<((s)-1)) ? (x) + (((-1)<<(s)) + 1) : (x))
+
+#else
+
+#define HUFF_EXTEND(x,s)  ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
+
+static const int extend_test[16] =   /* entry n is 2**(n-1) */
+  { 0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
+    0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000 };
+
+static const int extend_offset[16] = /* entry n is (-1 << n) + 1 */
+  { 0, ((-1)<<1) + 1, ((-1)<<2) + 1, ((-1)<<3) + 1, ((-1)<<4) + 1,
+    ((-1)<<5) + 1, ((-1)<<6) + 1, ((-1)<<7) + 1, ((-1)<<8) + 1,
+    ((-1)<<9) + 1, ((-1)<<10) + 1, ((-1)<<11) + 1, ((-1)<<12) + 1,
+    ((-1)<<13) + 1, ((-1)<<14) + 1, ((-1)<<15) + 1 };
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(boolean)
+process_restart (j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+  /* Re-init EOB run count, too */
+  entropy->saved.EOBRUN = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->pub.insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Huffman MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * Huffman-compressed coefficients. 
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ *
+ * We return FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * spectral selection, since we'll just re-assign them on the next call.
+ * Successive approximation AC refinement has to be more careful, however.)
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int Al = cinfo->Al;
+  register int s, r;
+  int blkn, ci;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  savable_state state;
+  d_derived_tbl * tbl;
+  jpeg_component_info * compptr;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(state, entropy->saved);
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      block = MCU_data[blkn];
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      tbl = entropy->derived_tbls[compptr->dc_tbl_no];
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      HUFF_DECODE(s, br_state, tbl, return FALSE, label1);
+      if (s) {
+	CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	r = GET_BITS(s);
+	s = HUFF_EXTEND(r, s);
+      }
+
+      /* Convert DC difference to actual value, update last_dc_val */
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      /* Scale and output the coefficient (assumes jpeg_natural_order[0]=0) */
+      (*block)[0] = (JCOEF) (s << Al);
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(entropy->saved, state);
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  register int s, k, r;
+  unsigned int EOBRUN;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  d_derived_tbl * tbl;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state.
+     * We can avoid loading/saving bitread state if in an EOB run.
+     */
+    EOBRUN = entropy->saved.EOBRUN;	/* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+
+    if (EOBRUN > 0)		/* if it's a band of zeroes... */
+      EOBRUN--;			/* ...process it now (we do nothing) */
+    else {
+      BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+      block = MCU_data[0];
+      tbl = entropy->ac_derived_tbl;
+
+      for (k = cinfo->Ss; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, tbl, return FALSE, label2);
+	r = s >> 4;
+	s &= 15;
+	if (s) {
+	  k += r;
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  r = GET_BITS(s);
+	  s = HUFF_EXTEND(r, s);
+	  /* Scale and output coefficient in natural (dezigzagged) order */
+	  (*block)[jpeg_natural_order[k]] = (JCOEF) (s << Al);
+	} else {
+	  if (r == 15) {	/* ZRL */
+	    k += 15;		/* skip 15 zeroes in band */
+	  } else {		/* EOBr, run length is 2^r + appended bits */
+	    EOBRUN = 1 << r;
+	    if (r) {		/* EOBr, r > 0 */
+	      CHECK_BIT_BUFFER(br_state, r, return FALSE);
+	      r = GET_BITS(r);
+	      EOBRUN += r;
+	    }
+	    EOBRUN--;		/* this band is processed at this moment */
+	    break;		/* force end-of-band */
+	  }
+	}
+      }
+
+      BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    }
+
+    /* Completed MCU, so update state */
+    entropy->saved.EOBRUN = EOBRUN;	/* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int p1 = 1 << cinfo->Al;	/* 1 in the bit position being coded */
+  int blkn;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* Not worth the cycles to check insufficient_data here,
+   * since we will not change the data anyway if we read zeroes.
+   */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    CHECK_BIT_BUFFER(br_state, 1, return FALSE);
+    if (GET_BITS(1))
+      (*block)[0] |= p1;
+    /* Note: since we use |=, repeating the assignment later is safe */
+  }
+
+  /* Completed MCU, so update state */
+  BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int Se = cinfo->Se;
+  int p1 = 1 << cinfo->Al;	/* 1 in the bit position being coded */
+  int m1 = (-1) << cinfo->Al;	/* -1 in the bit position being coded */
+  register int s, k, r;
+  unsigned int EOBRUN;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  BITREAD_STATE_VARS;
+  d_derived_tbl * tbl;
+  int num_newnz;
+  int newnz_pos[DCTSIZE2];
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, don't modify the MCU.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    EOBRUN = entropy->saved.EOBRUN; /* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+    block = MCU_data[0];
+    tbl = entropy->ac_derived_tbl;
+
+    /* If we are forced to suspend, we must undo the assignments to any newly
+     * nonzero coefficients in the block, because otherwise we'd get confused
+     * next time about which coefficients were already nonzero.
+     * But we need not undo addition of bits to already-nonzero coefficients;
+     * instead, we can test the current bit to see if we already did it.
+     */
+    num_newnz = 0;
+
+    /* initialize coefficient loop counter to start of band */
+    k = cinfo->Ss;
+
+    if (EOBRUN == 0) {
+      for (; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, tbl, goto undoit, label3);
+	r = s >> 4;
+	s &= 15;
+	if (s) {
+	  if (s != 1)		/* size of new coef should always be 1 */
+	    WARNMS(cinfo, JWRN_HUFF_BAD_CODE);
+	  CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	  if (GET_BITS(1))
+	    s = p1;		/* newly nonzero coef is positive */
+	  else
+	    s = m1;		/* newly nonzero coef is negative */
+	} else {
+	  if (r != 15) {
+	    EOBRUN = 1 << r;	/* EOBr, run length is 2^r + appended bits */
+	    if (r) {
+	      CHECK_BIT_BUFFER(br_state, r, goto undoit);
+	      r = GET_BITS(r);
+	      EOBRUN += r;
+	    }
+	    break;		/* rest of block is handled by EOB logic */
+	  }
+	  /* note s = 0 for processing ZRL */
+	}
+	/* Advance over already-nonzero coefs and r still-zero coefs,
+	 * appending correction bits to the nonzeroes.  A correction bit is 1
+	 * if the absolute value of the coefficient must be increased.
+	 */
+	do {
+	  thiscoef = *block + jpeg_natural_order[k];
+	  if (*thiscoef != 0) {
+	    CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	    if (GET_BITS(1)) {
+	      if ((*thiscoef & p1) == 0) { /* do nothing if already set it */
+		if (*thiscoef >= 0)
+		  *thiscoef += p1;
+		else
+		  *thiscoef += m1;
+	      }
+	    }
+	  } else {
+	    if (--r < 0)
+	      break;		/* reached target zero coefficient */
+	  }
+	  k++;
+	} while (k <= Se);
+	if (s) {
+	  int pos = jpeg_natural_order[k];
+	  /* Output newly nonzero coefficient */
+	  (*block)[pos] = (JCOEF) s;
+	  /* Remember its position in case we have to suspend */
+	  newnz_pos[num_newnz++] = pos;
+	}
+      }
+    }
+
+    if (EOBRUN > 0) {
+      /* Scan any remaining coefficient positions after the end-of-band
+       * (the last newly nonzero coefficient, if any).  Append a correction
+       * bit to each already-nonzero coefficient.  A correction bit is 1
+       * if the absolute value of the coefficient must be increased.
+       */
+      for (; k <= Se; k++) {
+	thiscoef = *block + jpeg_natural_order[k];
+	if (*thiscoef != 0) {
+	  CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	  if (GET_BITS(1)) {
+	    if ((*thiscoef & p1) == 0) { /* do nothing if already changed it */
+	      if (*thiscoef >= 0)
+		*thiscoef += p1;
+	      else
+		*thiscoef += m1;
+	    }
+	  }
+	}
+      }
+      /* Count one block completed in EOB run */
+      EOBRUN--;
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    entropy->saved.EOBRUN = EOBRUN; /* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+
+undoit:
+  /* Re-zero any output coefficients that we made newly nonzero */
+  while (num_newnz > 0)
+    (*block)[newnz_pos[--num_newnz]] = 0;
+
+  return FALSE;
+}
+
+
+/*
+ * Module initialization routine for progressive Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_phuff_decoder (j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy;
+  int *coef_bit_ptr;
+  int ci, i;
+
+  entropy = (phuff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(phuff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass_phuff_decoder;
+
+  /* Mark derived tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->derived_tbls[i] = NULL;
+  }
+
+  /* Create progression status table */
+  cinfo->coef_bits = (int (*)[DCTSIZE2])
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				cinfo->num_components*DCTSIZE2*SIZEOF(int));
+  coef_bit_ptr = & cinfo->coef_bits[0][0];
+  for (ci = 0; ci < cinfo->num_components; ci++) 
+    for (i = 0; i < DCTSIZE2; i++)
+      *coef_bit_ptr++ = -1;
+}
+
+#endif /* D_PROGRESSIVE_SUPPORTED */
diff --git a/jdpostct.c b/jdpostct.c
new file mode 100644
index 0000000..571563d
--- /dev/null
+++ b/jdpostct.c
@@ -0,0 +1,290 @@
+/*
+ * jdpostct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the decompression postprocessing controller.
+ * This controller manages the upsampling, color conversion, and color
+ * quantization/reduction steps; specifically, it controls the buffering
+ * between upsample/color conversion and color quantization/reduction.
+ *
+ * If no color quantization/reduction is required, then this module has no
+ * work to do, and it just hands off to the upsample/color conversion code.
+ * An integrated upsample/convert/quantize process would replace this module
+ * entirely.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_post_controller pub; /* public fields */
+
+  /* Color quantization source buffer: this holds output data from
+   * the upsample/color conversion step to be passed to the quantizer.
+   * For two-pass color quantization, we need a full-image buffer;
+   * for one-pass operation, a strip buffer is sufficient.
+   */
+  jvirt_sarray_ptr whole_image;	/* virtual array, or NULL if one-pass */
+  JSAMPARRAY buffer;		/* strip buffer, or current strip of virtual */
+  JDIMENSION strip_height;	/* buffer size in rows */
+  /* for two-pass mode only: */
+  JDIMENSION starting_row;	/* row # of first row in current strip */
+  JDIMENSION next_row;		/* index of next row to fill/empty in strip */
+} my_post_controller;
+
+typedef my_post_controller * my_post_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) post_process_1pass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) post_process_prepass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+METHODDEF(void) post_process_2pass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_dpost (j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->quantize_colors) {
+      /* Single-pass processing with color quantization. */
+      post->pub.post_process_data = post_process_1pass;
+      /* We could be doing buffered-image output before starting a 2-pass
+       * color quantization; in that case, jinit_d_post_controller did not
+       * allocate a strip buffer.  Use the virtual-array buffer as workspace.
+       */
+      if (post->buffer == NULL) {
+	post->buffer = (*cinfo->mem->access_virt_sarray)
+	  ((j_common_ptr) cinfo, post->whole_image,
+	   (JDIMENSION) 0, post->strip_height, TRUE);
+      }
+    } else {
+      /* For single-pass processing without color quantization,
+       * I have no work to do; just call the upsampler directly.
+       */
+      post->pub.post_process_data = cinfo->upsample->upsample;
+    }
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    /* First pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub.post_process_data = post_process_prepass;
+    break;
+  case JBUF_CRANK_DEST:
+    /* Second pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub.post_process_data = post_process_2pass;
+    break;
+#endif /* QUANT_2PASS_SUPPORTED */
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+  post->starting_row = post->next_row = 0;
+}
+
+
+/*
+ * Process some data in the one-pass (strip buffer) case.
+ * This is used for color precision reduction as well as one-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_1pass (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Fill the buffer, but not more than what we can dump out in one go. */
+  /* Note we rely on the upsampler to detect bottom of image. */
+  max_rows = out_rows_avail - *out_row_ctr;
+  if (max_rows > post->strip_height)
+    max_rows = post->strip_height;
+  num_rows = 0;
+  (*cinfo->upsample->upsample) (cinfo,
+		input_buf, in_row_group_ctr, in_row_groups_avail,
+		post->buffer, &num_rows, max_rows);
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->color_quantize) (cinfo,
+		post->buffer, output_buf + *out_row_ctr, (int) num_rows);
+  *out_row_ctr += num_rows;
+}
+
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+/*
+ * Process some data in the first pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_prepass (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		      JDIMENSION in_row_groups_avail,
+		      JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		      JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION old_next_row, num_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (*cinfo->mem->access_virt_sarray)
+	((j_common_ptr) cinfo, post->whole_image,
+	 post->starting_row, post->strip_height, TRUE);
+  }
+
+  /* Upsample some data (up to a strip height's worth). */
+  old_next_row = post->next_row;
+  (*cinfo->upsample->upsample) (cinfo,
+		input_buf, in_row_group_ctr, in_row_groups_avail,
+		post->buffer, &post->next_row, post->strip_height);
+
+  /* Allow quantizer to scan new data.  No data is emitted, */
+  /* but we advance out_row_ctr so outer loop can tell when we're done. */
+  if (post->next_row > old_next_row) {
+    num_rows = post->next_row - old_next_row;
+    (*cinfo->cquantize->color_quantize) (cinfo, post->buffer + old_next_row,
+					 (JSAMPARRAY) NULL, (int) num_rows);
+    *out_row_ctr += num_rows;
+  }
+
+  /* Advance if we filled the strip. */
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+
+/*
+ * Process some data in the second pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_2pass (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (*cinfo->mem->access_virt_sarray)
+	((j_common_ptr) cinfo, post->whole_image,
+	 post->starting_row, post->strip_height, FALSE);
+  }
+
+  /* Determine number of rows to emit. */
+  num_rows = post->strip_height - post->next_row; /* available in strip */
+  max_rows = out_rows_avail - *out_row_ctr; /* available in output area */
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+  /* We have to check bottom of image here, can't depend on upsampler. */
+  max_rows = cinfo->output_height - post->starting_row;
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->color_quantize) (cinfo,
+		post->buffer + post->next_row, output_buf + *out_row_ctr,
+		(int) num_rows);
+  *out_row_ctr += num_rows;
+
+  /* Advance if we filled the strip. */
+  post->next_row += num_rows;
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize postprocessing controller.
+ */
+
+GLOBAL(void)
+jinit_d_post_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_post_ptr post;
+
+  post = (my_post_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_post_controller));
+  cinfo->post = (struct jpeg_d_post_controller *) post;
+  post->pub.start_pass = start_pass_dpost;
+  post->whole_image = NULL;	/* flag for no virtual arrays */
+  post->buffer = NULL;		/* flag for no strip buffer */
+
+  /* Create the quantization buffer, if needed */
+  if (cinfo->quantize_colors) {
+    /* The buffer strip height is max_v_samp_factor, which is typically
+     * an efficient number of rows for upsampling to return.
+     * (In the presence of output rescaling, we might want to be smarter?)
+     */
+    post->strip_height = (JDIMENSION) cinfo->max_v_samp_factor;
+    if (need_full_buffer) {
+      /* Two-pass color quantization: need full-image storage. */
+      /* We round up the number of rows to a multiple of the strip height. */
+#ifdef QUANT_2PASS_SUPPORTED
+      post->whole_image = (*cinfo->mem->request_virt_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 cinfo->output_width * cinfo->out_color_components,
+	 (JDIMENSION) jround_up((long) cinfo->output_height,
+				(long) post->strip_height),
+	 post->strip_height);
+#else
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif /* QUANT_2PASS_SUPPORTED */
+    } else {
+      /* One-pass color quantization: just make a strip buffer. */
+      post->buffer = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 cinfo->output_width * cinfo->out_color_components,
+	 post->strip_height);
+    }
+  }
+}
diff --git a/jdsample.c b/jdsample.c
new file mode 100644
index 0000000..1864dd6
--- /dev/null
+++ b/jdsample.c
@@ -0,0 +1,496 @@
+/*
+ * jdsample.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains upsampling routines.
+ *
+ * Upsampling input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  Upsampling will normally produce
+ * max_v_samp_factor pixel rows from each row group (but this could vary
+ * if the upsampler is applying a scale factor of its own).
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+#include "jpegcomp.h"
+
+
+/* Pointer to routine to upsample a single component */
+typedef JMETHOD(void, upsample1_ptr,
+		(j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;	/* public fields */
+
+  /* Color conversion buffer.  When using separate upsampling and color
+   * conversion steps, this buffer holds one upsampled row group until it
+   * has been color converted and output.
+   * Note: we do not allocate any storage for component(s) which are full-size,
+   * ie do not need rescaling.  The corresponding entry of color_buf[] is
+   * simply set to point to the input data array, thereby avoiding copying.
+   */
+  JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  /* Per-component upsampling method pointers */
+  upsample1_ptr methods[MAX_COMPONENTS];
+
+  int next_row_out;		/* counts rows emitted from color_buf */
+  JDIMENSION rows_to_go;	/* counts rows remaining in image */
+
+  /* Height of an input row group for each component. */
+  int rowgroup_height[MAX_COMPONENTS];
+
+  /* These arrays save pixel expansion factors so that int_expand need not
+   * recompute them each time.  They are unused for other upsampling methods.
+   */
+  UINT8 h_expand[MAX_COMPONENTS];
+  UINT8 v_expand[MAX_COMPONENTS];
+} my_upsampler;
+
+typedef my_upsampler * my_upsample_ptr;
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_upsample (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Mark the conversion buffer empty */
+  upsample->next_row_out = cinfo->max_v_samp_factor;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * In this version we upsample each component independently.
+ * We upsample one row group into the conversion buffer, then apply
+ * color conversion a row at a time.
+ */
+
+METHODDEF(void)
+sep_upsample (j_decompress_ptr cinfo,
+	      JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	      JDIMENSION in_row_groups_avail,
+	      JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	      JDIMENSION out_rows_avail)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int ci;
+  jpeg_component_info * compptr;
+  JDIMENSION num_rows;
+
+  /* Fill the conversion buffer, if it's empty */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor) {
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      /* Invoke per-component upsample method.  Notice we pass a POINTER
+       * to color_buf[ci], so that fullsize_upsample can change it.
+       */
+      (*upsample->methods[ci]) (cinfo, compptr,
+	input_buf[ci] + (*in_row_group_ctr * upsample->rowgroup_height[ci]),
+	upsample->color_buf + ci);
+    }
+    upsample->next_row_out = 0;
+  }
+
+  /* Color-convert and emit rows */
+
+  /* How many we have in the buffer: */
+  num_rows = (JDIMENSION) (cinfo->max_v_samp_factor - upsample->next_row_out);
+  /* Not more than the distance to the end of the image.  Need this test
+   * in case the image height is not a multiple of max_v_samp_factor:
+   */
+  if (num_rows > upsample->rows_to_go) 
+    num_rows = upsample->rows_to_go;
+  /* And not more than what the client can accept: */
+  out_rows_avail -= *out_row_ctr;
+  if (num_rows > out_rows_avail)
+    num_rows = out_rows_avail;
+
+  (*cinfo->cconvert->color_convert) (cinfo, upsample->color_buf,
+				     (JDIMENSION) upsample->next_row_out,
+				     output_buf + *out_row_ctr,
+				     (int) num_rows);
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  upsample->next_row_out += num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor)
+    (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by sep_upsample to upsample pixel values
+ * of a single component.  One row group is processed per call.
+ */
+
+
+/*
+ * For full-size components, we just make color_buf[ci] point at the
+ * input buffer, and thus avoid copying any data.  Note that this is
+ * safe only because sep_upsample doesn't declare the input row group
+ * "consumed" until we are done color converting and emitting it.
+ */
+
+METHODDEF(void)
+fullsize_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		   JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  *output_data_ptr = input_data;
+}
+
+
+/*
+ * This is a no-op version used for "uninteresting" components.
+ * These components will not be referenced by color conversion.
+ */
+
+METHODDEF(void)
+noop_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  *output_data_ptr = NULL;	/* safety check */
+}
+
+
+/*
+ * This version handles any integral sampling ratios.
+ * This is not used for typical JPEG files, so it need not be fast.
+ * Nor, for that matter, is it particularly accurate: the algorithm is
+ * simple replication of the input pixel onto the corresponding output
+ * pixels.  The hi-falutin sampling literature refers to this as a
+ * "box filter".  A box filter tends to introduce visible artifacts,
+ * so if you are actually going to use 3:1 or 4:1 sampling ratios
+ * you would be well advised to improve this code.
+ */
+
+METHODDEF(void)
+int_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	      JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register JSAMPLE invalue;
+  register int h;
+  JSAMPROW outend;
+  int h_expand, v_expand;
+  int inrow, outrow;
+
+  h_expand = upsample->h_expand[compptr->component_index];
+  v_expand = upsample->v_expand[compptr->component_index];
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    /* Generate one output row with proper horizontal expansion */
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      for (h = h_expand; h > 0; h--) {
+	*outptr++ = invalue;
+      }
+    }
+    /* Generate any additional output rows by duplicating the first one */
+    if (v_expand > 1) {
+      jcopy_sample_rows(output_data, outrow, output_data, outrow+1,
+			v_expand-1, cinfo->output_width);
+    }
+    inrow++;
+    outrow += v_expand;
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v1_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register JSAMPLE invalue;
+  JSAMPROW outend;
+  int inrow;
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v2_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register JSAMPLE invalue;
+  JSAMPROW outend;
+  int inrow, outrow;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+    jcopy_sample_rows(output_data, outrow, output_data, outrow+1,
+		      1, cinfo->output_width);
+    inrow++;
+    outrow += 2;
+  }
+}
+
+
+/*
+ * Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+ *
+ * The upsampling algorithm is linear interpolation between pixel centers,
+ * also known as a "triangle filter".  This is a good compromise between
+ * speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+ * of the way between input pixel centers.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_fancy_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register int invalue;
+  register JDIMENSION colctr;
+  int inrow;
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    /* Special case for first column */
+    invalue = GETJSAMPLE(*inptr++);
+    *outptr++ = (JSAMPLE) invalue;
+    *outptr++ = (JSAMPLE) ((invalue * 3 + GETJSAMPLE(*inptr) + 2) >> 2);
+
+    for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
+      /* General case: 3/4 * nearer pixel + 1/4 * further pixel */
+      invalue = GETJSAMPLE(*inptr++) * 3;
+      *outptr++ = (JSAMPLE) ((invalue + GETJSAMPLE(inptr[-2]) + 1) >> 2);
+      *outptr++ = (JSAMPLE) ((invalue + GETJSAMPLE(*inptr) + 2) >> 2);
+    }
+
+    /* Special case for last column */
+    invalue = GETJSAMPLE(*inptr);
+    *outptr++ = (JSAMPLE) ((invalue * 3 + GETJSAMPLE(inptr[-1]) + 1) >> 2);
+    *outptr++ = (JSAMPLE) invalue;
+  }
+}
+
+
+/*
+ * Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * Again a triangle filter; see comments for h2v1 case, above.
+ *
+ * It is OK for us to reference the adjacent input rows because we demanded
+ * context from the main buffer controller (see initialization code).
+ */
+
+METHODDEF(void)
+h2v2_fancy_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr0, inptr1, outptr;
+#if BITS_IN_JSAMPLE == 8
+  register int thiscolsum, lastcolsum, nextcolsum;
+#else
+  register INT32 thiscolsum, lastcolsum, nextcolsum;
+#endif
+  register JDIMENSION colctr;
+  int inrow, outrow, v;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    for (v = 0; v < 2; v++) {
+      /* inptr0 points to nearest input row, inptr1 points to next nearest */
+      inptr0 = input_data[inrow];
+      if (v == 0)		/* next nearest is row above */
+	inptr1 = input_data[inrow-1];
+      else			/* next nearest is row below */
+	inptr1 = input_data[inrow+1];
+      outptr = output_data[outrow++];
+
+      /* Special case for first column */
+      thiscolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+      nextcolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+      *outptr++ = (JSAMPLE) ((thiscolsum * 4 + 8) >> 4);
+      *outptr++ = (JSAMPLE) ((thiscolsum * 3 + nextcolsum + 7) >> 4);
+      lastcolsum = thiscolsum; thiscolsum = nextcolsum;
+
+      for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
+	/* General case: 3/4 * nearer pixel + 1/4 * further pixel in each */
+	/* dimension, thus 9/16, 3/16, 3/16, 1/16 overall */
+	nextcolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+	*outptr++ = (JSAMPLE) ((thiscolsum * 3 + lastcolsum + 8) >> 4);
+	*outptr++ = (JSAMPLE) ((thiscolsum * 3 + nextcolsum + 7) >> 4);
+	lastcolsum = thiscolsum; thiscolsum = nextcolsum;
+      }
+
+      /* Special case for last column */
+      *outptr++ = (JSAMPLE) ((thiscolsum * 3 + lastcolsum + 8) >> 4);
+      *outptr++ = (JSAMPLE) ((thiscolsum * 4 + 7) >> 4);
+    }
+    inrow++;
+  }
+}
+
+
+/*
+ * Module initialization routine for upsampling.
+ */
+
+GLOBAL(void)
+jinit_upsampler (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+  int ci;
+  jpeg_component_info * compptr;
+  boolean need_buffer, do_fancy;
+  int h_in_group, v_in_group, h_out_group, v_out_group;
+
+  upsample = (my_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *) upsample;
+  upsample->pub.start_pass = start_pass_upsample;
+  upsample->pub.upsample = sep_upsample;
+  upsample->pub.need_context_rows = FALSE; /* until we find out differently */
+
+  if (cinfo->CCIR601_sampling)	/* this isn't supported */
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* jdmainct.c doesn't support context rows when min_DCT_scaled_size = 1,
+   * so don't ask for it.
+   */
+  do_fancy = cinfo->do_fancy_upsampling && cinfo->_min_DCT_scaled_size > 1;
+
+  /* Verify we can handle the sampling factors, select per-component methods,
+   * and create storage as needed.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Compute size of an "input group" after IDCT scaling.  This many samples
+     * are to be converted to max_h_samp_factor * max_v_samp_factor pixels.
+     */
+    h_in_group = (compptr->h_samp_factor * compptr->_DCT_scaled_size) /
+		 cinfo->_min_DCT_scaled_size;
+    v_in_group = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
+		 cinfo->_min_DCT_scaled_size;
+    h_out_group = cinfo->max_h_samp_factor;
+    v_out_group = cinfo->max_v_samp_factor;
+    upsample->rowgroup_height[ci] = v_in_group; /* save for use later */
+    need_buffer = TRUE;
+    if (! compptr->component_needed) {
+      /* Don't bother to upsample an uninteresting component. */
+      upsample->methods[ci] = noop_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group == h_out_group && v_in_group == v_out_group) {
+      /* Fullsize components can be processed without any work. */
+      upsample->methods[ci] = fullsize_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group * 2 == h_out_group &&
+	       v_in_group == v_out_group) {
+      /* Special cases for 2h1v upsampling */
+      if (do_fancy && compptr->downsampled_width > 2) {
+	if (jsimd_can_h2v1_fancy_upsample())
+	  upsample->methods[ci] = jsimd_h2v1_fancy_upsample;
+	else
+	  upsample->methods[ci] = h2v1_fancy_upsample;
+      } else {
+	if (jsimd_can_h2v1_upsample())
+	  upsample->methods[ci] = jsimd_h2v1_upsample;
+	else
+	  upsample->methods[ci] = h2v1_upsample;
+      }
+    } else if (h_in_group * 2 == h_out_group &&
+	       v_in_group * 2 == v_out_group) {
+      /* Special cases for 2h2v upsampling */
+      if (do_fancy && compptr->downsampled_width > 2) {
+	if (jsimd_can_h2v2_fancy_upsample())
+	  upsample->methods[ci] = jsimd_h2v2_fancy_upsample;
+	else
+	  upsample->methods[ci] = h2v2_fancy_upsample;
+	upsample->pub.need_context_rows = TRUE;
+      } else {
+	if (jsimd_can_h2v2_upsample())
+	  upsample->methods[ci] = jsimd_h2v2_upsample;
+	else
+	  upsample->methods[ci] = h2v2_upsample;
+      }
+    } else if ((h_out_group % h_in_group) == 0 &&
+	       (v_out_group % v_in_group) == 0) {
+      /* Generic integral-factors upsampling method */
+      upsample->methods[ci] = int_upsample;
+      upsample->h_expand[ci] = (UINT8) (h_out_group / h_in_group);
+      upsample->v_expand[ci] = (UINT8) (v_out_group / v_in_group);
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+    if (need_buffer) {
+      upsample->color_buf[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 (JDIMENSION) jround_up((long) cinfo->output_width,
+				(long) cinfo->max_h_samp_factor),
+	 (JDIMENSION) cinfo->max_v_samp_factor);
+    }
+  }
+}
diff --git a/jdtrans.c b/jdtrans.c
new file mode 100644
index 0000000..6c0ab71
--- /dev/null
+++ b/jdtrans.c
@@ -0,0 +1,143 @@
+/*
+ * jdtrans.c
+ *
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains library routines for transcoding decompression,
+ * that is, reading raw DCT coefficient arrays from an input JPEG file.
+ * The routines in jdapimin.c will also be needed by a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Forward declarations */
+LOCAL(void) transdecode_master_selection JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Read the coefficient arrays from a JPEG file.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * The entire image is read into a set of virtual coefficient-block arrays,
+ * one per component.  The return value is a pointer to the array of
+ * virtual-array descriptors.  These can be manipulated directly via the
+ * JPEG memory manager, or handed off to jpeg_write_coefficients().
+ * To release the memory occupied by the virtual arrays, call
+ * jpeg_finish_decompress() when done with the data.
+ *
+ * An alternative usage is to simply obtain access to the coefficient arrays
+ * during a buffered-image-mode decompression operation.  This is allowed
+ * after any jpeg_finish_output() call.  The arrays can be accessed until
+ * jpeg_finish_decompress() is called.  (Note that any call to the library
+ * may reposition the arrays, so don't rely on access_virt_barray() results
+ * to stay valid across library calls.)
+ *
+ * Returns NULL if suspended.  This case need be checked only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(jvirt_barray_ptr *)
+jpeg_read_coefficients (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize active modules */
+    transdecode_master_selection(cinfo);
+    cinfo->global_state = DSTATE_RDCOEFS;
+  }
+  if (cinfo->global_state == DSTATE_RDCOEFS) {
+    /* Absorb whole file into the coef buffer */
+    for (;;) {
+      int retcode;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL)
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      /* Absorb some more input */
+      retcode = (*cinfo->inputctl->consume_input) (cinfo);
+      if (retcode == JPEG_SUSPENDED)
+	return NULL;
+      if (retcode == JPEG_REACHED_EOI)
+	break;
+      /* Advance progress counter if appropriate */
+      if (cinfo->progress != NULL &&
+	  (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+	if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+	  /* startup underestimated number of scans; ratchet up one scan */
+	  cinfo->progress->pass_limit += (long) cinfo->total_iMCU_rows;
+	}
+      }
+    }
+    /* Set state so that jpeg_finish_decompress does the right thing */
+    cinfo->global_state = DSTATE_STOPPING;
+  }
+  /* At this point we should be in state DSTATE_STOPPING if being used
+   * standalone, or in state DSTATE_BUFIMAGE if being invoked to get access
+   * to the coefficients during a full buffered-image-mode decompression.
+   */
+  if ((cinfo->global_state == DSTATE_STOPPING ||
+       cinfo->global_state == DSTATE_BUFIMAGE) && cinfo->buffered_image) {
+    return cinfo->coef->coef_arrays;
+  }
+  /* Oops, improper usage */
+  ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return NULL;			/* keep compiler happy */
+}
+
+
+/*
+ * Master selection of decompression modules for transcoding.
+ * This substitutes for jdmaster.c's initialization of the full decompressor.
+ */
+
+LOCAL(void)
+transdecode_master_selection (j_decompress_ptr cinfo)
+{
+  /* This is effectively a buffered-image operation. */
+  cinfo->buffered_image = TRUE;
+
+  /* Entropy decoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef D_PROGRESSIVE_SUPPORTED
+      jinit_phuff_decoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_decoder(cinfo);
+  }
+
+  /* Always get a full-image coefficient buffer. */
+  jinit_d_coef_controller(cinfo, TRUE);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+  /* Initialize progress monitoring. */
+  if (cinfo->progress != NULL) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else if (cinfo->inputctl->has_multiple_scans) {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    } else {
+      nscans = 1;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = 1;
+  }
+}
diff --git a/jerror.c b/jerror.c
new file mode 100644
index 0000000..3da7be8
--- /dev/null
+++ b/jerror.c
@@ -0,0 +1,252 @@
+/*
+ * jerror.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains simple error-reporting and trace-message routines.
+ * These are suitable for Unix-like systems and others where writing to
+ * stderr is the right thing to do.  Many applications will want to replace
+ * some or all of these routines.
+ *
+ * If you define USE_WINDOWS_MESSAGEBOX in jconfig.h or in the makefile,
+ * you get a Windows-specific hack to display error messages in a dialog box.
+ * It ain't much, but it beats dropping error messages into the bit bucket,
+ * which is what happens to output to stderr under most Windows C compilers.
+ *
+ * These routines are used by both the compression and decompression code.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jversion.h"
+#include "jerror.h"
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+#include <windows.h>
+#endif
+
+#ifndef EXIT_FAILURE		/* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+
+
+/*
+ * Create the message string table.
+ * We do this from the master message list in jerror.h by re-reading
+ * jerror.h with a suitable definition for macro JMESSAGE.
+ * The message table is made an external symbol just in case any applications
+ * want to refer to it directly.
+ */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_std_message_table	jMsgTable
+#endif
+
+#define JMESSAGE(code,string)	string ,
+
+const char * const jpeg_std_message_table[] = {
+#include "jerror.h"
+  NULL
+};
+
+
+/*
+ * Error exit handler: must not return to caller.
+ *
+ * Applications may override this if they want to get control back after
+ * an error.  Typically one would longjmp somewhere instead of exiting.
+ * The setjmp buffer can be made a private field within an expanded error
+ * handler object.  Note that the info needed to generate an error message
+ * is stored in the error object, so you can generate the message now or
+ * later, at your convenience.
+ * You should make sure that the JPEG object is cleaned up (with jpeg_abort
+ * or jpeg_destroy) at some point.
+ */
+
+METHODDEF(void)
+error_exit (j_common_ptr cinfo)
+{
+  /* Always display the message */
+  (*cinfo->err->output_message) (cinfo);
+
+  /* Let the memory manager delete any temp files before we die */
+  jpeg_destroy(cinfo);
+
+  exit(EXIT_FAILURE);
+}
+
+
+/*
+ * Actual output of an error or trace message.
+ * Applications may override this method to send JPEG messages somewhere
+ * other than stderr.
+ *
+ * On Windows, printing to stderr is generally completely useless,
+ * so we provide optional code to produce an error-dialog popup.
+ * Most Windows applications will still prefer to override this routine,
+ * but if they don't, it'll do something at least marginally useful.
+ *
+ * NOTE: to use the library in an environment that doesn't support the
+ * C stdio library, you may have to delete the call to fprintf() entirely,
+ * not just not use this routine.
+ */
+
+METHODDEF(void)
+output_message (j_common_ptr cinfo)
+{
+  char buffer[JMSG_LENGTH_MAX];
+
+  /* Create the message */
+  (*cinfo->err->format_message) (cinfo, buffer);
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+  /* Display it in a message dialog box */
+  MessageBox(GetActiveWindow(), buffer, "JPEG Library Error",
+	     MB_OK | MB_ICONERROR);
+#else
+  /* Send it to stderr, adding a newline */
+  fprintf(stderr, "%s\n", buffer);
+#endif
+}
+
+
+/*
+ * Decide whether to emit a trace or warning message.
+ * msg_level is one of:
+ *   -1: recoverable corrupt-data warning, may want to abort.
+ *    0: important advisory messages (always display to user).
+ *    1: first level of tracing detail.
+ *    2,3,...: successively more detailed tracing messages.
+ * An application might override this method if it wanted to abort on warnings
+ * or change the policy about which messages to display.
+ */
+
+METHODDEF(void)
+emit_message (j_common_ptr cinfo, int msg_level)
+{
+  struct jpeg_error_mgr * err = cinfo->err;
+
+  if (msg_level < 0) {
+    /* It's a warning message.  Since corrupt files may generate many warnings,
+     * the policy implemented here is to show only the first warning,
+     * unless trace_level >= 3.
+     */
+    if (err->num_warnings == 0 || err->trace_level >= 3)
+      (*err->output_message) (cinfo);
+    /* Always count warnings in num_warnings. */
+    err->num_warnings++;
+  } else {
+    /* It's a trace message.  Show it if trace_level >= msg_level. */
+    if (err->trace_level >= msg_level)
+      (*err->output_message) (cinfo);
+  }
+}
+
+
+/*
+ * Format a message string for the most recent JPEG error or message.
+ * The message is stored into buffer, which should be at least JMSG_LENGTH_MAX
+ * characters.  Note that no '\n' character is added to the string.
+ * Few applications should need to override this method.
+ */
+
+METHODDEF(void)
+format_message (j_common_ptr cinfo, char * buffer)
+{
+  struct jpeg_error_mgr * err = cinfo->err;
+  int msg_code = err->msg_code;
+  const char * msgtext = NULL;
+  const char * msgptr;
+  char ch;
+  boolean isstring;
+
+  /* Look up message string in proper table */
+  if (msg_code > 0 && msg_code <= err->last_jpeg_message) {
+    msgtext = err->jpeg_message_table[msg_code];
+  } else if (err->addon_message_table != NULL &&
+	     msg_code >= err->first_addon_message &&
+	     msg_code <= err->last_addon_message) {
+    msgtext = err->addon_message_table[msg_code - err->first_addon_message];
+  }
+
+  /* Defend against bogus message number */
+  if (msgtext == NULL) {
+    err->msg_parm.i[0] = msg_code;
+    msgtext = err->jpeg_message_table[0];
+  }
+
+  /* Check for string parameter, as indicated by %s in the message text */
+  isstring = FALSE;
+  msgptr = msgtext;
+  while ((ch = *msgptr++) != '\0') {
+    if (ch == '%') {
+      if (*msgptr == 's') isstring = TRUE;
+      break;
+    }
+  }
+
+  /* Format the message into the passed buffer */
+  if (isstring)
+    sprintf(buffer, msgtext, err->msg_parm.s);
+  else
+    sprintf(buffer, msgtext,
+	    err->msg_parm.i[0], err->msg_parm.i[1],
+	    err->msg_parm.i[2], err->msg_parm.i[3],
+	    err->msg_parm.i[4], err->msg_parm.i[5],
+	    err->msg_parm.i[6], err->msg_parm.i[7]);
+}
+
+
+/*
+ * Reset error state variables at start of a new image.
+ * This is called during compression startup to reset trace/error
+ * processing to default state, without losing any application-specific
+ * method pointers.  An application might possibly want to override
+ * this method if it has additional error processing state.
+ */
+
+METHODDEF(void)
+reset_error_mgr (j_common_ptr cinfo)
+{
+  cinfo->err->num_warnings = 0;
+  /* trace_level is not reset since it is an application-supplied parameter */
+  cinfo->err->msg_code = 0;	/* may be useful as a flag for "no error" */
+}
+
+
+/*
+ * Fill in the standard error-handling methods in a jpeg_error_mgr object.
+ * Typical call is:
+ *	struct jpeg_compress_struct cinfo;
+ *	struct jpeg_error_mgr err;
+ *
+ *	cinfo.err = jpeg_std_error(&err);
+ * after which the application may override some of the methods.
+ */
+
+GLOBAL(struct jpeg_error_mgr *)
+jpeg_std_error (struct jpeg_error_mgr * err)
+{
+  err->error_exit = error_exit;
+  err->emit_message = emit_message;
+  err->output_message = output_message;
+  err->format_message = format_message;
+  err->reset_error_mgr = reset_error_mgr;
+
+  err->trace_level = 0;		/* default = no tracing */
+  err->num_warnings = 0;	/* no warnings emitted yet */
+  err->msg_code = 0;		/* may be useful as a flag for "no error" */
+
+  /* Initialize message table pointers */
+  err->jpeg_message_table = jpeg_std_message_table;
+  err->last_jpeg_message = (int) JMSG_LASTMSGCODE - 1;
+
+  err->addon_message_table = NULL;
+  err->first_addon_message = 0;	/* for safety */
+  err->last_addon_message = 0;
+
+  return err;
+}
diff --git a/jerror.h b/jerror.h
new file mode 100644
index 0000000..edbe903
--- /dev/null
+++ b/jerror.h
@@ -0,0 +1,293 @@
+/*
+ * jerror.h
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 1997-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file defines the error and message codes for the JPEG library.
+ * Edit this file to add new codes, or to translate the message strings to
+ * some other language.
+ * A set of error-reporting macros are defined too.  Some applications using
+ * the JPEG library may wish to include this file to get the error codes
+ * and/or the macros.
+ */
+
+/*
+ * To define the enum list of message codes, include this file without
+ * defining macro JMESSAGE.  To create a message string table, include it
+ * again with a suitable JMESSAGE definition (see jerror.c for an example).
+ */
+#ifndef JMESSAGE
+#ifndef JERROR_H
+/* First time through, define the enum list */
+#define JMAKE_ENUM_LIST
+#else
+/* Repeated inclusions of this file are no-ops unless JMESSAGE is defined */
+#define JMESSAGE(code,string)
+#endif /* JERROR_H */
+#endif /* JMESSAGE */
+
+#ifdef JMAKE_ENUM_LIST
+
+typedef enum {
+
+#define JMESSAGE(code,string)	code ,
+
+#endif /* JMAKE_ENUM_LIST */
+
+JMESSAGE(JMSG_NOMESSAGE, "Bogus message code %d") /* Must be first entry! */
+
+/* For maintenance convenience, list is alphabetical by message code name */
+JMESSAGE(JERR_ARITH_NOTIMPL,
+	 "Sorry, there are legal restrictions on arithmetic coding")
+JMESSAGE(JERR_BAD_ALIGN_TYPE, "ALIGN_TYPE is wrong, please fix")
+JMESSAGE(JERR_BAD_ALLOC_CHUNK, "MAX_ALLOC_CHUNK is wrong, please fix")
+JMESSAGE(JERR_BAD_BUFFER_MODE, "Bogus buffer control mode")
+JMESSAGE(JERR_BAD_COMPONENT_ID, "Invalid component ID %d in SOS")
+JMESSAGE(JERR_BAD_CROP_SPEC, "Invalid crop request")
+JMESSAGE(JERR_BAD_DCT_COEF, "DCT coefficient out of range")
+JMESSAGE(JERR_BAD_DCTSIZE, "IDCT output block size %d not supported")
+JMESSAGE(JERR_BAD_HUFF_TABLE, "Bogus Huffman table definition")
+JMESSAGE(JERR_BAD_IN_COLORSPACE, "Bogus input colorspace")
+JMESSAGE(JERR_BAD_J_COLORSPACE, "Bogus JPEG colorspace")
+JMESSAGE(JERR_BAD_LENGTH, "Bogus marker length")
+JMESSAGE(JERR_BAD_LIB_VERSION,
+	 "Wrong JPEG library version: library is %d, caller expects %d")
+JMESSAGE(JERR_BAD_MCU_SIZE, "Sampling factors too large for interleaved scan")
+JMESSAGE(JERR_BAD_POOL_ID, "Invalid memory pool code %d")
+JMESSAGE(JERR_BAD_PRECISION, "Unsupported JPEG data precision %d")
+JMESSAGE(JERR_BAD_PROGRESSION,
+	 "Invalid progressive parameters Ss=%d Se=%d Ah=%d Al=%d")
+JMESSAGE(JERR_BAD_PROG_SCRIPT,
+	 "Invalid progressive parameters at scan script entry %d")
+JMESSAGE(JERR_BAD_SAMPLING, "Bogus sampling factors")
+JMESSAGE(JERR_BAD_SCAN_SCRIPT, "Invalid scan script at entry %d")
+JMESSAGE(JERR_BAD_STATE, "Improper call to JPEG library in state %d")
+JMESSAGE(JERR_BAD_STRUCT_SIZE,
+	 "JPEG parameter struct mismatch: library thinks size is %u, caller expects %u")
+JMESSAGE(JERR_BAD_VIRTUAL_ACCESS, "Bogus virtual array access")
+JMESSAGE(JERR_BUFFER_SIZE, "Buffer passed to JPEG library is too small")
+JMESSAGE(JERR_CANT_SUSPEND, "Suspension not allowed here")
+JMESSAGE(JERR_CCIR601_NOTIMPL, "CCIR601 sampling not implemented yet")
+JMESSAGE(JERR_COMPONENT_COUNT, "Too many color components: %d, max %d")
+JMESSAGE(JERR_CONVERSION_NOTIMPL, "Unsupported color conversion request")
+JMESSAGE(JERR_DAC_INDEX, "Bogus DAC index %d")
+JMESSAGE(JERR_DAC_VALUE, "Bogus DAC value 0x%x")
+JMESSAGE(JERR_DHT_INDEX, "Bogus DHT index %d")
+JMESSAGE(JERR_DQT_INDEX, "Bogus DQT index %d")
+JMESSAGE(JERR_EMPTY_IMAGE, "Empty JPEG image (DNL not supported)")
+JMESSAGE(JERR_EMS_READ, "Read from EMS failed")
+JMESSAGE(JERR_EMS_WRITE, "Write to EMS failed")
+JMESSAGE(JERR_EOI_EXPECTED, "Didn't expect more than one scan")
+JMESSAGE(JERR_FILE_READ, "Input file read error")
+JMESSAGE(JERR_FILE_WRITE, "Output file write error --- out of disk space?")
+JMESSAGE(JERR_FRACT_SAMPLE_NOTIMPL, "Fractional sampling not implemented yet")
+JMESSAGE(JERR_HUFF_CLEN_OVERFLOW, "Huffman code size table overflow")
+JMESSAGE(JERR_HUFF_MISSING_CODE, "Missing Huffman code table entry")
+JMESSAGE(JERR_IMAGE_TOO_BIG, "Maximum supported image dimension is %u pixels")
+JMESSAGE(JERR_INPUT_EMPTY, "Empty input file")
+JMESSAGE(JERR_INPUT_EOF, "Premature end of input file")
+JMESSAGE(JERR_MISMATCHED_QUANT_TABLE,
+	 "Cannot transcode due to multiple use of quantization table %d")
+JMESSAGE(JERR_MISSING_DATA, "Scan script does not transmit all data")
+JMESSAGE(JERR_MODE_CHANGE, "Invalid color quantization mode change")
+JMESSAGE(JERR_NOTIMPL, "Not implemented yet")
+JMESSAGE(JERR_NOT_COMPILED, "Requested feature was omitted at compile time")
+JMESSAGE(JERR_NO_BACKING_STORE, "Backing store not supported")
+JMESSAGE(JERR_NO_HUFF_TABLE, "Huffman table 0x%02x was not defined")
+JMESSAGE(JERR_NO_IMAGE, "JPEG datastream contains no image")
+JMESSAGE(JERR_NO_QUANT_TABLE, "Quantization table 0x%02x was not defined")
+JMESSAGE(JERR_NO_SOI, "Not a JPEG file: starts with 0x%02x 0x%02x")
+JMESSAGE(JERR_OUT_OF_MEMORY, "Insufficient memory (case %d)")
+JMESSAGE(JERR_QUANT_COMPONENTS,
+	 "Cannot quantize more than %d color components")
+JMESSAGE(JERR_QUANT_FEW_COLORS, "Cannot quantize to fewer than %d colors")
+JMESSAGE(JERR_QUANT_MANY_COLORS, "Cannot quantize to more than %d colors")
+JMESSAGE(JERR_SOF_DUPLICATE, "Invalid JPEG file structure: two SOF markers")
+JMESSAGE(JERR_SOF_NO_SOS, "Invalid JPEG file structure: missing SOS marker")
+JMESSAGE(JERR_SOF_UNSUPPORTED, "Unsupported JPEG process: SOF type 0x%02x")
+JMESSAGE(JERR_SOI_DUPLICATE, "Invalid JPEG file structure: two SOI markers")
+JMESSAGE(JERR_SOS_NO_SOF, "Invalid JPEG file structure: SOS before SOF")
+JMESSAGE(JERR_TFILE_CREATE, "Failed to create temporary file %s")
+JMESSAGE(JERR_TFILE_READ, "Read failed on temporary file")
+JMESSAGE(JERR_TFILE_SEEK, "Seek failed on temporary file")
+JMESSAGE(JERR_TFILE_WRITE,
+	 "Write failed on temporary file --- out of disk space?")
+JMESSAGE(JERR_TOO_LITTLE_DATA, "Application transferred too few scanlines")
+JMESSAGE(JERR_UNKNOWN_MARKER, "Unsupported marker type 0x%02x")
+JMESSAGE(JERR_VIRTUAL_BUG, "Virtual array controller messed up")
+JMESSAGE(JERR_WIDTH_OVERFLOW, "Image too wide for this implementation")
+JMESSAGE(JERR_XMS_READ, "Read from XMS failed")
+JMESSAGE(JERR_XMS_WRITE, "Write to XMS failed")
+JMESSAGE(JMSG_COPYRIGHT, JCOPYRIGHT)
+JMESSAGE(JMSG_VERSION, JVERSION)
+JMESSAGE(JTRC_16BIT_TABLES,
+	 "Caution: quantization tables are too coarse for baseline JPEG")
+JMESSAGE(JTRC_ADOBE,
+	 "Adobe APP14 marker: version %d, flags 0x%04x 0x%04x, transform %d")
+JMESSAGE(JTRC_APP0, "Unknown APP0 marker (not JFIF), length %u")
+JMESSAGE(JTRC_APP14, "Unknown APP14 marker (not Adobe), length %u")
+JMESSAGE(JTRC_DAC, "Define Arithmetic Table 0x%02x: 0x%02x")
+JMESSAGE(JTRC_DHT, "Define Huffman Table 0x%02x")
+JMESSAGE(JTRC_DQT, "Define Quantization Table %d  precision %d")
+JMESSAGE(JTRC_DRI, "Define Restart Interval %u")
+JMESSAGE(JTRC_EMS_CLOSE, "Freed EMS handle %u")
+JMESSAGE(JTRC_EMS_OPEN, "Obtained EMS handle %u")
+JMESSAGE(JTRC_EOI, "End Of Image")
+JMESSAGE(JTRC_HUFFBITS, "        %3d %3d %3d %3d %3d %3d %3d %3d")
+JMESSAGE(JTRC_JFIF, "JFIF APP0 marker: version %d.%02d, density %dx%d  %d")
+JMESSAGE(JTRC_JFIF_BADTHUMBNAILSIZE,
+	 "Warning: thumbnail image size does not match data length %u")
+JMESSAGE(JTRC_JFIF_EXTENSION,
+	 "JFIF extension marker: type 0x%02x, length %u")
+JMESSAGE(JTRC_JFIF_THUMBNAIL, "    with %d x %d thumbnail image")
+JMESSAGE(JTRC_MISC_MARKER, "Miscellaneous marker 0x%02x, length %u")
+JMESSAGE(JTRC_PARMLESS_MARKER, "Unexpected marker 0x%02x")
+JMESSAGE(JTRC_QUANTVALS, "        %4u %4u %4u %4u %4u %4u %4u %4u")
+JMESSAGE(JTRC_QUANT_3_NCOLORS, "Quantizing to %d = %d*%d*%d colors")
+JMESSAGE(JTRC_QUANT_NCOLORS, "Quantizing to %d colors")
+JMESSAGE(JTRC_QUANT_SELECTED, "Selected %d colors for quantization")
+JMESSAGE(JTRC_RECOVERY_ACTION, "At marker 0x%02x, recovery action %d")
+JMESSAGE(JTRC_RST, "RST%d")
+JMESSAGE(JTRC_SMOOTH_NOTIMPL,
+	 "Smoothing not supported with nonstandard sampling ratios")
+JMESSAGE(JTRC_SOF, "Start Of Frame 0x%02x: width=%u, height=%u, components=%d")
+JMESSAGE(JTRC_SOF_COMPONENT, "    Component %d: %dhx%dv q=%d")
+JMESSAGE(JTRC_SOI, "Start of Image")
+JMESSAGE(JTRC_SOS, "Start Of Scan: %d components")
+JMESSAGE(JTRC_SOS_COMPONENT, "    Component %d: dc=%d ac=%d")
+JMESSAGE(JTRC_SOS_PARAMS, "  Ss=%d, Se=%d, Ah=%d, Al=%d")
+JMESSAGE(JTRC_TFILE_CLOSE, "Closed temporary file %s")
+JMESSAGE(JTRC_TFILE_OPEN, "Opened temporary file %s")
+JMESSAGE(JTRC_THUMB_JPEG,
+	 "JFIF extension marker: JPEG-compressed thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_PALETTE,
+	 "JFIF extension marker: palette thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_RGB,
+	 "JFIF extension marker: RGB thumbnail image, length %u")
+JMESSAGE(JTRC_UNKNOWN_IDS,
+	 "Unrecognized component IDs %d %d %d, assuming YCbCr")
+JMESSAGE(JTRC_XMS_CLOSE, "Freed XMS handle %u")
+JMESSAGE(JTRC_XMS_OPEN, "Obtained XMS handle %u")
+JMESSAGE(JWRN_ADOBE_XFORM, "Unknown Adobe color transform code %d")
+JMESSAGE(JWRN_BOGUS_PROGRESSION,
+	 "Inconsistent progression sequence for component %d coefficient %d")
+JMESSAGE(JWRN_EXTRANEOUS_DATA,
+	 "Corrupt JPEG data: %u extraneous bytes before marker 0x%02x")
+JMESSAGE(JWRN_HIT_MARKER, "Corrupt JPEG data: premature end of data segment")
+JMESSAGE(JWRN_HUFF_BAD_CODE, "Corrupt JPEG data: bad Huffman code")
+JMESSAGE(JWRN_JFIF_MAJOR, "Warning: unknown JFIF revision number %d.%02d")
+JMESSAGE(JWRN_JPEG_EOF, "Premature end of JPEG file")
+JMESSAGE(JWRN_MUST_RESYNC,
+	 "Corrupt JPEG data: found marker 0x%02x instead of RST%d")
+JMESSAGE(JWRN_NOT_SEQUENTIAL, "Invalid SOS parameters for sequential JPEG")
+JMESSAGE(JWRN_TOO_MUCH_DATA, "Application transferred too many scanlines")
+
+#ifdef JMAKE_ENUM_LIST
+
+  JMSG_LASTMSGCODE
+} J_MESSAGE_CODE;
+
+#undef JMAKE_ENUM_LIST
+#endif /* JMAKE_ENUM_LIST */
+
+/* Zap JMESSAGE macro so that future re-inclusions do nothing by default */
+#undef JMESSAGE
+
+
+#ifndef JERROR_H
+#define JERROR_H
+
+/* Macros to simplify using the error and trace message stuff */
+/* The first parameter is either type of cinfo pointer */
+
+/* Fatal errors (print message and exit) */
+#define ERREXIT(cinfo,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT1(cinfo,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT2(cinfo,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT3(cinfo,code,p1,p2,p3)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT4(cinfo,code,p1,p2,p3,p4)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (cinfo)->err->msg_parm.i[3] = (p4), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXITS(cinfo,code,str)  \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+
+#define MAKESTMT(stuff)		do { stuff } while (0)
+
+/* Nonfatal errors (we can keep going, but the data is probably corrupt) */
+#define WARNMS(cinfo,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+#define WARNMS1(cinfo,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+#define WARNMS2(cinfo,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+
+/* Informational/debugging messages */
+#define TRACEMS(cinfo,lvl,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS1(cinfo,lvl,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS2(cinfo,lvl,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS3(cinfo,lvl,code,p1,p2,p3)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS4(cinfo,lvl,code,p1,p2,p3,p4)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS5(cinfo,lvl,code,p1,p2,p3,p4,p5)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   _mp[4] = (p5); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS8(cinfo,lvl,code,p1,p2,p3,p4,p5,p6,p7,p8)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   _mp[4] = (p5); _mp[5] = (p6); _mp[6] = (p7); _mp[7] = (p8); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMSS(cinfo,lvl,code,str)  \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+
+#endif /* JERROR_H */
diff --git a/jfdctflt.c b/jfdctflt.c
new file mode 100644
index 0000000..79d7a00
--- /dev/null
+++ b/jfdctflt.c
@@ -0,0 +1,168 @@
+/*
+ * jfdctflt.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a floating-point implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * This implementation should be more accurate than either of the integer
+ * DCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_float (FAST_FLOAT * data)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
+  FAST_FLOAT *dataptr;
+  int ctr;
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+    
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;	/* phase 5 */
+    dataptr[6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;	/* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE*4] = tmp10 - tmp11;
+    
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
+    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE*6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE*3] = z13 - z2;
+    dataptr[DCTSIZE*1] = z11 + z4;
+    dataptr[DCTSIZE*7] = z11 - z4;
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
diff --git a/jfdctfst.c b/jfdctfst.c
new file mode 100644
index 0000000..ccb378a
--- /dev/null
+++ b/jfdctfst.c
@@ -0,0 +1,224 @@
+/*
+ * jfdctfst.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jfdctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * Again to save a few shifts, the intermediate results between pass 1 and
+ * pass 2 are not upscaled, but are represented only to integral precision.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#define CONST_BITS  8
+
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_0_382683433  ((INT32)   98)		/* FIX(0.382683433) */
+#define FIX_0_541196100  ((INT32)  139)		/* FIX(0.541196100) */
+#define FIX_0_707106781  ((INT32)  181)		/* FIX(0.707106781) */
+#define FIX_1_306562965  ((INT32)  334)		/* FIX(1.306562965) */
+#else
+#define FIX_0_382683433  FIX(0.382683433)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_707106781  FIX(0.707106781)
+#define FIX_1_306562965  FIX(1.306562965)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an INT32 constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_ifast (DCTELEM * data)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z1, z2, z3, z4, z5, z11, z13;
+  DCTELEM *dataptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;	/* phase 5 */
+    dataptr[6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;	/* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE*4] = tmp10 - tmp11;
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE*6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE*3] = z13 - z2;
+    dataptr[DCTSIZE*1] = z11 + z4;
+    dataptr[DCTSIZE*7] = z11 - z4;
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */
diff --git a/jfdctint.c b/jfdctint.c
new file mode 100644
index 0000000..0a78b64
--- /dev/null
+++ b/jfdctint.c
@@ -0,0 +1,283 @@
+/*
+ * jfdctint.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a slow-but-accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true DCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D DCT,
+ * because the y0 and y4 outputs need not be divided by sqrt(N).
+ * In the IJG code, this factor of 8 is removed by the quantization step
+ * (in jcdctmgr.c), NOT in this module.
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (For 12-bit sample data, the intermediate
+ * array is INT32 anyway.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((INT32)  2446)	/* FIX(0.298631336) */
+#define FIX_0_390180644  ((INT32)  3196)	/* FIX(0.390180644) */
+#define FIX_0_541196100  ((INT32)  4433)	/* FIX(0.541196100) */
+#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
+#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
+#define FIX_1_175875602  ((INT32)  9633)	/* FIX(1.175875602) */
+#define FIX_1_501321110  ((INT32)  12299)	/* FIX(1.501321110) */
+#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
+#define FIX_1_961570560  ((INT32)  16069)	/* FIX(1.961570560) */
+#define FIX_2_053119869  ((INT32)  16819)	/* FIX(2.053119869) */
+#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
+#define FIX_3_072711026  ((INT32)  25172)	/* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_islow (DCTELEM * data)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1, z2, z3, z4, z5;
+  DCTELEM *dataptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+    
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
+    dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+				   CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
+				   CONST_BITS-PASS1_BITS);
+    
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents cos(K*pi/16).
+     * i0..i3 in the paper are tmp4..tmp7 here.
+     */
+    
+    z1 = tmp4 + tmp7;
+    z2 = tmp5 + tmp6;
+    z3 = tmp4 + tmp6;
+    z4 = tmp5 + tmp7;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
+    dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
+    
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+    
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
+					   CONST_BITS+PASS1_BITS);
+    
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents cos(K*pi/16).
+     * i0..i3 in the paper are tmp4..tmp7 here.
+     */
+    
+    z1 = tmp4 + tmp7;
+    z2 = tmp5 + tmp6;
+    z3 = tmp4 + tmp6;
+    z4 = tmp5 + tmp7;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
+					   CONST_BITS+PASS1_BITS);
+    
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_ISLOW_SUPPORTED */
diff --git a/jidctflt.c b/jidctflt.c
new file mode 100644
index 0000000..0188ce3
--- /dev/null
+++ b/jidctflt.c
@@ -0,0 +1,242 @@
+/*
+ * jidctflt.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a floating-point implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * This implementation should be more accurate than either of the integer
+ * IDCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a float result.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((FAST_FLOAT) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  FLOAT_MULT_TYPE * quantptr;
+  FAST_FLOAT * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = tmp0 + tmp2;	/* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
+    tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;	/* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+    
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z13 = tmp6 + tmp5;		/* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
+
+    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
+    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
+    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    wsptr[DCTSIZE*0] = tmp0 + tmp7;
+    wsptr[DCTSIZE*7] = tmp0 - tmp7;
+    wsptr[DCTSIZE*1] = tmp1 + tmp6;
+    wsptr[DCTSIZE*6] = tmp1 - tmp6;
+    wsptr[DCTSIZE*2] = tmp2 + tmp5;
+    wsptr[DCTSIZE*5] = tmp2 - tmp5;
+    wsptr[DCTSIZE*4] = tmp3 + tmp4;
+    wsptr[DCTSIZE*3] = tmp3 - tmp4;
+
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * And testing floats for zero is relatively expensive, so we don't bother.
+     */
+    
+    /* Even part */
+
+    tmp10 = wsptr[0] + wsptr[4];
+    tmp11 = wsptr[0] - wsptr[4];
+
+    tmp13 = wsptr[2] + wsptr[6];
+    tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = wsptr[5] + wsptr[3];
+    z10 = wsptr[5] - wsptr[3];
+    z11 = wsptr[1] + wsptr[7];
+    z12 = wsptr[1] - wsptr[7];
+
+    tmp7 = z11 + z13;
+    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
+
+    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
+    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
+    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    /* Final output stage: scale down by a factor of 8 and range-limit */
+
+    outptr[0] = range_limit[(int) DESCALE((INT32) (tmp0 + tmp7), 3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) DESCALE((INT32) (tmp0 - tmp7), 3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE((INT32) (tmp1 + tmp6), 3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) DESCALE((INT32) (tmp1 - tmp6), 3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) DESCALE((INT32) (tmp2 + tmp5), 3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) DESCALE((INT32) (tmp2 - tmp5), 3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) DESCALE((INT32) (tmp3 + tmp4), 3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) DESCALE((INT32) (tmp3 - tmp4), 3)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
diff --git a/jidctfst.c b/jidctfst.c
new file mode 100644
index 0000000..dba4216
--- /dev/null
+++ b/jidctfst.c
@@ -0,0 +1,368 @@
+/*
+ * jidctfst.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jidctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * The dequantized coefficients are not integers because the AA&N scaling
+ * factors have been incorporated.  We represent them scaled up by PASS1_BITS,
+ * so that the first and second IDCT rounds have the same input scaling.
+ * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
+ * avoid a descaling shift; this compromises accuracy rather drastically
+ * for small quantization table entries, but it saves a lot of shifts.
+ * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
+ * so we use a much larger scaling factor to preserve accuracy.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  8
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  8
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_1_082392200  ((INT32)  277)		/* FIX(1.082392200) */
+#define FIX_1_414213562  ((INT32)  362)		/* FIX(1.414213562) */
+#define FIX_1_847759065  ((INT32)  473)		/* FIX(1.847759065) */
+#define FIX_2_613125930  ((INT32)  669)		/* FIX(2.613125930) */
+#else
+#define FIX_1_082392200  FIX(1.082392200)
+#define FIX_1_414213562  FIX(1.414213562)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_2_613125930  FIX(2.613125930)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an INT32 constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a DCTELEM result.  For 8-bit data a 16x16->16
+ * multiplication will do.  For 12-bit data, the multiplier table is
+ * declared INT32, so a 32-bit multiply will be used.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define DEQUANTIZE(coef,quantval)  (((IFAST_MULT_TYPE) (coef)) * (quantval))
+#else
+#define DEQUANTIZE(coef,quantval)  \
+	DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
+#endif
+
+
+/* Like DESCALE, but applies to a DCTELEM and produces an int.
+ * We assume that int right shift is unsigned if INT32 right shift is.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	DCTELEM ishift_temp;
+#if BITS_IN_JSAMPLE == 8
+#define DCTELEMBITS  16		/* DCTELEM may be 16 or 32 bits */
+#else
+#define DCTELEMBITS  32		/* DCTELEM must be 32 bits */
+#endif
+#define IRIGHT_SHIFT(x,shft)  \
+    ((ishift_temp = (x)) < 0 ? \
+     (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \
+     (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+#ifdef USE_ACCURATE_ROUNDING
+#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n))
+#else
+#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT(x, n))
+#endif
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  IFAST_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];	/* buffers data between passes */
+  SHIFT_TEMPS			/* for DESCALE */
+  ISHIFT_TEMPS			/* for IDESCALE */
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = tmp0 + tmp2;	/* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
+    tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;	/* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+    
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z13 = tmp6 + tmp5;		/* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
+    wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
+    wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
+    wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
+    wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
+    wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
+    wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);
+    wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);
+
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+    
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+
+    tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
+    tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
+
+    tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
+    tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562)
+	    - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
+    z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
+    z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
+    z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    /* Final output stage: scale down by a factor of 8 and range-limit */
+
+    outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */
diff --git a/jidctint.c b/jidctint.c
new file mode 100644
index 0000000..a72b320
--- /dev/null
+++ b/jidctint.c
@@ -0,0 +1,389 @@
+/*
+ * jidctint.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a slow-but-accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true IDCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D IDCT,
+ * because the y0 and y4 inputs need not be divided by sqrt(N).
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (To scale up 12-bit sample data further, an
+ * intermediate INT32 array would be needed.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((INT32)  2446)	/* FIX(0.298631336) */
+#define FIX_0_390180644  ((INT32)  3196)	/* FIX(0.390180644) */
+#define FIX_0_541196100  ((INT32)  4433)	/* FIX(0.541196100) */
+#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
+#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
+#define FIX_1_175875602  ((INT32)  9633)	/* FIX(1.175875602) */
+#define FIX_1_501321110  ((INT32)  12299)	/* FIX(1.501321110) */
+#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
+#define FIX_1_961570560  ((INT32)  16069)	/* FIX(1.961570560) */
+#define FIX_2_053119869  ((INT32)  16819)	/* FIX(2.053119869) */
+#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
+#define FIX_3_072711026  ((INT32)  25172)	/* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce an int result.  In this module, both inputs and result
+ * are 16 bits or less, so either int or short multiply will work.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1, z2, z3, z4, z5;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+  /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+    
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
+    tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+
+    tmp0 = (z2 + z3) << CONST_BITS;
+    tmp1 = (z2 - z3) << CONST_BITS;
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+    
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    
+    z1 = tmp0 + tmp3;
+    z2 = tmp1 + tmp2;
+    z3 = tmp0 + tmp2;
+    z4 = tmp1 + tmp3;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    tmp0 += z1 + z3;
+    tmp1 += z2 + z4;
+    tmp2 += z2 + z3;
+    tmp3 += z1 + z4;
+    
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+    
+    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
+    
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+    
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[6];
+    
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
+    tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    
+    tmp0 = ((INT32) wsptr[0] + (INT32) wsptr[4]) << CONST_BITS;
+    tmp1 = ((INT32) wsptr[0] - (INT32) wsptr[4]) << CONST_BITS;
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+    
+    tmp0 = (INT32) wsptr[7];
+    tmp1 = (INT32) wsptr[5];
+    tmp2 = (INT32) wsptr[3];
+    tmp3 = (INT32) wsptr[1];
+    
+    z1 = tmp0 + tmp3;
+    z2 = tmp1 + tmp2;
+    z3 = tmp0 + tmp2;
+    z4 = tmp1 + tmp3;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    tmp0 += z1 + z3;
+    tmp1 += z2 + z4;
+    tmp2 += z2 + z3;
+    tmp3 += z1 + z4;
+    
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+    
+    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp3,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) DESCALE(tmp10 - tmp3,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE(tmp11 + tmp2,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) DESCALE(tmp11 - tmp2,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) DESCALE(tmp12 + tmp1,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) DESCALE(tmp12 - tmp1,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) DESCALE(tmp13 + tmp0,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) DESCALE(tmp13 - tmp0,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_ISLOW_SUPPORTED */
diff --git a/jidctred.c b/jidctred.c
new file mode 100644
index 0000000..421f3c7
--- /dev/null
+++ b/jidctred.c
@@ -0,0 +1,398 @@
+/*
+ * jidctred.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains inverse-DCT routines that produce reduced-size output:
+ * either 4x4, 2x2, or 1x1 pixels from an 8x8 DCT block.
+ *
+ * The implementation is based on the Loeffler, Ligtenberg and Moschytz (LL&M)
+ * algorithm used in jidctint.c.  We simply replace each 8-to-8 1-D IDCT step
+ * with an 8-to-4 step that produces the four averages of two adjacent outputs
+ * (or an 8-to-2 step producing two averages of four outputs, for 2x2 output).
+ * These steps were derived by computing the corresponding values at the end
+ * of the normal LL&M code, then simplifying as much as possible.
+ *
+ * 1x1 is trivial: just take the DC coefficient divided by 8.
+ *
+ * See jidctint.c for additional comments.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling is the same as in jidctint.c. */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_211164243  ((INT32)  1730)	/* FIX(0.211164243) */
+#define FIX_0_509795579  ((INT32)  4176)	/* FIX(0.509795579) */
+#define FIX_0_601344887  ((INT32)  4926)	/* FIX(0.601344887) */
+#define FIX_0_720959822  ((INT32)  5906)	/* FIX(0.720959822) */
+#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
+#define FIX_0_850430095  ((INT32)  6967)	/* FIX(0.850430095) */
+#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
+#define FIX_1_061594337  ((INT32)  8697)	/* FIX(1.061594337) */
+#define FIX_1_272758580  ((INT32)  10426)	/* FIX(1.272758580) */
+#define FIX_1_451774981  ((INT32)  11893)	/* FIX(1.451774981) */
+#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
+#define FIX_2_172734803  ((INT32)  17799)	/* FIX(2.172734803) */
+#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
+#define FIX_3_624509785  ((INT32)  29692)	/* FIX(3.624509785) */
+#else
+#define FIX_0_211164243  FIX(0.211164243)
+#define FIX_0_509795579  FIX(0.509795579)
+#define FIX_0_601344887  FIX(0.601344887)
+#define FIX_0_720959822  FIX(0.720959822)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_850430095  FIX(0.850430095)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_061594337  FIX(1.061594337)
+#define FIX_1_272758580  FIX(1.272758580)
+#define FIX_1_451774981  FIX(1.451774981)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_2_172734803  FIX(2.172734803)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_624509785  FIX(3.624509785)
+#endif
+
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce an int result.  In this module, both inputs and result
+ * are 16 bits or less, so either int or short multiply will work.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 4x4 output block.
+ */
+
+GLOBAL(void)
+jpeg_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp2, tmp10, tmp12;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE*4];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
+    /* Don't bother to process column 4, because second pass won't use it */
+    if (ctr == DCTSIZE-4)
+      continue;
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*5] == 0 &&
+	inptr[DCTSIZE*6] == 0 && inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero; we need not examine term 4 for 4x4 output */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      
+      continue;
+    }
+    
+    /* Even part */
+    
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp0 <<= (CONST_BITS+1);
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp2 = MULTIPLY(z2, FIX_1_847759065) + MULTIPLY(z3, - FIX_0_765366865);
+    
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+    
+    /* Odd part */
+    
+    z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    
+    tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
+	 + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
+	 + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
+	 + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
+    
+    tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
+	 + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
+	 + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
+	 + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
+
+    /* Final output stage */
+    
+    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp2, CONST_BITS-PASS1_BITS+1);
+    wsptr[DCTSIZE*3] = (int) DESCALE(tmp10 - tmp2, CONST_BITS-PASS1_BITS+1);
+    wsptr[DCTSIZE*1] = (int) DESCALE(tmp12 + tmp0, CONST_BITS-PASS1_BITS+1);
+    wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 - tmp0, CONST_BITS-PASS1_BITS+1);
+  }
+  
+  /* Pass 2: process 4 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 4; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* It's not clear whether a zero row test is worthwhile here ... */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+    
+    tmp0 = ((INT32) wsptr[0]) << (CONST_BITS+1);
+    
+    tmp2 = MULTIPLY((INT32) wsptr[2], FIX_1_847759065)
+	 + MULTIPLY((INT32) wsptr[6], - FIX_0_765366865);
+    
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+    
+    /* Odd part */
+    
+    z1 = (INT32) wsptr[7];
+    z2 = (INT32) wsptr[5];
+    z3 = (INT32) wsptr[3];
+    z4 = (INT32) wsptr[1];
+    
+    tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
+	 + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
+	 + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
+	 + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
+    
+    tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
+	 + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
+	 + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
+	 + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
+
+    /* Final output stage */
+    
+    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp2,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) DESCALE(tmp10 - tmp2,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE(tmp12 + tmp0,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) DESCALE(tmp12 - tmp0,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 2x2 output block.
+ */
+
+GLOBAL(void)
+jpeg_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp10, z1;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE*2];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
+    /* Don't bother to process columns 2,4,6 */
+    if (ctr == DCTSIZE-2 || ctr == DCTSIZE-4 || ctr == DCTSIZE-6)
+      continue;
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*3] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero; we need not examine terms 2,4,6 for 2x2 output */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      
+      continue;
+    }
+    
+    /* Even part */
+    
+    z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp10 = z1 << (CONST_BITS+2);
+    
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp0 = MULTIPLY(z1, - FIX_0_720959822); /* sqrt(2) * (c7-c5+c3-c1) */
+    z1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp0 += MULTIPLY(z1, FIX_0_850430095); /* sqrt(2) * (-c1+c3+c5+c7) */
+    z1 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp0 += MULTIPLY(z1, - FIX_1_272758580); /* sqrt(2) * (-c1+c3-c5-c7) */
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp0 += MULTIPLY(z1, FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
+
+    /* Final output stage */
+    
+    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp0, CONST_BITS-PASS1_BITS+2);
+    wsptr[DCTSIZE*1] = (int) DESCALE(tmp10 - tmp0, CONST_BITS-PASS1_BITS+2);
+  }
+  
+  /* Pass 2: process 2 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 2; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* It's not clear whether a zero row test is worthwhile here ... */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[3] == 0 && wsptr[5] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+    
+    tmp10 = ((INT32) wsptr[0]) << (CONST_BITS+2);
+    
+    /* Odd part */
+
+    tmp0 = MULTIPLY((INT32) wsptr[7], - FIX_0_720959822) /* sqrt(2) * (c7-c5+c3-c1) */
+	 + MULTIPLY((INT32) wsptr[5], FIX_0_850430095) /* sqrt(2) * (-c1+c3+c5+c7) */
+	 + MULTIPLY((INT32) wsptr[3], - FIX_1_272758580) /* sqrt(2) * (-c1+c3-c5-c7) */
+	 + MULTIPLY((INT32) wsptr[1], FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
+
+    /* Final output stage */
+    
+    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp0,
+					  CONST_BITS+PASS1_BITS+3+2)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE(tmp10 - tmp0,
+					  CONST_BITS+PASS1_BITS+3+2)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 1x1 output block.
+ */
+
+GLOBAL(void)
+jpeg_idct_1x1 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  int dcval;
+  ISLOW_MULT_TYPE * quantptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  SHIFT_TEMPS
+
+  /* We hardly need an inverse DCT routine for this: just take the
+   * average pixel value, which is one-eighth of the DC coefficient.
+   */
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  dcval = DEQUANTIZE(coef_block[0], quantptr[0]);
+  dcval = (int) DESCALE((INT32) dcval, 3);
+
+  output_buf[0][output_col] = range_limit[dcval & RANGE_MASK];
+}
+
+#endif /* IDCT_SCALING_SUPPORTED */
diff --git a/jinclude.h b/jinclude.h
new file mode 100644
index 0000000..0a4f151
--- /dev/null
+++ b/jinclude.h
@@ -0,0 +1,91 @@
+/*
+ * jinclude.h
+ *
+ * Copyright (C) 1991-1994, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file exists to provide a single place to fix any problems with
+ * including the wrong system include files.  (Common problems are taken
+ * care of by the standard jconfig symbols, but on really weird systems
+ * you may have to edit this file.)
+ *
+ * NOTE: this file is NOT intended to be included by applications using the
+ * JPEG library.  Most applications need only include jpeglib.h.
+ */
+
+
+/* Include auto-config file to find out which system include files we need. */
+
+#include "jconfig.h"		/* auto configuration options */
+#define JCONFIG_INCLUDED	/* so that jpeglib.h doesn't do it again */
+
+/*
+ * We need the NULL macro and size_t typedef.
+ * On an ANSI-conforming system it is sufficient to include <stddef.h>.
+ * Otherwise, we get them from <stdlib.h> or <stdio.h>; we may have to
+ * pull in <sys/types.h> as well.
+ * Note that the core JPEG library does not require <stdio.h>;
+ * only the default error handler and data source/destination modules do.
+ * But we must pull it in because of the references to FILE in jpeglib.h.
+ * You can remove those references if you want to compile without <stdio.h>.
+ */
+
+#ifdef HAVE_STDDEF_H
+#include <stddef.h>
+#endif
+
+#ifdef HAVE_STDLIB_H
+#include <stdlib.h>
+#endif
+
+#ifdef NEED_SYS_TYPES_H
+#include <sys/types.h>
+#endif
+
+#include <stdio.h>
+
+/*
+ * We need memory copying and zeroing functions, plus strncpy().
+ * ANSI and System V implementations declare these in <string.h>.
+ * BSD doesn't have the mem() functions, but it does have bcopy()/bzero().
+ * Some systems may declare memset and memcpy in <memory.h>.
+ *
+ * NOTE: we assume the size parameters to these functions are of type size_t.
+ * Change the casts in these macros if not!
+ */
+
+#ifdef NEED_BSD_STRINGS
+
+#include <strings.h>
+#define MEMZERO(target,size)	bzero((void *)(target), (size_t)(size))
+#define MEMCOPY(dest,src,size)	bcopy((const void *)(src), (void *)(dest), (size_t)(size))
+
+#else /* not BSD, assume ANSI/SysV string lib */
+
+#include <string.h>
+#define MEMZERO(target,size)	memset((void *)(target), 0, (size_t)(size))
+#define MEMCOPY(dest,src,size)	memcpy((void *)(dest), (const void *)(src), (size_t)(size))
+
+#endif
+
+/*
+ * In ANSI C, and indeed any rational implementation, size_t is also the
+ * type returned by sizeof().  However, it seems there are some irrational
+ * implementations out there, in which sizeof() returns an int even though
+ * size_t is defined as long or unsigned long.  To ensure consistent results
+ * we always use this SIZEOF() macro in place of using sizeof() directly.
+ */
+
+#define SIZEOF(object)	((size_t) sizeof(object))
+
+/*
+ * The modules that use fread() and fwrite() always invoke them through
+ * these macros.  On some systems you may need to twiddle the argument casts.
+ * CAUTION: argument order is different from underlying functions!
+ */
+
+#define JFREAD(file,buf,sizeofbuf)  \
+  ((size_t) fread((void *) (buf), (size_t) 1, (size_t) (sizeofbuf), (file)))
+#define JFWRITE(file,buf,sizeofbuf)  \
+  ((size_t) fwrite((const void *) (buf), (size_t) 1, (size_t) (sizeofbuf), (file)))
diff --git a/jmemmgr.c b/jmemmgr.c
new file mode 100644
index 0000000..5458952
--- /dev/null
+++ b/jmemmgr.c
@@ -0,0 +1,1142 @@
+/*
+ * jmemmgr.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the JPEG system-independent memory management
+ * routines.  This code is usable across a wide variety of machines; most
+ * of the system dependencies have been isolated in a separate file.
+ * The major functions provided here are:
+ *   * pool-based allocation and freeing of memory;
+ *   * policy decisions about how to divide available memory among the
+ *     virtual arrays;
+ *   * control logic for swapping virtual arrays between main memory and
+ *     backing storage.
+ * The separate system-dependent file provides the actual backing-storage
+ * access code, and it contains the policy decision about how much total
+ * main memory to use.
+ * This file is system-dependent in the sense that some of its functions
+ * are unnecessary in some systems.  For example, if there is enough virtual
+ * memory so that backing storage will never be used, much of the virtual
+ * array control logic could be removed.  (Of course, if you have that much
+ * memory then you shouldn't care about a little bit of unused code...)
+ */
+
+#define JPEG_INTERNALS
+#define AM_MEMORY_MANAGER	/* we define jvirt_Xarray_control structs */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jmemsys.h"		/* import the system-dependent declarations */
+
+#ifndef NO_GETENV
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare getenv() */
+extern char * getenv JPP((const char * name));
+#endif
+#endif
+
+
+/*
+ * Some important notes:
+ *   The allocation routines provided here must never return NULL.
+ *   They should exit to error_exit if unsuccessful.
+ *
+ *   It's not a good idea to try to merge the sarray and barray routines,
+ *   even though they are textually almost the same, because samples are
+ *   usually stored as bytes while coefficients are shorts or ints.  Thus,
+ *   in machines where byte pointers have a different representation from
+ *   word pointers, the resulting machine code could not be the same.
+ */
+
+
+/*
+ * Many machines require storage alignment: longs must start on 4-byte
+ * boundaries, doubles on 8-byte boundaries, etc.  On such machines, malloc()
+ * always returns pointers that are multiples of the worst-case alignment
+ * requirement, and we had better do so too.
+ * There isn't any really portable way to determine the worst-case alignment
+ * requirement.  This module assumes that the alignment requirement is
+ * multiples of ALIGN_SIZE.
+ * By default, we define ALIGN_SIZE as sizeof(double).  This is necessary on some
+ * workstations (where doubles really do need 8-byte alignment) and will work
+ * fine on nearly everything.  If your machine has lesser alignment needs,
+ * you can save a few bytes by making ALIGN_SIZE smaller.
+ * The only place I know of where this will NOT work is certain Macintosh
+ * 680x0 compilers that define double as a 10-byte IEEE extended float.
+ * Doing 10-byte alignment is counterproductive because longwords won't be
+ * aligned well.  Put "#define ALIGN_SIZE 4" in jconfig.h if you have
+ * such a compiler.
+ */
+
+#ifndef ALIGN_SIZE		/* so can override from jconfig.h */
+#ifndef WITH_SIMD
+#define ALIGN_SIZE  SIZEOF(double)
+#else
+#define ALIGN_SIZE  16 /* Most SIMD implementations require this */
+#endif
+#endif
+
+/*
+ * We allocate objects from "pools", where each pool is gotten with a single
+ * request to jpeg_get_small() or jpeg_get_large().  There is no per-object
+ * overhead within a pool, except for alignment padding.  Each pool has a
+ * header with a link to the next pool of the same class.
+ * Small and large pool headers are identical except that the latter's
+ * link pointer must be FAR on 80x86 machines.
+ */
+
+typedef struct small_pool_struct * small_pool_ptr;
+
+typedef struct small_pool_struct {
+  small_pool_ptr next;	/* next in list of pools */
+  size_t bytes_used;		/* how many bytes already used within pool */
+  size_t bytes_left;		/* bytes still available in this pool */
+} small_pool_hdr;
+
+typedef struct large_pool_struct FAR * large_pool_ptr;
+
+typedef struct large_pool_struct {
+  large_pool_ptr next;	/* next in list of pools */
+  size_t bytes_used;		/* how many bytes already used within pool */
+  size_t bytes_left;		/* bytes still available in this pool */
+} large_pool_hdr;
+
+/*
+ * Here is the full definition of a memory manager object.
+ */
+
+typedef struct {
+  struct jpeg_memory_mgr pub;	/* public fields */
+
+  /* Each pool identifier (lifetime class) names a linked list of pools. */
+  small_pool_ptr small_list[JPOOL_NUMPOOLS];
+  large_pool_ptr large_list[JPOOL_NUMPOOLS];
+
+  /* Since we only have one lifetime class of virtual arrays, only one
+   * linked list is necessary (for each datatype).  Note that the virtual
+   * array control blocks being linked together are actually stored somewhere
+   * in the small-pool list.
+   */
+  jvirt_sarray_ptr virt_sarray_list;
+  jvirt_barray_ptr virt_barray_list;
+
+  /* This counts total space obtained from jpeg_get_small/large */
+  size_t total_space_allocated;
+
+  /* alloc_sarray and alloc_barray set this value for use by virtual
+   * array routines.
+   */
+  JDIMENSION last_rowsperchunk;	/* from most recent alloc_sarray/barray */
+} my_memory_mgr;
+
+typedef my_memory_mgr * my_mem_ptr;
+
+
+/*
+ * The control blocks for virtual arrays.
+ * Note that these blocks are allocated in the "small" pool area.
+ * System-dependent info for the associated backing store (if any) is hidden
+ * inside the backing_store_info struct.
+ */
+
+struct jvirt_sarray_control {
+  JSAMPARRAY mem_buffer;	/* => the in-memory buffer */
+  JDIMENSION rows_in_array;	/* total virtual array height */
+  JDIMENSION samplesperrow;	/* width of array (and of memory buffer) */
+  JDIMENSION maxaccess;		/* max rows accessed by access_virt_sarray */
+  JDIMENSION rows_in_mem;	/* height of memory buffer */
+  JDIMENSION rowsperchunk;	/* allocation chunk size in mem_buffer */
+  JDIMENSION cur_start_row;	/* first logical row # in the buffer */
+  JDIMENSION first_undef_row;	/* row # of first uninitialized row */
+  boolean pre_zero;		/* pre-zero mode requested? */
+  boolean dirty;		/* do current buffer contents need written? */
+  boolean b_s_open;		/* is backing-store data valid? */
+  jvirt_sarray_ptr next;	/* link to next virtual sarray control block */
+  backing_store_info b_s_info;	/* System-dependent control info */
+};
+
+struct jvirt_barray_control {
+  JBLOCKARRAY mem_buffer;	/* => the in-memory buffer */
+  JDIMENSION rows_in_array;	/* total virtual array height */
+  JDIMENSION blocksperrow;	/* width of array (and of memory buffer) */
+  JDIMENSION maxaccess;		/* max rows accessed by access_virt_barray */
+  JDIMENSION rows_in_mem;	/* height of memory buffer */
+  JDIMENSION rowsperchunk;	/* allocation chunk size in mem_buffer */
+  JDIMENSION cur_start_row;	/* first logical row # in the buffer */
+  JDIMENSION first_undef_row;	/* row # of first uninitialized row */
+  boolean pre_zero;		/* pre-zero mode requested? */
+  boolean dirty;		/* do current buffer contents need written? */
+  boolean b_s_open;		/* is backing-store data valid? */
+  jvirt_barray_ptr next;	/* link to next virtual barray control block */
+  backing_store_info b_s_info;	/* System-dependent control info */
+};
+
+
+#ifdef MEM_STATS		/* optional extra stuff for statistics */
+
+LOCAL(void)
+print_mem_stats (j_common_ptr cinfo, int pool_id)
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  small_pool_ptr shdr_ptr;
+  large_pool_ptr lhdr_ptr;
+
+  /* Since this is only a debugging stub, we can cheat a little by using
+   * fprintf directly rather than going through the trace message code.
+   * This is helpful because message parm array can't handle longs.
+   */
+  fprintf(stderr, "Freeing pool %d, total space = %ld\n",
+	  pool_id, mem->total_space_allocated);
+
+  for (lhdr_ptr = mem->large_list[pool_id]; lhdr_ptr != NULL;
+       lhdr_ptr = lhdr_ptr->next) {
+    fprintf(stderr, "  Large chunk used %ld\n",
+	    (long) lhdr_ptr->bytes_used);
+  }
+
+  for (shdr_ptr = mem->small_list[pool_id]; shdr_ptr != NULL;
+       shdr_ptr = shdr_ptr->next) {
+    fprintf(stderr, "  Small chunk used %ld free %ld\n",
+	    (long) shdr_ptr->bytes_used,
+	    (long) shdr_ptr->bytes_left);
+  }
+}
+
+#endif /* MEM_STATS */
+
+
+LOCAL(void)
+out_of_memory (j_common_ptr cinfo, int which)
+/* Report an out-of-memory error and stop execution */
+/* If we compiled MEM_STATS support, report alloc requests before dying */
+{
+#ifdef MEM_STATS
+  cinfo->err->trace_level = 2;	/* force self_destruct to report stats */
+#endif
+  ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, which);
+}
+
+
+/*
+ * Allocation of "small" objects.
+ *
+ * For these, we use pooled storage.  When a new pool must be created,
+ * we try to get enough space for the current request plus a "slop" factor,
+ * where the slop will be the amount of leftover space in the new pool.
+ * The speed vs. space tradeoff is largely determined by the slop values.
+ * A different slop value is provided for each pool class (lifetime),
+ * and we also distinguish the first pool of a class from later ones.
+ * NOTE: the values given work fairly well on both 16- and 32-bit-int
+ * machines, but may be too small if longs are 64 bits or more.
+ *
+ * Since we do not know what alignment malloc() gives us, we have to
+ * allocate ALIGN_SIZE-1 extra space per pool to have room for alignment
+ * adjustment.
+ */
+
+static const size_t first_pool_slop[JPOOL_NUMPOOLS] = 
+{
+	1600,			/* first PERMANENT pool */
+	16000			/* first IMAGE pool */
+};
+
+static const size_t extra_pool_slop[JPOOL_NUMPOOLS] = 
+{
+	0,			/* additional PERMANENT pools */
+	5000			/* additional IMAGE pools */
+};
+
+#define MIN_SLOP  50		/* greater than 0 to avoid futile looping */
+
+
+METHODDEF(void *)
+alloc_small (j_common_ptr cinfo, int pool_id, size_t sizeofobject)
+/* Allocate a "small" object */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  small_pool_ptr hdr_ptr, prev_hdr_ptr;
+  char * data_ptr;
+  size_t min_request, slop;
+
+  /*
+   * Round up the requested size to a multiple of ALIGN_SIZE in order
+   * to assure alignment for the next object allocated in the same pool
+   * and so that algorithms can straddle outside the proper area up
+   * to the next alignment.
+   */
+  sizeofobject = jround_up(sizeofobject, ALIGN_SIZE);
+
+  /* Check for unsatisfiable request (do now to ensure no overflow below) */
+  if ((SIZEOF(small_pool_hdr) + sizeofobject + ALIGN_SIZE - 1) > MAX_ALLOC_CHUNK)
+    out_of_memory(cinfo, 1);	/* request exceeds malloc's ability */
+
+  /* See if space is available in any existing pool */
+  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+  prev_hdr_ptr = NULL;
+  hdr_ptr = mem->small_list[pool_id];
+  while (hdr_ptr != NULL) {
+    if (hdr_ptr->bytes_left >= sizeofobject)
+      break;			/* found pool with enough space */
+    prev_hdr_ptr = hdr_ptr;
+    hdr_ptr = hdr_ptr->next;
+  }
+
+  /* Time to make a new pool? */
+  if (hdr_ptr == NULL) {
+    /* min_request is what we need now, slop is what will be leftover */
+    min_request = SIZEOF(small_pool_hdr) + sizeofobject + ALIGN_SIZE - 1;
+    if (prev_hdr_ptr == NULL)	/* first pool in class? */
+      slop = first_pool_slop[pool_id];
+    else
+      slop = extra_pool_slop[pool_id];
+    /* Don't ask for more than MAX_ALLOC_CHUNK */
+    if (slop > (size_t) (MAX_ALLOC_CHUNK-min_request))
+      slop = (size_t) (MAX_ALLOC_CHUNK-min_request);
+    /* Try to get space, if fail reduce slop and try again */
+    for (;;) {
+      hdr_ptr = (small_pool_ptr) jpeg_get_small(cinfo, min_request + slop);
+      if (hdr_ptr != NULL)
+	break;
+      slop /= 2;
+      if (slop < MIN_SLOP)	/* give up when it gets real small */
+	out_of_memory(cinfo, 2); /* jpeg_get_small failed */
+    }
+    mem->total_space_allocated += min_request + slop;
+    /* Success, initialize the new pool header and add to end of list */
+    hdr_ptr->next = NULL;
+    hdr_ptr->bytes_used = 0;
+    hdr_ptr->bytes_left = sizeofobject + slop;
+    if (prev_hdr_ptr == NULL)	/* first pool in class? */
+      mem->small_list[pool_id] = hdr_ptr;
+    else
+      prev_hdr_ptr->next = hdr_ptr;
+  }
+
+  /* OK, allocate the object from the current pool */
+  data_ptr = (char *) hdr_ptr; /* point to first data byte in pool... */
+  data_ptr += SIZEOF(small_pool_hdr); /* ...by skipping the header... */
+  if ((size_t)data_ptr % ALIGN_SIZE) /* ...and adjust for alignment */
+    data_ptr += ALIGN_SIZE - (size_t)data_ptr % ALIGN_SIZE;
+  data_ptr += hdr_ptr->bytes_used; /* point to place for object */
+  hdr_ptr->bytes_used += sizeofobject;
+  hdr_ptr->bytes_left -= sizeofobject;
+
+  return (void *) data_ptr;
+}
+
+
+/*
+ * Allocation of "large" objects.
+ *
+ * The external semantics of these are the same as "small" objects,
+ * except that FAR pointers are used on 80x86.  However the pool
+ * management heuristics are quite different.  We assume that each
+ * request is large enough that it may as well be passed directly to
+ * jpeg_get_large; the pool management just links everything together
+ * so that we can free it all on demand.
+ * Note: the major use of "large" objects is in JSAMPARRAY and JBLOCKARRAY
+ * structures.  The routines that create these structures (see below)
+ * deliberately bunch rows together to ensure a large request size.
+ */
+
+METHODDEF(void FAR *)
+alloc_large (j_common_ptr cinfo, int pool_id, size_t sizeofobject)
+/* Allocate a "large" object */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  large_pool_ptr hdr_ptr;
+  char FAR * data_ptr;
+
+  /*
+   * Round up the requested size to a multiple of ALIGN_SIZE so that
+   * algorithms can straddle outside the proper area up to the next
+   * alignment.
+   */
+  sizeofobject = jround_up(sizeofobject, ALIGN_SIZE);
+
+  /* Check for unsatisfiable request (do now to ensure no overflow below) */
+  if ((SIZEOF(large_pool_hdr) + sizeofobject + ALIGN_SIZE - 1) > MAX_ALLOC_CHUNK)
+    out_of_memory(cinfo, 3);	/* request exceeds malloc's ability */
+
+  /* Always make a new pool */
+  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+  hdr_ptr = (large_pool_ptr) jpeg_get_large(cinfo, sizeofobject +
+					    SIZEOF(large_pool_hdr) +
+					    ALIGN_SIZE - 1);
+  if (hdr_ptr == NULL)
+    out_of_memory(cinfo, 4);	/* jpeg_get_large failed */
+  mem->total_space_allocated += sizeofobject + SIZEOF(large_pool_hdr) + ALIGN_SIZE - 1;
+
+  /* Success, initialize the new pool header and add to list */
+  hdr_ptr->next = mem->large_list[pool_id];
+  /* We maintain space counts in each pool header for statistical purposes,
+   * even though they are not needed for allocation.
+   */
+  hdr_ptr->bytes_used = sizeofobject;
+  hdr_ptr->bytes_left = 0;
+  mem->large_list[pool_id] = hdr_ptr;
+
+  data_ptr = (char *) hdr_ptr; /* point to first data byte in pool... */
+  data_ptr += SIZEOF(small_pool_hdr); /* ...by skipping the header... */
+  if ((size_t)data_ptr % ALIGN_SIZE) /* ...and adjust for alignment */
+    data_ptr += ALIGN_SIZE - (size_t)data_ptr % ALIGN_SIZE;
+
+  return (void FAR *) data_ptr;
+}
+
+
+/*
+ * Creation of 2-D sample arrays.
+ * The pointers are in near heap, the samples themselves in FAR heap.
+ *
+ * To minimize allocation overhead and to allow I/O of large contiguous
+ * blocks, we allocate the sample rows in groups of as many rows as possible
+ * without exceeding MAX_ALLOC_CHUNK total bytes per allocation request.
+ * NB: the virtual array control routines, later in this file, know about
+ * this chunking of rows.  The rowsperchunk value is left in the mem manager
+ * object so that it can be saved away if this sarray is the workspace for
+ * a virtual array.
+ *
+ * Since we are often upsampling with a factor 2, we align the size (not
+ * the start) to 2 * ALIGN_SIZE so that the upsampling routines don't have
+ * to be as careful about size.
+ */
+
+METHODDEF(JSAMPARRAY)
+alloc_sarray (j_common_ptr cinfo, int pool_id,
+	      JDIMENSION samplesperrow, JDIMENSION numrows)
+/* Allocate a 2-D sample array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  JSAMPARRAY result;
+  JSAMPROW workspace;
+  JDIMENSION rowsperchunk, currow, i;
+  long ltemp;
+
+  /* Make sure each row is properly aligned */
+  if ((ALIGN_SIZE % SIZEOF(JSAMPLE)) != 0)
+    out_of_memory(cinfo, 5);	/* safety check */
+  samplesperrow = (JDIMENSION)jround_up(samplesperrow, (2 * ALIGN_SIZE) / SIZEOF(JSAMPLE));
+
+  /* Calculate max # of rows allowed in one allocation chunk */
+  ltemp = (MAX_ALLOC_CHUNK-SIZEOF(large_pool_hdr)) /
+	  ((long) samplesperrow * SIZEOF(JSAMPLE));
+  if (ltemp <= 0)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+  if (ltemp < (long) numrows)
+    rowsperchunk = (JDIMENSION) ltemp;
+  else
+    rowsperchunk = numrows;
+  mem->last_rowsperchunk = rowsperchunk;
+
+  /* Get space for row pointers (small object) */
+  result = (JSAMPARRAY) alloc_small(cinfo, pool_id,
+				    (size_t) (numrows * SIZEOF(JSAMPROW)));
+
+  /* Get the rows themselves (large objects) */
+  currow = 0;
+  while (currow < numrows) {
+    rowsperchunk = MIN(rowsperchunk, numrows - currow);
+    workspace = (JSAMPROW) alloc_large(cinfo, pool_id,
+	(size_t) ((size_t) rowsperchunk * (size_t) samplesperrow
+		  * SIZEOF(JSAMPLE)));
+    for (i = rowsperchunk; i > 0; i--) {
+      result[currow++] = workspace;
+      workspace += samplesperrow;
+    }
+  }
+
+  return result;
+}
+
+
+/*
+ * Creation of 2-D coefficient-block arrays.
+ * This is essentially the same as the code for sample arrays, above.
+ */
+
+METHODDEF(JBLOCKARRAY)
+alloc_barray (j_common_ptr cinfo, int pool_id,
+	      JDIMENSION blocksperrow, JDIMENSION numrows)
+/* Allocate a 2-D coefficient-block array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  JBLOCKARRAY result;
+  JBLOCKROW workspace;
+  JDIMENSION rowsperchunk, currow, i;
+  long ltemp;
+
+  /* Make sure each row is properly aligned */
+  if ((SIZEOF(JBLOCK) % ALIGN_SIZE) != 0)
+    out_of_memory(cinfo, 6);	/* safety check */
+
+  /* Calculate max # of rows allowed in one allocation chunk */
+  ltemp = (MAX_ALLOC_CHUNK-SIZEOF(large_pool_hdr)) /
+	  ((long) blocksperrow * SIZEOF(JBLOCK));
+  if (ltemp <= 0)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+  if (ltemp < (long) numrows)
+    rowsperchunk = (JDIMENSION) ltemp;
+  else
+    rowsperchunk = numrows;
+  mem->last_rowsperchunk = rowsperchunk;
+
+  /* Get space for row pointers (small object) */
+  result = (JBLOCKARRAY) alloc_small(cinfo, pool_id,
+				     (size_t) (numrows * SIZEOF(JBLOCKROW)));
+
+  /* Get the rows themselves (large objects) */
+  currow = 0;
+  while (currow < numrows) {
+    rowsperchunk = MIN(rowsperchunk, numrows - currow);
+    workspace = (JBLOCKROW) alloc_large(cinfo, pool_id,
+	(size_t) ((size_t) rowsperchunk * (size_t) blocksperrow
+		  * SIZEOF(JBLOCK)));
+    for (i = rowsperchunk; i > 0; i--) {
+      result[currow++] = workspace;
+      workspace += blocksperrow;
+    }
+  }
+
+  return result;
+}
+
+
+/*
+ * About virtual array management:
+ *
+ * The above "normal" array routines are only used to allocate strip buffers
+ * (as wide as the image, but just a few rows high).  Full-image-sized buffers
+ * are handled as "virtual" arrays.  The array is still accessed a strip at a
+ * time, but the memory manager must save the whole array for repeated
+ * accesses.  The intended implementation is that there is a strip buffer in
+ * memory (as high as is possible given the desired memory limit), plus a
+ * backing file that holds the rest of the array.
+ *
+ * The request_virt_array routines are told the total size of the image and
+ * the maximum number of rows that will be accessed at once.  The in-memory
+ * buffer must be at least as large as the maxaccess value.
+ *
+ * The request routines create control blocks but not the in-memory buffers.
+ * That is postponed until realize_virt_arrays is called.  At that time the
+ * total amount of space needed is known (approximately, anyway), so free
+ * memory can be divided up fairly.
+ *
+ * The access_virt_array routines are responsible for making a specific strip
+ * area accessible (after reading or writing the backing file, if necessary).
+ * Note that the access routines are told whether the caller intends to modify
+ * the accessed strip; during a read-only pass this saves having to rewrite
+ * data to disk.  The access routines are also responsible for pre-zeroing
+ * any newly accessed rows, if pre-zeroing was requested.
+ *
+ * In current usage, the access requests are usually for nonoverlapping
+ * strips; that is, successive access start_row numbers differ by exactly
+ * num_rows = maxaccess.  This means we can get good performance with simple
+ * buffer dump/reload logic, by making the in-memory buffer be a multiple
+ * of the access height; then there will never be accesses across bufferload
+ * boundaries.  The code will still work with overlapping access requests,
+ * but it doesn't handle bufferload overlaps very efficiently.
+ */
+
+
+METHODDEF(jvirt_sarray_ptr)
+request_virt_sarray (j_common_ptr cinfo, int pool_id, boolean pre_zero,
+		     JDIMENSION samplesperrow, JDIMENSION numrows,
+		     JDIMENSION maxaccess)
+/* Request a virtual 2-D sample array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  jvirt_sarray_ptr result;
+
+  /* Only IMAGE-lifetime virtual arrays are currently supported */
+  if (pool_id != JPOOL_IMAGE)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+  /* get control block */
+  result = (jvirt_sarray_ptr) alloc_small(cinfo, pool_id,
+					  SIZEOF(struct jvirt_sarray_control));
+
+  result->mem_buffer = NULL;	/* marks array not yet realized */
+  result->rows_in_array = numrows;
+  result->samplesperrow = samplesperrow;
+  result->maxaccess = maxaccess;
+  result->pre_zero = pre_zero;
+  result->b_s_open = FALSE;	/* no associated backing-store object */
+  result->next = mem->virt_sarray_list; /* add to list of virtual arrays */
+  mem->virt_sarray_list = result;
+
+  return result;
+}
+
+
+METHODDEF(jvirt_barray_ptr)
+request_virt_barray (j_common_ptr cinfo, int pool_id, boolean pre_zero,
+		     JDIMENSION blocksperrow, JDIMENSION numrows,
+		     JDIMENSION maxaccess)
+/* Request a virtual 2-D coefficient-block array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  jvirt_barray_ptr result;
+
+  /* Only IMAGE-lifetime virtual arrays are currently supported */
+  if (pool_id != JPOOL_IMAGE)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+  /* get control block */
+  result = (jvirt_barray_ptr) alloc_small(cinfo, pool_id,
+					  SIZEOF(struct jvirt_barray_control));
+
+  result->mem_buffer = NULL;	/* marks array not yet realized */
+  result->rows_in_array = numrows;
+  result->blocksperrow = blocksperrow;
+  result->maxaccess = maxaccess;
+  result->pre_zero = pre_zero;
+  result->b_s_open = FALSE;	/* no associated backing-store object */
+  result->next = mem->virt_barray_list; /* add to list of virtual arrays */
+  mem->virt_barray_list = result;
+
+  return result;
+}
+
+
+METHODDEF(void)
+realize_virt_arrays (j_common_ptr cinfo)
+/* Allocate the in-memory buffers for any unrealized virtual arrays */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  size_t space_per_minheight, maximum_space, avail_mem;
+  size_t minheights, max_minheights;
+  jvirt_sarray_ptr sptr;
+  jvirt_barray_ptr bptr;
+
+  /* Compute the minimum space needed (maxaccess rows in each buffer)
+   * and the maximum space needed (full image height in each buffer).
+   * These may be of use to the system-dependent jpeg_mem_available routine.
+   */
+  space_per_minheight = 0;
+  maximum_space = 0;
+  for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
+    if (sptr->mem_buffer == NULL) { /* if not realized yet */
+      space_per_minheight += (long) sptr->maxaccess *
+			     (long) sptr->samplesperrow * SIZEOF(JSAMPLE);
+      maximum_space += (long) sptr->rows_in_array *
+		       (long) sptr->samplesperrow * SIZEOF(JSAMPLE);
+    }
+  }
+  for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
+    if (bptr->mem_buffer == NULL) { /* if not realized yet */
+      space_per_minheight += (long) bptr->maxaccess *
+			     (long) bptr->blocksperrow * SIZEOF(JBLOCK);
+      maximum_space += (long) bptr->rows_in_array *
+		       (long) bptr->blocksperrow * SIZEOF(JBLOCK);
+    }
+  }
+
+  if (space_per_minheight <= 0)
+    return;			/* no unrealized arrays, no work */
+
+  /* Determine amount of memory to actually use; this is system-dependent. */
+  avail_mem = jpeg_mem_available(cinfo, space_per_minheight, maximum_space,
+				 mem->total_space_allocated);
+
+  /* If the maximum space needed is available, make all the buffers full
+   * height; otherwise parcel it out with the same number of minheights
+   * in each buffer.
+   */
+  if (avail_mem >= maximum_space)
+    max_minheights = 1000000000L;
+  else {
+    max_minheights = avail_mem / space_per_minheight;
+    /* If there doesn't seem to be enough space, try to get the minimum
+     * anyway.  This allows a "stub" implementation of jpeg_mem_available().
+     */
+    if (max_minheights <= 0)
+      max_minheights = 1;
+  }
+
+  /* Allocate the in-memory buffers and initialize backing store as needed. */
+
+  for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
+    if (sptr->mem_buffer == NULL) { /* if not realized yet */
+      minheights = ((long) sptr->rows_in_array - 1L) / sptr->maxaccess + 1L;
+      if (minheights <= max_minheights) {
+	/* This buffer fits in memory */
+	sptr->rows_in_mem = sptr->rows_in_array;
+      } else {
+	/* It doesn't fit in memory, create backing store. */
+	sptr->rows_in_mem = (JDIMENSION) (max_minheights * sptr->maxaccess);
+	jpeg_open_backing_store(cinfo, & sptr->b_s_info,
+				(long) sptr->rows_in_array *
+				(long) sptr->samplesperrow *
+				(long) SIZEOF(JSAMPLE));
+	sptr->b_s_open = TRUE;
+      }
+      sptr->mem_buffer = alloc_sarray(cinfo, JPOOL_IMAGE,
+				      sptr->samplesperrow, sptr->rows_in_mem);
+      sptr->rowsperchunk = mem->last_rowsperchunk;
+      sptr->cur_start_row = 0;
+      sptr->first_undef_row = 0;
+      sptr->dirty = FALSE;
+    }
+  }
+
+  for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
+    if (bptr->mem_buffer == NULL) { /* if not realized yet */
+      minheights = ((long) bptr->rows_in_array - 1L) / bptr->maxaccess + 1L;
+      if (minheights <= max_minheights) {
+	/* This buffer fits in memory */
+	bptr->rows_in_mem = bptr->rows_in_array;
+      } else {
+	/* It doesn't fit in memory, create backing store. */
+	bptr->rows_in_mem = (JDIMENSION) (max_minheights * bptr->maxaccess);
+	jpeg_open_backing_store(cinfo, & bptr->b_s_info,
+				(long) bptr->rows_in_array *
+				(long) bptr->blocksperrow *
+				(long) SIZEOF(JBLOCK));
+	bptr->b_s_open = TRUE;
+      }
+      bptr->mem_buffer = alloc_barray(cinfo, JPOOL_IMAGE,
+				      bptr->blocksperrow, bptr->rows_in_mem);
+      bptr->rowsperchunk = mem->last_rowsperchunk;
+      bptr->cur_start_row = 0;
+      bptr->first_undef_row = 0;
+      bptr->dirty = FALSE;
+    }
+  }
+}
+
+
+LOCAL(void)
+do_sarray_io (j_common_ptr cinfo, jvirt_sarray_ptr ptr, boolean writing)
+/* Do backing store read or write of a virtual sample array */
+{
+  long bytesperrow, file_offset, byte_count, rows, thisrow, i;
+
+  bytesperrow = (long) ptr->samplesperrow * SIZEOF(JSAMPLE);
+  file_offset = ptr->cur_start_row * bytesperrow;
+  /* Loop to read or write each allocation chunk in mem_buffer */
+  for (i = 0; i < (long) ptr->rows_in_mem; i += ptr->rowsperchunk) {
+    /* One chunk, but check for short chunk at end of buffer */
+    rows = MIN((long) ptr->rowsperchunk, (long) ptr->rows_in_mem - i);
+    /* Transfer no more than is currently defined */
+    thisrow = (long) ptr->cur_start_row + i;
+    rows = MIN(rows, (long) ptr->first_undef_row - thisrow);
+    /* Transfer no more than fits in file */
+    rows = MIN(rows, (long) ptr->rows_in_array - thisrow);
+    if (rows <= 0)		/* this chunk might be past end of file! */
+      break;
+    byte_count = rows * bytesperrow;
+    if (writing)
+      (*ptr->b_s_info.write_backing_store) (cinfo, & ptr->b_s_info,
+					    (void FAR *) ptr->mem_buffer[i],
+					    file_offset, byte_count);
+    else
+      (*ptr->b_s_info.read_backing_store) (cinfo, & ptr->b_s_info,
+					   (void FAR *) ptr->mem_buffer[i],
+					   file_offset, byte_count);
+    file_offset += byte_count;
+  }
+}
+
+
+LOCAL(void)
+do_barray_io (j_common_ptr cinfo, jvirt_barray_ptr ptr, boolean writing)
+/* Do backing store read or write of a virtual coefficient-block array */
+{
+  long bytesperrow, file_offset, byte_count, rows, thisrow, i;
+
+  bytesperrow = (long) ptr->blocksperrow * SIZEOF(JBLOCK);
+  file_offset = ptr->cur_start_row * bytesperrow;
+  /* Loop to read or write each allocation chunk in mem_buffer */
+  for (i = 0; i < (long) ptr->rows_in_mem; i += ptr->rowsperchunk) {
+    /* One chunk, but check for short chunk at end of buffer */
+    rows = MIN((long) ptr->rowsperchunk, (long) ptr->rows_in_mem - i);
+    /* Transfer no more than is currently defined */
+    thisrow = (long) ptr->cur_start_row + i;
+    rows = MIN(rows, (long) ptr->first_undef_row - thisrow);
+    /* Transfer no more than fits in file */
+    rows = MIN(rows, (long) ptr->rows_in_array - thisrow);
+    if (rows <= 0)		/* this chunk might be past end of file! */
+      break;
+    byte_count = rows * bytesperrow;
+    if (writing)
+      (*ptr->b_s_info.write_backing_store) (cinfo, & ptr->b_s_info,
+					    (void FAR *) ptr->mem_buffer[i],
+					    file_offset, byte_count);
+    else
+      (*ptr->b_s_info.read_backing_store) (cinfo, & ptr->b_s_info,
+					   (void FAR *) ptr->mem_buffer[i],
+					   file_offset, byte_count);
+    file_offset += byte_count;
+  }
+}
+
+
+METHODDEF(JSAMPARRAY)
+access_virt_sarray (j_common_ptr cinfo, jvirt_sarray_ptr ptr,
+		    JDIMENSION start_row, JDIMENSION num_rows,
+		    boolean writable)
+/* Access the part of a virtual sample array starting at start_row */
+/* and extending for num_rows rows.  writable is true if  */
+/* caller intends to modify the accessed area. */
+{
+  JDIMENSION end_row = start_row + num_rows;
+  JDIMENSION undef_row;
+
+  /* debugging check */
+  if (end_row > ptr->rows_in_array || num_rows > ptr->maxaccess ||
+      ptr->mem_buffer == NULL)
+    ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+
+  /* Make the desired part of the virtual array accessible */
+  if (start_row < ptr->cur_start_row ||
+      end_row > ptr->cur_start_row+ptr->rows_in_mem) {
+    if (! ptr->b_s_open)
+      ERREXIT(cinfo, JERR_VIRTUAL_BUG);
+    /* Flush old buffer contents if necessary */
+    if (ptr->dirty) {
+      do_sarray_io(cinfo, ptr, TRUE);
+      ptr->dirty = FALSE;
+    }
+    /* Decide what part of virtual array to access.
+     * Algorithm: if target address > current window, assume forward scan,
+     * load starting at target address.  If target address < current window,
+     * assume backward scan, load so that target area is top of window.
+     * Note that when switching from forward write to forward read, will have
+     * start_row = 0, so the limiting case applies and we load from 0 anyway.
+     */
+    if (start_row > ptr->cur_start_row) {
+      ptr->cur_start_row = start_row;
+    } else {
+      /* use long arithmetic here to avoid overflow & unsigned problems */
+      long ltemp;
+
+      ltemp = (long) end_row - (long) ptr->rows_in_mem;
+      if (ltemp < 0)
+	ltemp = 0;		/* don't fall off front end of file */
+      ptr->cur_start_row = (JDIMENSION) ltemp;
+    }
+    /* Read in the selected part of the array.
+     * During the initial write pass, we will do no actual read
+     * because the selected part is all undefined.
+     */
+    do_sarray_io(cinfo, ptr, FALSE);
+  }
+  /* Ensure the accessed part of the array is defined; prezero if needed.
+   * To improve locality of access, we only prezero the part of the array
+   * that the caller is about to access, not the entire in-memory array.
+   */
+  if (ptr->first_undef_row < end_row) {
+    if (ptr->first_undef_row < start_row) {
+      if (writable)		/* writer skipped over a section of array */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+      undef_row = start_row;	/* but reader is allowed to read ahead */
+    } else {
+      undef_row = ptr->first_undef_row;
+    }
+    if (writable)
+      ptr->first_undef_row = end_row;
+    if (ptr->pre_zero) {
+      size_t bytesperrow = (size_t) ptr->samplesperrow * SIZEOF(JSAMPLE);
+      undef_row -= ptr->cur_start_row; /* make indexes relative to buffer */
+      end_row -= ptr->cur_start_row;
+      while (undef_row < end_row) {
+	jzero_far((void FAR *) ptr->mem_buffer[undef_row], bytesperrow);
+	undef_row++;
+      }
+    } else {
+      if (! writable)		/* reader looking at undefined data */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+    }
+  }
+  /* Flag the buffer dirty if caller will write in it */
+  if (writable)
+    ptr->dirty = TRUE;
+  /* Return address of proper part of the buffer */
+  return ptr->mem_buffer + (start_row - ptr->cur_start_row);
+}
+
+
+METHODDEF(JBLOCKARRAY)
+access_virt_barray (j_common_ptr cinfo, jvirt_barray_ptr ptr,
+		    JDIMENSION start_row, JDIMENSION num_rows,
+		    boolean writable)
+/* Access the part of a virtual block array starting at start_row */
+/* and extending for num_rows rows.  writable is true if  */
+/* caller intends to modify the accessed area. */
+{
+  JDIMENSION end_row = start_row + num_rows;
+  JDIMENSION undef_row;
+
+  /* debugging check */
+  if (end_row > ptr->rows_in_array || num_rows > ptr->maxaccess ||
+      ptr->mem_buffer == NULL)
+    ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+
+  /* Make the desired part of the virtual array accessible */
+  if (start_row < ptr->cur_start_row ||
+      end_row > ptr->cur_start_row+ptr->rows_in_mem) {
+    if (! ptr->b_s_open)
+      ERREXIT(cinfo, JERR_VIRTUAL_BUG);
+    /* Flush old buffer contents if necessary */
+    if (ptr->dirty) {
+      do_barray_io(cinfo, ptr, TRUE);
+      ptr->dirty = FALSE;
+    }
+    /* Decide what part of virtual array to access.
+     * Algorithm: if target address > current window, assume forward scan,
+     * load starting at target address.  If target address < current window,
+     * assume backward scan, load so that target area is top of window.
+     * Note that when switching from forward write to forward read, will have
+     * start_row = 0, so the limiting case applies and we load from 0 anyway.
+     */
+    if (start_row > ptr->cur_start_row) {
+      ptr->cur_start_row = start_row;
+    } else {
+      /* use long arithmetic here to avoid overflow & unsigned problems */
+      long ltemp;
+
+      ltemp = (long) end_row - (long) ptr->rows_in_mem;
+      if (ltemp < 0)
+	ltemp = 0;		/* don't fall off front end of file */
+      ptr->cur_start_row = (JDIMENSION) ltemp;
+    }
+    /* Read in the selected part of the array.
+     * During the initial write pass, we will do no actual read
+     * because the selected part is all undefined.
+     */
+    do_barray_io(cinfo, ptr, FALSE);
+  }
+  /* Ensure the accessed part of the array is defined; prezero if needed.
+   * To improve locality of access, we only prezero the part of the array
+   * that the caller is about to access, not the entire in-memory array.
+   */
+  if (ptr->first_undef_row < end_row) {
+    if (ptr->first_undef_row < start_row) {
+      if (writable)		/* writer skipped over a section of array */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+      undef_row = start_row;	/* but reader is allowed to read ahead */
+    } else {
+      undef_row = ptr->first_undef_row;
+    }
+    if (writable)
+      ptr->first_undef_row = end_row;
+    if (ptr->pre_zero) {
+      size_t bytesperrow = (size_t) ptr->blocksperrow * SIZEOF(JBLOCK);
+      undef_row -= ptr->cur_start_row; /* make indexes relative to buffer */
+      end_row -= ptr->cur_start_row;
+      while (undef_row < end_row) {
+	jzero_far((void FAR *) ptr->mem_buffer[undef_row], bytesperrow);
+	undef_row++;
+      }
+    } else {
+      if (! writable)		/* reader looking at undefined data */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+    }
+  }
+  /* Flag the buffer dirty if caller will write in it */
+  if (writable)
+    ptr->dirty = TRUE;
+  /* Return address of proper part of the buffer */
+  return ptr->mem_buffer + (start_row - ptr->cur_start_row);
+}
+
+
+/*
+ * Release all objects belonging to a specified pool.
+ */
+
+METHODDEF(void)
+free_pool (j_common_ptr cinfo, int pool_id)
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  small_pool_ptr shdr_ptr;
+  large_pool_ptr lhdr_ptr;
+  size_t space_freed;
+
+  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+#ifdef MEM_STATS
+  if (cinfo->err->trace_level > 1)
+    print_mem_stats(cinfo, pool_id); /* print pool's memory usage statistics */
+#endif
+
+  /* If freeing IMAGE pool, close any virtual arrays first */
+  if (pool_id == JPOOL_IMAGE) {
+    jvirt_sarray_ptr sptr;
+    jvirt_barray_ptr bptr;
+
+    for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
+      if (sptr->b_s_open) {	/* there may be no backing store */
+	sptr->b_s_open = FALSE;	/* prevent recursive close if error */
+	(*sptr->b_s_info.close_backing_store) (cinfo, & sptr->b_s_info);
+      }
+    }
+    mem->virt_sarray_list = NULL;
+    for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
+      if (bptr->b_s_open) {	/* there may be no backing store */
+	bptr->b_s_open = FALSE;	/* prevent recursive close if error */
+	(*bptr->b_s_info.close_backing_store) (cinfo, & bptr->b_s_info);
+      }
+    }
+    mem->virt_barray_list = NULL;
+  }
+
+  /* Release large objects */
+  lhdr_ptr = mem->large_list[pool_id];
+  mem->large_list[pool_id] = NULL;
+
+  while (lhdr_ptr != NULL) {
+    large_pool_ptr next_lhdr_ptr = lhdr_ptr->next;
+    space_freed = lhdr_ptr->bytes_used +
+		  lhdr_ptr->bytes_left +
+		  SIZEOF(large_pool_hdr);
+    jpeg_free_large(cinfo, (void FAR *) lhdr_ptr, space_freed);
+    mem->total_space_allocated -= space_freed;
+    lhdr_ptr = next_lhdr_ptr;
+  }
+
+  /* Release small objects */
+  shdr_ptr = mem->small_list[pool_id];
+  mem->small_list[pool_id] = NULL;
+
+  while (shdr_ptr != NULL) {
+    small_pool_ptr next_shdr_ptr = shdr_ptr->next;
+    space_freed = shdr_ptr->bytes_used +
+		  shdr_ptr->bytes_left +
+		  SIZEOF(small_pool_hdr);
+    jpeg_free_small(cinfo, (void *) shdr_ptr, space_freed);
+    mem->total_space_allocated -= space_freed;
+    shdr_ptr = next_shdr_ptr;
+  }
+}
+
+
+/*
+ * Close up shop entirely.
+ * Note that this cannot be called unless cinfo->mem is non-NULL.
+ */
+
+METHODDEF(void)
+self_destruct (j_common_ptr cinfo)
+{
+  int pool;
+
+  /* Close all backing store, release all memory.
+   * Releasing pools in reverse order might help avoid fragmentation
+   * with some (brain-damaged) malloc libraries.
+   */
+  for (pool = JPOOL_NUMPOOLS-1; pool >= JPOOL_PERMANENT; pool--) {
+    free_pool(cinfo, pool);
+  }
+
+  /* Release the memory manager control block too. */
+  jpeg_free_small(cinfo, (void *) cinfo->mem, SIZEOF(my_memory_mgr));
+  cinfo->mem = NULL;		/* ensures I will be called only once */
+
+  jpeg_mem_term(cinfo);		/* system-dependent cleanup */
+}
+
+
+/*
+ * Memory manager initialization.
+ * When this is called, only the error manager pointer is valid in cinfo!
+ */
+
+GLOBAL(void)
+jinit_memory_mgr (j_common_ptr cinfo)
+{
+  my_mem_ptr mem;
+  long max_to_use;
+  int pool;
+  size_t test_mac;
+
+  cinfo->mem = NULL;		/* for safety if init fails */
+
+  /* Check for configuration errors.
+   * SIZEOF(ALIGN_TYPE) should be a power of 2; otherwise, it probably
+   * doesn't reflect any real hardware alignment requirement.
+   * The test is a little tricky: for X>0, X and X-1 have no one-bits
+   * in common if and only if X is a power of 2, ie has only one one-bit.
+   * Some compilers may give an "unreachable code" warning here; ignore it.
+   */
+  if ((ALIGN_SIZE & (ALIGN_SIZE-1)) != 0)
+    ERREXIT(cinfo, JERR_BAD_ALIGN_TYPE);
+  /* MAX_ALLOC_CHUNK must be representable as type size_t, and must be
+   * a multiple of ALIGN_SIZE.
+   * Again, an "unreachable code" warning may be ignored here.
+   * But a "constant too large" warning means you need to fix MAX_ALLOC_CHUNK.
+   */
+  test_mac = (size_t) MAX_ALLOC_CHUNK;
+  if ((long) test_mac != MAX_ALLOC_CHUNK ||
+      (MAX_ALLOC_CHUNK % ALIGN_SIZE) != 0)
+    ERREXIT(cinfo, JERR_BAD_ALLOC_CHUNK);
+
+  max_to_use = jpeg_mem_init(cinfo); /* system-dependent initialization */
+
+  /* Attempt to allocate memory manager's control block */
+  mem = (my_mem_ptr) jpeg_get_small(cinfo, SIZEOF(my_memory_mgr));
+
+  if (mem == NULL) {
+    jpeg_mem_term(cinfo);	/* system-dependent cleanup */
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 0);
+  }
+
+  /* OK, fill in the method pointers */
+  mem->pub.alloc_small = alloc_small;
+  mem->pub.alloc_large = alloc_large;
+  mem->pub.alloc_sarray = alloc_sarray;
+  mem->pub.alloc_barray = alloc_barray;
+  mem->pub.request_virt_sarray = request_virt_sarray;
+  mem->pub.request_virt_barray = request_virt_barray;
+  mem->pub.realize_virt_arrays = realize_virt_arrays;
+  mem->pub.access_virt_sarray = access_virt_sarray;
+  mem->pub.access_virt_barray = access_virt_barray;
+  mem->pub.free_pool = free_pool;
+  mem->pub.self_destruct = self_destruct;
+
+  /* Make MAX_ALLOC_CHUNK accessible to other modules */
+  mem->pub.max_alloc_chunk = MAX_ALLOC_CHUNK;
+
+  /* Initialize working state */
+  mem->pub.max_memory_to_use = max_to_use;
+
+  for (pool = JPOOL_NUMPOOLS-1; pool >= JPOOL_PERMANENT; pool--) {
+    mem->small_list[pool] = NULL;
+    mem->large_list[pool] = NULL;
+  }
+  mem->virt_sarray_list = NULL;
+  mem->virt_barray_list = NULL;
+
+  mem->total_space_allocated = SIZEOF(my_memory_mgr);
+
+  /* Declare ourselves open for business */
+  cinfo->mem = & mem->pub;
+
+  /* Check for an environment variable JPEGMEM; if found, override the
+   * default max_memory setting from jpeg_mem_init.  Note that the
+   * surrounding application may again override this value.
+   * If your system doesn't support getenv(), define NO_GETENV to disable
+   * this feature.
+   */
+#ifndef NO_GETENV
+  { char * memenv;
+
+    if ((memenv = getenv("JPEGMEM")) != NULL) {
+      char ch = 'x';
+
+      if (sscanf(memenv, "%ld%c", &max_to_use, &ch) > 0) {
+	if (ch == 'm' || ch == 'M')
+	  max_to_use *= 1000L;
+	mem->pub.max_memory_to_use = max_to_use * 1000L;
+      }
+    }
+  }
+#endif
+
+}
diff --git a/jmemnobs.c b/jmemnobs.c
new file mode 100644
index 0000000..34b1895
--- /dev/null
+++ b/jmemnobs.c
@@ -0,0 +1,109 @@
+/*
+ * jmemnobs.c
+ *
+ * Copyright (C) 1992-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file provides a really simple implementation of the system-
+ * dependent portion of the JPEG memory manager.  This implementation
+ * assumes that no backing-store files are needed: all required space
+ * can be obtained from malloc().
+ * This is very portable in the sense that it'll compile on almost anything,
+ * but you'd better have lots of main memory (or virtual memory) if you want
+ * to process big images.
+ * Note that the max_memory_to_use option is ignored by this implementation.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jmemsys.h"		/* import the system-dependent declarations */
+
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare malloc(),free() */
+extern void * malloc JPP((size_t size));
+extern void free JPP((void *ptr));
+#endif
+
+
+/*
+ * Memory allocation and freeing are controlled by the regular library
+ * routines malloc() and free().
+ */
+
+GLOBAL(void *)
+jpeg_get_small (j_common_ptr cinfo, size_t sizeofobject)
+{
+  return (void *) malloc(sizeofobject);
+}
+
+GLOBAL(void)
+jpeg_free_small (j_common_ptr cinfo, void * object, size_t sizeofobject)
+{
+  free(object);
+}
+
+
+/*
+ * "Large" objects are treated the same as "small" ones.
+ * NB: although we include FAR keywords in the routine declarations,
+ * this file won't actually work in 80x86 small/medium model; at least,
+ * you probably won't be able to process useful-size images in only 64KB.
+ */
+
+GLOBAL(void FAR *)
+jpeg_get_large (j_common_ptr cinfo, size_t sizeofobject)
+{
+  return (void FAR *) malloc(sizeofobject);
+}
+
+GLOBAL(void)
+jpeg_free_large (j_common_ptr cinfo, void FAR * object, size_t sizeofobject)
+{
+  free(object);
+}
+
+
+/*
+ * This routine computes the total memory space available for allocation.
+ * Here we always say, "we got all you want bud!"
+ */
+
+GLOBAL(size_t)
+jpeg_mem_available (j_common_ptr cinfo, size_t min_bytes_needed,
+		    size_t max_bytes_needed, size_t already_allocated)
+{
+  return max_bytes_needed;
+}
+
+
+/*
+ * Backing store (temporary file) management.
+ * Since jpeg_mem_available always promised the moon,
+ * this should never be called and we can just error out.
+ */
+
+GLOBAL(void)
+jpeg_open_backing_store (j_common_ptr cinfo, backing_store_ptr info,
+			 long total_bytes_needed)
+{
+  ERREXIT(cinfo, JERR_NO_BACKING_STORE);
+}
+
+
+/*
+ * These routines take care of any system-dependent initialization and
+ * cleanup required.  Here, there isn't any.
+ */
+
+GLOBAL(long)
+jpeg_mem_init (j_common_ptr cinfo)
+{
+  return 0;			/* just set max_memory_to_use to 0 */
+}
+
+GLOBAL(void)
+jpeg_mem_term (j_common_ptr cinfo)
+{
+  /* no work */
+}
diff --git a/jmemsys.h b/jmemsys.h
new file mode 100644
index 0000000..b190945
--- /dev/null
+++ b/jmemsys.h
@@ -0,0 +1,198 @@
+/*
+ * jmemsys.h
+ *
+ * Copyright (C) 1992-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This include file defines the interface between the system-independent
+ * and system-dependent portions of the JPEG memory manager.  No other
+ * modules need include it.  (The system-independent portion is jmemmgr.c;
+ * there are several different versions of the system-dependent portion.)
+ *
+ * This file works as-is for the system-dependent memory managers supplied
+ * in the IJG distribution.  You may need to modify it if you write a
+ * custom memory manager.  If system-dependent changes are needed in
+ * this file, the best method is to #ifdef them based on a configuration
+ * symbol supplied in jconfig.h, as we have done with USE_MSDOS_MEMMGR
+ * and USE_MAC_MEMMGR.
+ */
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_get_small		jGetSmall
+#define jpeg_free_small		jFreeSmall
+#define jpeg_get_large		jGetLarge
+#define jpeg_free_large		jFreeLarge
+#define jpeg_mem_available	jMemAvail
+#define jpeg_open_backing_store	jOpenBackStore
+#define jpeg_mem_init		jMemInit
+#define jpeg_mem_term		jMemTerm
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/*
+ * These two functions are used to allocate and release small chunks of
+ * memory.  (Typically the total amount requested through jpeg_get_small is
+ * no more than 20K or so; this will be requested in chunks of a few K each.)
+ * Behavior should be the same as for the standard library functions malloc
+ * and free; in particular, jpeg_get_small must return NULL on failure.
+ * On most systems, these ARE malloc and free.  jpeg_free_small is passed the
+ * size of the object being freed, just in case it's needed.
+ * On an 80x86 machine using small-data memory model, these manage near heap.
+ */
+
+EXTERN(void *) jpeg_get_small JPP((j_common_ptr cinfo, size_t sizeofobject));
+EXTERN(void) jpeg_free_small JPP((j_common_ptr cinfo, void * object,
+				  size_t sizeofobject));
+
+/*
+ * These two functions are used to allocate and release large chunks of
+ * memory (up to the total free space designated by jpeg_mem_available).
+ * The interface is the same as above, except that on an 80x86 machine,
+ * far pointers are used.  On most other machines these are identical to
+ * the jpeg_get/free_small routines; but we keep them separate anyway,
+ * in case a different allocation strategy is desirable for large chunks.
+ */
+
+EXTERN(void FAR *) jpeg_get_large JPP((j_common_ptr cinfo,
+				       size_t sizeofobject));
+EXTERN(void) jpeg_free_large JPP((j_common_ptr cinfo, void FAR * object,
+				  size_t sizeofobject));
+
+/*
+ * The macro MAX_ALLOC_CHUNK designates the maximum number of bytes that may
+ * be requested in a single call to jpeg_get_large (and jpeg_get_small for that
+ * matter, but that case should never come into play).  This macro is needed
+ * to model the 64Kb-segment-size limit of far addressing on 80x86 machines.
+ * On those machines, we expect that jconfig.h will provide a proper value.
+ * On machines with 32-bit flat address spaces, any large constant may be used.
+ *
+ * NB: jmemmgr.c expects that MAX_ALLOC_CHUNK will be representable as type
+ * size_t and will be a multiple of sizeof(align_type).
+ */
+
+#ifndef MAX_ALLOC_CHUNK		/* may be overridden in jconfig.h */
+#define MAX_ALLOC_CHUNK  1000000000L
+#endif
+
+/*
+ * This routine computes the total space still available for allocation by
+ * jpeg_get_large.  If more space than this is needed, backing store will be
+ * used.  NOTE: any memory already allocated must not be counted.
+ *
+ * There is a minimum space requirement, corresponding to the minimum
+ * feasible buffer sizes; jmemmgr.c will request that much space even if
+ * jpeg_mem_available returns zero.  The maximum space needed, enough to hold
+ * all working storage in memory, is also passed in case it is useful.
+ * Finally, the total space already allocated is passed.  If no better
+ * method is available, cinfo->mem->max_memory_to_use - already_allocated
+ * is often a suitable calculation.
+ *
+ * It is OK for jpeg_mem_available to underestimate the space available
+ * (that'll just lead to more backing-store access than is really necessary).
+ * However, an overestimate will lead to failure.  Hence it's wise to subtract
+ * a slop factor from the true available space.  5% should be enough.
+ *
+ * On machines with lots of virtual memory, any large constant may be returned.
+ * Conversely, zero may be returned to always use the minimum amount of memory.
+ */
+
+EXTERN(size_t) jpeg_mem_available JPP((j_common_ptr cinfo,
+				     size_t min_bytes_needed,
+				     size_t max_bytes_needed,
+				     size_t already_allocated));
+
+
+/*
+ * This structure holds whatever state is needed to access a single
+ * backing-store object.  The read/write/close method pointers are called
+ * by jmemmgr.c to manipulate the backing-store object; all other fields
+ * are private to the system-dependent backing store routines.
+ */
+
+#define TEMP_NAME_LENGTH   64	/* max length of a temporary file's name */
+
+
+#ifdef USE_MSDOS_MEMMGR		/* DOS-specific junk */
+
+typedef unsigned short XMSH;	/* type of extended-memory handles */
+typedef unsigned short EMSH;	/* type of expanded-memory handles */
+
+typedef union {
+  short file_handle;		/* DOS file handle if it's a temp file */
+  XMSH xms_handle;		/* handle if it's a chunk of XMS */
+  EMSH ems_handle;		/* handle if it's a chunk of EMS */
+} handle_union;
+
+#endif /* USE_MSDOS_MEMMGR */
+
+#ifdef USE_MAC_MEMMGR		/* Mac-specific junk */
+#include <Files.h>
+#endif /* USE_MAC_MEMMGR */
+
+
+typedef struct backing_store_struct * backing_store_ptr;
+
+typedef struct backing_store_struct {
+  /* Methods for reading/writing/closing this backing-store object */
+  JMETHOD(void, read_backing_store, (j_common_ptr cinfo,
+				     backing_store_ptr info,
+				     void FAR * buffer_address,
+				     long file_offset, long byte_count));
+  JMETHOD(void, write_backing_store, (j_common_ptr cinfo,
+				      backing_store_ptr info,
+				      void FAR * buffer_address,
+				      long file_offset, long byte_count));
+  JMETHOD(void, close_backing_store, (j_common_ptr cinfo,
+				      backing_store_ptr info));
+
+  /* Private fields for system-dependent backing-store management */
+#ifdef USE_MSDOS_MEMMGR
+  /* For the MS-DOS manager (jmemdos.c), we need: */
+  handle_union handle;		/* reference to backing-store storage object */
+  char temp_name[TEMP_NAME_LENGTH]; /* name if it's a file */
+#else
+#ifdef USE_MAC_MEMMGR
+  /* For the Mac manager (jmemmac.c), we need: */
+  short temp_file;		/* file reference number to temp file */
+  FSSpec tempSpec;		/* the FSSpec for the temp file */
+  char temp_name[TEMP_NAME_LENGTH]; /* name if it's a file */
+#else
+  /* For a typical implementation with temp files, we need: */
+  FILE * temp_file;		/* stdio reference to temp file */
+  char temp_name[TEMP_NAME_LENGTH]; /* name of temp file */
+#endif
+#endif
+} backing_store_info;
+
+
+/*
+ * Initial opening of a backing-store object.  This must fill in the
+ * read/write/close pointers in the object.  The read/write routines
+ * may take an error exit if the specified maximum file size is exceeded.
+ * (If jpeg_mem_available always returns a large value, this routine can
+ * just take an error exit.)
+ */
+
+EXTERN(void) jpeg_open_backing_store JPP((j_common_ptr cinfo,
+					  backing_store_ptr info,
+					  long total_bytes_needed));
+
+
+/*
+ * These routines take care of any system-dependent initialization and
+ * cleanup required.  jpeg_mem_init will be called before anything is
+ * allocated (and, therefore, nothing in cinfo is of use except the error
+ * manager pointer).  It should return a suitable default value for
+ * max_memory_to_use; this may subsequently be overridden by the surrounding
+ * application.  (Note that max_memory_to_use is only important if
+ * jpeg_mem_available chooses to consult it ... no one else will.)
+ * jpeg_mem_term may assume that all requested memory has been freed and that
+ * all opened backing-store objects have been closed.
+ */
+
+EXTERN(long) jpeg_mem_init JPP((j_common_ptr cinfo));
+EXTERN(void) jpeg_mem_term JPP((j_common_ptr cinfo));
diff --git a/jmorecfg.h b/jmorecfg.h
new file mode 100644
index 0000000..0e7fb72
--- /dev/null
+++ b/jmorecfg.h
@@ -0,0 +1,370 @@
+/*
+ * jmorecfg.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Copyright (C) 2009, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains additional configuration options that customize the
+ * JPEG software for special applications or support machine-dependent
+ * optimizations.  Most users will not need to touch this file.
+ */
+
+
+/*
+ * Define BITS_IN_JSAMPLE as either
+ *   8   for 8-bit sample values (the usual setting)
+ *   12  for 12-bit sample values
+ * Only 8 and 12 are legal data precisions for lossy JPEG according to the
+ * JPEG standard, and the IJG code does not support anything else!
+ * We do not support run-time selection of data precision, sorry.
+ */
+
+#define BITS_IN_JSAMPLE  8	/* use 8 or 12 */
+
+
+/*
+ * Maximum number of components (color channels) allowed in JPEG image.
+ * To meet the letter of the JPEG spec, set this to 255.  However, darn
+ * few applications need more than 4 channels (maybe 5 for CMYK + alpha
+ * mask).  We recommend 10 as a reasonable compromise; use 4 if you are
+ * really short on memory.  (Each allowed component costs a hundred or so
+ * bytes of storage, whether actually used in an image or not.)
+ */
+
+#define MAX_COMPONENTS  10	/* maximum number of image components */
+
+
+/*
+ * Basic data types.
+ * You may need to change these if you have a machine with unusual data
+ * type sizes; for example, "char" not 8 bits, "short" not 16 bits,
+ * or "long" not 32 bits.  We don't care whether "int" is 16 or 32 bits,
+ * but it had better be at least 16.
+ */
+
+/* Representation of a single sample (pixel element value).
+ * We frequently allocate large arrays of these, so it's important to keep
+ * them small.  But if you have memory to burn and access to char or short
+ * arrays is very slow on your hardware, you might want to change these.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+/* JSAMPLE should be the smallest type that will hold the values 0..255.
+ * You can use a signed char by having GETJSAMPLE mask it with 0xFF.
+ */
+
+#ifdef HAVE_UNSIGNED_CHAR
+
+typedef unsigned char JSAMPLE;
+#define GETJSAMPLE(value)  ((int) (value))
+
+#else /* not HAVE_UNSIGNED_CHAR */
+
+typedef char JSAMPLE;
+#ifdef __CHAR_UNSIGNED__
+#define GETJSAMPLE(value)  ((int) (value))
+#else
+#define GETJSAMPLE(value)  ((int) (value) & 0xFF)
+#endif /* __CHAR_UNSIGNED__ */
+
+#endif /* HAVE_UNSIGNED_CHAR */
+
+#define MAXJSAMPLE	255
+#define CENTERJSAMPLE	128
+
+#endif /* BITS_IN_JSAMPLE == 8 */
+
+
+#if BITS_IN_JSAMPLE == 12
+/* JSAMPLE should be the smallest type that will hold the values 0..4095.
+ * On nearly all machines "short" will do nicely.
+ */
+
+typedef short JSAMPLE;
+#define GETJSAMPLE(value)  ((int) (value))
+
+#define MAXJSAMPLE	4095
+#define CENTERJSAMPLE	2048
+
+#endif /* BITS_IN_JSAMPLE == 12 */
+
+
+/* Representation of a DCT frequency coefficient.
+ * This should be a signed value of at least 16 bits; "short" is usually OK.
+ * Again, we allocate large arrays of these, but you can change to int
+ * if you have memory to burn and "short" is really slow.
+ */
+
+typedef short JCOEF;
+
+
+/* Compressed datastreams are represented as arrays of JOCTET.
+ * These must be EXACTLY 8 bits wide, at least once they are written to
+ * external storage.  Note that when using the stdio data source/destination
+ * managers, this is also the data type passed to fread/fwrite.
+ */
+
+#ifdef HAVE_UNSIGNED_CHAR
+
+typedef unsigned char JOCTET;
+#define GETJOCTET(value)  (value)
+
+#else /* not HAVE_UNSIGNED_CHAR */
+
+typedef char JOCTET;
+#ifdef __CHAR_UNSIGNED__
+#define GETJOCTET(value)  (value)
+#else
+#define GETJOCTET(value)  ((value) & 0xFF)
+#endif /* __CHAR_UNSIGNED__ */
+
+#endif /* HAVE_UNSIGNED_CHAR */
+
+
+/* These typedefs are used for various table entries and so forth.
+ * They must be at least as wide as specified; but making them too big
+ * won't cost a huge amount of memory, so we don't provide special
+ * extraction code like we did for JSAMPLE.  (In other words, these
+ * typedefs live at a different point on the speed/space tradeoff curve.)
+ */
+
+/* UINT8 must hold at least the values 0..255. */
+
+#ifdef HAVE_UNSIGNED_CHAR
+typedef unsigned char UINT8;
+#else /* not HAVE_UNSIGNED_CHAR */
+#ifdef __CHAR_UNSIGNED__
+typedef char UINT8;
+#else /* not __CHAR_UNSIGNED__ */
+typedef short UINT8;
+#endif /* __CHAR_UNSIGNED__ */
+#endif /* HAVE_UNSIGNED_CHAR */
+
+/* UINT16 must hold at least the values 0..65535. */
+
+#ifdef HAVE_UNSIGNED_SHORT
+typedef unsigned short UINT16;
+#else /* not HAVE_UNSIGNED_SHORT */
+typedef unsigned int UINT16;
+#endif /* HAVE_UNSIGNED_SHORT */
+
+/* INT16 must hold at least the values -32768..32767. */
+
+#ifndef XMD_H			/* X11/xmd.h correctly defines INT16 */
+typedef short INT16;
+#endif
+
+/* INT32 must hold at least signed 32-bit values. */
+
+#ifndef XMD_H			/* X11/xmd.h correctly defines INT32 */
+typedef long INT32;
+#endif
+
+/* Datatype used for image dimensions.  The JPEG standard only supports
+ * images up to 64K*64K due to 16-bit fields in SOF markers.  Therefore
+ * "unsigned int" is sufficient on all machines.  However, if you need to
+ * handle larger images and you don't mind deviating from the spec, you
+ * can change this datatype.
+ */
+
+typedef unsigned int JDIMENSION;
+
+#define JPEG_MAX_DIMENSION  65500L  /* a tad under 64K to prevent overflows */
+
+
+/* These macros are used in all function definitions and extern declarations.
+ * You could modify them if you need to change function linkage conventions;
+ * in particular, you'll need to do that to make the library a Windows DLL.
+ * Another application is to make all functions global for use with debuggers
+ * or code profilers that require it.
+ */
+
+/* a function called through method pointers: */
+#define METHODDEF(type)		static type
+/* a function used only in its module: */
+#define LOCAL(type)		static type
+/* a function referenced thru EXTERNs: */
+#define GLOBAL(type)		type
+/* a reference to a GLOBAL function: */
+#define EXTERN(type)		extern type
+
+
+/* This macro is used to declare a "method", that is, a function pointer.
+ * We want to supply prototype parameters if the compiler can cope.
+ * Note that the arglist parameter must be parenthesized!
+ * Again, you can customize this if you need special linkage keywords.
+ */
+
+#ifdef HAVE_PROTOTYPES
+#define JMETHOD(type,methodname,arglist)  type (*methodname) arglist
+#else
+#define JMETHOD(type,methodname,arglist)  type (*methodname) ()
+#endif
+
+
+/* Here is the pseudo-keyword for declaring pointers that must be "far"
+ * on 80x86 machines.  Most of the specialized coding for 80x86 is handled
+ * by just saying "FAR *" where such a pointer is needed.  In a few places
+ * explicit coding is needed; see uses of the NEED_FAR_POINTERS symbol.
+ */
+
+#ifdef NEED_FAR_POINTERS
+#define FAR  far
+#else
+#define FAR
+#endif
+
+
+/*
+ * On a few systems, type boolean and/or its values FALSE, TRUE may appear
+ * in standard header files.  Or you may have conflicts with application-
+ * specific header files that you want to include together with these files.
+ * Defining HAVE_BOOLEAN before including jpeglib.h should make it work.
+ */
+
+#ifndef HAVE_BOOLEAN
+typedef int boolean;
+#endif
+#ifndef FALSE			/* in case these macros already exist */
+#define FALSE	0		/* values of boolean */
+#endif
+#ifndef TRUE
+#define TRUE	1
+#endif
+
+
+/*
+ * The remaining options affect code selection within the JPEG library,
+ * but they don't need to be visible to most applications using the library.
+ * To minimize application namespace pollution, the symbols won't be
+ * defined unless JPEG_INTERNALS or JPEG_INTERNAL_OPTIONS has been defined.
+ */
+
+#ifdef JPEG_INTERNALS
+#define JPEG_INTERNAL_OPTIONS
+#endif
+
+#ifdef JPEG_INTERNAL_OPTIONS
+
+
+/*
+ * These defines indicate whether to include various optional functions.
+ * Undefining some of these symbols will produce a smaller but less capable
+ * library.  Note that you can leave certain source files out of the
+ * compilation/linking process if you've #undef'd the corresponding symbols.
+ * (You may HAVE to do that if your compiler doesn't like null source files.)
+ */
+
+/* Arithmetic coding is unsupported for legal reasons.  Complaints to IBM. */
+
+/* Capability options common to encoder and decoder: */
+
+#define DCT_ISLOW_SUPPORTED	/* slow but accurate integer algorithm */
+#define DCT_IFAST_SUPPORTED	/* faster, less accurate integer method */
+#define DCT_FLOAT_SUPPORTED	/* floating-point: accurate, fast on fast HW */
+
+/* Encoder capability options: */
+
+#undef  C_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
+#define C_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
+#define C_PROGRESSIVE_SUPPORTED	    /* Progressive JPEG? (Requires MULTISCAN)*/
+#define ENTROPY_OPT_SUPPORTED	    /* Optimization of entropy coding parms? */
+/* Note: if you selected 12-bit data precision, it is dangerous to turn off
+ * ENTROPY_OPT_SUPPORTED.  The standard Huffman tables are only good for 8-bit
+ * precision, so jchuff.c normally uses entropy optimization to compute
+ * usable tables for higher precision.  If you don't want to do optimization,
+ * you'll have to supply different default Huffman tables.
+ * The exact same statements apply for progressive JPEG: the default tables
+ * don't work for progressive mode.  (This may get fixed, however.)
+ */
+#define INPUT_SMOOTHING_SUPPORTED   /* Input image smoothing option? */
+
+/* Decoder capability options: */
+
+#undef  D_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
+#define D_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
+#define D_PROGRESSIVE_SUPPORTED	    /* Progressive JPEG? (Requires MULTISCAN)*/
+#define SAVE_MARKERS_SUPPORTED	    /* jpeg_save_markers() needed? */
+#define BLOCK_SMOOTHING_SUPPORTED   /* Block smoothing? (Progressive only) */
+#define IDCT_SCALING_SUPPORTED	    /* Output rescaling via IDCT? */
+#undef  UPSAMPLE_SCALING_SUPPORTED  /* Output rescaling at upsample stage? */
+#define UPSAMPLE_MERGING_SUPPORTED  /* Fast path for sloppy upsampling? */
+#define QUANT_1PASS_SUPPORTED	    /* 1-pass color quantization? */
+#define QUANT_2PASS_SUPPORTED	    /* 2-pass color quantization? */
+
+/* more capability options later, no doubt */
+
+
+/*
+ * Ordering of RGB data in scanlines passed to or from the application.
+ * If your application wants to deal with data in the order B,G,R, just
+ * change these macros.  You can also deal with formats such as R,G,B,X
+ * (one extra byte per pixel) by changing RGB_PIXELSIZE.  Note that changing
+ * the offsets will also change the order in which colormap data is organized.
+ * RESTRICTIONS:
+ * 1. The sample applications cjpeg,djpeg do NOT support modified RGB formats.
+ * 2. These macros only affect RGB<=>YCbCr color conversion, so they are not
+ *    useful if you are using JPEG color spaces other than YCbCr or grayscale.
+ * 3. The color quantizer modules will not behave desirably if RGB_PIXELSIZE
+ *    is not 3 (they don't understand about dummy color components!).  So you
+ *    can't use color quantization if you change that value.
+ */
+
+#define RGB_RED		0	/* Offset of Red in an RGB scanline element */
+#define RGB_GREEN	1	/* Offset of Green */
+#define RGB_BLUE	2	/* Offset of Blue */
+#define RGB_PIXELSIZE	3	/* JSAMPLEs per RGB scanline element */
+
+#define JPEG_NUMCS 12
+
+static const int rgb_red[JPEG_NUMCS] = {
+	-1, -1, RGB_RED, -1, -1, -1, 0, 0, 2, 2, 3, 1
+};
+
+static const int rgb_green[JPEG_NUMCS] = {
+	-1, -1, RGB_GREEN, -1, -1, -1, 1, 1, 1, 1, 2, 2
+};
+
+static const int rgb_blue[JPEG_NUMCS] = {
+	-1, -1, RGB_BLUE, -1, -1, -1, 2, 2, 0, 0, 1, 3
+};
+
+static const int rgb_pixelsize[JPEG_NUMCS] = {
+	-1, -1, RGB_PIXELSIZE, -1, -1, -1, 3, 4, 3, 4, 4, 4
+};
+
+/* Definitions for speed-related optimizations. */
+
+/* On some machines (notably 68000 series) "int" is 32 bits, but multiplying
+ * two 16-bit shorts is faster than multiplying two ints.  Define MULTIPLIER
+ * as short on such a machine.  MULTIPLIER must be at least 16 bits wide.
+ */
+
+#ifndef MULTIPLIER
+#ifndef WITH_SIMD
+#define MULTIPLIER  int		/* type for fastest integer multiply */
+#else
+#define MULTIPLIER short  /* prefer 16-bit with SIMD for parellelism */
+#endif
+#endif
+
+
+/* FAST_FLOAT should be either float or double, whichever is done faster
+ * by your compiler.  (Note that this type is only used in the floating point
+ * DCT routines, so it only matters if you've defined DCT_FLOAT_SUPPORTED.)
+ * Typically, float is faster in ANSI C compilers, while double is faster in
+ * pre-ANSI compilers (because they insist on converting to double anyway).
+ * The code below therefore chooses float if we have ANSI-style prototypes.
+ */
+
+#ifndef FAST_FLOAT
+#ifdef HAVE_PROTOTYPES
+#define FAST_FLOAT  float
+#else
+#define FAST_FLOAT  double
+#endif
+#endif
+
+#endif /* JPEG_INTERNAL_OPTIONS */
diff --git a/jpegcomp.h b/jpegcomp.h
new file mode 100644
index 0000000..1b9e0a4
--- /dev/null
+++ b/jpegcomp.h
@@ -0,0 +1,26 @@
+/*
+ * jpegcomp.h
+ *
+ * Copyright (C) 2010, D. R. Commander
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * JPEG compatibility macros
+ * These declarations are considered internal to the JPEG library; most
+ * applications using the library shouldn't need to include this file.
+ */
+
+#if JPEG_LIB_VERSION >= 70
+#define _DCT_scaled_size DCT_h_scaled_size
+#define _min_DCT_scaled_size min_DCT_h_scaled_size
+#define _min_DCT_h_scaled_size min_DCT_h_scaled_size
+#define _min_DCT_v_scaled_size min_DCT_v_scaled_size
+#define _jpeg_width jpeg_width
+#define _jpeg_height jpeg_height
+#else
+#define _DCT_scaled_size DCT_scaled_size
+#define _min_DCT_scaled_size min_DCT_scaled_size
+#define _min_DCT_h_scaled_size min_DCT_scaled_size
+#define _min_DCT_v_scaled_size min_DCT_scaled_size
+#define _jpeg_width image_width
+#define _jpeg_height image_height
+#endif
diff --git a/jpegint.h b/jpegint.h
new file mode 100644
index 0000000..7a31f51
--- /dev/null
+++ b/jpegint.h
@@ -0,0 +1,392 @@
+/*
+ * jpegint.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file provides common declarations for the various JPEG modules.
+ * These declarations are considered internal to the JPEG library; most
+ * applications using the library shouldn't need to include this file.
+ */
+
+
+/* Declarations for both compression & decompression */
+
+typedef enum {			/* Operating modes for buffer controllers */
+	JBUF_PASS_THRU,		/* Plain stripwise operation */
+	/* Remaining modes require a full-image buffer to have been created */
+	JBUF_SAVE_SOURCE,	/* Run source subobject only, save output */
+	JBUF_CRANK_DEST,	/* Run dest subobject only, using saved data */
+	JBUF_SAVE_AND_PASS	/* Run both subobjects, save output */
+} J_BUF_MODE;
+
+/* Values of global_state field (jdapi.c has some dependencies on ordering!) */
+#define CSTATE_START	100	/* after create_compress */
+#define CSTATE_SCANNING	101	/* start_compress done, write_scanlines OK */
+#define CSTATE_RAW_OK	102	/* start_compress done, write_raw_data OK */
+#define CSTATE_WRCOEFS	103	/* jpeg_write_coefficients done */
+#define DSTATE_START	200	/* after create_decompress */
+#define DSTATE_INHEADER	201	/* reading header markers, no SOS yet */
+#define DSTATE_READY	202	/* found SOS, ready for start_decompress */
+#define DSTATE_PRELOAD	203	/* reading multiscan file in start_decompress*/
+#define DSTATE_PRESCAN	204	/* performing dummy pass for 2-pass quant */
+#define DSTATE_SCANNING	205	/* start_decompress done, read_scanlines OK */
+#define DSTATE_RAW_OK	206	/* start_decompress done, read_raw_data OK */
+#define DSTATE_BUFIMAGE	207	/* expecting jpeg_start_output */
+#define DSTATE_BUFPOST	208	/* looking for SOS/EOI in jpeg_finish_output */
+#define DSTATE_RDCOEFS	209	/* reading file in jpeg_read_coefficients */
+#define DSTATE_STOPPING	210	/* looking for EOI in jpeg_finish_decompress */
+
+
+/* Declarations for compression modules */
+
+/* Master control module */
+struct jpeg_comp_master {
+  JMETHOD(void, prepare_for_pass, (j_compress_ptr cinfo));
+  JMETHOD(void, pass_startup, (j_compress_ptr cinfo));
+  JMETHOD(void, finish_pass, (j_compress_ptr cinfo));
+
+  /* State variables made visible to other modules */
+  boolean call_pass_startup;	/* True if pass_startup must be called */
+  boolean is_last_pass;		/* True during last pass */
+};
+
+/* Main buffer control (downsampled-data buffer) */
+struct jpeg_c_main_controller {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, process_data, (j_compress_ptr cinfo,
+			       JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			       JDIMENSION in_rows_avail));
+};
+
+/* Compression preprocessing (downsampling input buffer control) */
+struct jpeg_c_prep_controller {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, pre_process_data, (j_compress_ptr cinfo,
+				   JSAMPARRAY input_buf,
+				   JDIMENSION *in_row_ctr,
+				   JDIMENSION in_rows_avail,
+				   JSAMPIMAGE output_buf,
+				   JDIMENSION *out_row_group_ctr,
+				   JDIMENSION out_row_groups_avail));
+};
+
+/* Coefficient buffer control */
+struct jpeg_c_coef_controller {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(boolean, compress_data, (j_compress_ptr cinfo,
+				   JSAMPIMAGE input_buf));
+};
+
+/* Colorspace conversion */
+struct jpeg_color_converter {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo));
+  JMETHOD(void, color_convert, (j_compress_ptr cinfo,
+				JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+				JDIMENSION output_row, int num_rows));
+};
+
+/* Downsampling */
+struct jpeg_downsampler {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo));
+  JMETHOD(void, downsample, (j_compress_ptr cinfo,
+			     JSAMPIMAGE input_buf, JDIMENSION in_row_index,
+			     JSAMPIMAGE output_buf,
+			     JDIMENSION out_row_group_index));
+
+  boolean need_context_rows;	/* TRUE if need rows above & below */
+};
+
+/* Forward DCT (also controls coefficient quantization) */
+struct jpeg_forward_dct {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo));
+  /* perhaps this should be an array??? */
+  JMETHOD(void, forward_DCT, (j_compress_ptr cinfo,
+			      jpeg_component_info * compptr,
+			      JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+			      JDIMENSION start_row, JDIMENSION start_col,
+			      JDIMENSION num_blocks));
+};
+
+/* Entropy encoding */
+struct jpeg_entropy_encoder {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, boolean gather_statistics));
+  JMETHOD(boolean, encode_mcu, (j_compress_ptr cinfo, JBLOCKROW *MCU_data));
+  JMETHOD(void, finish_pass, (j_compress_ptr cinfo));
+};
+
+/* Marker writing */
+struct jpeg_marker_writer {
+  JMETHOD(void, write_file_header, (j_compress_ptr cinfo));
+  JMETHOD(void, write_frame_header, (j_compress_ptr cinfo));
+  JMETHOD(void, write_scan_header, (j_compress_ptr cinfo));
+  JMETHOD(void, write_file_trailer, (j_compress_ptr cinfo));
+  JMETHOD(void, write_tables_only, (j_compress_ptr cinfo));
+  /* These routines are exported to allow insertion of extra markers */
+  /* Probably only COM and APPn markers should be written this way */
+  JMETHOD(void, write_marker_header, (j_compress_ptr cinfo, int marker,
+				      unsigned int datalen));
+  JMETHOD(void, write_marker_byte, (j_compress_ptr cinfo, int val));
+};
+
+
+/* Declarations for decompression modules */
+
+/* Master control module */
+struct jpeg_decomp_master {
+  JMETHOD(void, prepare_for_output_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, finish_output_pass, (j_decompress_ptr cinfo));
+
+  /* State variables made visible to other modules */
+  boolean is_dummy_pass;	/* True during 1st pass for 2-pass quant */
+};
+
+/* Input control module */
+struct jpeg_input_controller {
+  JMETHOD(int, consume_input, (j_decompress_ptr cinfo));
+  JMETHOD(void, reset_input_controller, (j_decompress_ptr cinfo));
+  JMETHOD(void, start_input_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, finish_input_pass, (j_decompress_ptr cinfo));
+
+  /* State variables made visible to other modules */
+  boolean has_multiple_scans;	/* True if file has multiple scans */
+  boolean eoi_reached;		/* True when EOI has been consumed */
+};
+
+/* Main buffer control (downsampled-data buffer) */
+struct jpeg_d_main_controller {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, process_data, (j_decompress_ptr cinfo,
+			       JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			       JDIMENSION out_rows_avail));
+};
+
+/* Coefficient buffer control */
+struct jpeg_d_coef_controller {
+  JMETHOD(void, start_input_pass, (j_decompress_ptr cinfo));
+  JMETHOD(int, consume_data, (j_decompress_ptr cinfo));
+  JMETHOD(void, start_output_pass, (j_decompress_ptr cinfo));
+  JMETHOD(int, decompress_data, (j_decompress_ptr cinfo,
+				 JSAMPIMAGE output_buf));
+  /* Pointer to array of coefficient virtual arrays, or NULL if none */
+  jvirt_barray_ptr *coef_arrays;
+};
+
+/* Decompression postprocessing (color quantization buffer control) */
+struct jpeg_d_post_controller {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, post_process_data, (j_decompress_ptr cinfo,
+				    JSAMPIMAGE input_buf,
+				    JDIMENSION *in_row_group_ctr,
+				    JDIMENSION in_row_groups_avail,
+				    JSAMPARRAY output_buf,
+				    JDIMENSION *out_row_ctr,
+				    JDIMENSION out_rows_avail));
+};
+
+/* Marker reading & parsing */
+struct jpeg_marker_reader {
+  JMETHOD(void, reset_marker_reader, (j_decompress_ptr cinfo));
+  /* Read markers until SOS or EOI.
+   * Returns same codes as are defined for jpeg_consume_input:
+   * JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+   */
+  JMETHOD(int, read_markers, (j_decompress_ptr cinfo));
+  /* Read a restart marker --- exported for use by entropy decoder only */
+  jpeg_marker_parser_method read_restart_marker;
+
+  /* State of marker reader --- nominally internal, but applications
+   * supplying COM or APPn handlers might like to know the state.
+   */
+  boolean saw_SOI;		/* found SOI? */
+  boolean saw_SOF;		/* found SOF? */
+  int next_restart_num;		/* next restart number expected (0-7) */
+  unsigned int discarded_bytes;	/* # of bytes skipped looking for a marker */
+};
+
+/* Entropy decoding */
+struct jpeg_entropy_decoder {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  JMETHOD(boolean, decode_mcu, (j_decompress_ptr cinfo,
+				JBLOCKROW *MCU_data));
+
+  /* This is here to share code between baseline and progressive decoders; */
+  /* other modules probably should not use it */
+  boolean insufficient_data;	/* set TRUE after emitting warning */
+};
+
+/* Inverse DCT (also performs dequantization) */
+typedef JMETHOD(void, inverse_DCT_method_ptr,
+		(j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col));
+
+struct jpeg_inverse_dct {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  /* It is useful to allow each component to have a separate IDCT method. */
+  inverse_DCT_method_ptr inverse_DCT[MAX_COMPONENTS];
+};
+
+/* Upsampling (note that upsampler must also call color converter) */
+struct jpeg_upsampler {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, upsample, (j_decompress_ptr cinfo,
+			   JSAMPIMAGE input_buf,
+			   JDIMENSION *in_row_group_ctr,
+			   JDIMENSION in_row_groups_avail,
+			   JSAMPARRAY output_buf,
+			   JDIMENSION *out_row_ctr,
+			   JDIMENSION out_rows_avail));
+
+  boolean need_context_rows;	/* TRUE if need rows above & below */
+};
+
+/* Colorspace conversion */
+struct jpeg_color_deconverter {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, color_convert, (j_decompress_ptr cinfo,
+				JSAMPIMAGE input_buf, JDIMENSION input_row,
+				JSAMPARRAY output_buf, int num_rows));
+};
+
+/* Color quantization or color precision reduction */
+struct jpeg_color_quantizer {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo, boolean is_pre_scan));
+  JMETHOD(void, color_quantize, (j_decompress_ptr cinfo,
+				 JSAMPARRAY input_buf, JSAMPARRAY output_buf,
+				 int num_rows));
+  JMETHOD(void, finish_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, new_color_map, (j_decompress_ptr cinfo));
+};
+
+
+/* Miscellaneous useful macros */
+
+#undef MAX
+#define MAX(a,b)	((a) > (b) ? (a) : (b))
+#undef MIN
+#define MIN(a,b)	((a) < (b) ? (a) : (b))
+
+
+/* We assume that right shift corresponds to signed division by 2 with
+ * rounding towards minus infinity.  This is correct for typical "arithmetic
+ * shift" instructions that shift in copies of the sign bit.  But some
+ * C compilers implement >> with an unsigned shift.  For these machines you
+ * must define RIGHT_SHIFT_IS_UNSIGNED.
+ * RIGHT_SHIFT provides a proper signed right shift of an INT32 quantity.
+ * It is only applied with constant shift counts.  SHIFT_TEMPS must be
+ * included in the variables of any routine using RIGHT_SHIFT.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define SHIFT_TEMPS	INT32 shift_temp;
+#define RIGHT_SHIFT(x,shft)  \
+	((shift_temp = (x)) < 0 ? \
+	 (shift_temp >> (shft)) | ((~((INT32) 0)) << (32-(shft))) : \
+	 (shift_temp >> (shft)))
+#else
+#define SHIFT_TEMPS
+#define RIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jinit_compress_master	jICompress
+#define jinit_c_master_control	jICMaster
+#define jinit_c_main_controller	jICMainC
+#define jinit_c_prep_controller	jICPrepC
+#define jinit_c_coef_controller	jICCoefC
+#define jinit_color_converter	jICColor
+#define jinit_downsampler	jIDownsampler
+#define jinit_forward_dct	jIFDCT
+#define jinit_huff_encoder	jIHEncoder
+#define jinit_phuff_encoder	jIPHEncoder
+#define jinit_marker_writer	jIMWriter
+#define jinit_master_decompress	jIDMaster
+#define jinit_d_main_controller	jIDMainC
+#define jinit_d_coef_controller	jIDCoefC
+#define jinit_d_post_controller	jIDPostC
+#define jinit_input_controller	jIInCtlr
+#define jinit_marker_reader	jIMReader
+#define jinit_huff_decoder	jIHDecoder
+#define jinit_phuff_decoder	jIPHDecoder
+#define jinit_inverse_dct	jIIDCT
+#define jinit_upsampler		jIUpsampler
+#define jinit_color_deconverter	jIDColor
+#define jinit_1pass_quantizer	jI1Quant
+#define jinit_2pass_quantizer	jI2Quant
+#define jinit_merged_upsampler	jIMUpsampler
+#define jinit_memory_mgr	jIMemMgr
+#define jdiv_round_up		jDivRound
+#define jround_up		jRound
+#define jcopy_sample_rows	jCopySamples
+#define jcopy_block_row		jCopyBlocks
+#define jzero_far		jZeroFar
+#define jpeg_zigzag_order	jZIGTable
+#define jpeg_natural_order	jZAGTable
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/* Compression module initialization routines */
+EXTERN(void) jinit_compress_master JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_c_master_control JPP((j_compress_ptr cinfo,
+					 boolean transcode_only));
+EXTERN(void) jinit_c_main_controller JPP((j_compress_ptr cinfo,
+					  boolean need_full_buffer));
+EXTERN(void) jinit_c_prep_controller JPP((j_compress_ptr cinfo,
+					  boolean need_full_buffer));
+EXTERN(void) jinit_c_coef_controller JPP((j_compress_ptr cinfo,
+					  boolean need_full_buffer));
+EXTERN(void) jinit_color_converter JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_downsampler JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_forward_dct JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_huff_encoder JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_phuff_encoder JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_marker_writer JPP((j_compress_ptr cinfo));
+/* Decompression module initialization routines */
+EXTERN(void) jinit_master_decompress JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_d_main_controller JPP((j_decompress_ptr cinfo,
+					  boolean need_full_buffer));
+EXTERN(void) jinit_d_coef_controller JPP((j_decompress_ptr cinfo,
+					  boolean need_full_buffer));
+EXTERN(void) jinit_d_post_controller JPP((j_decompress_ptr cinfo,
+					  boolean need_full_buffer));
+EXTERN(void) jinit_input_controller JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_marker_reader JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_huff_decoder JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_phuff_decoder JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_inverse_dct JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_upsampler JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_color_deconverter JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_1pass_quantizer JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_2pass_quantizer JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_merged_upsampler JPP((j_decompress_ptr cinfo));
+/* Memory manager initialization */
+EXTERN(void) jinit_memory_mgr JPP((j_common_ptr cinfo));
+
+/* Utility routines in jutils.c */
+EXTERN(long) jdiv_round_up JPP((long a, long b));
+EXTERN(size_t) jround_up JPP((size_t a, size_t b));
+EXTERN(void) jcopy_sample_rows JPP((JSAMPARRAY input_array, int source_row,
+				    JSAMPARRAY output_array, int dest_row,
+				    int num_rows, JDIMENSION num_cols));
+EXTERN(void) jcopy_block_row JPP((JBLOCKROW input_row, JBLOCKROW output_row,
+				  JDIMENSION num_blocks));
+EXTERN(void) jzero_far JPP((void FAR * target, size_t bytestozero));
+/* Constant tables in jutils.c */
+#if 0				/* This table is not actually needed in v6a */
+extern const int jpeg_zigzag_order[]; /* natural coef order to zigzag order */
+#endif
+extern const int jpeg_natural_order[]; /* zigzag coef order to natural order */
+
+/* Suppress undefined-structure complaints if necessary. */
+
+#ifdef INCOMPLETE_TYPES_BROKEN
+#ifndef AM_MEMORY_MANAGER	/* only jmemmgr.c defines these */
+struct jvirt_sarray_control { long dummy; };
+struct jvirt_barray_control { long dummy; };
+#endif
+#endif /* INCOMPLETE_TYPES_BROKEN */
diff --git a/jpeglib.h b/jpeglib.h
new file mode 100644
index 0000000..cb3acaf
--- /dev/null
+++ b/jpeglib.h
@@ -0,0 +1,1201 @@
+/*
+ * jpeglib.h
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2002-2009 by Guido Vollbeding.
+ * Copyright (C) 2009-2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file defines the application interface for the JPEG library.
+ * Most applications using the library need only include this file,
+ * and perhaps jerror.h if they want to know the exact error codes.
+ */
+
+#ifndef JPEGLIB_H
+#define JPEGLIB_H
+
+/*
+ * First we include the configuration files that record how this
+ * installation of the JPEG library is set up.  jconfig.h can be
+ * generated automatically for many systems.  jmorecfg.h contains
+ * manual configuration options that most people need not worry about.
+ */
+
+#ifndef JCONFIG_INCLUDED	/* in case jinclude.h already did */
+#include "jconfig.h"		/* widely used configuration options */
+#endif
+#include "jmorecfg.h"		/* seldom changed options */
+
+
+#ifdef __cplusplus
+#ifndef DONT_USE_EXTERN_C
+extern "C" {
+#endif
+#endif
+
+
+/* Various constants determining the sizes of things.
+ * All of these are specified by the JPEG standard, so don't change them
+ * if you want to be compatible.
+ */
+
+#define DCTSIZE		    8	/* The basic DCT block is 8x8 samples */
+#define DCTSIZE2	    64	/* DCTSIZE squared; # of elements in a block */
+#define NUM_QUANT_TBLS      4	/* Quantization tables are numbered 0..3 */
+#define NUM_HUFF_TBLS       4	/* Huffman tables are numbered 0..3 */
+#define NUM_ARITH_TBLS      16	/* Arith-coding tables are numbered 0..15 */
+#define MAX_COMPS_IN_SCAN   4	/* JPEG limit on # of components in one scan */
+#define MAX_SAMP_FACTOR     4	/* JPEG limit on sampling factors */
+/* Unfortunately, some bozo at Adobe saw no reason to be bound by the standard;
+ * the PostScript DCT filter can emit files with many more than 10 blocks/MCU.
+ * If you happen to run across such a file, you can up D_MAX_BLOCKS_IN_MCU
+ * to handle it.  We even let you do this from the jconfig.h file.  However,
+ * we strongly discourage changing C_MAX_BLOCKS_IN_MCU; just because Adobe
+ * sometimes emits noncompliant files doesn't mean you should too.
+ */
+#define C_MAX_BLOCKS_IN_MCU   10 /* compressor's limit on blocks per MCU */
+#ifndef D_MAX_BLOCKS_IN_MCU
+#define D_MAX_BLOCKS_IN_MCU   10 /* decompressor's limit on blocks per MCU */
+#endif
+
+
+/* Data structures for images (arrays of samples and of DCT coefficients).
+ * On 80x86 machines, the image arrays are too big for near pointers,
+ * but the pointer arrays can fit in near memory.
+ */
+
+typedef JSAMPLE FAR *JSAMPROW;	/* ptr to one image row of pixel samples. */
+typedef JSAMPROW *JSAMPARRAY;	/* ptr to some rows (a 2-D sample array) */
+typedef JSAMPARRAY *JSAMPIMAGE;	/* a 3-D sample array: top index is color */
+
+typedef JCOEF JBLOCK[DCTSIZE2];	/* one block of coefficients */
+typedef JBLOCK FAR *JBLOCKROW;	/* pointer to one row of coefficient blocks */
+typedef JBLOCKROW *JBLOCKARRAY;		/* a 2-D array of coefficient blocks */
+typedef JBLOCKARRAY *JBLOCKIMAGE;	/* a 3-D array of coefficient blocks */
+
+typedef JCOEF FAR *JCOEFPTR;	/* useful in a couple of places */
+
+
+/* Types for JPEG compression parameters and working tables. */
+
+
+/* DCT coefficient quantization tables. */
+
+typedef struct {
+  /* This array gives the coefficient quantizers in natural array order
+   * (not the zigzag order in which they are stored in a JPEG DQT marker).
+   * CAUTION: IJG versions prior to v6a kept this array in zigzag order.
+   */
+  UINT16 quantval[DCTSIZE2];	/* quantization step for each coefficient */
+  /* This field is used only during compression.  It's initialized FALSE when
+   * the table is created, and set TRUE when it's been output to the file.
+   * You could suppress output of a table by setting this to TRUE.
+   * (See jpeg_suppress_tables for an example.)
+   */
+  boolean sent_table;		/* TRUE when table has been output */
+} JQUANT_TBL;
+
+
+/* Huffman coding tables. */
+
+typedef struct {
+  /* These two fields directly represent the contents of a JPEG DHT marker */
+  UINT8 bits[17];		/* bits[k] = # of symbols with codes of */
+				/* length k bits; bits[0] is unused */
+  UINT8 huffval[256];		/* The symbols, in order of incr code length */
+  /* This field is used only during compression.  It's initialized FALSE when
+   * the table is created, and set TRUE when it's been output to the file.
+   * You could suppress output of a table by setting this to TRUE.
+   * (See jpeg_suppress_tables for an example.)
+   */
+  boolean sent_table;		/* TRUE when table has been output */
+} JHUFF_TBL;
+
+
+/* Basic info about one component (color channel). */
+
+typedef struct {
+  /* These values are fixed over the whole image. */
+  /* For compression, they must be supplied by parameter setup; */
+  /* for decompression, they are read from the SOF marker. */
+  int component_id;		/* identifier for this component (0..255) */
+  int component_index;		/* its index in SOF or cinfo->comp_info[] */
+  int h_samp_factor;		/* horizontal sampling factor (1..4) */
+  int v_samp_factor;		/* vertical sampling factor (1..4) */
+  int quant_tbl_no;		/* quantization table selector (0..3) */
+  /* These values may vary between scans. */
+  /* For compression, they must be supplied by parameter setup; */
+  /* for decompression, they are read from the SOS marker. */
+  /* The decompressor output side may not use these variables. */
+  int dc_tbl_no;		/* DC entropy table selector (0..3) */
+  int ac_tbl_no;		/* AC entropy table selector (0..3) */
+  
+  /* Remaining fields should be treated as private by applications. */
+  
+  /* These values are computed during compression or decompression startup: */
+  /* Component's size in DCT blocks.
+   * Any dummy blocks added to complete an MCU are not counted; therefore
+   * these values do not depend on whether a scan is interleaved or not.
+   */
+  JDIMENSION width_in_blocks;
+  JDIMENSION height_in_blocks;
+  /* Size of a DCT block in samples.  Always DCTSIZE for compression.
+   * For decompression this is the size of the output from one DCT block,
+   * reflecting any scaling we choose to apply during the IDCT step.
+   * Values of 1,2,4,8 are likely to be supported.  Note that different
+   * components may receive different IDCT scalings.
+   */
+#if JPEG_LIB_VERSION >= 70
+  int DCT_h_scaled_size;
+  int DCT_v_scaled_size;
+#else
+  int DCT_scaled_size;
+#endif
+  /* The downsampled dimensions are the component's actual, unpadded number
+   * of samples at the main buffer (preprocessing/compression interface), thus
+   * downsampled_width = ceil(image_width * Hi/Hmax)
+   * and similarly for height.  For decompression, IDCT scaling is included, so
+   * downsampled_width = ceil(image_width * Hi/Hmax * DCT_[h_]scaled_size/DCTSIZE)
+   */
+  JDIMENSION downsampled_width;	 /* actual width in samples */
+  JDIMENSION downsampled_height; /* actual height in samples */
+  /* This flag is used only for decompression.  In cases where some of the
+   * components will be ignored (eg grayscale output from YCbCr image),
+   * we can skip most computations for the unused components.
+   */
+  boolean component_needed;	/* do we need the value of this component? */
+
+  /* These values are computed before starting a scan of the component. */
+  /* The decompressor output side may not use these variables. */
+  int MCU_width;		/* number of blocks per MCU, horizontally */
+  int MCU_height;		/* number of blocks per MCU, vertically */
+  int MCU_blocks;		/* MCU_width * MCU_height */
+  int MCU_sample_width;		/* MCU width in samples, MCU_width*DCT_[h_]scaled_size */
+  int last_col_width;		/* # of non-dummy blocks across in last MCU */
+  int last_row_height;		/* # of non-dummy blocks down in last MCU */
+
+  /* Saved quantization table for component; NULL if none yet saved.
+   * See jdinput.c comments about the need for this information.
+   * This field is currently used only for decompression.
+   */
+  JQUANT_TBL * quant_table;
+
+  /* Private per-component storage for DCT or IDCT subsystem. */
+  void * dct_table;
+} jpeg_component_info;
+
+
+/* The script for encoding a multiple-scan file is an array of these: */
+
+typedef struct {
+  int comps_in_scan;		/* number of components encoded in this scan */
+  int component_index[MAX_COMPS_IN_SCAN]; /* their SOF/comp_info[] indexes */
+  int Ss, Se;			/* progressive JPEG spectral selection parms */
+  int Ah, Al;			/* progressive JPEG successive approx. parms */
+} jpeg_scan_info;
+
+/* The decompressor can save APPn and COM markers in a list of these: */
+
+typedef struct jpeg_marker_struct FAR * jpeg_saved_marker_ptr;
+
+struct jpeg_marker_struct {
+  jpeg_saved_marker_ptr next;	/* next in list, or NULL */
+  UINT8 marker;			/* marker code: JPEG_COM, or JPEG_APP0+n */
+  unsigned int original_length;	/* # bytes of data in the file */
+  unsigned int data_length;	/* # bytes of data saved at data[] */
+  JOCTET FAR * data;		/* the data contained in the marker */
+  /* the marker length word is not counted in data_length or original_length */
+};
+
+/* Known color spaces. */
+
+#define JCS_EXTENSIONS 1
+
+typedef enum {
+	JCS_UNKNOWN,		/* error/unspecified */
+	JCS_GRAYSCALE,		/* monochrome */
+	JCS_RGB,		/* red/green/blue as specified by the RGB_RED, RGB_GREEN,
+                 RGB_BLUE, and RGB_PIXELSIZE macros */
+	JCS_YCbCr,		/* Y/Cb/Cr (also known as YUV) */
+	JCS_CMYK,		/* C/M/Y/K */
+	JCS_YCCK,		/* Y/Cb/Cr/K */
+	JCS_EXT_RGB,		/* red/green/blue */
+	JCS_EXT_RGBX,		/* red/green/blue/x */
+	JCS_EXT_BGR,		/* blue/green/red */
+	JCS_EXT_BGRX,		/* blue/green/red/x */
+	JCS_EXT_XBGR,		/* x/blue/green/red */
+	JCS_EXT_XRGB		/* x/red/green/blue */
+} J_COLOR_SPACE;
+
+/* DCT/IDCT algorithm options. */
+
+typedef enum {
+	JDCT_ISLOW,		/* slow but accurate integer algorithm */
+	JDCT_IFAST,		/* faster, less accurate integer method */
+	JDCT_FLOAT		/* floating-point: accurate, fast on fast HW */
+} J_DCT_METHOD;
+
+#ifndef JDCT_DEFAULT		/* may be overridden in jconfig.h */
+#define JDCT_DEFAULT  JDCT_ISLOW
+#endif
+#ifndef JDCT_FASTEST		/* may be overridden in jconfig.h */
+#define JDCT_FASTEST  JDCT_IFAST
+#endif
+
+/* Dithering options for decompression. */
+
+typedef enum {
+	JDITHER_NONE,		/* no dithering */
+	JDITHER_ORDERED,	/* simple ordered dither */
+	JDITHER_FS		/* Floyd-Steinberg error diffusion dither */
+} J_DITHER_MODE;
+
+
+/* Common fields between JPEG compression and decompression master structs. */
+
+#define jpeg_common_fields \
+  struct jpeg_error_mgr * err;	/* Error handler module */\
+  struct jpeg_memory_mgr * mem;	/* Memory manager module */\
+  struct jpeg_progress_mgr * progress; /* Progress monitor, or NULL if none */\
+  void * client_data;		/* Available for use by application */\
+  boolean is_decompressor;	/* So common code can tell which is which */\
+  int global_state		/* For checking call sequence validity */
+
+/* Routines that are to be used by both halves of the library are declared
+ * to receive a pointer to this structure.  There are no actual instances of
+ * jpeg_common_struct, only of jpeg_compress_struct and jpeg_decompress_struct.
+ */
+struct jpeg_common_struct {
+  jpeg_common_fields;		/* Fields common to both master struct types */
+  /* Additional fields follow in an actual jpeg_compress_struct or
+   * jpeg_decompress_struct.  All three structs must agree on these
+   * initial fields!  (This would be a lot cleaner in C++.)
+   */
+};
+
+typedef struct jpeg_common_struct * j_common_ptr;
+typedef struct jpeg_compress_struct * j_compress_ptr;
+typedef struct jpeg_decompress_struct * j_decompress_ptr;
+
+
+/* Master record for a compression instance */
+
+struct jpeg_compress_struct {
+  jpeg_common_fields;		/* Fields shared with jpeg_decompress_struct */
+
+  /* Destination for compressed data */
+  struct jpeg_destination_mgr * dest;
+
+  /* Description of source image --- these fields must be filled in by
+   * outer application before starting compression.  in_color_space must
+   * be correct before you can even call jpeg_set_defaults().
+   */
+
+  JDIMENSION image_width;	/* input image width */
+  JDIMENSION image_height;	/* input image height */
+  int input_components;		/* # of color components in input image */
+  J_COLOR_SPACE in_color_space;	/* colorspace of input image */
+
+  double input_gamma;		/* image gamma of input image */
+
+  /* Compression parameters --- these fields must be set before calling
+   * jpeg_start_compress().  We recommend calling jpeg_set_defaults() to
+   * initialize everything to reasonable defaults, then changing anything
+   * the application specifically wants to change.  That way you won't get
+   * burnt when new parameters are added.  Also note that there are several
+   * helper routines to simplify changing parameters.
+   */
+
+#if JPEG_LIB_VERSION >= 70
+  unsigned int scale_num, scale_denom; /* fraction by which to scale image */
+
+  JDIMENSION jpeg_width;	/* scaled JPEG image width */
+  JDIMENSION jpeg_height;	/* scaled JPEG image height */
+  /* Dimensions of actual JPEG image that will be written to file,
+   * derived from input dimensions by scaling factors above.
+   * These fields are computed by jpeg_start_compress().
+   * You can also use jpeg_calc_jpeg_dimensions() to determine these values
+   * in advance of calling jpeg_start_compress().
+   */
+#endif
+
+  int data_precision;		/* bits of precision in image data */
+
+  int num_components;		/* # of color components in JPEG image */
+  J_COLOR_SPACE jpeg_color_space; /* colorspace of JPEG image */
+
+  jpeg_component_info * comp_info;
+  /* comp_info[i] describes component that appears i'th in SOF */
+
+  JQUANT_TBL * quant_tbl_ptrs[NUM_QUANT_TBLS];
+#if JPEG_LIB_VERSION >= 70
+  int q_scale_factor[NUM_QUANT_TBLS];
+#endif
+  /* ptrs to coefficient quantization tables, or NULL if not defined,
+   * and corresponding scale factors (percentage, initialized 100).
+   */
+
+  JHUFF_TBL * dc_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  JHUFF_TBL * ac_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  /* ptrs to Huffman coding tables, or NULL if not defined */
+
+  UINT8 arith_dc_L[NUM_ARITH_TBLS]; /* L values for DC arith-coding tables */
+  UINT8 arith_dc_U[NUM_ARITH_TBLS]; /* U values for DC arith-coding tables */
+  UINT8 arith_ac_K[NUM_ARITH_TBLS]; /* Kx values for AC arith-coding tables */
+
+  int num_scans;		/* # of entries in scan_info array */
+  const jpeg_scan_info * scan_info; /* script for multi-scan file, or NULL */
+  /* The default value of scan_info is NULL, which causes a single-scan
+   * sequential JPEG file to be emitted.  To create a multi-scan file,
+   * set num_scans and scan_info to point to an array of scan definitions.
+   */
+
+  boolean raw_data_in;		/* TRUE=caller supplies downsampled data */
+  boolean arith_code;		/* TRUE=arithmetic coding, FALSE=Huffman */
+  boolean optimize_coding;	/* TRUE=optimize entropy encoding parms */
+  boolean CCIR601_sampling;	/* TRUE=first samples are cosited */
+#if JPEG_LIB_VERSION >= 70
+  boolean do_fancy_downsampling; /* TRUE=apply fancy downsampling */
+#endif
+  int smoothing_factor;		/* 1..100, or 0 for no input smoothing */
+  J_DCT_METHOD dct_method;	/* DCT algorithm selector */
+
+  /* The restart interval can be specified in absolute MCUs by setting
+   * restart_interval, or in MCU rows by setting restart_in_rows
+   * (in which case the correct restart_interval will be figured
+   * for each scan).
+   */
+  unsigned int restart_interval; /* MCUs per restart, or 0 for no restart */
+  int restart_in_rows;		/* if > 0, MCU rows per restart interval */
+
+  /* Parameters controlling emission of special markers. */
+
+  boolean write_JFIF_header;	/* should a JFIF marker be written? */
+  UINT8 JFIF_major_version;	/* What to write for the JFIF version number */
+  UINT8 JFIF_minor_version;
+  /* These three values are not used by the JPEG code, merely copied */
+  /* into the JFIF APP0 marker.  density_unit can be 0 for unknown, */
+  /* 1 for dots/inch, or 2 for dots/cm.  Note that the pixel aspect */
+  /* ratio is defined by X_density/Y_density even when density_unit=0. */
+  UINT8 density_unit;		/* JFIF code for pixel size units */
+  UINT16 X_density;		/* Horizontal pixel density */
+  UINT16 Y_density;		/* Vertical pixel density */
+  boolean write_Adobe_marker;	/* should an Adobe marker be written? */
+  
+  /* State variable: index of next scanline to be written to
+   * jpeg_write_scanlines().  Application may use this to control its
+   * processing loop, e.g., "while (next_scanline < image_height)".
+   */
+
+  JDIMENSION next_scanline;	/* 0 .. image_height-1  */
+
+  /* Remaining fields are known throughout compressor, but generally
+   * should not be touched by a surrounding application.
+   */
+
+  /*
+   * These fields are computed during compression startup
+   */
+  boolean progressive_mode;	/* TRUE if scan script uses progressive mode */
+  int max_h_samp_factor;	/* largest h_samp_factor */
+  int max_v_samp_factor;	/* largest v_samp_factor */
+
+#if JPEG_LIB_VERSION >= 70
+  int min_DCT_h_scaled_size;	/* smallest DCT_h_scaled_size of any component */
+  int min_DCT_v_scaled_size;	/* smallest DCT_v_scaled_size of any component */
+#endif
+
+  JDIMENSION total_iMCU_rows;	/* # of iMCU rows to be input to coef ctlr */
+  /* The coefficient controller receives data in units of MCU rows as defined
+   * for fully interleaved scans (whether the JPEG file is interleaved or not).
+   * There are v_samp_factor * DCTSIZE sample rows of each component in an
+   * "iMCU" (interleaved MCU) row.
+   */
+  
+  /*
+   * These fields are valid during any one scan.
+   * They describe the components and MCUs actually appearing in the scan.
+   */
+  int comps_in_scan;		/* # of JPEG components in this scan */
+  jpeg_component_info * cur_comp_info[MAX_COMPS_IN_SCAN];
+  /* *cur_comp_info[i] describes component that appears i'th in SOS */
+  
+  JDIMENSION MCUs_per_row;	/* # of MCUs across the image */
+  JDIMENSION MCU_rows_in_scan;	/* # of MCU rows in the image */
+  
+  int blocks_in_MCU;		/* # of DCT blocks per MCU */
+  int MCU_membership[C_MAX_BLOCKS_IN_MCU];
+  /* MCU_membership[i] is index in cur_comp_info of component owning */
+  /* i'th block in an MCU */
+
+  int Ss, Se, Ah, Al;		/* progressive JPEG parameters for scan */
+
+#if JPEG_LIB_VERSION >= 80
+  int block_size;		/* the basic DCT block size: 1..16 */
+  const int * natural_order;	/* natural-order position array */
+  int lim_Se;			/* min( Se, DCTSIZE2-1 ) */
+#endif
+
+  /*
+   * Links to compression subobjects (methods and private variables of modules)
+   */
+  struct jpeg_comp_master * master;
+  struct jpeg_c_main_controller * main;
+  struct jpeg_c_prep_controller * prep;
+  struct jpeg_c_coef_controller * coef;
+  struct jpeg_marker_writer * marker;
+  struct jpeg_color_converter * cconvert;
+  struct jpeg_downsampler * downsample;
+  struct jpeg_forward_dct * fdct;
+  struct jpeg_entropy_encoder * entropy;
+  jpeg_scan_info * script_space; /* workspace for jpeg_simple_progression */
+  int script_space_size;
+};
+
+
+/* Master record for a decompression instance */
+
+struct jpeg_decompress_struct {
+  jpeg_common_fields;		/* Fields shared with jpeg_compress_struct */
+
+  /* Source of compressed data */
+  struct jpeg_source_mgr * src;
+
+  /* Basic description of image --- filled in by jpeg_read_header(). */
+  /* Application may inspect these values to decide how to process image. */
+
+  JDIMENSION image_width;	/* nominal image width (from SOF marker) */
+  JDIMENSION image_height;	/* nominal image height */
+  int num_components;		/* # of color components in JPEG image */
+  J_COLOR_SPACE jpeg_color_space; /* colorspace of JPEG image */
+
+  /* Decompression processing parameters --- these fields must be set before
+   * calling jpeg_start_decompress().  Note that jpeg_read_header() initializes
+   * them to default values.
+   */
+
+  J_COLOR_SPACE out_color_space; /* colorspace for output */
+
+  unsigned int scale_num, scale_denom; /* fraction by which to scale image */
+
+  double output_gamma;		/* image gamma wanted in output */
+
+  boolean buffered_image;	/* TRUE=multiple output passes */
+  boolean raw_data_out;		/* TRUE=downsampled data wanted */
+
+  J_DCT_METHOD dct_method;	/* IDCT algorithm selector */
+  boolean do_fancy_upsampling;	/* TRUE=apply fancy upsampling */
+  boolean do_block_smoothing;	/* TRUE=apply interblock smoothing */
+
+  boolean quantize_colors;	/* TRUE=colormapped output wanted */
+  /* the following are ignored if not quantize_colors: */
+  J_DITHER_MODE dither_mode;	/* type of color dithering to use */
+  boolean two_pass_quantize;	/* TRUE=use two-pass color quantization */
+  int desired_number_of_colors;	/* max # colors to use in created colormap */
+  /* these are significant only in buffered-image mode: */
+  boolean enable_1pass_quant;	/* enable future use of 1-pass quantizer */
+  boolean enable_external_quant;/* enable future use of external colormap */
+  boolean enable_2pass_quant;	/* enable future use of 2-pass quantizer */
+
+  /* Description of actual output image that will be returned to application.
+   * These fields are computed by jpeg_start_decompress().
+   * You can also use jpeg_calc_output_dimensions() to determine these values
+   * in advance of calling jpeg_start_decompress().
+   */
+
+  JDIMENSION output_width;	/* scaled image width */
+  JDIMENSION output_height;	/* scaled image height */
+  int out_color_components;	/* # of color components in out_color_space */
+  int output_components;	/* # of color components returned */
+  /* output_components is 1 (a colormap index) when quantizing colors;
+   * otherwise it equals out_color_components.
+   */
+  int rec_outbuf_height;	/* min recommended height of scanline buffer */
+  /* If the buffer passed to jpeg_read_scanlines() is less than this many rows
+   * high, space and time will be wasted due to unnecessary data copying.
+   * Usually rec_outbuf_height will be 1 or 2, at most 4.
+   */
+
+  /* When quantizing colors, the output colormap is described by these fields.
+   * The application can supply a colormap by setting colormap non-NULL before
+   * calling jpeg_start_decompress; otherwise a colormap is created during
+   * jpeg_start_decompress or jpeg_start_output.
+   * The map has out_color_components rows and actual_number_of_colors columns.
+   */
+  int actual_number_of_colors;	/* number of entries in use */
+  JSAMPARRAY colormap;		/* The color map as a 2-D pixel array */
+
+  /* State variables: these variables indicate the progress of decompression.
+   * The application may examine these but must not modify them.
+   */
+
+  /* Row index of next scanline to be read from jpeg_read_scanlines().
+   * Application may use this to control its processing loop, e.g.,
+   * "while (output_scanline < output_height)".
+   */
+  JDIMENSION output_scanline;	/* 0 .. output_height-1  */
+
+  /* Current input scan number and number of iMCU rows completed in scan.
+   * These indicate the progress of the decompressor input side.
+   */
+  int input_scan_number;	/* Number of SOS markers seen so far */
+  JDIMENSION input_iMCU_row;	/* Number of iMCU rows completed */
+
+  /* The "output scan number" is the notional scan being displayed by the
+   * output side.  The decompressor will not allow output scan/row number
+   * to get ahead of input scan/row, but it can fall arbitrarily far behind.
+   */
+  int output_scan_number;	/* Nominal scan number being displayed */
+  JDIMENSION output_iMCU_row;	/* Number of iMCU rows read */
+
+  /* Current progression status.  coef_bits[c][i] indicates the precision
+   * with which component c's DCT coefficient i (in zigzag order) is known.
+   * It is -1 when no data has yet been received, otherwise it is the point
+   * transform (shift) value for the most recent scan of the coefficient
+   * (thus, 0 at completion of the progression).
+   * This pointer is NULL when reading a non-progressive file.
+   */
+  int (*coef_bits)[DCTSIZE2];	/* -1 or current Al value for each coef */
+
+  /* Internal JPEG parameters --- the application usually need not look at
+   * these fields.  Note that the decompressor output side may not use
+   * any parameters that can change between scans.
+   */
+
+  /* Quantization and Huffman tables are carried forward across input
+   * datastreams when processing abbreviated JPEG datastreams.
+   */
+
+  JQUANT_TBL * quant_tbl_ptrs[NUM_QUANT_TBLS];
+  /* ptrs to coefficient quantization tables, or NULL if not defined */
+
+  JHUFF_TBL * dc_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  JHUFF_TBL * ac_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  /* ptrs to Huffman coding tables, or NULL if not defined */
+
+  /* These parameters are never carried across datastreams, since they
+   * are given in SOF/SOS markers or defined to be reset by SOI.
+   */
+
+  int data_precision;		/* bits of precision in image data */
+
+  jpeg_component_info * comp_info;
+  /* comp_info[i] describes component that appears i'th in SOF */
+
+#if JPEG_LIB_VERSION >= 80
+  boolean is_baseline;		/* TRUE if Baseline SOF0 encountered */
+#endif
+  boolean progressive_mode;	/* TRUE if SOFn specifies progressive mode */
+  boolean arith_code;		/* TRUE=arithmetic coding, FALSE=Huffman */
+
+  UINT8 arith_dc_L[NUM_ARITH_TBLS]; /* L values for DC arith-coding tables */
+  UINT8 arith_dc_U[NUM_ARITH_TBLS]; /* U values for DC arith-coding tables */
+  UINT8 arith_ac_K[NUM_ARITH_TBLS]; /* Kx values for AC arith-coding tables */
+
+  unsigned int restart_interval; /* MCUs per restart interval, or 0 for no restart */
+
+  /* These fields record data obtained from optional markers recognized by
+   * the JPEG library.
+   */
+  boolean saw_JFIF_marker;	/* TRUE iff a JFIF APP0 marker was found */
+  /* Data copied from JFIF marker; only valid if saw_JFIF_marker is TRUE: */
+  UINT8 JFIF_major_version;	/* JFIF version number */
+  UINT8 JFIF_minor_version;
+  UINT8 density_unit;		/* JFIF code for pixel size units */
+  UINT16 X_density;		/* Horizontal pixel density */
+  UINT16 Y_density;		/* Vertical pixel density */
+  boolean saw_Adobe_marker;	/* TRUE iff an Adobe APP14 marker was found */
+  UINT8 Adobe_transform;	/* Color transform code from Adobe marker */
+
+  boolean CCIR601_sampling;	/* TRUE=first samples are cosited */
+
+  /* Aside from the specific data retained from APPn markers known to the
+   * library, the uninterpreted contents of any or all APPn and COM markers
+   * can be saved in a list for examination by the application.
+   */
+  jpeg_saved_marker_ptr marker_list; /* Head of list of saved markers */
+
+  /* Remaining fields are known throughout decompressor, but generally
+   * should not be touched by a surrounding application.
+   */
+
+  /*
+   * These fields are computed during decompression startup
+   */
+  int max_h_samp_factor;	/* largest h_samp_factor */
+  int max_v_samp_factor;	/* largest v_samp_factor */
+
+#if JPEG_LIB_VERSION >= 70
+  int min_DCT_h_scaled_size;	/* smallest DCT_h_scaled_size of any component */
+  int min_DCT_v_scaled_size;	/* smallest DCT_v_scaled_size of any component */
+#else
+  int min_DCT_scaled_size;	/* smallest DCT_scaled_size of any component */
+#endif
+
+  JDIMENSION total_iMCU_rows;	/* # of iMCU rows in image */
+  /* The coefficient controller's input and output progress is measured in
+   * units of "iMCU" (interleaved MCU) rows.  These are the same as MCU rows
+   * in fully interleaved JPEG scans, but are used whether the scan is
+   * interleaved or not.  We define an iMCU row as v_samp_factor DCT block
+   * rows of each component.  Therefore, the IDCT output contains
+   * v_samp_factor*DCT_[v_]scaled_size sample rows of a component per iMCU row.
+   */
+
+  JSAMPLE * sample_range_limit; /* table for fast range-limiting */
+
+  /*
+   * These fields are valid during any one scan.
+   * They describe the components and MCUs actually appearing in the scan.
+   * Note that the decompressor output side must not use these fields.
+   */
+  int comps_in_scan;		/* # of JPEG components in this scan */
+  jpeg_component_info * cur_comp_info[MAX_COMPS_IN_SCAN];
+  /* *cur_comp_info[i] describes component that appears i'th in SOS */
+
+  JDIMENSION MCUs_per_row;	/* # of MCUs across the image */
+  JDIMENSION MCU_rows_in_scan;	/* # of MCU rows in the image */
+
+  int blocks_in_MCU;		/* # of DCT blocks per MCU */
+  int MCU_membership[D_MAX_BLOCKS_IN_MCU];
+  /* MCU_membership[i] is index in cur_comp_info of component owning */
+  /* i'th block in an MCU */
+
+  int Ss, Se, Ah, Al;		/* progressive JPEG parameters for scan */
+
+#if JPEG_LIB_VERSION >= 80
+  /* These fields are derived from Se of first SOS marker.
+   */
+  int block_size;		/* the basic DCT block size: 1..16 */
+  const int * natural_order; /* natural-order position array for entropy decode */
+  int lim_Se;			/* min( Se, DCTSIZE2-1 ) for entropy decode */
+#endif
+
+  /* This field is shared between entropy decoder and marker parser.
+   * It is either zero or the code of a JPEG marker that has been
+   * read from the data source, but has not yet been processed.
+   */
+  int unread_marker;
+
+  /*
+   * Links to decompression subobjects (methods, private variables of modules)
+   */
+  struct jpeg_decomp_master * master;
+  struct jpeg_d_main_controller * main;
+  struct jpeg_d_coef_controller * coef;
+  struct jpeg_d_post_controller * post;
+  struct jpeg_input_controller * inputctl;
+  struct jpeg_marker_reader * marker;
+  struct jpeg_entropy_decoder * entropy;
+  struct jpeg_inverse_dct * idct;
+  struct jpeg_upsampler * upsample;
+  struct jpeg_color_deconverter * cconvert;
+  struct jpeg_color_quantizer * cquantize;
+};
+
+
+/* "Object" declarations for JPEG modules that may be supplied or called
+ * directly by the surrounding application.
+ * As with all objects in the JPEG library, these structs only define the
+ * publicly visible methods and state variables of a module.  Additional
+ * private fields may exist after the public ones.
+ */
+
+
+/* Error handler object */
+
+struct jpeg_error_mgr {
+  /* Error exit handler: does not return to caller */
+  JMETHOD(void, error_exit, (j_common_ptr cinfo));
+  /* Conditionally emit a trace or warning message */
+  JMETHOD(void, emit_message, (j_common_ptr cinfo, int msg_level));
+  /* Routine that actually outputs a trace or error message */
+  JMETHOD(void, output_message, (j_common_ptr cinfo));
+  /* Format a message string for the most recent JPEG error or message */
+  JMETHOD(void, format_message, (j_common_ptr cinfo, char * buffer));
+#define JMSG_LENGTH_MAX  200	/* recommended size of format_message buffer */
+  /* Reset error state variables at start of a new image */
+  JMETHOD(void, reset_error_mgr, (j_common_ptr cinfo));
+  
+  /* The message ID code and any parameters are saved here.
+   * A message can have one string parameter or up to 8 int parameters.
+   */
+  int msg_code;
+#define JMSG_STR_PARM_MAX  80
+  union {
+    int i[8];
+    char s[JMSG_STR_PARM_MAX];
+  } msg_parm;
+  
+  /* Standard state variables for error facility */
+  
+  int trace_level;		/* max msg_level that will be displayed */
+  
+  /* For recoverable corrupt-data errors, we emit a warning message,
+   * but keep going unless emit_message chooses to abort.  emit_message
+   * should count warnings in num_warnings.  The surrounding application
+   * can check for bad data by seeing if num_warnings is nonzero at the
+   * end of processing.
+   */
+  long num_warnings;		/* number of corrupt-data warnings */
+
+  /* These fields point to the table(s) of error message strings.
+   * An application can change the table pointer to switch to a different
+   * message list (typically, to change the language in which errors are
+   * reported).  Some applications may wish to add additional error codes
+   * that will be handled by the JPEG library error mechanism; the second
+   * table pointer is used for this purpose.
+   *
+   * First table includes all errors generated by JPEG library itself.
+   * Error code 0 is reserved for a "no such error string" message.
+   */
+  const char * const * jpeg_message_table; /* Library errors */
+  int last_jpeg_message;    /* Table contains strings 0..last_jpeg_message */
+  /* Second table can be added by application (see cjpeg/djpeg for example).
+   * It contains strings numbered first_addon_message..last_addon_message.
+   */
+  const char * const * addon_message_table; /* Non-library errors */
+  int first_addon_message;	/* code for first string in addon table */
+  int last_addon_message;	/* code for last string in addon table */
+};
+
+
+/* Progress monitor object */
+
+struct jpeg_progress_mgr {
+  JMETHOD(void, progress_monitor, (j_common_ptr cinfo));
+
+  long pass_counter;		/* work units completed in this pass */
+  long pass_limit;		/* total number of work units in this pass */
+  int completed_passes;		/* passes completed so far */
+  int total_passes;		/* total number of passes expected */
+};
+
+
+/* Data destination object for compression */
+
+struct jpeg_destination_mgr {
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+
+  JMETHOD(void, init_destination, (j_compress_ptr cinfo));
+  JMETHOD(boolean, empty_output_buffer, (j_compress_ptr cinfo));
+  JMETHOD(void, term_destination, (j_compress_ptr cinfo));
+};
+
+
+/* Data source object for decompression */
+
+struct jpeg_source_mgr {
+  const JOCTET * next_input_byte; /* => next byte to read from buffer */
+  size_t bytes_in_buffer;	/* # of bytes remaining in buffer */
+
+  JMETHOD(void, init_source, (j_decompress_ptr cinfo));
+  JMETHOD(boolean, fill_input_buffer, (j_decompress_ptr cinfo));
+  JMETHOD(void, skip_input_data, (j_decompress_ptr cinfo, long num_bytes));
+  JMETHOD(boolean, resync_to_restart, (j_decompress_ptr cinfo, int desired));
+  JMETHOD(void, term_source, (j_decompress_ptr cinfo));
+};
+
+
+/* Memory manager object.
+ * Allocates "small" objects (a few K total), "large" objects (tens of K),
+ * and "really big" objects (virtual arrays with backing store if needed).
+ * The memory manager does not allow individual objects to be freed; rather,
+ * each created object is assigned to a pool, and whole pools can be freed
+ * at once.  This is faster and more convenient than remembering exactly what
+ * to free, especially where malloc()/free() are not too speedy.
+ * NB: alloc routines never return NULL.  They exit to error_exit if not
+ * successful.
+ */
+
+#define JPOOL_PERMANENT	0	/* lasts until master record is destroyed */
+#define JPOOL_IMAGE	1	/* lasts until done with image/datastream */
+#define JPOOL_NUMPOOLS	2
+
+typedef struct jvirt_sarray_control * jvirt_sarray_ptr;
+typedef struct jvirt_barray_control * jvirt_barray_ptr;
+
+
+struct jpeg_memory_mgr {
+  /* Method pointers */
+  JMETHOD(void *, alloc_small, (j_common_ptr cinfo, int pool_id,
+				size_t sizeofobject));
+  JMETHOD(void FAR *, alloc_large, (j_common_ptr cinfo, int pool_id,
+				     size_t sizeofobject));
+  JMETHOD(JSAMPARRAY, alloc_sarray, (j_common_ptr cinfo, int pool_id,
+				     JDIMENSION samplesperrow,
+				     JDIMENSION numrows));
+  JMETHOD(JBLOCKARRAY, alloc_barray, (j_common_ptr cinfo, int pool_id,
+				      JDIMENSION blocksperrow,
+				      JDIMENSION numrows));
+  JMETHOD(jvirt_sarray_ptr, request_virt_sarray, (j_common_ptr cinfo,
+						  int pool_id,
+						  boolean pre_zero,
+						  JDIMENSION samplesperrow,
+						  JDIMENSION numrows,
+						  JDIMENSION maxaccess));
+  JMETHOD(jvirt_barray_ptr, request_virt_barray, (j_common_ptr cinfo,
+						  int pool_id,
+						  boolean pre_zero,
+						  JDIMENSION blocksperrow,
+						  JDIMENSION numrows,
+						  JDIMENSION maxaccess));
+  JMETHOD(void, realize_virt_arrays, (j_common_ptr cinfo));
+  JMETHOD(JSAMPARRAY, access_virt_sarray, (j_common_ptr cinfo,
+					   jvirt_sarray_ptr ptr,
+					   JDIMENSION start_row,
+					   JDIMENSION num_rows,
+					   boolean writable));
+  JMETHOD(JBLOCKARRAY, access_virt_barray, (j_common_ptr cinfo,
+					    jvirt_barray_ptr ptr,
+					    JDIMENSION start_row,
+					    JDIMENSION num_rows,
+					    boolean writable));
+  JMETHOD(void, free_pool, (j_common_ptr cinfo, int pool_id));
+  JMETHOD(void, self_destruct, (j_common_ptr cinfo));
+
+  /* Limit on memory allocation for this JPEG object.  (Note that this is
+   * merely advisory, not a guaranteed maximum; it only affects the space
+   * used for virtual-array buffers.)  May be changed by outer application
+   * after creating the JPEG object.
+   */
+  long max_memory_to_use;
+
+  /* Maximum allocation request accepted by alloc_large. */
+  long max_alloc_chunk;
+};
+
+
+/* Routine signature for application-supplied marker processing methods.
+ * Need not pass marker code since it is stored in cinfo->unread_marker.
+ */
+typedef JMETHOD(boolean, jpeg_marker_parser_method, (j_decompress_ptr cinfo));
+
+
+/* Declarations for routines called by application.
+ * The JPP macro hides prototype parameters from compilers that can't cope.
+ * Note JPP requires double parentheses.
+ */
+
+#ifdef HAVE_PROTOTYPES
+#define JPP(arglist)	arglist
+#else
+#define JPP(arglist)	()
+#endif
+
+
+/* Short forms of external names for systems with brain-damaged linkers.
+ * We shorten external names to be unique in the first six letters, which
+ * is good enough for all known systems.
+ * (If your compiler itself needs names to be unique in less than 15 
+ * characters, you are out of luck.  Get a better compiler.)
+ */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_std_error		jStdError
+#define jpeg_CreateCompress	jCreaCompress
+#define jpeg_CreateDecompress	jCreaDecompress
+#define jpeg_destroy_compress	jDestCompress
+#define jpeg_destroy_decompress	jDestDecompress
+#define jpeg_stdio_dest		jStdDest
+#define jpeg_stdio_src		jStdSrc
+#if JPEG_LIB_VERSION >= 80
+#define jpeg_mem_dest		jMemDest
+#define jpeg_mem_src		jMemSrc
+#endif
+#define jpeg_set_defaults	jSetDefaults
+#define jpeg_set_colorspace	jSetColorspace
+#define jpeg_default_colorspace	jDefColorspace
+#define jpeg_set_quality	jSetQuality
+#define jpeg_set_linear_quality	jSetLQuality
+#if JPEG_LIB_VERSION >= 70
+#define jpeg_default_qtables	jDefQTables
+#endif
+#define jpeg_add_quant_table	jAddQuantTable
+#define jpeg_quality_scaling	jQualityScaling
+#define jpeg_simple_progression	jSimProgress
+#define jpeg_suppress_tables	jSuppressTables
+#define jpeg_alloc_quant_table	jAlcQTable
+#define jpeg_alloc_huff_table	jAlcHTable
+#define jpeg_start_compress	jStrtCompress
+#define jpeg_write_scanlines	jWrtScanlines
+#define jpeg_finish_compress	jFinCompress
+#if JPEG_LIB_VERSION >= 70
+#define jpeg_calc_jpeg_dimensions	jCjpegDimensions
+#endif
+#define jpeg_write_raw_data	jWrtRawData
+#define jpeg_write_marker	jWrtMarker
+#define jpeg_write_m_header	jWrtMHeader
+#define jpeg_write_m_byte	jWrtMByte
+#define jpeg_write_tables	jWrtTables
+#define jpeg_read_header	jReadHeader
+#define jpeg_start_decompress	jStrtDecompress
+#define jpeg_read_scanlines	jReadScanlines
+#define jpeg_finish_decompress	jFinDecompress
+#define jpeg_read_raw_data	jReadRawData
+#define jpeg_has_multiple_scans	jHasMultScn
+#define jpeg_start_output	jStrtOutput
+#define jpeg_finish_output	jFinOutput
+#define jpeg_input_complete	jInComplete
+#define jpeg_new_colormap	jNewCMap
+#define jpeg_consume_input	jConsumeInput
+#if JPEG_LIB_VERSION >= 80
+#define jpeg_core_output_dimensions	jCoreDimensions
+#endif
+#define jpeg_calc_output_dimensions	jCalcDimensions
+#define jpeg_save_markers	jSaveMarkers
+#define jpeg_set_marker_processor	jSetMarker
+#define jpeg_read_coefficients	jReadCoefs
+#define jpeg_write_coefficients	jWrtCoefs
+#define jpeg_copy_critical_parameters	jCopyCrit
+#define jpeg_abort_compress	jAbrtCompress
+#define jpeg_abort_decompress	jAbrtDecompress
+#define jpeg_abort		jAbort
+#define jpeg_destroy		jDestroy
+#define jpeg_resync_to_restart	jResyncRestart
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/* Default error-management setup */
+EXTERN(struct jpeg_error_mgr *) jpeg_std_error
+	JPP((struct jpeg_error_mgr * err));
+
+/* Initialization of JPEG compression objects.
+ * jpeg_create_compress() and jpeg_create_decompress() are the exported
+ * names that applications should call.  These expand to calls on
+ * jpeg_CreateCompress and jpeg_CreateDecompress with additional information
+ * passed for version mismatch checking.
+ * NB: you must set up the error-manager BEFORE calling jpeg_create_xxx.
+ */
+#define jpeg_create_compress(cinfo) \
+    jpeg_CreateCompress((cinfo), JPEG_LIB_VERSION, \
+			(size_t) sizeof(struct jpeg_compress_struct))
+#define jpeg_create_decompress(cinfo) \
+    jpeg_CreateDecompress((cinfo), JPEG_LIB_VERSION, \
+			  (size_t) sizeof(struct jpeg_decompress_struct))
+EXTERN(void) jpeg_CreateCompress JPP((j_compress_ptr cinfo,
+				      int version, size_t structsize));
+EXTERN(void) jpeg_CreateDecompress JPP((j_decompress_ptr cinfo,
+					int version, size_t structsize));
+/* Destruction of JPEG compression objects */
+EXTERN(void) jpeg_destroy_compress JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_destroy_decompress JPP((j_decompress_ptr cinfo));
+
+/* Standard data source and destination managers: stdio streams. */
+/* Caller is responsible for opening the file before and closing after. */
+EXTERN(void) jpeg_stdio_dest JPP((j_compress_ptr cinfo, FILE * outfile));
+EXTERN(void) jpeg_stdio_src JPP((j_decompress_ptr cinfo, FILE * infile));
+
+#if JPEG_LIB_VERSION >= 80
+/* Data source and destination managers: memory buffers. */
+EXTERN(void) jpeg_mem_dest JPP((j_compress_ptr cinfo,
+			       unsigned char ** outbuffer,
+			       unsigned long * outsize));
+EXTERN(void) jpeg_mem_src JPP((j_decompress_ptr cinfo,
+			      unsigned char * inbuffer,
+			      unsigned long insize));
+#endif
+
+/* Default parameter setup for compression */
+EXTERN(void) jpeg_set_defaults JPP((j_compress_ptr cinfo));
+/* Compression parameter setup aids */
+EXTERN(void) jpeg_set_colorspace JPP((j_compress_ptr cinfo,
+				      J_COLOR_SPACE colorspace));
+EXTERN(void) jpeg_default_colorspace JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_set_quality JPP((j_compress_ptr cinfo, int quality,
+				   boolean force_baseline));
+EXTERN(void) jpeg_set_linear_quality JPP((j_compress_ptr cinfo,
+					  int scale_factor,
+					  boolean force_baseline));
+#if JPEG_LIB_VERSION >= 70
+EXTERN(void) jpeg_default_qtables JPP((j_compress_ptr cinfo,
+				       boolean force_baseline));
+#endif
+EXTERN(void) jpeg_add_quant_table JPP((j_compress_ptr cinfo, int which_tbl,
+				       const unsigned int *basic_table,
+				       int scale_factor,
+				       boolean force_baseline));
+EXTERN(int) jpeg_quality_scaling JPP((int quality));
+EXTERN(void) jpeg_simple_progression JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_suppress_tables JPP((j_compress_ptr cinfo,
+				       boolean suppress));
+EXTERN(JQUANT_TBL *) jpeg_alloc_quant_table JPP((j_common_ptr cinfo));
+EXTERN(JHUFF_TBL *) jpeg_alloc_huff_table JPP((j_common_ptr cinfo));
+
+/* Main entry points for compression */
+EXTERN(void) jpeg_start_compress JPP((j_compress_ptr cinfo,
+				      boolean write_all_tables));
+EXTERN(JDIMENSION) jpeg_write_scanlines JPP((j_compress_ptr cinfo,
+					     JSAMPARRAY scanlines,
+					     JDIMENSION num_lines));
+EXTERN(void) jpeg_finish_compress JPP((j_compress_ptr cinfo));
+
+#if JPEG_LIB_VERSION >= 70
+/* Precalculate JPEG dimensions for current compression parameters. */
+EXTERN(void) jpeg_calc_jpeg_dimensions JPP((j_compress_ptr cinfo));
+#endif
+
+/* Replaces jpeg_write_scanlines when writing raw downsampled data. */
+EXTERN(JDIMENSION) jpeg_write_raw_data JPP((j_compress_ptr cinfo,
+					    JSAMPIMAGE data,
+					    JDIMENSION num_lines));
+
+/* Write a special marker.  See libjpeg.txt concerning safe usage. */
+EXTERN(void) jpeg_write_marker
+	JPP((j_compress_ptr cinfo, int marker,
+	     const JOCTET * dataptr, unsigned int datalen));
+/* Same, but piecemeal. */
+EXTERN(void) jpeg_write_m_header
+	JPP((j_compress_ptr cinfo, int marker, unsigned int datalen));
+EXTERN(void) jpeg_write_m_byte
+	JPP((j_compress_ptr cinfo, int val));
+
+/* Alternate compression function: just write an abbreviated table file */
+EXTERN(void) jpeg_write_tables JPP((j_compress_ptr cinfo));
+
+/* Decompression startup: read start of JPEG datastream to see what's there */
+EXTERN(int) jpeg_read_header JPP((j_decompress_ptr cinfo,
+				  boolean require_image));
+/* Return value is one of: */
+#define JPEG_SUSPENDED		0 /* Suspended due to lack of input data */
+#define JPEG_HEADER_OK		1 /* Found valid image datastream */
+#define JPEG_HEADER_TABLES_ONLY	2 /* Found valid table-specs-only datastream */
+/* If you pass require_image = TRUE (normal case), you need not check for
+ * a TABLES_ONLY return code; an abbreviated file will cause an error exit.
+ * JPEG_SUSPENDED is only possible if you use a data source module that can
+ * give a suspension return (the stdio source module doesn't).
+ */
+
+/* Main entry points for decompression */
+EXTERN(boolean) jpeg_start_decompress JPP((j_decompress_ptr cinfo));
+EXTERN(JDIMENSION) jpeg_read_scanlines JPP((j_decompress_ptr cinfo,
+					    JSAMPARRAY scanlines,
+					    JDIMENSION max_lines));
+EXTERN(boolean) jpeg_finish_decompress JPP((j_decompress_ptr cinfo));
+
+/* Replaces jpeg_read_scanlines when reading raw downsampled data. */
+EXTERN(JDIMENSION) jpeg_read_raw_data JPP((j_decompress_ptr cinfo,
+					   JSAMPIMAGE data,
+					   JDIMENSION max_lines));
+
+/* Additional entry points for buffered-image mode. */
+EXTERN(boolean) jpeg_has_multiple_scans JPP((j_decompress_ptr cinfo));
+EXTERN(boolean) jpeg_start_output JPP((j_decompress_ptr cinfo,
+				       int scan_number));
+EXTERN(boolean) jpeg_finish_output JPP((j_decompress_ptr cinfo));
+EXTERN(boolean) jpeg_input_complete JPP((j_decompress_ptr cinfo));
+EXTERN(void) jpeg_new_colormap JPP((j_decompress_ptr cinfo));
+EXTERN(int) jpeg_consume_input JPP((j_decompress_ptr cinfo));
+/* Return value is one of: */
+/* #define JPEG_SUSPENDED	0    Suspended due to lack of input data */
+#define JPEG_REACHED_SOS	1 /* Reached start of new scan */
+#define JPEG_REACHED_EOI	2 /* Reached end of image */
+#define JPEG_ROW_COMPLETED	3 /* Completed one iMCU row */
+#define JPEG_SCAN_COMPLETED	4 /* Completed last iMCU row of a scan */
+
+/* Precalculate output dimensions for current decompression parameters. */
+#if JPEG_LIB_VERSION >= 80
+EXTERN(void) jpeg_core_output_dimensions JPP((j_decompress_ptr cinfo));
+#endif
+EXTERN(void) jpeg_calc_output_dimensions JPP((j_decompress_ptr cinfo));
+
+/* Control saving of COM and APPn markers into marker_list. */
+EXTERN(void) jpeg_save_markers
+	JPP((j_decompress_ptr cinfo, int marker_code,
+	     unsigned int length_limit));
+
+/* Install a special processing method for COM or APPn markers. */
+EXTERN(void) jpeg_set_marker_processor
+	JPP((j_decompress_ptr cinfo, int marker_code,
+	     jpeg_marker_parser_method routine));
+
+/* Read or write raw DCT coefficients --- useful for lossless transcoding. */
+EXTERN(jvirt_barray_ptr *) jpeg_read_coefficients JPP((j_decompress_ptr cinfo));
+EXTERN(void) jpeg_write_coefficients JPP((j_compress_ptr cinfo,
+					  jvirt_barray_ptr * coef_arrays));
+EXTERN(void) jpeg_copy_critical_parameters JPP((j_decompress_ptr srcinfo,
+						j_compress_ptr dstinfo));
+
+/* If you choose to abort compression or decompression before completing
+ * jpeg_finish_(de)compress, then you need to clean up to release memory,
+ * temporary files, etc.  You can just call jpeg_destroy_(de)compress
+ * if you're done with the JPEG object, but if you want to clean it up and
+ * reuse it, call this:
+ */
+EXTERN(void) jpeg_abort_compress JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_abort_decompress JPP((j_decompress_ptr cinfo));
+
+/* Generic versions of jpeg_abort and jpeg_destroy that work on either
+ * flavor of JPEG object.  These may be more convenient in some places.
+ */
+EXTERN(void) jpeg_abort JPP((j_common_ptr cinfo));
+EXTERN(void) jpeg_destroy JPP((j_common_ptr cinfo));
+
+/* Default restart-marker-resync procedure for use by data source modules */
+EXTERN(boolean) jpeg_resync_to_restart JPP((j_decompress_ptr cinfo,
+					    int desired));
+
+
+/* These marker codes are exported since applications and data source modules
+ * are likely to want to use them.
+ */
+
+#define JPEG_RST0	0xD0	/* RST0 marker code */
+#define JPEG_EOI	0xD9	/* EOI marker code */
+#define JPEG_APP0	0xE0	/* APP0 marker code */
+#define JPEG_COM	0xFE	/* COM marker code */
+
+
+/* If we have a brain-damaged compiler that emits warnings (or worse, errors)
+ * for structure definitions that are never filled in, keep it quiet by
+ * supplying dummy definitions for the various substructures.
+ */
+
+#ifdef INCOMPLETE_TYPES_BROKEN
+#ifndef JPEG_INTERNALS		/* will be defined in jpegint.h */
+struct jvirt_sarray_control { long dummy; };
+struct jvirt_barray_control { long dummy; };
+struct jpeg_comp_master { long dummy; };
+struct jpeg_c_main_controller { long dummy; };
+struct jpeg_c_prep_controller { long dummy; };
+struct jpeg_c_coef_controller { long dummy; };
+struct jpeg_marker_writer { long dummy; };
+struct jpeg_color_converter { long dummy; };
+struct jpeg_downsampler { long dummy; };
+struct jpeg_forward_dct { long dummy; };
+struct jpeg_entropy_encoder { long dummy; };
+struct jpeg_decomp_master { long dummy; };
+struct jpeg_d_main_controller { long dummy; };
+struct jpeg_d_coef_controller { long dummy; };
+struct jpeg_d_post_controller { long dummy; };
+struct jpeg_input_controller { long dummy; };
+struct jpeg_marker_reader { long dummy; };
+struct jpeg_entropy_decoder { long dummy; };
+struct jpeg_inverse_dct { long dummy; };
+struct jpeg_upsampler { long dummy; };
+struct jpeg_color_deconverter { long dummy; };
+struct jpeg_color_quantizer { long dummy; };
+#endif /* JPEG_INTERNALS */
+#endif /* INCOMPLETE_TYPES_BROKEN */
+
+
+/*
+ * The JPEG library modules define JPEG_INTERNALS before including this file.
+ * The internal structure declarations are read only when that is true.
+ * Applications using the library should not include jpegint.h, but may wish
+ * to include jerror.h.
+ */
+
+#ifdef JPEG_INTERNALS
+#include "jpegint.h"		/* fetch private declarations */
+#include "jerror.h"		/* fetch error codes too */
+#endif
+
+#ifdef __cplusplus
+#ifndef DONT_USE_EXTERN_C
+}
+#endif
+#endif
+
+#endif /* JPEGLIB_H */
diff --git a/jpegtran.1 b/jpegtran.1
new file mode 100644
index 0000000..66d2f9f
--- /dev/null
+++ b/jpegtran.1
@@ -0,0 +1,262 @@
+.TH JPEGTRAN 1 "11 October 2010"
+.SH NAME
+jpegtran \- lossless transformation of JPEG files
+.SH SYNOPSIS
+.B jpegtran
+[
+.I options
+]
+[
+.I filename
+]
+.LP
+.SH DESCRIPTION
+.LP
+.B jpegtran
+performs various useful transformations of JPEG files.
+It can translate the coded representation from one variant of JPEG to another,
+for example from baseline JPEG to progressive JPEG or vice versa.  It can also
+perform some rearrangements of the image data, for example turning an image
+from landscape to portrait format by rotation.
+.PP
+.B jpegtran
+works by rearranging the compressed data (DCT coefficients), without
+ever fully decoding the image.  Therefore, its transformations are lossless:
+there is no image degradation at all, which would not be true if you used
+.B djpeg
+followed by
+.B cjpeg
+to accomplish the same conversion.  But by the same token,
+.B jpegtran
+cannot perform lossy operations such as changing the image quality.
+.PP
+.B jpegtran
+reads the named JPEG/JFIF file, or the standard input if no file is
+named, and produces a JPEG/JFIF file on the standard output.
+.SH OPTIONS
+All switch names may be abbreviated; for example,
+.B \-optimize
+may be written
+.B \-opt
+or
+.BR \-o .
+Upper and lower case are equivalent.
+British spellings are also accepted (e.g.,
+.BR \-optimise ),
+though for brevity these are not mentioned below.
+.PP
+To specify the coded JPEG representation used in the output file,
+.B jpegtran
+accepts a subset of the switches recognized by
+.BR cjpeg :
+.TP
+.B \-optimize
+Perform optimization of entropy encoding parameters.
+.TP
+.B \-progressive
+Create progressive JPEG file.
+.TP
+.BI \-restart " N"
+Emit a JPEG restart marker every N MCU rows, or every N MCU blocks if "B" is
+attached to the number.
+.TP
+.BI \-scans " file"
+Use the scan script given in the specified text file.
+.PP
+See
+.BR cjpeg (1)
+for more details about these switches.
+If you specify none of these switches, you get a plain baseline-JPEG output
+file.  The quality setting and so forth are determined by the input file.
+.PP
+The image can be losslessly transformed by giving one of these switches:
+.TP
+.B \-flip horizontal
+Mirror image horizontally (left-right).
+.TP
+.B \-flip vertical
+Mirror image vertically (top-bottom).
+.TP
+.B \-rotate 90
+Rotate image 90 degrees clockwise.
+.TP
+.B \-rotate 180
+Rotate image 180 degrees.
+.TP
+.B \-rotate 270
+Rotate image 270 degrees clockwise (or 90 ccw).
+.TP
+.B \-transpose
+Transpose image (across UL-to-LR axis).
+.TP
+.B \-transverse
+Transverse transpose (across UR-to-LL axis).
+.PP
+The transpose transformation has no restrictions regarding image dimensions.
+The other transformations operate rather oddly if the image dimensions are not
+a multiple of the iMCU size (usually 8 or 16 pixels), because they can only
+transform complete blocks of DCT coefficient data in the desired way.
+.PP
+.BR jpegtran 's
+default behavior when transforming an odd-size image is designed
+to preserve exact reversibility and mathematical consistency of the
+transformation set.  As stated, transpose is able to flip the entire image
+area.  Horizontal mirroring leaves any partial iMCU column at the right edge
+untouched, but is able to flip all rows of the image.  Similarly, vertical
+mirroring leaves any partial iMCU row at the bottom edge untouched, but is
+able to flip all columns.  The other transforms can be built up as sequences
+of transpose and flip operations; for consistency, their actions on edge
+pixels are defined to be the same as the end result of the corresponding
+transpose-and-flip sequence.
+.PP
+For practical use, you may prefer to discard any untransformable edge pixels
+rather than having a strange-looking strip along the right and/or bottom edges
+of a transformed image.  To do this, add the
+.B \-trim
+switch:
+.TP
+.B \-trim
+Drop non-transformable edge blocks.
+.IP
+Obviously, a transformation with
+.B \-trim
+is not reversible, so strictly speaking
+.B jpegtran
+with this switch is not lossless.  Also, the expected mathematical
+equivalences between the transformations no longer hold.  For example,
+.B \-rot 270 -trim
+trims only the bottom edge, but
+.B \-rot 90 -trim
+followed by
+.B \-rot 180 -trim
+trims both edges.
+.TP
+.B \-perfect
+If you are only interested in perfect transformations, add the
+.B \-perfect
+switch.  This causes
+.B jpegtran
+to fail with an error if the transformation is not perfect.
+.IP
+For example, you may want to do
+.IP
+.B (jpegtran \-rot 90 -perfect
+.I foo.jpg
+.B || djpeg
+.I foo.jpg
+.B | pnmflip \-r90 | cjpeg)
+.IP
+to do a perfect rotation, if available, or an approximated one if not.
+.TP
+.B \-crop WxH+X+Y
+Crop the image to a rectangular region of width W and height H, starting at
+point X,Y.  The lossless crop feature discards data outside of a given image
+region but losslessly preserves what is inside.  Like the rotate and flip
+transforms, lossless crop is restricted by the current JPEG format; the upper
+left corner of the selected region must fall on an iMCU boundary.  If it
+doesn't, then it is silently moved up and/or left to the nearest iMCU boundary
+(the lower right corner is unchanged.)
+.PP
+Other not-strictly-lossless transformation switches are:
+.TP
+.B \-grayscale
+Force grayscale output.
+.IP
+This option discards the chrominance channels if the input image is YCbCr
+(ie, a standard color JPEG), resulting in a grayscale JPEG file.  The
+luminance channel is preserved exactly, so this is a better method of reducing
+to grayscale than decompression, conversion, and recompression.  This switch
+is particularly handy for fixing a monochrome picture that was mistakenly
+encoded as a color JPEG.  (In such a case, the space savings from getting rid
+of the near-empty chroma channels won't be large; but the decoding time for
+a grayscale JPEG is substantially less than that for a color JPEG.)
+.PP
+.B jpegtran
+also recognizes these switches that control what to do with "extra" markers,
+such as comment blocks:
+.TP
+.B \-copy none
+Copy no extra markers from source file.  This setting suppresses all
+comments and other excess baggage present in the source file.
+.TP
+.B \-copy comments
+Copy only comment markers.  This setting copies comments from the source file
+but discards any other data which is inessential for image display.
+.TP
+.B \-copy all
+Copy all extra markers.  This setting preserves miscellaneous markers
+found in the source file, such as JFIF thumbnails, Exif data, and Photoshop
+settings.  In some files, these extra markers can be sizable.
+.PP
+The default behavior is \fB-copy comments\fR.  (Note: in IJG releases v6 and
+v6a, \fBjpegtran\fR always did the equivalent of \fB-copy none\fR.)
+.PP
+Additional switches recognized by jpegtran are:
+.TP
+.BI \-maxmemory " N"
+Set limit for amount of memory to use in processing large images.  Value is
+in thousands of bytes, or millions of bytes if "M" is attached to the
+number.  For example,
+.B \-max 4m
+selects 4000000 bytes.  If more space is needed, temporary files will be used.
+.TP
+.BI \-outfile " name"
+Send output image to the named file, not to standard output.
+.TP
+.B \-verbose
+Enable debug printout.  More
+.BR \-v 's
+give more output.  Also, version information is printed at startup.
+.TP
+.B \-debug
+Same as
+.BR \-verbose .
+.SH EXAMPLES
+.LP
+This example converts a baseline JPEG file to progressive form:
+.IP
+.B jpegtran \-progressive
+.I foo.jpg
+.B >
+.I fooprog.jpg
+.PP
+This example rotates an image 90 degrees clockwise, discarding any
+unrotatable edge pixels:
+.IP
+.B jpegtran \-rot 90 -trim
+.I foo.jpg
+.B >
+.I foo90.jpg
+.SH ENVIRONMENT
+.TP
+.B JPEGMEM
+If this environment variable is set, its value is the default memory limit.
+The value is specified as described for the
+.B \-maxmemory
+switch.
+.B JPEGMEM
+overrides the default value specified when the program was compiled, and
+itself is overridden by an explicit
+.BR \-maxmemory .
+.SH SEE ALSO
+.BR cjpeg (1),
+.BR djpeg (1),
+.BR rdjpgcom (1),
+.BR wrjpgcom (1)
+.br
+Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+Communications of the ACM, April 1991 (vol. 34, no. 4), pp. 30-44.
+.SH AUTHOR
+Independent JPEG Group
+.SH BUGS
+Arithmetic coding is not supported for legal reasons.
+.PP
+The transform options can't transform odd-size images perfectly.  Use
+.B \-trim
+or
+.B \-perfect
+if you don't like the results.
+.PP
+The entire image is read into memory and then written out again, even in
+cases where this isn't really necessary.  Expect swapping on large images,
+especially when using the more complex transform options.
diff --git a/jpegtran.c b/jpegtran.c
new file mode 100644
index 0000000..73dbe53
--- /dev/null
+++ b/jpegtran.c
@@ -0,0 +1,549 @@
+/*
+ * jpegtran.c
+ *
+ * Copyright (C) 1995-2010, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a command-line user interface for JPEG transcoding.
+ * It is very similar to cjpeg.c, and partly to djpeg.c, but provides
+ * lossless transcoding between different JPEG file formats.  It also
+ * provides some lossless and sort-of-lossless transformations of JPEG data.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+#include "transupp.h"		/* Support routines for jpegtran */
+#include "jversion.h"		/* for version message */
+#include "config.h"
+
+#ifdef USE_CCOMMAND		/* command-line reader for Macintosh */
+#ifdef __MWERKS__
+#include <SIOUX.h>              /* Metrowerks needs this */
+#include <console.h>		/* ... and this */
+#endif
+#ifdef THINK_C
+#include <console.h>		/* Think declares it here */
+#endif
+#endif
+
+
+/*
+ * Argument-parsing code.
+ * The switch parser is designed to be useful with DOS-style command line
+ * syntax, ie, intermixed switches and file names, where only the switches
+ * to the left of a given file name affect processing of that file.
+ * The main program in this file doesn't actually use this capability...
+ */
+
+
+static const char * progname;	/* program name for error messages */
+static char * outfilename;	/* for -outfile switch */
+static JCOPY_OPTION copyoption;	/* -copy switch */
+static jpeg_transform_info transformoption; /* image transformation options */
+
+
+LOCAL(void)
+usage (void)
+/* complain about bad command line */
+{
+  fprintf(stderr, "usage: %s [switches] ", progname);
+#ifdef TWO_FILE_COMMANDLINE
+  fprintf(stderr, "inputfile outputfile\n");
+#else
+  fprintf(stderr, "[inputfile]\n");
+#endif
+
+  fprintf(stderr, "Switches (names may be abbreviated):\n");
+  fprintf(stderr, "  -copy none     Copy no extra markers from source file\n");
+  fprintf(stderr, "  -copy comments Copy only comment markers (default)\n");
+  fprintf(stderr, "  -copy all      Copy all extra markers\n");
+#ifdef ENTROPY_OPT_SUPPORTED
+  fprintf(stderr, "  -optimize      Optimize Huffman table (smaller file, but slow compression)\n");
+#endif
+#ifdef C_PROGRESSIVE_SUPPORTED
+  fprintf(stderr, "  -progressive   Create progressive JPEG file\n");
+#endif
+  fprintf(stderr, "Switches for modifying the image:\n");
+#if TRANSFORMS_SUPPORTED
+  fprintf(stderr, "  -crop WxH+X+Y  Crop to a rectangular subarea\n");
+  fprintf(stderr, "  -grayscale     Reduce to grayscale (omit color data)\n");
+  fprintf(stderr, "  -flip [horizontal|vertical]  Mirror image (left-right or top-bottom)\n");
+  fprintf(stderr, "  -perfect       Fail if there is non-transformable edge blocks\n");
+  fprintf(stderr, "  -rotate [90|180|270]         Rotate image (degrees clockwise)\n");
+#endif
+#if TRANSFORMS_SUPPORTED
+  fprintf(stderr, "  -transpose     Transpose image\n");
+  fprintf(stderr, "  -transverse    Transverse transpose image\n");
+  fprintf(stderr, "  -trim          Drop non-transformable edge blocks\n");
+#endif
+  fprintf(stderr, "Switches for advanced users:\n");
+  fprintf(stderr, "  -restart N     Set restart interval in rows, or in blocks with B\n");
+  fprintf(stderr, "  -maxmemory N   Maximum memory to use (in kbytes)\n");
+  fprintf(stderr, "  -outfile name  Specify name for output file\n");
+  fprintf(stderr, "  -verbose  or  -debug   Emit debug output\n");
+  fprintf(stderr, "Switches for wizards:\n");
+#ifdef C_ARITH_CODING_SUPPORTED
+  fprintf(stderr, "  -arithmetic    Use arithmetic coding\n");
+#endif
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+  fprintf(stderr, "  -scans file    Create multi-scan JPEG per script file\n");
+#endif
+  exit(EXIT_FAILURE);
+}
+
+
+LOCAL(void)
+select_transform (JXFORM_CODE transform)
+/* Silly little routine to detect multiple transform options,
+ * which we can't handle.
+ */
+{
+#if TRANSFORMS_SUPPORTED
+  if (transformoption.transform == JXFORM_NONE ||
+      transformoption.transform == transform) {
+    transformoption.transform = transform;
+  } else {
+    fprintf(stderr, "%s: can only do one image transformation at a time\n",
+	    progname);
+    usage();
+  }
+#else
+  fprintf(stderr, "%s: sorry, image transformation was not compiled\n",
+	  progname);
+  exit(EXIT_FAILURE);
+#endif
+}
+
+
+LOCAL(int)
+parse_switches (j_compress_ptr cinfo, int argc, char **argv,
+		int last_file_arg_seen, boolean for_real)
+/* Parse optional switches.
+ * Returns argv[] index of first file-name argument (== argc if none).
+ * Any file names with indexes <= last_file_arg_seen are ignored;
+ * they have presumably been processed in a previous iteration.
+ * (Pass 0 for last_file_arg_seen on the first or only iteration.)
+ * for_real is FALSE on the first (dummy) pass; we may skip any expensive
+ * processing.
+ */
+{
+  int argn;
+  char * arg;
+  boolean simple_progressive;
+  char * scansarg = NULL;	/* saves -scans parm if any */
+
+  /* Set up default JPEG parameters. */
+  simple_progressive = FALSE;
+  outfilename = NULL;
+  copyoption = JCOPYOPT_DEFAULT;
+  transformoption.transform = JXFORM_NONE;
+  transformoption.perfect = FALSE;
+  transformoption.trim = FALSE;
+  transformoption.force_grayscale = FALSE;
+  transformoption.crop = FALSE;
+  cinfo->err->trace_level = 0;
+
+  /* Scan command line options, adjust parameters */
+
+  for (argn = 1; argn < argc; argn++) {
+    arg = argv[argn];
+    if (*arg != '-') {
+      /* Not a switch, must be a file name argument */
+      if (argn <= last_file_arg_seen) {
+	outfilename = NULL;	/* -outfile applies to just one input file */
+	continue;		/* ignore this name if previously processed */
+      }
+      break;			/* else done parsing switches */
+    }
+    arg++;			/* advance past switch marker character */
+
+    if (keymatch(arg, "arithmetic", 1)) {
+      /* Use arithmetic coding. */
+#ifdef C_ARITH_CODING_SUPPORTED
+      cinfo->arith_code = TRUE;
+#else
+      fprintf(stderr, "%s: sorry, arithmetic coding not supported\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "copy", 2)) {
+      /* Select which extra markers to copy. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (keymatch(argv[argn], "none", 1)) {
+	copyoption = JCOPYOPT_NONE;
+      } else if (keymatch(argv[argn], "comments", 1)) {
+	copyoption = JCOPYOPT_COMMENTS;
+      } else if (keymatch(argv[argn], "all", 1)) {
+	copyoption = JCOPYOPT_ALL;
+      } else
+	usage();
+
+    } else if (keymatch(arg, "crop", 2)) {
+      /* Perform lossless cropping. */
+#if TRANSFORMS_SUPPORTED
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (! jtransform_parse_crop_spec(&transformoption, argv[argn])) {
+	fprintf(stderr, "%s: bogus -crop argument '%s'\n",
+		progname, argv[argn]);
+	exit(EXIT_FAILURE);
+      }
+#else
+      select_transform(JXFORM_NONE);	/* force an error */
+#endif
+
+    } else if (keymatch(arg, "debug", 1) || keymatch(arg, "verbose", 1)) {
+      /* Enable debug printouts. */
+      /* On first -d, print version identification */
+      static boolean printed_version = FALSE;
+
+      if (! printed_version) {
+	fprintf(stderr, "%s version %s (build %s)\n",
+		PACKAGE_NAME, VERSION, BUILD);
+	fprintf(stderr, "%s\n\n", LJTCOPYRIGHT);
+	fprintf(stderr, "Based on Independent JPEG Group's libjpeg, version %s\n%s\n\n",
+		JVERSION, JCOPYRIGHT);
+	printed_version = TRUE;
+      }
+      cinfo->err->trace_level++;
+
+    } else if (keymatch(arg, "flip", 1)) {
+      /* Mirror left-right or top-bottom. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (keymatch(argv[argn], "horizontal", 1))
+	select_transform(JXFORM_FLIP_H);
+      else if (keymatch(argv[argn], "vertical", 1))
+	select_transform(JXFORM_FLIP_V);
+      else
+	usage();
+
+    } else if (keymatch(arg, "grayscale", 1) || keymatch(arg, "greyscale",1)) {
+      /* Force to grayscale. */
+#if TRANSFORMS_SUPPORTED
+      transformoption.force_grayscale = TRUE;
+#else
+      select_transform(JXFORM_NONE);	/* force an error */
+#endif
+
+    } else if (keymatch(arg, "maxmemory", 3)) {
+      /* Maximum memory in Kb (or Mb with 'm'). */
+      long lval;
+      char ch = 'x';
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%ld%c", &lval, &ch) < 1)
+	usage();
+      if (ch == 'm' || ch == 'M')
+	lval *= 1000L;
+      cinfo->mem->max_memory_to_use = lval * 1000L;
+
+    } else if (keymatch(arg, "optimize", 1) || keymatch(arg, "optimise", 1)) {
+      /* Enable entropy parm optimization. */
+#ifdef ENTROPY_OPT_SUPPORTED
+      cinfo->optimize_coding = TRUE;
+#else
+      fprintf(stderr, "%s: sorry, entropy optimization was not compiled\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "outfile", 4)) {
+      /* Set output file name. */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      outfilename = argv[argn];	/* save it away for later use */
+
+    } else if (keymatch(arg, "perfect", 2)) {
+      /* Fail if there is any partial edge MCUs that the transform can't
+       * handle. */
+      transformoption.perfect = TRUE;
+
+    } else if (keymatch(arg, "progressive", 2)) {
+      /* Select simple progressive mode. */
+#ifdef C_PROGRESSIVE_SUPPORTED
+      simple_progressive = TRUE;
+      /* We must postpone execution until num_components is known. */
+#else
+      fprintf(stderr, "%s: sorry, progressive output was not compiled\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "restart", 1)) {
+      /* Restart interval in MCU rows (or in MCUs with 'b'). */
+      long lval;
+      char ch = 'x';
+
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (sscanf(argv[argn], "%ld%c", &lval, &ch) < 1)
+	usage();
+      if (lval < 0 || lval > 65535L)
+	usage();
+      if (ch == 'b' || ch == 'B') {
+	cinfo->restart_interval = (unsigned int) lval;
+	cinfo->restart_in_rows = 0; /* else prior '-restart n' overrides me */
+      } else {
+	cinfo->restart_in_rows = (int) lval;
+	/* restart_interval will be computed during startup */
+      }
+
+    } else if (keymatch(arg, "rotate", 2)) {
+      /* Rotate 90, 180, or 270 degrees (measured clockwise). */
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      if (keymatch(argv[argn], "90", 2))
+	select_transform(JXFORM_ROT_90);
+      else if (keymatch(argv[argn], "180", 3))
+	select_transform(JXFORM_ROT_180);
+      else if (keymatch(argv[argn], "270", 3))
+	select_transform(JXFORM_ROT_270);
+      else
+	usage();
+
+    } else if (keymatch(arg, "scans", 1)) {
+      /* Set scan script. */
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+      if (++argn >= argc)	/* advance to next argument */
+	usage();
+      scansarg = argv[argn];
+      /* We must postpone reading the file in case -progressive appears. */
+#else
+      fprintf(stderr, "%s: sorry, multi-scan output was not compiled\n",
+	      progname);
+      exit(EXIT_FAILURE);
+#endif
+
+    } else if (keymatch(arg, "transpose", 1)) {
+      /* Transpose (across UL-to-LR axis). */
+      select_transform(JXFORM_TRANSPOSE);
+
+    } else if (keymatch(arg, "transverse", 6)) {
+      /* Transverse transpose (across UR-to-LL axis). */
+      select_transform(JXFORM_TRANSVERSE);
+
+    } else if (keymatch(arg, "trim", 3)) {
+      /* Trim off any partial edge MCUs that the transform can't handle. */
+      transformoption.trim = TRUE;
+
+    } else {
+      usage();			/* bogus switch */
+    }
+  }
+
+  /* Post-switch-scanning cleanup */
+
+  if (for_real) {
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+    if (simple_progressive)	/* process -progressive; -scans can override */
+      jpeg_simple_progression(cinfo);
+#endif
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+    if (scansarg != NULL)	/* process -scans if it was present */
+      if (! read_scan_script(cinfo, scansarg))
+	usage();
+#endif
+  }
+
+  return argn;			/* return index of next arg (file name) */
+}
+
+
+/*
+ * The main program.
+ */
+
+int
+main (int argc, char **argv)
+{
+  struct jpeg_decompress_struct srcinfo;
+  struct jpeg_compress_struct dstinfo;
+  struct jpeg_error_mgr jsrcerr, jdsterr;
+#ifdef PROGRESS_REPORT
+  struct cdjpeg_progress_mgr progress;
+#endif
+  jvirt_barray_ptr * src_coef_arrays;
+  jvirt_barray_ptr * dst_coef_arrays;
+  int file_index;
+  /* We assume all-in-memory processing and can therefore use only a
+   * single file pointer for sequential input and output operation. 
+   */
+  FILE * fp;
+
+  /* On Mac, fetch a command line. */
+#ifdef USE_CCOMMAND
+  argc = ccommand(&argv);
+#endif
+
+  progname = argv[0];
+  if (progname == NULL || progname[0] == 0)
+    progname = "jpegtran";	/* in case C library doesn't provide it */
+
+  /* Initialize the JPEG decompression object with default error handling. */
+  srcinfo.err = jpeg_std_error(&jsrcerr);
+  jpeg_create_decompress(&srcinfo);
+  /* Initialize the JPEG compression object with default error handling. */
+  dstinfo.err = jpeg_std_error(&jdsterr);
+  jpeg_create_compress(&dstinfo);
+
+  /* Now safe to enable signal catcher.
+   * Note: we assume only the decompression object will have virtual arrays.
+   */
+#ifdef NEED_SIGNAL_CATCHER
+  enable_signal_catcher((j_common_ptr) &srcinfo);
+#endif
+
+  /* Scan command line to find file names.
+   * It is convenient to use just one switch-parsing routine, but the switch
+   * values read here are mostly ignored; we will rescan the switches after
+   * opening the input file.  Also note that most of the switches affect the
+   * destination JPEG object, so we parse into that and then copy over what
+   * needs to affects the source too.
+   */
+
+  file_index = parse_switches(&dstinfo, argc, argv, 0, FALSE);
+  jsrcerr.trace_level = jdsterr.trace_level;
+  srcinfo.mem->max_memory_to_use = dstinfo.mem->max_memory_to_use;
+
+#ifdef TWO_FILE_COMMANDLINE
+  /* Must have either -outfile switch or explicit output file name */
+  if (outfilename == NULL) {
+    if (file_index != argc-2) {
+      fprintf(stderr, "%s: must name one input and one output file\n",
+	      progname);
+      usage();
+    }
+    outfilename = argv[file_index+1];
+  } else {
+    if (file_index != argc-1) {
+      fprintf(stderr, "%s: must name one input and one output file\n",
+	      progname);
+      usage();
+    }
+  }
+#else
+  /* Unix style: expect zero or one file name */
+  if (file_index < argc-1) {
+    fprintf(stderr, "%s: only one input file\n", progname);
+    usage();
+  }
+#endif /* TWO_FILE_COMMANDLINE */
+
+  /* Open the input file. */
+  if (file_index < argc) {
+    if ((fp = fopen(argv[file_index], READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s for reading\n", progname, argv[file_index]);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default input file is stdin */
+    fp = read_stdin();
+  }
+
+#ifdef PROGRESS_REPORT
+  start_progress_monitor((j_common_ptr) &dstinfo, &progress);
+#endif
+
+  /* Specify data source for decompression */
+  jpeg_stdio_src(&srcinfo, fp);
+
+  /* Enable saving of extra markers that we want to copy */
+  jcopy_markers_setup(&srcinfo, copyoption);
+
+  /* Read file header */
+  (void) jpeg_read_header(&srcinfo, TRUE);
+
+  /* Any space needed by a transform option must be requested before
+   * jpeg_read_coefficients so that memory allocation will be done right.
+   */
+#if TRANSFORMS_SUPPORTED
+  /* Fail right away if -perfect is given and transformation is not perfect.
+   */
+  if (!jtransform_request_workspace(&srcinfo, &transformoption)) {
+    fprintf(stderr, "%s: transformation is not perfect\n", progname);
+    exit(EXIT_FAILURE);
+  }
+#endif
+
+  /* Read source file as DCT coefficients */
+  src_coef_arrays = jpeg_read_coefficients(&srcinfo);
+
+  /* Initialize destination compression parameters from source values */
+  jpeg_copy_critical_parameters(&srcinfo, &dstinfo);
+
+  /* Adjust destination parameters if required by transform options;
+   * also find out which set of coefficient arrays will hold the output.
+   */
+#if TRANSFORMS_SUPPORTED
+  dst_coef_arrays = jtransform_adjust_parameters(&srcinfo, &dstinfo,
+						 src_coef_arrays,
+						 &transformoption);
+#else
+  dst_coef_arrays = src_coef_arrays;
+#endif
+
+  /* Close input file, if we opened it.
+   * Note: we assume that jpeg_read_coefficients consumed all input
+   * until JPEG_REACHED_EOI, and that jpeg_finish_decompress will
+   * only consume more while (! cinfo->inputctl->eoi_reached).
+   * We cannot call jpeg_finish_decompress here since we still need the
+   * virtual arrays allocated from the source object for processing.
+   */
+  if (fp != stdin)
+    fclose(fp);
+
+  /* Open the output file. */
+  if (outfilename != NULL) {
+    if ((fp = fopen(outfilename, WRITE_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s for writing\n", progname, outfilename);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default output file is stdout */
+    fp = write_stdout();
+  }
+
+  /* Adjust default compression parameters by re-parsing the options */
+  file_index = parse_switches(&dstinfo, argc, argv, 0, TRUE);
+
+  /* Specify data destination for compression */
+  jpeg_stdio_dest(&dstinfo, fp);
+
+  /* Start compressor (note no image data is actually written here) */
+  jpeg_write_coefficients(&dstinfo, dst_coef_arrays);
+
+  /* Copy to the output file any extra markers that we want to preserve */
+  jcopy_markers_execute(&srcinfo, &dstinfo, copyoption);
+
+  /* Execute image transformation, if any */
+#if TRANSFORMS_SUPPORTED
+  jtransform_execute_transformation(&srcinfo, &dstinfo,
+				    src_coef_arrays,
+				    &transformoption);
+#endif
+
+  /* Finish compression and release memory */
+  jpeg_finish_compress(&dstinfo);
+  jpeg_destroy_compress(&dstinfo);
+  (void) jpeg_finish_decompress(&srcinfo);
+  jpeg_destroy_decompress(&srcinfo);
+
+  /* Close output file, if we opened it */
+  if (fp != stdout)
+    fclose(fp);
+
+#ifdef PROGRESS_REPORT
+  end_progress_monitor((j_common_ptr) &dstinfo);
+#endif
+
+  /* All done. */
+  exit(jsrcerr.num_warnings + jdsterr.num_warnings ?EXIT_WARNING:EXIT_SUCCESS);
+  return 0;			/* suppress no-return-value warnings */
+}
diff --git a/jpegut.c b/jpegut.c
new file mode 100644
index 0000000..db46b6d
--- /dev/null
+++ b/jpegut.c
@@ -0,0 +1,415 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005 Sun Microsystems, Inc.
+ * Copyright (C)2009 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./rrtimer.h"
+#include "./turbojpeg.h"
+
+#define _catch(f) {if((f)==-1) {printf("TJPEG: %s\n", tjGetErrorStr());  bailout();}}
+
+const char *_subnamel[NUMSUBOPT]={"4:4:4", "4:2:2", "4:2:0", "GRAY"};
+const char *_subnames[NUMSUBOPT]={"444", "422", "420", "GRAY"};
+
+int exitstatus=0;
+#define bailout() {exitstatus=-1;  goto finally;}
+
+int pixels[9][3]=
+{
+	{0, 255, 0},
+	{255, 0, 255},
+	{255, 255, 0},
+	{0, 0, 255},
+	{0, 255, 255},
+	{255, 0, 0},
+	{255, 255, 255},
+	{0, 0, 0},
+	{255, 0, 0}
+};
+
+void initbuf(unsigned char *buf, int w, int h, int ps, int flags)
+{
+	int roffset=(flags&TJ_BGR)?2:0, goffset=1, boffset=(flags&TJ_BGR)?0:2, i,
+		_i, j;
+	if(flags&TJ_ALPHAFIRST) {roffset++;  goffset++;  boffset++;}
+	memset(buf, 0, w*h*ps);
+	if(ps==1)
+	{
+		for(_i=0; _i<16; _i++)
+		{
+			if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+			for(j=0; j<w; j++)
+			{
+				if(((_i/8)+(j/8))%2==0) buf[w*i+j]=255;
+				else buf[w*i+j]=76;
+			}
+		}
+		for(_i=16; _i<h; _i++)
+		{
+			if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+			for(j=0; j<w; j++)
+			{
+				if(((_i/8)+(j/8))%2==0) buf[w*i+j]=0;
+				else buf[w*i+j]=226;
+			}
+		}
+		return;
+	}
+	for(_i=0; _i<16; _i++)
+	{
+		if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+		for(j=0; j<w; j++)
+		{
+			buf[(w*i+j)*ps+roffset]=255;
+			if(((_i/8)+(j/8))%2==0)
+			{
+				buf[(w*i+j)*ps+goffset]=255;
+				buf[(w*i+j)*ps+boffset]=255;
+			}
+		}
+	}
+	for(_i=16; _i<h; _i++)
+	{
+		if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+		for(j=0; j<w; j++)
+		{
+			if(((_i/8)+(j/8))%2!=0)
+			{
+				buf[(w*i+j)*ps+roffset]=255;
+				buf[(w*i+j)*ps+goffset]=255;
+			}
+		}
+	}
+}
+
+void dumpbuf(unsigned char *buf, int w, int h, int ps, int flags)
+{
+	int roffset=(flags&TJ_BGR)?2:0, goffset=1, boffset=(flags&TJ_BGR)?0:2, i,
+		j;
+	for(i=0; i<h; i++)
+	{
+		for(j=0; j<w; j++)
+		{
+			printf("%.3d/%.3d/%.3d ", buf[(w*i+j)*ps+roffset],
+				buf[(w*i+j)*ps+roffset], buf[(w*i+j)*ps+roffset]);
+		}
+		printf("\n");
+	}
+}
+
+int checkbuf(unsigned char *buf, int w, int h, int ps, int subsamp, int flags)
+{
+	int roffset=(flags&TJ_BGR)?2:0, goffset=1, boffset=(flags&TJ_BGR)?0:2, i,
+		_i, j;
+	if(flags&TJ_ALPHAFIRST) {roffset++;  goffset++;  boffset++;}
+	if(ps==1) roffset=goffset=boffset=0;
+	if(subsamp==TJ_GRAYSCALE)
+	{
+		for(_i=0; _i<16; _i++)
+		{
+			if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+			for(j=0; j<w; j++)
+			{
+				unsigned char r=buf[(w*i+j)*ps+roffset],
+					g=buf[(w*i+j)*ps+goffset],
+					b=buf[(w*i+j)*ps+boffset];
+				if(((_i/8)+(j/8))%2==0)
+				{
+					if(r<253 || g<253 || b<253) return 0;
+				}
+				else
+				{
+					if(r<74 || r>78 || g<74 || g>78 || b<74 || b>78) return 0;
+				}
+			}
+		}
+		for(_i=16; _i<h; _i++)
+		{
+			if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+			for(j=0; j<w; j++)
+			{
+				unsigned char r=buf[(w*i+j)*ps+roffset],
+					g=buf[(w*i+j)*ps+goffset],
+					b=buf[(w*i+j)*ps+boffset];
+				if(((_i/8)+(j/8))%2==0)
+				{
+					if(r>2 || g>2 || b>2) return 0;
+				}
+				else
+				{
+					if(r<224 || r>228 || g<224 || g>228 || b<224 || b>228) return 0;
+				}
+			}
+		}
+	}
+	else
+	{
+		for(_i=0; _i<16; _i++)
+		{
+			if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+			for(j=0; j<w; j++)
+			{
+				if(buf[(w*i+j)*ps+roffset]<253) return 0;
+				if(((_i/8)+(j/8))%2==0)
+				{
+					if(buf[(w*i+j)*ps+goffset]<253) return 0;
+					if(buf[(w*i+j)*ps+boffset]<253) return 0;
+				}
+				else
+				{
+					if(buf[(w*i+j)*ps+goffset]>2) return 0;
+					if(buf[(w*i+j)*ps+boffset]>2) return 0;
+				}
+			}
+		}
+		for(_i=16; _i<h; _i++)
+		{
+			if(flags&TJ_BOTTOMUP) i=h-_i-1;  else i=_i;
+			for(j=0; j<w; j++)
+			{
+				if(buf[(w*i+j)*ps+boffset]>2) return 0;
+				if(((_i/8)+(j/8))%2==0)
+				{
+					if(buf[(w*i+j)*ps+roffset]>2) return 0;
+					if(buf[(w*i+j)*ps+goffset]>2) return 0;
+				}
+				else
+				{
+					if(buf[(w*i+j)*ps+roffset]<253) return 0;
+					if(buf[(w*i+j)*ps+goffset]<253) return 0;
+				}
+			}
+		}
+	}
+	return 1;
+}
+
+void writejpeg(unsigned char *jpegbuf, unsigned long jpgbufsize, char *filename)
+{
+	FILE *outfile=NULL;
+	if((outfile=fopen(filename, "wb"))==NULL)
+	{
+		printf("ERROR: Could not open %s for writing.\n", filename);
+		bailout();
+	}
+	if(fwrite(jpegbuf, jpgbufsize, 1, outfile)!=1)
+	{
+		printf("ERROR: Could not write to %s.\n", filename);
+		bailout();
+	}
+
+	finally:
+	if(outfile) fclose(outfile);
+}
+
+void gentestjpeg(tjhandle hnd, unsigned char *jpegbuf, unsigned long *size,
+	int w, int h, int ps, char *basefilename, int subsamp, int qual, int flags)
+{
+	char tempstr[1024];  unsigned char *bmpbuf=NULL;
+	const char *pixformat;  double t;
+
+	if(flags&TJ_BGR)
+	{
+		if(ps==3) pixformat="BGR";
+		else {if(flags&TJ_ALPHAFIRST) pixformat="ABGR";  else pixformat="BGRA";}
+	}
+	else
+	{
+		if(ps==3) pixformat="RGB";
+		else {if(flags&TJ_ALPHAFIRST) pixformat="ARGB";  else pixformat="RGBA";}
+	}
+	if(ps==1) pixformat="Grayscale";
+	printf("%s %s -> %s Q%d ... ", pixformat,
+		(flags&TJ_BOTTOMUP)?"Bottom-Up":"Top-Down ", _subnamel[subsamp], qual);
+
+	if((bmpbuf=(unsigned char *)malloc(w*h*ps+1))==NULL)
+	{
+		printf("ERROR: Could not allocate buffer\n");  bailout();
+	}
+	initbuf(bmpbuf, w, h, ps, flags);
+	memset(jpegbuf, 0, TJBUFSIZE(w, h));
+
+	t=rrtime();
+	_catch(tjCompress(hnd, bmpbuf, w, 0, h, ps, jpegbuf, size, subsamp, qual, flags));
+	t=rrtime()-t;
+
+	sprintf(tempstr, "%s_enc_%s_%s_%sQ%d.jpg", basefilename, pixformat,
+		(flags&TJ_BOTTOMUP)? "BU":"TD", _subnames[subsamp], qual);
+	writejpeg(jpegbuf, *size, tempstr);
+	printf("Done.  %f ms\n  Result in %s\n", t*1000., tempstr);
+
+	finally:
+	if(bmpbuf) free(bmpbuf);
+}
+
+void gentestbmp(tjhandle hnd, unsigned char *jpegbuf, unsigned long jpegsize,
+	int w, int h, int ps, char *basefilename, int subsamp, int qual, int flags)
+{
+	unsigned char *bmpbuf=NULL;
+	const char *pixformat;  int _w=0, _h=0;  double t;
+
+	if(flags&TJ_BGR)
+	{
+		if(ps==3) pixformat="BGR";
+		else {if(flags&TJ_ALPHAFIRST) pixformat="ABGR";  else pixformat="BGRA";}
+	}
+	else
+	{
+		if(ps==3) pixformat="RGB";
+		else {if(flags&TJ_ALPHAFIRST) pixformat="ARGB";  else pixformat="RGBA";}
+	}
+	if(ps==1) pixformat="Grayscale";
+	printf("JPEG -> %s %s ... ", pixformat, (flags&TJ_BOTTOMUP)?"Bottom-Up":"Top-Down ");
+
+	_catch(tjDecompressHeader(hnd, jpegbuf, jpegsize, &_w, &_h));
+	if(_w!=w || _h!=h)
+	{
+		printf("Incorrect JPEG header\n");  bailout();
+	}
+
+	if((bmpbuf=(unsigned char *)malloc(w*h*ps+1))==NULL)
+	{
+		printf("ERROR: Could not allocate buffer\n");  bailout();
+	}
+	memset(bmpbuf, 0, w*ps*h);
+
+	t=rrtime();
+	_catch(tjDecompress(hnd, jpegbuf, jpegsize, bmpbuf, w, w*ps, h, ps, flags));
+	t=rrtime()-t;
+
+	if(checkbuf(bmpbuf, w, h, ps, subsamp, flags)) printf("Passed.");
+	else {printf("FAILED!");  dumpbuf(bmpbuf, w, h, ps, flags);}
+
+	printf("  %f ms\n\n", t*1000.);
+
+	finally:
+	if(bmpbuf) free(bmpbuf);
+}
+
+void dotest(int w, int h, int ps, int subsamp, char *basefilename)
+{
+	tjhandle hnd=NULL, dhnd=NULL;  unsigned char *jpegbuf=NULL;
+	unsigned long size;
+
+	if((jpegbuf=(unsigned char *)malloc(TJBUFSIZE(w, h))) == NULL)
+	{
+		puts("ERROR: Could not allocate buffer.");  bailout();
+	}
+
+	if((hnd=tjInitCompress())==NULL)
+		{printf("Error in tjInitCompress():\n%s\n", tjGetErrorStr());  bailout();}
+	if((dhnd=tjInitDecompress())==NULL)
+		{printf("Error in tjInitDecompress():\n%s\n", tjGetErrorStr());  bailout();}
+
+	gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, 0);
+	gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, 0);
+
+	if(ps==1) goto finally;
+
+	gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_BGR);
+	gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_BGR);
+
+	gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_BOTTOMUP);
+	gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_BOTTOMUP);
+
+	gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_BGR|TJ_BOTTOMUP);
+	gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_BGR|TJ_BOTTOMUP);
+
+	if(ps==4)
+	{
+		gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST);
+		gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST);
+
+		gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST|TJ_BGR);
+		gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST|TJ_BGR);
+
+		gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST|TJ_BOTTOMUP);
+		gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST|TJ_BOTTOMUP);
+
+		gentestjpeg(hnd, jpegbuf, &size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST|TJ_BGR|TJ_BOTTOMUP);
+		gentestbmp(dhnd, jpegbuf, size, w, h, ps, basefilename, subsamp, 100, TJ_ALPHAFIRST|TJ_BGR|TJ_BOTTOMUP);
+	}
+
+	finally:
+	if(hnd) tjDestroy(hnd);
+	if(dhnd) tjDestroy(dhnd);
+
+	if(jpegbuf) free(jpegbuf);
+}
+
+#define MAXLENGTH 2048
+
+void dotest1(void)
+{
+	int i, j, i2;  unsigned char *bmpbuf=NULL, *jpgbuf=NULL;
+	tjhandle hnd=NULL;  unsigned long size;
+	if((hnd=tjInitCompress())==NULL)
+		{printf("Error in tjInitCompress():\n%s\n", tjGetErrorStr());  bailout();}
+	printf("Buffer size regression test\n");
+	for(j=1; j<48; j++)
+	{
+		for(i=1; i<(j==1?MAXLENGTH:48); i++)
+		{
+			if(i%100==0) printf("%.4d x %.4d\b\b\b\b\b\b\b\b\b\b\b", i, j);
+			if((bmpbuf=(unsigned char *)malloc(i*j*4))==NULL
+			|| (jpgbuf=(unsigned char *)malloc(TJBUFSIZE(i, j)))==NULL)
+			{
+				printf("Memory allocation failure\n");  bailout();
+			}
+			memset(bmpbuf, 0, i*j*4);
+			for(i2=0; i2<i*j; i2++)
+			{
+				bmpbuf[i2*4]=pixels[i2%9][2];
+				bmpbuf[i2*4+1]=pixels[i2%9][1];
+				bmpbuf[i2*2+2]=pixels[i2%9][0];
+			}
+			_catch(tjCompress(hnd, bmpbuf, i, i*4, j, 4,
+				jpgbuf, &size, TJ_444, 100, TJ_BGR));
+			free(bmpbuf);  bmpbuf=NULL;  free(jpgbuf);  jpgbuf=NULL;
+
+			if((bmpbuf=(unsigned char *)malloc(j*i*4))==NULL
+			|| (jpgbuf=(unsigned char *)malloc(TJBUFSIZE(j, i)))==NULL)
+			{
+				printf("Memory allocation failure\n");  bailout();
+			}
+			for(i2=0; i2<j*i*4; i2++)
+			{
+				if(i2%2==0) bmpbuf[i2]=0xFF;
+				else bmpbuf[i2]=0;
+			}
+			_catch(tjCompress(hnd, bmpbuf, j, j*4, i, 4,
+				jpgbuf, &size, TJ_444, 100, TJ_BGR));
+			free(bmpbuf);  bmpbuf=NULL;  free(jpgbuf);  jpgbuf=NULL;
+		}
+	}
+	printf("Done.      \n");
+
+	finally:
+	if(bmpbuf) free(bmpbuf);  if(jpgbuf) free(jpgbuf);
+	if(hnd) tjDestroy(hnd);
+}
+
+int main(int argc, char *argv[])
+{
+	dotest(35, 41, 3, TJ_444, "test");
+	dotest(35, 41, 4, TJ_444, "test");
+	dotest(35, 41, 1, TJ_GRAYSCALE, "test");
+	dotest(35, 41, 3, TJ_GRAYSCALE, "test");
+	dotest(35, 41, 4, TJ_GRAYSCALE, "test");
+	dotest1();
+
+	return exitstatus;
+}
diff --git a/jpgtest.cxx b/jpgtest.cxx
new file mode 100644
index 0000000..b1c5e1a
--- /dev/null
+++ b/jpgtest.cxx
@@ -0,0 +1,392 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005, 2006 Sun Microsystems, Inc.
+ * Copyright (C)2009 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "./bmp.h"
+#include "./rrutil.h"
+#include "./rrtimer.h"
+#include "./turbojpeg.h"
+
+#define _catch(f) {if((f)==-1) {printf("Error in %s:\n%s\n", #f, tjGetErrorStr());  goto bailout;}}
+
+int forcemmx=0, forcesse=0, forcesse2=0, forcesse3=0, fastupsample=0;
+const int _ps[BMPPIXELFORMATS]={3, 4, 3, 4, 4, 4};
+const int _flags[BMPPIXELFORMATS]={0, 0, TJ_BGR, TJ_BGR,
+	TJ_BGR|TJ_ALPHAFIRST, TJ_ALPHAFIRST};
+const int _rindex[BMPPIXELFORMATS]={0, 0, 2, 2, 3, 1};
+const int _gindex[BMPPIXELFORMATS]={1, 1, 1, 1, 2, 2};
+const int _bindex[BMPPIXELFORMATS]={2, 2, 0, 0, 1, 3};
+const char *_pfname[]={"RGB", "RGBA", "BGR", "BGRA", "ABGR", "ARGB"};
+const char *_subnamel[NUMSUBOPT]={"4:4:4", "4:2:2", "4:2:0", "GRAY"};
+const char *_subnames[NUMSUBOPT]={"444", "422", "420", "GRAY"};
+
+void printsigfig(double val, int figs)
+{
+	char format[80];
+	double _l=log10(val);  int l;
+	if(_l<0.)
+	{
+		l=(int)fabs(_l);
+		sprintf(format, "%%%d.%df", figs+l+2, figs+l);
+	}
+	else
+	{
+		l=(int)_l+1;
+		if(figs<=l) sprintf(format, "%%.0f");
+		else sprintf(format, "%%%d.%df", figs+1, figs-l);
+	}	
+	printf(format, val);
+}
+
+void dotest(unsigned char *srcbuf, int w, int h, BMPPIXELFORMAT pf, int bu,
+	int jpegsub, int qual, char *filename, int dotile, int useppm, int quiet)
+{
+	char tempstr[1024];
+	FILE *outfile;  tjhandle hnd;
+	unsigned char **jpegbuf=NULL, *rgbbuf=NULL;
+	rrtimer timer; double elapsed;
+	int jpgbufsize=0, i, j, tilesizex, tilesizey, numtilesx, numtilesy, ITER;
+	unsigned long *comptilesize=NULL;
+	int flags=(forcemmx?TJ_FORCEMMX:0)|(forcesse?TJ_FORCESSE:0)
+		|(forcesse2?TJ_FORCESSE2:0)|(forcesse3?TJ_FORCESSE3:0)
+		|(fastupsample?TJ_FASTUPSAMPLE:0);
+	int ps=_ps[pf];
+	int pitch=w*ps;
+
+	flags |= _flags[pf];
+	if(bu) flags |= TJ_BOTTOMUP;
+
+	if((rgbbuf=(unsigned char *)malloc(pitch*h)) == NULL)
+	{
+		puts("ERROR: Could not allocate image buffer.");
+		exit(1);
+	}
+
+	if(!quiet) printf("\n>>>>>  %s (%s) <--> JPEG %s Q%d  <<<<<\n", _pfname[pf],
+		bu?"Bottom-up":"Top-down", _subnamel[jpegsub], qual);
+	if(dotile) {tilesizex=tilesizey=4;}  else {tilesizex=w;  tilesizey=h;}
+
+	do
+	{
+		tilesizex*=2;  if(tilesizex>w) tilesizex=w;
+		tilesizey*=2;  if(tilesizey>h) tilesizey=h;
+		numtilesx=(w+tilesizex-1)/tilesizex;
+		numtilesy=(h+tilesizey-1)/tilesizey;
+		if((comptilesize=(unsigned long *)malloc(sizeof(unsigned long)*numtilesx*numtilesy)) == NULL
+		|| (jpegbuf=(unsigned char **)malloc(sizeof(unsigned char *)*numtilesx*numtilesy)) == NULL)
+		{
+			puts("ERROR: Could not allocate image buffers.");
+			goto bailout;
+		}
+		memset(jpegbuf, 0, sizeof(unsigned char *)*numtilesx*numtilesy);
+		for(i=0; i<numtilesx*numtilesy; i++)
+		{
+			if((jpegbuf[i]=(unsigned char *)malloc(TJBUFSIZE(tilesizex, tilesizey))) == NULL)
+			{
+				puts("ERROR: Could not allocate image buffers.");
+				goto bailout;
+			}
+		}
+
+		// Compression test
+		if(quiet) printf("%s\t%s\t%s\t%d\t",  _pfname[pf], bu?"BU":"TD",
+			_subnamel[jpegsub], qual);
+		for(i=0; i<h; i++) memcpy(&rgbbuf[pitch*i], &srcbuf[w*ps*i], w*ps);
+		if((hnd=tjInitCompress())==NULL)
+		{
+			printf("Error in tjInitCompress():\n%s\n", tjGetErrorStr());
+			goto bailout;
+		}
+		_catch(tjCompress(hnd, rgbbuf, tilesizex, pitch, tilesizey, ps,
+			jpegbuf[0], &comptilesize[0], jpegsub, qual, flags));
+		ITER=0;
+		timer.start();
+		do
+		{
+			jpgbufsize=0;  int tilen=0;
+			for(i=0; i<h; i+=tilesizey)
+			{
+				for(j=0; j<w; j+=tilesizex)
+				{
+					int tempw=min(tilesizex, w-j), temph=min(tilesizey, h-i);
+					_catch(tjCompress(hnd, &rgbbuf[pitch*i+j*ps], tempw, pitch,
+						temph, ps, jpegbuf[tilen], &comptilesize[tilen], jpegsub, qual,
+						flags));
+					jpgbufsize+=comptilesize[tilen];
+					tilen++;
+				}
+			}
+			ITER++;
+		} while((elapsed=timer.elapsed())<5.);
+		_catch(tjDestroy(hnd));
+		if(quiet)
+		{
+			if(tilesizex==w && tilesizey==h) printf("Full     \t");
+			else printf("%-4d %-4d\t", tilesizex, tilesizey);
+			printsigfig((double)(w*h)/1000000.*(double)ITER/elapsed, 4);
+			printf("\t");
+			printsigfig((double)(w*h*ps)/(double)jpgbufsize, 4);
+			printf("\t");
+		}
+		else
+		{
+			if(tilesizex==w && tilesizey==h) printf("\nFull image\n");
+			else printf("\nTile size: %d x %d\n", tilesizex, tilesizey);
+			printf("C--> Frame rate:           %f fps\n", (double)ITER/elapsed);
+			printf("     Output image size:    %d bytes\n", jpgbufsize);
+			printf("     Compression ratio:    %f:1\n",
+				(double)(w*h*ps)/(double)jpgbufsize);
+			printf("     Source throughput:    %f Megapixels/sec\n",
+				(double)(w*h)/1000000.*(double)ITER/elapsed);
+			printf("     Output bit stream:    %f Megabits/sec\n",
+				(double)jpgbufsize*8./1000000.*(double)ITER/elapsed);
+		}
+		if(tilesizex==w && tilesizey==h)
+		{
+			sprintf(tempstr, "%s_%sQ%d.jpg", filename, _subnames[jpegsub], qual);
+			if((outfile=fopen(tempstr, "wb"))==NULL)
+			{
+				puts("ERROR: Could not open reference image");
+				exit(1);
+			}
+			if(fwrite(jpegbuf[0], jpgbufsize, 1, outfile)!=1)
+			{
+				puts("ERROR: Could not write reference image");
+				exit(1);
+			}
+			fclose(outfile);
+			if(!quiet) printf("Reference image written to %s\n", tempstr);
+		}
+
+		// Decompression test
+		memset(rgbbuf, 127, pitch*h);  // Grey image means decompressor did nothing
+		if((hnd=tjInitDecompress())==NULL)
+		{
+			printf("Error in tjInitDecompress():\n%s\n", tjGetErrorStr());
+			goto bailout;
+		}
+		_catch(tjDecompress(hnd, jpegbuf[0], jpgbufsize, rgbbuf, tilesizex, pitch,
+			tilesizey, ps, flags));
+		ITER=0;
+		timer.start();
+		do
+		{
+			int tilen=0;
+			for(i=0; i<h; i+=tilesizey)
+			{
+				for(j=0; j<w; j+=tilesizex)
+				{
+					int tempw=min(tilesizex, w-j), temph=min(tilesizey, h-i);
+					_catch(tjDecompress(hnd, jpegbuf[tilen], comptilesize[tilen],
+						&rgbbuf[pitch*i+ps*j], tempw, pitch, temph, ps, flags));
+					tilen++;
+				}
+			}
+			ITER++;
+		}	while((elapsed=timer.elapsed())<5.);
+		_catch(tjDestroy(hnd));
+		if(quiet)
+		{
+			printsigfig((double)(w*h)/1000000.*(double)ITER/elapsed, 4);
+			printf("\n");
+		}
+		else
+		{
+			printf("D--> Frame rate:           %f fps\n", (double)ITER/elapsed);
+			printf("     Dest. throughput:     %f Megapixels/sec\n",
+				(double)(w*h)/1000000.*(double)ITER/elapsed);
+		}
+		if(tilesizex==w && tilesizey==h)
+			sprintf(tempstr, "%s_%sQ%d_full.%s", filename, _subnames[jpegsub], qual,
+				useppm?"ppm":"bmp");
+		else sprintf(tempstr, "%s_%sQ%d_%dx%d.%s", filename, _subnames[jpegsub],
+			qual, tilesizex, tilesizey, useppm?"ppm":"bmp");
+		if(savebmp(tempstr, rgbbuf, w, h, pf, pitch, bu)==-1)
+		{
+			printf("ERROR saving bitmap: %s\n", bmpgeterr());
+			goto bailout;
+		}
+		sprintf(strrchr(tempstr, '.'), "-err.%s", useppm?"ppm":"bmp");
+		if(!quiet)
+			printf("Computing compression error and saving to %s.\n", tempstr);
+		if(jpegsub==TJ_GRAYSCALE)
+		{
+			for(j=0; j<h; j++)
+			{
+				for(i=0; i<w*ps; i+=ps)
+				{
+					int y=(int)((double)srcbuf[w*ps*j+i+_rindex[pf]]*0.299
+						+ (double)srcbuf[w*ps*j+i+_gindex[pf]]*0.587
+						+ (double)srcbuf[w*ps*j+i+_bindex[pf]]*0.114 + 0.5);
+					if(y>255) y=255;  if(y<0) y=0;
+					rgbbuf[pitch*j+i+_rindex[pf]]=abs(rgbbuf[pitch*j+i+_rindex[pf]]-y);
+					rgbbuf[pitch*j+i+_gindex[pf]]=abs(rgbbuf[pitch*j+i+_gindex[pf]]-y);
+					rgbbuf[pitch*j+i+_bindex[pf]]=abs(rgbbuf[pitch*j+i+_bindex[pf]]-y);
+				}
+			}
+		}		
+		else
+		{
+			for(j=0; j<h; j++) for(i=0; i<w*ps; i++)
+				rgbbuf[pitch*j+i]=abs(rgbbuf[pitch*j+i]-srcbuf[w*ps*j+i]);
+		}
+		if(savebmp(tempstr, rgbbuf, w, h, pf, pitch, bu)==-1)
+		{
+			printf("ERROR saving bitmap: %s\n", bmpgeterr());
+			goto bailout;
+		}
+
+		// Cleanup
+		if(jpegbuf)
+		{
+			for(i=0; i<numtilesx*numtilesy; i++)
+				{if(jpegbuf[i]) free(jpegbuf[i]);  jpegbuf[i]=NULL;}
+			free(jpegbuf);  jpegbuf=NULL;
+		}
+		if(comptilesize) {free(comptilesize);  comptilesize=NULL;}
+	} while(tilesizex<w || tilesizey<h);
+
+	if(rgbbuf) {free(rgbbuf);  rgbbuf=NULL;}
+	return;
+
+	bailout:
+	if(jpegbuf)
+	{
+		for(i=0; i<numtilesx*numtilesy; i++)
+			{if(jpegbuf[i]) free(jpegbuf[i]);  jpegbuf[i]=NULL;}
+		free(jpegbuf);  jpegbuf=NULL;
+	}
+	if(comptilesize) {free(comptilesize);  comptilesize=NULL;}
+	if(rgbbuf) {free(rgbbuf);  rgbbuf=NULL;}
+	return;
+}
+
+
+int main(int argc, char *argv[])
+{
+	unsigned char *bmpbuf=NULL;  int w, h, i, useppm=0;
+	int qual, dotile=0, quiet=0, hiqual=-1;  char *temp;
+	BMPPIXELFORMAT pf=BMP_BGR;
+	int bu=0;
+
+	printf("\n");
+
+	if(argc<3)
+	{
+		printf("USAGE: %s <Inputfile (BMP|PPM)> <%% Quality>\n\n", argv[0]);
+		printf("       [-tile]\n");
+		printf("       Test performance of the codec when the image is encoded\n");
+		printf("       as separate tiles of varying sizes.\n\n");
+		printf("       [-forcemmx] [-forcesse] [-forcesse2] [-forcesse3]\n");
+		printf("       Force MMX, SSE, or SSE2 code paths in Intel codec\n\n");
+		printf("       [-rgb | -bgr | -rgba | -bgra | -abgr | -argb]\n");
+		printf("       Test the specified color conversion path in the codec (default: BGR)\n\n");
+		printf("       [-fastupsample]\n");
+		printf("       Use fast, inaccurate upsampling code to perform 4:2:2 and 4:2:0\n");
+		printf("       YUV decoding in libjpeg decompressor\n\n");
+		printf("       [-quiet]\n");
+		printf("       Output in tabular rather than verbose format\n\n");
+		printf("       NOTE: If the quality is specified as a range, i.e. 90-100, a separate\n");
+		printf("       test will be performed for all quality values in the range.\n");
+		exit(1);
+	}
+	if((qual=atoi(argv[2]))<1 || qual>100)
+	{
+		puts("ERROR: Quality must be between 1 and 100.");
+		exit(1);
+	}
+	if((temp=strchr(argv[2], '-'))!=NULL && strlen(temp)>1
+		&& sscanf(&temp[1], "%d", &hiqual)==1 && hiqual>qual && hiqual>=1
+		&& hiqual<=100) {}
+	else hiqual=qual;
+
+	if(argc>3)
+	{
+		for(i=3; i<argc; i++)
+		{
+			if(!stricmp(argv[i], "-tile")) dotile=1;
+			if(!stricmp(argv[i], "-forcesse3"))
+			{
+				printf("Using SSE3 code\n");
+				forcesse3=1;
+			}
+			if(!stricmp(argv[i], "-forcesse2"))
+			{
+				printf("Using SSE2 code\n");
+				forcesse2=1;
+			}
+			if(!stricmp(argv[i], "-forcesse"))
+			{
+				printf("Using SSE code\n");
+				forcesse=1;
+			}
+			if(!stricmp(argv[i], "-forcemmx"))
+			{
+				printf("Using MMX code\n");
+				forcemmx=1;
+			}
+			if(!stricmp(argv[i], "-fastupsample"))
+			{
+				printf("Using fast upsampling code\n");
+				fastupsample=1;
+			}
+			if(!stricmp(argv[i], "-rgb")) pf=BMP_RGB;
+			if(!stricmp(argv[i], "-rgba")) pf=BMP_RGBA;
+			if(!stricmp(argv[i], "-bgr")) pf=BMP_BGR;
+			if(!stricmp(argv[i], "-bgra")) pf=BMP_BGRA;
+			if(!stricmp(argv[i], "-abgr")) pf=BMP_ABGR;
+			if(!stricmp(argv[i], "-argb")) pf=BMP_ARGB;
+			if(!stricmp(argv[i], "-bottomup")) bu=1;
+			if(!stricmp(argv[i], "-quiet")) quiet=1;
+		}
+	}
+
+	if(loadbmp(argv[1], &bmpbuf, &w, &h, pf, 1, bu)==-1)
+	{
+		printf("ERROR loading bitmap: %s\n", bmpgeterr());  exit(1);
+	}
+
+	temp=strrchr(argv[1], '.');
+	if(temp!=NULL)
+	{
+		if(!stricmp(temp, ".ppm")) useppm=1;
+		*temp='\0';
+	}
+
+	if(quiet)
+	{
+		printf("All performance values in Mpixels/sec\n\n");
+		printf("Bitmap\tBitmap\tJPEG\tJPEG\tTile Size\tCompr\tCompr\tDecomp\n");
+		printf("Format\tOrder\tFormat\tQual\t X    Y  \tPerf \tRatio\tPerf\n\n");
+	}
+
+	for(i=hiqual; i>=qual; i--)
+		dotest(bmpbuf, w, h, pf, bu, TJ_GRAYSCALE, i, argv[1], dotile, useppm, quiet);
+	if(quiet) printf("\n");
+	for(i=hiqual; i>=qual; i--)
+		dotest(bmpbuf, w, h, pf, bu, TJ_420, i, argv[1], dotile, useppm, quiet);
+	if(quiet) printf("\n");
+	for(i=hiqual; i>=qual; i--)
+		dotest(bmpbuf, w, h, pf, bu, TJ_422, i, argv[1], dotile, useppm, quiet);
+	if(quiet) printf("\n");
+	for(i=hiqual; i>=qual; i--)
+		dotest(bmpbuf, w, h, pf, bu, TJ_444, i, argv[1], dotile, useppm, quiet);
+
+	if(bmpbuf) free(bmpbuf);
+	return 0;
+}
diff --git a/jquant1.c b/jquant1.c
new file mode 100644
index 0000000..362bb1e
--- /dev/null
+++ b/jquant1.c
@@ -0,0 +1,860 @@
+/*
+ * jquant1.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Copyright (C) 2009, D. R. Commander
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains 1-pass color quantization (color mapping) routines.
+ * These routines provide mapping to a fixed color map using equally spaced
+ * color values.  Optional Floyd-Steinberg or ordered dithering is available.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+#ifdef QUANT_1PASS_SUPPORTED
+
+
+/*
+ * The main purpose of 1-pass quantization is to provide a fast, if not very
+ * high quality, colormapped output capability.  A 2-pass quantizer usually
+ * gives better visual quality; however, for quantized grayscale output this
+ * quantizer is perfectly adequate.  Dithering is highly recommended with this
+ * quantizer, though you can turn it off if you really want to.
+ *
+ * In 1-pass quantization the colormap must be chosen in advance of seeing the
+ * image.  We use a map consisting of all combinations of Ncolors[i] color
+ * values for the i'th component.  The Ncolors[] values are chosen so that
+ * their product, the total number of colors, is no more than that requested.
+ * (In most cases, the product will be somewhat less.)
+ *
+ * Since the colormap is orthogonal, the representative value for each color
+ * component can be determined without considering the other components;
+ * then these indexes can be combined into a colormap index by a standard
+ * N-dimensional-array-subscript calculation.  Most of the arithmetic involved
+ * can be precalculated and stored in the lookup table colorindex[].
+ * colorindex[i][j] maps pixel value j in component i to the nearest
+ * representative value (grid plane) for that component; this index is
+ * multiplied by the array stride for component i, so that the
+ * index of the colormap entry closest to a given pixel value is just
+ *    sum( colorindex[component-number][pixel-component-value] )
+ * Aside from being fast, this scheme allows for variable spacing between
+ * representative values with no additional lookup cost.
+ *
+ * If gamma correction has been applied in color conversion, it might be wise
+ * to adjust the color grid spacing so that the representative colors are
+ * equidistant in linear space.  At this writing, gamma correction is not
+ * implemented by jdcolor, so nothing is done here.
+ */
+
+
+/* Declarations for ordered dithering.
+ *
+ * We use a standard 16x16 ordered dither array.  The basic concept of ordered
+ * dithering is described in many references, for instance Dale Schumacher's
+ * chapter II.2 of Graphics Gems II (James Arvo, ed. Academic Press, 1991).
+ * In place of Schumacher's comparisons against a "threshold" value, we add a
+ * "dither" value to the input pixel and then round the result to the nearest
+ * output value.  The dither value is equivalent to (0.5 - threshold) times
+ * the distance between output values.  For ordered dithering, we assume that
+ * the output colors are equally spaced; if not, results will probably be
+ * worse, since the dither may be too much or too little at a given point.
+ *
+ * The normal calculation would be to form pixel value + dither, range-limit
+ * this to 0..MAXJSAMPLE, and then index into the colorindex table as usual.
+ * We can skip the separate range-limiting step by extending the colorindex
+ * table in both directions.
+ */
+
+#define ODITHER_SIZE  16	/* dimension of dither matrix */
+/* NB: if ODITHER_SIZE is not a power of 2, ODITHER_MASK uses will break */
+#define ODITHER_CELLS (ODITHER_SIZE*ODITHER_SIZE)	/* # cells in matrix */
+#define ODITHER_MASK  (ODITHER_SIZE-1) /* mask for wrapping around counters */
+
+typedef int ODITHER_MATRIX[ODITHER_SIZE][ODITHER_SIZE];
+typedef int (*ODITHER_MATRIX_PTR)[ODITHER_SIZE];
+
+static const UINT8 base_dither_matrix[ODITHER_SIZE][ODITHER_SIZE] = {
+  /* Bayer's order-4 dither array.  Generated by the code given in
+   * Stephen Hawley's article "Ordered Dithering" in Graphics Gems I.
+   * The values in this array must range from 0 to ODITHER_CELLS-1.
+   */
+  {   0,192, 48,240, 12,204, 60,252,  3,195, 51,243, 15,207, 63,255 },
+  { 128, 64,176,112,140, 76,188,124,131, 67,179,115,143, 79,191,127 },
+  {  32,224, 16,208, 44,236, 28,220, 35,227, 19,211, 47,239, 31,223 },
+  { 160, 96,144, 80,172,108,156, 92,163, 99,147, 83,175,111,159, 95 },
+  {   8,200, 56,248,  4,196, 52,244, 11,203, 59,251,  7,199, 55,247 },
+  { 136, 72,184,120,132, 68,180,116,139, 75,187,123,135, 71,183,119 },
+  {  40,232, 24,216, 36,228, 20,212, 43,235, 27,219, 39,231, 23,215 },
+  { 168,104,152, 88,164,100,148, 84,171,107,155, 91,167,103,151, 87 },
+  {   2,194, 50,242, 14,206, 62,254,  1,193, 49,241, 13,205, 61,253 },
+  { 130, 66,178,114,142, 78,190,126,129, 65,177,113,141, 77,189,125 },
+  {  34,226, 18,210, 46,238, 30,222, 33,225, 17,209, 45,237, 29,221 },
+  { 162, 98,146, 82,174,110,158, 94,161, 97,145, 81,173,109,157, 93 },
+  {  10,202, 58,250,  6,198, 54,246,  9,201, 57,249,  5,197, 53,245 },
+  { 138, 74,186,122,134, 70,182,118,137, 73,185,121,133, 69,181,117 },
+  {  42,234, 26,218, 38,230, 22,214, 41,233, 25,217, 37,229, 21,213 },
+  { 170,106,154, 90,166,102,150, 86,169,105,153, 89,165,101,149, 85 }
+};
+
+
+/* Declarations for Floyd-Steinberg dithering.
+ *
+ * Errors are accumulated into the array fserrors[], at a resolution of
+ * 1/16th of a pixel count.  The error at a given pixel is propagated
+ * to its not-yet-processed neighbors using the standard F-S fractions,
+ *		...	(here)	7/16
+ *		3/16	5/16	1/16
+ * We work left-to-right on even rows, right-to-left on odd rows.
+ *
+ * We can get away with a single array (holding one row's worth of errors)
+ * by using it to store the current row's errors at pixel columns not yet
+ * processed, but the next row's errors at columns already processed.  We
+ * need only a few extra variables to hold the errors immediately around the
+ * current column.  (If we are lucky, those variables are in registers, but
+ * even if not, they're probably cheaper to access than array elements are.)
+ *
+ * The fserrors[] array is indexed [component#][position].
+ * We provide (#columns + 2) entries per component; the extra entry at each
+ * end saves us from special-casing the first and last pixels.
+ *
+ * Note: on a wide image, we might not have enough room in a PC's near data
+ * segment to hold the error array; so it is allocated with alloc_large.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+typedef INT16 FSERROR;		/* 16 bits should be enough */
+typedef int LOCFSERROR;		/* use 'int' for calculation temps */
+#else
+typedef INT32 FSERROR;		/* may need more than 16 bits */
+typedef INT32 LOCFSERROR;	/* be sure calculation temps are big enough */
+#endif
+
+typedef FSERROR FAR *FSERRPTR;	/* pointer to error array (in FAR storage!) */
+
+
+/* Private subobject */
+
+#define MAX_Q_COMPS 4		/* max components I can handle */
+
+typedef struct {
+  struct jpeg_color_quantizer pub; /* public fields */
+
+  /* Initially allocated colormap is saved here */
+  JSAMPARRAY sv_colormap;	/* The color map as a 2-D pixel array */
+  int sv_actual;		/* number of entries in use */
+
+  JSAMPARRAY colorindex;	/* Precomputed mapping for speed */
+  /* colorindex[i][j] = index of color closest to pixel value j in component i,
+   * premultiplied as described above.  Since colormap indexes must fit into
+   * JSAMPLEs, the entries of this array will too.
+   */
+  boolean is_padded;		/* is the colorindex padded for odither? */
+
+  int Ncolors[MAX_Q_COMPS];	/* # of values alloced to each component */
+
+  /* Variables for ordered dithering */
+  int row_index;		/* cur row's vertical index in dither matrix */
+  ODITHER_MATRIX_PTR odither[MAX_Q_COMPS]; /* one dither array per component */
+
+  /* Variables for Floyd-Steinberg dithering */
+  FSERRPTR fserrors[MAX_Q_COMPS]; /* accumulated errors */
+  boolean on_odd_row;		/* flag to remember which row we are on */
+} my_cquantizer;
+
+typedef my_cquantizer * my_cquantize_ptr;
+
+
+/*
+ * Policy-making subroutines for create_colormap and create_colorindex.
+ * These routines determine the colormap to be used.  The rest of the module
+ * only assumes that the colormap is orthogonal.
+ *
+ *  * select_ncolors decides how to divvy up the available colors
+ *    among the components.
+ *  * output_value defines the set of representative values for a component.
+ *  * largest_input_value defines the mapping from input values to
+ *    representative values for a component.
+ * Note that the latter two routines may impose different policies for
+ * different components, though this is not currently done.
+ */
+
+
+LOCAL(int)
+select_ncolors (j_decompress_ptr cinfo, int Ncolors[])
+/* Determine allocation of desired colors to components, */
+/* and fill in Ncolors[] array to indicate choice. */
+/* Return value is total number of colors (product of Ncolors[] values). */
+{
+  int nc = cinfo->out_color_components; /* number of color components */
+  int max_colors = cinfo->desired_number_of_colors;
+  int total_colors, iroot, i, j;
+  boolean changed;
+  long temp;
+  int RGB_order[3] = { RGB_GREEN, RGB_RED, RGB_BLUE };
+  RGB_order[0] = rgb_green[cinfo->out_color_space];
+  RGB_order[1] = rgb_red[cinfo->out_color_space];
+  RGB_order[2] = rgb_blue[cinfo->out_color_space];
+
+  /* We can allocate at least the nc'th root of max_colors per component. */
+  /* Compute floor(nc'th root of max_colors). */
+  iroot = 1;
+  do {
+    iroot++;
+    temp = iroot;		/* set temp = iroot ** nc */
+    for (i = 1; i < nc; i++)
+      temp *= iroot;
+  } while (temp <= (long) max_colors); /* repeat till iroot exceeds root */
+  iroot--;			/* now iroot = floor(root) */
+
+  /* Must have at least 2 color values per component */
+  if (iroot < 2)
+    ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, (int) temp);
+
+  /* Initialize to iroot color values for each component */
+  total_colors = 1;
+  for (i = 0; i < nc; i++) {
+    Ncolors[i] = iroot;
+    total_colors *= iroot;
+  }
+  /* We may be able to increment the count for one or more components without
+   * exceeding max_colors, though we know not all can be incremented.
+   * Sometimes, the first component can be incremented more than once!
+   * (Example: for 16 colors, we start at 2*2*2, go to 3*2*2, then 4*2*2.)
+   * In RGB colorspace, try to increment G first, then R, then B.
+   */
+  do {
+    changed = FALSE;
+    for (i = 0; i < nc; i++) {
+      j = (cinfo->out_color_space == JCS_RGB ? RGB_order[i] : i);
+      /* calculate new total_colors if Ncolors[j] is incremented */
+      temp = total_colors / Ncolors[j];
+      temp *= Ncolors[j]+1;	/* done in long arith to avoid oflo */
+      if (temp > (long) max_colors)
+	break;			/* won't fit, done with this pass */
+      Ncolors[j]++;		/* OK, apply the increment */
+      total_colors = (int) temp;
+      changed = TRUE;
+    }
+  } while (changed);
+
+  return total_colors;
+}
+
+
+LOCAL(int)
+output_value (j_decompress_ptr cinfo, int ci, int j, int maxj)
+/* Return j'th output value, where j will range from 0 to maxj */
+/* The output values must fall in 0..MAXJSAMPLE in increasing order */
+{
+  /* We always provide values 0 and MAXJSAMPLE for each component;
+   * any additional values are equally spaced between these limits.
+   * (Forcing the upper and lower values to the limits ensures that
+   * dithering can't produce a color outside the selected gamut.)
+   */
+  return (int) (((INT32) j * MAXJSAMPLE + maxj/2) / maxj);
+}
+
+
+LOCAL(int)
+largest_input_value (j_decompress_ptr cinfo, int ci, int j, int maxj)
+/* Return largest input value that should map to j'th output value */
+/* Must have largest(j=0) >= 0, and largest(j=maxj) >= MAXJSAMPLE */
+{
+  /* Breakpoints are halfway between values returned by output_value */
+  return (int) (((INT32) (2*j + 1) * MAXJSAMPLE + maxj) / (2*maxj));
+}
+
+
+/*
+ * Create the colormap.
+ */
+
+LOCAL(void)
+create_colormap (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPARRAY colormap;		/* Created colormap */
+  int total_colors;		/* Number of distinct output colors */
+  int i,j,k, nci, blksize, blkdist, ptr, val;
+
+  /* Select number of colors for each component */
+  total_colors = select_ncolors(cinfo, cquantize->Ncolors);
+
+  /* Report selected color counts */
+  if (cinfo->out_color_components == 3)
+    TRACEMS4(cinfo, 1, JTRC_QUANT_3_NCOLORS,
+	     total_colors, cquantize->Ncolors[0],
+	     cquantize->Ncolors[1], cquantize->Ncolors[2]);
+  else
+    TRACEMS1(cinfo, 1, JTRC_QUANT_NCOLORS, total_colors);
+
+  /* Allocate and fill in the colormap. */
+  /* The colors are ordered in the map in standard row-major order, */
+  /* i.e. rightmost (highest-indexed) color changes most rapidly. */
+
+  colormap = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     (JDIMENSION) total_colors, (JDIMENSION) cinfo->out_color_components);
+
+  /* blksize is number of adjacent repeated entries for a component */
+  /* blkdist is distance between groups of identical entries for a component */
+  blkdist = total_colors;
+
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    /* fill in colormap entries for i'th color component */
+    nci = cquantize->Ncolors[i]; /* # of distinct values for this color */
+    blksize = blkdist / nci;
+    for (j = 0; j < nci; j++) {
+      /* Compute j'th output value (out of nci) for component */
+      val = output_value(cinfo, i, j, nci-1);
+      /* Fill in all colormap entries that have this value of this component */
+      for (ptr = j * blksize; ptr < total_colors; ptr += blkdist) {
+	/* fill in blksize entries beginning at ptr */
+	for (k = 0; k < blksize; k++)
+	  colormap[i][ptr+k] = (JSAMPLE) val;
+      }
+    }
+    blkdist = blksize;		/* blksize of this color is blkdist of next */
+  }
+
+  /* Save the colormap in private storage,
+   * where it will survive color quantization mode changes.
+   */
+  cquantize->sv_colormap = colormap;
+  cquantize->sv_actual = total_colors;
+}
+
+
+/*
+ * Create the color index table.
+ */
+
+LOCAL(void)
+create_colorindex (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPROW indexptr;
+  int i,j,k, nci, blksize, val, pad;
+
+  /* For ordered dither, we pad the color index tables by MAXJSAMPLE in
+   * each direction (input index values can be -MAXJSAMPLE .. 2*MAXJSAMPLE).
+   * This is not necessary in the other dithering modes.  However, we
+   * flag whether it was done in case user changes dithering mode.
+   */
+  if (cinfo->dither_mode == JDITHER_ORDERED) {
+    pad = MAXJSAMPLE*2;
+    cquantize->is_padded = TRUE;
+  } else {
+    pad = 0;
+    cquantize->is_padded = FALSE;
+  }
+
+  cquantize->colorindex = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     (JDIMENSION) (MAXJSAMPLE+1 + pad),
+     (JDIMENSION) cinfo->out_color_components);
+
+  /* blksize is number of adjacent repeated entries for a component */
+  blksize = cquantize->sv_actual;
+
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    /* fill in colorindex entries for i'th color component */
+    nci = cquantize->Ncolors[i]; /* # of distinct values for this color */
+    blksize = blksize / nci;
+
+    /* adjust colorindex pointers to provide padding at negative indexes. */
+    if (pad)
+      cquantize->colorindex[i] += MAXJSAMPLE;
+
+    /* in loop, val = index of current output value, */
+    /* and k = largest j that maps to current val */
+    indexptr = cquantize->colorindex[i];
+    val = 0;
+    k = largest_input_value(cinfo, i, 0, nci-1);
+    for (j = 0; j <= MAXJSAMPLE; j++) {
+      while (j > k)		/* advance val if past boundary */
+	k = largest_input_value(cinfo, i, ++val, nci-1);
+      /* premultiply so that no multiplication needed in main processing */
+      indexptr[j] = (JSAMPLE) (val * blksize);
+    }
+    /* Pad at both ends if necessary */
+    if (pad)
+      for (j = 1; j <= MAXJSAMPLE; j++) {
+	indexptr[-j] = indexptr[0];
+	indexptr[MAXJSAMPLE+j] = indexptr[MAXJSAMPLE];
+      }
+  }
+}
+
+
+/*
+ * Create an ordered-dither array for a component having ncolors
+ * distinct output values.
+ */
+
+LOCAL(ODITHER_MATRIX_PTR)
+make_odither_array (j_decompress_ptr cinfo, int ncolors)
+{
+  ODITHER_MATRIX_PTR odither;
+  int j,k;
+  INT32 num,den;
+
+  odither = (ODITHER_MATRIX_PTR)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(ODITHER_MATRIX));
+  /* The inter-value distance for this color is MAXJSAMPLE/(ncolors-1).
+   * Hence the dither value for the matrix cell with fill order f
+   * (f=0..N-1) should be (N-1-2*f)/(2*N) * MAXJSAMPLE/(ncolors-1).
+   * On 16-bit-int machine, be careful to avoid overflow.
+   */
+  den = 2 * ODITHER_CELLS * ((INT32) (ncolors - 1));
+  for (j = 0; j < ODITHER_SIZE; j++) {
+    for (k = 0; k < ODITHER_SIZE; k++) {
+      num = ((INT32) (ODITHER_CELLS-1 - 2*((int)base_dither_matrix[j][k])))
+	    * MAXJSAMPLE;
+      /* Ensure round towards zero despite C's lack of consistency
+       * about rounding negative values in integer division...
+       */
+      odither[j][k] = (int) (num<0 ? -((-num)/den) : num/den);
+    }
+  }
+  return odither;
+}
+
+
+/*
+ * Create the ordered-dither tables.
+ * Components having the same number of representative colors may 
+ * share a dither table.
+ */
+
+LOCAL(void)
+create_odither_tables (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  ODITHER_MATRIX_PTR odither;
+  int i, j, nci;
+
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    nci = cquantize->Ncolors[i]; /* # of distinct values for this color */
+    odither = NULL;		/* search for matching prior component */
+    for (j = 0; j < i; j++) {
+      if (nci == cquantize->Ncolors[j]) {
+	odither = cquantize->odither[j];
+	break;
+      }
+    }
+    if (odither == NULL)	/* need a new table? */
+      odither = make_odither_array(cinfo, nci);
+    cquantize->odither[i] = odither;
+  }
+}
+
+
+/*
+ * Map some rows of pixels to the output colormapped representation.
+ */
+
+METHODDEF(void)
+color_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		JSAMPARRAY output_buf, int num_rows)
+/* General case, no dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPARRAY colorindex = cquantize->colorindex;
+  register int pixcode, ci;
+  register JSAMPROW ptrin, ptrout;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+  register int nc = cinfo->out_color_components;
+
+  for (row = 0; row < num_rows; row++) {
+    ptrin = input_buf[row];
+    ptrout = output_buf[row];
+    for (col = width; col > 0; col--) {
+      pixcode = 0;
+      for (ci = 0; ci < nc; ci++) {
+	pixcode += GETJSAMPLE(colorindex[ci][GETJSAMPLE(*ptrin++)]);
+      }
+      *ptrout++ = (JSAMPLE) pixcode;
+    }
+  }
+}
+
+
+METHODDEF(void)
+color_quantize3 (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		 JSAMPARRAY output_buf, int num_rows)
+/* Fast path for out_color_components==3, no dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  register int pixcode;
+  register JSAMPROW ptrin, ptrout;
+  JSAMPROW colorindex0 = cquantize->colorindex[0];
+  JSAMPROW colorindex1 = cquantize->colorindex[1];
+  JSAMPROW colorindex2 = cquantize->colorindex[2];
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    ptrin = input_buf[row];
+    ptrout = output_buf[row];
+    for (col = width; col > 0; col--) {
+      pixcode  = GETJSAMPLE(colorindex0[GETJSAMPLE(*ptrin++)]);
+      pixcode += GETJSAMPLE(colorindex1[GETJSAMPLE(*ptrin++)]);
+      pixcode += GETJSAMPLE(colorindex2[GETJSAMPLE(*ptrin++)]);
+      *ptrout++ = (JSAMPLE) pixcode;
+    }
+  }
+}
+
+
+METHODDEF(void)
+quantize_ord_dither (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		     JSAMPARRAY output_buf, int num_rows)
+/* General case, with ordered dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  register JSAMPROW input_ptr;
+  register JSAMPROW output_ptr;
+  JSAMPROW colorindex_ci;
+  int * dither;			/* points to active row of dither matrix */
+  int row_index, col_index;	/* current indexes into dither matrix */
+  int nc = cinfo->out_color_components;
+  int ci;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    /* Initialize output values to 0 so can process components separately */
+    jzero_far((void FAR *) output_buf[row],
+	      (size_t) (width * SIZEOF(JSAMPLE)));
+    row_index = cquantize->row_index;
+    for (ci = 0; ci < nc; ci++) {
+      input_ptr = input_buf[row] + ci;
+      output_ptr = output_buf[row];
+      colorindex_ci = cquantize->colorindex[ci];
+      dither = cquantize->odither[ci][row_index];
+      col_index = 0;
+
+      for (col = width; col > 0; col--) {
+	/* Form pixel value + dither, range-limit to 0..MAXJSAMPLE,
+	 * select output value, accumulate into output code for this pixel.
+	 * Range-limiting need not be done explicitly, as we have extended
+	 * the colorindex table to produce the right answers for out-of-range
+	 * inputs.  The maximum dither is +- MAXJSAMPLE; this sets the
+	 * required amount of padding.
+	 */
+	*output_ptr += colorindex_ci[GETJSAMPLE(*input_ptr)+dither[col_index]];
+	input_ptr += nc;
+	output_ptr++;
+	col_index = (col_index + 1) & ODITHER_MASK;
+      }
+    }
+    /* Advance row index for next row */
+    row_index = (row_index + 1) & ODITHER_MASK;
+    cquantize->row_index = row_index;
+  }
+}
+
+
+METHODDEF(void)
+quantize3_ord_dither (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		      JSAMPARRAY output_buf, int num_rows)
+/* Fast path for out_color_components==3, with ordered dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  register int pixcode;
+  register JSAMPROW input_ptr;
+  register JSAMPROW output_ptr;
+  JSAMPROW colorindex0 = cquantize->colorindex[0];
+  JSAMPROW colorindex1 = cquantize->colorindex[1];
+  JSAMPROW colorindex2 = cquantize->colorindex[2];
+  int * dither0;		/* points to active row of dither matrix */
+  int * dither1;
+  int * dither2;
+  int row_index, col_index;	/* current indexes into dither matrix */
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    row_index = cquantize->row_index;
+    input_ptr = input_buf[row];
+    output_ptr = output_buf[row];
+    dither0 = cquantize->odither[0][row_index];
+    dither1 = cquantize->odither[1][row_index];
+    dither2 = cquantize->odither[2][row_index];
+    col_index = 0;
+
+    for (col = width; col > 0; col--) {
+      pixcode  = GETJSAMPLE(colorindex0[GETJSAMPLE(*input_ptr++) +
+					dither0[col_index]]);
+      pixcode += GETJSAMPLE(colorindex1[GETJSAMPLE(*input_ptr++) +
+					dither1[col_index]]);
+      pixcode += GETJSAMPLE(colorindex2[GETJSAMPLE(*input_ptr++) +
+					dither2[col_index]]);
+      *output_ptr++ = (JSAMPLE) pixcode;
+      col_index = (col_index + 1) & ODITHER_MASK;
+    }
+    row_index = (row_index + 1) & ODITHER_MASK;
+    cquantize->row_index = row_index;
+  }
+}
+
+
+METHODDEF(void)
+quantize_fs_dither (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		    JSAMPARRAY output_buf, int num_rows)
+/* General case, with Floyd-Steinberg dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  register LOCFSERROR cur;	/* current error or pixel value */
+  LOCFSERROR belowerr;		/* error for pixel below cur */
+  LOCFSERROR bpreverr;		/* error for below/prev col */
+  LOCFSERROR bnexterr;		/* error for below/next col */
+  LOCFSERROR delta;
+  register FSERRPTR errorptr;	/* => fserrors[] at column before current */
+  register JSAMPROW input_ptr;
+  register JSAMPROW output_ptr;
+  JSAMPROW colorindex_ci;
+  JSAMPROW colormap_ci;
+  int pixcode;
+  int nc = cinfo->out_color_components;
+  int dir;			/* 1 for left-to-right, -1 for right-to-left */
+  int dirnc;			/* dir * nc */
+  int ci;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+  JSAMPLE *range_limit = cinfo->sample_range_limit;
+  SHIFT_TEMPS
+
+  for (row = 0; row < num_rows; row++) {
+    /* Initialize output values to 0 so can process components separately */
+    jzero_far((void FAR *) output_buf[row],
+	      (size_t) (width * SIZEOF(JSAMPLE)));
+    for (ci = 0; ci < nc; ci++) {
+      input_ptr = input_buf[row] + ci;
+      output_ptr = output_buf[row];
+      if (cquantize->on_odd_row) {
+	/* work right to left in this row */
+	input_ptr += (width-1) * nc; /* so point to rightmost pixel */
+	output_ptr += width-1;
+	dir = -1;
+	dirnc = -nc;
+	errorptr = cquantize->fserrors[ci] + (width+1); /* => entry after last column */
+      } else {
+	/* work left to right in this row */
+	dir = 1;
+	dirnc = nc;
+	errorptr = cquantize->fserrors[ci]; /* => entry before first column */
+      }
+      colorindex_ci = cquantize->colorindex[ci];
+      colormap_ci = cquantize->sv_colormap[ci];
+      /* Preset error values: no error propagated to first pixel from left */
+      cur = 0;
+      /* and no error propagated to row below yet */
+      belowerr = bpreverr = 0;
+
+      for (col = width; col > 0; col--) {
+	/* cur holds the error propagated from the previous pixel on the
+	 * current line.  Add the error propagated from the previous line
+	 * to form the complete error correction term for this pixel, and
+	 * round the error term (which is expressed * 16) to an integer.
+	 * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
+	 * for either sign of the error value.
+	 * Note: errorptr points to *previous* column's array entry.
+	 */
+	cur = RIGHT_SHIFT(cur + errorptr[dir] + 8, 4);
+	/* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
+	 * The maximum error is +- MAXJSAMPLE; this sets the required size
+	 * of the range_limit array.
+	 */
+	cur += GETJSAMPLE(*input_ptr);
+	cur = GETJSAMPLE(range_limit[cur]);
+	/* Select output value, accumulate into output code for this pixel */
+	pixcode = GETJSAMPLE(colorindex_ci[cur]);
+	*output_ptr += (JSAMPLE) pixcode;
+	/* Compute actual representation error at this pixel */
+	/* Note: we can do this even though we don't have the final */
+	/* pixel code, because the colormap is orthogonal. */
+	cur -= GETJSAMPLE(colormap_ci[pixcode]);
+	/* Compute error fractions to be propagated to adjacent pixels.
+	 * Add these into the running sums, and simultaneously shift the
+	 * next-line error sums left by 1 column.
+	 */
+	bnexterr = cur;
+	delta = cur * 2;
+	cur += delta;		/* form error * 3 */
+	errorptr[0] = (FSERROR) (bpreverr + cur);
+	cur += delta;		/* form error * 5 */
+	bpreverr = belowerr + cur;
+	belowerr = bnexterr;
+	cur += delta;		/* form error * 7 */
+	/* At this point cur contains the 7/16 error value to be propagated
+	 * to the next pixel on the current line, and all the errors for the
+	 * next line have been shifted over. We are therefore ready to move on.
+	 */
+	input_ptr += dirnc;	/* advance input ptr to next column */
+	output_ptr += dir;	/* advance output ptr to next column */
+	errorptr += dir;	/* advance errorptr to current column */
+      }
+      /* Post-loop cleanup: we must unload the final error value into the
+       * final fserrors[] entry.  Note we need not unload belowerr because
+       * it is for the dummy column before or after the actual array.
+       */
+      errorptr[0] = (FSERROR) bpreverr; /* unload prev err into array */
+    }
+    cquantize->on_odd_row = (cquantize->on_odd_row ? FALSE : TRUE);
+  }
+}
+
+
+/*
+ * Allocate workspace for Floyd-Steinberg errors.
+ */
+
+LOCAL(void)
+alloc_fs_workspace (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  size_t arraysize;
+  int i;
+
+  arraysize = (size_t) ((cinfo->output_width + 2) * SIZEOF(FSERROR));
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    cquantize->fserrors[i] = (FSERRPTR)
+      (*cinfo->mem->alloc_large)((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
+  }
+}
+
+
+/*
+ * Initialize for one-pass color quantization.
+ */
+
+METHODDEF(void)
+start_pass_1_quant (j_decompress_ptr cinfo, boolean is_pre_scan)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  size_t arraysize;
+  int i;
+
+  /* Install my colormap. */
+  cinfo->colormap = cquantize->sv_colormap;
+  cinfo->actual_number_of_colors = cquantize->sv_actual;
+
+  /* Initialize for desired dithering mode. */
+  switch (cinfo->dither_mode) {
+  case JDITHER_NONE:
+    if (cinfo->out_color_components == 3)
+      cquantize->pub.color_quantize = color_quantize3;
+    else
+      cquantize->pub.color_quantize = color_quantize;
+    break;
+  case JDITHER_ORDERED:
+    if (cinfo->out_color_components == 3)
+      cquantize->pub.color_quantize = quantize3_ord_dither;
+    else
+      cquantize->pub.color_quantize = quantize_ord_dither;
+    cquantize->row_index = 0;	/* initialize state for ordered dither */
+    /* If user changed to ordered dither from another mode,
+     * we must recreate the color index table with padding.
+     * This will cost extra space, but probably isn't very likely.
+     */
+    if (! cquantize->is_padded)
+      create_colorindex(cinfo);
+    /* Create ordered-dither tables if we didn't already. */
+    if (cquantize->odither[0] == NULL)
+      create_odither_tables(cinfo);
+    break;
+  case JDITHER_FS:
+    cquantize->pub.color_quantize = quantize_fs_dither;
+    cquantize->on_odd_row = FALSE; /* initialize state for F-S dither */
+    /* Allocate Floyd-Steinberg workspace if didn't already. */
+    if (cquantize->fserrors[0] == NULL)
+      alloc_fs_workspace(cinfo);
+    /* Initialize the propagated errors to zero. */
+    arraysize = (size_t) ((cinfo->output_width + 2) * SIZEOF(FSERROR));
+    for (i = 0; i < cinfo->out_color_components; i++)
+      jzero_far((void FAR *) cquantize->fserrors[i], arraysize);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+}
+
+
+/*
+ * Finish up at the end of the pass.
+ */
+
+METHODDEF(void)
+finish_pass_1_quant (j_decompress_ptr cinfo)
+{
+  /* no work in 1-pass case */
+}
+
+
+/*
+ * Switch to a new external colormap between output passes.
+ * Shouldn't get to this module!
+ */
+
+METHODDEF(void)
+new_color_map_1_quant (j_decompress_ptr cinfo)
+{
+  ERREXIT(cinfo, JERR_MODE_CHANGE);
+}
+
+
+/*
+ * Module initialization routine for 1-pass color quantization.
+ */
+
+GLOBAL(void)
+jinit_1pass_quantizer (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize;
+
+  cquantize = (my_cquantize_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_cquantizer));
+  cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
+  cquantize->pub.start_pass = start_pass_1_quant;
+  cquantize->pub.finish_pass = finish_pass_1_quant;
+  cquantize->pub.new_color_map = new_color_map_1_quant;
+  cquantize->fserrors[0] = NULL; /* Flag FS workspace not allocated */
+  cquantize->odither[0] = NULL;	/* Also flag odither arrays not allocated */
+
+  /* Make sure my internal arrays won't overflow */
+  if (cinfo->out_color_components > MAX_Q_COMPS)
+    ERREXIT1(cinfo, JERR_QUANT_COMPONENTS, MAX_Q_COMPS);
+  /* Make sure colormap indexes can be represented by JSAMPLEs */
+  if (cinfo->desired_number_of_colors > (MAXJSAMPLE+1))
+    ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXJSAMPLE+1);
+
+  /* Create the colormap and color index table. */
+  create_colormap(cinfo);
+  create_colorindex(cinfo);
+
+  /* Allocate Floyd-Steinberg workspace now if requested.
+   * We do this now since it is FAR storage and may affect the memory
+   * manager's space calculations.  If the user changes to FS dither
+   * mode in a later pass, we will allocate the space then, and will
+   * possibly overrun the max_memory_to_use setting.
+   */
+  if (cinfo->dither_mode == JDITHER_FS)
+    alloc_fs_workspace(cinfo);
+}
+
+#endif /* QUANT_1PASS_SUPPORTED */
diff --git a/jquant2.c b/jquant2.c
new file mode 100644
index 0000000..da964f7
--- /dev/null
+++ b/jquant2.c
@@ -0,0 +1,1293 @@
+/*
+ * jquant2.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Copyright (C) 2009, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains 2-pass color quantization (color mapping) routines.
+ * These routines provide selection of a custom color map for an image,
+ * followed by mapping of the image to that color map, with optional
+ * Floyd-Steinberg dithering.
+ * It is also possible to use just the second pass to map to an arbitrary
+ * externally-given color map.
+ *
+ * Note: ordered dithering is not supported, since there isn't any fast
+ * way to compute intercolor distances; it's unclear that ordered dither's
+ * fundamental assumptions even hold with an irregularly spaced color map.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+
+/*
+ * This module implements the well-known Heckbert paradigm for color
+ * quantization.  Most of the ideas used here can be traced back to
+ * Heckbert's seminal paper
+ *   Heckbert, Paul.  "Color Image Quantization for Frame Buffer Display",
+ *   Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304.
+ *
+ * In the first pass over the image, we accumulate a histogram showing the
+ * usage count of each possible color.  To keep the histogram to a reasonable
+ * size, we reduce the precision of the input; typical practice is to retain
+ * 5 or 6 bits per color, so that 8 or 4 different input values are counted
+ * in the same histogram cell.
+ *
+ * Next, the color-selection step begins with a box representing the whole
+ * color space, and repeatedly splits the "largest" remaining box until we
+ * have as many boxes as desired colors.  Then the mean color in each
+ * remaining box becomes one of the possible output colors.
+ * 
+ * The second pass over the image maps each input pixel to the closest output
+ * color (optionally after applying a Floyd-Steinberg dithering correction).
+ * This mapping is logically trivial, but making it go fast enough requires
+ * considerable care.
+ *
+ * Heckbert-style quantizers vary a good deal in their policies for choosing
+ * the "largest" box and deciding where to cut it.  The particular policies
+ * used here have proved out well in experimental comparisons, but better ones
+ * may yet be found.
+ *
+ * In earlier versions of the IJG code, this module quantized in YCbCr color
+ * space, processing the raw upsampled data without a color conversion step.
+ * This allowed the color conversion math to be done only once per colormap
+ * entry, not once per pixel.  However, that optimization precluded other
+ * useful optimizations (such as merging color conversion with upsampling)
+ * and it also interfered with desired capabilities such as quantizing to an
+ * externally-supplied colormap.  We have therefore abandoned that approach.
+ * The present code works in the post-conversion color space, typically RGB.
+ *
+ * To improve the visual quality of the results, we actually work in scaled
+ * RGB space, giving G distances more weight than R, and R in turn more than
+ * B.  To do everything in integer math, we must use integer scale factors.
+ * The 2/3/1 scale factors used here correspond loosely to the relative
+ * weights of the colors in the NTSC grayscale equation.
+ * If you want to use this code to quantize a non-RGB color space, you'll
+ * probably need to change these scale factors.
+ */
+
+#define R_SCALE 2		/* scale R distances by this much */
+#define G_SCALE 3		/* scale G distances by this much */
+#define B_SCALE 1		/* and B by this much */
+
+static const int c_scales[3]={R_SCALE, G_SCALE, B_SCALE};
+#define C0_SCALE c_scales[rgb_red[cinfo->out_color_space]]
+#define C1_SCALE c_scales[rgb_green[cinfo->out_color_space]]
+#define C2_SCALE c_scales[rgb_blue[cinfo->out_color_space]]
+
+/*
+ * First we have the histogram data structure and routines for creating it.
+ *
+ * The number of bits of precision can be adjusted by changing these symbols.
+ * We recommend keeping 6 bits for G and 5 each for R and B.
+ * If you have plenty of memory and cycles, 6 bits all around gives marginally
+ * better results; if you are short of memory, 5 bits all around will save
+ * some space but degrade the results.
+ * To maintain a fully accurate histogram, we'd need to allocate a "long"
+ * (preferably unsigned long) for each cell.  In practice this is overkill;
+ * we can get by with 16 bits per cell.  Few of the cell counts will overflow,
+ * and clamping those that do overflow to the maximum value will give close-
+ * enough results.  This reduces the recommended histogram size from 256Kb
+ * to 128Kb, which is a useful savings on PC-class machines.
+ * (In the second pass the histogram space is re-used for pixel mapping data;
+ * in that capacity, each cell must be able to store zero to the number of
+ * desired colors.  16 bits/cell is plenty for that too.)
+ * Since the JPEG code is intended to run in small memory model on 80x86
+ * machines, we can't just allocate the histogram in one chunk.  Instead
+ * of a true 3-D array, we use a row of pointers to 2-D arrays.  Each
+ * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and
+ * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries.  Note that
+ * on 80x86 machines, the pointer row is in near memory but the actual
+ * arrays are in far memory (same arrangement as we use for image arrays).
+ */
+
+#define MAXNUMCOLORS  (MAXJSAMPLE+1) /* maximum size of colormap */
+
+/* These will do the right thing for either R,G,B or B,G,R color order,
+ * but you may not like the results for other color orders.
+ */
+#define HIST_C0_BITS  5		/* bits of precision in R/B histogram */
+#define HIST_C1_BITS  6		/* bits of precision in G histogram */
+#define HIST_C2_BITS  5		/* bits of precision in B/R histogram */
+
+/* Number of elements along histogram axes. */
+#define HIST_C0_ELEMS  (1<<HIST_C0_BITS)
+#define HIST_C1_ELEMS  (1<<HIST_C1_BITS)
+#define HIST_C2_ELEMS  (1<<HIST_C2_BITS)
+
+/* These are the amounts to shift an input value to get a histogram index. */
+#define C0_SHIFT  (BITS_IN_JSAMPLE-HIST_C0_BITS)
+#define C1_SHIFT  (BITS_IN_JSAMPLE-HIST_C1_BITS)
+#define C2_SHIFT  (BITS_IN_JSAMPLE-HIST_C2_BITS)
+
+
+typedef UINT16 histcell;	/* histogram cell; prefer an unsigned type */
+
+typedef histcell FAR * histptr;	/* for pointers to histogram cells */
+
+typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
+typedef hist1d FAR * hist2d;	/* type for the 2nd-level pointers */
+typedef hist2d * hist3d;	/* type for top-level pointer */
+
+
+/* Declarations for Floyd-Steinberg dithering.
+ *
+ * Errors are accumulated into the array fserrors[], at a resolution of
+ * 1/16th of a pixel count.  The error at a given pixel is propagated
+ * to its not-yet-processed neighbors using the standard F-S fractions,
+ *		...	(here)	7/16
+ *		3/16	5/16	1/16
+ * We work left-to-right on even rows, right-to-left on odd rows.
+ *
+ * We can get away with a single array (holding one row's worth of errors)
+ * by using it to store the current row's errors at pixel columns not yet
+ * processed, but the next row's errors at columns already processed.  We
+ * need only a few extra variables to hold the errors immediately around the
+ * current column.  (If we are lucky, those variables are in registers, but
+ * even if not, they're probably cheaper to access than array elements are.)
+ *
+ * The fserrors[] array has (#columns + 2) entries; the extra entry at
+ * each end saves us from special-casing the first and last pixels.
+ * Each entry is three values long, one value for each color component.
+ *
+ * Note: on a wide image, we might not have enough room in a PC's near data
+ * segment to hold the error array; so it is allocated with alloc_large.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+typedef INT16 FSERROR;		/* 16 bits should be enough */
+typedef int LOCFSERROR;		/* use 'int' for calculation temps */
+#else
+typedef INT32 FSERROR;		/* may need more than 16 bits */
+typedef INT32 LOCFSERROR;	/* be sure calculation temps are big enough */
+#endif
+
+typedef FSERROR FAR *FSERRPTR;	/* pointer to error array (in FAR storage!) */
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_quantizer pub; /* public fields */
+
+  /* Space for the eventually created colormap is stashed here */
+  JSAMPARRAY sv_colormap;	/* colormap allocated at init time */
+  int desired;			/* desired # of colors = size of colormap */
+
+  /* Variables for accumulating image statistics */
+  hist3d histogram;		/* pointer to the histogram */
+
+  boolean needs_zeroed;		/* TRUE if next pass must zero histogram */
+
+  /* Variables for Floyd-Steinberg dithering */
+  FSERRPTR fserrors;		/* accumulated errors */
+  boolean on_odd_row;		/* flag to remember which row we are on */
+  int * error_limiter;		/* table for clamping the applied error */
+} my_cquantizer;
+
+typedef my_cquantizer * my_cquantize_ptr;
+
+
+/*
+ * Prescan some rows of pixels.
+ * In this module the prescan simply updates the histogram, which has been
+ * initialized to zeroes by start_pass.
+ * An output_buf parameter is required by the method signature, but no data
+ * is actually output (in fact the buffer controller is probably passing a
+ * NULL pointer).
+ */
+
+METHODDEF(void)
+prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		  JSAMPARRAY output_buf, int num_rows)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  register JSAMPROW ptr;
+  register histptr histp;
+  register hist3d histogram = cquantize->histogram;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    ptr = input_buf[row];
+    for (col = width; col > 0; col--) {
+      /* get pixel value and index into the histogram */
+      histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
+			 [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
+			 [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
+      /* increment, check for overflow and undo increment if so. */
+      if (++(*histp) <= 0)
+	(*histp)--;
+      ptr += 3;
+    }
+  }
+}
+
+
+/*
+ * Next we have the really interesting routines: selection of a colormap
+ * given the completed histogram.
+ * These routines work with a list of "boxes", each representing a rectangular
+ * subset of the input color space (to histogram precision).
+ */
+
+typedef struct {
+  /* The bounds of the box (inclusive); expressed as histogram indexes */
+  int c0min, c0max;
+  int c1min, c1max;
+  int c2min, c2max;
+  /* The volume (actually 2-norm) of the box */
+  INT32 volume;
+  /* The number of nonzero histogram cells within this box */
+  long colorcount;
+} box;
+
+typedef box * boxptr;
+
+
+LOCAL(boxptr)
+find_biggest_color_pop (boxptr boxlist, int numboxes)
+/* Find the splittable box with the largest color population */
+/* Returns NULL if no splittable boxes remain */
+{
+  register boxptr boxp;
+  register int i;
+  register long maxc = 0;
+  boxptr which = NULL;
+  
+  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
+    if (boxp->colorcount > maxc && boxp->volume > 0) {
+      which = boxp;
+      maxc = boxp->colorcount;
+    }
+  }
+  return which;
+}
+
+
+LOCAL(boxptr)
+find_biggest_volume (boxptr boxlist, int numboxes)
+/* Find the splittable box with the largest (scaled) volume */
+/* Returns NULL if no splittable boxes remain */
+{
+  register boxptr boxp;
+  register int i;
+  register INT32 maxv = 0;
+  boxptr which = NULL;
+  
+  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
+    if (boxp->volume > maxv) {
+      which = boxp;
+      maxv = boxp->volume;
+    }
+  }
+  return which;
+}
+
+
+LOCAL(void)
+update_box (j_decompress_ptr cinfo, boxptr boxp)
+/* Shrink the min/max bounds of a box to enclose only nonzero elements, */
+/* and recompute its volume and population */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  histptr histp;
+  int c0,c1,c2;
+  int c0min,c0max,c1min,c1max,c2min,c2max;
+  INT32 dist0,dist1,dist2;
+  long ccount;
+  
+  c0min = boxp->c0min;  c0max = boxp->c0max;
+  c1min = boxp->c1min;  c1max = boxp->c1max;
+  c2min = boxp->c2min;  c2max = boxp->c2max;
+  
+  if (c0max > c0min)
+    for (c0 = c0min; c0 <= c0max; c0++)
+      for (c1 = c1min; c1 <= c1max; c1++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c0min = c0min = c0;
+	    goto have_c0min;
+	  }
+      }
+ have_c0min:
+  if (c0max > c0min)
+    for (c0 = c0max; c0 >= c0min; c0--)
+      for (c1 = c1min; c1 <= c1max; c1++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c0max = c0max = c0;
+	    goto have_c0max;
+	  }
+      }
+ have_c0max:
+  if (c1max > c1min)
+    for (c1 = c1min; c1 <= c1max; c1++)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c1min = c1min = c1;
+	    goto have_c1min;
+	  }
+      }
+ have_c1min:
+  if (c1max > c1min)
+    for (c1 = c1max; c1 >= c1min; c1--)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c1max = c1max = c1;
+	    goto have_c1max;
+	  }
+      }
+ have_c1max:
+  if (c2max > c2min)
+    for (c2 = c2min; c2 <= c2max; c2++)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1min][c2];
+	for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
+	  if (*histp != 0) {
+	    boxp->c2min = c2min = c2;
+	    goto have_c2min;
+	  }
+      }
+ have_c2min:
+  if (c2max > c2min)
+    for (c2 = c2max; c2 >= c2min; c2--)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1min][c2];
+	for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
+	  if (*histp != 0) {
+	    boxp->c2max = c2max = c2;
+	    goto have_c2max;
+	  }
+      }
+ have_c2max:
+
+  /* Update box volume.
+   * We use 2-norm rather than real volume here; this biases the method
+   * against making long narrow boxes, and it has the side benefit that
+   * a box is splittable iff norm > 0.
+   * Since the differences are expressed in histogram-cell units,
+   * we have to shift back to JSAMPLE units to get consistent distances;
+   * after which, we scale according to the selected distance scale factors.
+   */
+  dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE;
+  dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE;
+  dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE;
+  boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2;
+  
+  /* Now scan remaining volume of box and compute population */
+  ccount = 0;
+  for (c0 = c0min; c0 <= c0max; c0++)
+    for (c1 = c1min; c1 <= c1max; c1++) {
+      histp = & histogram[c0][c1][c2min];
+      for (c2 = c2min; c2 <= c2max; c2++, histp++)
+	if (*histp != 0) {
+	  ccount++;
+	}
+    }
+  boxp->colorcount = ccount;
+}
+
+
+LOCAL(int)
+median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes,
+	    int desired_colors)
+/* Repeatedly select and split the largest box until we have enough boxes */
+{
+  int n,lb;
+  int c0,c1,c2,cmax;
+  register boxptr b1,b2;
+
+  while (numboxes < desired_colors) {
+    /* Select box to split.
+     * Current algorithm: by population for first half, then by volume.
+     */
+    if (numboxes*2 <= desired_colors) {
+      b1 = find_biggest_color_pop(boxlist, numboxes);
+    } else {
+      b1 = find_biggest_volume(boxlist, numboxes);
+    }
+    if (b1 == NULL)		/* no splittable boxes left! */
+      break;
+    b2 = &boxlist[numboxes];	/* where new box will go */
+    /* Copy the color bounds to the new box. */
+    b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max;
+    b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min;
+    /* Choose which axis to split the box on.
+     * Current algorithm: longest scaled axis.
+     * See notes in update_box about scaling distances.
+     */
+    c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE;
+    c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE;
+    c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE;
+    /* We want to break any ties in favor of green, then red, blue last.
+     * This code does the right thing for R,G,B or B,G,R color orders only.
+     */
+    if (rgb_red[cinfo->out_color_space] == 0) {
+      cmax = c1; n = 1;
+      if (c0 > cmax) { cmax = c0; n = 0; }
+      if (c2 > cmax) { n = 2; }
+    }
+    else {
+      cmax = c1; n = 1;
+      if (c2 > cmax) { cmax = c2; n = 2; }
+      if (c0 > cmax) { n = 0; }
+    }
+    /* Choose split point along selected axis, and update box bounds.
+     * Current algorithm: split at halfway point.
+     * (Since the box has been shrunk to minimum volume,
+     * any split will produce two nonempty subboxes.)
+     * Note that lb value is max for lower box, so must be < old max.
+     */
+    switch (n) {
+    case 0:
+      lb = (b1->c0max + b1->c0min) / 2;
+      b1->c0max = lb;
+      b2->c0min = lb+1;
+      break;
+    case 1:
+      lb = (b1->c1max + b1->c1min) / 2;
+      b1->c1max = lb;
+      b2->c1min = lb+1;
+      break;
+    case 2:
+      lb = (b1->c2max + b1->c2min) / 2;
+      b1->c2max = lb;
+      b2->c2min = lb+1;
+      break;
+    }
+    /* Update stats for boxes */
+    update_box(cinfo, b1);
+    update_box(cinfo, b2);
+    numboxes++;
+  }
+  return numboxes;
+}
+
+
+LOCAL(void)
+compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor)
+/* Compute representative color for a box, put it in colormap[icolor] */
+{
+  /* Current algorithm: mean weighted by pixels (not colors) */
+  /* Note it is important to get the rounding correct! */
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  histptr histp;
+  int c0,c1,c2;
+  int c0min,c0max,c1min,c1max,c2min,c2max;
+  long count;
+  long total = 0;
+  long c0total = 0;
+  long c1total = 0;
+  long c2total = 0;
+  
+  c0min = boxp->c0min;  c0max = boxp->c0max;
+  c1min = boxp->c1min;  c1max = boxp->c1max;
+  c2min = boxp->c2min;  c2max = boxp->c2max;
+  
+  for (c0 = c0min; c0 <= c0max; c0++)
+    for (c1 = c1min; c1 <= c1max; c1++) {
+      histp = & histogram[c0][c1][c2min];
+      for (c2 = c2min; c2 <= c2max; c2++) {
+	if ((count = *histp++) != 0) {
+	  total += count;
+	  c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count;
+	  c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count;
+	  c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count;
+	}
+      }
+    }
+  
+  cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total);
+  cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total);
+  cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total);
+}
+
+
+LOCAL(void)
+select_colors (j_decompress_ptr cinfo, int desired_colors)
+/* Master routine for color selection */
+{
+  boxptr boxlist;
+  int numboxes;
+  int i;
+
+  /* Allocate workspace for box list */
+  boxlist = (boxptr) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box));
+  /* Initialize one box containing whole space */
+  numboxes = 1;
+  boxlist[0].c0min = 0;
+  boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT;
+  boxlist[0].c1min = 0;
+  boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT;
+  boxlist[0].c2min = 0;
+  boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT;
+  /* Shrink it to actually-used volume and set its statistics */
+  update_box(cinfo, & boxlist[0]);
+  /* Perform median-cut to produce final box list */
+  numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors);
+  /* Compute the representative color for each box, fill colormap */
+  for (i = 0; i < numboxes; i++)
+    compute_color(cinfo, & boxlist[i], i);
+  cinfo->actual_number_of_colors = numboxes;
+  TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes);
+}
+
+
+/*
+ * These routines are concerned with the time-critical task of mapping input
+ * colors to the nearest color in the selected colormap.
+ *
+ * We re-use the histogram space as an "inverse color map", essentially a
+ * cache for the results of nearest-color searches.  All colors within a
+ * histogram cell will be mapped to the same colormap entry, namely the one
+ * closest to the cell's center.  This may not be quite the closest entry to
+ * the actual input color, but it's almost as good.  A zero in the cache
+ * indicates we haven't found the nearest color for that cell yet; the array
+ * is cleared to zeroes before starting the mapping pass.  When we find the
+ * nearest color for a cell, its colormap index plus one is recorded in the
+ * cache for future use.  The pass2 scanning routines call fill_inverse_cmap
+ * when they need to use an unfilled entry in the cache.
+ *
+ * Our method of efficiently finding nearest colors is based on the "locally
+ * sorted search" idea described by Heckbert and on the incremental distance
+ * calculation described by Spencer W. Thomas in chapter III.1 of Graphics
+ * Gems II (James Arvo, ed.  Academic Press, 1991).  Thomas points out that
+ * the distances from a given colormap entry to each cell of the histogram can
+ * be computed quickly using an incremental method: the differences between
+ * distances to adjacent cells themselves differ by a constant.  This allows a
+ * fairly fast implementation of the "brute force" approach of computing the
+ * distance from every colormap entry to every histogram cell.  Unfortunately,
+ * it needs a work array to hold the best-distance-so-far for each histogram
+ * cell (because the inner loop has to be over cells, not colormap entries).
+ * The work array elements have to be INT32s, so the work array would need
+ * 256Kb at our recommended precision.  This is not feasible in DOS machines.
+ *
+ * To get around these problems, we apply Thomas' method to compute the
+ * nearest colors for only the cells within a small subbox of the histogram.
+ * The work array need be only as big as the subbox, so the memory usage
+ * problem is solved.  Furthermore, we need not fill subboxes that are never
+ * referenced in pass2; many images use only part of the color gamut, so a
+ * fair amount of work is saved.  An additional advantage of this
+ * approach is that we can apply Heckbert's locality criterion to quickly
+ * eliminate colormap entries that are far away from the subbox; typically
+ * three-fourths of the colormap entries are rejected by Heckbert's criterion,
+ * and we need not compute their distances to individual cells in the subbox.
+ * The speed of this approach is heavily influenced by the subbox size: too
+ * small means too much overhead, too big loses because Heckbert's criterion
+ * can't eliminate as many colormap entries.  Empirically the best subbox
+ * size seems to be about 1/512th of the histogram (1/8th in each direction).
+ *
+ * Thomas' article also describes a refined method which is asymptotically
+ * faster than the brute-force method, but it is also far more complex and
+ * cannot efficiently be applied to small subboxes.  It is therefore not
+ * useful for programs intended to be portable to DOS machines.  On machines
+ * with plenty of memory, filling the whole histogram in one shot with Thomas'
+ * refined method might be faster than the present code --- but then again,
+ * it might not be any faster, and it's certainly more complicated.
+ */
+
+
+/* log2(histogram cells in update box) for each axis; this can be adjusted */
+#define BOX_C0_LOG  (HIST_C0_BITS-3)
+#define BOX_C1_LOG  (HIST_C1_BITS-3)
+#define BOX_C2_LOG  (HIST_C2_BITS-3)
+
+#define BOX_C0_ELEMS  (1<<BOX_C0_LOG) /* # of hist cells in update box */
+#define BOX_C1_ELEMS  (1<<BOX_C1_LOG)
+#define BOX_C2_ELEMS  (1<<BOX_C2_LOG)
+
+#define BOX_C0_SHIFT  (C0_SHIFT + BOX_C0_LOG)
+#define BOX_C1_SHIFT  (C1_SHIFT + BOX_C1_LOG)
+#define BOX_C2_SHIFT  (C2_SHIFT + BOX_C2_LOG)
+
+
+/*
+ * The next three routines implement inverse colormap filling.  They could
+ * all be folded into one big routine, but splitting them up this way saves
+ * some stack space (the mindist[] and bestdist[] arrays need not coexist)
+ * and may allow some compilers to produce better code by registerizing more
+ * inner-loop variables.
+ */
+
+LOCAL(int)
+find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
+		    JSAMPLE colorlist[])
+/* Locate the colormap entries close enough to an update box to be candidates
+ * for the nearest entry to some cell(s) in the update box.  The update box
+ * is specified by the center coordinates of its first cell.  The number of
+ * candidate colormap entries is returned, and their colormap indexes are
+ * placed in colorlist[].
+ * This routine uses Heckbert's "locally sorted search" criterion to select
+ * the colors that need further consideration.
+ */
+{
+  int numcolors = cinfo->actual_number_of_colors;
+  int maxc0, maxc1, maxc2;
+  int centerc0, centerc1, centerc2;
+  int i, x, ncolors;
+  INT32 minmaxdist, min_dist, max_dist, tdist;
+  INT32 mindist[MAXNUMCOLORS];	/* min distance to colormap entry i */
+
+  /* Compute true coordinates of update box's upper corner and center.
+   * Actually we compute the coordinates of the center of the upper-corner
+   * histogram cell, which are the upper bounds of the volume we care about.
+   * Note that since ">>" rounds down, the "center" values may be closer to
+   * min than to max; hence comparisons to them must be "<=", not "<".
+   */
+  maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT));
+  centerc0 = (minc0 + maxc0) >> 1;
+  maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT));
+  centerc1 = (minc1 + maxc1) >> 1;
+  maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT));
+  centerc2 = (minc2 + maxc2) >> 1;
+
+  /* For each color in colormap, find:
+   *  1. its minimum squared-distance to any point in the update box
+   *     (zero if color is within update box);
+   *  2. its maximum squared-distance to any point in the update box.
+   * Both of these can be found by considering only the corners of the box.
+   * We save the minimum distance for each color in mindist[];
+   * only the smallest maximum distance is of interest.
+   */
+  minmaxdist = 0x7FFFFFFFL;
+
+  for (i = 0; i < numcolors; i++) {
+    /* We compute the squared-c0-distance term, then add in the other two. */
+    x = GETJSAMPLE(cinfo->colormap[0][i]);
+    if (x < minc0) {
+      tdist = (x - minc0) * C0_SCALE;
+      min_dist = tdist*tdist;
+      tdist = (x - maxc0) * C0_SCALE;
+      max_dist = tdist*tdist;
+    } else if (x > maxc0) {
+      tdist = (x - maxc0) * C0_SCALE;
+      min_dist = tdist*tdist;
+      tdist = (x - minc0) * C0_SCALE;
+      max_dist = tdist*tdist;
+    } else {
+      /* within cell range so no contribution to min_dist */
+      min_dist = 0;
+      if (x <= centerc0) {
+	tdist = (x - maxc0) * C0_SCALE;
+	max_dist = tdist*tdist;
+      } else {
+	tdist = (x - minc0) * C0_SCALE;
+	max_dist = tdist*tdist;
+      }
+    }
+
+    x = GETJSAMPLE(cinfo->colormap[1][i]);
+    if (x < minc1) {
+      tdist = (x - minc1) * C1_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - maxc1) * C1_SCALE;
+      max_dist += tdist*tdist;
+    } else if (x > maxc1) {
+      tdist = (x - maxc1) * C1_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - minc1) * C1_SCALE;
+      max_dist += tdist*tdist;
+    } else {
+      /* within cell range so no contribution to min_dist */
+      if (x <= centerc1) {
+	tdist = (x - maxc1) * C1_SCALE;
+	max_dist += tdist*tdist;
+      } else {
+	tdist = (x - minc1) * C1_SCALE;
+	max_dist += tdist*tdist;
+      }
+    }
+
+    x = GETJSAMPLE(cinfo->colormap[2][i]);
+    if (x < minc2) {
+      tdist = (x - minc2) * C2_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - maxc2) * C2_SCALE;
+      max_dist += tdist*tdist;
+    } else if (x > maxc2) {
+      tdist = (x - maxc2) * C2_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - minc2) * C2_SCALE;
+      max_dist += tdist*tdist;
+    } else {
+      /* within cell range so no contribution to min_dist */
+      if (x <= centerc2) {
+	tdist = (x - maxc2) * C2_SCALE;
+	max_dist += tdist*tdist;
+      } else {
+	tdist = (x - minc2) * C2_SCALE;
+	max_dist += tdist*tdist;
+      }
+    }
+
+    mindist[i] = min_dist;	/* save away the results */
+    if (max_dist < minmaxdist)
+      minmaxdist = max_dist;
+  }
+
+  /* Now we know that no cell in the update box is more than minmaxdist
+   * away from some colormap entry.  Therefore, only colors that are
+   * within minmaxdist of some part of the box need be considered.
+   */
+  ncolors = 0;
+  for (i = 0; i < numcolors; i++) {
+    if (mindist[i] <= minmaxdist)
+      colorlist[ncolors++] = (JSAMPLE) i;
+  }
+  return ncolors;
+}
+
+
+LOCAL(void)
+find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
+		  int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[])
+/* Find the closest colormap entry for each cell in the update box,
+ * given the list of candidate colors prepared by find_nearby_colors.
+ * Return the indexes of the closest entries in the bestcolor[] array.
+ * This routine uses Thomas' incremental distance calculation method to
+ * find the distance from a colormap entry to successive cells in the box.
+ */
+{
+  int ic0, ic1, ic2;
+  int i, icolor;
+  register INT32 * bptr;	/* pointer into bestdist[] array */
+  JSAMPLE * cptr;		/* pointer into bestcolor[] array */
+  INT32 dist0, dist1;		/* initial distance values */
+  register INT32 dist2;		/* current distance in inner loop */
+  INT32 xx0, xx1;		/* distance increments */
+  register INT32 xx2;
+  INT32 inc0, inc1, inc2;	/* initial values for increments */
+  /* This array holds the distance to the nearest-so-far color for each cell */
+  INT32 bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
+
+  /* Initialize best-distance for each cell of the update box */
+  bptr = bestdist;
+  for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--)
+    *bptr++ = 0x7FFFFFFFL;
+  
+  /* For each color selected by find_nearby_colors,
+   * compute its distance to the center of each cell in the box.
+   * If that's less than best-so-far, update best distance and color number.
+   */
+  
+  /* Nominal steps between cell centers ("x" in Thomas article) */
+#define STEP_C0  ((1 << C0_SHIFT) * C0_SCALE)
+#define STEP_C1  ((1 << C1_SHIFT) * C1_SCALE)
+#define STEP_C2  ((1 << C2_SHIFT) * C2_SCALE)
+  
+  for (i = 0; i < numcolors; i++) {
+    icolor = GETJSAMPLE(colorlist[i]);
+    /* Compute (square of) distance from minc0/c1/c2 to this color */
+    inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
+    dist0 = inc0*inc0;
+    inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
+    dist0 += inc1*inc1;
+    inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
+    dist0 += inc2*inc2;
+    /* Form the initial difference increments */
+    inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
+    inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1;
+    inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2;
+    /* Now loop over all cells in box, updating distance per Thomas method */
+    bptr = bestdist;
+    cptr = bestcolor;
+    xx0 = inc0;
+    for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) {
+      dist1 = dist0;
+      xx1 = inc1;
+      for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) {
+	dist2 = dist1;
+	xx2 = inc2;
+	for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) {
+	  if (dist2 < *bptr) {
+	    *bptr = dist2;
+	    *cptr = (JSAMPLE) icolor;
+	  }
+	  dist2 += xx2;
+	  xx2 += 2 * STEP_C2 * STEP_C2;
+	  bptr++;
+	  cptr++;
+	}
+	dist1 += xx1;
+	xx1 += 2 * STEP_C1 * STEP_C1;
+      }
+      dist0 += xx0;
+      xx0 += 2 * STEP_C0 * STEP_C0;
+    }
+  }
+}
+
+
+LOCAL(void)
+fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2)
+/* Fill the inverse-colormap entries in the update box that contains */
+/* histogram cell c0/c1/c2.  (Only that one cell MUST be filled, but */
+/* we can fill as many others as we wish.) */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  int minc0, minc1, minc2;	/* lower left corner of update box */
+  int ic0, ic1, ic2;
+  register JSAMPLE * cptr;	/* pointer into bestcolor[] array */
+  register histptr cachep;	/* pointer into main cache array */
+  /* This array lists the candidate colormap indexes. */
+  JSAMPLE colorlist[MAXNUMCOLORS];
+  int numcolors;		/* number of candidate colors */
+  /* This array holds the actually closest colormap index for each cell. */
+  JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
+
+  /* Convert cell coordinates to update box ID */
+  c0 >>= BOX_C0_LOG;
+  c1 >>= BOX_C1_LOG;
+  c2 >>= BOX_C2_LOG;
+
+  /* Compute true coordinates of update box's origin corner.
+   * Actually we compute the coordinates of the center of the corner
+   * histogram cell, which are the lower bounds of the volume we care about.
+   */
+  minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1);
+  minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1);
+  minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1);
+  
+  /* Determine which colormap entries are close enough to be candidates
+   * for the nearest entry to some cell in the update box.
+   */
+  numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist);
+
+  /* Determine the actually nearest colors. */
+  find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist,
+		   bestcolor);
+
+  /* Save the best color numbers (plus 1) in the main cache array */
+  c0 <<= BOX_C0_LOG;		/* convert ID back to base cell indexes */
+  c1 <<= BOX_C1_LOG;
+  c2 <<= BOX_C2_LOG;
+  cptr = bestcolor;
+  for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) {
+    for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
+      cachep = & histogram[c0+ic0][c1+ic1][c2];
+      for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
+	*cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1);
+      }
+    }
+  }
+}
+
+
+/*
+ * Map some rows of pixels to the output colormapped representation.
+ */
+
+METHODDEF(void)
+pass2_no_dither (j_decompress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
+/* This version performs no dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  register JSAMPROW inptr, outptr;
+  register histptr cachep;
+  register int c0, c1, c2;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    inptr = input_buf[row];
+    outptr = output_buf[row];
+    for (col = width; col > 0; col--) {
+      /* get pixel value and index into the cache */
+      c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
+      c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
+      c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
+      cachep = & histogram[c0][c1][c2];
+      /* If we have not seen this color before, find nearest colormap entry */
+      /* and update the cache */
+      if (*cachep == 0)
+	fill_inverse_cmap(cinfo, c0,c1,c2);
+      /* Now emit the colormap index for this cell */
+      *outptr++ = (JSAMPLE) (*cachep - 1);
+    }
+  }
+}
+
+
+METHODDEF(void)
+pass2_fs_dither (j_decompress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
+/* This version performs Floyd-Steinberg dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  register LOCFSERROR cur0, cur1, cur2;	/* current error or pixel value */
+  LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */
+  LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */
+  register FSERRPTR errorptr;	/* => fserrors[] at column before current */
+  JSAMPROW inptr;		/* => current input pixel */
+  JSAMPROW outptr;		/* => current output pixel */
+  histptr cachep;
+  int dir;			/* +1 or -1 depending on direction */
+  int dir3;			/* 3*dir, for advancing inptr & errorptr */
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+  JSAMPLE *range_limit = cinfo->sample_range_limit;
+  int *error_limit = cquantize->error_limiter;
+  JSAMPROW colormap0 = cinfo->colormap[0];
+  JSAMPROW colormap1 = cinfo->colormap[1];
+  JSAMPROW colormap2 = cinfo->colormap[2];
+  SHIFT_TEMPS
+
+  for (row = 0; row < num_rows; row++) {
+    inptr = input_buf[row];
+    outptr = output_buf[row];
+    if (cquantize->on_odd_row) {
+      /* work right to left in this row */
+      inptr += (width-1) * 3;	/* so point to rightmost pixel */
+      outptr += width-1;
+      dir = -1;
+      dir3 = -3;
+      errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */
+      cquantize->on_odd_row = FALSE; /* flip for next time */
+    } else {
+      /* work left to right in this row */
+      dir = 1;
+      dir3 = 3;
+      errorptr = cquantize->fserrors; /* => entry before first real column */
+      cquantize->on_odd_row = TRUE; /* flip for next time */
+    }
+    /* Preset error values: no error propagated to first pixel from left */
+    cur0 = cur1 = cur2 = 0;
+    /* and no error propagated to row below yet */
+    belowerr0 = belowerr1 = belowerr2 = 0;
+    bpreverr0 = bpreverr1 = bpreverr2 = 0;
+
+    for (col = width; col > 0; col--) {
+      /* curN holds the error propagated from the previous pixel on the
+       * current line.  Add the error propagated from the previous line
+       * to form the complete error correction term for this pixel, and
+       * round the error term (which is expressed * 16) to an integer.
+       * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
+       * for either sign of the error value.
+       * Note: errorptr points to *previous* column's array entry.
+       */
+      cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4);
+      cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4);
+      cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4);
+      /* Limit the error using transfer function set by init_error_limit.
+       * See comments with init_error_limit for rationale.
+       */
+      cur0 = error_limit[cur0];
+      cur1 = error_limit[cur1];
+      cur2 = error_limit[cur2];
+      /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
+       * The maximum error is +- MAXJSAMPLE (or less with error limiting);
+       * this sets the required size of the range_limit array.
+       */
+      cur0 += GETJSAMPLE(inptr[0]);
+      cur1 += GETJSAMPLE(inptr[1]);
+      cur2 += GETJSAMPLE(inptr[2]);
+      cur0 = GETJSAMPLE(range_limit[cur0]);
+      cur1 = GETJSAMPLE(range_limit[cur1]);
+      cur2 = GETJSAMPLE(range_limit[cur2]);
+      /* Index into the cache with adjusted pixel value */
+      cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT];
+      /* If we have not seen this color before, find nearest colormap */
+      /* entry and update the cache */
+      if (*cachep == 0)
+	fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT);
+      /* Now emit the colormap index for this cell */
+      { register int pixcode = *cachep - 1;
+	*outptr = (JSAMPLE) pixcode;
+	/* Compute representation error for this pixel */
+	cur0 -= GETJSAMPLE(colormap0[pixcode]);
+	cur1 -= GETJSAMPLE(colormap1[pixcode]);
+	cur2 -= GETJSAMPLE(colormap2[pixcode]);
+      }
+      /* Compute error fractions to be propagated to adjacent pixels.
+       * Add these into the running sums, and simultaneously shift the
+       * next-line error sums left by 1 column.
+       */
+      { register LOCFSERROR bnexterr, delta;
+
+	bnexterr = cur0;	/* Process component 0 */
+	delta = cur0 * 2;
+	cur0 += delta;		/* form error * 3 */
+	errorptr[0] = (FSERROR) (bpreverr0 + cur0);
+	cur0 += delta;		/* form error * 5 */
+	bpreverr0 = belowerr0 + cur0;
+	belowerr0 = bnexterr;
+	cur0 += delta;		/* form error * 7 */
+	bnexterr = cur1;	/* Process component 1 */
+	delta = cur1 * 2;
+	cur1 += delta;		/* form error * 3 */
+	errorptr[1] = (FSERROR) (bpreverr1 + cur1);
+	cur1 += delta;		/* form error * 5 */
+	bpreverr1 = belowerr1 + cur1;
+	belowerr1 = bnexterr;
+	cur1 += delta;		/* form error * 7 */
+	bnexterr = cur2;	/* Process component 2 */
+	delta = cur2 * 2;
+	cur2 += delta;		/* form error * 3 */
+	errorptr[2] = (FSERROR) (bpreverr2 + cur2);
+	cur2 += delta;		/* form error * 5 */
+	bpreverr2 = belowerr2 + cur2;
+	belowerr2 = bnexterr;
+	cur2 += delta;		/* form error * 7 */
+      }
+      /* At this point curN contains the 7/16 error value to be propagated
+       * to the next pixel on the current line, and all the errors for the
+       * next line have been shifted over.  We are therefore ready to move on.
+       */
+      inptr += dir3;		/* Advance pixel pointers to next column */
+      outptr += dir;
+      errorptr += dir3;		/* advance errorptr to current column */
+    }
+    /* Post-loop cleanup: we must unload the final error values into the
+     * final fserrors[] entry.  Note we need not unload belowerrN because
+     * it is for the dummy column before or after the actual array.
+     */
+    errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */
+    errorptr[1] = (FSERROR) bpreverr1;
+    errorptr[2] = (FSERROR) bpreverr2;
+  }
+}
+
+
+/*
+ * Initialize the error-limiting transfer function (lookup table).
+ * The raw F-S error computation can potentially compute error values of up to
+ * +- MAXJSAMPLE.  But we want the maximum correction applied to a pixel to be
+ * much less, otherwise obviously wrong pixels will be created.  (Typical
+ * effects include weird fringes at color-area boundaries, isolated bright
+ * pixels in a dark area, etc.)  The standard advice for avoiding this problem
+ * is to ensure that the "corners" of the color cube are allocated as output
+ * colors; then repeated errors in the same direction cannot cause cascading
+ * error buildup.  However, that only prevents the error from getting
+ * completely out of hand; Aaron Giles reports that error limiting improves
+ * the results even with corner colors allocated.
+ * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty
+ * well, but the smoother transfer function used below is even better.  Thanks
+ * to Aaron Giles for this idea.
+ */
+
+LOCAL(void)
+init_error_limit (j_decompress_ptr cinfo)
+/* Allocate and fill in the error_limiter table */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  int * table;
+  int in, out;
+
+  table = (int *) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int));
+  table += MAXJSAMPLE;		/* so can index -MAXJSAMPLE .. +MAXJSAMPLE */
+  cquantize->error_limiter = table;
+
+#define STEPSIZE ((MAXJSAMPLE+1)/16)
+  /* Map errors 1:1 up to +- MAXJSAMPLE/16 */
+  out = 0;
+  for (in = 0; in < STEPSIZE; in++, out++) {
+    table[in] = out; table[-in] = -out;
+  }
+  /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */
+  for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) {
+    table[in] = out; table[-in] = -out;
+  }
+  /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */
+  for (; in <= MAXJSAMPLE; in++) {
+    table[in] = out; table[-in] = -out;
+  }
+#undef STEPSIZE
+}
+
+
+/*
+ * Finish up at the end of each pass.
+ */
+
+METHODDEF(void)
+finish_pass1 (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+
+  /* Select the representative colors and fill in cinfo->colormap */
+  cinfo->colormap = cquantize->sv_colormap;
+  select_colors(cinfo, cquantize->desired);
+  /* Force next pass to zero the color index table */
+  cquantize->needs_zeroed = TRUE;
+}
+
+
+METHODDEF(void)
+finish_pass2 (j_decompress_ptr cinfo)
+{
+  /* no work */
+}
+
+
+/*
+ * Initialize for each processing pass.
+ */
+
+METHODDEF(void)
+start_pass_2_quant (j_decompress_ptr cinfo, boolean is_pre_scan)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  int i;
+
+  /* Only F-S dithering or no dithering is supported. */
+  /* If user asks for ordered dither, give him F-S. */
+  if (cinfo->dither_mode != JDITHER_NONE)
+    cinfo->dither_mode = JDITHER_FS;
+
+  if (is_pre_scan) {
+    /* Set up method pointers */
+    cquantize->pub.color_quantize = prescan_quantize;
+    cquantize->pub.finish_pass = finish_pass1;
+    cquantize->needs_zeroed = TRUE; /* Always zero histogram */
+  } else {
+    /* Set up method pointers */
+    if (cinfo->dither_mode == JDITHER_FS)
+      cquantize->pub.color_quantize = pass2_fs_dither;
+    else
+      cquantize->pub.color_quantize = pass2_no_dither;
+    cquantize->pub.finish_pass = finish_pass2;
+
+    /* Make sure color count is acceptable */
+    i = cinfo->actual_number_of_colors;
+    if (i < 1)
+      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1);
+    if (i > MAXNUMCOLORS)
+      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
+
+    if (cinfo->dither_mode == JDITHER_FS) {
+      size_t arraysize = (size_t) ((cinfo->output_width + 2) *
+				   (3 * SIZEOF(FSERROR)));
+      /* Allocate Floyd-Steinberg workspace if we didn't already. */
+      if (cquantize->fserrors == NULL)
+	cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
+      /* Initialize the propagated errors to zero. */
+      jzero_far((void FAR *) cquantize->fserrors, arraysize);
+      /* Make the error-limit table if we didn't already. */
+      if (cquantize->error_limiter == NULL)
+	init_error_limit(cinfo);
+      cquantize->on_odd_row = FALSE;
+    }
+
+  }
+  /* Zero the histogram or inverse color map, if necessary */
+  if (cquantize->needs_zeroed) {
+    for (i = 0; i < HIST_C0_ELEMS; i++) {
+      jzero_far((void FAR *) histogram[i],
+		HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
+    }
+    cquantize->needs_zeroed = FALSE;
+  }
+}
+
+
+/*
+ * Switch to a new external colormap between output passes.
+ */
+
+METHODDEF(void)
+new_color_map_2_quant (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+
+  /* Reset the inverse color map */
+  cquantize->needs_zeroed = TRUE;
+}
+
+
+/*
+ * Module initialization routine for 2-pass color quantization.
+ */
+
+GLOBAL(void)
+jinit_2pass_quantizer (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize;
+  int i;
+
+  cquantize = (my_cquantize_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_cquantizer));
+  cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
+  cquantize->pub.start_pass = start_pass_2_quant;
+  cquantize->pub.new_color_map = new_color_map_2_quant;
+  cquantize->fserrors = NULL;	/* flag optional arrays not allocated */
+  cquantize->error_limiter = NULL;
+
+  /* Make sure jdmaster didn't give me a case I can't handle */
+  if (cinfo->out_color_components != 3)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  /* Allocate the histogram/inverse colormap storage */
+  cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d));
+  for (i = 0; i < HIST_C0_ELEMS; i++) {
+    cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
+  }
+  cquantize->needs_zeroed = TRUE; /* histogram is garbage now */
+
+  /* Allocate storage for the completed colormap, if required.
+   * We do this now since it is FAR storage and may affect
+   * the memory manager's space calculations.
+   */
+  if (cinfo->enable_2pass_quant) {
+    /* Make sure color count is acceptable */
+    int desired = cinfo->desired_number_of_colors;
+    /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */
+    if (desired < 8)
+      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8);
+    /* Make sure colormap indexes can be represented by JSAMPLEs */
+    if (desired > MAXNUMCOLORS)
+      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
+    cquantize->sv_colormap = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3);
+    cquantize->desired = desired;
+  } else
+    cquantize->sv_colormap = NULL;
+
+  /* Only F-S dithering or no dithering is supported. */
+  /* If user asks for ordered dither, give him F-S. */
+  if (cinfo->dither_mode != JDITHER_NONE)
+    cinfo->dither_mode = JDITHER_FS;
+
+  /* Allocate Floyd-Steinberg workspace if necessary.
+   * This isn't really needed until pass 2, but again it is FAR storage.
+   * Although we will cope with a later change in dither_mode,
+   * we do not promise to honor max_memory_to_use if dither_mode changes.
+   */
+  if (cinfo->dither_mode == JDITHER_FS) {
+    cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR))));
+    /* Might as well create the error-limiting table too. */
+    init_error_limit(cinfo);
+  }
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
diff --git a/jsimd.h b/jsimd.h
new file mode 100644
index 0000000..90be19e
--- /dev/null
+++ b/jsimd.h
@@ -0,0 +1,89 @@
+/*
+ * jsimd.h
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * 
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ *
+ */
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jsimd_can_rgb_ycc                 jSCanRgbYcc
+#define jsimd_can_ycc_rgb                 jSCanYccRgb
+#define jsimd_rgb_ycc_convert             jSRgbYccConv
+#define jsimd_ycc_rgb_convert             jSYccRgbConv
+#define jsimd_can_h2v2_downsample         jSCanH2V2Down
+#define jsimd_can_h2v1_downsample         jSCanH2V1Down
+#define jsimd_h2v2_downsample             jSH2V2Down
+#define jsimd_h2v1_downsample             jSH2V1Down
+#define jsimd_can_h2v2_upsample           jSCanH2V2Up
+#define jsimd_can_h2v1_upsample           jSCanH2V1Up
+#define jsimd_h2v2_upsample               jSH2V2Up
+#define jsimd_h2v1_upsample               jSH2V1Up
+#define jsimd_can_h2v2_fancy_upsample     jSCanH2V2FUp
+#define jsimd_can_h2v1_fancy_upsample     jSCanH2V1FUp
+#define jsimd_h2v2_fancy_upsample         jSH2V2FUp
+#define jsimd_h2v1_fancy_upsample         jSH2V1FUp
+#define jsimd_can_h2v2_merged_upsample    jSCanH2V2MUp
+#define jsimd_can_h2v1_merged_upsample    jSCanH2V1MUp
+#define jsimd_h2v2_merged_upsample        jSH2V2MUp
+#define jsimd_h2v1_merged_upsample        jSH2V1MUp
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+EXTERN(int) jsimd_can_rgb_ycc JPP((void));
+EXTERN(int) jsimd_can_ycc_rgb JPP((void));
+
+EXTERN(void) jsimd_rgb_ycc_convert
+        JPP((j_compress_ptr cinfo,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_ycc_rgb_convert
+        JPP((j_decompress_ptr cinfo,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+
+EXTERN(int) jsimd_can_h2v2_downsample JPP((void));
+EXTERN(int) jsimd_can_h2v1_downsample JPP((void));
+
+EXTERN(void) jsimd_h2v2_downsample
+        JPP((j_compress_ptr cinfo, jpeg_component_info * compptr,
+             JSAMPARRAY input_data, JSAMPARRAY output_data));
+EXTERN(void) jsimd_h2v1_downsample
+        JPP((j_compress_ptr cinfo, jpeg_component_info * compptr,
+             JSAMPARRAY input_data, JSAMPARRAY output_data));
+
+EXTERN(int) jsimd_can_h2v2_upsample JPP((void));
+EXTERN(int) jsimd_can_h2v1_upsample JPP((void));
+
+EXTERN(void) jsimd_h2v2_upsample
+        JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+EXTERN(void) jsimd_h2v1_upsample
+        JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+EXTERN(int) jsimd_can_h2v2_fancy_upsample JPP((void));
+EXTERN(int) jsimd_can_h2v1_fancy_upsample JPP((void));
+
+EXTERN(void) jsimd_h2v2_fancy_upsample
+        JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+EXTERN(void) jsimd_h2v1_fancy_upsample
+        JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+EXTERN(int) jsimd_can_h2v2_merged_upsample JPP((void));
+EXTERN(int) jsimd_can_h2v1_merged_upsample JPP((void));
+
+EXTERN(void) jsimd_h2v2_merged_upsample
+        JPP((j_decompress_ptr cinfo,
+             JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+             JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_merged_upsample
+        JPP((j_decompress_ptr cinfo,
+             JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+             JSAMPARRAY output_buf));
+
diff --git a/jsimd_none.c b/jsimd_none.c
new file mode 100644
index 0000000..8960802
--- /dev/null
+++ b/jsimd_none.c
@@ -0,0 +1,299 @@
+/*
+ * jsimd_none.c
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright 2009 D. R. Commander
+ * 
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ *
+ * This file contains stubs for when there is no SIMD support available.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jsimd.h"
+#include "jdct.h"
+#include "jsimddct.h"
+
+GLOBAL(int)
+jsimd_can_rgb_ycc (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
+                       JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+                       JDIMENSION output_row, int num_rows)
+{
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
+                       JSAMPIMAGE input_buf, JDIMENSION input_row,
+                       JSAMPARRAY output_buf, int num_rows)
+{
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_downsample (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_downsample (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+                       JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+}
+
+GLOBAL(void)
+jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+                       JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_upsample (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_upsample (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_upsample (j_decompress_ptr cinfo,
+                     jpeg_component_info * compptr, 
+                     JSAMPARRAY input_data,
+                     JSAMPARRAY * output_data_ptr)
+{
+}
+
+GLOBAL(void)
+jsimd_h2v1_upsample (j_decompress_ptr cinfo,
+                     jpeg_component_info * compptr, 
+                     JSAMPARRAY input_data,
+                     JSAMPARRAY * output_data_ptr)
+{
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_fancy_upsample (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_fancy_upsample (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
+                           jpeg_component_info * compptr, 
+                           JSAMPARRAY input_data,
+                           JSAMPARRAY * output_data_ptr)
+{
+}
+
+GLOBAL(void)
+jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
+                           jpeg_component_info * compptr, 
+                           JSAMPARRAY input_data,
+                           JSAMPARRAY * output_data_ptr)
+{
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_merged_upsample (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_merged_upsample (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
+                            JSAMPIMAGE input_buf,
+                            JDIMENSION in_row_group_ctr,
+                            JSAMPARRAY output_buf)
+{
+}
+
+GLOBAL(void)
+jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
+                            JSAMPIMAGE input_buf,
+                            JDIMENSION in_row_group_ctr,
+                            JSAMPARRAY output_buf)
+{
+}
+
+GLOBAL(int)
+jsimd_can_convsamp (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_convsamp_float (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
+                DCTELEM * workspace)
+{
+}
+
+GLOBAL(void)
+jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
+                      FAST_FLOAT * workspace)
+{
+}
+
+GLOBAL(int)
+jsimd_can_fdct_islow (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_ifast (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_float (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_fdct_islow (DCTELEM * data)
+{
+}
+
+GLOBAL(void)
+jsimd_fdct_ifast (DCTELEM * data)
+{
+}
+
+GLOBAL(void)
+jsimd_fdct_float (FAST_FLOAT * data)
+{
+}
+
+GLOBAL(int)
+jsimd_can_quantize (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_quantize_float (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
+                DCTELEM * workspace)
+{
+}
+
+GLOBAL(void)
+jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+                      FAST_FLOAT * workspace)
+{
+}
+
+GLOBAL(int)
+jsimd_can_idct_2x2 (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_4x4 (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+}
+
+GLOBAL(void)
+jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+}
+
+GLOBAL(int)
+jsimd_can_idct_islow (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_ifast (void)
+{
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_float (void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+}
+
+GLOBAL(void)
+jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+}
+
+GLOBAL(void)
+jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+}
+
diff --git a/jsimddct.h b/jsimddct.h
new file mode 100644
index 0000000..d73d0c4
--- /dev/null
+++ b/jsimddct.h
@@ -0,0 +1,101 @@
+/*
+ * jsimddct.h
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * 
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ *
+ */
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jsimd_can_convsamp                jSCanConv
+#define jsimd_can_convsamp_float          jSCanConvF
+#define jsimd_convsamp                    jSConv
+#define jsimd_convsamp_float              jSConvF
+#define jsimd_can_fdct_islow              jSCanFDCTIS
+#define jsimd_can_fdct_ifast              jSCanFDCTIF
+#define jsimd_can_fdct_float              jSCanFDCTFl
+#define jsimd_fdct_islow                  jSFDCTIS
+#define jsimd_fdct_ifast                  jSFDCTIF
+#define jsimd_fdct_float                  jSFDCTFl
+#define jsimd_can_quantize                jSCanQuant
+#define jsimd_can_quantize_float          jSCanQuantF
+#define jsimd_quantize                    jSQuant
+#define jsimd_quantize_float              jSQuantF
+#define jsimd_can_idct_2x2                jSCanIDCT22
+#define jsimd_can_idct_4x4                jSCanIDCT44
+#define jsimd_idct_2x2                    jSIDCT22
+#define jsimd_idct_4x4                    jSIDCT44
+#define jsimd_can_idct_islow              jSCanIDCTIS
+#define jsimd_can_idct_ifast              jSCanIDCTIF
+#define jsimd_can_idct_float              jSCanIDCTFl
+#define jsimd_idct_islow                  jSIDCTIS
+#define jsimd_idct_ifast                  jSIDCTIF
+#define jsimd_idct_float                  jSIDCTFl
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+EXTERN(int) jsimd_can_convsamp JPP((void));
+EXTERN(int) jsimd_can_convsamp_float JPP((void));
+
+EXTERN(void) jsimd_convsamp JPP((JSAMPARRAY sample_data,
+                                 JDIMENSION start_col,
+                                 DCTELEM * workspace));
+EXTERN(void) jsimd_convsamp_float JPP((JSAMPARRAY sample_data,
+                                       JDIMENSION start_col,
+                                       FAST_FLOAT * workspace));
+
+EXTERN(int) jsimd_can_fdct_islow JPP((void));
+EXTERN(int) jsimd_can_fdct_ifast JPP((void));
+EXTERN(int) jsimd_can_fdct_float JPP((void));
+
+EXTERN(void) jsimd_fdct_islow JPP((DCTELEM * data));
+EXTERN(void) jsimd_fdct_ifast JPP((DCTELEM * data));
+EXTERN(void) jsimd_fdct_float JPP((FAST_FLOAT * data));
+
+EXTERN(int) jsimd_can_quantize JPP((void));
+EXTERN(int) jsimd_can_quantize_float JPP((void));
+
+EXTERN(void) jsimd_quantize JPP((JCOEFPTR coef_block,
+                                 DCTELEM * divisors,
+                                 DCTELEM * workspace));
+EXTERN(void) jsimd_quantize_float JPP((JCOEFPTR coef_block,
+                                       FAST_FLOAT * divisors,
+                                       FAST_FLOAT * workspace));
+
+EXTERN(int) jsimd_can_idct_2x2 JPP((void));
+EXTERN(int) jsimd_can_idct_4x4 JPP((void));
+
+EXTERN(void) jsimd_idct_2x2 JPP((j_decompress_ptr cinfo,
+                                 jpeg_component_info * compptr,
+                                 JCOEFPTR coef_block,
+                                 JSAMPARRAY output_buf,
+                                 JDIMENSION output_col));
+EXTERN(void) jsimd_idct_4x4 JPP((j_decompress_ptr cinfo,
+                                 jpeg_component_info * compptr,
+                                 JCOEFPTR coef_block,
+                                 JSAMPARRAY output_buf,
+                                 JDIMENSION output_col));
+
+EXTERN(int) jsimd_can_idct_islow JPP((void));
+EXTERN(int) jsimd_can_idct_ifast JPP((void));
+EXTERN(int) jsimd_can_idct_float JPP((void));
+
+EXTERN(void) jsimd_idct_islow JPP((j_decompress_ptr cinfo,
+                                   jpeg_component_info * compptr,
+                                   JCOEFPTR coef_block,
+                                   JSAMPARRAY output_buf,
+                                   JDIMENSION output_col));
+EXTERN(void) jsimd_idct_ifast JPP((j_decompress_ptr cinfo,
+                                   jpeg_component_info * compptr,
+                                   JCOEFPTR coef_block,
+                                   JSAMPARRAY output_buf,
+                                   JDIMENSION output_col));
+EXTERN(void) jsimd_idct_float JPP((j_decompress_ptr cinfo,
+                                   jpeg_component_info * compptr,
+                                   JCOEFPTR coef_block,
+                                   JSAMPARRAY output_buf,
+                                   JDIMENSION output_col));
+
diff --git a/jutils.c b/jutils.c
new file mode 100644
index 0000000..98b54f5
--- /dev/null
+++ b/jutils.c
@@ -0,0 +1,179 @@
+/*
+ * jutils.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains tables and miscellaneous utility routines needed
+ * for both compression and decompression.
+ * Note we prefix all global names with "j" to minimize conflicts with
+ * a surrounding application.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * jpeg_zigzag_order[i] is the zigzag-order position of the i'th element
+ * of a DCT block read in natural order (left to right, top to bottom).
+ */
+
+#if 0				/* This table is not actually needed in v6a */
+
+const int jpeg_zigzag_order[DCTSIZE2] = {
+   0,  1,  5,  6, 14, 15, 27, 28,
+   2,  4,  7, 13, 16, 26, 29, 42,
+   3,  8, 12, 17, 25, 30, 41, 43,
+   9, 11, 18, 24, 31, 40, 44, 53,
+  10, 19, 23, 32, 39, 45, 52, 54,
+  20, 22, 33, 38, 46, 51, 55, 60,
+  21, 34, 37, 47, 50, 56, 59, 61,
+  35, 36, 48, 49, 57, 58, 62, 63
+};
+
+#endif
+
+/*
+ * jpeg_natural_order[i] is the natural-order position of the i'th element
+ * of zigzag order.
+ *
+ * When reading corrupted data, the Huffman decoders could attempt
+ * to reference an entry beyond the end of this array (if the decoded
+ * zero run length reaches past the end of the block).  To prevent
+ * wild stores without adding an inner-loop test, we put some extra
+ * "63"s after the real entries.  This will cause the extra coefficient
+ * to be stored in location 63 of the block, not somewhere random.
+ * The worst case would be a run-length of 15, which means we need 16
+ * fake entries.
+ */
+
+const int jpeg_natural_order[DCTSIZE2+16] = {
+  0,  1,  8, 16,  9,  2,  3, 10,
+ 17, 24, 32, 25, 18, 11,  4,  5,
+ 12, 19, 26, 33, 40, 48, 41, 34,
+ 27, 20, 13,  6,  7, 14, 21, 28,
+ 35, 42, 49, 56, 57, 50, 43, 36,
+ 29, 22, 15, 23, 30, 37, 44, 51,
+ 58, 59, 52, 45, 38, 31, 39, 46,
+ 53, 60, 61, 54, 47, 55, 62, 63,
+ 63, 63, 63, 63, 63, 63, 63, 63, /* extra entries for safety in decoder */
+ 63, 63, 63, 63, 63, 63, 63, 63
+};
+
+
+/*
+ * Arithmetic utilities
+ */
+
+GLOBAL(long)
+jdiv_round_up (long a, long b)
+/* Compute a/b rounded up to next integer, ie, ceil(a/b) */
+/* Assumes a >= 0, b > 0 */
+{
+  return (a + b - 1L) / b;
+}
+
+
+GLOBAL(size_t)
+jround_up (size_t a, size_t b)
+/* Compute a rounded up to next multiple of b, ie, ceil(a/b)*b */
+/* Assumes a >= 0, b > 0 */
+{
+  a += b - 1L;
+  return a - (a % b);
+}
+
+
+/* On normal machines we can apply MEMCOPY() and MEMZERO() to sample arrays
+ * and coefficient-block arrays.  This won't work on 80x86 because the arrays
+ * are FAR and we're assuming a small-pointer memory model.  However, some
+ * DOS compilers provide far-pointer versions of memcpy() and memset() even
+ * in the small-model libraries.  These will be used if USE_FMEM is defined.
+ * Otherwise, the routines below do it the hard way.  (The performance cost
+ * is not all that great, because these routines aren't very heavily used.)
+ */
+
+#ifndef NEED_FAR_POINTERS	/* normal case, same as regular macros */
+#define FMEMCOPY(dest,src,size)	MEMCOPY(dest,src,size)
+#define FMEMZERO(target,size)	MEMZERO(target,size)
+#else				/* 80x86 case, define if we can */
+#ifdef USE_FMEM
+#define FMEMCOPY(dest,src,size)	_fmemcpy((void FAR *)(dest), (const void FAR *)(src), (size_t)(size))
+#define FMEMZERO(target,size)	_fmemset((void FAR *)(target), 0, (size_t)(size))
+#endif
+#endif
+
+
+GLOBAL(void)
+jcopy_sample_rows (JSAMPARRAY input_array, int source_row,
+		   JSAMPARRAY output_array, int dest_row,
+		   int num_rows, JDIMENSION num_cols)
+/* Copy some rows of samples from one place to another.
+ * num_rows rows are copied from input_array[source_row++]
+ * to output_array[dest_row++]; these areas may overlap for duplication.
+ * The source and destination arrays must be at least as wide as num_cols.
+ */
+{
+  register JSAMPROW inptr, outptr;
+#ifdef FMEMCOPY
+  register size_t count = (size_t) (num_cols * SIZEOF(JSAMPLE));
+#else
+  register JDIMENSION count;
+#endif
+  register int row;
+
+  input_array += source_row;
+  output_array += dest_row;
+
+  for (row = num_rows; row > 0; row--) {
+    inptr = *input_array++;
+    outptr = *output_array++;
+#ifdef FMEMCOPY
+    FMEMCOPY(outptr, inptr, count);
+#else
+    for (count = num_cols; count > 0; count--)
+      *outptr++ = *inptr++;	/* needn't bother with GETJSAMPLE() here */
+#endif
+  }
+}
+
+
+GLOBAL(void)
+jcopy_block_row (JBLOCKROW input_row, JBLOCKROW output_row,
+		 JDIMENSION num_blocks)
+/* Copy a row of coefficient blocks from one place to another. */
+{
+#ifdef FMEMCOPY
+  FMEMCOPY(output_row, input_row, num_blocks * (DCTSIZE2 * SIZEOF(JCOEF)));
+#else
+  register JCOEFPTR inptr, outptr;
+  register long count;
+
+  inptr = (JCOEFPTR) input_row;
+  outptr = (JCOEFPTR) output_row;
+  for (count = (long) num_blocks * DCTSIZE2; count > 0; count--) {
+    *outptr++ = *inptr++;
+  }
+#endif
+}
+
+
+GLOBAL(void)
+jzero_far (void FAR * target, size_t bytestozero)
+/* Zero out a chunk of FAR memory. */
+/* This might be sample-array data, block-array data, or alloc_large data. */
+{
+#ifdef FMEMZERO
+  FMEMZERO(target, bytestozero);
+#else
+  register char FAR * ptr = (char FAR *) target;
+  register size_t count;
+
+  for (count = bytestozero; count > 0; count--) {
+    *ptr++ = 0;
+  }
+#endif
+}
diff --git a/jversion.h b/jversion.h
new file mode 100644
index 0000000..82a7e0d
--- /dev/null
+++ b/jversion.h
@@ -0,0 +1,37 @@
+/*
+ * jversion.h
+ *
+ * Copyright (C) 1991-2010, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains software version identification.
+ */
+
+
+#if JPEG_LIB_VERSION >= 80
+
+#define JVERSION	"8b  16-May-2010"
+
+#define JCOPYRIGHT	"Copyright (C) 2010, Thomas G. Lane, Guido Vollbeding"
+
+#elif JPEG_LIB_VERSION >= 70
+
+#define JVERSION        "7  27-Jun-2009"
+
+#define JCOPYRIGHT      "Copyright (C) 2009, Thomas G. Lane, Guido Vollbeding"
+
+#else
+
+#define JVERSION	"6b  27-Mar-1998"
+
+#define JCOPYRIGHT	"Copyright (C) 1998, Thomas G. Lane"
+
+#endif
+
+#define LJTCOPYRIGHT	"Copyright (C) 1999-2006 MIYASAKA Masaru\n" \
+			"Copyright (C) 2004 Landmark Graphics Corporation\n" \
+			"Copyright (C) 2005-2007 Sun Microsystems, Inc.\n" \
+			"Copyright (C) 2009 Pierre Ossman for Cendio AB\n" \
+			"Copyright (C) 2009-2010 D. R. Commander"
diff --git a/libjpeg.map.in b/libjpeg.map.in
new file mode 100644
index 0000000..f3e7ff8
--- /dev/null
+++ b/libjpeg.map.in
@@ -0,0 +1,4 @@
+LIBJPEG_@JPEG_LIB_VERSION_DECIMAL@ {
+  global:
+    *;
+};
diff --git a/libjpeg.txt b/libjpeg.txt
new file mode 100644
index 0000000..f8ec519
--- /dev/null
+++ b/libjpeg.txt
@@ -0,0 +1,3049 @@
+USING THE IJG JPEG LIBRARY
+
+Copyright (C) 1994-2010, Thomas G. Lane, Guido Vollbeding, D. R. Commander.
+This file is part of the Independent JPEG Group's software.
+For conditions of distribution and use, see the accompanying README file.
+
+
+This file describes how to use the IJG JPEG library within an application
+program.  Read it if you want to write a program that uses the library.
+
+The file example.c provides heavily commented skeleton code for calling the
+JPEG library.  Also see jpeglib.h (the include file to be used by application
+programs) for full details about data structures and function parameter lists.
+The library source code, of course, is the ultimate reference.
+
+Note that there have been *major* changes from the application interface
+presented by IJG version 4 and earlier versions.  The old design had several
+inherent limitations, and it had accumulated a lot of cruft as we added
+features while trying to minimize application-interface changes.  We have
+sacrificed backward compatibility in the version 5 rewrite, but we think the
+improvements justify this.
+
+
+TABLE OF CONTENTS
+-----------------
+
+Overview:
+	Functions provided by the library
+	Outline of typical usage
+Basic library usage:
+	Data formats
+	Compression details
+	Decompression details
+	Mechanics of usage: include files, linking, etc
+Advanced features:
+	Compression parameter selection
+	Decompression parameter selection
+	Special color spaces
+	Error handling
+	Compressed data handling (source and destination managers)
+	I/O suspension
+	Progressive JPEG support
+	Buffered-image mode
+	Abbreviated datastreams and multiple images
+	Special markers
+	Raw (downsampled) image data
+	Really raw data: DCT coefficients
+	Progress monitoring
+	Memory management
+	Memory usage
+	Library compile-time options
+	Portability considerations
+	Notes for MS-DOS implementors
+
+You should read at least the overview and basic usage sections before trying
+to program with the library.  The sections on advanced features can be read
+if and when you need them.
+
+
+OVERVIEW
+========
+
+Functions provided by the library
+---------------------------------
+
+The IJG JPEG library provides C code to read and write JPEG-compressed image
+files.  The surrounding application program receives or supplies image data a
+scanline at a time, using a straightforward uncompressed image format.  All
+details of color conversion and other preprocessing/postprocessing can be
+handled by the library.
+
+The library includes a substantial amount of code that is not covered by the
+JPEG standard but is necessary for typical applications of JPEG.  These
+functions preprocess the image before JPEG compression or postprocess it after
+decompression.  They include colorspace conversion, downsampling/upsampling,
+and color quantization.  The application indirectly selects use of this code
+by specifying the format in which it wishes to supply or receive image data.
+For example, if colormapped output is requested, then the decompression
+library automatically invokes color quantization.
+
+A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
+and even more so in decompression postprocessing.  The decompression library
+provides multiple implementations that cover most of the useful tradeoffs,
+ranging from very-high-quality down to fast-preview operation.  On the
+compression side we have generally not provided low-quality choices, since
+compression is normally less time-critical.  It should be understood that the
+low-quality modes may not meet the JPEG standard's accuracy requirements;
+nonetheless, they are useful for viewers.
+
+A word about functions *not* provided by the library.  We handle a subset of
+the ISO JPEG standard; most baseline, extended-sequential, and progressive
+JPEG processes are supported.  (Our subset includes all features now in common
+use.)  Unsupported ISO options include:
+	* Hierarchical storage
+	* Lossless JPEG
+	* Arithmetic entropy coding (unsupported for legal reasons)
+	* DNL marker
+	* Nonintegral subsampling ratios
+We support both 8- and 12-bit data precision, but this is a compile-time
+choice rather than a run-time choice; hence it is difficult to use both
+precisions in a single application.
+
+By itself, the library handles only interchange JPEG datastreams --- in
+particular the widely used JFIF file format.  The library can be used by
+surrounding code to process interchange or abbreviated JPEG datastreams that
+are embedded in more complex file formats.  (For example, this library is
+used by the free LIBTIFF library to support JPEG compression in TIFF.)
+
+
+Outline of typical usage
+------------------------
+
+The rough outline of a JPEG compression operation is:
+
+	Allocate and initialize a JPEG compression object
+	Specify the destination for the compressed data (eg, a file)
+	Set parameters for compression, including image size & colorspace
+	jpeg_start_compress(...);
+	while (scan lines remain to be written)
+		jpeg_write_scanlines(...);
+	jpeg_finish_compress(...);
+	Release the JPEG compression object
+
+A JPEG compression object holds parameters and working state for the JPEG
+library.  We make creation/destruction of the object separate from starting
+or finishing compression of an image; the same object can be re-used for a
+series of image compression operations.  This makes it easy to re-use the
+same parameter settings for a sequence of images.  Re-use of a JPEG object
+also has important implications for processing abbreviated JPEG datastreams,
+as discussed later.
+
+The image data to be compressed is supplied to jpeg_write_scanlines() from
+in-memory buffers.  If the application is doing file-to-file compression,
+reading image data from the source file is the application's responsibility.
+The library emits compressed data by calling a "data destination manager",
+which typically will write the data into a file; but the application can
+provide its own destination manager to do something else.
+
+Similarly, the rough outline of a JPEG decompression operation is:
+
+	Allocate and initialize a JPEG decompression object
+	Specify the source of the compressed data (eg, a file)
+	Call jpeg_read_header() to obtain image info
+	Set parameters for decompression
+	jpeg_start_decompress(...);
+	while (scan lines remain to be read)
+		jpeg_read_scanlines(...);
+	jpeg_finish_decompress(...);
+	Release the JPEG decompression object
+
+This is comparable to the compression outline except that reading the
+datastream header is a separate step.  This is helpful because information
+about the image's size, colorspace, etc is available when the application
+selects decompression parameters.  For example, the application can choose an
+output scaling ratio that will fit the image into the available screen size.
+
+The decompression library obtains compressed data by calling a data source
+manager, which typically will read the data from a file; but other behaviors
+can be obtained with a custom source manager.  Decompressed data is delivered
+into in-memory buffers passed to jpeg_read_scanlines().
+
+It is possible to abort an incomplete compression or decompression operation
+by calling jpeg_abort(); or, if you do not need to retain the JPEG object,
+simply release it by calling jpeg_destroy().
+
+JPEG compression and decompression objects are two separate struct types.
+However, they share some common fields, and certain routines such as
+jpeg_destroy() can work on either type of object.
+
+The JPEG library has no static variables: all state is in the compression
+or decompression object.  Therefore it is possible to process multiple
+compression and decompression operations concurrently, using multiple JPEG
+objects.
+
+Both compression and decompression can be done in an incremental memory-to-
+memory fashion, if suitable source/destination managers are used.  See the
+section on "I/O suspension" for more details.
+
+
+BASIC LIBRARY USAGE
+===================
+
+Data formats
+------------
+
+Before diving into procedural details, it is helpful to understand the
+image data format that the JPEG library expects or returns.
+
+The standard input image format is a rectangular array of pixels, with each
+pixel having the same number of "component" or "sample" values (color
+channels).  You must specify how many components there are and the colorspace
+interpretation of the components.  Most applications will use RGB data
+(three components per pixel) or grayscale data (one component per pixel).
+PLEASE NOTE THAT RGB DATA IS THREE SAMPLES PER PIXEL, GRAYSCALE ONLY ONE.
+A remarkable number of people manage to miss this, only to find that their
+programs don't work with grayscale JPEG files.
+
+There is no provision for colormapped input.  JPEG files are always full-color
+or full grayscale (or sometimes another colorspace such as CMYK).  You can
+feed in a colormapped image by expanding it to full-color format.  However
+JPEG often doesn't work very well with source data that has been colormapped,
+because of dithering noise.  This is discussed in more detail in the JPEG FAQ
+and the other references mentioned in the README file.
+
+Pixels are stored by scanlines, with each scanline running from left to
+right.  The component values for each pixel are adjacent in the row; for
+example, R,G,B,R,G,B,R,G,B,... for 24-bit RGB color.  Each scanline is an
+array of data type JSAMPLE --- which is typically "unsigned char", unless
+you've changed jmorecfg.h.  (You can also change the RGB pixel layout, say
+to B,G,R order, by modifying jmorecfg.h.  But see the restrictions listed in
+that file before doing so.)
+
+A 2-D array of pixels is formed by making a list of pointers to the starts of
+scanlines; so the scanlines need not be physically adjacent in memory.  Even
+if you process just one scanline at a time, you must make a one-element
+pointer array to conform to this structure.  Pointers to JSAMPLE rows are of
+type JSAMPROW, and the pointer to the pointer array is of type JSAMPARRAY.
+
+The library accepts or supplies one or more complete scanlines per call.
+It is not possible to process part of a row at a time.  Scanlines are always
+processed top-to-bottom.  You can process an entire image in one call if you
+have it all in memory, but usually it's simplest to process one scanline at
+a time.
+
+For best results, source data values should have the precision specified by
+BITS_IN_JSAMPLE (normally 8 bits).  For instance, if you choose to compress
+data that's only 6 bits/channel, you should left-justify each value in a
+byte before passing it to the compressor.  If you need to compress data
+that has more than 8 bits/channel, compile with BITS_IN_JSAMPLE = 12.
+(See "Library compile-time options", later.)
+
+
+The data format returned by the decompressor is the same in all details,
+except that colormapped output is supported.  (Again, a JPEG file is never
+colormapped.  But you can ask the decompressor to perform on-the-fly color
+quantization to deliver colormapped output.)  If you request colormapped
+output then the returned data array contains a single JSAMPLE per pixel;
+its value is an index into a color map.  The color map is represented as
+a 2-D JSAMPARRAY in which each row holds the values of one color component,
+that is, colormap[i][j] is the value of the i'th color component for pixel
+value (map index) j.  Note that since the colormap indexes are stored in
+JSAMPLEs, the maximum number of colors is limited by the size of JSAMPLE
+(ie, at most 256 colors for an 8-bit JPEG library).
+
+
+Compression details
+-------------------
+
+Here we revisit the JPEG compression outline given in the overview.
+
+1. Allocate and initialize a JPEG compression object.
+
+A JPEG compression object is a "struct jpeg_compress_struct".  (It also has
+a bunch of subsidiary structures which are allocated via malloc(), but the
+application doesn't control those directly.)  This struct can be just a local
+variable in the calling routine, if a single routine is going to execute the
+whole JPEG compression sequence.  Otherwise it can be static or allocated
+from malloc().
+
+You will also need a structure representing a JPEG error handler.  The part
+of this that the library cares about is a "struct jpeg_error_mgr".  If you
+are providing your own error handler, you'll typically want to embed the
+jpeg_error_mgr struct in a larger structure; this is discussed later under
+"Error handling".  For now we'll assume you are just using the default error
+handler.  The default error handler will print JPEG error/warning messages
+on stderr, and it will call exit() if a fatal error occurs.
+
+You must initialize the error handler structure, store a pointer to it into
+the JPEG object's "err" field, and then call jpeg_create_compress() to
+initialize the rest of the JPEG object.
+
+Typical code for this step, if you are using the default error handler, is
+
+	struct jpeg_compress_struct cinfo;
+	struct jpeg_error_mgr jerr;
+	...
+	cinfo.err = jpeg_std_error(&jerr);
+	jpeg_create_compress(&cinfo);
+
+jpeg_create_compress allocates a small amount of memory, so it could fail
+if you are out of memory.  In that case it will exit via the error handler;
+that's why the error handler must be initialized first.
+
+
+2. Specify the destination for the compressed data (eg, a file).
+
+As previously mentioned, the JPEG library delivers compressed data to a
+"data destination" module.  The library includes one data destination
+module which knows how to write to a stdio stream.  You can use your own
+destination module if you want to do something else, as discussed later.
+
+If you use the standard destination module, you must open the target stdio
+stream beforehand.  Typical code for this step looks like:
+
+	FILE * outfile;
+	...
+	if ((outfile = fopen(filename, "wb")) == NULL) {
+	    fprintf(stderr, "can't open %s\n", filename);
+	    exit(1);
+	}
+	jpeg_stdio_dest(&cinfo, outfile);
+
+where the last line invokes the standard destination module.
+
+WARNING: it is critical that the binary compressed data be delivered to the
+output file unchanged.  On non-Unix systems the stdio library may perform
+newline translation or otherwise corrupt binary data.  To suppress this
+behavior, you may need to use a "b" option to fopen (as shown above), or use
+setmode() or another routine to put the stdio stream in binary mode.  See
+cjpeg.c and djpeg.c for code that has been found to work on many systems.
+
+You can select the data destination after setting other parameters (step 3),
+if that's more convenient.  You may not change the destination between
+calling jpeg_start_compress() and jpeg_finish_compress().
+
+
+3. Set parameters for compression, including image size & colorspace.
+
+You must supply information about the source image by setting the following
+fields in the JPEG object (cinfo structure):
+
+	image_width		Width of image, in pixels
+	image_height		Height of image, in pixels
+	input_components	Number of color channels (samples per pixel)
+	in_color_space		Color space of source image
+
+The image dimensions are, hopefully, obvious.  JPEG supports image dimensions
+of 1 to 64K pixels in either direction.  The input color space is typically
+RGB or grayscale, and input_components is 3 or 1 accordingly.  (See "Special
+color spaces", later, for more info.)  The in_color_space field must be
+assigned one of the J_COLOR_SPACE enum constants, typically JCS_RGB or
+JCS_GRAYSCALE.
+
+JPEG has a large number of compression parameters that determine how the
+image is encoded.  Most applications don't need or want to know about all
+these parameters.  You can set all the parameters to reasonable defaults by
+calling jpeg_set_defaults(); then, if there are particular values you want
+to change, you can do so after that.  The "Compression parameter selection"
+section tells about all the parameters.
+
+You must set in_color_space correctly before calling jpeg_set_defaults(),
+because the defaults depend on the source image colorspace.  However the
+other three source image parameters need not be valid until you call
+jpeg_start_compress().  There's no harm in calling jpeg_set_defaults() more
+than once, if that happens to be convenient.
+
+Typical code for a 24-bit RGB source image is
+
+	cinfo.image_width = Width; 	/* image width and height, in pixels */
+	cinfo.image_height = Height;
+	cinfo.input_components = 3;	/* # of color components per pixel */
+	cinfo.in_color_space = JCS_RGB; /* colorspace of input image */
+
+	jpeg_set_defaults(&cinfo);
+	/* Make optional parameter settings here */
+
+
+4. jpeg_start_compress(...);
+
+After you have established the data destination and set all the necessary
+source image info and other parameters, call jpeg_start_compress() to begin
+a compression cycle.  This will initialize internal state, allocate working
+storage, and emit the first few bytes of the JPEG datastream header.
+
+Typical code:
+
+	jpeg_start_compress(&cinfo, TRUE);
+
+The "TRUE" parameter ensures that a complete JPEG interchange datastream
+will be written.  This is appropriate in most cases.  If you think you might
+want to use an abbreviated datastream, read the section on abbreviated
+datastreams, below.
+
+Once you have called jpeg_start_compress(), you may not alter any JPEG
+parameters or other fields of the JPEG object until you have completed
+the compression cycle.
+
+
+5. while (scan lines remain to be written)
+	jpeg_write_scanlines(...);
+
+Now write all the required image data by calling jpeg_write_scanlines()
+one or more times.  You can pass one or more scanlines in each call, up
+to the total image height.  In most applications it is convenient to pass
+just one or a few scanlines at a time.  The expected format for the passed
+data is discussed under "Data formats", above.
+
+Image data should be written in top-to-bottom scanline order.  The JPEG spec
+contains some weasel wording about how top and bottom are application-defined
+terms (a curious interpretation of the English language...) but if you want
+your files to be compatible with everyone else's, you WILL use top-to-bottom
+order.  If the source data must be read in bottom-to-top order, you can use
+the JPEG library's virtual array mechanism to invert the data efficiently.
+Examples of this can be found in the sample application cjpeg.
+
+The library maintains a count of the number of scanlines written so far
+in the next_scanline field of the JPEG object.  Usually you can just use
+this variable as the loop counter, so that the loop test looks like
+"while (cinfo.next_scanline < cinfo.image_height)".
+
+Code for this step depends heavily on the way that you store the source data.
+example.c shows the following code for the case of a full-size 2-D source
+array containing 3-byte RGB pixels:
+
+	JSAMPROW row_pointer[1];	/* pointer to a single row */
+	int row_stride;			/* physical row width in buffer */
+
+	row_stride = image_width * 3;	/* JSAMPLEs per row in image_buffer */
+
+	while (cinfo.next_scanline < cinfo.image_height) {
+	    row_pointer[0] = & image_buffer[cinfo.next_scanline * row_stride];
+	    jpeg_write_scanlines(&cinfo, row_pointer, 1);
+	}
+
+jpeg_write_scanlines() returns the number of scanlines actually written.
+This will normally be equal to the number passed in, so you can usually
+ignore the return value.  It is different in just two cases:
+  * If you try to write more scanlines than the declared image height,
+    the additional scanlines are ignored.
+  * If you use a suspending data destination manager, output buffer overrun
+    will cause the compressor to return before accepting all the passed lines.
+    This feature is discussed under "I/O suspension", below.  The normal
+    stdio destination manager will NOT cause this to happen.
+In any case, the return value is the same as the change in the value of
+next_scanline.
+
+
+6. jpeg_finish_compress(...);
+
+After all the image data has been written, call jpeg_finish_compress() to
+complete the compression cycle.  This step is ESSENTIAL to ensure that the
+last bufferload of data is written to the data destination.
+jpeg_finish_compress() also releases working memory associated with the JPEG
+object.
+
+Typical code:
+
+	jpeg_finish_compress(&cinfo);
+
+If using the stdio destination manager, don't forget to close the output
+stdio stream (if necessary) afterwards.
+
+If you have requested a multi-pass operating mode, such as Huffman code
+optimization, jpeg_finish_compress() will perform the additional passes using
+data buffered by the first pass.  In this case jpeg_finish_compress() may take
+quite a while to complete.  With the default compression parameters, this will
+not happen.
+
+It is an error to call jpeg_finish_compress() before writing the necessary
+total number of scanlines.  If you wish to abort compression, call
+jpeg_abort() as discussed below.
+
+After completing a compression cycle, you may dispose of the JPEG object
+as discussed next, or you may use it to compress another image.  In that case
+return to step 2, 3, or 4 as appropriate.  If you do not change the
+destination manager, the new datastream will be written to the same target.
+If you do not change any JPEG parameters, the new datastream will be written
+with the same parameters as before.  Note that you can change the input image
+dimensions freely between cycles, but if you change the input colorspace, you
+should call jpeg_set_defaults() to adjust for the new colorspace; and then
+you'll need to repeat all of step 3.
+
+
+7. Release the JPEG compression object.
+
+When you are done with a JPEG compression object, destroy it by calling
+jpeg_destroy_compress().  This will free all subsidiary memory (regardless of
+the previous state of the object).  Or you can call jpeg_destroy(), which
+works for either compression or decompression objects --- this may be more
+convenient if you are sharing code between compression and decompression
+cases.  (Actually, these routines are equivalent except for the declared type
+of the passed pointer.  To avoid gripes from ANSI C compilers, jpeg_destroy()
+should be passed a j_common_ptr.)
+
+If you allocated the jpeg_compress_struct structure from malloc(), freeing
+it is your responsibility --- jpeg_destroy() won't.  Ditto for the error
+handler structure.
+
+Typical code:
+
+	jpeg_destroy_compress(&cinfo);
+
+
+8. Aborting.
+
+If you decide to abort a compression cycle before finishing, you can clean up
+in either of two ways:
+
+* If you don't need the JPEG object any more, just call
+  jpeg_destroy_compress() or jpeg_destroy() to release memory.  This is
+  legitimate at any point after calling jpeg_create_compress() --- in fact,
+  it's safe even if jpeg_create_compress() fails.
+
+* If you want to re-use the JPEG object, call jpeg_abort_compress(), or call
+  jpeg_abort() which works on both compression and decompression objects.
+  This will return the object to an idle state, releasing any working memory.
+  jpeg_abort() is allowed at any time after successful object creation.
+
+Note that cleaning up the data destination, if required, is your
+responsibility; neither of these routines will call term_destination().
+(See "Compressed data handling", below, for more about that.)
+
+jpeg_destroy() and jpeg_abort() are the only safe calls to make on a JPEG
+object that has reported an error by calling error_exit (see "Error handling"
+for more info).  The internal state of such an object is likely to be out of
+whack.  Either of these two routines will return the object to a known state.
+
+
+Decompression details
+---------------------
+
+Here we revisit the JPEG decompression outline given in the overview.
+
+1. Allocate and initialize a JPEG decompression object.
+
+This is just like initialization for compression, as discussed above,
+except that the object is a "struct jpeg_decompress_struct" and you
+call jpeg_create_decompress().  Error handling is exactly the same.
+
+Typical code:
+
+	struct jpeg_decompress_struct cinfo;
+	struct jpeg_error_mgr jerr;
+	...
+	cinfo.err = jpeg_std_error(&jerr);
+	jpeg_create_decompress(&cinfo);
+
+(Both here and in the IJG code, we usually use variable name "cinfo" for
+both compression and decompression objects.)
+
+
+2. Specify the source of the compressed data (eg, a file).
+
+As previously mentioned, the JPEG library reads compressed data from a "data
+source" module.  The library includes one data source module which knows how
+to read from a stdio stream.  You can use your own source module if you want
+to do something else, as discussed later.
+
+If you use the standard source module, you must open the source stdio stream
+beforehand.  Typical code for this step looks like:
+
+	FILE * infile;
+	...
+	if ((infile = fopen(filename, "rb")) == NULL) {
+	    fprintf(stderr, "can't open %s\n", filename);
+	    exit(1);
+	}
+	jpeg_stdio_src(&cinfo, infile);
+
+where the last line invokes the standard source module.
+
+WARNING: it is critical that the binary compressed data be read unchanged.
+On non-Unix systems the stdio library may perform newline translation or
+otherwise corrupt binary data.  To suppress this behavior, you may need to use
+a "b" option to fopen (as shown above), or use setmode() or another routine to
+put the stdio stream in binary mode.  See cjpeg.c and djpeg.c for code that
+has been found to work on many systems.
+
+You may not change the data source between calling jpeg_read_header() and
+jpeg_finish_decompress().  If you wish to read a series of JPEG images from
+a single source file, you should repeat the jpeg_read_header() to
+jpeg_finish_decompress() sequence without reinitializing either the JPEG
+object or the data source module; this prevents buffered input data from
+being discarded.
+
+
+3. Call jpeg_read_header() to obtain image info.
+
+Typical code for this step is just
+
+	jpeg_read_header(&cinfo, TRUE);
+
+This will read the source datastream header markers, up to the beginning
+of the compressed data proper.  On return, the image dimensions and other
+info have been stored in the JPEG object.  The application may wish to
+consult this information before selecting decompression parameters.
+
+More complex code is necessary if
+  * A suspending data source is used --- in that case jpeg_read_header()
+    may return before it has read all the header data.  See "I/O suspension",
+    below.  The normal stdio source manager will NOT cause this to happen.
+  * Abbreviated JPEG files are to be processed --- see the section on
+    abbreviated datastreams.  Standard applications that deal only in
+    interchange JPEG files need not be concerned with this case either.
+
+It is permissible to stop at this point if you just wanted to find out the
+image dimensions and other header info for a JPEG file.  In that case,
+call jpeg_destroy() when you are done with the JPEG object, or call
+jpeg_abort() to return it to an idle state before selecting a new data
+source and reading another header.
+
+
+4. Set parameters for decompression.
+
+jpeg_read_header() sets appropriate default decompression parameters based on
+the properties of the image (in particular, its colorspace).  However, you
+may well want to alter these defaults before beginning the decompression.
+For example, the default is to produce full color output from a color file.
+If you want colormapped output you must ask for it.  Other options allow the
+returned image to be scaled and allow various speed/quality tradeoffs to be
+selected.  "Decompression parameter selection", below, gives details.
+
+If the defaults are appropriate, nothing need be done at this step.
+
+Note that all default values are set by each call to jpeg_read_header().
+If you reuse a decompression object, you cannot expect your parameter
+settings to be preserved across cycles, as you can for compression.
+You must set desired parameter values each time.
+
+
+5. jpeg_start_decompress(...);
+
+Once the parameter values are satisfactory, call jpeg_start_decompress() to
+begin decompression.  This will initialize internal state, allocate working
+memory, and prepare for returning data.
+
+Typical code is just
+
+	jpeg_start_decompress(&cinfo);
+
+If you have requested a multi-pass operating mode, such as 2-pass color
+quantization, jpeg_start_decompress() will do everything needed before data
+output can begin.  In this case jpeg_start_decompress() may take quite a while
+to complete.  With a single-scan (non progressive) JPEG file and default
+decompression parameters, this will not happen; jpeg_start_decompress() will
+return quickly.
+
+After this call, the final output image dimensions, including any requested
+scaling, are available in the JPEG object; so is the selected colormap, if
+colormapped output has been requested.  Useful fields include
+
+	output_width		image width and height, as scaled
+	output_height
+	out_color_components	# of color components in out_color_space
+	output_components	# of color components returned per pixel
+	colormap		the selected colormap, if any
+	actual_number_of_colors		number of entries in colormap
+
+output_components is 1 (a colormap index) when quantizing colors; otherwise it
+equals out_color_components.  It is the number of JSAMPLE values that will be
+emitted per pixel in the output arrays.
+
+Typically you will need to allocate data buffers to hold the incoming image.
+You will need output_width * output_components JSAMPLEs per scanline in your
+output buffer, and a total of output_height scanlines will be returned.
+
+Note: if you are using the JPEG library's internal memory manager to allocate
+data buffers (as djpeg does), then the manager's protocol requires that you
+request large buffers *before* calling jpeg_start_decompress().  This is a
+little tricky since the output_XXX fields are not normally valid then.  You
+can make them valid by calling jpeg_calc_output_dimensions() after setting the
+relevant parameters (scaling, output color space, and quantization flag).
+
+
+6. while (scan lines remain to be read)
+	jpeg_read_scanlines(...);
+
+Now you can read the decompressed image data by calling jpeg_read_scanlines()
+one or more times.  At each call, you pass in the maximum number of scanlines
+to be read (ie, the height of your working buffer); jpeg_read_scanlines()
+will return up to that many lines.  The return value is the number of lines
+actually read.  The format of the returned data is discussed under "Data
+formats", above.  Don't forget that grayscale and color JPEGs will return
+different data formats!
+
+Image data is returned in top-to-bottom scanline order.  If you must write
+out the image in bottom-to-top order, you can use the JPEG library's virtual
+array mechanism to invert the data efficiently.  Examples of this can be
+found in the sample application djpeg.
+
+The library maintains a count of the number of scanlines returned so far
+in the output_scanline field of the JPEG object.  Usually you can just use
+this variable as the loop counter, so that the loop test looks like
+"while (cinfo.output_scanline < cinfo.output_height)".  (Note that the test
+should NOT be against image_height, unless you never use scaling.  The
+image_height field is the height of the original unscaled image.)
+The return value always equals the change in the value of output_scanline.
+
+If you don't use a suspending data source, it is safe to assume that
+jpeg_read_scanlines() reads at least one scanline per call, until the
+bottom of the image has been reached.
+
+If you use a buffer larger than one scanline, it is NOT safe to assume that
+jpeg_read_scanlines() fills it.  (The current implementation returns only a
+few scanlines per call, no matter how large a buffer you pass.)  So you must
+always provide a loop that calls jpeg_read_scanlines() repeatedly until the
+whole image has been read.
+
+
+7. jpeg_finish_decompress(...);
+
+After all the image data has been read, call jpeg_finish_decompress() to
+complete the decompression cycle.  This causes working memory associated
+with the JPEG object to be released.
+
+Typical code:
+
+	jpeg_finish_decompress(&cinfo);
+
+If using the stdio source manager, don't forget to close the source stdio
+stream if necessary.
+
+It is an error to call jpeg_finish_decompress() before reading the correct
+total number of scanlines.  If you wish to abort decompression, call
+jpeg_abort() as discussed below.
+
+After completing a decompression cycle, you may dispose of the JPEG object as
+discussed next, or you may use it to decompress another image.  In that case
+return to step 2 or 3 as appropriate.  If you do not change the source
+manager, the next image will be read from the same source.
+
+
+8. Release the JPEG decompression object.
+
+When you are done with a JPEG decompression object, destroy it by calling
+jpeg_destroy_decompress() or jpeg_destroy().  The previous discussion of
+destroying compression objects applies here too.
+
+Typical code:
+
+	jpeg_destroy_decompress(&cinfo);
+
+
+9. Aborting.
+
+You can abort a decompression cycle by calling jpeg_destroy_decompress() or
+jpeg_destroy() if you don't need the JPEG object any more, or
+jpeg_abort_decompress() or jpeg_abort() if you want to reuse the object.
+The previous discussion of aborting compression cycles applies here too.
+
+
+Mechanics of usage: include files, linking, etc
+-----------------------------------------------
+
+Applications using the JPEG library should include the header file jpeglib.h
+to obtain declarations of data types and routines.  Before including
+jpeglib.h, include system headers that define at least the typedefs FILE and
+size_t.  On ANSI-conforming systems, including <stdio.h> is sufficient; on
+older Unix systems, you may need <sys/types.h> to define size_t.
+
+If the application needs to refer to individual JPEG library error codes, also
+include jerror.h to define those symbols.
+
+jpeglib.h indirectly includes the files jconfig.h and jmorecfg.h.  If you are
+installing the JPEG header files in a system directory, you will want to
+install all four files: jpeglib.h, jerror.h, jconfig.h, jmorecfg.h.
+
+The most convenient way to include the JPEG code into your executable program
+is to prepare a library file ("libjpeg.a", or a corresponding name on non-Unix
+machines) and reference it at your link step.  If you use only half of the
+library (only compression or only decompression), only that much code will be
+included from the library, unless your linker is hopelessly brain-damaged.
+The supplied makefiles build libjpeg.a automatically (see install.txt).
+
+While you can build the JPEG library as a shared library if the whim strikes
+you, we don't really recommend it.  The trouble with shared libraries is that
+at some point you'll probably try to substitute a new version of the library
+without recompiling the calling applications.  That generally doesn't work
+because the parameter struct declarations usually change with each new
+version.  In other words, the library's API is *not* guaranteed binary
+compatible across versions; we only try to ensure source-code compatibility.
+(In hindsight, it might have been smarter to hide the parameter structs from
+applications and introduce a ton of access functions instead.  Too late now,
+however.)
+
+On some systems your application may need to set up a signal handler to ensure
+that temporary files are deleted if the program is interrupted.  This is most
+critical if you are on MS-DOS and use the jmemdos.c memory manager back end;
+it will try to grab extended memory for temp files, and that space will NOT be
+freed automatically.  See cjpeg.c or djpeg.c for an example signal handler.
+
+It may be worth pointing out that the core JPEG library does not actually
+require the stdio library: only the default source/destination managers and
+error handler need it.  You can use the library in a stdio-less environment
+if you replace those modules and use jmemnobs.c (or another memory manager of
+your own devising).  More info about the minimum system library requirements
+may be found in jinclude.h.
+
+
+ADVANCED FEATURES
+=================
+
+Compression parameter selection
+-------------------------------
+
+This section describes all the optional parameters you can set for JPEG
+compression, as well as the "helper" routines provided to assist in this
+task.  Proper setting of some parameters requires detailed understanding
+of the JPEG standard; if you don't know what a parameter is for, it's best
+not to mess with it!  See REFERENCES in the README file for pointers to
+more info about JPEG.
+
+It's a good idea to call jpeg_set_defaults() first, even if you plan to set
+all the parameters; that way your code is more likely to work with future JPEG
+libraries that have additional parameters.  For the same reason, we recommend
+you use a helper routine where one is provided, in preference to twiddling
+cinfo fields directly.
+
+The helper routines are:
+
+jpeg_set_defaults (j_compress_ptr cinfo)
+	This routine sets all JPEG parameters to reasonable defaults, using
+	only the input image's color space (field in_color_space, which must
+	already be set in cinfo).  Many applications will only need to use
+	this routine and perhaps jpeg_set_quality().
+
+jpeg_set_colorspace (j_compress_ptr cinfo, J_COLOR_SPACE colorspace)
+	Sets the JPEG file's colorspace (field jpeg_color_space) as specified,
+	and sets other color-space-dependent parameters appropriately.  See
+	"Special color spaces", below, before using this.  A large number of
+	parameters, including all per-component parameters, are set by this
+	routine; if you want to twiddle individual parameters you should call
+	jpeg_set_colorspace() before rather than after.
+
+jpeg_default_colorspace (j_compress_ptr cinfo)
+	Selects an appropriate JPEG colorspace based on cinfo->in_color_space,
+	and calls jpeg_set_colorspace().  This is actually a subroutine of
+	jpeg_set_defaults().  It's broken out in case you want to change
+	just the colorspace-dependent JPEG parameters.
+
+jpeg_set_quality (j_compress_ptr cinfo, int quality, boolean force_baseline)
+	Constructs JPEG quantization tables appropriate for the indicated
+	quality setting.  The quality value is expressed on the 0..100 scale
+	recommended by IJG (cjpeg's "-quality" switch uses this routine).
+	Note that the exact mapping from quality values to tables may change
+	in future IJG releases as more is learned about DCT quantization.
+	If the force_baseline parameter is TRUE, then the quantization table
+	entries are constrained to the range 1..255 for full JPEG baseline
+	compatibility.  In the current implementation, this only makes a
+	difference for quality settings below 25, and it effectively prevents
+	very small/low quality files from being generated.  The IJG decoder
+	is capable of reading the non-baseline files generated at low quality
+	settings when force_baseline is FALSE, but other decoders may not be.
+
+jpeg_set_linear_quality (j_compress_ptr cinfo, int scale_factor,
+			 boolean force_baseline)
+	Same as jpeg_set_quality() except that the generated tables are the
+	sample tables given in the JPEC spec section K.1, multiplied by the
+	specified scale factor (which is expressed as a percentage; thus
+	scale_factor = 100 reproduces the spec's tables).  Note that larger
+	scale factors give lower quality.  This entry point is useful for
+	conforming to the Adobe PostScript DCT conventions, but we do not
+	recommend linear scaling as a user-visible quality scale otherwise.
+	force_baseline again constrains the computed table entries to 1..255.
+
+int jpeg_quality_scaling (int quality)
+	Converts a value on the IJG-recommended quality scale to a linear
+	scaling percentage.  Note that this routine may change or go away
+	in future releases --- IJG may choose to adopt a scaling method that
+	can't be expressed as a simple scalar multiplier, in which case the
+	premise of this routine collapses.  Caveat user.
+
+jpeg_default_qtables (j_compress_ptr cinfo, boolean force_baseline)
+	[libjpeg v7/v8 only] Set default quantization tables with linear
+	q_scale_factor[] values (see below).
+
+jpeg_add_quant_table (j_compress_ptr cinfo, int which_tbl,
+		      const unsigned int *basic_table,
+		      int scale_factor, boolean force_baseline)
+	Allows an arbitrary quantization table to be created.  which_tbl
+	indicates which table slot to fill.  basic_table points to an array
+	of 64 unsigned ints given in normal array order.  These values are
+	multiplied by scale_factor/100 and then clamped to the range 1..65535
+	(or to 1..255 if force_baseline is TRUE).
+	CAUTION: prior to library version 6a, jpeg_add_quant_table expected
+	the basic table to be given in JPEG zigzag order.  If you need to
+	write code that works with either older or newer versions of this
+	routine, you must check the library version number.  Something like
+	"#if JPEG_LIB_VERSION >= 61" is the right test.
+
+jpeg_simple_progression (j_compress_ptr cinfo)
+	Generates a default scan script for writing a progressive-JPEG file.
+	This is the recommended method of creating a progressive file,
+	unless you want to make a custom scan sequence.  You must ensure that
+	the JPEG color space is set correctly before calling this routine.
+
+
+Compression parameters (cinfo fields) include:
+
+J_DCT_METHOD dct_method
+	Selects the algorithm used for the DCT step.  Choices are:
+		JDCT_ISLOW: slow but accurate integer algorithm
+		JDCT_IFAST: faster, less accurate integer method
+		JDCT_FLOAT: floating-point method
+		JDCT_DEFAULT: default method (normally JDCT_ISLOW)
+		JDCT_FASTEST: fastest method (normally JDCT_IFAST)
+	The FLOAT method is very slightly more accurate than the ISLOW method,
+	but may give different results on different machines due to varying
+	roundoff behavior.  The integer methods should give the same results
+	on all machines.  On machines with sufficiently fast FP hardware, the
+	floating-point method may also be the fastest.  The IFAST method is
+	considerably less accurate than the other two; its use is not
+	recommended if high quality is a concern.  JDCT_DEFAULT and
+	JDCT_FASTEST are macros configurable by each installation.
+
+J_COLOR_SPACE jpeg_color_space
+int num_components
+	The JPEG color space and corresponding number of components; see
+	"Special color spaces", below, for more info.  We recommend using
+	jpeg_set_color_space() if you want to change these.
+
+boolean optimize_coding
+	TRUE causes the compressor to compute optimal Huffman coding tables
+	for the image.  This requires an extra pass over the data and
+	therefore costs a good deal of space and time.  The default is
+	FALSE, which tells the compressor to use the supplied or default
+	Huffman tables.  In most cases optimal tables save only a few percent
+	of file size compared to the default tables.  Note that when this is
+	TRUE, you need not supply Huffman tables at all, and any you do
+	supply will be overwritten.
+
+unsigned int restart_interval
+int restart_in_rows
+	To emit restart markers in the JPEG file, set one of these nonzero.
+	Set restart_interval to specify the exact interval in MCU blocks.
+	Set restart_in_rows to specify the interval in MCU rows.  (If
+	restart_in_rows is not 0, then restart_interval is set after the
+	image width in MCUs is computed.)  Defaults are zero (no restarts).
+	One restart marker per MCU row is often a good choice.
+	NOTE: the overhead of restart markers is higher in grayscale JPEG
+	files than in color files, and MUCH higher in progressive JPEGs.
+	If you use restarts, you may want to use larger intervals in those
+	cases.
+
+const jpeg_scan_info * scan_info
+int num_scans
+	By default, scan_info is NULL; this causes the compressor to write a
+	single-scan sequential JPEG file.  If not NULL, scan_info points to
+	an array of scan definition records of length num_scans.  The
+	compressor will then write a JPEG file having one scan for each scan
+	definition record.  This is used to generate noninterleaved or
+	progressive JPEG files.  The library checks that the scan array
+	defines a valid JPEG scan sequence.  (jpeg_simple_progression creates
+	a suitable scan definition array for progressive JPEG.)  This is
+	discussed further under "Progressive JPEG support".
+
+int smoothing_factor
+	If non-zero, the input image is smoothed; the value should be 1 for
+	minimal smoothing to 100 for maximum smoothing.  Consult jcsample.c
+	for details of the smoothing algorithm.  The default is zero.
+
+boolean write_JFIF_header
+	If TRUE, a JFIF APP0 marker is emitted.  jpeg_set_defaults() and
+	jpeg_set_colorspace() set this TRUE if a JFIF-legal JPEG color space
+	(ie, YCbCr or grayscale) is selected, otherwise FALSE.
+
+UINT8 JFIF_major_version
+UINT8 JFIF_minor_version
+	The version number to be written into the JFIF marker.
+	jpeg_set_defaults() initializes the version to 1.01 (major=minor=1).
+	You should set it to 1.02 (major=1, minor=2) if you plan to write
+	any JFIF 1.02 extension markers.
+
+UINT8 density_unit
+UINT16 X_density
+UINT16 Y_density
+	The resolution information to be written into the JFIF marker;
+	not used otherwise.  density_unit may be 0 for unknown,
+	1 for dots/inch, or 2 for dots/cm.  The default values are 0,1,1
+	indicating square pixels of unknown size.
+
+boolean write_Adobe_marker
+	If TRUE, an Adobe APP14 marker is emitted.  jpeg_set_defaults() and
+	jpeg_set_colorspace() set this TRUE if JPEG color space RGB, CMYK,
+	or YCCK is selected, otherwise FALSE.  It is generally a bad idea
+	to set both write_JFIF_header and write_Adobe_marker.  In fact,
+	you probably shouldn't change the default settings at all --- the
+	default behavior ensures that the JPEG file's color space can be
+	recognized by the decoder.
+
+JQUANT_TBL * quant_tbl_ptrs[NUM_QUANT_TBLS]
+	Pointers to coefficient quantization tables, one per table slot,
+	or NULL if no table is defined for a slot.  Usually these should
+	be set via one of the above helper routines; jpeg_add_quant_table()
+	is general enough to define any quantization table.  The other
+	routines will set up table slot 0 for luminance quality and table
+	slot 1 for chrominance.
+
+int q_scale_factor[NUM_QUANT_TBLS]
+	[libjpeg v7+ only] Linear quantization scaling factors (0-100, default
+	100) for use with jpeg_default_qtables().
+	See rdswitch.c and cjpeg.c for an example of usage.
+	Note that the q_scale_factor[] values use "linear" scales, so JPEG
+	quality levels chosen by the user must be converted to these scales
+	using jpeg_quality_scaling().  Here is an example which corresponds to
+	cjpeg -quality 90,70:
+
+		jpeg_set_defaults(cinfo);
+
+		/* Set luminance quality 90. */
+		cinfo->q_scale_factor[0] = jpeg_quality_scaling(90);
+		/* Set chrominance quality 70. */
+		cinfo->q_scale_factor[1] = jpeg_quality_scaling(70);
+
+		jpeg_default_qtables(cinfo, force_baseline);
+
+	CAUTION: Setting separate quality levels for chrominance and luminance
+	is mainly only useful if chrominance subsampling is disabled.  2x2
+	chrominance subsampling (AKA "4:2:0") is the default, but you can
+	explicitly disable subsampling as follows:
+
+		cinfo->comp_info[0].v_samp_factor = 1;
+		cinfo->comp_info[0].h_samp_factor = 1;
+
+JHUFF_TBL * dc_huff_tbl_ptrs[NUM_HUFF_TBLS]
+JHUFF_TBL * ac_huff_tbl_ptrs[NUM_HUFF_TBLS]
+	Pointers to Huffman coding tables, one per table slot, or NULL if
+	no table is defined for a slot.  Slots 0 and 1 are filled with the
+	JPEG sample tables by jpeg_set_defaults().  If you need to allocate
+	more table structures, jpeg_alloc_huff_table() may be used.
+	Note that optimal Huffman tables can be computed for an image
+	by setting optimize_coding, as discussed above; there's seldom
+	any need to mess with providing your own Huffman tables.
+
+
+[libjpeg v7+ only] The actual dimensions of the JPEG image that will be written
+to the file are given by the following fields.  These are computed from the
+input image dimensions and the compression parameters by jpeg_start_compress().
+You can also call jpeg_calc_jpeg_dimensions() to obtain the values that will
+result from the current parameter settings.
+
+JDIMENSION jpeg_width		Actual dimensions of output image.
+JDIMENSION jpeg_height
+
+
+There are some additional cinfo fields which are not documented here
+because you currently can't change them; for example, you can't set
+arith_code TRUE because arithmetic coding is unsupported.
+
+
+Per-component parameters are stored in the struct cinfo.comp_info[i] for
+component number i.  Note that components here refer to components of the
+JPEG color space, *not* the source image color space.  A suitably large
+comp_info[] array is allocated by jpeg_set_defaults(); if you choose not
+to use that routine, it's up to you to allocate the array.
+
+int component_id
+	The one-byte identifier code to be recorded in the JPEG file for
+	this component.  For the standard color spaces, we recommend you
+	leave the default values alone.
+
+int h_samp_factor
+int v_samp_factor
+	Horizontal and vertical sampling factors for the component; must
+	be 1..4 according to the JPEG standard.  Note that larger sampling
+	factors indicate a higher-resolution component; many people find
+	this behavior quite unintuitive.  The default values are 2,2 for
+	luminance components and 1,1 for chrominance components, except
+	for grayscale where 1,1 is used.
+
+int quant_tbl_no
+	Quantization table number for component.  The default value is
+	0 for luminance components and 1 for chrominance components.
+
+int dc_tbl_no
+int ac_tbl_no
+	DC and AC entropy coding table numbers.  The default values are
+	0 for luminance components and 1 for chrominance components.
+
+int component_index
+	Must equal the component's index in comp_info[].  (Beginning in
+	release v6, the compressor library will fill this in automatically;
+	you don't have to.)
+
+
+Decompression parameter selection
+---------------------------------
+
+Decompression parameter selection is somewhat simpler than compression
+parameter selection, since all of the JPEG internal parameters are
+recorded in the source file and need not be supplied by the application.
+(Unless you are working with abbreviated files, in which case see
+"Abbreviated datastreams", below.)  Decompression parameters control
+the postprocessing done on the image to deliver it in a format suitable
+for the application's use.  Many of the parameters control speed/quality
+tradeoffs, in which faster decompression may be obtained at the price of
+a poorer-quality image.  The defaults select the highest quality (slowest)
+processing.
+
+The following fields in the JPEG object are set by jpeg_read_header() and
+may be useful to the application in choosing decompression parameters:
+
+JDIMENSION image_width			Width and height of image
+JDIMENSION image_height
+int num_components			Number of color components
+J_COLOR_SPACE jpeg_color_space		Colorspace of image
+boolean saw_JFIF_marker			TRUE if a JFIF APP0 marker was seen
+  UINT8 JFIF_major_version		Version information from JFIF marker
+  UINT8 JFIF_minor_version
+  UINT8 density_unit			Resolution data from JFIF marker
+  UINT16 X_density
+  UINT16 Y_density
+boolean saw_Adobe_marker		TRUE if an Adobe APP14 marker was seen
+  UINT8 Adobe_transform			Color transform code from Adobe marker
+
+The JPEG color space, unfortunately, is something of a guess since the JPEG
+standard proper does not provide a way to record it.  In practice most files
+adhere to the JFIF or Adobe conventions, and the decoder will recognize these
+correctly.  See "Special color spaces", below, for more info.
+
+
+The decompression parameters that determine the basic properties of the
+returned image are:
+
+J_COLOR_SPACE out_color_space
+	Output color space.  jpeg_read_header() sets an appropriate default
+	based on jpeg_color_space; typically it will be RGB or grayscale.
+	The application can change this field to request output in a different
+	colorspace.  For example, set it to JCS_GRAYSCALE to get grayscale
+	output from a color file.  (This is useful for previewing: grayscale
+	output is faster than full color since the color components need not
+	be processed.)  Note that not all possible color space transforms are
+	currently implemented; you may need to extend jdcolor.c if you want an
+	unusual conversion.
+
+unsigned int scale_num, scale_denom
+	Scale the image by the fraction scale_num/scale_denom.  Default is
+	1/1, or no scaling.  Currently, the only supported scaling ratios
+	are 1/1, 1/2, 1/4, and 1/8.  (The library design allows for arbitrary
+	scaling ratios but this is not likely to be implemented any time soon.)
+	Smaller scaling ratios permit significantly faster decoding since
+	fewer pixels need be processed and a simpler IDCT method can be used.
+
+boolean quantize_colors
+	If set TRUE, colormapped output will be delivered.  Default is FALSE,
+	meaning that full-color output will be delivered.
+
+The next three parameters are relevant only if quantize_colors is TRUE.
+
+int desired_number_of_colors
+	Maximum number of colors to use in generating a library-supplied color
+	map (the actual number of colors is returned in a different field).
+	Default 256.  Ignored when the application supplies its own color map.
+
+boolean two_pass_quantize
+	If TRUE, an extra pass over the image is made to select a custom color
+	map for the image.  This usually looks a lot better than the one-size-
+	fits-all colormap that is used otherwise.  Default is TRUE.  Ignored
+	when the application supplies its own color map.
+
+J_DITHER_MODE dither_mode
+	Selects color dithering method.  Supported values are:
+		JDITHER_NONE	no dithering: fast, very low quality
+		JDITHER_ORDERED	ordered dither: moderate speed and quality
+		JDITHER_FS	Floyd-Steinberg dither: slow, high quality
+	Default is JDITHER_FS.  (At present, ordered dither is implemented
+	only in the single-pass, standard-colormap case.  If you ask for
+	ordered dither when two_pass_quantize is TRUE or when you supply
+	an external color map, you'll get F-S dithering.)
+
+When quantize_colors is TRUE, the target color map is described by the next
+two fields.  colormap is set to NULL by jpeg_read_header().  The application
+can supply a color map by setting colormap non-NULL and setting
+actual_number_of_colors to the map size.  Otherwise, jpeg_start_decompress()
+selects a suitable color map and sets these two fields itself.
+[Implementation restriction: at present, an externally supplied colormap is
+only accepted for 3-component output color spaces.]
+
+JSAMPARRAY colormap
+	The color map, represented as a 2-D pixel array of out_color_components
+	rows and actual_number_of_colors columns.  Ignored if not quantizing.
+	CAUTION: if the JPEG library creates its own colormap, the storage
+	pointed to by this field is released by jpeg_finish_decompress().
+	Copy the colormap somewhere else first, if you want to save it.
+
+int actual_number_of_colors
+	The number of colors in the color map.
+
+Additional decompression parameters that the application may set include:
+
+J_DCT_METHOD dct_method
+	Selects the algorithm used for the DCT step.  Choices are the same
+	as described above for compression.
+
+boolean do_fancy_upsampling
+	If TRUE, do careful upsampling of chroma components.  If FALSE,
+	a faster but sloppier method is used.  Default is TRUE.  The visual
+	impact of the sloppier method is often very small.
+
+boolean do_block_smoothing
+	If TRUE, interblock smoothing is applied in early stages of decoding
+	progressive JPEG files; if FALSE, not.  Default is TRUE.  Early
+	progression stages look "fuzzy" with smoothing, "blocky" without.
+	In any case, block smoothing ceases to be applied after the first few
+	AC coefficients are known to full accuracy, so it is relevant only
+	when using buffered-image mode for progressive images.
+
+boolean enable_1pass_quant
+boolean enable_external_quant
+boolean enable_2pass_quant
+	These are significant only in buffered-image mode, which is
+	described in its own section below.
+
+
+The output image dimensions are given by the following fields.  These are
+computed from the source image dimensions and the decompression parameters
+by jpeg_start_decompress().  You can also call jpeg_calc_output_dimensions()
+to obtain the values that will result from the current parameter settings.
+This can be useful if you are trying to pick a scaling ratio that will get
+close to a desired target size.  It's also important if you are using the
+JPEG library's memory manager to allocate output buffer space, because you
+are supposed to request such buffers *before* jpeg_start_decompress().
+
+JDIMENSION output_width		Actual dimensions of output image.
+JDIMENSION output_height
+int out_color_components	Number of color components in out_color_space.
+int output_components		Number of color components returned.
+int rec_outbuf_height		Recommended height of scanline buffer.
+
+When quantizing colors, output_components is 1, indicating a single color map
+index per pixel.  Otherwise it equals out_color_components.  The output arrays
+are required to be output_width * output_components JSAMPLEs wide.
+
+rec_outbuf_height is the recommended minimum height (in scanlines) of the
+buffer passed to jpeg_read_scanlines().  If the buffer is smaller, the
+library will still work, but time will be wasted due to unnecessary data
+copying.  In high-quality modes, rec_outbuf_height is always 1, but some
+faster, lower-quality modes set it to larger values (typically 2 to 4).
+If you are going to ask for a high-speed processing mode, you may as well
+go to the trouble of honoring rec_outbuf_height so as to avoid data copying.
+(An output buffer larger than rec_outbuf_height lines is OK, but won't
+provide any material speed improvement over that height.)
+
+
+Special color spaces
+--------------------
+
+The JPEG standard itself is "color blind" and doesn't specify any particular
+color space.  It is customary to convert color data to a luminance/chrominance
+color space before compressing, since this permits greater compression.  The
+existing de-facto JPEG file format standards specify YCbCr or grayscale data
+(JFIF), or grayscale, RGB, YCbCr, CMYK, or YCCK (Adobe).  For special
+applications such as multispectral images, other color spaces can be used,
+but it must be understood that such files will be unportable.
+
+The JPEG library can handle the most common colorspace conversions (namely
+RGB <=> YCbCr and CMYK <=> YCCK).  It can also deal with data of an unknown
+color space, passing it through without conversion.  If you deal extensively
+with an unusual color space, you can easily extend the library to understand
+additional color spaces and perform appropriate conversions.
+
+For compression, the source data's color space is specified by field
+in_color_space.  This is transformed to the JPEG file's color space given
+by jpeg_color_space.  jpeg_set_defaults() chooses a reasonable JPEG color
+space depending on in_color_space, but you can override this by calling
+jpeg_set_colorspace().  Of course you must select a supported transformation.
+jccolor.c currently supports the following transformations:
+	RGB => YCbCr
+	RGB => GRAYSCALE
+	YCbCr => GRAYSCALE
+	CMYK => YCCK
+plus the null transforms: GRAYSCALE => GRAYSCALE, RGB => RGB,
+YCbCr => YCbCr, CMYK => CMYK, YCCK => YCCK, and UNKNOWN => UNKNOWN.
+
+The de-facto file format standards (JFIF and Adobe) specify APPn markers that
+indicate the color space of the JPEG file.  It is important to ensure that
+these are written correctly, or omitted if the JPEG file's color space is not
+one of the ones supported by the de-facto standards.  jpeg_set_colorspace()
+will set the compression parameters to include or omit the APPn markers
+properly, so long as it is told the truth about the JPEG color space.
+For example, if you are writing some random 3-component color space without
+conversion, don't try to fake out the library by setting in_color_space and
+jpeg_color_space to JCS_YCbCr; use JCS_UNKNOWN.  You may want to write an
+APPn marker of your own devising to identify the colorspace --- see "Special
+markers", below.
+
+When told that the color space is UNKNOWN, the library will default to using
+luminance-quality compression parameters for all color components.  You may
+well want to change these parameters.  See the source code for
+jpeg_set_colorspace(), in jcparam.c, for details.
+
+For decompression, the JPEG file's color space is given in jpeg_color_space,
+and this is transformed to the output color space out_color_space.
+jpeg_read_header's setting of jpeg_color_space can be relied on if the file
+conforms to JFIF or Adobe conventions, but otherwise it is no better than a
+guess.  If you know the JPEG file's color space for certain, you can override
+jpeg_read_header's guess by setting jpeg_color_space.  jpeg_read_header also
+selects a default output color space based on (its guess of) jpeg_color_space;
+set out_color_space to override this.  Again, you must select a supported
+transformation.  jdcolor.c currently supports
+	YCbCr => GRAYSCALE
+	YCbCr => RGB
+	GRAYSCALE => RGB
+	YCCK => CMYK
+as well as the null transforms.  (Since GRAYSCALE=>RGB is provided, an
+application can force grayscale JPEGs to look like color JPEGs if it only
+wants to handle one case.)
+
+The two-pass color quantizer, jquant2.c, is specialized to handle RGB data
+(it weights distances appropriately for RGB colors).  You'll need to modify
+the code if you want to use it for non-RGB output color spaces.  Note that
+jquant2.c is used to map to an application-supplied colormap as well as for
+the normal two-pass colormap selection process.
+
+CAUTION: it appears that Adobe Photoshop writes inverted data in CMYK JPEG
+files: 0 represents 100% ink coverage, rather than 0% ink as you'd expect.
+This is arguably a bug in Photoshop, but if you need to work with Photoshop
+CMYK files, you will have to deal with it in your application.  We cannot
+"fix" this in the library by inverting the data during the CMYK<=>YCCK
+transform, because that would break other applications, notably Ghostscript.
+Photoshop versions prior to 3.0 write EPS files containing JPEG-encoded CMYK
+data in the same inverted-YCCK representation used in bare JPEG files, but
+the surrounding PostScript code performs an inversion using the PS image
+operator.  I am told that Photoshop 3.0 will write uninverted YCCK in
+EPS/JPEG files, and will omit the PS-level inversion.  (But the data
+polarity used in bare JPEG files will not change in 3.0.)  In either case,
+the JPEG library must not invert the data itself, or else Ghostscript would
+read these EPS files incorrectly.
+
+
+Error handling
+--------------
+
+When the default error handler is used, any error detected inside the JPEG
+routines will cause a message to be printed on stderr, followed by exit().
+You can supply your own error handling routines to override this behavior
+and to control the treatment of nonfatal warnings and trace/debug messages.
+The file example.c illustrates the most common case, which is to have the
+application regain control after an error rather than exiting.
+
+The JPEG library never writes any message directly; it always goes through
+the error handling routines.  Three classes of messages are recognized:
+  * Fatal errors: the library cannot continue.
+  * Warnings: the library can continue, but the data is corrupt, and a
+    damaged output image is likely to result.
+  * Trace/informational messages.  These come with a trace level indicating
+    the importance of the message; you can control the verbosity of the
+    program by adjusting the maximum trace level that will be displayed.
+
+You may, if you wish, simply replace the entire JPEG error handling module
+(jerror.c) with your own code.  However, you can avoid code duplication by
+only replacing some of the routines depending on the behavior you need.
+This is accomplished by calling jpeg_std_error() as usual, but then overriding
+some of the method pointers in the jpeg_error_mgr struct, as illustrated by
+example.c.
+
+All of the error handling routines will receive a pointer to the JPEG object
+(a j_common_ptr which points to either a jpeg_compress_struct or a
+jpeg_decompress_struct; if you need to tell which, test the is_decompressor
+field).  This struct includes a pointer to the error manager struct in its
+"err" field.  Frequently, custom error handler routines will need to access
+additional data which is not known to the JPEG library or the standard error
+handler.  The most convenient way to do this is to embed either the JPEG
+object or the jpeg_error_mgr struct in a larger structure that contains
+additional fields; then casting the passed pointer provides access to the
+additional fields.  Again, see example.c for one way to do it.  (Beginning
+with IJG version 6b, there is also a void pointer "client_data" in each
+JPEG object, which the application can also use to find related data.
+The library does not touch client_data at all.)
+
+The individual methods that you might wish to override are:
+
+error_exit (j_common_ptr cinfo)
+	Receives control for a fatal error.  Information sufficient to
+	generate the error message has been stored in cinfo->err; call
+	output_message to display it.  Control must NOT return to the caller;
+	generally this routine will exit() or longjmp() somewhere.
+	Typically you would override this routine to get rid of the exit()
+	default behavior.  Note that if you continue processing, you should
+	clean up the JPEG object with jpeg_abort() or jpeg_destroy().
+
+output_message (j_common_ptr cinfo)
+	Actual output of any JPEG message.  Override this to send messages
+	somewhere other than stderr.  Note that this method does not know
+	how to generate a message, only where to send it.
+
+format_message (j_common_ptr cinfo, char * buffer)
+	Constructs a readable error message string based on the error info
+	stored in cinfo->err.  This method is called by output_message.  Few
+	applications should need to override this method.  One possible
+	reason for doing so is to implement dynamic switching of error message
+	language.
+
+emit_message (j_common_ptr cinfo, int msg_level)
+	Decide whether or not to emit a warning or trace message; if so,
+	calls output_message.  The main reason for overriding this method
+	would be to abort on warnings.  msg_level is -1 for warnings,
+	0 and up for trace messages.
+
+Only error_exit() and emit_message() are called from the rest of the JPEG
+library; the other two are internal to the error handler.
+
+The actual message texts are stored in an array of strings which is pointed to
+by the field err->jpeg_message_table.  The messages are numbered from 0 to
+err->last_jpeg_message, and it is these code numbers that are used in the
+JPEG library code.  You could replace the message texts (for instance, with
+messages in French or German) by changing the message table pointer.  See
+jerror.h for the default texts.  CAUTION: this table will almost certainly
+change or grow from one library version to the next.
+
+It may be useful for an application to add its own message texts that are
+handled by the same mechanism.  The error handler supports a second "add-on"
+message table for this purpose.  To define an addon table, set the pointer
+err->addon_message_table and the message numbers err->first_addon_message and
+err->last_addon_message.  If you number the addon messages beginning at 1000
+or so, you won't have to worry about conflicts with the library's built-in
+messages.  See the sample applications cjpeg/djpeg for an example of using
+addon messages (the addon messages are defined in cderror.h).
+
+Actual invocation of the error handler is done via macros defined in jerror.h:
+	ERREXITn(...)	for fatal errors
+	WARNMSn(...)	for corrupt-data warnings
+	TRACEMSn(...)	for trace and informational messages.
+These macros store the message code and any additional parameters into the
+error handler struct, then invoke the error_exit() or emit_message() method.
+The variants of each macro are for varying numbers of additional parameters.
+The additional parameters are inserted into the generated message using
+standard printf() format codes.
+
+See jerror.h and jerror.c for further details.
+
+
+Compressed data handling (source and destination managers)
+----------------------------------------------------------
+
+The JPEG compression library sends its compressed data to a "destination
+manager" module.  The default destination manager just writes the data to a
+memory buffer or to a stdio stream, but you can provide your own manager to
+do something else.  Similarly, the decompression library calls a "source
+manager" to obtain the compressed data; you can provide your own source
+manager if you want the data to come from somewhere other than a memory
+buffer or a stdio stream.
+
+In both cases, compressed data is processed a bufferload at a time: the
+destination or source manager provides a work buffer, and the library invokes
+the manager only when the buffer is filled or emptied.  (You could define a
+one-character buffer to force the manager to be invoked for each byte, but
+that would be rather inefficient.)  The buffer's size and location are
+controlled by the manager, not by the library.  For example, the memory
+source manager just makes the buffer pointer and length point to the original
+data in memory.  In this case the buffer-reload procedure will be invoked
+only if the decompressor ran off the end of the datastream, which would
+indicate an erroneous datastream.
+
+The work buffer is defined as an array of datatype JOCTET, which is generally
+"char" or "unsigned char".  On a machine where char is not exactly 8 bits
+wide, you must define JOCTET as a wider data type and then modify the data
+source and destination modules to transcribe the work arrays into 8-bit units
+on external storage.
+
+A data destination manager struct contains a pointer and count defining the
+next byte to write in the work buffer and the remaining free space:
+
+	JOCTET * next_output_byte;  /* => next byte to write in buffer */
+	size_t free_in_buffer;      /* # of byte spaces remaining in buffer */
+
+The library increments the pointer and decrements the count until the buffer
+is filled.  The manager's empty_output_buffer method must reset the pointer
+and count.  The manager is expected to remember the buffer's starting address
+and total size in private fields not visible to the library.
+
+A data destination manager provides three methods:
+
+init_destination (j_compress_ptr cinfo)
+	Initialize destination.  This is called by jpeg_start_compress()
+	before any data is actually written.  It must initialize
+	next_output_byte and free_in_buffer.  free_in_buffer must be
+	initialized to a positive value.
+
+empty_output_buffer (j_compress_ptr cinfo)
+	This is called whenever the buffer has filled (free_in_buffer
+	reaches zero).  In typical applications, it should write out the
+	*entire* buffer (use the saved start address and buffer length;
+	ignore the current state of next_output_byte and free_in_buffer).
+	Then reset the pointer & count to the start of the buffer, and
+	return TRUE indicating that the buffer has been dumped.
+	free_in_buffer must be set to a positive value when TRUE is
+	returned.  A FALSE return should only be used when I/O suspension is
+	desired (this operating mode is discussed in the next section).
+
+term_destination (j_compress_ptr cinfo)
+	Terminate destination --- called by jpeg_finish_compress() after all
+	data has been written.  In most applications, this must flush any
+	data remaining in the buffer.  Use either next_output_byte or
+	free_in_buffer to determine how much data is in the buffer.
+
+term_destination() is NOT called by jpeg_abort() or jpeg_destroy().  If you
+want the destination manager to be cleaned up during an abort, you must do it
+yourself.
+
+You will also need code to create a jpeg_destination_mgr struct, fill in its
+method pointers, and insert a pointer to the struct into the "dest" field of
+the JPEG compression object.  This can be done in-line in your setup code if
+you like, but it's probably cleaner to provide a separate routine similar to
+the jpeg_stdio_dest() or jpeg_mem_dest() routines of the supplied destination
+managers.
+
+Decompression source managers follow a parallel design, but with some
+additional frammishes.  The source manager struct contains a pointer and count
+defining the next byte to read from the work buffer and the number of bytes
+remaining:
+
+	const JOCTET * next_input_byte; /* => next byte to read from buffer */
+	size_t bytes_in_buffer;         /* # of bytes remaining in buffer */
+
+The library increments the pointer and decrements the count until the buffer
+is emptied.  The manager's fill_input_buffer method must reset the pointer and
+count.  In most applications, the manager must remember the buffer's starting
+address and total size in private fields not visible to the library.
+
+A data source manager provides five methods:
+
+init_source (j_decompress_ptr cinfo)
+	Initialize source.  This is called by jpeg_read_header() before any
+	data is actually read.  Unlike init_destination(), it may leave
+	bytes_in_buffer set to 0 (in which case a fill_input_buffer() call
+	will occur immediately).
+
+fill_input_buffer (j_decompress_ptr cinfo)
+	This is called whenever bytes_in_buffer has reached zero and more
+	data is wanted.  In typical applications, it should read fresh data
+	into the buffer (ignoring the current state of next_input_byte and
+	bytes_in_buffer), reset the pointer & count to the start of the
+	buffer, and return TRUE indicating that the buffer has been reloaded.
+	It is not necessary to fill the buffer entirely, only to obtain at
+	least one more byte.  bytes_in_buffer MUST be set to a positive value
+	if TRUE is returned.  A FALSE return should only be used when I/O
+	suspension is desired (this mode is discussed in the next section).
+
+skip_input_data (j_decompress_ptr cinfo, long num_bytes)
+	Skip num_bytes worth of data.  The buffer pointer and count should
+	be advanced over num_bytes input bytes, refilling the buffer as
+	needed.  This is used to skip over a potentially large amount of
+	uninteresting data (such as an APPn marker).  In some applications
+	it may be possible to optimize away the reading of the skipped data,
+	but it's not clear that being smart is worth much trouble; large
+	skips are uncommon.  bytes_in_buffer may be zero on return.
+	A zero or negative skip count should be treated as a no-op.
+
+resync_to_restart (j_decompress_ptr cinfo, int desired)
+	This routine is called only when the decompressor has failed to find
+	a restart (RSTn) marker where one is expected.  Its mission is to
+	find a suitable point for resuming decompression.  For most
+	applications, we recommend that you just use the default resync
+	procedure, jpeg_resync_to_restart().  However, if you are able to back
+	up in the input data stream, or if you have a-priori knowledge about
+	the likely location of restart markers, you may be able to do better.
+	Read the read_restart_marker() and jpeg_resync_to_restart() routines
+	in jdmarker.c if you think you'd like to implement your own resync
+	procedure.
+
+term_source (j_decompress_ptr cinfo)
+	Terminate source --- called by jpeg_finish_decompress() after all
+	data has been read.  Often a no-op.
+
+For both fill_input_buffer() and skip_input_data(), there is no such thing
+as an EOF return.  If the end of the file has been reached, the routine has
+a choice of exiting via ERREXIT() or inserting fake data into the buffer.
+In most cases, generating a warning message and inserting a fake EOI marker
+is the best course of action --- this will allow the decompressor to output
+however much of the image is there.  In pathological cases, the decompressor
+may swallow the EOI and again demand data ... just keep feeding it fake EOIs.
+jdatasrc.c illustrates the recommended error recovery behavior.
+
+term_source() is NOT called by jpeg_abort() or jpeg_destroy().  If you want
+the source manager to be cleaned up during an abort, you must do it yourself.
+
+You will also need code to create a jpeg_source_mgr struct, fill in its method
+pointers, and insert a pointer to the struct into the "src" field of the JPEG
+decompression object.  This can be done in-line in your setup code if you
+like, but it's probably cleaner to provide a separate routine similar to the
+jpeg_stdio_src() or jpeg_mem_src() routines of the supplied source managers.
+
+For more information, consult the memory and stdio source and destination
+managers in jdatasrc.c and jdatadst.c.
+
+
+I/O suspension
+--------------
+
+Some applications need to use the JPEG library as an incremental memory-to-
+memory filter: when the compressed data buffer is filled or emptied, they want
+control to return to the outer loop, rather than expecting that the buffer can
+be emptied or reloaded within the data source/destination manager subroutine.
+The library supports this need by providing an "I/O suspension" mode, which we
+describe in this section.
+
+The I/O suspension mode is not a panacea: nothing is guaranteed about the
+maximum amount of time spent in any one call to the library, so it will not
+eliminate response-time problems in single-threaded applications.  If you
+need guaranteed response time, we suggest you "bite the bullet" and implement
+a real multi-tasking capability.
+
+To use I/O suspension, cooperation is needed between the calling application
+and the data source or destination manager; you will always need a custom
+source/destination manager.  (Please read the previous section if you haven't
+already.)  The basic idea is that the empty_output_buffer() or
+fill_input_buffer() routine is a no-op, merely returning FALSE to indicate
+that it has done nothing.  Upon seeing this, the JPEG library suspends
+operation and returns to its caller.  The surrounding application is
+responsible for emptying or refilling the work buffer before calling the
+JPEG library again.
+
+Compression suspension:
+
+For compression suspension, use an empty_output_buffer() routine that returns
+FALSE; typically it will not do anything else.  This will cause the
+compressor to return to the caller of jpeg_write_scanlines(), with the return
+value indicating that not all the supplied scanlines have been accepted.
+The application must make more room in the output buffer, adjust the output
+buffer pointer/count appropriately, and then call jpeg_write_scanlines()
+again, pointing to the first unconsumed scanline.
+
+When forced to suspend, the compressor will backtrack to a convenient stopping
+point (usually the start of the current MCU); it will regenerate some output
+data when restarted.  Therefore, although empty_output_buffer() is only
+called when the buffer is filled, you should NOT write out the entire buffer
+after a suspension.  Write only the data up to the current position of
+next_output_byte/free_in_buffer.  The data beyond that point will be
+regenerated after resumption.
+
+Because of the backtracking behavior, a good-size output buffer is essential
+for efficiency; you don't want the compressor to suspend often.  (In fact, an
+overly small buffer could lead to infinite looping, if a single MCU required
+more data than would fit in the buffer.)  We recommend a buffer of at least
+several Kbytes.  You may want to insert explicit code to ensure that you don't
+call jpeg_write_scanlines() unless there is a reasonable amount of space in
+the output buffer; in other words, flush the buffer before trying to compress
+more data.
+
+The compressor does not allow suspension while it is trying to write JPEG
+markers at the beginning and end of the file.  This means that:
+  * At the beginning of a compression operation, there must be enough free
+    space in the output buffer to hold the header markers (typically 600 or
+    so bytes).  The recommended buffer size is bigger than this anyway, so
+    this is not a problem as long as you start with an empty buffer.  However,
+    this restriction might catch you if you insert large special markers, such
+    as a JFIF thumbnail image, without flushing the buffer afterwards.
+  * When you call jpeg_finish_compress(), there must be enough space in the
+    output buffer to emit any buffered data and the final EOI marker.  In the
+    current implementation, half a dozen bytes should suffice for this, but
+    for safety's sake we recommend ensuring that at least 100 bytes are free
+    before calling jpeg_finish_compress().
+
+A more significant restriction is that jpeg_finish_compress() cannot suspend.
+This means you cannot use suspension with multi-pass operating modes, namely
+Huffman code optimization and multiple-scan output.  Those modes write the
+whole file during jpeg_finish_compress(), which will certainly result in
+buffer overrun.  (Note that this restriction applies only to compression,
+not decompression.  The decompressor supports input suspension in all of its
+operating modes.)
+
+Decompression suspension:
+
+For decompression suspension, use a fill_input_buffer() routine that simply
+returns FALSE (except perhaps during error recovery, as discussed below).
+This will cause the decompressor to return to its caller with an indication
+that suspension has occurred.  This can happen at four places:
+  * jpeg_read_header(): will return JPEG_SUSPENDED.
+  * jpeg_start_decompress(): will return FALSE, rather than its usual TRUE.
+  * jpeg_read_scanlines(): will return the number of scanlines already
+	completed (possibly 0).
+  * jpeg_finish_decompress(): will return FALSE, rather than its usual TRUE.
+The surrounding application must recognize these cases, load more data into
+the input buffer, and repeat the call.  In the case of jpeg_read_scanlines(),
+increment the passed pointers past any scanlines successfully read.
+
+Just as with compression, the decompressor will typically backtrack to a
+convenient restart point before suspending.  When fill_input_buffer() is
+called, next_input_byte/bytes_in_buffer point to the current restart point,
+which is where the decompressor will backtrack to if FALSE is returned.
+The data beyond that position must NOT be discarded if you suspend; it needs
+to be re-read upon resumption.  In most implementations, you'll need to shift
+this data down to the start of your work buffer and then load more data after
+it.  Again, this behavior means that a several-Kbyte work buffer is essential
+for decent performance; furthermore, you should load a reasonable amount of
+new data before resuming decompression.  (If you loaded, say, only one new
+byte each time around, you could waste a LOT of cycles.)
+
+The skip_input_data() source manager routine requires special care in a
+suspension scenario.  This routine is NOT granted the ability to suspend the
+decompressor; it can decrement bytes_in_buffer to zero, but no more.  If the
+requested skip distance exceeds the amount of data currently in the input
+buffer, then skip_input_data() must set bytes_in_buffer to zero and record the
+additional skip distance somewhere else.  The decompressor will immediately
+call fill_input_buffer(), which should return FALSE, which will cause a
+suspension return.  The surrounding application must then arrange to discard
+the recorded number of bytes before it resumes loading the input buffer.
+(Yes, this design is rather baroque, but it avoids complexity in the far more
+common case where a non-suspending source manager is used.)
+
+If the input data has been exhausted, we recommend that you emit a warning
+and insert dummy EOI markers just as a non-suspending data source manager
+would do.  This can be handled either in the surrounding application logic or
+within fill_input_buffer(); the latter is probably more efficient.  If
+fill_input_buffer() knows that no more data is available, it can set the
+pointer/count to point to a dummy EOI marker and then return TRUE just as
+though it had read more data in a non-suspending situation.
+
+The decompressor does not attempt to suspend within standard JPEG markers;
+instead it will backtrack to the start of the marker and reprocess the whole
+marker next time.  Hence the input buffer must be large enough to hold the
+longest standard marker in the file.  Standard JPEG markers should normally
+not exceed a few hundred bytes each (DHT tables are typically the longest).
+We recommend at least a 2K buffer for performance reasons, which is much
+larger than any correct marker is likely to be.  For robustness against
+damaged marker length counts, you may wish to insert a test in your
+application for the case that the input buffer is completely full and yet
+the decoder has suspended without consuming any data --- otherwise, if this
+situation did occur, it would lead to an endless loop.  (The library can't
+provide this test since it has no idea whether "the buffer is full", or
+even whether there is a fixed-size input buffer.)
+
+The input buffer would need to be 64K to allow for arbitrary COM or APPn
+markers, but these are handled specially: they are either saved into allocated
+memory, or skipped over by calling skip_input_data().  In the former case,
+suspension is handled correctly, and in the latter case, the problem of
+buffer overrun is placed on skip_input_data's shoulders, as explained above.
+Note that if you provide your own marker handling routine for large markers,
+you should consider how to deal with buffer overflow.
+
+Multiple-buffer management:
+
+In some applications it is desirable to store the compressed data in a linked
+list of buffer areas, so as to avoid data copying.  This can be handled by
+having empty_output_buffer() or fill_input_buffer() set the pointer and count
+to reference the next available buffer; FALSE is returned only if no more
+buffers are available.  Although seemingly straightforward, there is a
+pitfall in this approach: the backtrack that occurs when FALSE is returned
+could back up into an earlier buffer.  For example, when fill_input_buffer()
+is called, the current pointer & count indicate the backtrack restart point.
+Since fill_input_buffer() will set the pointer and count to refer to a new
+buffer, the restart position must be saved somewhere else.  Suppose a second
+call to fill_input_buffer() occurs in the same library call, and no
+additional input data is available, so fill_input_buffer must return FALSE.
+If the JPEG library has not moved the pointer/count forward in the current
+buffer, then *the correct restart point is the saved position in the prior
+buffer*.  Prior buffers may be discarded only after the library establishes
+a restart point within a later buffer.  Similar remarks apply for output into
+a chain of buffers.
+
+The library will never attempt to backtrack over a skip_input_data() call,
+so any skipped data can be permanently discarded.  You still have to deal
+with the case of skipping not-yet-received data, however.
+
+It's much simpler to use only a single buffer; when fill_input_buffer() is
+called, move any unconsumed data (beyond the current pointer/count) down to
+the beginning of this buffer and then load new data into the remaining buffer
+space.  This approach requires a little more data copying but is far easier
+to get right.
+
+
+Progressive JPEG support
+------------------------
+
+Progressive JPEG rearranges the stored data into a series of scans of
+increasing quality.  In situations where a JPEG file is transmitted across a
+slow communications link, a decoder can generate a low-quality image very
+quickly from the first scan, then gradually improve the displayed quality as
+more scans are received.  The final image after all scans are complete is
+identical to that of a regular (sequential) JPEG file of the same quality
+setting.  Progressive JPEG files are often slightly smaller than equivalent
+sequential JPEG files, but the possibility of incremental display is the main
+reason for using progressive JPEG.
+
+The IJG encoder library generates progressive JPEG files when given a
+suitable "scan script" defining how to divide the data into scans.
+Creation of progressive JPEG files is otherwise transparent to the encoder.
+Progressive JPEG files can also be read transparently by the decoder library.
+If the decoding application simply uses the library as defined above, it
+will receive a final decoded image without any indication that the file was
+progressive.  Of course, this approach does not allow incremental display.
+To perform incremental display, an application needs to use the decoder
+library's "buffered-image" mode, in which it receives a decoded image
+multiple times.
+
+Each displayed scan requires about as much work to decode as a full JPEG
+image of the same size, so the decoder must be fairly fast in relation to the
+data transmission rate in order to make incremental display useful.  However,
+it is possible to skip displaying the image and simply add the incoming bits
+to the decoder's coefficient buffer.  This is fast because only Huffman
+decoding need be done, not IDCT, upsampling, colorspace conversion, etc.
+The IJG decoder library allows the application to switch dynamically between
+displaying the image and simply absorbing the incoming bits.  A properly
+coded application can automatically adapt the number of display passes to
+suit the time available as the image is received.  Also, a final
+higher-quality display cycle can be performed from the buffered data after
+the end of the file is reached.
+
+Progressive compression:
+
+To create a progressive JPEG file (or a multiple-scan sequential JPEG file),
+set the scan_info cinfo field to point to an array of scan descriptors, and
+perform compression as usual.  Instead of constructing your own scan list,
+you can call the jpeg_simple_progression() helper routine to create a
+recommended progression sequence; this method should be used by all
+applications that don't want to get involved in the nitty-gritty of
+progressive scan sequence design.  (If you want to provide user control of
+scan sequences, you may wish to borrow the scan script reading code found
+in rdswitch.c, so that you can read scan script files just like cjpeg's.)
+When scan_info is not NULL, the compression library will store DCT'd data
+into a buffer array as jpeg_write_scanlines() is called, and will emit all
+the requested scans during jpeg_finish_compress().  This implies that
+multiple-scan output cannot be created with a suspending data destination
+manager, since jpeg_finish_compress() does not support suspension.  We
+should also note that the compressor currently forces Huffman optimization
+mode when creating a progressive JPEG file, because the default Huffman
+tables are unsuitable for progressive files.
+
+Progressive decompression:
+
+When buffered-image mode is not used, the decoder library will read all of
+a multi-scan file during jpeg_start_decompress(), so that it can provide a
+final decoded image.  (Here "multi-scan" means either progressive or
+multi-scan sequential.)  This makes multi-scan files transparent to the
+decoding application.  However, existing applications that used suspending
+input with version 5 of the IJG library will need to be modified to check
+for a suspension return from jpeg_start_decompress().
+
+To perform incremental display, an application must use the library's
+buffered-image mode.  This is described in the next section.
+
+
+Buffered-image mode
+-------------------
+
+In buffered-image mode, the library stores the partially decoded image in a
+coefficient buffer, from which it can be read out as many times as desired.
+This mode is typically used for incremental display of progressive JPEG files,
+but it can be used with any JPEG file.  Each scan of a progressive JPEG file
+adds more data (more detail) to the buffered image.  The application can
+display in lockstep with the source file (one display pass per input scan),
+or it can allow input processing to outrun display processing.  By making
+input and display processing run independently, it is possible for the
+application to adapt progressive display to a wide range of data transmission
+rates.
+
+The basic control flow for buffered-image decoding is
+
+	jpeg_create_decompress()
+	set data source
+	jpeg_read_header()
+	set overall decompression parameters
+	cinfo.buffered_image = TRUE;	/* select buffered-image mode */
+	jpeg_start_decompress()
+	for (each output pass) {
+	    adjust output decompression parameters if required
+	    jpeg_start_output()		/* start a new output pass */
+	    for (all scanlines in image) {
+	        jpeg_read_scanlines()
+	        display scanlines
+	    }
+	    jpeg_finish_output()	/* terminate output pass */
+	}
+	jpeg_finish_decompress()
+	jpeg_destroy_decompress()
+
+This differs from ordinary unbuffered decoding in that there is an additional
+level of looping.  The application can choose how many output passes to make
+and how to display each pass.
+
+The simplest approach to displaying progressive images is to do one display
+pass for each scan appearing in the input file.  In this case the outer loop
+condition is typically
+	while (! jpeg_input_complete(&cinfo))
+and the start-output call should read
+	jpeg_start_output(&cinfo, cinfo.input_scan_number);
+The second parameter to jpeg_start_output() indicates which scan of the input
+file is to be displayed; the scans are numbered starting at 1 for this
+purpose.  (You can use a loop counter starting at 1 if you like, but using
+the library's input scan counter is easier.)  The library automatically reads
+data as necessary to complete each requested scan, and jpeg_finish_output()
+advances to the next scan or end-of-image marker (hence input_scan_number
+will be incremented by the time control arrives back at jpeg_start_output()).
+With this technique, data is read from the input file only as needed, and
+input and output processing run in lockstep.
+
+After reading the final scan and reaching the end of the input file, the
+buffered image remains available; it can be read additional times by
+repeating the jpeg_start_output()/jpeg_read_scanlines()/jpeg_finish_output()
+sequence.  For example, a useful technique is to use fast one-pass color
+quantization for display passes made while the image is arriving, followed by
+a final display pass using two-pass quantization for highest quality.  This
+is done by changing the library parameters before the final output pass.
+Changing parameters between passes is discussed in detail below.
+
+In general the last scan of a progressive file cannot be recognized as such
+until after it is read, so a post-input display pass is the best approach if
+you want special processing in the final pass.
+
+When done with the image, be sure to call jpeg_finish_decompress() to release
+the buffered image (or just use jpeg_destroy_decompress()).
+
+If input data arrives faster than it can be displayed, the application can
+cause the library to decode input data in advance of what's needed to produce
+output.  This is done by calling the routine jpeg_consume_input().
+The return value is one of the following:
+	JPEG_REACHED_SOS:    reached an SOS marker (the start of a new scan)
+	JPEG_REACHED_EOI:    reached the EOI marker (end of image)
+	JPEG_ROW_COMPLETED:  completed reading one MCU row of compressed data
+	JPEG_SCAN_COMPLETED: completed reading last MCU row of current scan
+	JPEG_SUSPENDED:      suspended before completing any of the above
+(JPEG_SUSPENDED can occur only if a suspending data source is used.)  This
+routine can be called at any time after initializing the JPEG object.  It
+reads some additional data and returns when one of the indicated significant
+events occurs.  (If called after the EOI marker is reached, it will
+immediately return JPEG_REACHED_EOI without attempting to read more data.)
+
+The library's output processing will automatically call jpeg_consume_input()
+whenever the output processing overtakes the input; thus, simple lockstep
+display requires no direct calls to jpeg_consume_input().  But by adding
+calls to jpeg_consume_input(), you can absorb data in advance of what is
+being displayed.  This has two benefits:
+  * You can limit buildup of unprocessed data in your input buffer.
+  * You can eliminate extra display passes by paying attention to the
+    state of the library's input processing.
+
+The first of these benefits only requires interspersing calls to
+jpeg_consume_input() with your display operations and any other processing
+you may be doing.  To avoid wasting cycles due to backtracking, it's best to
+call jpeg_consume_input() only after a hundred or so new bytes have arrived.
+This is discussed further under "I/O suspension", above.  (Note: the JPEG
+library currently is not thread-safe.  You must not call jpeg_consume_input()
+from one thread of control if a different library routine is working on the
+same JPEG object in another thread.)
+
+When input arrives fast enough that more than one new scan is available
+before you start a new output pass, you may as well skip the output pass
+corresponding to the completed scan.  This occurs for free if you pass
+cinfo.input_scan_number as the target scan number to jpeg_start_output().
+The input_scan_number field is simply the index of the scan currently being
+consumed by the input processor.  You can ensure that this is up-to-date by
+emptying the input buffer just before calling jpeg_start_output(): call
+jpeg_consume_input() repeatedly until it returns JPEG_SUSPENDED or
+JPEG_REACHED_EOI.
+
+The target scan number passed to jpeg_start_output() is saved in the
+cinfo.output_scan_number field.  The library's output processing calls
+jpeg_consume_input() whenever the current input scan number and row within
+that scan is less than or equal to the current output scan number and row.
+Thus, input processing can "get ahead" of the output processing but is not
+allowed to "fall behind".  You can achieve several different effects by
+manipulating this interlock rule.  For example, if you pass a target scan
+number greater than the current input scan number, the output processor will
+wait until that scan starts to arrive before producing any output.  (To avoid
+an infinite loop, the target scan number is automatically reset to the last
+scan number when the end of image is reached.  Thus, if you specify a large
+target scan number, the library will just absorb the entire input file and
+then perform an output pass.  This is effectively the same as what
+jpeg_start_decompress() does when you don't select buffered-image mode.)
+When you pass a target scan number equal to the current input scan number,
+the image is displayed no faster than the current input scan arrives.  The
+final possibility is to pass a target scan number less than the current input
+scan number; this disables the input/output interlock and causes the output
+processor to simply display whatever it finds in the image buffer, without
+waiting for input.  (However, the library will not accept a target scan
+number less than one, so you can't avoid waiting for the first scan.)
+
+When data is arriving faster than the output display processing can advance
+through the image, jpeg_consume_input() will store data into the buffered
+image beyond the point at which the output processing is reading data out
+again.  If the input arrives fast enough, it may "wrap around" the buffer to
+the point where the input is more than one whole scan ahead of the output.
+If the output processing simply proceeds through its display pass without
+paying attention to the input, the effect seen on-screen is that the lower
+part of the image is one or more scans better in quality than the upper part.
+Then, when the next output scan is started, you have a choice of what target
+scan number to use.  The recommended choice is to use the current input scan
+number at that time, which implies that you've skipped the output scans
+corresponding to the input scans that were completed while you processed the
+previous output scan.  In this way, the decoder automatically adapts its
+speed to the arriving data, by skipping output scans as necessary to keep up
+with the arriving data.
+
+When using this strategy, you'll want to be sure that you perform a final
+output pass after receiving all the data; otherwise your last display may not
+be full quality across the whole screen.  So the right outer loop logic is
+something like this:
+	do {
+	    absorb any waiting input by calling jpeg_consume_input()
+	    final_pass = jpeg_input_complete(&cinfo);
+	    adjust output decompression parameters if required
+	    jpeg_start_output(&cinfo, cinfo.input_scan_number);
+	    ...
+	    jpeg_finish_output()
+	} while (! final_pass);
+rather than quitting as soon as jpeg_input_complete() returns TRUE.  This
+arrangement makes it simple to use higher-quality decoding parameters
+for the final pass.  But if you don't want to use special parameters for
+the final pass, the right loop logic is like this:
+	for (;;) {
+	    absorb any waiting input by calling jpeg_consume_input()
+	    jpeg_start_output(&cinfo, cinfo.input_scan_number);
+	    ...
+	    jpeg_finish_output()
+	    if (jpeg_input_complete(&cinfo) &&
+	        cinfo.input_scan_number == cinfo.output_scan_number)
+	      break;
+	}
+In this case you don't need to know in advance whether an output pass is to
+be the last one, so it's not necessary to have reached EOF before starting
+the final output pass; rather, what you want to test is whether the output
+pass was performed in sync with the final input scan.  This form of the loop
+will avoid an extra output pass whenever the decoder is able (or nearly able)
+to keep up with the incoming data.
+
+When the data transmission speed is high, you might begin a display pass,
+then find that much or all of the file has arrived before you can complete
+the pass.  (You can detect this by noting the JPEG_REACHED_EOI return code
+from jpeg_consume_input(), or equivalently by testing jpeg_input_complete().)
+In this situation you may wish to abort the current display pass and start a
+new one using the newly arrived information.  To do so, just call
+jpeg_finish_output() and then start a new pass with jpeg_start_output().
+
+A variant strategy is to abort and restart display if more than one complete
+scan arrives during an output pass; this can be detected by noting
+JPEG_REACHED_SOS returns and/or examining cinfo.input_scan_number.  This
+idea should be employed with caution, however, since the display process
+might never get to the bottom of the image before being aborted, resulting
+in the lower part of the screen being several passes worse than the upper.
+In most cases it's probably best to abort an output pass only if the whole
+file has arrived and you want to begin the final output pass immediately.
+
+When receiving data across a communication link, we recommend always using
+the current input scan number for the output target scan number; if a
+higher-quality final pass is to be done, it should be started (aborting any
+incomplete output pass) as soon as the end of file is received.  However,
+many other strategies are possible.  For example, the application can examine
+the parameters of the current input scan and decide whether to display it or
+not.  If the scan contains only chroma data, one might choose not to use it
+as the target scan, expecting that the scan will be small and will arrive
+quickly.  To skip to the next scan, call jpeg_consume_input() until it
+returns JPEG_REACHED_SOS or JPEG_REACHED_EOI.  Or just use the next higher
+number as the target scan for jpeg_start_output(); but that method doesn't
+let you inspect the next scan's parameters before deciding to display it.
+
+
+In buffered-image mode, jpeg_start_decompress() never performs input and
+thus never suspends.  An application that uses input suspension with
+buffered-image mode must be prepared for suspension returns from these
+routines:
+* jpeg_start_output() performs input only if you request 2-pass quantization
+  and the target scan isn't fully read yet.  (This is discussed below.)
+* jpeg_read_scanlines(), as always, returns the number of scanlines that it
+  was able to produce before suspending.
+* jpeg_finish_output() will read any markers following the target scan,
+  up to the end of the file or the SOS marker that begins another scan.
+  (But it reads no input if jpeg_consume_input() has already reached the
+  end of the file or a SOS marker beyond the target output scan.)
+* jpeg_finish_decompress() will read until the end of file, and thus can
+  suspend if the end hasn't already been reached (as can be tested by
+  calling jpeg_input_complete()).
+jpeg_start_output(), jpeg_finish_output(), and jpeg_finish_decompress()
+all return TRUE if they completed their tasks, FALSE if they had to suspend.
+In the event of a FALSE return, the application must load more input data
+and repeat the call.  Applications that use non-suspending data sources need
+not check the return values of these three routines.
+
+
+It is possible to change decoding parameters between output passes in the
+buffered-image mode.  The decoder library currently supports only very
+limited changes of parameters.  ONLY THE FOLLOWING parameter changes are
+allowed after jpeg_start_decompress() is called:
+* dct_method can be changed before each call to jpeg_start_output().
+  For example, one could use a fast DCT method for early scans, changing
+  to a higher quality method for the final scan.
+* dither_mode can be changed before each call to jpeg_start_output();
+  of course this has no impact if not using color quantization.  Typically
+  one would use ordered dither for initial passes, then switch to
+  Floyd-Steinberg dither for the final pass.  Caution: changing dither mode
+  can cause more memory to be allocated by the library.  Although the amount
+  of memory involved is not large (a scanline or so), it may cause the
+  initial max_memory_to_use specification to be exceeded, which in the worst
+  case would result in an out-of-memory failure.
+* do_block_smoothing can be changed before each call to jpeg_start_output().
+  This setting is relevant only when decoding a progressive JPEG image.
+  During the first DC-only scan, block smoothing provides a very "fuzzy" look
+  instead of the very "blocky" look seen without it; which is better seems a
+  matter of personal taste.  But block smoothing is nearly always a win
+  during later stages, especially when decoding a successive-approximation
+  image: smoothing helps to hide the slight blockiness that otherwise shows
+  up on smooth gradients until the lowest coefficient bits are sent.
+* Color quantization mode can be changed under the rules described below.
+  You *cannot* change between full-color and quantized output (because that
+  would alter the required I/O buffer sizes), but you can change which
+  quantization method is used.
+
+When generating color-quantized output, changing quantization method is a
+very useful way of switching between high-speed and high-quality display.
+The library allows you to change among its three quantization methods:
+1. Single-pass quantization to a fixed color cube.
+   Selected by cinfo.two_pass_quantize = FALSE and cinfo.colormap = NULL.
+2. Single-pass quantization to an application-supplied colormap.
+   Selected by setting cinfo.colormap to point to the colormap (the value of
+   two_pass_quantize is ignored); also set cinfo.actual_number_of_colors.
+3. Two-pass quantization to a colormap chosen specifically for the image.
+   Selected by cinfo.two_pass_quantize = TRUE and cinfo.colormap = NULL.
+   (This is the default setting selected by jpeg_read_header, but it is
+   probably NOT what you want for the first pass of progressive display!)
+These methods offer successively better quality and lesser speed.  However,
+only the first method is available for quantizing in non-RGB color spaces.
+
+IMPORTANT: because the different quantizer methods have very different
+working-storage requirements, the library requires you to indicate which
+one(s) you intend to use before you call jpeg_start_decompress().  (If we did
+not require this, the max_memory_to_use setting would be a complete fiction.)
+You do this by setting one or more of these three cinfo fields to TRUE:
+	enable_1pass_quant		Fixed color cube colormap
+	enable_external_quant		Externally-supplied colormap
+	enable_2pass_quant		Two-pass custom colormap
+All three are initialized FALSE by jpeg_read_header().  But
+jpeg_start_decompress() automatically sets TRUE the one selected by the
+current two_pass_quantize and colormap settings, so you only need to set the
+enable flags for any other quantization methods you plan to change to later.
+
+After setting the enable flags correctly at jpeg_start_decompress() time, you
+can change to any enabled quantization method by setting two_pass_quantize
+and colormap properly just before calling jpeg_start_output().  The following
+special rules apply:
+1. You must explicitly set cinfo.colormap to NULL when switching to 1-pass
+   or 2-pass mode from a different mode, or when you want the 2-pass
+   quantizer to be re-run to generate a new colormap.
+2. To switch to an external colormap, or to change to a different external
+   colormap than was used on the prior pass, you must call
+   jpeg_new_colormap() after setting cinfo.colormap.
+NOTE: if you want to use the same colormap as was used in the prior pass,
+you should not do either of these things.  This will save some nontrivial
+switchover costs.
+(These requirements exist because cinfo.colormap will always be non-NULL
+after completing a prior output pass, since both the 1-pass and 2-pass
+quantizers set it to point to their output colormaps.  Thus you have to
+do one of these two things to notify the library that something has changed.
+Yup, it's a bit klugy, but it's necessary to do it this way for backwards
+compatibility.)
+
+Note that in buffered-image mode, the library generates any requested colormap
+during jpeg_start_output(), not during jpeg_start_decompress().
+
+When using two-pass quantization, jpeg_start_output() makes a pass over the
+buffered image to determine the optimum color map; it therefore may take a
+significant amount of time, whereas ordinarily it does little work.  The
+progress monitor hook is called during this pass, if defined.  It is also
+important to realize that if the specified target scan number is greater than
+or equal to the current input scan number, jpeg_start_output() will attempt
+to consume input as it makes this pass.  If you use a suspending data source,
+you need to check for a FALSE return from jpeg_start_output() under these
+conditions.  The combination of 2-pass quantization and a not-yet-fully-read
+target scan is the only case in which jpeg_start_output() will consume input.
+
+
+Application authors who support buffered-image mode may be tempted to use it
+for all JPEG images, even single-scan ones.  This will work, but it is
+inefficient: there is no need to create an image-sized coefficient buffer for
+single-scan images.  Requesting buffered-image mode for such an image wastes
+memory.  Worse, it can cost time on large images, since the buffered data has
+to be swapped out or written to a temporary file.  If you are concerned about
+maximum performance on baseline JPEG files, you should use buffered-image
+mode only when the incoming file actually has multiple scans.  This can be
+tested by calling jpeg_has_multiple_scans(), which will return a correct
+result at any time after jpeg_read_header() completes.
+
+It is also worth noting that when you use jpeg_consume_input() to let input
+processing get ahead of output processing, the resulting pattern of access to
+the coefficient buffer is quite nonsequential.  It's best to use the memory
+manager jmemnobs.c if you can (ie, if you have enough real or virtual main
+memory).  If not, at least make sure that max_memory_to_use is set as high as
+possible.  If the JPEG memory manager has to use a temporary file, you will
+probably see a lot of disk traffic and poor performance.  (This could be
+improved with additional work on the memory manager, but we haven't gotten
+around to it yet.)
+
+In some applications it may be convenient to use jpeg_consume_input() for all
+input processing, including reading the initial markers; that is, you may
+wish to call jpeg_consume_input() instead of jpeg_read_header() during
+startup.  This works, but note that you must check for JPEG_REACHED_SOS and
+JPEG_REACHED_EOI return codes as the equivalent of jpeg_read_header's codes.
+Once the first SOS marker has been reached, you must call
+jpeg_start_decompress() before jpeg_consume_input() will consume more input;
+it'll just keep returning JPEG_REACHED_SOS until you do.  If you read a
+tables-only file this way, jpeg_consume_input() will return JPEG_REACHED_EOI
+without ever returning JPEG_REACHED_SOS; be sure to check for this case.
+If this happens, the decompressor will not read any more input until you call
+jpeg_abort() to reset it.  It is OK to call jpeg_consume_input() even when not
+using buffered-image mode, but in that case it's basically a no-op after the
+initial markers have been read: it will just return JPEG_SUSPENDED.
+
+
+Abbreviated datastreams and multiple images
+-------------------------------------------
+
+A JPEG compression or decompression object can be reused to process multiple
+images.  This saves a small amount of time per image by eliminating the
+"create" and "destroy" operations, but that isn't the real purpose of the
+feature.  Rather, reuse of an object provides support for abbreviated JPEG
+datastreams.  Object reuse can also simplify processing a series of images in
+a single input or output file.  This section explains these features.
+
+A JPEG file normally contains several hundred bytes worth of quantization
+and Huffman tables.  In a situation where many images will be stored or
+transmitted with identical tables, this may represent an annoying overhead.
+The JPEG standard therefore permits tables to be omitted.  The standard
+defines three classes of JPEG datastreams:
+  * "Interchange" datastreams contain an image and all tables needed to decode
+     the image.  These are the usual kind of JPEG file.
+  * "Abbreviated image" datastreams contain an image, but are missing some or
+    all of the tables needed to decode that image.
+  * "Abbreviated table specification" (henceforth "tables-only") datastreams
+    contain only table specifications.
+To decode an abbreviated image, it is necessary to load the missing table(s)
+into the decoder beforehand.  This can be accomplished by reading a separate
+tables-only file.  A variant scheme uses a series of images in which the first
+image is an interchange (complete) datastream, while subsequent ones are
+abbreviated and rely on the tables loaded by the first image.  It is assumed
+that once the decoder has read a table, it will remember that table until a
+new definition for the same table number is encountered.
+
+It is the application designer's responsibility to figure out how to associate
+the correct tables with an abbreviated image.  While abbreviated datastreams
+can be useful in a closed environment, their use is strongly discouraged in
+any situation where data exchange with other applications might be needed.
+Caveat designer.
+
+The JPEG library provides support for reading and writing any combination of
+tables-only datastreams and abbreviated images.  In both compression and
+decompression objects, a quantization or Huffman table will be retained for
+the lifetime of the object, unless it is overwritten by a new table definition.
+
+
+To create abbreviated image datastreams, it is only necessary to tell the
+compressor not to emit some or all of the tables it is using.  Each
+quantization and Huffman table struct contains a boolean field "sent_table",
+which normally is initialized to FALSE.  For each table used by the image, the
+header-writing process emits the table and sets sent_table = TRUE unless it is
+already TRUE.  (In normal usage, this prevents outputting the same table
+definition multiple times, as would otherwise occur because the chroma
+components typically share tables.)  Thus, setting this field to TRUE before
+calling jpeg_start_compress() will prevent the table from being written at
+all.
+
+If you want to create a "pure" abbreviated image file containing no tables,
+just call "jpeg_suppress_tables(&cinfo, TRUE)" after constructing all the
+tables.  If you want to emit some but not all tables, you'll need to set the
+individual sent_table fields directly.
+
+To create an abbreviated image, you must also call jpeg_start_compress()
+with a second parameter of FALSE, not TRUE.  Otherwise jpeg_start_compress()
+will force all the sent_table fields to FALSE.  (This is a safety feature to
+prevent abbreviated images from being created accidentally.)
+
+To create a tables-only file, perform the same parameter setup that you
+normally would, but instead of calling jpeg_start_compress() and so on, call
+jpeg_write_tables(&cinfo).  This will write an abbreviated datastream
+containing only SOI, DQT and/or DHT markers, and EOI.  All the quantization
+and Huffman tables that are currently defined in the compression object will
+be emitted unless their sent_tables flag is already TRUE, and then all the
+sent_tables flags will be set TRUE.
+
+A sure-fire way to create matching tables-only and abbreviated image files
+is to proceed as follows:
+
+	create JPEG compression object
+	set JPEG parameters
+	set destination to tables-only file
+	jpeg_write_tables(&cinfo);
+	set destination to image file
+	jpeg_start_compress(&cinfo, FALSE);
+	write data...
+	jpeg_finish_compress(&cinfo);
+
+Since the JPEG parameters are not altered between writing the table file and
+the abbreviated image file, the same tables are sure to be used.  Of course,
+you can repeat the jpeg_start_compress() ... jpeg_finish_compress() sequence
+many times to produce many abbreviated image files matching the table file.
+
+You cannot suppress output of the computed Huffman tables when Huffman
+optimization is selected.  (If you could, there'd be no way to decode the
+image...)  Generally, you don't want to set optimize_coding = TRUE when
+you are trying to produce abbreviated files.
+
+In some cases you might want to compress an image using tables which are
+not stored in the application, but are defined in an interchange or
+tables-only file readable by the application.  This can be done by setting up
+a JPEG decompression object to read the specification file, then copying the
+tables into your compression object.  See jpeg_copy_critical_parameters()
+for an example of copying quantization tables.
+
+
+To read abbreviated image files, you simply need to load the proper tables
+into the decompression object before trying to read the abbreviated image.
+If the proper tables are stored in the application program, you can just
+allocate the table structs and fill in their contents directly.  For example,
+to load a fixed quantization table into table slot "n":
+
+    if (cinfo.quant_tbl_ptrs[n] == NULL)
+      cinfo.quant_tbl_ptrs[n] = jpeg_alloc_quant_table((j_common_ptr) &cinfo);
+    quant_ptr = cinfo.quant_tbl_ptrs[n];	/* quant_ptr is JQUANT_TBL* */
+    for (i = 0; i < 64; i++) {
+      /* Qtable[] is desired quantization table, in natural array order */
+      quant_ptr->quantval[i] = Qtable[i];
+    }
+
+Code to load a fixed Huffman table is typically (for AC table "n"):
+
+    if (cinfo.ac_huff_tbl_ptrs[n] == NULL)
+      cinfo.ac_huff_tbl_ptrs[n] = jpeg_alloc_huff_table((j_common_ptr) &cinfo);
+    huff_ptr = cinfo.ac_huff_tbl_ptrs[n];	/* huff_ptr is JHUFF_TBL* */
+    for (i = 1; i <= 16; i++) {
+      /* counts[i] is number of Huffman codes of length i bits, i=1..16 */
+      huff_ptr->bits[i] = counts[i];
+    }
+    for (i = 0; i < 256; i++) {
+      /* symbols[] is the list of Huffman symbols, in code-length order */
+      huff_ptr->huffval[i] = symbols[i];
+    }
+
+(Note that trying to set cinfo.quant_tbl_ptrs[n] to point directly at a
+constant JQUANT_TBL object is not safe.  If the incoming file happened to
+contain a quantization table definition, your master table would get
+overwritten!  Instead allocate a working table copy and copy the master table
+into it, as illustrated above.  Ditto for Huffman tables, of course.)
+
+You might want to read the tables from a tables-only file, rather than
+hard-wiring them into your application.  The jpeg_read_header() call is
+sufficient to read a tables-only file.  You must pass a second parameter of
+FALSE to indicate that you do not require an image to be present.  Thus, the
+typical scenario is
+
+	create JPEG decompression object
+	set source to tables-only file
+	jpeg_read_header(&cinfo, FALSE);
+	set source to abbreviated image file
+	jpeg_read_header(&cinfo, TRUE);
+	set decompression parameters
+	jpeg_start_decompress(&cinfo);
+	read data...
+	jpeg_finish_decompress(&cinfo);
+
+In some cases, you may want to read a file without knowing whether it contains
+an image or just tables.  In that case, pass FALSE and check the return value
+from jpeg_read_header(): it will be JPEG_HEADER_OK if an image was found,
+JPEG_HEADER_TABLES_ONLY if only tables were found.  (A third return value,
+JPEG_SUSPENDED, is possible when using a suspending data source manager.)
+Note that jpeg_read_header() will not complain if you read an abbreviated
+image for which you haven't loaded the missing tables; the missing-table check
+occurs later, in jpeg_start_decompress().
+
+
+It is possible to read a series of images from a single source file by
+repeating the jpeg_read_header() ... jpeg_finish_decompress() sequence,
+without releasing/recreating the JPEG object or the data source module.
+(If you did reinitialize, any partial bufferload left in the data source
+buffer at the end of one image would be discarded, causing you to lose the
+start of the next image.)  When you use this method, stored tables are
+automatically carried forward, so some of the images can be abbreviated images
+that depend on tables from earlier images.
+
+If you intend to write a series of images into a single destination file,
+you might want to make a specialized data destination module that doesn't
+flush the output buffer at term_destination() time.  This would speed things
+up by some trifling amount.  Of course, you'd need to remember to flush the
+buffer after the last image.  You can make the later images be abbreviated
+ones by passing FALSE to jpeg_start_compress().
+
+
+Special markers
+---------------
+
+Some applications may need to insert or extract special data in the JPEG
+datastream.  The JPEG standard provides marker types "COM" (comment) and
+"APP0" through "APP15" (application) to hold application-specific data.
+Unfortunately, the use of these markers is not specified by the standard.
+COM markers are fairly widely used to hold user-supplied text.  The JFIF file
+format spec uses APP0 markers with specified initial strings to hold certain
+data.  Adobe applications use APP14 markers beginning with the string "Adobe"
+for miscellaneous data.  Other APPn markers are rarely seen, but might
+contain almost anything.
+
+If you wish to store user-supplied text, we recommend you use COM markers
+and place readable 7-bit ASCII text in them.  Newline conventions are not
+standardized --- expect to find LF (Unix style), CR/LF (DOS style), or CR
+(Mac style).  A robust COM reader should be able to cope with random binary
+garbage, including nulls, since some applications generate COM markers
+containing non-ASCII junk.  (But yours should not be one of them.)
+
+For program-supplied data, use an APPn marker, and be sure to begin it with an
+identifying string so that you can tell whether the marker is actually yours.
+It's probably best to avoid using APP0 or APP14 for any private markers.
+(NOTE: the upcoming SPIFF standard will use APP8 markers; we recommend you
+not use APP8 markers for any private purposes, either.)
+
+Keep in mind that at most 65533 bytes can be put into one marker, but you
+can have as many markers as you like.
+
+By default, the IJG compression library will write a JFIF APP0 marker if the
+selected JPEG colorspace is grayscale or YCbCr, or an Adobe APP14 marker if
+the selected colorspace is RGB, CMYK, or YCCK.  You can disable this, but
+we don't recommend it.  The decompression library will recognize JFIF and
+Adobe markers and will set the JPEG colorspace properly when one is found.
+
+
+You can write special markers immediately following the datastream header by
+calling jpeg_write_marker() after jpeg_start_compress() and before the first
+call to jpeg_write_scanlines().  When you do this, the markers appear after
+the SOI and the JFIF APP0 and Adobe APP14 markers (if written), but before
+all else.  Specify the marker type parameter as "JPEG_COM" for COM or
+"JPEG_APP0 + n" for APPn.  (Actually, jpeg_write_marker will let you write
+any marker type, but we don't recommend writing any other kinds of marker.)
+For example, to write a user comment string pointed to by comment_text:
+	jpeg_write_marker(cinfo, JPEG_COM, comment_text, strlen(comment_text));
+
+If it's not convenient to store all the marker data in memory at once,
+you can instead call jpeg_write_m_header() followed by multiple calls to
+jpeg_write_m_byte().  If you do it this way, it's your responsibility to
+call jpeg_write_m_byte() exactly the number of times given in the length
+parameter to jpeg_write_m_header().  (This method lets you empty the
+output buffer partway through a marker, which might be important when
+using a suspending data destination module.  In any case, if you are using
+a suspending destination, you should flush its buffer after inserting
+any special markers.  See "I/O suspension".)
+
+Or, if you prefer to synthesize the marker byte sequence yourself,
+you can just cram it straight into the data destination module.
+
+If you are writing JFIF 1.02 extension markers (thumbnail images), don't
+forget to set cinfo.JFIF_minor_version = 2 so that the encoder will write the
+correct JFIF version number in the JFIF header marker.  The library's default
+is to write version 1.01, but that's wrong if you insert any 1.02 extension
+markers.  (We could probably get away with just defaulting to 1.02, but there
+used to be broken decoders that would complain about unknown minor version
+numbers.  To reduce compatibility risks it's safest not to write 1.02 unless
+you are actually using 1.02 extensions.)
+
+
+When reading, two methods of handling special markers are available:
+1. You can ask the library to save the contents of COM and/or APPn markers
+into memory, and then examine them at your leisure afterwards.
+2. You can supply your own routine to process COM and/or APPn markers
+on-the-fly as they are read.
+The first method is simpler to use, especially if you are using a suspending
+data source; writing a marker processor that copes with input suspension is
+not easy (consider what happens if the marker is longer than your available
+input buffer).  However, the second method conserves memory since the marker
+data need not be kept around after it's been processed.
+
+For either method, you'd normally set up marker handling after creating a
+decompression object and before calling jpeg_read_header(), because the
+markers of interest will typically be near the head of the file and so will
+be scanned by jpeg_read_header.  Once you've established a marker handling
+method, it will be used for the life of that decompression object
+(potentially many datastreams), unless you change it.  Marker handling is
+determined separately for COM markers and for each APPn marker code.
+
+
+To save the contents of special markers in memory, call
+	jpeg_save_markers(cinfo, marker_code, length_limit)
+where marker_code is the marker type to save, JPEG_COM or JPEG_APP0+n.
+(To arrange to save all the special marker types, you need to call this
+routine 17 times, for COM and APP0-APP15.)  If the incoming marker is longer
+than length_limit data bytes, only length_limit bytes will be saved; this
+parameter allows you to avoid chewing up memory when you only need to see the
+first few bytes of a potentially large marker.  If you want to save all the
+data, set length_limit to 0xFFFF; that is enough since marker lengths are only
+16 bits.  As a special case, setting length_limit to 0 prevents that marker
+type from being saved at all.  (That is the default behavior, in fact.)
+
+After jpeg_read_header() completes, you can examine the special markers by
+following the cinfo->marker_list pointer chain.  All the special markers in
+the file appear in this list, in order of their occurrence in the file (but
+omitting any markers of types you didn't ask for).  Both the original data
+length and the saved data length are recorded for each list entry; the latter
+will not exceed length_limit for the particular marker type.  Note that these
+lengths exclude the marker length word, whereas the stored representation
+within the JPEG file includes it.  (Hence the maximum data length is really
+only 65533.)
+
+It is possible that additional special markers appear in the file beyond the
+SOS marker at which jpeg_read_header stops; if so, the marker list will be
+extended during reading of the rest of the file.  This is not expected to be
+common, however.  If you are short on memory you may want to reset the length
+limit to zero for all marker types after finishing jpeg_read_header, to
+ensure that the max_memory_to_use setting cannot be exceeded due to addition
+of later markers.
+
+The marker list remains stored until you call jpeg_finish_decompress or
+jpeg_abort, at which point the memory is freed and the list is set to empty.
+(jpeg_destroy also releases the storage, of course.)
+
+Note that the library is internally interested in APP0 and APP14 markers;
+if you try to set a small nonzero length limit on these types, the library
+will silently force the length up to the minimum it wants.  (But you can set
+a zero length limit to prevent them from being saved at all.)  Also, in a
+16-bit environment, the maximum length limit may be constrained to less than
+65533 by malloc() limitations.  It is therefore best not to assume that the
+effective length limit is exactly what you set it to be.
+
+
+If you want to supply your own marker-reading routine, you do it by calling
+jpeg_set_marker_processor().  A marker processor routine must have the
+signature
+	boolean jpeg_marker_parser_method (j_decompress_ptr cinfo)
+Although the marker code is not explicitly passed, the routine can find it
+in cinfo->unread_marker.  At the time of call, the marker proper has been
+read from the data source module.  The processor routine is responsible for
+reading the marker length word and the remaining parameter bytes, if any.
+Return TRUE to indicate success.  (FALSE should be returned only if you are
+using a suspending data source and it tells you to suspend.  See the standard
+marker processors in jdmarker.c for appropriate coding methods if you need to
+use a suspending data source.)
+
+If you override the default APP0 or APP14 processors, it is up to you to
+recognize JFIF and Adobe markers if you want colorspace recognition to occur
+properly.  We recommend copying and extending the default processors if you
+want to do that.  (A better idea is to save these marker types for later
+examination by calling jpeg_save_markers(); that method doesn't interfere
+with the library's own processing of these markers.)
+
+jpeg_set_marker_processor() and jpeg_save_markers() are mutually exclusive
+--- if you call one it overrides any previous call to the other, for the
+particular marker type specified.
+
+A simple example of an external COM processor can be found in djpeg.c.
+Also, see jpegtran.c for an example of using jpeg_save_markers.
+
+
+Raw (downsampled) image data
+----------------------------
+
+Some applications need to supply already-downsampled image data to the JPEG
+compressor, or to receive raw downsampled data from the decompressor.  The
+library supports this requirement by allowing the application to write or
+read raw data, bypassing the normal preprocessing or postprocessing steps.
+The interface is different from the standard one and is somewhat harder to
+use.  If your interest is merely in bypassing color conversion, we recommend
+that you use the standard interface and simply set jpeg_color_space =
+in_color_space (or jpeg_color_space = out_color_space for decompression).
+The mechanism described in this section is necessary only to supply or
+receive downsampled image data, in which not all components have the same
+dimensions.
+
+
+To compress raw data, you must supply the data in the colorspace to be used
+in the JPEG file (please read the earlier section on Special color spaces)
+and downsampled to the sampling factors specified in the JPEG parameters.
+You must supply the data in the format used internally by the JPEG library,
+namely a JSAMPIMAGE array.  This is an array of pointers to two-dimensional
+arrays, each of type JSAMPARRAY.  Each 2-D array holds the values for one
+color component.  This structure is necessary since the components are of
+different sizes.  If the image dimensions are not a multiple of the MCU size,
+you must also pad the data correctly (usually, this is done by replicating
+the last column and/or row).  The data must be padded to a multiple of a DCT
+block in each component: that is, each downsampled row must contain a
+multiple of 8 valid samples, and there must be a multiple of 8 sample rows
+for each component.  (For applications such as conversion of digital TV
+images, the standard image size is usually a multiple of the DCT block size,
+so that no padding need actually be done.)
+
+The procedure for compression of raw data is basically the same as normal
+compression, except that you call jpeg_write_raw_data() in place of
+jpeg_write_scanlines().  Before calling jpeg_start_compress(), you must do
+the following:
+  * Set cinfo->raw_data_in to TRUE.  (It is set FALSE by jpeg_set_defaults().)
+    This notifies the library that you will be supplying raw data.
+  * Ensure jpeg_color_space is correct --- an explicit jpeg_set_colorspace()
+    call is a good idea.  Note that since color conversion is bypassed,
+    in_color_space is ignored, except that jpeg_set_defaults() uses it to
+    choose the default jpeg_color_space setting.
+  * Ensure the sampling factors, cinfo->comp_info[i].h_samp_factor and
+    cinfo->comp_info[i].v_samp_factor, are correct.  Since these indicate the
+    dimensions of the data you are supplying, it's wise to set them
+    explicitly, rather than assuming the library's defaults are what you want.
+
+To pass raw data to the library, call jpeg_write_raw_data() in place of
+jpeg_write_scanlines().  The two routines work similarly except that
+jpeg_write_raw_data takes a JSAMPIMAGE data array rather than JSAMPARRAY.
+The scanlines count passed to and returned from jpeg_write_raw_data is
+measured in terms of the component with the largest v_samp_factor.
+
+jpeg_write_raw_data() processes one MCU row per call, which is to say
+v_samp_factor*DCTSIZE sample rows of each component.  The passed num_lines
+value must be at least max_v_samp_factor*DCTSIZE, and the return value will
+be exactly that amount (or possibly some multiple of that amount, in future
+library versions).  This is true even on the last call at the bottom of the
+image; don't forget to pad your data as necessary.
+
+The required dimensions of the supplied data can be computed for each
+component as
+	cinfo->comp_info[i].width_in_blocks*DCTSIZE  samples per row
+	cinfo->comp_info[i].height_in_blocks*DCTSIZE rows in image
+after jpeg_start_compress() has initialized those fields.  If the valid data
+is smaller than this, it must be padded appropriately.  For some sampling
+factors and image sizes, additional dummy DCT blocks are inserted to make
+the image a multiple of the MCU dimensions.  The library creates such dummy
+blocks itself; it does not read them from your supplied data.  Therefore you
+need never pad by more than DCTSIZE samples.  An example may help here.
+Assume 2h2v downsampling of YCbCr data, that is
+	cinfo->comp_info[0].h_samp_factor = 2		for Y
+	cinfo->comp_info[0].v_samp_factor = 2
+	cinfo->comp_info[1].h_samp_factor = 1		for Cb
+	cinfo->comp_info[1].v_samp_factor = 1
+	cinfo->comp_info[2].h_samp_factor = 1		for Cr
+	cinfo->comp_info[2].v_samp_factor = 1
+and suppose that the nominal image dimensions (cinfo->image_width and
+cinfo->image_height) are 101x101 pixels.  Then jpeg_start_compress() will
+compute downsampled_width = 101 and width_in_blocks = 13 for Y,
+downsampled_width = 51 and width_in_blocks = 7 for Cb and Cr (and the same
+for the height fields).  You must pad the Y data to at least 13*8 = 104
+columns and rows, the Cb/Cr data to at least 7*8 = 56 columns and rows.  The
+MCU height is max_v_samp_factor = 2 DCT rows so you must pass at least 16
+scanlines on each call to jpeg_write_raw_data(), which is to say 16 actual
+sample rows of Y and 8 each of Cb and Cr.  A total of 7 MCU rows are needed,
+so you must pass a total of 7*16 = 112 "scanlines".  The last DCT block row
+of Y data is dummy, so it doesn't matter what you pass for it in the data
+arrays, but the scanlines count must total up to 112 so that all of the Cb
+and Cr data gets passed.
+
+Output suspension is supported with raw-data compression: if the data
+destination module suspends, jpeg_write_raw_data() will return 0.
+In this case the same data rows must be passed again on the next call.
+
+
+Decompression with raw data output implies bypassing all postprocessing:
+you cannot ask for rescaling or color quantization, for instance.  More
+seriously, you must deal with the color space and sampling factors present in
+the incoming file.  If your application only handles, say, 2h1v YCbCr data,
+you must check for and fail on other color spaces or other sampling factors.
+The library will not convert to a different color space for you.
+
+To obtain raw data output, set cinfo->raw_data_out = TRUE before
+jpeg_start_decompress() (it is set FALSE by jpeg_read_header()).  Be sure to
+verify that the color space and sampling factors are ones you can handle.
+Then call jpeg_read_raw_data() in place of jpeg_read_scanlines().  The
+decompression process is otherwise the same as usual.
+
+jpeg_read_raw_data() returns one MCU row per call, and thus you must pass a
+buffer of at least max_v_samp_factor*DCTSIZE scanlines (scanline counting is
+the same as for raw-data compression).  The buffer you pass must be large
+enough to hold the actual data plus padding to DCT-block boundaries.  As with
+compression, any entirely dummy DCT blocks are not processed so you need not
+allocate space for them, but the total scanline count includes them.  The
+above example of computing buffer dimensions for raw-data compression is
+equally valid for decompression.
+
+Input suspension is supported with raw-data decompression: if the data source
+module suspends, jpeg_read_raw_data() will return 0.  You can also use
+buffered-image mode to read raw data in multiple passes.
+
+
+Really raw data: DCT coefficients
+---------------------------------
+
+It is possible to read or write the contents of a JPEG file as raw DCT
+coefficients.  This facility is mainly intended for use in lossless
+transcoding between different JPEG file formats.  Other possible applications
+include lossless cropping of a JPEG image, lossless reassembly of a
+multi-strip or multi-tile TIFF/JPEG file into a single JPEG datastream, etc.
+
+To read the contents of a JPEG file as DCT coefficients, open the file and do
+jpeg_read_header() as usual.  But instead of calling jpeg_start_decompress()
+and jpeg_read_scanlines(), call jpeg_read_coefficients().  This will read the
+entire image into a set of virtual coefficient-block arrays, one array per
+component.  The return value is a pointer to an array of virtual-array
+descriptors.  Each virtual array can be accessed directly using the JPEG
+memory manager's access_virt_barray method (see Memory management, below,
+and also read structure.txt's discussion of virtual array handling).  Or,
+for simple transcoding to a different JPEG file format, the array list can
+just be handed directly to jpeg_write_coefficients().
+
+Each block in the block arrays contains quantized coefficient values in
+normal array order (not JPEG zigzag order).  The block arrays contain only
+DCT blocks containing real data; any entirely-dummy blocks added to fill out
+interleaved MCUs at the right or bottom edges of the image are discarded
+during reading and are not stored in the block arrays.  (The size of each
+block array can be determined from the width_in_blocks and height_in_blocks
+fields of the component's comp_info entry.)  This is also the data format
+expected by jpeg_write_coefficients().
+
+When you are done using the virtual arrays, call jpeg_finish_decompress()
+to release the array storage and return the decompression object to an idle
+state; or just call jpeg_destroy() if you don't need to reuse the object.
+
+If you use a suspending data source, jpeg_read_coefficients() will return
+NULL if it is forced to suspend; a non-NULL return value indicates successful
+completion.  You need not test for a NULL return value when using a
+non-suspending data source.
+
+It is also possible to call jpeg_read_coefficients() to obtain access to the
+decoder's coefficient arrays during a normal decode cycle in buffered-image
+mode.  This frammish might be useful for progressively displaying an incoming
+image and then re-encoding it without loss.  To do this, decode in buffered-
+image mode as discussed previously, then call jpeg_read_coefficients() after
+the last jpeg_finish_output() call.  The arrays will be available for your use
+until you call jpeg_finish_decompress().
+
+
+To write the contents of a JPEG file as DCT coefficients, you must provide
+the DCT coefficients stored in virtual block arrays.  You can either pass
+block arrays read from an input JPEG file by jpeg_read_coefficients(), or
+allocate virtual arrays from the JPEG compression object and fill them
+yourself.  In either case, jpeg_write_coefficients() is substituted for
+jpeg_start_compress() and jpeg_write_scanlines().  Thus the sequence is
+  * Create compression object
+  * Set all compression parameters as necessary
+  * Request virtual arrays if needed
+  * jpeg_write_coefficients()
+  * jpeg_finish_compress()
+  * Destroy or re-use compression object
+jpeg_write_coefficients() is passed a pointer to an array of virtual block
+array descriptors; the number of arrays is equal to cinfo.num_components.
+
+The virtual arrays need only have been requested, not realized, before
+jpeg_write_coefficients() is called.  A side-effect of
+jpeg_write_coefficients() is to realize any virtual arrays that have been
+requested from the compression object's memory manager.  Thus, when obtaining
+the virtual arrays from the compression object, you should fill the arrays
+after calling jpeg_write_coefficients().  The data is actually written out
+when you call jpeg_finish_compress(); jpeg_write_coefficients() only writes
+the file header.
+
+When writing raw DCT coefficients, it is crucial that the JPEG quantization
+tables and sampling factors match the way the data was encoded, or the
+resulting file will be invalid.  For transcoding from an existing JPEG file,
+we recommend using jpeg_copy_critical_parameters().  This routine initializes
+all the compression parameters to default values (like jpeg_set_defaults()),
+then copies the critical information from a source decompression object.
+The decompression object should have just been used to read the entire
+JPEG input file --- that is, it should be awaiting jpeg_finish_decompress().
+
+jpeg_write_coefficients() marks all tables stored in the compression object
+as needing to be written to the output file (thus, it acts like
+jpeg_start_compress(cinfo, TRUE)).  This is for safety's sake, to avoid
+emitting abbreviated JPEG files by accident.  If you really want to emit an
+abbreviated JPEG file, call jpeg_suppress_tables(), or set the tables'
+individual sent_table flags, between calling jpeg_write_coefficients() and
+jpeg_finish_compress().
+
+
+Progress monitoring
+-------------------
+
+Some applications may need to regain control from the JPEG library every so
+often.  The typical use of this feature is to produce a percent-done bar or
+other progress display.  (For a simple example, see cjpeg.c or djpeg.c.)
+Although you do get control back frequently during the data-transferring pass
+(the jpeg_read_scanlines or jpeg_write_scanlines loop), any additional passes
+will occur inside jpeg_finish_compress or jpeg_start_decompress; those
+routines may take a long time to execute, and you don't get control back
+until they are done.
+
+You can define a progress-monitor routine which will be called periodically
+by the library.  No guarantees are made about how often this call will occur,
+so we don't recommend you use it for mouse tracking or anything like that.
+At present, a call will occur once per MCU row, scanline, or sample row
+group, whichever unit is convenient for the current processing mode; so the
+wider the image, the longer the time between calls.  During the data
+transferring pass, only one call occurs per call of jpeg_read_scanlines or
+jpeg_write_scanlines, so don't pass a large number of scanlines at once if
+you want fine resolution in the progress count.  (If you really need to use
+the callback mechanism for time-critical tasks like mouse tracking, you could
+insert additional calls inside some of the library's inner loops.)
+
+To establish a progress-monitor callback, create a struct jpeg_progress_mgr,
+fill in its progress_monitor field with a pointer to your callback routine,
+and set cinfo->progress to point to the struct.  The callback will be called
+whenever cinfo->progress is non-NULL.  (This pointer is set to NULL by
+jpeg_create_compress or jpeg_create_decompress; the library will not change
+it thereafter.  So if you allocate dynamic storage for the progress struct,
+make sure it will live as long as the JPEG object does.  Allocating from the
+JPEG memory manager with lifetime JPOOL_PERMANENT will work nicely.)  You
+can use the same callback routine for both compression and decompression.
+
+The jpeg_progress_mgr struct contains four fields which are set by the library:
+	long pass_counter;	/* work units completed in this pass */
+	long pass_limit;	/* total number of work units in this pass */
+	int completed_passes;	/* passes completed so far */
+	int total_passes;	/* total number of passes expected */
+During any one pass, pass_counter increases from 0 up to (not including)
+pass_limit; the step size is usually but not necessarily 1.  The pass_limit
+value may change from one pass to another.  The expected total number of
+passes is in total_passes, and the number of passes already completed is in
+completed_passes.  Thus the fraction of work completed may be estimated as
+		completed_passes + (pass_counter/pass_limit)
+		--------------------------------------------
+				total_passes
+ignoring the fact that the passes may not be equal amounts of work.
+
+When decompressing, pass_limit can even change within a pass, because it
+depends on the number of scans in the JPEG file, which isn't always known in
+advance.  The computed fraction-of-work-done may jump suddenly (if the library
+discovers it has overestimated the number of scans) or even decrease (in the
+opposite case).  It is not wise to put great faith in the work estimate.
+
+When using the decompressor's buffered-image mode, the progress monitor work
+estimate is likely to be completely unhelpful, because the library has no way
+to know how many output passes will be demanded of it.  Currently, the library
+sets total_passes based on the assumption that there will be one more output
+pass if the input file end hasn't yet been read (jpeg_input_complete() isn't
+TRUE), but no more output passes if the file end has been reached when the
+output pass is started.  This means that total_passes will rise as additional
+output passes are requested.  If you have a way of determining the input file
+size, estimating progress based on the fraction of the file that's been read
+will probably be more useful than using the library's value.
+
+
+Memory management
+-----------------
+
+This section covers some key facts about the JPEG library's built-in memory
+manager.  For more info, please read structure.txt's section about the memory
+manager, and consult the source code if necessary.
+
+All memory and temporary file allocation within the library is done via the
+memory manager.  If necessary, you can replace the "back end" of the memory
+manager to control allocation yourself (for example, if you don't want the
+library to use malloc() and free() for some reason).
+
+Some data is allocated "permanently" and will not be freed until the JPEG
+object is destroyed.  Most data is allocated "per image" and is freed by
+jpeg_finish_compress, jpeg_finish_decompress, or jpeg_abort.  You can call the
+memory manager yourself to allocate structures that will automatically be
+freed at these times.  Typical code for this is
+  ptr = (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE, size);
+Use JPOOL_PERMANENT to get storage that lasts as long as the JPEG object.
+Use alloc_large instead of alloc_small for anything bigger than a few Kbytes.
+There are also alloc_sarray and alloc_barray routines that automatically
+build 2-D sample or block arrays.
+
+The library's minimum space requirements to process an image depend on the
+image's width, but not on its height, because the library ordinarily works
+with "strip" buffers that are as wide as the image but just a few rows high.
+Some operating modes (eg, two-pass color quantization) require full-image
+buffers.  Such buffers are treated as "virtual arrays": only the current strip
+need be in memory, and the rest can be swapped out to a temporary file.
+
+If you use the simplest memory manager back end (jmemnobs.c), then no
+temporary files are used; virtual arrays are simply malloc()'d.  Images bigger
+than memory can be processed only if your system supports virtual memory.
+The other memory manager back ends support temporary files of various flavors
+and thus work in machines without virtual memory.  They may also be useful on
+Unix machines if you need to process images that exceed available swap space.
+
+When using temporary files, the library will make the in-memory buffers for
+its virtual arrays just big enough to stay within a "maximum memory" setting.
+Your application can set this limit by setting cinfo->mem->max_memory_to_use
+after creating the JPEG object.  (Of course, there is still a minimum size for
+the buffers, so the max-memory setting is effective only if it is bigger than
+the minimum space needed.)  If you allocate any large structures yourself, you
+must allocate them before jpeg_start_compress() or jpeg_start_decompress() in
+order to have them counted against the max memory limit.  Also keep in mind
+that space allocated with alloc_small() is ignored, on the assumption that
+it's too small to be worth worrying about; so a reasonable safety margin
+should be left when setting max_memory_to_use.
+
+If you use the jmemname.c or jmemdos.c memory manager back end, it is
+important to clean up the JPEG object properly to ensure that the temporary
+files get deleted.  (This is especially crucial with jmemdos.c, where the
+"temporary files" may be extended-memory segments; if they are not freed,
+DOS will require a reboot to recover the memory.)  Thus, with these memory
+managers, it's a good idea to provide a signal handler that will trap any
+early exit from your program.  The handler should call either jpeg_abort()
+or jpeg_destroy() for any active JPEG objects.  A handler is not needed with
+jmemnobs.c, and shouldn't be necessary with jmemansi.c or jmemmac.c either,
+since the C library is supposed to take care of deleting files made with
+tmpfile().
+
+
+Memory usage
+------------
+
+Working memory requirements while performing compression or decompression
+depend on image dimensions, image characteristics (such as colorspace and
+JPEG process), and operating mode (application-selected options).
+
+As of v6b, the decompressor requires:
+ 1. About 24K in more-or-less-fixed-size data.  This varies a bit depending
+    on operating mode and image characteristics (particularly color vs.
+    grayscale), but it doesn't depend on image dimensions.
+ 2. Strip buffers (of size proportional to the image width) for IDCT and
+    upsampling results.  The worst case for commonly used sampling factors
+    is about 34 bytes * width in pixels for a color image.  A grayscale image
+    only needs about 8 bytes per pixel column.
+ 3. A full-image DCT coefficient buffer is needed to decode a multi-scan JPEG
+    file (including progressive JPEGs), or whenever you select buffered-image
+    mode.  This takes 2 bytes/coefficient.  At typical 2x2 sampling, that's
+    3 bytes per pixel for a color image.  Worst case (1x1 sampling) requires
+    6 bytes/pixel.  For grayscale, figure 2 bytes/pixel.
+ 4. To perform 2-pass color quantization, the decompressor also needs a
+    128K color lookup table and a full-image pixel buffer (3 bytes/pixel).
+This does not count any memory allocated by the application, such as a
+buffer to hold the final output image.
+
+The above figures are valid for 8-bit JPEG data precision and a machine with
+32-bit ints.  For 12-bit JPEG data, double the size of the strip buffers and
+quantization pixel buffer.  The "fixed-size" data will be somewhat smaller
+with 16-bit ints, larger with 64-bit ints.  Also, CMYK or other unusual
+color spaces will require different amounts of space.
+
+The full-image coefficient and pixel buffers, if needed at all, do not
+have to be fully RAM resident; you can have the library use temporary
+files instead when the total memory usage would exceed a limit you set.
+(But if your OS supports virtual memory, it's probably better to just use
+jmemnobs and let the OS do the swapping.)
+
+The compressor's memory requirements are similar, except that it has no need
+for color quantization.  Also, it needs a full-image DCT coefficient buffer
+if Huffman-table optimization is asked for, even if progressive mode is not
+requested.
+
+If you need more detailed information about memory usage in a particular
+situation, you can enable the MEM_STATS code in jmemmgr.c.
+
+
+Library compile-time options
+----------------------------
+
+A number of compile-time options are available by modifying jmorecfg.h.
+
+The JPEG standard provides for both the baseline 8-bit DCT process and
+a 12-bit DCT process.  The IJG code supports 12-bit lossy JPEG if you define
+BITS_IN_JSAMPLE as 12 rather than 8.  Note that this causes JSAMPLE to be
+larger than a char, so it affects the surrounding application's image data.
+The sample applications cjpeg and djpeg can support 12-bit mode only for PPM
+and GIF file formats; you must disable the other file formats to compile a
+12-bit cjpeg or djpeg.  (install.txt has more information about that.)
+At present, a 12-bit library can handle *only* 12-bit images, not both
+precisions.  (If you need to include both 8- and 12-bit libraries in a single
+application, you could probably do it by defining NEED_SHORT_EXTERNAL_NAMES
+for just one of the copies.  You'd have to access the 8-bit and 12-bit copies
+from separate application source files.  This is untested ... if you try it,
+we'd like to hear whether it works!)
+
+Note that a 12-bit library always compresses in Huffman optimization mode,
+in order to generate valid Huffman tables.  This is necessary because our
+default Huffman tables only cover 8-bit data.  If you need to output 12-bit
+files in one pass, you'll have to supply suitable default Huffman tables.
+You may also want to supply your own DCT quantization tables; the existing
+quality-scaling code has been developed for 8-bit use, and probably doesn't
+generate especially good tables for 12-bit.
+
+The maximum number of components (color channels) in the image is determined
+by MAX_COMPONENTS.  The JPEG standard allows up to 255 components, but we
+expect that few applications will need more than four or so.
+
+On machines with unusual data type sizes, you may be able to improve
+performance or reduce memory space by tweaking the various typedefs in
+jmorecfg.h.  In particular, on some RISC CPUs, access to arrays of "short"s
+is quite slow; consider trading memory for speed by making JCOEF, INT16, and
+UINT16 be "int" or "unsigned int".  UINT8 is also a candidate to become int.
+You probably don't want to make JSAMPLE be int unless you have lots of memory
+to burn.
+
+You can reduce the size of the library by compiling out various optional
+functions.  To do this, undefine xxx_SUPPORTED symbols as necessary.
+
+You can also save a few K by not having text error messages in the library;
+the standard error message table occupies about 5Kb.  This is particularly
+reasonable for embedded applications where there's no good way to display 
+a message anyway.  To do this, remove the creation of the message table
+(jpeg_std_message_table[]) from jerror.c, and alter format_message to do
+something reasonable without it.  You could output the numeric value of the
+message code number, for example.  If you do this, you can also save a couple
+more K by modifying the TRACEMSn() macros in jerror.h to expand to nothing;
+you don't need trace capability anyway, right?
+
+
+Portability considerations
+--------------------------
+
+The JPEG library has been written to be extremely portable; the sample
+applications cjpeg and djpeg are slightly less so.  This section summarizes
+the design goals in this area.  (If you encounter any bugs that cause the
+library to be less portable than is claimed here, we'd appreciate hearing
+about them.)
+
+The code works fine on ANSI C, C++, and pre-ANSI C compilers, using any of
+the popular system include file setups, and some not-so-popular ones too.
+See install.txt for configuration procedures.
+
+The code is not dependent on the exact sizes of the C data types.  As
+distributed, we make the assumptions that
+	char	is at least 8 bits wide
+	short	is at least 16 bits wide
+	int	is at least 16 bits wide
+	long	is at least 32 bits wide
+(These are the minimum requirements of the ANSI C standard.)  Wider types will
+work fine, although memory may be used inefficiently if char is much larger
+than 8 bits or short is much bigger than 16 bits.  The code should work
+equally well with 16- or 32-bit ints.
+
+In a system where these assumptions are not met, you may be able to make the
+code work by modifying the typedefs in jmorecfg.h.  However, you will probably
+have difficulty if int is less than 16 bits wide, since references to plain
+int abound in the code.
+
+char can be either signed or unsigned, although the code runs faster if an
+unsigned char type is available.  If char is wider than 8 bits, you will need
+to redefine JOCTET and/or provide custom data source/destination managers so
+that JOCTET represents exactly 8 bits of data on external storage.
+
+The JPEG library proper does not assume ASCII representation of characters.
+But some of the image file I/O modules in cjpeg/djpeg do have ASCII
+dependencies in file-header manipulation; so does cjpeg's select_file_type()
+routine.
+
+The JPEG library does not rely heavily on the C library.  In particular, C
+stdio is used only by the data source/destination modules and the error
+handler, all of which are application-replaceable.  (cjpeg/djpeg are more
+heavily dependent on stdio.)  malloc and free are called only from the memory
+manager "back end" module, so you can use a different memory allocator by
+replacing that one file.
+
+The code generally assumes that C names must be unique in the first 15
+characters.  However, global function names can be made unique in the
+first 6 characters by defining NEED_SHORT_EXTERNAL_NAMES.
+
+More info about porting the code may be gleaned by reading jconfig.txt,
+jmorecfg.h, and jinclude.h.
+
+
+Notes for MS-DOS implementors
+-----------------------------
+
+The IJG code is designed to work efficiently in 80x86 "small" or "medium"
+memory models (i.e., data pointers are 16 bits unless explicitly declared
+"far"; code pointers can be either size).  You may be able to use small
+model to compile cjpeg or djpeg by itself, but you will probably have to use
+medium model for any larger application.  This won't make much difference in
+performance.  You *will* take a noticeable performance hit if you use a
+large-data memory model (perhaps 10%-25%), and you should avoid "huge" model
+if at all possible.
+
+The JPEG library typically needs 2Kb-3Kb of stack space.  It will also
+malloc about 20K-30K of near heap space while executing (and lots of far
+heap, but that doesn't count in this calculation).  This figure will vary
+depending on selected operating mode, and to a lesser extent on image size.
+There is also about 5Kb-6Kb of constant data which will be allocated in the
+near data segment (about 4Kb of this is the error message table).
+Thus you have perhaps 20K available for other modules' static data and near
+heap space before you need to go to a larger memory model.  The C library's
+static data will account for several K of this, but that still leaves a good
+deal for your needs.  (If you are tight on space, you could reduce the sizes
+of the I/O buffers allocated by jdatasrc.c and jdatadst.c, say from 4K to
+1K.  Another possibility is to move the error message table to far memory;
+this should be doable with only localized hacking on jerror.c.)
+
+About 2K of the near heap space is "permanent" memory that will not be
+released until you destroy the JPEG object.  This is only an issue if you
+save a JPEG object between compression or decompression operations.
+
+Far data space may also be a tight resource when you are dealing with large
+images.  The most memory-intensive case is decompression with two-pass color
+quantization, or single-pass quantization to an externally supplied color
+map.  This requires a 128Kb color lookup table plus strip buffers amounting
+to about 40 bytes per column for typical sampling ratios (eg, about 25600
+bytes for a 640-pixel-wide image).  You may not be able to process wide
+images if you have large data structures of your own.
+
+Of course, all of these concerns vanish if you use a 32-bit flat-memory-model
+compiler, such as DJGPP or Watcom C.  We highly recommend flat model if you
+can use it; the JPEG library is significantly faster in flat model.
diff --git a/rdbmp.c b/rdbmp.c
new file mode 100644
index 0000000..fd773d4
--- /dev/null
+++ b/rdbmp.c
@@ -0,0 +1,480 @@
+/*
+ * rdbmp.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to read input images in Microsoft "BMP"
+ * format (MS Windows 3.x, OS/2 1.x, and OS/2 2.x flavors).
+ * Currently, only 8-bit and 24-bit images are supported, not 1-bit or
+ * 4-bit (feeding such low-depth images into JPEG would be silly anyway).
+ * Also, we don't support RLE-compressed files.
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume input from
+ * an ordinary stdio stream.  They further assume that reading begins
+ * at the start of the file; start_input may need work if the
+ * user interface has already read some data (e.g., to determine that
+ * the file is indeed BMP format).
+ *
+ * This code contributed by James Arthur Boucher.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef BMP_SUPPORTED
+
+
+/* Macros to deal with unsigned chars as efficiently as compiler allows */
+
+#ifdef HAVE_UNSIGNED_CHAR
+typedef unsigned char U_CHAR;
+#define UCH(x)	((int) (x))
+#else /* !HAVE_UNSIGNED_CHAR */
+#ifdef CHAR_IS_UNSIGNED
+typedef char U_CHAR;
+#define UCH(x)	((int) (x))
+#else
+typedef char U_CHAR;
+#define UCH(x)	((int) (x) & 0xFF)
+#endif
+#endif /* HAVE_UNSIGNED_CHAR */
+
+
+#define	ReadOK(file,buffer,len)	(JFREAD(file,buffer,len) == ((size_t) (len)))
+
+
+/* Private version of data source object */
+
+typedef struct _bmp_source_struct * bmp_source_ptr;
+
+typedef struct _bmp_source_struct {
+  struct cjpeg_source_struct pub; /* public fields */
+
+  j_compress_ptr cinfo;		/* back link saves passing separate parm */
+
+  JSAMPARRAY colormap;		/* BMP colormap (converted to my format) */
+
+  jvirt_sarray_ptr whole_image;	/* Needed to reverse row order */
+  JDIMENSION source_row;	/* Current source row number */
+  JDIMENSION row_width;		/* Physical width of scanlines in file */
+
+  int bits_per_pixel;		/* remembers 8- or 24-bit format */
+} bmp_source_struct;
+
+
+LOCAL(int)
+read_byte (bmp_source_ptr sinfo)
+/* Read next byte from BMP file */
+{
+  register FILE *infile = sinfo->pub.input_file;
+  register int c;
+
+  if ((c = getc(infile)) == EOF)
+    ERREXIT(sinfo->cinfo, JERR_INPUT_EOF);
+  return c;
+}
+
+
+LOCAL(void)
+read_colormap (bmp_source_ptr sinfo, int cmaplen, int mapentrysize)
+/* Read the colormap from a BMP file */
+{
+  int i;
+
+  switch (mapentrysize) {
+  case 3:
+    /* BGR format (occurs in OS/2 files) */
+    for (i = 0; i < cmaplen; i++) {
+      sinfo->colormap[2][i] = (JSAMPLE) read_byte(sinfo);
+      sinfo->colormap[1][i] = (JSAMPLE) read_byte(sinfo);
+      sinfo->colormap[0][i] = (JSAMPLE) read_byte(sinfo);
+    }
+    break;
+  case 4:
+    /* BGR0 format (occurs in MS Windows files) */
+    for (i = 0; i < cmaplen; i++) {
+      sinfo->colormap[2][i] = (JSAMPLE) read_byte(sinfo);
+      sinfo->colormap[1][i] = (JSAMPLE) read_byte(sinfo);
+      sinfo->colormap[0][i] = (JSAMPLE) read_byte(sinfo);
+      (void) read_byte(sinfo);
+    }
+    break;
+  default:
+    ERREXIT(sinfo->cinfo, JERR_BMP_BADCMAP);
+    break;
+  }
+}
+
+
+/*
+ * Read one row of pixels.
+ * The image has been read into the whole_image array, but is otherwise
+ * unprocessed.  We must read it out in top-to-bottom row order, and if
+ * it is an 8-bit image, we must expand colormapped pixels to 24bit format.
+ */
+
+METHODDEF(JDIMENSION)
+get_8bit_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 8-bit colormap indexes */
+{
+  bmp_source_ptr source = (bmp_source_ptr) sinfo;
+  register JSAMPARRAY colormap = source->colormap;
+  JSAMPARRAY image_ptr;
+  register int t;
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+
+  /* Fetch next row from virtual array */
+  source->source_row--;
+  image_ptr = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, source->whole_image,
+     source->source_row, (JDIMENSION) 1, FALSE);
+
+  /* Expand the colormap indexes to real data */
+  inptr = image_ptr[0];
+  outptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    t = GETJSAMPLE(*inptr++);
+    *outptr++ = colormap[0][t];	/* can omit GETJSAMPLE() safely */
+    *outptr++ = colormap[1][t];
+    *outptr++ = colormap[2][t];
+  }
+
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_24bit_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 24-bit pixels */
+{
+  bmp_source_ptr source = (bmp_source_ptr) sinfo;
+  JSAMPARRAY image_ptr;
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+
+  /* Fetch next row from virtual array */
+  source->source_row--;
+  image_ptr = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, source->whole_image,
+     source->source_row, (JDIMENSION) 1, FALSE);
+
+  /* Transfer data.  Note source values are in BGR order
+   * (even though Microsoft's own documents say the opposite).
+   */
+  inptr = image_ptr[0];
+  outptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    outptr[2] = *inptr++;	/* can omit GETJSAMPLE() safely */
+    outptr[1] = *inptr++;
+    outptr[0] = *inptr++;
+    outptr += 3;
+  }
+
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_32bit_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 32-bit pixels */
+{
+  bmp_source_ptr source = (bmp_source_ptr) sinfo;
+  JSAMPARRAY image_ptr;
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+
+  /* Fetch next row from virtual array */
+  source->source_row--;
+  image_ptr = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, source->whole_image,
+     source->source_row, (JDIMENSION) 1, FALSE);
+  /* Transfer data.  Note source values are in BGR order
+   * (even though Microsoft's own documents say the opposite).
+   */
+  inptr = image_ptr[0];
+  outptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    outptr[2] = *inptr++;	/* can omit GETJSAMPLE() safely */
+    outptr[1] = *inptr++;
+    outptr[0] = *inptr++;
+    inptr++;			/* skip the 4th byte (Alpha channel) */
+    outptr += 3;
+  }
+
+  return 1;
+}
+
+
+/*
+ * This method loads the image into whole_image during the first call on
+ * get_pixel_rows.  The get_pixel_rows pointer is then adjusted to call
+ * get_8bit_row, get_24bit_row, or get_32bit_row on subsequent calls.
+ */
+
+METHODDEF(JDIMENSION)
+preload_image (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  bmp_source_ptr source = (bmp_source_ptr) sinfo;
+  register FILE *infile = source->pub.input_file;
+  register int c;
+  register JSAMPROW out_ptr;
+  JSAMPARRAY image_ptr;
+  JDIMENSION row, col;
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+
+  /* Read the data into a virtual array in input-file row order. */
+  for (row = 0; row < cinfo->image_height; row++) {
+    if (progress != NULL) {
+      progress->pub.pass_counter = (long) row;
+      progress->pub.pass_limit = (long) cinfo->image_height;
+      (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+    }
+    image_ptr = (*cinfo->mem->access_virt_sarray)
+      ((j_common_ptr) cinfo, source->whole_image,
+       row, (JDIMENSION) 1, TRUE);
+    out_ptr = image_ptr[0];
+    for (col = source->row_width; col > 0; col--) {
+      /* inline copy of read_byte() for speed */
+      if ((c = getc(infile)) == EOF)
+	ERREXIT(cinfo, JERR_INPUT_EOF);
+      *out_ptr++ = (JSAMPLE) c;
+    }
+  }
+  if (progress != NULL)
+    progress->completed_extra_passes++;
+
+  /* Set up to read from the virtual array in top-to-bottom order */
+  switch (source->bits_per_pixel) {
+  case 8:
+    source->pub.get_pixel_rows = get_8bit_row;
+    break;
+  case 24:
+    source->pub.get_pixel_rows = get_24bit_row;
+    break;
+  case 32:
+    source->pub.get_pixel_rows = get_32bit_row;
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BMP_BADDEPTH);
+  }
+  source->source_row = cinfo->image_height;
+
+  /* And read the first row */
+  return (*source->pub.get_pixel_rows) (cinfo, sinfo);
+}
+
+
+/*
+ * Read the file header; return image size and component count.
+ */
+
+METHODDEF(void)
+start_input_bmp (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  bmp_source_ptr source = (bmp_source_ptr) sinfo;
+  U_CHAR bmpfileheader[14];
+  U_CHAR bmpinfoheader[64];
+#define GET_2B(array,offset)  ((unsigned int) UCH(array[offset]) + \
+			       (((unsigned int) UCH(array[offset+1])) << 8))
+#define GET_4B(array,offset)  ((INT32) UCH(array[offset]) + \
+			       (((INT32) UCH(array[offset+1])) << 8) + \
+			       (((INT32) UCH(array[offset+2])) << 16) + \
+			       (((INT32) UCH(array[offset+3])) << 24))
+  INT32 bfOffBits;
+  INT32 headerSize;
+  INT32 biWidth;
+  INT32 biHeight;
+  unsigned int biPlanes;
+  INT32 biCompression;
+  INT32 biXPelsPerMeter,biYPelsPerMeter;
+  INT32 biClrUsed = 0;
+  int mapentrysize = 0;		/* 0 indicates no colormap */
+  INT32 bPad;
+  JDIMENSION row_width;
+
+  /* Read and verify the bitmap file header */
+  if (! ReadOK(source->pub.input_file, bmpfileheader, 14))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  if (GET_2B(bmpfileheader,0) != 0x4D42) /* 'BM' */
+    ERREXIT(cinfo, JERR_BMP_NOT);
+  bfOffBits = (INT32) GET_4B(bmpfileheader,10);
+  /* We ignore the remaining fileheader fields */
+
+  /* The infoheader might be 12 bytes (OS/2 1.x), 40 bytes (Windows),
+   * or 64 bytes (OS/2 2.x).  Check the first 4 bytes to find out which.
+   */
+  if (! ReadOK(source->pub.input_file, bmpinfoheader, 4))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  headerSize = (INT32) GET_4B(bmpinfoheader,0);
+  if (headerSize < 12 || headerSize > 64)
+    ERREXIT(cinfo, JERR_BMP_BADHEADER);
+  if (! ReadOK(source->pub.input_file, bmpinfoheader+4, headerSize-4))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+
+  switch ((int) headerSize) {
+  case 12:
+    /* Decode OS/2 1.x header (Microsoft calls this a BITMAPCOREHEADER) */
+    biWidth = (INT32) GET_2B(bmpinfoheader,4);
+    biHeight = (INT32) GET_2B(bmpinfoheader,6);
+    biPlanes = GET_2B(bmpinfoheader,8);
+    source->bits_per_pixel = (int) GET_2B(bmpinfoheader,10);
+
+    switch (source->bits_per_pixel) {
+    case 8:			/* colormapped image */
+      mapentrysize = 3;		/* OS/2 uses RGBTRIPLE colormap */
+      TRACEMS2(cinfo, 1, JTRC_BMP_OS2_MAPPED, (int) biWidth, (int) biHeight);
+      break;
+    case 24:			/* RGB image */
+      TRACEMS2(cinfo, 1, JTRC_BMP_OS2, (int) biWidth, (int) biHeight);
+      break;
+    default:
+      ERREXIT(cinfo, JERR_BMP_BADDEPTH);
+      break;
+    }
+    break;
+  case 40:
+  case 64:
+    /* Decode Windows 3.x header (Microsoft calls this a BITMAPINFOHEADER) */
+    /* or OS/2 2.x header, which has additional fields that we ignore */
+    biWidth = GET_4B(bmpinfoheader,4);
+    biHeight = GET_4B(bmpinfoheader,8);
+    biPlanes = GET_2B(bmpinfoheader,12);
+    source->bits_per_pixel = (int) GET_2B(bmpinfoheader,14);
+    biCompression = GET_4B(bmpinfoheader,16);
+    biXPelsPerMeter = GET_4B(bmpinfoheader,24);
+    biYPelsPerMeter = GET_4B(bmpinfoheader,28);
+    biClrUsed = GET_4B(bmpinfoheader,32);
+    /* biSizeImage, biClrImportant fields are ignored */
+
+    switch (source->bits_per_pixel) {
+    case 8:			/* colormapped image */
+      mapentrysize = 4;		/* Windows uses RGBQUAD colormap */
+      TRACEMS2(cinfo, 1, JTRC_BMP_MAPPED, (int) biWidth, (int) biHeight);
+      break;
+    case 24:			/* RGB image */
+      TRACEMS2(cinfo, 1, JTRC_BMP, (int) biWidth, (int) biHeight);
+      break;
+    case 32:			/* RGB image + Alpha channel */
+      TRACEMS2(cinfo, 1, JTRC_BMP, (int) biWidth, (int) biHeight);
+      break;
+    default:
+      ERREXIT(cinfo, JERR_BMP_BADDEPTH);
+      break;
+    }
+    if (biCompression != 0)
+      ERREXIT(cinfo, JERR_BMP_COMPRESSED);
+
+    if (biXPelsPerMeter > 0 && biYPelsPerMeter > 0) {
+      /* Set JFIF density parameters from the BMP data */
+      cinfo->X_density = (UINT16) (biXPelsPerMeter/100); /* 100 cm per meter */
+      cinfo->Y_density = (UINT16) (biYPelsPerMeter/100);
+      cinfo->density_unit = 2;	/* dots/cm */
+    }
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BMP_BADHEADER);
+    return;
+  }
+
+  if (biWidth <= 0 || biHeight <= 0)
+    ERREXIT(cinfo, JERR_BMP_EMPTY);
+  if (biPlanes != 1)
+    ERREXIT(cinfo, JERR_BMP_BADPLANES);
+
+  /* Compute distance to bitmap data --- will adjust for colormap below */
+  bPad = bfOffBits - (headerSize + 14);
+
+  /* Read the colormap, if any */
+  if (mapentrysize > 0) {
+    if (biClrUsed <= 0)
+      biClrUsed = 256;		/* assume it's 256 */
+    else if (biClrUsed > 256)
+      ERREXIT(cinfo, JERR_BMP_BADCMAP);
+    /* Allocate space to store the colormap */
+    source->colormap = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) biClrUsed, (JDIMENSION) 3);
+    /* and read it from the file */
+    read_colormap(source, (int) biClrUsed, mapentrysize);
+    /* account for size of colormap */
+    bPad -= biClrUsed * mapentrysize;
+  }
+
+  /* Skip any remaining pad bytes */
+  if (bPad < 0)			/* incorrect bfOffBits value? */
+    ERREXIT(cinfo, JERR_BMP_BADHEADER);
+  while (--bPad >= 0) {
+    (void) read_byte(source);
+  }
+
+  /* Compute row width in file, including padding to 4-byte boundary */
+  if (source->bits_per_pixel == 24)
+    row_width = (JDIMENSION) (biWidth * 3);
+  else if (source->bits_per_pixel == 32)
+    row_width = (JDIMENSION) (biWidth * 4);
+  else
+    row_width = (JDIMENSION) biWidth;
+  while ((row_width & 3) != 0) row_width++;
+  source->row_width = row_width;
+
+  /* Allocate space for inversion array, prepare for preload pass */
+  source->whole_image = (*cinfo->mem->request_virt_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+     row_width, (JDIMENSION) biHeight, (JDIMENSION) 1);
+  source->pub.get_pixel_rows = preload_image;
+  if (cinfo->progress != NULL) {
+    cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+    progress->total_extra_passes++; /* count file input as separate pass */
+  }
+
+  /* Allocate one-row buffer for returned data */
+  source->pub.buffer = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     (JDIMENSION) (biWidth * 3), (JDIMENSION) 1);
+  source->pub.buffer_height = 1;
+
+  cinfo->in_color_space = JCS_RGB;
+  cinfo->input_components = 3;
+  cinfo->data_precision = 8;
+  cinfo->image_width = (JDIMENSION) biWidth;
+  cinfo->image_height = (JDIMENSION) biHeight;
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_input_bmp (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  /* no work */
+}
+
+
+/*
+ * The module selection routine for BMP format input.
+ */
+
+GLOBAL(cjpeg_source_ptr)
+jinit_read_bmp (j_compress_ptr cinfo)
+{
+  bmp_source_ptr source;
+
+  /* Create module interface object */
+  source = (bmp_source_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(bmp_source_struct));
+  source->cinfo = cinfo;	/* make back link for subroutines */
+  /* Fill in method ptrs, except get_pixel_rows which start_input sets */
+  source->pub.start_input = start_input_bmp;
+  source->pub.finish_input = finish_input_bmp;
+
+  return (cjpeg_source_ptr) source;
+}
+
+#endif /* BMP_SUPPORTED */
diff --git a/rdcolmap.c b/rdcolmap.c
new file mode 100644
index 0000000..42b3437
--- /dev/null
+++ b/rdcolmap.c
@@ -0,0 +1,253 @@
+/*
+ * rdcolmap.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file implements djpeg's "-map file" switch.  It reads a source image
+ * and constructs a colormap to be supplied to the JPEG decompressor.
+ *
+ * Currently, these file formats are supported for the map file:
+ *   GIF: the contents of the GIF's global colormap are used.
+ *   PPM (either text or raw flavor): the entire file is read and
+ *      each unique pixel value is entered in the map.
+ * Note that reading a large PPM file will be horrendously slow.
+ * Typically, a PPM-format map file should contain just one pixel
+ * of each desired color.  Such a file can be extracted from an
+ * ordinary image PPM file with ppmtomap(1).
+ *
+ * Rescaling a PPM that has a maxval unequal to MAXJSAMPLE is not
+ * currently implemented.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef QUANT_2PASS_SUPPORTED	/* otherwise can't quantize to supplied map */
+
+/* Portions of this code are based on the PBMPLUS library, which is:
+**
+** Copyright (C) 1988 by Jef Poskanzer.
+**
+** Permission to use, copy, modify, and distribute this software and its
+** documentation for any purpose and without fee is hereby granted, provided
+** that the above copyright notice appear in all copies and that both that
+** copyright notice and this permission notice appear in supporting
+** documentation.  This software is provided "as is" without express or
+** implied warranty.
+*/
+
+
+/*
+ * Add a (potentially) new color to the color map.
+ */
+
+LOCAL(void)
+add_map_entry (j_decompress_ptr cinfo, int R, int G, int B)
+{
+  JSAMPROW colormap0 = cinfo->colormap[0];
+  JSAMPROW colormap1 = cinfo->colormap[1];
+  JSAMPROW colormap2 = cinfo->colormap[2];
+  int ncolors = cinfo->actual_number_of_colors;
+  int index;
+
+  /* Check for duplicate color. */
+  for (index = 0; index < ncolors; index++) {
+    if (GETJSAMPLE(colormap0[index]) == R &&
+	GETJSAMPLE(colormap1[index]) == G &&
+	GETJSAMPLE(colormap2[index]) == B)
+      return;			/* color is already in map */
+  }
+
+  /* Check for map overflow. */
+  if (ncolors >= (MAXJSAMPLE+1))
+    ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, (MAXJSAMPLE+1));
+
+  /* OK, add color to map. */
+  colormap0[ncolors] = (JSAMPLE) R;
+  colormap1[ncolors] = (JSAMPLE) G;
+  colormap2[ncolors] = (JSAMPLE) B;
+  cinfo->actual_number_of_colors++;
+}
+
+
+/*
+ * Extract color map from a GIF file.
+ */
+
+LOCAL(void)
+read_gif_map (j_decompress_ptr cinfo, FILE * infile)
+{
+  int header[13];
+  int i, colormaplen;
+  int R, G, B;
+
+  /* Initial 'G' has already been read by read_color_map */
+  /* Read the rest of the GIF header and logical screen descriptor */
+  for (i = 1; i < 13; i++) {
+    if ((header[i] = getc(infile)) == EOF)
+      ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+  }
+
+  /* Verify GIF Header */
+  if (header[1] != 'I' || header[2] != 'F')
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+
+  /* There must be a global color map. */
+  if ((header[10] & 0x80) == 0)
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+
+  /* OK, fetch it. */
+  colormaplen = 2 << (header[10] & 0x07);
+
+  for (i = 0; i < colormaplen; i++) {
+    R = getc(infile);
+    G = getc(infile);
+    B = getc(infile);
+    if (R == EOF || G == EOF || B == EOF)
+      ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+    add_map_entry(cinfo,
+		  R << (BITS_IN_JSAMPLE-8),
+		  G << (BITS_IN_JSAMPLE-8),
+		  B << (BITS_IN_JSAMPLE-8));
+  }
+}
+
+
+/* Support routines for reading PPM */
+
+
+LOCAL(int)
+pbm_getc (FILE * infile)
+/* Read next char, skipping over any comments */
+/* A comment/newline sequence is returned as a newline */
+{
+  register int ch;
+  
+  ch = getc(infile);
+  if (ch == '#') {
+    do {
+      ch = getc(infile);
+    } while (ch != '\n' && ch != EOF);
+  }
+  return ch;
+}
+
+
+LOCAL(unsigned int)
+read_pbm_integer (j_decompress_ptr cinfo, FILE * infile)
+/* Read an unsigned decimal integer from the PPM file */
+/* Swallows one trailing character after the integer */
+/* Note that on a 16-bit-int machine, only values up to 64k can be read. */
+/* This should not be a problem in practice. */
+{
+  register int ch;
+  register unsigned int val;
+  
+  /* Skip any leading whitespace */
+  do {
+    ch = pbm_getc(infile);
+    if (ch == EOF)
+      ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+  } while (ch == ' ' || ch == '\t' || ch == '\n' || ch == '\r');
+  
+  if (ch < '0' || ch > '9')
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+  
+  val = ch - '0';
+  while ((ch = pbm_getc(infile)) >= '0' && ch <= '9') {
+    val *= 10;
+    val += ch - '0';
+  }
+  return val;
+}
+
+
+/*
+ * Extract color map from a PPM file.
+ */
+
+LOCAL(void)
+read_ppm_map (j_decompress_ptr cinfo, FILE * infile)
+{
+  int c;
+  unsigned int w, h, maxval, row, col;
+  int R, G, B;
+
+  /* Initial 'P' has already been read by read_color_map */
+  c = getc(infile);		/* save format discriminator for a sec */
+
+  /* while we fetch the remaining header info */
+  w = read_pbm_integer(cinfo, infile);
+  h = read_pbm_integer(cinfo, infile);
+  maxval = read_pbm_integer(cinfo, infile);
+
+  if (w <= 0 || h <= 0 || maxval <= 0) /* error check */
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+
+  /* For now, we don't support rescaling from an unusual maxval. */
+  if (maxval != (unsigned int) MAXJSAMPLE)
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+
+  switch (c) {
+  case '3':			/* it's a text-format PPM file */
+    for (row = 0; row < h; row++) {
+      for (col = 0; col < w; col++) {
+	R = read_pbm_integer(cinfo, infile);
+	G = read_pbm_integer(cinfo, infile);
+	B = read_pbm_integer(cinfo, infile);
+	add_map_entry(cinfo, R, G, B);
+      }
+    }
+    break;
+
+  case '6':			/* it's a raw-format PPM file */
+    for (row = 0; row < h; row++) {
+      for (col = 0; col < w; col++) {
+	R = getc(infile);
+	G = getc(infile);
+	B = getc(infile);
+	if (R == EOF || G == EOF || B == EOF)
+	  ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+	add_map_entry(cinfo, R, G, B);
+      }
+    }
+    break;
+
+  default:
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+    break;
+  }
+}
+
+
+/*
+ * Main entry point from djpeg.c.
+ *  Input: opened input file (from file name argument on command line).
+ *  Output: colormap and actual_number_of_colors fields are set in cinfo.
+ */
+
+GLOBAL(void)
+read_color_map (j_decompress_ptr cinfo, FILE * infile)
+{
+  /* Allocate space for a color map of maximum supported size. */
+  cinfo->colormap = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     (JDIMENSION) (MAXJSAMPLE+1), (JDIMENSION) 3);
+  cinfo->actual_number_of_colors = 0; /* initialize map to empty */
+
+  /* Read first byte to determine file format */
+  switch (getc(infile)) {
+  case 'G':
+    read_gif_map(cinfo, infile);
+    break;
+  case 'P':
+    read_ppm_map(cinfo, infile);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_CMAP_FILE);
+    break;
+  }
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
diff --git a/rdgif.c b/rdgif.c
new file mode 100644
index 0000000..b27c167
--- /dev/null
+++ b/rdgif.c
@@ -0,0 +1,38 @@
+/*
+ * rdgif.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to read input images in GIF format.
+ *
+ *****************************************************************************
+ * NOTE: to avoid entanglements with Unisys' patent on LZW compression,      *
+ * the ability to read GIF files has been removed from the IJG distribution. *
+ * Sorry about that.                                                         *
+ *****************************************************************************
+ *
+ * We are required to state that
+ *    "The Graphics Interchange Format(c) is the Copyright property of
+ *    CompuServe Incorporated. GIF(sm) is a Service Mark property of
+ *    CompuServe Incorporated."
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef GIF_SUPPORTED
+
+/*
+ * The module selection routine for GIF format input.
+ */
+
+GLOBAL(cjpeg_source_ptr)
+jinit_read_gif (j_compress_ptr cinfo)
+{
+  fprintf(stderr, "GIF input is unsupported for legal reasons.  Sorry.\n");
+  exit(EXIT_FAILURE);
+  return NULL;			/* keep compiler happy */
+}
+
+#endif /* GIF_SUPPORTED */
diff --git a/rdjpgcom.1 b/rdjpgcom.1
new file mode 100644
index 0000000..97611df
--- /dev/null
+++ b/rdjpgcom.1
@@ -0,0 +1,63 @@
+.TH RDJPGCOM 1 "02 April 2009"
+.SH NAME
+rdjpgcom \- display text comments from a JPEG file
+.SH SYNOPSIS
+.B rdjpgcom
+[
+.B \-raw
+]
+[
+.B \-verbose
+]
+[
+.I filename
+]
+.LP
+.SH DESCRIPTION
+.LP
+.B rdjpgcom
+reads the named JPEG/JFIF file, or the standard input if no file is named,
+and prints any text comments found in the file on the standard output.
+.PP
+The JPEG standard allows "comment" (COM) blocks to occur within a JPEG file.
+Although the standard doesn't actually define what COM blocks are for, they
+are widely used to hold user-supplied text strings.  This lets you add
+annotations, titles, index terms, etc to your JPEG files, and later retrieve
+them as text.  COM blocks do not interfere with the image stored in the JPEG
+file.  The maximum size of a COM block is 64K, but you can have as many of
+them as you like in one JPEG file.
+.SH OPTIONS
+.TP
+.B \-raw
+Normally
+.B rdjpgcom
+escapes non-printable characters in comments, for security reasons.
+This option avoids that.
+.PP
+.B \-verbose
+Causes
+.B rdjpgcom
+to also display the JPEG image dimensions.
+.PP
+Switch names may be abbreviated, and are not case sensitive.
+.SH HINTS
+.B rdjpgcom
+does not depend on the IJG JPEG library.  Its source code is intended as an
+illustration of the minimum amount of code required to parse a JPEG file
+header correctly.
+.PP
+In
+.B \-verbose
+mode,
+.B rdjpgcom
+will also attempt to print the contents of any "APP12" markers as text.
+Some digital cameras produce APP12 markers containing useful textual
+information.  If you like, you can modify the source code to print
+other APPn marker types as well.
+.SH SEE ALSO
+.BR cjpeg (1),
+.BR djpeg (1),
+.BR jpegtran (1),
+.BR wrjpgcom (1)
+.SH AUTHOR
+Independent JPEG Group
diff --git a/rdjpgcom.c b/rdjpgcom.c
new file mode 100644
index 0000000..3719154
--- /dev/null
+++ b/rdjpgcom.c
@@ -0,0 +1,515 @@
+/*
+ * rdjpgcom.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 2009 by Bill Allombert, Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a very simple stand-alone application that displays
+ * the text in COM (comment) markers in a JFIF file.
+ * This may be useful as an example of the minimum logic needed to parse
+ * JPEG markers.
+ */
+
+#define JPEG_CJPEG_DJPEG	/* to get the command-line config symbols */
+#include "jinclude.h"		/* get auto-config symbols, <stdio.h> */
+
+#ifdef HAVE_LOCALE_H
+#include <locale.h>		/* Bill Allombert: use locale for isprint */
+#endif
+#include <ctype.h>		/* to declare isupper(), tolower() */
+#ifdef USE_SETMODE
+#include <fcntl.h>		/* to declare setmode()'s parameter macros */
+/* If you have setmode() but not <io.h>, just delete this line: */
+#include <io.h>			/* to declare setmode() */
+#endif
+
+#ifdef USE_CCOMMAND		/* command-line reader for Macintosh */
+#ifdef __MWERKS__
+#include <SIOUX.h>              /* Metrowerks needs this */
+#include <console.h>		/* ... and this */
+#endif
+#ifdef THINK_C
+#include <console.h>		/* Think declares it here */
+#endif
+#endif
+
+#ifdef DONT_USE_B_MODE		/* define mode parameters for fopen() */
+#define READ_BINARY	"r"
+#else
+#ifdef VMS			/* VMS is very nonstandard */
+#define READ_BINARY	"rb", "ctx=stm"
+#else				/* standard ANSI-compliant case */
+#define READ_BINARY	"rb"
+#endif
+#endif
+
+#ifndef EXIT_FAILURE		/* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+#ifndef EXIT_SUCCESS
+#ifdef VMS
+#define EXIT_SUCCESS  1		/* VMS is very nonstandard */
+#else
+#define EXIT_SUCCESS  0
+#endif
+#endif
+
+
+/*
+ * These macros are used to read the input file.
+ * To reuse this code in another application, you might need to change these.
+ */
+
+static FILE * infile;		/* input JPEG file */
+
+/* Return next input byte, or EOF if no more */
+#define NEXTBYTE()  getc(infile)
+
+
+/* Error exit handler */
+#define ERREXIT(msg)  (fprintf(stderr, "%s\n", msg), exit(EXIT_FAILURE))
+
+
+/* Read one byte, testing for EOF */
+static int
+read_1_byte (void)
+{
+  int c;
+
+  c = NEXTBYTE();
+  if (c == EOF)
+    ERREXIT("Premature EOF in JPEG file");
+  return c;
+}
+
+/* Read 2 bytes, convert to unsigned int */
+/* All 2-byte quantities in JPEG markers are MSB first */
+static unsigned int
+read_2_bytes (void)
+{
+  int c1, c2;
+
+  c1 = NEXTBYTE();
+  if (c1 == EOF)
+    ERREXIT("Premature EOF in JPEG file");
+  c2 = NEXTBYTE();
+  if (c2 == EOF)
+    ERREXIT("Premature EOF in JPEG file");
+  return (((unsigned int) c1) << 8) + ((unsigned int) c2);
+}
+
+
+/*
+ * JPEG markers consist of one or more 0xFF bytes, followed by a marker
+ * code byte (which is not an FF).  Here are the marker codes of interest
+ * in this program.  (See jdmarker.c for a more complete list.)
+ */
+
+#define M_SOF0  0xC0		/* Start Of Frame N */
+#define M_SOF1  0xC1		/* N indicates which compression process */
+#define M_SOF2  0xC2		/* Only SOF0-SOF2 are now in common use */
+#define M_SOF3  0xC3
+#define M_SOF5  0xC5		/* NB: codes C4 and CC are NOT SOF markers */
+#define M_SOF6  0xC6
+#define M_SOF7  0xC7
+#define M_SOF9  0xC9
+#define M_SOF10 0xCA
+#define M_SOF11 0xCB
+#define M_SOF13 0xCD
+#define M_SOF14 0xCE
+#define M_SOF15 0xCF
+#define M_SOI   0xD8		/* Start Of Image (beginning of datastream) */
+#define M_EOI   0xD9		/* End Of Image (end of datastream) */
+#define M_SOS   0xDA		/* Start Of Scan (begins compressed data) */
+#define M_APP0	0xE0		/* Application-specific marker, type N */
+#define M_APP12	0xEC		/* (we don't bother to list all 16 APPn's) */
+#define M_COM   0xFE		/* COMment */
+
+
+/*
+ * Find the next JPEG marker and return its marker code.
+ * We expect at least one FF byte, possibly more if the compressor used FFs
+ * to pad the file.
+ * There could also be non-FF garbage between markers.  The treatment of such
+ * garbage is unspecified; we choose to skip over it but emit a warning msg.
+ * NB: this routine must not be used after seeing SOS marker, since it will
+ * not deal correctly with FF/00 sequences in the compressed image data...
+ */
+
+static int
+next_marker (void)
+{
+  int c;
+  int discarded_bytes = 0;
+
+  /* Find 0xFF byte; count and skip any non-FFs. */
+  c = read_1_byte();
+  while (c != 0xFF) {
+    discarded_bytes++;
+    c = read_1_byte();
+  }
+  /* Get marker code byte, swallowing any duplicate FF bytes.  Extra FFs
+   * are legal as pad bytes, so don't count them in discarded_bytes.
+   */
+  do {
+    c = read_1_byte();
+  } while (c == 0xFF);
+
+  if (discarded_bytes != 0) {
+    fprintf(stderr, "Warning: garbage data found in JPEG file\n");
+  }
+
+  return c;
+}
+
+
+/*
+ * Read the initial marker, which should be SOI.
+ * For a JFIF file, the first two bytes of the file should be literally
+ * 0xFF M_SOI.  To be more general, we could use next_marker, but if the
+ * input file weren't actually JPEG at all, next_marker might read the whole
+ * file and then return a misleading error message...
+ */
+
+static int
+first_marker (void)
+{
+  int c1, c2;
+
+  c1 = NEXTBYTE();
+  c2 = NEXTBYTE();
+  if (c1 != 0xFF || c2 != M_SOI)
+    ERREXIT("Not a JPEG file");
+  return c2;
+}
+
+
+/*
+ * Most types of marker are followed by a variable-length parameter segment.
+ * This routine skips over the parameters for any marker we don't otherwise
+ * want to process.
+ * Note that we MUST skip the parameter segment explicitly in order not to
+ * be fooled by 0xFF bytes that might appear within the parameter segment;
+ * such bytes do NOT introduce new markers.
+ */
+
+static void
+skip_variable (void)
+/* Skip over an unknown or uninteresting variable-length marker */
+{
+  unsigned int length;
+
+  /* Get the marker parameter length count */
+  length = read_2_bytes();
+  /* Length includes itself, so must be at least 2 */
+  if (length < 2)
+    ERREXIT("Erroneous JPEG marker length");
+  length -= 2;
+  /* Skip over the remaining bytes */
+  while (length > 0) {
+    (void) read_1_byte();
+    length--;
+  }
+}
+
+
+/*
+ * Process a COM marker.
+ * We want to print out the marker contents as legible text;
+ * we must guard against non-text junk and varying newline representations.
+ */
+
+static void
+process_COM (int raw)
+{
+  unsigned int length;
+  int ch;
+  int lastch = 0;
+
+  /* Bill Allombert: set locale properly for isprint */
+#ifdef HAVE_LOCALE_H
+  setlocale(LC_CTYPE, "");
+#endif
+
+  /* Get the marker parameter length count */
+  length = read_2_bytes();
+  /* Length includes itself, so must be at least 2 */
+  if (length < 2)
+    ERREXIT("Erroneous JPEG marker length");
+  length -= 2;
+
+  while (length > 0) {
+    ch = read_1_byte();
+    if (raw) {
+      putc(ch, stdout);
+    /* Emit the character in a readable form.
+     * Nonprintables are converted to \nnn form,
+     * while \ is converted to \\.
+     * Newlines in CR, CR/LF, or LF form will be printed as one newline.
+     */
+    } else if (ch == '\r') {
+      printf("\n");
+    } else if (ch == '\n') {
+      if (lastch != '\r')
+	printf("\n");
+    } else if (ch == '\\') {
+      printf("\\\\");
+    } else if (isprint(ch)) {
+      putc(ch, stdout);
+    } else {
+      printf("\\%03o", ch);
+    }
+    lastch = ch;
+    length--;
+  }
+  printf("\n");
+
+  /* Bill Allombert: revert to C locale */
+#ifdef HAVE_LOCALE_H
+  setlocale(LC_CTYPE, "C");
+#endif
+}
+
+
+/*
+ * Process a SOFn marker.
+ * This code is only needed if you want to know the image dimensions...
+ */
+
+static void
+process_SOFn (int marker)
+{
+  unsigned int length;
+  unsigned int image_height, image_width;
+  int data_precision, num_components;
+  const char * process;
+  int ci;
+
+  length = read_2_bytes();	/* usual parameter length count */
+
+  data_precision = read_1_byte();
+  image_height = read_2_bytes();
+  image_width = read_2_bytes();
+  num_components = read_1_byte();
+
+  switch (marker) {
+  case M_SOF0:	process = "Baseline";  break;
+  case M_SOF1:	process = "Extended sequential";  break;
+  case M_SOF2:	process = "Progressive";  break;
+  case M_SOF3:	process = "Lossless";  break;
+  case M_SOF5:	process = "Differential sequential";  break;
+  case M_SOF6:	process = "Differential progressive";  break;
+  case M_SOF7:	process = "Differential lossless";  break;
+  case M_SOF9:	process = "Extended sequential, arithmetic coding";  break;
+  case M_SOF10:	process = "Progressive, arithmetic coding";  break;
+  case M_SOF11:	process = "Lossless, arithmetic coding";  break;
+  case M_SOF13:	process = "Differential sequential, arithmetic coding";  break;
+  case M_SOF14:	process = "Differential progressive, arithmetic coding"; break;
+  case M_SOF15:	process = "Differential lossless, arithmetic coding";  break;
+  default:	process = "Unknown";  break;
+  }
+
+  printf("JPEG image is %uw * %uh, %d color components, %d bits per sample\n",
+	 image_width, image_height, num_components, data_precision);
+  printf("JPEG process: %s\n", process);
+
+  if (length != (unsigned int) (8 + num_components * 3))
+    ERREXIT("Bogus SOF marker length");
+
+  for (ci = 0; ci < num_components; ci++) {
+    (void) read_1_byte();	/* Component ID code */
+    (void) read_1_byte();	/* H, V sampling factors */
+    (void) read_1_byte();	/* Quantization table number */
+  }
+}
+
+
+/*
+ * Parse the marker stream until SOS or EOI is seen;
+ * display any COM markers.
+ * While the companion program wrjpgcom will always insert COM markers before
+ * SOFn, other implementations might not, so we scan to SOS before stopping.
+ * If we were only interested in the image dimensions, we would stop at SOFn.
+ * (Conversely, if we only cared about COM markers, there would be no need
+ * for special code to handle SOFn; we could treat it like other markers.)
+ */
+
+static int
+scan_JPEG_header (int verbose, int raw)
+{
+  int marker;
+
+  /* Expect SOI at start of file */
+  if (first_marker() != M_SOI)
+    ERREXIT("Expected SOI marker first");
+
+  /* Scan miscellaneous markers until we reach SOS. */
+  for (;;) {
+    marker = next_marker();
+    switch (marker) {
+      /* Note that marker codes 0xC4, 0xC8, 0xCC are not, and must not be,
+       * treated as SOFn.  C4 in particular is actually DHT.
+       */
+    case M_SOF0:		/* Baseline */
+    case M_SOF1:		/* Extended sequential, Huffman */
+    case M_SOF2:		/* Progressive, Huffman */
+    case M_SOF3:		/* Lossless, Huffman */
+    case M_SOF5:		/* Differential sequential, Huffman */
+    case M_SOF6:		/* Differential progressive, Huffman */
+    case M_SOF7:		/* Differential lossless, Huffman */
+    case M_SOF9:		/* Extended sequential, arithmetic */
+    case M_SOF10:		/* Progressive, arithmetic */
+    case M_SOF11:		/* Lossless, arithmetic */
+    case M_SOF13:		/* Differential sequential, arithmetic */
+    case M_SOF14:		/* Differential progressive, arithmetic */
+    case M_SOF15:		/* Differential lossless, arithmetic */
+      if (verbose)
+	process_SOFn(marker);
+      else
+	skip_variable();
+      break;
+
+    case M_SOS:			/* stop before hitting compressed data */
+      return marker;
+
+    case M_EOI:			/* in case it's a tables-only JPEG stream */
+      return marker;
+
+    case M_COM:
+      process_COM(raw);
+      break;
+
+    case M_APP12:
+      /* Some digital camera makers put useful textual information into
+       * APP12 markers, so we print those out too when in -verbose mode.
+       */
+      if (verbose) {
+	printf("APP12 contains:\n");
+	process_COM(raw);
+      } else
+	skip_variable();
+      break;
+
+    default:			/* Anything else just gets skipped */
+      skip_variable();		/* we assume it has a parameter count... */
+      break;
+    }
+  } /* end loop */
+}
+
+
+/* Command line parsing code */
+
+static const char * progname;	/* program name for error messages */
+
+
+static void
+usage (void)
+/* complain about bad command line */
+{
+  fprintf(stderr, "rdjpgcom displays any textual comments in a JPEG file.\n");
+
+  fprintf(stderr, "Usage: %s [switches] [inputfile]\n", progname);
+
+  fprintf(stderr, "Switches (names may be abbreviated):\n");
+  fprintf(stderr, "  -raw        Display non-printable characters in comments (unsafe)\n");
+  fprintf(stderr, "  -verbose    Also display dimensions of JPEG image\n");
+
+  exit(EXIT_FAILURE);
+}
+
+
+static int
+keymatch (char * arg, const char * keyword, int minchars)
+/* Case-insensitive matching of (possibly abbreviated) keyword switches. */
+/* keyword is the constant keyword (must be lower case already), */
+/* minchars is length of minimum legal abbreviation. */
+{
+  register int ca, ck;
+  register int nmatched = 0;
+
+  while ((ca = *arg++) != '\0') {
+    if ((ck = *keyword++) == '\0')
+      return 0;			/* arg longer than keyword, no good */
+    if (isupper(ca))		/* force arg to lcase (assume ck is already) */
+      ca = tolower(ca);
+    if (ca != ck)
+      return 0;			/* no good */
+    nmatched++;			/* count matched characters */
+  }
+  /* reached end of argument; fail if it's too short for unique abbrev */
+  if (nmatched < minchars)
+    return 0;
+  return 1;			/* A-OK */
+}
+
+
+/*
+ * The main program.
+ */
+
+int
+main (int argc, char **argv)
+{
+  int argn;
+  char * arg;
+  int verbose = 0, raw = 0;
+
+  /* On Mac, fetch a command line. */
+#ifdef USE_CCOMMAND
+  argc = ccommand(&argv);
+#endif
+
+  progname = argv[0];
+  if (progname == NULL || progname[0] == 0)
+    progname = "rdjpgcom";	/* in case C library doesn't provide it */
+
+  /* Parse switches, if any */
+  for (argn = 1; argn < argc; argn++) {
+    arg = argv[argn];
+    if (arg[0] != '-')
+      break;			/* not switch, must be file name */
+    arg++;			/* advance over '-' */
+    if (keymatch(arg, "verbose", 1)) {
+      verbose++;
+    } else if (keymatch(arg, "raw", 1)) {
+      raw = 1;
+    } else
+      usage();
+  }
+
+  /* Open the input file. */
+  /* Unix style: expect zero or one file name */
+  if (argn < argc-1) {
+    fprintf(stderr, "%s: only one input file\n", progname);
+    usage();
+  }
+  if (argn < argc) {
+    if ((infile = fopen(argv[argn], READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s\n", progname, argv[argn]);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default input file is stdin */
+#ifdef USE_SETMODE		/* need to hack file mode? */
+    setmode(fileno(stdin), O_BINARY);
+#endif
+#ifdef USE_FDOPEN		/* need to re-open in binary mode? */
+    if ((infile = fdopen(fileno(stdin), READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open stdin\n", progname);
+      exit(EXIT_FAILURE);
+    }
+#else
+    infile = stdin;
+#endif
+  }
+
+  /* Scan the JPEG headers. */
+  (void) scan_JPEG_header(verbose, raw);
+
+  /* All done. */
+  exit(EXIT_SUCCESS);
+  return 0;			/* suppress no-return-value warnings */
+}
diff --git a/rdppm.c b/rdppm.c
new file mode 100644
index 0000000..a757022
--- /dev/null
+++ b/rdppm.c
@@ -0,0 +1,459 @@
+/*
+ * rdppm.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2009 by Bill Allombert, Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to read input images in PPM/PGM format.
+ * The extended 2-byte-per-sample raw PPM/PGM formats are supported.
+ * The PBMPLUS library is NOT required to compile this software
+ * (but it is highly useful as a set of PPM image manipulation programs).
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume input from
+ * an ordinary stdio stream.  They further assume that reading begins
+ * at the start of the file; start_input may need work if the
+ * user interface has already read some data (e.g., to determine that
+ * the file is indeed PPM format).
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef PPM_SUPPORTED
+
+
+/* Portions of this code are based on the PBMPLUS library, which is:
+**
+** Copyright (C) 1988 by Jef Poskanzer.
+**
+** Permission to use, copy, modify, and distribute this software and its
+** documentation for any purpose and without fee is hereby granted, provided
+** that the above copyright notice appear in all copies and that both that
+** copyright notice and this permission notice appear in supporting
+** documentation.  This software is provided "as is" without express or
+** implied warranty.
+*/
+
+
+/* Macros to deal with unsigned chars as efficiently as compiler allows */
+
+#ifdef HAVE_UNSIGNED_CHAR
+typedef unsigned char U_CHAR;
+#define UCH(x)	((int) (x))
+#else /* !HAVE_UNSIGNED_CHAR */
+#ifdef CHAR_IS_UNSIGNED
+typedef char U_CHAR;
+#define UCH(x)	((int) (x))
+#else
+typedef char U_CHAR;
+#define UCH(x)	((int) (x) & 0xFF)
+#endif
+#endif /* HAVE_UNSIGNED_CHAR */
+
+
+#define	ReadOK(file,buffer,len)	(JFREAD(file,buffer,len) == ((size_t) (len)))
+
+
+/*
+ * On most systems, reading individual bytes with getc() is drastically less
+ * efficient than buffering a row at a time with fread().  On PCs, we must
+ * allocate the buffer in near data space, because we are assuming small-data
+ * memory model, wherein fread() can't reach far memory.  If you need to
+ * process very wide images on a PC, you might have to compile in large-memory
+ * model, or else replace fread() with a getc() loop --- which will be much
+ * slower.
+ */
+
+
+/* Private version of data source object */
+
+typedef struct {
+  struct cjpeg_source_struct pub; /* public fields */
+
+  U_CHAR *iobuffer;		/* non-FAR pointer to I/O buffer */
+  JSAMPROW pixrow;		/* FAR pointer to same */
+  size_t buffer_width;		/* width of I/O buffer */
+  JSAMPLE *rescale;		/* => maxval-remapping array, or NULL */
+} ppm_source_struct;
+
+typedef ppm_source_struct * ppm_source_ptr;
+
+
+LOCAL(int)
+pbm_getc (FILE * infile)
+/* Read next char, skipping over any comments */
+/* A comment/newline sequence is returned as a newline */
+{
+  register int ch;
+
+  ch = getc(infile);
+  if (ch == '#') {
+    do {
+      ch = getc(infile);
+    } while (ch != '\n' && ch != EOF);
+  }
+  return ch;
+}
+
+
+LOCAL(unsigned int)
+read_pbm_integer (j_compress_ptr cinfo, FILE * infile)
+/* Read an unsigned decimal integer from the PPM file */
+/* Swallows one trailing character after the integer */
+/* Note that on a 16-bit-int machine, only values up to 64k can be read. */
+/* This should not be a problem in practice. */
+{
+  register int ch;
+  register unsigned int val;
+
+  /* Skip any leading whitespace */
+  do {
+    ch = pbm_getc(infile);
+    if (ch == EOF)
+      ERREXIT(cinfo, JERR_INPUT_EOF);
+  } while (ch == ' ' || ch == '\t' || ch == '\n' || ch == '\r');
+
+  if (ch < '0' || ch > '9')
+    ERREXIT(cinfo, JERR_PPM_NONNUMERIC);
+
+  val = ch - '0';
+  while ((ch = pbm_getc(infile)) >= '0' && ch <= '9') {
+    val *= 10;
+    val += ch - '0';
+  }
+  return val;
+}
+
+
+/*
+ * Read one row of pixels.
+ *
+ * We provide several different versions depending on input file format.
+ * In all cases, input is scaled to the size of JSAMPLE.
+ *
+ * A really fast path is provided for reading byte/sample raw files with
+ * maxval = MAXJSAMPLE, which is the normal case for 8-bit data.
+ */
+
+
+METHODDEF(JDIMENSION)
+get_text_gray_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading text-format PGM files with any maxval */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  FILE * infile = source->pub.input_file;
+  register JSAMPROW ptr;
+  register JSAMPLE *rescale = source->rescale;
+  JDIMENSION col;
+
+  ptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    *ptr++ = rescale[read_pbm_integer(cinfo, infile)];
+  }
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_text_rgb_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading text-format PPM files with any maxval */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  FILE * infile = source->pub.input_file;
+  register JSAMPROW ptr;
+  register JSAMPLE *rescale = source->rescale;
+  JDIMENSION col;
+
+  ptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    *ptr++ = rescale[read_pbm_integer(cinfo, infile)];
+    *ptr++ = rescale[read_pbm_integer(cinfo, infile)];
+    *ptr++ = rescale[read_pbm_integer(cinfo, infile)];
+  }
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_scaled_gray_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading raw-byte-format PGM files with any maxval */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  register JSAMPROW ptr;
+  register U_CHAR * bufferptr;
+  register JSAMPLE *rescale = source->rescale;
+  JDIMENSION col;
+
+  if (! ReadOK(source->pub.input_file, source->iobuffer, source->buffer_width))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  ptr = source->pub.buffer[0];
+  bufferptr = source->iobuffer;
+  for (col = cinfo->image_width; col > 0; col--) {
+    *ptr++ = rescale[UCH(*bufferptr++)];
+  }
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_scaled_rgb_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading raw-byte-format PPM files with any maxval */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  register JSAMPROW ptr;
+  register U_CHAR * bufferptr;
+  register JSAMPLE *rescale = source->rescale;
+  JDIMENSION col;
+
+  if (! ReadOK(source->pub.input_file, source->iobuffer, source->buffer_width))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  ptr = source->pub.buffer[0];
+  bufferptr = source->iobuffer;
+  for (col = cinfo->image_width; col > 0; col--) {
+    *ptr++ = rescale[UCH(*bufferptr++)];
+    *ptr++ = rescale[UCH(*bufferptr++)];
+    *ptr++ = rescale[UCH(*bufferptr++)];
+  }
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_raw_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading raw-byte-format files with maxval = MAXJSAMPLE.
+ * In this case we just read right into the JSAMPLE buffer!
+ * Note that same code works for PPM and PGM files.
+ */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+
+  if (! ReadOK(source->pub.input_file, source->iobuffer, source->buffer_width))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_word_gray_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading raw-word-format PGM files with any maxval */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  register JSAMPROW ptr;
+  register U_CHAR * bufferptr;
+  register JSAMPLE *rescale = source->rescale;
+  JDIMENSION col;
+
+  if (! ReadOK(source->pub.input_file, source->iobuffer, source->buffer_width))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  ptr = source->pub.buffer[0];
+  bufferptr = source->iobuffer;
+  for (col = cinfo->image_width; col > 0; col--) {
+    register int temp;
+    temp  = UCH(*bufferptr++) << 8;
+    temp |= UCH(*bufferptr++);
+    *ptr++ = rescale[temp];
+  }
+  return 1;
+}
+
+
+METHODDEF(JDIMENSION)
+get_word_rgb_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading raw-word-format PPM files with any maxval */
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  register JSAMPROW ptr;
+  register U_CHAR * bufferptr;
+  register JSAMPLE *rescale = source->rescale;
+  JDIMENSION col;
+
+  if (! ReadOK(source->pub.input_file, source->iobuffer, source->buffer_width))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+  ptr = source->pub.buffer[0];
+  bufferptr = source->iobuffer;
+  for (col = cinfo->image_width; col > 0; col--) {
+    register int temp;
+    temp  = UCH(*bufferptr++) << 8;
+    temp |= UCH(*bufferptr++);
+    *ptr++ = rescale[temp];
+    temp  = UCH(*bufferptr++) << 8;
+    temp |= UCH(*bufferptr++);
+    *ptr++ = rescale[temp];
+    temp  = UCH(*bufferptr++) << 8;
+    temp |= UCH(*bufferptr++);
+    *ptr++ = rescale[temp];
+  }
+  return 1;
+}
+
+
+/*
+ * Read the file header; return image size and component count.
+ */
+
+METHODDEF(void)
+start_input_ppm (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  ppm_source_ptr source = (ppm_source_ptr) sinfo;
+  int c;
+  unsigned int w, h, maxval;
+  boolean need_iobuffer, use_raw_buffer, need_rescale;
+
+  if (getc(source->pub.input_file) != 'P')
+    ERREXIT(cinfo, JERR_PPM_NOT);
+
+  c = getc(source->pub.input_file); /* subformat discriminator character */
+
+  /* detect unsupported variants (ie, PBM) before trying to read header */
+  switch (c) {
+  case '2':			/* it's a text-format PGM file */
+  case '3':			/* it's a text-format PPM file */
+  case '5':			/* it's a raw-format PGM file */
+  case '6':			/* it's a raw-format PPM file */
+    break;
+  default:
+    ERREXIT(cinfo, JERR_PPM_NOT);
+    break;
+  }
+
+  /* fetch the remaining header info */
+  w = read_pbm_integer(cinfo, source->pub.input_file);
+  h = read_pbm_integer(cinfo, source->pub.input_file);
+  maxval = read_pbm_integer(cinfo, source->pub.input_file);
+
+  if (w <= 0 || h <= 0 || maxval <= 0) /* error check */
+    ERREXIT(cinfo, JERR_PPM_NOT);
+
+  cinfo->data_precision = BITS_IN_JSAMPLE; /* we always rescale data to this */
+  cinfo->image_width = (JDIMENSION) w;
+  cinfo->image_height = (JDIMENSION) h;
+
+  /* initialize flags to most common settings */
+  need_iobuffer = TRUE;		/* do we need an I/O buffer? */
+  use_raw_buffer = FALSE;	/* do we map input buffer onto I/O buffer? */
+  need_rescale = TRUE;		/* do we need a rescale array? */
+
+  switch (c) {
+  case '2':			/* it's a text-format PGM file */
+    cinfo->input_components = 1;
+    cinfo->in_color_space = JCS_GRAYSCALE;
+    TRACEMS2(cinfo, 1, JTRC_PGM_TEXT, w, h);
+    source->pub.get_pixel_rows = get_text_gray_row;
+    need_iobuffer = FALSE;
+    break;
+
+  case '3':			/* it's a text-format PPM file */
+    cinfo->input_components = 3;
+    cinfo->in_color_space = JCS_RGB;
+    TRACEMS2(cinfo, 1, JTRC_PPM_TEXT, w, h);
+    source->pub.get_pixel_rows = get_text_rgb_row;
+    need_iobuffer = FALSE;
+    break;
+
+  case '5':			/* it's a raw-format PGM file */
+    cinfo->input_components = 1;
+    cinfo->in_color_space = JCS_GRAYSCALE;
+    TRACEMS2(cinfo, 1, JTRC_PGM, w, h);
+    if (maxval > 255) {
+      source->pub.get_pixel_rows = get_word_gray_row;
+    } else if (maxval == MAXJSAMPLE && SIZEOF(JSAMPLE) == SIZEOF(U_CHAR)) {
+      source->pub.get_pixel_rows = get_raw_row;
+      use_raw_buffer = TRUE;
+      need_rescale = FALSE;
+    } else {
+      source->pub.get_pixel_rows = get_scaled_gray_row;
+    }
+    break;
+
+  case '6':			/* it's a raw-format PPM file */
+    cinfo->input_components = 3;
+    cinfo->in_color_space = JCS_RGB;
+    TRACEMS2(cinfo, 1, JTRC_PPM, w, h);
+    if (maxval > 255) {
+      source->pub.get_pixel_rows = get_word_rgb_row;
+    } else if (maxval == MAXJSAMPLE && SIZEOF(JSAMPLE) == SIZEOF(U_CHAR)) {
+      source->pub.get_pixel_rows = get_raw_row;
+      use_raw_buffer = TRUE;
+      need_rescale = FALSE;
+    } else {
+      source->pub.get_pixel_rows = get_scaled_rgb_row;
+    }
+    break;
+  }
+
+  /* Allocate space for I/O buffer: 1 or 3 bytes or words/pixel. */
+  if (need_iobuffer) {
+    source->buffer_width = (size_t) w * cinfo->input_components *
+      ((maxval<=255) ? SIZEOF(U_CHAR) : (2*SIZEOF(U_CHAR)));
+    source->iobuffer = (U_CHAR *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  source->buffer_width);
+  }
+
+  /* Create compressor input buffer. */
+  if (use_raw_buffer) {
+    /* For unscaled raw-input case, we can just map it onto the I/O buffer. */
+    /* Synthesize a JSAMPARRAY pointer structure */
+    /* Cast here implies near->far pointer conversion on PCs */
+    source->pixrow = (JSAMPROW) source->iobuffer;
+    source->pub.buffer = & source->pixrow;
+    source->pub.buffer_height = 1;
+  } else {
+    /* Need to translate anyway, so make a separate sample buffer. */
+    source->pub.buffer = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) w * cinfo->input_components, (JDIMENSION) 1);
+    source->pub.buffer_height = 1;
+  }
+
+  /* Compute the rescaling array if required. */
+  if (need_rescale) {
+    INT32 val, half_maxval;
+
+    /* On 16-bit-int machines we have to be careful of maxval = 65535 */
+    source->rescale = (JSAMPLE *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  (size_t) (((long) maxval + 1L) * SIZEOF(JSAMPLE)));
+    half_maxval = maxval / 2;
+    for (val = 0; val <= (INT32) maxval; val++) {
+      /* The multiplication here must be done in 32 bits to avoid overflow */
+      source->rescale[val] = (JSAMPLE) ((val*MAXJSAMPLE + half_maxval)/maxval);
+    }
+  }
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_input_ppm (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  /* no work */
+}
+
+
+/*
+ * The module selection routine for PPM format input.
+ */
+
+GLOBAL(cjpeg_source_ptr)
+jinit_read_ppm (j_compress_ptr cinfo)
+{
+  ppm_source_ptr source;
+
+  /* Create module interface object */
+  source = (ppm_source_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(ppm_source_struct));
+  /* Fill in method ptrs, except get_pixel_rows which start_input sets */
+  source->pub.start_input = start_input_ppm;
+  source->pub.finish_input = finish_input_ppm;
+
+  return (cjpeg_source_ptr) source;
+}
+
+#endif /* PPM_SUPPORTED */
diff --git a/rdrle.c b/rdrle.c
new file mode 100644
index 0000000..542bc37
--- /dev/null
+++ b/rdrle.c
@@ -0,0 +1,387 @@
+/*
+ * rdrle.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to read input images in Utah RLE format.
+ * The Utah Raster Toolkit library is required (version 3.1 or later).
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume input from
+ * an ordinary stdio stream.  They further assume that reading begins
+ * at the start of the file; start_input may need work if the
+ * user interface has already read some data (e.g., to determine that
+ * the file is indeed RLE format).
+ *
+ * Based on code contributed by Mike Lijewski,
+ * with updates from Robert Hutchinson.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef RLE_SUPPORTED
+
+/* rle.h is provided by the Utah Raster Toolkit. */
+
+#include <rle.h>
+
+/*
+ * We assume that JSAMPLE has the same representation as rle_pixel,
+ * to wit, "unsigned char".  Hence we can't cope with 12- or 16-bit samples.
+ */
+
+#if BITS_IN_JSAMPLE != 8
+  Sorry, this code only copes with 8-bit JSAMPLEs. /* deliberate syntax err */
+#endif
+
+/*
+ * We support the following types of RLE files:
+ *   
+ *   GRAYSCALE   - 8 bits, no colormap
+ *   MAPPEDGRAY  - 8 bits, 1 channel colomap
+ *   PSEUDOCOLOR - 8 bits, 3 channel colormap
+ *   TRUECOLOR   - 24 bits, 3 channel colormap
+ *   DIRECTCOLOR - 24 bits, no colormap
+ *
+ * For now, we ignore any alpha channel in the image.
+ */
+
+typedef enum
+  { GRAYSCALE, MAPPEDGRAY, PSEUDOCOLOR, TRUECOLOR, DIRECTCOLOR } rle_kind;
+
+
+/*
+ * Since RLE stores scanlines bottom-to-top, we have to invert the image
+ * to conform to JPEG's top-to-bottom order.  To do this, we read the
+ * incoming image into a virtual array on the first get_pixel_rows call,
+ * then fetch the required row from the virtual array on subsequent calls.
+ */
+
+typedef struct _rle_source_struct * rle_source_ptr;
+
+typedef struct _rle_source_struct {
+  struct cjpeg_source_struct pub; /* public fields */
+
+  rle_kind visual;              /* actual type of input file */
+  jvirt_sarray_ptr image;       /* virtual array to hold the image */
+  JDIMENSION row;		/* current row # in the virtual array */
+  rle_hdr header;               /* Input file information */
+  rle_pixel** rle_row;          /* holds a row returned by rle_getrow() */
+
+} rle_source_struct;
+
+
+/*
+ * Read the file header; return image size and component count.
+ */
+
+METHODDEF(void)
+start_input_rle (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  rle_source_ptr source = (rle_source_ptr) sinfo;
+  JDIMENSION width, height;
+#ifdef PROGRESS_REPORT
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+#endif
+
+  /* Use RLE library routine to get the header info */
+  source->header = *rle_hdr_init(NULL);
+  source->header.rle_file = source->pub.input_file;
+  switch (rle_get_setup(&(source->header))) {
+  case RLE_SUCCESS:
+    /* A-OK */
+    break;
+  case RLE_NOT_RLE:
+    ERREXIT(cinfo, JERR_RLE_NOT);
+    break;
+  case RLE_NO_SPACE:
+    ERREXIT(cinfo, JERR_RLE_MEM);
+    break;
+  case RLE_EMPTY:
+    ERREXIT(cinfo, JERR_RLE_EMPTY);
+    break;
+  case RLE_EOF:
+    ERREXIT(cinfo, JERR_RLE_EOF);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_RLE_BADERROR);
+    break;
+  }
+
+  /* Figure out what we have, set private vars and return values accordingly */
+  
+  width  = source->header.xmax - source->header.xmin + 1;
+  height = source->header.ymax - source->header.ymin + 1;
+  source->header.xmin = 0;		/* realign horizontally */
+  source->header.xmax = width-1;
+
+  cinfo->image_width      = width;
+  cinfo->image_height     = height;
+  cinfo->data_precision   = 8;  /* we can only handle 8 bit data */
+
+  if (source->header.ncolors == 1 && source->header.ncmap == 0) {
+    source->visual     = GRAYSCALE;
+    TRACEMS2(cinfo, 1, JTRC_RLE_GRAY, width, height);
+  } else if (source->header.ncolors == 1 && source->header.ncmap == 1) {
+    source->visual     = MAPPEDGRAY;
+    TRACEMS3(cinfo, 1, JTRC_RLE_MAPGRAY, width, height,
+             1 << source->header.cmaplen);
+  } else if (source->header.ncolors == 1 && source->header.ncmap == 3) {
+    source->visual     = PSEUDOCOLOR;
+    TRACEMS3(cinfo, 1, JTRC_RLE_MAPPED, width, height,
+	     1 << source->header.cmaplen);
+  } else if (source->header.ncolors == 3 && source->header.ncmap == 3) {
+    source->visual     = TRUECOLOR;
+    TRACEMS3(cinfo, 1, JTRC_RLE_FULLMAP, width, height,
+	     1 << source->header.cmaplen);
+  } else if (source->header.ncolors == 3 && source->header.ncmap == 0) {
+    source->visual     = DIRECTCOLOR;
+    TRACEMS2(cinfo, 1, JTRC_RLE, width, height);
+  } else
+    ERREXIT(cinfo, JERR_RLE_UNSUPPORTED);
+  
+  if (source->visual == GRAYSCALE || source->visual == MAPPEDGRAY) {
+    cinfo->in_color_space   = JCS_GRAYSCALE;
+    cinfo->input_components = 1;
+  } else {
+    cinfo->in_color_space   = JCS_RGB;
+    cinfo->input_components = 3;
+  }
+
+  /*
+   * A place to hold each scanline while it's converted.
+   * (GRAYSCALE scanlines don't need converting)
+   */
+  if (source->visual != GRAYSCALE) {
+    source->rle_row = (rle_pixel**) (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) width, (JDIMENSION) cinfo->input_components);
+  }
+
+  /* request a virtual array to hold the image */
+  source->image = (*cinfo->mem->request_virt_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+     (JDIMENSION) (width * source->header.ncolors),
+     (JDIMENSION) height, (JDIMENSION) 1);
+
+#ifdef PROGRESS_REPORT
+  if (progress != NULL) {
+    /* count file input as separate pass */
+    progress->total_extra_passes++;
+  }
+#endif
+
+  source->pub.buffer_height = 1;
+}
+
+
+/*
+ * Read one row of pixels.
+ * Called only after load_image has read the image into the virtual array.
+ * Used for GRAYSCALE, MAPPEDGRAY, TRUECOLOR, and DIRECTCOLOR images.
+ */
+
+METHODDEF(JDIMENSION)
+get_rle_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  rle_source_ptr source = (rle_source_ptr) sinfo;
+
+  source->row--;
+  source->pub.buffer = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, source->image, source->row, (JDIMENSION) 1, FALSE);
+
+  return 1;
+}
+
+/*
+ * Read one row of pixels.
+ * Called only after load_image has read the image into the virtual array.
+ * Used for PSEUDOCOLOR images.
+ */
+
+METHODDEF(JDIMENSION)
+get_pseudocolor_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  rle_source_ptr source = (rle_source_ptr) sinfo;
+  JSAMPROW src_row, dest_row;
+  JDIMENSION col;
+  rle_map *colormap;
+  int val;
+
+  colormap = source->header.cmap;
+  dest_row = source->pub.buffer[0];
+  source->row--;
+  src_row = * (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, source->image, source->row, (JDIMENSION) 1, FALSE);
+
+  for (col = cinfo->image_width; col > 0; col--) {
+    val = GETJSAMPLE(*src_row++);
+    *dest_row++ = (JSAMPLE) (colormap[val      ] >> 8);
+    *dest_row++ = (JSAMPLE) (colormap[val + 256] >> 8);
+    *dest_row++ = (JSAMPLE) (colormap[val + 512] >> 8);
+  }
+
+  return 1;
+}
+
+
+/*
+ * Load the image into a virtual array.  We have to do this because RLE
+ * files start at the lower left while the JPEG standard has them starting
+ * in the upper left.  This is called the first time we want to get a row
+ * of input.  What we do is load the RLE data into the array and then call
+ * the appropriate routine to read one row from the array.  Before returning,
+ * we set source->pub.get_pixel_rows so that subsequent calls go straight to
+ * the appropriate row-reading routine.
+ */
+
+METHODDEF(JDIMENSION)
+load_image (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  rle_source_ptr source = (rle_source_ptr) sinfo;
+  JDIMENSION row, col;
+  JSAMPROW  scanline, red_ptr, green_ptr, blue_ptr;
+  rle_pixel **rle_row;
+  rle_map *colormap;
+  char channel;
+#ifdef PROGRESS_REPORT
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+#endif
+
+  colormap = source->header.cmap;
+  rle_row = source->rle_row;
+
+  /* Read the RLE data into our virtual array.
+   * We assume here that (a) rle_pixel is represented the same as JSAMPLE,
+   * and (b) we are not on a machine where FAR pointers differ from regular.
+   */
+  RLE_CLR_BIT(source->header, RLE_ALPHA); /* don't read the alpha channel */
+
+#ifdef PROGRESS_REPORT
+  if (progress != NULL) {
+    progress->pub.pass_limit = cinfo->image_height;
+    progress->pub.pass_counter = 0;
+    (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+  }
+#endif
+
+  switch (source->visual) {
+
+  case GRAYSCALE:
+  case PSEUDOCOLOR:
+    for (row = 0; row < cinfo->image_height; row++) {
+      rle_row = (rle_pixel **) (*cinfo->mem->access_virt_sarray)
+         ((j_common_ptr) cinfo, source->image, row, (JDIMENSION) 1, TRUE);
+      rle_getrow(&source->header, rle_row);
+#ifdef PROGRESS_REPORT
+      if (progress != NULL) {
+        progress->pub.pass_counter++;
+        (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+      }
+#endif
+    }
+    break;
+
+  case MAPPEDGRAY:
+  case TRUECOLOR:
+    for (row = 0; row < cinfo->image_height; row++) {
+      scanline = * (*cinfo->mem->access_virt_sarray)
+        ((j_common_ptr) cinfo, source->image, row, (JDIMENSION) 1, TRUE);
+      rle_row = source->rle_row;
+      rle_getrow(&source->header, rle_row);
+
+      for (col = 0; col < cinfo->image_width; col++) {
+        for (channel = 0; channel < source->header.ncolors; channel++) {
+          *scanline++ = (JSAMPLE)
+            (colormap[GETJSAMPLE(rle_row[channel][col]) + 256 * channel] >> 8);
+        }
+      }
+
+#ifdef PROGRESS_REPORT
+      if (progress != NULL) {
+        progress->pub.pass_counter++;
+        (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+      }
+#endif
+    }
+    break;
+
+  case DIRECTCOLOR:
+    for (row = 0; row < cinfo->image_height; row++) {
+      scanline = * (*cinfo->mem->access_virt_sarray)
+        ((j_common_ptr) cinfo, source->image, row, (JDIMENSION) 1, TRUE);
+      rle_getrow(&source->header, rle_row);
+
+      red_ptr   = rle_row[0];
+      green_ptr = rle_row[1];
+      blue_ptr  = rle_row[2];
+
+      for (col = cinfo->image_width; col > 0; col--) {
+        *scanline++ = *red_ptr++;
+        *scanline++ = *green_ptr++;
+        *scanline++ = *blue_ptr++;
+      }
+
+#ifdef PROGRESS_REPORT
+      if (progress != NULL) {
+        progress->pub.pass_counter++;
+        (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+      }
+#endif
+    }
+  }
+
+#ifdef PROGRESS_REPORT
+  if (progress != NULL)
+    progress->completed_extra_passes++;
+#endif
+
+  /* Set up to call proper row-extraction routine in future */
+  if (source->visual == PSEUDOCOLOR) {
+    source->pub.buffer = source->rle_row;
+    source->pub.get_pixel_rows = get_pseudocolor_row;
+  } else {
+    source->pub.get_pixel_rows = get_rle_row;
+  }
+  source->row = cinfo->image_height;
+
+  /* And fetch the topmost (bottommost) row */
+  return (*source->pub.get_pixel_rows) (cinfo, sinfo);   
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_input_rle (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  /* no work */
+}
+
+
+/*
+ * The module selection routine for RLE format input.
+ */
+
+GLOBAL(cjpeg_source_ptr)
+jinit_read_rle (j_compress_ptr cinfo)
+{
+  rle_source_ptr source;
+
+  /* Create module interface object */
+  source = (rle_source_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+                                  SIZEOF(rle_source_struct));
+  /* Fill in method ptrs */
+  source->pub.start_input = start_input_rle;
+  source->pub.finish_input = finish_input_rle;
+  source->pub.get_pixel_rows = load_image;
+
+  return (cjpeg_source_ptr) source;
+}
+
+#endif /* RLE_SUPPORTED */
diff --git a/rdswitch.c b/rdswitch.c
new file mode 100644
index 0000000..d406755
--- /dev/null
+++ b/rdswitch.c
@@ -0,0 +1,421 @@
+/*
+ * rdswitch.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to process some of cjpeg's more complicated
+ * command-line switches.  Switches processed here are:
+ *	-qtables file		Read quantization tables from text file
+ *	-scans file		Read scan script from text file
+ *	-quality N[,N,...]	Set quality ratings
+ *	-qslots N[,N,...]	Set component quantization table selectors
+ *	-sample HxV[,HxV,...]	Set component sampling factors
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+#include <ctype.h>		/* to declare isdigit(), isspace() */
+
+
+LOCAL(int)
+text_getc (FILE * file)
+/* Read next char, skipping over any comments (# to end of line) */
+/* A comment/newline sequence is returned as a newline */
+{
+  register int ch;
+  
+  ch = getc(file);
+  if (ch == '#') {
+    do {
+      ch = getc(file);
+    } while (ch != '\n' && ch != EOF);
+  }
+  return ch;
+}
+
+
+LOCAL(boolean)
+read_text_integer (FILE * file, long * result, int * termchar)
+/* Read an unsigned decimal integer from a file, store it in result */
+/* Reads one trailing character after the integer; returns it in termchar */
+{
+  register int ch;
+  register long val;
+  
+  /* Skip any leading whitespace, detect EOF */
+  do {
+    ch = text_getc(file);
+    if (ch == EOF) {
+      *termchar = ch;
+      return FALSE;
+    }
+  } while (isspace(ch));
+  
+  if (! isdigit(ch)) {
+    *termchar = ch;
+    return FALSE;
+  }
+
+  val = ch - '0';
+  while ((ch = text_getc(file)) != EOF) {
+    if (! isdigit(ch))
+      break;
+    val *= 10;
+    val += ch - '0';
+  }
+  *result = val;
+  *termchar = ch;
+  return TRUE;
+}
+
+
+#if JPEG_LIB_VERSION < 70
+static int q_scale_factor[NUM_QUANT_TBLS] = {100, 100, 100, 100};
+#endif
+
+GLOBAL(boolean)
+read_quant_tables (j_compress_ptr cinfo, char * filename, boolean force_baseline)
+/* Read a set of quantization tables from the specified file.
+ * The file is plain ASCII text: decimal numbers with whitespace between.
+ * Comments preceded by '#' may be included in the file.
+ * There may be one to NUM_QUANT_TBLS tables in the file, each of 64 values.
+ * The tables are implicitly numbered 0,1,etc.
+ * NOTE: does not affect the qslots mapping, which will default to selecting
+ * table 0 for luminance (or primary) components, 1 for chrominance components.
+ * You must use -qslots if you want a different component->table mapping.
+ */
+{
+  FILE * fp;
+  int tblno, i, termchar;
+  long val;
+  unsigned int table[DCTSIZE2];
+
+  if ((fp = fopen(filename, "r")) == NULL) {
+    fprintf(stderr, "Can't open table file %s\n", filename);
+    return FALSE;
+  }
+  tblno = 0;
+
+  while (read_text_integer(fp, &val, &termchar)) { /* read 1st element of table */
+    if (tblno >= NUM_QUANT_TBLS) {
+      fprintf(stderr, "Too many tables in file %s\n", filename);
+      fclose(fp);
+      return FALSE;
+    }
+    table[0] = (unsigned int) val;
+    for (i = 1; i < DCTSIZE2; i++) {
+      if (! read_text_integer(fp, &val, &termchar)) {
+	fprintf(stderr, "Invalid table data in file %s\n", filename);
+	fclose(fp);
+	return FALSE;
+      }
+      table[i] = (unsigned int) val;
+    }
+#if JPEG_LIB_VERSION >= 70
+    jpeg_add_quant_table(cinfo, tblno, table, cinfo->q_scale_factor[tblno],
+			 force_baseline);
+#else
+    jpeg_add_quant_table(cinfo, tblno, table, q_scale_factor[tblno],
+                         force_baseline);
+#endif
+    tblno++;
+  }
+
+  if (termchar != EOF) {
+    fprintf(stderr, "Non-numeric data in file %s\n", filename);
+    fclose(fp);
+    return FALSE;
+  }
+
+  fclose(fp);
+  return TRUE;
+}
+
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+
+LOCAL(boolean)
+read_scan_integer (FILE * file, long * result, int * termchar)
+/* Variant of read_text_integer that always looks for a non-space termchar;
+ * this simplifies parsing of punctuation in scan scripts.
+ */
+{
+  register int ch;
+
+  if (! read_text_integer(file, result, termchar))
+    return FALSE;
+  ch = *termchar;
+  while (ch != EOF && isspace(ch))
+    ch = text_getc(file);
+  if (isdigit(ch)) {		/* oops, put it back */
+    if (ungetc(ch, file) == EOF)
+      return FALSE;
+    ch = ' ';
+  } else {
+    /* Any separators other than ';' and ':' are ignored;
+     * this allows user to insert commas, etc, if desired.
+     */
+    if (ch != EOF && ch != ';' && ch != ':')
+      ch = ' ';
+  }
+  *termchar = ch;
+  return TRUE;
+}
+
+
+GLOBAL(boolean)
+read_scan_script (j_compress_ptr cinfo, char * filename)
+/* Read a scan script from the specified text file.
+ * Each entry in the file defines one scan to be emitted.
+ * Entries are separated by semicolons ';'.
+ * An entry contains one to four component indexes,
+ * optionally followed by a colon ':' and four progressive-JPEG parameters.
+ * The component indexes denote which component(s) are to be transmitted
+ * in the current scan.  The first component has index 0.
+ * Sequential JPEG is used if the progressive-JPEG parameters are omitted.
+ * The file is free format text: any whitespace may appear between numbers
+ * and the ':' and ';' punctuation marks.  Also, other punctuation (such
+ * as commas or dashes) can be placed between numbers if desired.
+ * Comments preceded by '#' may be included in the file.
+ * Note: we do very little validity checking here;
+ * jcmaster.c will validate the script parameters.
+ */
+{
+  FILE * fp;
+  int scanno, ncomps, termchar;
+  long val;
+  jpeg_scan_info * scanptr;
+#define MAX_SCANS  100		/* quite arbitrary limit */
+  jpeg_scan_info scans[MAX_SCANS];
+
+  if ((fp = fopen(filename, "r")) == NULL) {
+    fprintf(stderr, "Can't open scan definition file %s\n", filename);
+    return FALSE;
+  }
+  scanptr = scans;
+  scanno = 0;
+
+  while (read_scan_integer(fp, &val, &termchar)) {
+    if (scanno >= MAX_SCANS) {
+      fprintf(stderr, "Too many scans defined in file %s\n", filename);
+      fclose(fp);
+      return FALSE;
+    }
+    scanptr->component_index[0] = (int) val;
+    ncomps = 1;
+    while (termchar == ' ') {
+      if (ncomps >= MAX_COMPS_IN_SCAN) {
+	fprintf(stderr, "Too many components in one scan in file %s\n",
+		filename);
+	fclose(fp);
+	return FALSE;
+      }
+      if (! read_scan_integer(fp, &val, &termchar))
+	goto bogus;
+      scanptr->component_index[ncomps] = (int) val;
+      ncomps++;
+    }
+    scanptr->comps_in_scan = ncomps;
+    if (termchar == ':') {
+      if (! read_scan_integer(fp, &val, &termchar) || termchar != ' ')
+	goto bogus;
+      scanptr->Ss = (int) val;
+      if (! read_scan_integer(fp, &val, &termchar) || termchar != ' ')
+	goto bogus;
+      scanptr->Se = (int) val;
+      if (! read_scan_integer(fp, &val, &termchar) || termchar != ' ')
+	goto bogus;
+      scanptr->Ah = (int) val;
+      if (! read_scan_integer(fp, &val, &termchar))
+	goto bogus;
+      scanptr->Al = (int) val;
+    } else {
+      /* set non-progressive parameters */
+      scanptr->Ss = 0;
+      scanptr->Se = DCTSIZE2-1;
+      scanptr->Ah = 0;
+      scanptr->Al = 0;
+    }
+    if (termchar != ';' && termchar != EOF) {
+bogus:
+      fprintf(stderr, "Invalid scan entry format in file %s\n", filename);
+      fclose(fp);
+      return FALSE;
+    }
+    scanptr++, scanno++;
+  }
+
+  if (termchar != EOF) {
+    fprintf(stderr, "Non-numeric data in file %s\n", filename);
+    fclose(fp);
+    return FALSE;
+  }
+
+  if (scanno > 0) {
+    /* Stash completed scan list in cinfo structure.
+     * NOTE: for cjpeg's use, JPOOL_IMAGE is the right lifetime for this data,
+     * but if you want to compress multiple images you'd want JPOOL_PERMANENT.
+     */
+    scanptr = (jpeg_scan_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  scanno * SIZEOF(jpeg_scan_info));
+    MEMCOPY(scanptr, scans, scanno * SIZEOF(jpeg_scan_info));
+    cinfo->scan_info = scanptr;
+    cinfo->num_scans = scanno;
+  }
+
+  fclose(fp);
+  return TRUE;
+}
+
+#endif /* C_MULTISCAN_FILES_SUPPORTED */
+
+
+#if JPEG_LIB_VERSION < 70
+/* These are the sample quantization tables given in JPEG spec section K.1.
+ * The spec says that the values given produce "good" quality, and
+ * when divided by 2, "very good" quality.
+ */
+static const unsigned int std_luminance_quant_tbl[DCTSIZE2] = {
+  16,  11,  10,  16,  24,  40,  51,  61,
+  12,  12,  14,  19,  26,  58,  60,  55,
+  14,  13,  16,  24,  40,  57,  69,  56,
+  14,  17,  22,  29,  51,  87,  80,  62,
+  18,  22,  37,  56,  68, 109, 103,  77,
+  24,  35,  55,  64,  81, 104, 113,  92,
+  49,  64,  78,  87, 103, 121, 120, 101,
+  72,  92,  95,  98, 112, 100, 103,  99
+};
+static const unsigned int std_chrominance_quant_tbl[DCTSIZE2] = {
+  17,  18,  24,  47,  99,  99,  99,  99,
+  18,  21,  26,  66,  99,  99,  99,  99,
+  24,  26,  56,  99,  99,  99,  99,  99,
+  47,  66,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99
+};
+
+
+LOCAL(void)
+jpeg_default_qtables (j_compress_ptr cinfo, boolean force_baseline)
+{
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+		       q_scale_factor[0], force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+		       q_scale_factor[1], force_baseline);
+}
+#endif
+
+
+GLOBAL(boolean)
+set_quality_ratings (j_compress_ptr cinfo, char *arg, boolean force_baseline)
+/* Process a quality-ratings parameter string, of the form
+ *     N[,N,...]
+ * If there are more q-table slots than parameters, the last value is replicated.
+ */
+{
+  int val = 75;			/* default value */
+  int tblno;
+  char ch;
+
+  for (tblno = 0; tblno < NUM_QUANT_TBLS; tblno++) {
+    if (*arg) {
+      ch = ',';			/* if not set by sscanf, will be ',' */
+      if (sscanf(arg, "%d%c", &val, &ch) < 1)
+	return FALSE;
+      if (ch != ',')		/* syntax check */
+	return FALSE;
+      /* Convert user 0-100 rating to percentage scaling */
+#if JPEG_LIB_VERSION >= 70
+      cinfo->q_scale_factor[tblno] = jpeg_quality_scaling(val);
+#else
+      q_scale_factor[tblno] = jpeg_quality_scaling(val);
+#endif
+      while (*arg && *arg++ != ',') /* advance to next segment of arg string */
+	;
+    } else {
+      /* reached end of parameter, set remaining factors to last value */
+#if JPEG_LIB_VERSION >= 70
+      cinfo->q_scale_factor[tblno] = jpeg_quality_scaling(val);
+#else
+      q_scale_factor[tblno] = jpeg_quality_scaling(val);
+#endif
+    }
+  }
+  jpeg_default_qtables(cinfo, force_baseline);
+  return TRUE;
+}
+
+
+GLOBAL(boolean)
+set_quant_slots (j_compress_ptr cinfo, char *arg)
+/* Process a quantization-table-selectors parameter string, of the form
+ *     N[,N,...]
+ * If there are more components than parameters, the last value is replicated.
+ */
+{
+  int val = 0;			/* default table # */
+  int ci;
+  char ch;
+
+  for (ci = 0; ci < MAX_COMPONENTS; ci++) {
+    if (*arg) {
+      ch = ',';			/* if not set by sscanf, will be ',' */
+      if (sscanf(arg, "%d%c", &val, &ch) < 1)
+	return FALSE;
+      if (ch != ',')		/* syntax check */
+	return FALSE;
+      if (val < 0 || val >= NUM_QUANT_TBLS) {
+	fprintf(stderr, "JPEG quantization tables are numbered 0..%d\n",
+		NUM_QUANT_TBLS-1);
+	return FALSE;
+      }
+      cinfo->comp_info[ci].quant_tbl_no = val;
+      while (*arg && *arg++ != ',') /* advance to next segment of arg string */
+	;
+    } else {
+      /* reached end of parameter, set remaining components to last table */
+      cinfo->comp_info[ci].quant_tbl_no = val;
+    }
+  }
+  return TRUE;
+}
+
+
+GLOBAL(boolean)
+set_sample_factors (j_compress_ptr cinfo, char *arg)
+/* Process a sample-factors parameter string, of the form
+ *     HxV[,HxV,...]
+ * If there are more components than parameters, "1x1" is assumed for the rest.
+ */
+{
+  int ci, val1, val2;
+  char ch1, ch2;
+
+  for (ci = 0; ci < MAX_COMPONENTS; ci++) {
+    if (*arg) {
+      ch2 = ',';		/* if not set by sscanf, will be ',' */
+      if (sscanf(arg, "%d%c%d%c", &val1, &ch1, &val2, &ch2) < 3)
+	return FALSE;
+      if ((ch1 != 'x' && ch1 != 'X') || ch2 != ',') /* syntax check */
+	return FALSE;
+      if (val1 <= 0 || val1 > 4 || val2 <= 0 || val2 > 4) {
+	fprintf(stderr, "JPEG sampling factors must be 1..4\n");
+	return FALSE;
+      }
+      cinfo->comp_info[ci].h_samp_factor = val1;
+      cinfo->comp_info[ci].v_samp_factor = val2;
+      while (*arg && *arg++ != ',') /* advance to next segment of arg string */
+	;
+    } else {
+      /* reached end of parameter, set remaining components to 1x1 sampling */
+      cinfo->comp_info[ci].h_samp_factor = 1;
+      cinfo->comp_info[ci].v_samp_factor = 1;
+    }
+  }
+  return TRUE;
+}
diff --git a/rdtarga.c b/rdtarga.c
new file mode 100644
index 0000000..4c2cd26
--- /dev/null
+++ b/rdtarga.c
@@ -0,0 +1,500 @@
+/*
+ * rdtarga.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to read input images in Targa format.
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume input from
+ * an ordinary stdio stream.  They further assume that reading begins
+ * at the start of the file; start_input may need work if the
+ * user interface has already read some data (e.g., to determine that
+ * the file is indeed Targa format).
+ *
+ * Based on code contributed by Lee Daniel Crocker.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef TARGA_SUPPORTED
+
+
+/* Macros to deal with unsigned chars as efficiently as compiler allows */
+
+#ifdef HAVE_UNSIGNED_CHAR
+typedef unsigned char U_CHAR;
+#define UCH(x)	((int) (x))
+#else /* !HAVE_UNSIGNED_CHAR */
+#ifdef CHAR_IS_UNSIGNED
+typedef char U_CHAR;
+#define UCH(x)	((int) (x))
+#else
+typedef char U_CHAR;
+#define UCH(x)	((int) (x) & 0xFF)
+#endif
+#endif /* HAVE_UNSIGNED_CHAR */
+
+
+#define	ReadOK(file,buffer,len)	(JFREAD(file,buffer,len) == ((size_t) (len)))
+
+
+/* Private version of data source object */
+
+typedef struct _tga_source_struct * tga_source_ptr;
+
+typedef struct _tga_source_struct {
+  struct cjpeg_source_struct pub; /* public fields */
+
+  j_compress_ptr cinfo;		/* back link saves passing separate parm */
+
+  JSAMPARRAY colormap;		/* Targa colormap (converted to my format) */
+
+  jvirt_sarray_ptr whole_image;	/* Needed if funny input row order */
+  JDIMENSION current_row;	/* Current logical row number to read */
+
+  /* Pointer to routine to extract next Targa pixel from input file */
+  JMETHOD(void, read_pixel, (tga_source_ptr sinfo));
+
+  /* Result of read_pixel is delivered here: */
+  U_CHAR tga_pixel[4];
+
+  int pixel_size;		/* Bytes per Targa pixel (1 to 4) */
+
+  /* State info for reading RLE-coded pixels; both counts must be init to 0 */
+  int block_count;		/* # of pixels remaining in RLE block */
+  int dup_pixel_count;		/* # of times to duplicate previous pixel */
+
+  /* This saves the correct pixel-row-expansion method for preload_image */
+  JMETHOD(JDIMENSION, get_pixel_rows, (j_compress_ptr cinfo,
+				       cjpeg_source_ptr sinfo));
+} tga_source_struct;
+
+
+/* For expanding 5-bit pixel values to 8-bit with best rounding */
+
+static const UINT8 c5to8bits[32] = {
+    0,   8,  16,  25,  33,  41,  49,  58,
+   66,  74,  82,  90,  99, 107, 115, 123,
+  132, 140, 148, 156, 165, 173, 181, 189,
+  197, 206, 214, 222, 230, 239, 247, 255
+};
+
+
+
+LOCAL(int)
+read_byte (tga_source_ptr sinfo)
+/* Read next byte from Targa file */
+{
+  register FILE *infile = sinfo->pub.input_file;
+  register int c;
+
+  if ((c = getc(infile)) == EOF)
+    ERREXIT(sinfo->cinfo, JERR_INPUT_EOF);
+  return c;
+}
+
+
+LOCAL(void)
+read_colormap (tga_source_ptr sinfo, int cmaplen, int mapentrysize)
+/* Read the colormap from a Targa file */
+{
+  int i;
+
+  /* Presently only handles 24-bit BGR format */
+  if (mapentrysize != 24)
+    ERREXIT(sinfo->cinfo, JERR_TGA_BADCMAP);
+
+  for (i = 0; i < cmaplen; i++) {
+    sinfo->colormap[2][i] = (JSAMPLE) read_byte(sinfo);
+    sinfo->colormap[1][i] = (JSAMPLE) read_byte(sinfo);
+    sinfo->colormap[0][i] = (JSAMPLE) read_byte(sinfo);
+  }
+}
+
+
+/*
+ * read_pixel methods: get a single pixel from Targa file into tga_pixel[]
+ */
+
+METHODDEF(void)
+read_non_rle_pixel (tga_source_ptr sinfo)
+/* Read one Targa pixel from the input file; no RLE expansion */
+{
+  register FILE *infile = sinfo->pub.input_file;
+  register int i;
+
+  for (i = 0; i < sinfo->pixel_size; i++) {
+    sinfo->tga_pixel[i] = (U_CHAR) getc(infile);
+  }
+}
+
+
+METHODDEF(void)
+read_rle_pixel (tga_source_ptr sinfo)
+/* Read one Targa pixel from the input file, expanding RLE data as needed */
+{
+  register FILE *infile = sinfo->pub.input_file;
+  register int i;
+
+  /* Duplicate previously read pixel? */
+  if (sinfo->dup_pixel_count > 0) {
+    sinfo->dup_pixel_count--;
+    return;
+  }
+
+  /* Time to read RLE block header? */
+  if (--sinfo->block_count < 0) { /* decrement pixels remaining in block */
+    i = read_byte(sinfo);
+    if (i & 0x80) {		/* Start of duplicate-pixel block? */
+      sinfo->dup_pixel_count = i & 0x7F; /* number of dups after this one */
+      sinfo->block_count = 0;	/* then read new block header */
+    } else {
+      sinfo->block_count = i & 0x7F; /* number of pixels after this one */
+    }
+  }
+
+  /* Read next pixel */
+  for (i = 0; i < sinfo->pixel_size; i++) {
+    sinfo->tga_pixel[i] = (U_CHAR) getc(infile);
+  }
+}
+
+
+/*
+ * Read one row of pixels.
+ *
+ * We provide several different versions depending on input file format.
+ */
+
+
+METHODDEF(JDIMENSION)
+get_8bit_gray_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 8-bit grayscale pixels */
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  register JSAMPROW ptr;
+  register JDIMENSION col;
+  
+  ptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    (*source->read_pixel) (source); /* Load next pixel into tga_pixel */
+    *ptr++ = (JSAMPLE) UCH(source->tga_pixel[0]);
+  }
+  return 1;
+}
+
+METHODDEF(JDIMENSION)
+get_8bit_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 8-bit colormap indexes */
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  register int t;
+  register JSAMPROW ptr;
+  register JDIMENSION col;
+  register JSAMPARRAY colormap = source->colormap;
+
+  ptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    (*source->read_pixel) (source); /* Load next pixel into tga_pixel */
+    t = UCH(source->tga_pixel[0]);
+    *ptr++ = colormap[0][t];
+    *ptr++ = colormap[1][t];
+    *ptr++ = colormap[2][t];
+  }
+  return 1;
+}
+
+METHODDEF(JDIMENSION)
+get_16bit_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 16-bit pixels */
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  register int t;
+  register JSAMPROW ptr;
+  register JDIMENSION col;
+  
+  ptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    (*source->read_pixel) (source); /* Load next pixel into tga_pixel */
+    t = UCH(source->tga_pixel[0]);
+    t += UCH(source->tga_pixel[1]) << 8;
+    /* We expand 5 bit data to 8 bit sample width.
+     * The format of the 16-bit (LSB first) input word is
+     *     xRRRRRGGGGGBBBBB
+     */
+    ptr[2] = (JSAMPLE) c5to8bits[t & 0x1F];
+    t >>= 5;
+    ptr[1] = (JSAMPLE) c5to8bits[t & 0x1F];
+    t >>= 5;
+    ptr[0] = (JSAMPLE) c5to8bits[t & 0x1F];
+    ptr += 3;
+  }
+  return 1;
+}
+
+METHODDEF(JDIMENSION)
+get_24bit_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+/* This version is for reading 24-bit pixels */
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  register JSAMPROW ptr;
+  register JDIMENSION col;
+  
+  ptr = source->pub.buffer[0];
+  for (col = cinfo->image_width; col > 0; col--) {
+    (*source->read_pixel) (source); /* Load next pixel into tga_pixel */
+    *ptr++ = (JSAMPLE) UCH(source->tga_pixel[2]); /* change BGR to RGB order */
+    *ptr++ = (JSAMPLE) UCH(source->tga_pixel[1]);
+    *ptr++ = (JSAMPLE) UCH(source->tga_pixel[0]);
+  }
+  return 1;
+}
+
+/*
+ * Targa also defines a 32-bit pixel format with order B,G,R,A.
+ * We presently ignore the attribute byte, so the code for reading
+ * these pixels is identical to the 24-bit routine above.
+ * This works because the actual pixel length is only known to read_pixel.
+ */
+
+#define get_32bit_row  get_24bit_row
+
+
+/*
+ * This method is for re-reading the input data in standard top-down
+ * row order.  The entire image has already been read into whole_image
+ * with proper conversion of pixel format, but it's in a funny row order.
+ */
+
+METHODDEF(JDIMENSION)
+get_memory_row (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  JDIMENSION source_row;
+
+  /* Compute row of source that maps to current_row of normal order */
+  /* For now, assume image is bottom-up and not interlaced. */
+  /* NEEDS WORK to support interlaced images! */
+  source_row = cinfo->image_height - source->current_row - 1;
+
+  /* Fetch that row from virtual array */
+  source->pub.buffer = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, source->whole_image,
+     source_row, (JDIMENSION) 1, FALSE);
+
+  source->current_row++;
+  return 1;
+}
+
+
+/*
+ * This method loads the image into whole_image during the first call on
+ * get_pixel_rows.  The get_pixel_rows pointer is then adjusted to call
+ * get_memory_row on subsequent calls.
+ */
+
+METHODDEF(JDIMENSION)
+preload_image (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  JDIMENSION row;
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+
+  /* Read the data into a virtual array in input-file row order. */
+  for (row = 0; row < cinfo->image_height; row++) {
+    if (progress != NULL) {
+      progress->pub.pass_counter = (long) row;
+      progress->pub.pass_limit = (long) cinfo->image_height;
+      (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+    }
+    source->pub.buffer = (*cinfo->mem->access_virt_sarray)
+      ((j_common_ptr) cinfo, source->whole_image, row, (JDIMENSION) 1, TRUE);
+    (*source->get_pixel_rows) (cinfo, sinfo);
+  }
+  if (progress != NULL)
+    progress->completed_extra_passes++;
+
+  /* Set up to read from the virtual array in unscrambled order */
+  source->pub.get_pixel_rows = get_memory_row;
+  source->current_row = 0;
+  /* And read the first row */
+  return get_memory_row(cinfo, sinfo);
+}
+
+
+/*
+ * Read the file header; return image size and component count.
+ */
+
+METHODDEF(void)
+start_input_tga (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  tga_source_ptr source = (tga_source_ptr) sinfo;
+  U_CHAR targaheader[18];
+  int idlen, cmaptype, subtype, flags, interlace_type, components;
+  unsigned int width, height, maplen;
+  boolean is_bottom_up;
+
+#define GET_2B(offset)	((unsigned int) UCH(targaheader[offset]) + \
+			 (((unsigned int) UCH(targaheader[offset+1])) << 8))
+
+  if (! ReadOK(source->pub.input_file, targaheader, 18))
+    ERREXIT(cinfo, JERR_INPUT_EOF);
+
+  /* Pretend "15-bit" pixels are 16-bit --- we ignore attribute bit anyway */
+  if (targaheader[16] == 15)
+    targaheader[16] = 16;
+
+  idlen = UCH(targaheader[0]);
+  cmaptype = UCH(targaheader[1]);
+  subtype = UCH(targaheader[2]);
+  maplen = GET_2B(5);
+  width = GET_2B(12);
+  height = GET_2B(14);
+  source->pixel_size = UCH(targaheader[16]) >> 3;
+  flags = UCH(targaheader[17]);	/* Image Descriptor byte */
+
+  is_bottom_up = ((flags & 0x20) == 0);	/* bit 5 set => top-down */
+  interlace_type = flags >> 6;	/* bits 6/7 are interlace code */
+
+  if (cmaptype > 1 ||		/* cmaptype must be 0 or 1 */
+      source->pixel_size < 1 || source->pixel_size > 4 ||
+      (UCH(targaheader[16]) & 7) != 0 || /* bits/pixel must be multiple of 8 */
+      interlace_type != 0)	/* currently don't allow interlaced image */
+    ERREXIT(cinfo, JERR_TGA_BADPARMS);
+  
+  if (subtype > 8) {
+    /* It's an RLE-coded file */
+    source->read_pixel = read_rle_pixel;
+    source->block_count = source->dup_pixel_count = 0;
+    subtype -= 8;
+  } else {
+    /* Non-RLE file */
+    source->read_pixel = read_non_rle_pixel;
+  }
+
+  /* Now should have subtype 1, 2, or 3 */
+  components = 3;		/* until proven different */
+  cinfo->in_color_space = JCS_RGB;
+
+  switch (subtype) {
+  case 1:			/* Colormapped image */
+    if (source->pixel_size == 1 && cmaptype == 1)
+      source->get_pixel_rows = get_8bit_row;
+    else
+      ERREXIT(cinfo, JERR_TGA_BADPARMS);
+    TRACEMS2(cinfo, 1, JTRC_TGA_MAPPED, width, height);
+    break;
+  case 2:			/* RGB image */
+    switch (source->pixel_size) {
+    case 2:
+      source->get_pixel_rows = get_16bit_row;
+      break;
+    case 3:
+      source->get_pixel_rows = get_24bit_row;
+      break;
+    case 4:
+      source->get_pixel_rows = get_32bit_row;
+      break;
+    default:
+      ERREXIT(cinfo, JERR_TGA_BADPARMS);
+      break;
+    }
+    TRACEMS2(cinfo, 1, JTRC_TGA, width, height);
+    break;
+  case 3:			/* Grayscale image */
+    components = 1;
+    cinfo->in_color_space = JCS_GRAYSCALE;
+    if (source->pixel_size == 1)
+      source->get_pixel_rows = get_8bit_gray_row;
+    else
+      ERREXIT(cinfo, JERR_TGA_BADPARMS);
+    TRACEMS2(cinfo, 1, JTRC_TGA_GRAY, width, height);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_TGA_BADPARMS);
+    break;
+  }
+
+  if (is_bottom_up) {
+    /* Create a virtual array to buffer the upside-down image. */
+    source->whole_image = (*cinfo->mem->request_virt_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+       (JDIMENSION) width * components, (JDIMENSION) height, (JDIMENSION) 1);
+    if (cinfo->progress != NULL) {
+      cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+      progress->total_extra_passes++; /* count file input as separate pass */
+    }
+    /* source->pub.buffer will point to the virtual array. */
+    source->pub.buffer_height = 1; /* in case anyone looks at it */
+    source->pub.get_pixel_rows = preload_image;
+  } else {
+    /* Don't need a virtual array, but do need a one-row input buffer. */
+    source->whole_image = NULL;
+    source->pub.buffer = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) width * components, (JDIMENSION) 1);
+    source->pub.buffer_height = 1;
+    source->pub.get_pixel_rows = source->get_pixel_rows;
+  }
+  
+  while (idlen--)		/* Throw away ID field */
+    (void) read_byte(source);
+
+  if (maplen > 0) {
+    if (maplen > 256 || GET_2B(3) != 0)
+      ERREXIT(cinfo, JERR_TGA_BADCMAP);
+    /* Allocate space to store the colormap */
+    source->colormap = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE, (JDIMENSION) maplen, (JDIMENSION) 3);
+    /* and read it from the file */
+    read_colormap(source, (int) maplen, UCH(targaheader[7]));
+  } else {
+    if (cmaptype)		/* but you promised a cmap! */
+      ERREXIT(cinfo, JERR_TGA_BADPARMS);
+    source->colormap = NULL;
+  }
+
+  cinfo->input_components = components;
+  cinfo->data_precision = 8;
+  cinfo->image_width = width;
+  cinfo->image_height = height;
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_input_tga (j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+  /* no work */
+}
+
+
+/*
+ * The module selection routine for Targa format input.
+ */
+
+GLOBAL(cjpeg_source_ptr)
+jinit_read_targa (j_compress_ptr cinfo)
+{
+  tga_source_ptr source;
+
+  /* Create module interface object */
+  source = (tga_source_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(tga_source_struct));
+  source->cinfo = cinfo;	/* make back link for subroutines */
+  /* Fill in method ptrs, except get_pixel_rows which start_input sets */
+  source->pub.start_input = start_input_tga;
+  source->pub.finish_input = finish_input_tga;
+
+  return (cjpeg_source_ptr) source;
+}
+
+#endif /* TARGA_SUPPORTED */
diff --git a/release/Description.plist.tmpl b/release/Description.plist.tmpl
new file mode 100644
index 0000000..4c5875c
--- /dev/null
+++ b/release/Description.plist.tmpl
@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>IFPkgDescriptionDescription</key>
+	<string>A SIMD-accelerated JPEG codec which provides both the libjpeg and TurboJPEG APIs</string>
+	<key>IFPkgDescriptionTitle</key>
+	<string>{__APPNAME}</string>
+	<key>IFPkgDescriptionVersion</key>
+	<string>{__VERSION}</string>
+</dict>
+</plist>
diff --git a/release/Info.plist.tmpl b/release/Info.plist.tmpl
new file mode 100755
index 0000000..a33daae
--- /dev/null
+++ b/release/Info.plist.tmpl
@@ -0,0 +1,44 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple Computer//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>CFBundleGetInfoString</key>
+	<string>{__VERSION}, The libjpeg-turbo Project</string>
+	<key>CFBundleIdentifier</key>
+	<string>com.libjpeg-turbo.libjpeg-turbo</string>
+	<key>CFBundleShortVersionString</key>
+	<string>{__VERSION}</string>
+	<key>IFMajorVersion</key>
+	<integer>1</integer>
+	<key>IFMinorVersion</key>
+	<integer>{__BUILD}</integer>
+	<key>IFPkgFlagAllowBackRev</key>
+	<false/>
+	<key>IFPkgFlagAuthorizationAction</key>
+	<string>AdminAuthorization</string>
+	<key>IFPkgFlagBackgroundAlignment</key>
+	<string>topleft</string>
+	<key>IFPkgFlagBackgroundScaling</key>
+	<string>none</string>
+	<key>IFPkgFlagDefaultLocation</key>
+	<string>/</string>
+	<key>IFPkgFlagFollowLinks</key>
+	<true/>
+	<key>IFPkgFlagInstallFat</key>
+	<false/>
+	<key>IFPkgFlagIsRequired</key>
+	<false/>
+	<key>IFPkgFlagOverwritePermissions</key>
+	<false/>
+	<key>IFPkgFlagRelocatable</key>
+	<false/>
+	<key>IFPkgFlagRestartAction</key>
+	<string>NoRestart</string>
+	<key>IFPkgFlagRootVolumeOnly</key>
+	<true/>
+	<key>IFPkgFlagUpdateInstalledLanguages</key>
+	<false/>
+	<key>IFPkgFormatVersion</key>
+	<real>0.10000000149011612</real>
+</dict>
+</plist>
diff --git a/release/License.rtf b/release/License.rtf
new file mode 100755
index 0000000..30d15a5
--- /dev/null
+++ b/release/License.rtf
@@ -0,0 +1,465 @@
+{\rtf1\mac\ansicpg10000\cocoartf824\cocoasubrtf470
+{\fonttbl\f0\fswiss\fcharset77 Helvetica;\f1\fnil\fcharset77 AndaleMono;\f2\froman\fcharset77 TimesNewRomanPSMT;
+}
+{\colortbl;\red255\green255\blue255;}
+\margl1440\margr1440\vieww9000\viewh8400\viewkind0
+\deftab720
+\pard\pardeftab720\qc
+
+\f0\fs24 \cf0 wxWindows Library Licence, Version 3.1\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\qc
+\cf0 Copyright (C) 1998-2005 Julian Smart, Robert Roebling et al\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\qc
+\cf0 Everyone is permitted to copy and distribute verbatim copies of this licence document, but changing it is not allowed.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\qc
+\cf0 WXWINDOWS LIBRARY LICENCE\uc0\u8232 TERMS AND CONDITIONS FOR COPYING,\u8232 DISTRIBUTION AND MODIFICATION\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 This library is free software; you can redistribute it and/or modify it under the terms of the GNU Library General Public Licence as published by the Free Software Foundation; either version 2 of the Licence, or (at your option) any later version.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0   \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Library General Public Licence for more details.
+\f2 \
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 You should have received a copy of the GNU Library General Public Licence along with this software, usually in a file named COPYING.LIB.  If not, write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
+\f1 \
+\
+
+\f0 EXCEPTION NOTICE
+\f1 \
+\
+
+\f0 1. As a special exception, the copyright holders of this library give permission for additional uses of the text contained in this release of the library as licenced under the wxWindows Library Licence, applying either version 3.1 of the Licence, or (at your option) any later version of the Licence as published by the copyright holders of version 3.1 of the Licence document.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 2. The exception is that you may use, copy, link, modify and distribute under your own terms, binary object code versions of works based on the Library.
+\f1 \
+\
+
+\f0 3. If you copy code from files distributed under the terms of the GNU General Public Licence or the GNU Library General Public Licence into a copy of this library, as this licence permits, the exception does not apply to the code that you add in this way.  To avoid misleading anyone as to the status of such modified files, you must delete this exception notice from such code and/or adjust the licensing conditions notice accordingly.
+\f1 \
+\
+
+\f0 4. If you write modifications of your own for this library, it is your choice whether to permit this exception to apply to your modifications.  If you do not wish that, you must delete the exception notice from such code and/or adjust the licensing conditions notice accordingly.
+\f1 \
+\
+\
+\
+\
+\
+\pard\pardeftab720\qc
+
+\f0 \cf0 GNU LESSER GENERAL PUBLIC LICENSE\uc0\u8232 Version 2.1, February 1999\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\qc
+\cf0 Copyright (C) 1991, 1999 Free Software Foundation, Inc.\uc0\u8232 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.\
+\
+[This is the first released version of the Lesser GPL.  It also counts as the successor of the GNU Library Public License, version 2, hence the version number 2.1.]
+\f1 \
+\
+\pard\pardeftab720\qc
+
+\f0 \cf0 Preamble\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 The licenses for most software are designed to take away your freedom to share and change it.  By contrast, the GNU General Public Licenses are intended to guarantee your freedom to share and change free software--to make sure the software is free for all its users.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 This license, the Lesser General Public License, applies to some specially designated software packages--typically libraries--of the Free Software Foundation and other authors who decide to use it.  You can use it too, but we suggest you first think carefully about whether this license or the ordinary General Public License is the better strategy to use in any particular case, based on the explanations below.
+\f1 \
+\
+
+\f0 When we speak of free software, we are referring to freedom of use, not price.  Our General Public Licenses are designed to make sure that you have the freedom to distribute copies of free software (and charge for this service if you wish); that you receive source code or can get it if you want it; that you can change the software and use pieces of it in new free programs; and that you are informed that you can do these things.
+\f1 \
+\
+
+\f0 To protect your rights, we need to make restrictions that forbid distributors to deny you these rights or to ask you to surrender these rights.  These restrictions translate to certain responsibilities for you if you distribute copies of the library or if you modify it.
+\f1 \
+\
+
+\f0 For example, if you distribute copies of the library, whether gratis or for a fee, you must give the recipients all the rights that we gave you.  You must make sure that they, too, receive or can get the source code.  If you link other code with the library, you must provide complete object files to the recipients, so that they can relink them with the library after making changes to the library and recompiling it.  And you must show them these terms so they know their rights.
+\f1 \
+\
+
+\f0 We protect your rights with a two-step method: (1) we copyright the library, and (2) we offer you this license, which gives you legal permission to copy, distribute and/or modify the library.
+\f1 \
+\
+
+\f0 To protect each distributor, we want to make it very clear that there is no warranty for the free library.  Also, if the library is modified by someone else and passed on, the recipients should know that what they have is not the original version, so that the original author's reputation will not be affected by problems that might be introduced by others.
+\f1 \
+\page 
+\f0 \
+Finally, software patents pose a constant threat to the existence of any free program.  We wish to make sure that a company cannot effectively restrict the users of a free program by obtaining a restrictive license from a patent holder.  Therefore, we insist that any patent license obtained for a version of the library must be consistent with the full freedom of use specified in this license.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 Most GNU software, including some libraries, is covered by the ordinary GNU General Public License.  This license, the GNU Lesser General Public License, applies to certain designated libraries, and is quite different from the ordinary General Public License.  We use this license for certain libraries in order to permit linking those libraries into non-free programs.
+\f1 \
+\
+
+\f0 When a program is linked with a library, whether statically or using a shared library, the combination of the two is legally speaking a combined work, a derivative of the original library.  The ordinary General Public License therefore permits such linking only if the entire combination fits its criteria of freedom.  The Lesser General Public License permits more lax criteria for linking other code with the library.
+\f1 \
+\
+
+\f0 We call this license the "Lesser" General Public License because it does Less to protect the user's freedom than the ordinary General Public License.  It also provides other free software developers Less of an advantage over competing non-free programs.  These disadvantages are the reason we use the ordinary General Public License for many libraries.  However, the Lesser license provides advantages in certain special circumstances.
+\f1 \
+\
+
+\f0 For example, on rare occasions, there may be a special need to encourage the widest possible use of a certain library, so that it becomes a de-facto standard.  To achieve this, non-free programs must be allowed to use the library.  A more frequent case is that a free library does the same job as widely used non-free libraries.  In this case, there is little to gain by limiting the free library to free software only, so we use the Lesser General Public License.
+\f1 \
+\
+
+\f0 In other cases, permission to use a particular library in non-free programs enables a greater number of people to use a large body of free software.  For example, permission to use the GNU C Library in non-free programs enables many more people to use the whole GNU operating system, as well as its variant, the GNU/Linux operating system.
+\f1 \
+\
+
+\f0 Although the Lesser General Public License is Less protective of the users' freedom, it does ensure that the user of a program that is linked with the Library has the freedom and the wherewithal to run that program using a modified version of the Library.
+\f1 \
+\
+
+\f0 The precise terms and conditions for copying, distribution and modification follow.  Pay close attention to the difference between a "work based on the library" and a "work that uses the library".  The former contains code derived from the library, whereas the latter must be combined with the library in order to run.
+\f1 \
+\page 
+\f0 \
+\pard\pardeftab720\qc
+\cf0 GNU LESSER GENERAL PUBLIC LICENSE\uc0\u8232 TERMS AND CONDITIONS FOR COPYING,\u8232 DISTRIBUTION AND MODIFICATION\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 0. This License Agreement applies to any software library or other program which contains a notice placed by the copyright holder or other authorized party saying it may be distributed under the terms of this Lesser General Public License (also called "this License"). Each licensee is addressed as "you".\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 A "library" means a collection of software functions and/or data prepared so as to be conveniently linked with application programs (which use some of those functions and data) to form executables.
+\f1 \
+\
+
+\f0 The "Library", below, refers to any such software library or work which has been distributed under these terms.  A "work based on the Library" means either the Library or any derivative work under copyright law: that is to say, a work containing the Library or a portion of it, either verbatim or with modifications and/or translated straightforwardly into another language.  (Hereinafter, translation is included without limitation in the term "modification".)
+\f1 \
+\
+
+\f0 "Source code" for a work means the preferred form of the work for making modifications to it.  For a library, complete source code means all the source code for all modules it contains, plus any associated interface definition files, plus the scripts used to control compilation and installation of the library.
+\f1 \
+\
+
+\f0 Activities other than copying, distribution and modification are not covered by this License; they are outside its scope.  The act of running a program using the Library is not restricted, and output from such a program is covered only if its contents constitute a work based on the Library (independent of the use of the Library in a tool for writing it).  Whether that is true depends on what the Library does and what the program that uses the Library does.
+\f1 \
+  \
+
+\f0 1. You may copy and distribute verbatim copies of the Library's complete source code as you receive it, in any medium, provided that you conspicuously and appropriately publish on each copy an appropriate copyright notice and disclaimer of warranty; keep intact all the notices that refer to this License and to the absence of any warranty; and distribute a copy of this License along with the Library.
+\f2 \
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 You may charge a fee for the physical act of transferring a copy, and you may at your option offer warranty protection in exchange for a fee.
+\f1 \
+\page 
+\f0 \
+2. You may modify your copy or copies of the Library or any portion of it, thus forming a work based on the Library, and copy and distribute such modifications or work under the terms of Section 1 above, provided that you also meet all of these conditions:\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+
+\f0 \cf0 a) The modified work must itself be a software library.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 b) You must cause the files modified to carry prominent notices stating that you changed the files and the date of any change.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 c) You must cause the whole of the work to be licensed at no charge to all third parties under the terms of this License.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 d) If a facility in the modified Library refers to a function or a table of data to be supplied by an application program that uses the facility, other than as an argument passed when the facility is invoked, then you must make a good faith effort to ensure that, in the event an application does not supply such function or table, the facility still operates, and performs whatever part of its purpose remains meaningful.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 (For example, a function in a library to compute square roots has a purpose that is entirely well-defined independent of the application.  Therefore, Subsection 2d requires that any application-supplied function or table used by this function must be optional: if the application does not supply it, the square root function must still compute square roots.)
+\f1 \
+\
+
+\f0 These requirements apply to the modified work as a whole.  If identifiable sections of that work are not derived from the Library, and can be reasonably considered independent and separate works in themselves, then this License, and its terms, do not apply to those sections when you distribute them as separate works.  But when you distribute the same sections as part of a whole which is a work based on the Library, the distribution of the whole must be on the terms of this License, whose permissions for other licensees extend to the entire whole, and thus to each and every part regardless of who wrote it.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 Thus, it is not the intent of this section to claim rights or contest your rights to work written entirely by you; rather, the intent is to exercise the right to control the distribution of derivative or collective works based on the Library.
+\f1 \
+\
+
+\f0 In addition, mere aggregation of another work not based on the Library with the Library (or with a work based on the Library) on a volume of a storage or distribution medium does not bring the other work under the scope of this License.
+\f1 \
+\
+
+\f0 3. You may opt to apply the terms of the ordinary GNU General Public License instead of this License to a given copy of the Library.  To do this, you must alter all the notices that refer to this License, so that they refer to the ordinary GNU General Public License, version 2, instead of to this License.  (If a newer version than version 2 of the ordinary GNU General Public License has appeared, then you can specify that version instead if you wish.)  Do not make any other change in these notices.
+\f1 \
+\page 
+\f0 \
+Once this change is made in a given copy, it is irreversible for that copy, so the ordinary GNU General Public License applies to all subsequent copies and derivative works made from that copy.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 This option is useful when you wish to copy part of the code of the Library into a program that is not a library.
+\f1 \
+\
+
+\f0 4. You may copy and distribute the Library (or a portion or derivative of it, under Section 2) in object code or executable form under the terms of Sections 1 and 2 above provided that you accompany it with the complete corresponding machine-readable source code, which must be distributed under the terms of Sections 1 and 2 above on a
+\f1 \
+
+\f0 medium customarily used for software interchange.
+\f1 \
+\
+
+\f0 If distribution of object code is made by offering access to copy from a designated place, then offering equivalent access to copy the source code from the same place satisfies the requirement to distribute the source code, even though third parties are not
+\f1 \
+
+\f0 compelled to copy the source along with the object code.
+\f1 \
+\
+
+\f0 5. A program that contains no derivative of any portion of the Library, but is designed to work with the Library by being compiled or linked with it, is called a "work that uses the Library".  Such a work, in isolation, is not a derivative work of the Library, and therefore falls outside the scope of this License.
+\f1 \
+\
+
+\f0 However, linking a "work that uses the Library" with the Library creates an executable that is a derivative of the Library (because it contains portions of the Library), rather than a "work that uses the library".  The executable is therefore covered by this License. Section 6 states terms for distribution of such executables.
+\f1 \
+\
+
+\f0 When a "work that uses the Library" uses material from a header file that is part of the Library, the object code for the work may be a derivative work of the Library even though the source code is not. Whether this is true is especially significant if the work can be linked without the Library, or if the work is itself a library.  The threshold for this to be true is not precisely defined by law.
+\f1 \
+\
+
+\f0 If such an object file uses only numerical parameters, data structure layouts and accessors, and small macros and small inline functions (ten lines or less in length), then the use of the object file is unrestricted, regardless of whether it is legally a derivative
+\f1 \
+
+\f0 work.  (Executables containing this object code plus portions of the Library will still fall under Section 6.)
+\f1 \
+\
+
+\f0 Otherwise, if the work is a derivative of the Library, you may distribute the object code for the work under the terms of Section 6. Any executables containing that work also fall under Section 6, whether or not they are linked directly with the Library itself.
+\f1 \
+\page 
+\f0 \
+6. As an exception to the Sections above, you may also combine or link a "work that uses the Library" with the Library to produce a work containing portions of the Library, and distribute that work under terms of your choice, provided that the terms permit modification of the work for the customer's own use and reverse engineering for debugging such modifications.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 You must give prominent notice with each copy of the work that the Library is used in it and that the Library and its use are covered by this License.  You must supply a copy of this License.  If the work during execution displays copyright notices, you must include the copyright notice for the Library among them, as well as a reference directing the user to the copy of this License.  Also, you must do one of these things:
+\f1 \
+\
+\pard\pardeftab720\li720\ri0\ql\qnatural
+
+\f0 \cf0 a) Accompany the work with the complete corresponding machine-readable source code for the Library including whatever changes were used in the work (which must be distributed under Sections 1 and 2 above); and, if the work is an executable linked with the Library, with the complete machine-readable "work that uses the Library", as object code and/or source code, so that the user can modify the Library and then relink to produce a modified executable containing the modified Library.  (It is understood that the user who changes the contents of definitions files in the Library will not necessarily be able to recompile the application to use the modified definitions.)\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 b) Use a suitable shared library mechanism for linking with the Library.  A suitable mechanism is one that (1) uses at run time a copy of the library already present on the user's computer system, rather than copying library functions into the executable, and (2) will operate properly with a modified version of the library, if the user installs one, as long as the modified version is interface-compatible with the version that the work was made with.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 c) Accompany the work with a written offer, valid for at least three years, to give the same user the materials specified in Subsection 6a, above, for a charge no more than the cost of performing this distribution.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 d) If distribution of the work is made by offering access to copy from a designated place, offer equivalent access to copy the above specified materials from the same place.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 e) Verify that the user has already received a copy of these materials or that you have already sent this user a copy.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+For an executable, the required form of the "work that uses the Library" must include any data and utility programs needed for reproducing the executable from it.  However, as a special exception, the materials to be distributed need not include anything that is normally distributed (in either source or binary form) with the major components (compiler, kernel, and so on) of the operating system on which the executable runs, unless that component itself accompanies the executable.
+\f1 \
+\
+
+\f0 It may happen that this requirement contradicts the license restrictions of other proprietary libraries that do not normally accompany the operating system.  Such a contradiction means you cannot use both them and the Library together in an executable that you distribute.
+\f1 \
+\page 
+\f0 \
+7. You may place library facilities that are a work based on the Library side-by-side in a single library together with other library facilities not covered by this License, and distribute such a combined library, provided that the separate distribution of the work based on the Library and of the other library facilities is otherwise permitted, and provided that you do these two things:\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+
+\f0 \cf0 a) Accompany the combined library with a copy of the same work based on the Library, uncombined with any other library facilities.  This must be distributed under the terms of the Sections above.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 b) Give prominent notice with the combined library of the fact that part of it is a work based on the Library, and explaining where to find the accompanying uncombined form of the same work.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+8. You may not copy, modify, sublicense, link with, or distribute the Library except as expressly provided under this License.  Any attempt otherwise to copy, modify, sublicense, link with, or distribute the Library is void, and will automatically terminate your rights under this License.  However, parties who have received copies, or rights, from you under this License will not have their licenses terminated so long as such parties remain in full compliance.
+\f1 \
+\
+
+\f0 9. You are not required to accept this License, since you have not signed it.  However, nothing else grants you permission to modify or distribute the Library or its derivative works.  These actions are prohibited by law if you do not accept this License.  Therefore, by modifying or distributing the Library (or any work based on the Library), you indicate your acceptance of this License to do so, and all its terms and conditions for copying, distributing or modifying the Library or works based on it.
+\f1 \
+\
+
+\f0 10. Each time you redistribute the Library (or any work based on the Library), the recipient automatically receives a license from the original licensor to copy, distribute, link with or modify the Library subject to these terms and conditions.  You may not impose any further restrictions on the recipients' exercise of the rights granted herein. You are not responsible for enforcing compliance by third parties with this License.
+\f1 \
+\page 
+\f0 \
+11. If, as a consequence of a court judgment or allegation of patent infringement or for any other reason (not limited to patent issues), conditions are imposed on you (whether by court order, agreement or otherwise) that contradict the conditions of this License, they do not excuse you from the conditions of this License.  If you cannot distribute so as to satisfy simultaneously your obligations under this License and any other pertinent obligations, then as a consequence you may not distribute the Library at all.  For example, if a patent license would not permit royalty-free redistribution of the Library by all those who receive copies directly or indirectly through you, then the only way you could satisfy both it and this License would be to refrain entirely from distribution of the Library.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 If any portion of this section is held invalid or unenforceable under any particular circumstance, the balance of the section is intended to apply, and the section as a whole is intended to apply in other circumstances.
+\f1 \
+\
+
+\f0 It is not the purpose of this section to induce you to infringe any patents or other property right claims or to contest validity of any such claims; this section has the sole purpose of protecting the integrity of the free software distribution system which is
+\f1 \
+
+\f0 implemented by public license practices.  Many people have made generous contributions to the wide range of software distributed through that system in reliance on consistent application of that system; it is up to the author/donor to decide if he or she is willing to distribute software through any other system and a licensee cannot impose that choice.
+\f1 \
+\
+
+\f0 This section is intended to make thoroughly clear what is believed to be a consequence of the rest of this License.
+\f1 \
+\
+
+\f0 12. If the distribution and/or use of the Library is restricted in certain countries either by patents or by copyrighted interfaces, the original copyright holder who places the Library under this License may add an explicit geographical distribution limitation excluding those countries, so that distribution is permitted only in or among countries not thus excluded.  In such case, this License incorporates the limitation as if written in the body of this License.
+\f1 \
+\
+
+\f0 13. The Free Software Foundation may publish revised and/or new versions of the Lesser General Public License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns.
+\f1 \
+\
+
+\f0 Each version is given a distinguishing version number.  If the Library specifies a version number of this License which applies to it and "any later version", you have the option of following the terms and conditions either of that version or of any later version published by the Free Software Foundation.  If the Library does not specify a license version number, you may choose any version ever published by the Free Software Foundation.
+\f1 \
+\page 
+\f0 \
+14. If you wish to incorporate parts of the Library into other free programs whose distribution conditions are incompatible with these, write to the author to ask for permission.  For software which is copyrighted by the Free Software Foundation, write to the Free Software Foundation; we sometimes make exceptions for this.  Our decision will be guided by the two goals of preserving the free status of all derivatives of our free software and of promoting the sharing and reuse of software generally.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\qc
+
+\f0 \cf0 NO WARRANTY\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 15. BECAUSE THE LIBRARY IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY FOR THE LIBRARY, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES PROVIDE THE LIBRARY "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE LIBRARY IS WITH YOU.  SHOULD THE LIBRARY PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 16. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR REDISTRIBUTE THE LIBRARY AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE LIBRARY (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A FAILURE OF THE LIBRARY TO OPERATE WITH ANY OTHER SOFTWARE), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
+\f1 \
+\
+\pard\pardeftab720\qc
+
+\f0 \cf0 END OF TERMS AND CONDITIONS\
+\page \
+How to Apply These Terms to Your New Libraries\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 If you develop a new library, and you want it to be of the greatest possible use to the public, we recommend making it free software that everyone can redistribute and change.  You can do so by permitting redistribution under these terms (or, alternatively, under the terms of the ordinary General Public License).\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 To apply these terms, attach the following notices to the library.  It is safest to attach them to the start of each source file to most effectively convey the exclusion of warranty; and each file should have at least the "copyright" line and a pointer to where the full notice is found.
+\f1 \
+\
+\pard\pardeftab720\li720\ri0\ql\qnatural
+
+\f0 \cf0 <one line to give the library's name and a brief idea of what it does.>\uc0\u8232 Copyright (C) <year>  <name of author>\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either version 2.1 of the License, or (at your option) any later version.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for more details.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 Also add information on how to contact you by electronic and paper mail.\
+\pard\pardeftab720\ql\qnatural
+
+\f1 \cf0 \
+\pard\pardeftab720\ql\qnatural
+
+\f0 \cf0 You should also get your employer (if you work as a programmer) or your school, if any, to sign a "copyright disclaimer" for the library, if necessary.  Here is a sample; alter the names:
+\f1 \
+\
+\pard\pardeftab720\li720\ri0\ql\qnatural
+
+\f0 \cf0 Yoyodyne, Inc., hereby disclaims all copyright interest in the library `Frob' (a library for tweaking knobs) written by James Random Hacker.\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+\pard\pardeftab720\li720\ri0\ql\qnatural
+\cf0 <signature of Ty Coon>, 1 April 1990\uc0\u8232 Ty Coon, President of Vice\
+\pard\pardeftab720\ql\qnatural
+\cf0 \
+That's all there is to it!
+\f1 \
+\
+}
\ No newline at end of file
diff --git a/release/ReadMe.rtf b/release/ReadMe.rtf
new file mode 100644
index 0000000..b7f156e
--- /dev/null
+++ b/release/ReadMe.rtf
@@ -0,0 +1,11 @@
+{\rtf1\ansi\ansicpg1252\cocoartf1038\cocoasubrtf250
+{\fonttbl\f0\fswiss\fcharset0 Helvetica;}
+{\colortbl;\red255\green255\blue255;}
+\margl1440\margr1440\vieww15200\viewh9600\viewkind0
+\deftab720
+\pard\pardeftab720\ql\qnatural
+
+\f0\fs24 \cf0 libjpeg-turbo is a high-speed version of libjpeg for x86 and x86-64 processors which uses SIMD instructions (MMX, SSE2, etc.) to accelerate baseline JPEG compression and decompression.  libjpeg-turbo is generally 2-4x as fast as the unmodified version of libjpeg, all else being equal.  libjpeg-turbo also includes a wrapper library for the TurboJPEG API used by VirtualGL and TurboVNC.\
+\
+libjpeg-turbo was originally based on libjpeg/SIMD by Miyasaka Masaru, but the TigerVNC and VirtualGL projects made numerous enhancements to the codec, including improved support for Mac OS X, 64-bit support, support for 32-bit and big endian pixel formats, accelerated Huffman encoding/decoding, and various bug fixes.  The goal was to produce a fully open source codec that could replace the partially closed source TurboJPEG/IPP codec used by VirtualGL and TurboVNC.  libjpeg-turbo generally performs in the range of 80-120% of TurboJPEG/IPP.  It is faster in some areas but slower in others.\
+}
\ No newline at end of file
diff --git a/release/Welcome.rtf b/release/Welcome.rtf
new file mode 100755
index 0000000..c047f1c
--- /dev/null
+++ b/release/Welcome.rtf
@@ -0,0 +1,17 @@
+{\rtf1\ansi\ansicpg1252\cocoartf1038\cocoasubrtf250
+{\fonttbl\f0\fswiss\fcharset0 Helvetica;\f1\fmodern\fcharset0 CourierNewPSMT;}
+{\colortbl;\red255\green255\blue255;}
+\margl1440\margr1440\vieww9000\viewh8400\viewkind0
+\deftab720
+\pard\pardeftab720\ql\qnatural
+
+\f0\fs24 \cf0 This installer will install the libjpeg-turbo SDK and run-time libraries onto your computer so that you can use libjpeg-turbo to build new applications or accelerate existing ones.  To remove the libjpeg-turbo package, you can either use the "Uninstall libjpeg-turbo" application, which can be found in the same archive as this package, or you can run\
+\
+
+\f1   /opt/libjpeg-turbo/bin/uninstall\
+
+\f0 \
+from the command line.\
+\
+If TurboJPEG/IPP is currently installed, then it must be removed prior to installing libjpeg-turbo.\
+}
\ No newline at end of file
diff --git a/release/copyright b/release/copyright
new file mode 100755
index 0000000..13db231
--- /dev/null
+++ b/release/copyright
@@ -0,0 +1 @@
+Use is subject to the terms of the wxWindows Library License, version 3.1
diff --git a/release/deb-control.tmpl b/release/deb-control.tmpl
new file mode 100644
index 0000000..5a83dc3
--- /dev/null
+++ b/release/deb-control.tmpl
@@ -0,0 +1,23 @@
+Package: {__PKGNAME}
+Version: {__VERSION}-{__BUILD}
+Section: misc
+Priority: optional
+Architecture: {__ARCH}
+Essential: no
+Maintainer: The libjpeg-turbo Project [http://libjpeg-turbo.VirtualGL.org]
+Description: A toolkit for displaying OpenGL applications to thin clients
+ libjpeg-turbo is a high-speed version of libjpeg for x86 and x86-64 processors
+ which uses SIMD instructions (MMX, SSE2, etc.) to accelerate baseline JPEG
+ compression and decompression.  libjpeg-turbo is generally 2-4x as fast
+ as the unmodified version of libjpeg, all else being equal.  libjpeg-turbo
+ also includes a wrapper library for the TurboJPEG API used by VirtualGL and
+ TurboVNC.
+ .
+ libjpeg-turbo was originally based on libjpeg/SIMD by Miyasaka Masaru, but
+ the TigerVNC and VirtualGL projects made numerous enhancements to the codec,
+ including improved support for Mac OS X, 64-bit support, support for 32-bit
+ and big endian pixel formats, accelerated Huffman encoding/decoding, and
+ various bug fixes.  The goal was to produce a fully open source codec that
+ could replace the partially closed source TurboJPEG/IPP codec used by
+ VirtualGL and TurboVNC.  libjpeg-turbo generally performs in the range of
+ 80-120% of TurboJPEG/IPP.  It is faster in some areas but slower in others.
diff --git a/release/libjpeg-turbo.nsi.in b/release/libjpeg-turbo.nsi.in
new file mode 100755
index 0000000..39b85d2
--- /dev/null
+++ b/release/libjpeg-turbo.nsi.in
@@ -0,0 +1,126 @@
+!include x64.nsh
+Name "@CMAKE_PROJECT_NAME@ SDK for @INST_PLATFORM@"
+OutFile "@CMAKE_BINARY_DIR@\${BUILDDIR}@INST_NAME@.exe"
+InstallDir c:\@INST_NAME@
+
+SetCompressor bzip2
+
+Page directory
+Page instfiles
+
+UninstPage uninstConfirm
+UninstPage instfiles
+
+Section "@CMAKE_PROJECT_NAME@ SDK for @INST_PLATFORM@ (required)"
+!ifdef WIN64
+	${If} ${RunningX64}
+	${DisableX64FSRedirection}
+	${Endif}
+!endif
+	SectionIn RO
+!ifdef GCC
+	IfFileExists $SYSDIR/libturbojpeg.dll exists 0
+!else
+	IfFileExists $SYSDIR/turbojpeg.dll exists 0
+!endif
+	goto notexists
+	exists:
+!ifdef GCC
+	MessageBox MB_OK "An existing version of the @CMAKE_PROJECT_NAME@ SDK for @INST_PLATFORM@ is already installed.  Please uninstall it first."
+!else
+	MessageBox MB_OK "An existing version of the @CMAKE_PROJECT_NAME@ SDK for @INST_PLATFORM@ or the TurboJPEG SDK is already installed.  Please uninstall it first."
+!endif
+	quit
+
+	notexists:
+	SetOutPath $SYSDIR
+!ifdef GCC
+	File "@CMAKE_BINARY_DIR@\libturbojpeg.dll"
+!else
+	File "@CMAKE_BINARY_DIR@\${BUILDDIR}turbojpeg.dll"
+!endif
+	SetOutPath $INSTDIR\bin
+!ifdef GCC
+	File "/oname=libjpeg-@DLL_VERSION@.dll" "@CMAKE_BINARY_DIR@\sharedlib\libjpeg-*.dll" 
+!else
+	File "@CMAKE_BINARY_DIR@\sharedlib\${BUILDDIR}jpeg@DLL_VERSION@.dll"
+!endif
+	SetOutPath $INSTDIR\lib
+!ifdef GCC
+	File "@CMAKE_BINARY_DIR@\libturbojpeg.dll.a"
+	File "@CMAKE_BINARY_DIR@\libturbojpeg.a"
+	File "@CMAKE_BINARY_DIR@\sharedlib\libjpeg.dll.a"
+	File "@CMAKE_BINARY_DIR@\libjpeg.a"
+!else
+	File "@CMAKE_BINARY_DIR@\${BUILDDIR}turbojpeg.lib"
+	File "@CMAKE_BINARY_DIR@\${BUILDDIR}turbojpeg-static.lib"
+	File "@CMAKE_BINARY_DIR@\sharedlib\${BUILDDIR}jpeg.lib"
+	File "@CMAKE_BINARY_DIR@\${BUILDDIR}jpeg-static.lib"
+!endif
+	SetOutPath $INSTDIR\include
+	File "@CMAKE_BINARY_DIR@\jconfig.h"
+	File "@CMAKE_SOURCE_DIR@\jerror.h"
+	File "@CMAKE_SOURCE_DIR@\jmorecfg.h"
+	File "@CMAKE_SOURCE_DIR@\jpeglib.h"
+	File "@CMAKE_SOURCE_DIR@\turbojpeg.h"
+	SetOutPath $INSTDIR
+	File "@CMAKE_SOURCE_DIR@\README"
+	File "@CMAKE_SOURCE_DIR@\README-turbo.txt"
+	File "@CMAKE_SOURCE_DIR@\libjpeg.txt"
+	File "@CMAKE_SOURCE_DIR@\LGPL.txt"
+	File "@CMAKE_SOURCE_DIR@\LICENSE.txt"
+
+	WriteRegStr HKLM "SOFTWARE\@INST_NAME@ @VERSION@" "Install_Dir" "$INSTDIR"
+
+	WriteRegStr HKLM "Software\Microsoft\Windows\CurrentVersion\Uninstall\@INST_NAME@ @VERSION@" "DisplayName" "@CMAKE_PROJECT_NAME@ SDK v@VERSION@ for @INST_PLATFORM@"
+	WriteRegStr HKLM "Software\Microsoft\Windows\CurrentVersion\Uninstall\@INST_NAME@ @VERSION@" "UninstallString" '"$INSTDIR\uninstall_@VERSION@.exe"'
+	WriteRegDWORD HKLM "Software\Microsoft\Windows\CurrentVersion\Uninstall\@INST_NAME@ @VERSION@" "NoModify" 1
+	WriteRegDWORD HKLM "Software\Microsoft\Windows\CurrentVersion\Uninstall\@INST_NAME@ @VERSION@" "NoRepair" 1
+	WriteUninstaller "uninstall_@VERSION@.exe"
+SectionEnd
+
+Section "Uninstall"
+!ifdef WIN64
+	${If} ${RunningX64}
+	${DisableX64FSRedirection}
+	${Endif}
+!endif
+
+	SetShellVarContext all
+
+	DeleteRegKey HKLM "Software\Microsoft\Windows\CurrentVersion\Uninstall\@INST_NAME@ @VERSION@"
+	DeleteRegKey HKLM "SOFTWARE\@INST_NAME@ @VERSION@"
+
+!ifdef GCC
+	Delete $INSTDIR\bin\libjpeg-@DLL_VERSION@.dll
+	Delete $SYSDIR\libturbojpeg.dll
+	Delete $INSTDIR\lib\libturbojpeg.dll.a"
+	Delete $INSTDIR\lib\libturbojpeg.a"
+	Delete $INSTDIR\lib\libjpeg.dll.a"
+	Delete $INSTDIR\lib\libjpeg.a"
+!else
+	Delete $INSTDIR\bin\jpeg@DLL_VERSION@.dll
+	Delete $SYSDIR\turbojpeg.dll
+	Delete $INSTDIR\lib\jpeg.lib
+	Delete $INSTDIR\lib\jpeg-static.lib
+	Delete $INSTDIR\lib\turbojpeg.lib
+	Delete $INSTDIR\lib\turbojpeg-static.lib
+!endif
+	Delete $INSTDIR\include\jconfig.h"
+	Delete $INSTDIR\include\jerror.h"
+	Delete $INSTDIR\include\jmorecfg.h"
+	Delete $INSTDIR\include\jpeglib.h"
+	Delete $INSTDIR\include\turbojpeg.h"
+	Delete $INSTDIR\uninstall_@VERSION@.exe
+	Delete $INSTDIR\README
+	Delete $INSTDIR\README-turbo.txt
+	Delete $INSTDIR\libjpeg.txt
+	Delete $INSTDIR\LGPL.txt
+	Delete $INSTDIR\LICENSE.txt
+
+	RMDir "$INSTDIR\include"
+	RMDir "$INSTDIR\lib"
+	RMDir "$INSTDIR\bin"
+	RMDir "$INSTDIR"
+
+SectionEnd
diff --git a/release/libjpeg-turbo.spec.in b/release/libjpeg-turbo.spec.in
new file mode 100644
index 0000000..bbf2879
--- /dev/null
+++ b/release/libjpeg-turbo.spec.in
@@ -0,0 +1,107 @@
+%ifarch x86_64
+%define __lib lib64
+%else
+%define __lib lib
+%endif
+
+Summary: A SIMD-accelerated JPEG codec which provides both the libjpeg and TurboJPEG APIs
+Name: @PACKAGE_NAME@
+Version: @VERSION@
+Vendor: The libjpeg-turbo Project
+URL: http://libjpeg-turbo.virtualgl.org
+Group: System Environment/Libraries
+#-->Source0: http://prdownloads.sourceforge.net/libjpeg-turbo/libjpeg-turbo-%{version}.tar.gz
+Release: @BUILD@
+License: wxWindows Library License, v3.1
+BuildRoot: %{_blddir}/%{name}-buildroot-%{version}-%{release}
+Prereq: /sbin/ldconfig
+Provides: %{name} = %{version}-%{release}, turbojpeg = 2.00
+Obsoletes: turbojpeg
+
+%description
+libjpeg-turbo is a high-speed version of libjpeg for x86 and x86-64 processors
+which uses SIMD instructions (MMX, SSE2, etc.) to accelerate baseline JPEG
+compression and decompression.  libjpeg-turbo is generally 2-4x as fast
+as the unmodified version of libjpeg, all else being equal.  libjpeg-turbo also
+includes a wrapper library for the TurboJPEG API used by VirtualGL and
+TurboVNC.
+
+libjpeg-turbo was originally based on libjpeg/SIMD by Miyasaka Masaru, but
+the TigerVNC and VirtualGL projects made numerous enhancements to the codec,
+including improved support for Mac OS X, 64-bit support, support for 32-bit
+and big endian pixel formats, accelerated Huffman encoding/decoding, and
+various bug fixes.  The goal was to produce a fully open source codec that
+could replace the partially closed source TurboJPEG/IPP codec used by VirtualGL
+and TurboVNC.  libjpeg-turbo generally performs in the range of 80-120% of
+TurboJPEG/IPP.  It is faster in some areas but slower in others.
+
+#-->%prep
+#-->%setup -q
+
+#-->%build
+#-->configure libdir=/opt/%{name}/%{__lib} mandir=/opt/%{name}/man JPEG_LIB_VERSION=@JPEG_LIB_VERSION@ SO_MAJOR_VERSION=@SO_MAJOR_VERSION@ SO_MINOR_VERSION=@SO_MINOR_VERSION@ --with-pic
+#-->make DESTDIR=$RPM_BUILD_ROOT libdir=/opt/%{name}/%{__lib} mandir=/opt/%{name}/man
+
+%install
+
+rm -rf $RPM_BUILD_ROOT
+make install DESTDIR=$RPM_BUILD_ROOT libdir=/opt/%{name}/%{__lib} mandir=/opt/%{name}/man
+rm -f $RPM_BUILD_ROOT/opt/%{name}/%{__lib}/*.la
+mkdir -p $RPM_BUILD_ROOT/usr/%{__lib}
+mv $RPM_BUILD_ROOT/opt/%{name}/%{__lib}/libturbojpeg.* $RPM_BUILD_ROOT/usr/%{__lib}
+/sbin/ldconfig -n $RPM_BUILD_ROOT/opt/%{name}/%{__lib}
+/sbin/ldconfig -n $RPM_BUILD_ROOT/usr/%{__lib}
+mkdir -p $RPM_BUILD_ROOT/usr/include
+mv $RPM_BUILD_ROOT/opt/%{name}/include/turbojpeg.h $RPM_BUILD_ROOT/usr/include
+ln -fs /usr/include/turbojpeg.h $RPM_BUILD_ROOT/opt/%{name}/include/
+ln -fs /usr/%{__lib}/libturbojpeg.a $RPM_BUILD_ROOT/opt/%{name}/%{__lib}/
+%ifarch x86_64
+%else
+ln -fs %{__lib} $RPM_BUILD_ROOT/opt/%{name}/lib32
+%endif
+
+%post -p /sbin/ldconfig
+
+%postun -p /sbin/ldconfig
+
+%clean
+rm -rf $RPM_BUILD_ROOT
+
+%files
+%defattr(-,root,root)
+%doc %{_srcdir}/README-turbo.txt %{_srcdir}/README %{_srcdir}/libjpeg.txt %{_srcdir}/usage.txt %{_srcdir}/LICENSE.txt %{_srcdir}/LGPL.txt
+%dir /opt/%{name}
+%dir /opt/%{name}/bin
+/opt/%{name}/bin/cjpeg
+/opt/%{name}/bin/djpeg
+/opt/%{name}/bin/jpegtran
+/opt/%{name}/bin/rdjpgcom
+/opt/%{name}/bin/wrjpgcom
+%dir /opt/%{name}/%{__lib}
+%ifarch x86_64
+%else
+/opt/%{name}/lib32
+%endif
+/opt/%{name}/%{__lib}/libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@
+/opt/%{name}/%{__lib}/libjpeg.so.@SO_MAJOR_VERSION@
+/opt/%{name}/%{__lib}/libjpeg.so
+/opt/%{name}/%{__lib}/libjpeg.a
+/opt/%{name}/%{__lib}/libturbojpeg.a
+/usr/%{__lib}/libturbojpeg.so
+/usr/%{__lib}/libturbojpeg.a
+/usr/include/turbojpeg.h
+%dir /opt/%{name}/include
+/opt/%{name}/include/jconfig.h
+/opt/%{name}/include/jerror.h
+/opt/%{name}/include/jmorecfg.h
+/opt/%{name}/include/jpeglib.h
+/opt/%{name}/include/turbojpeg.h
+%dir /opt/%{name}/man
+%dir /opt/%{name}/man/man1
+/opt/%{name}/man/man1/cjpeg.1*
+/opt/%{name}/man/man1/djpeg.1*
+/opt/%{name}/man/man1/jpegtran.1*
+/opt/%{name}/man/man1/rdjpgcom.1*
+/opt/%{name}/man/man1/wrjpgcom.1*
+
+%changelog
diff --git a/release/makecygwinpkg b/release/makecygwinpkg
new file mode 100755
index 0000000..3ed1a40
--- /dev/null
+++ b/release/makecygwinpkg
@@ -0,0 +1,49 @@
+#!/bin/sh
+
+set -u
+set -e
+trap onexit INT
+trap onexit TERM
+trap onexit EXIT
+
+TMPDIR=
+
+onexit()
+{
+	if [ ! "$TMPDIR" = "" ]; then
+		rm -rf $TMPDIR
+	fi
+}
+
+usage()
+{
+	echo "$0 <package name> <version> <source dir.>"
+	exit 1
+}
+
+if [ $# -lt 3 ]; then usage $0; fi
+PACKAGE_NAME=$1
+VERSION=$2
+SRCDIR=$3
+
+umask 022
+rm -f $PACKAGE_NAME-$VERSION-cygwin.tar.bz2
+TMPDIR=`mktemp -d /tmp/ljtbuild.XXXXXX`
+__PWD=`pwd`
+make install DESTDIR=$TMPDIR/pkg mandir=/opt/$PACKAGE_NAME/man
+rm $TMPDIR/pkg/opt/$PACKAGE_NAME/lib/*.la
+DOCDIR=$TMPDIR/pkg/usr/share/doc/$PACKAGE_NAME-$VERSION
+mkdir -p $DOCDIR
+install -m 644 $SRCDIR/README-turbo.txt $DOCDIR
+install -m 644 $SRCDIR/README $DOCDIR
+install -m 644 $SRCDIR/libjpeg.txt $DOCDIR
+install -m 644 $SRCDIR/usage.txt $DOCDIR
+install -m 644 $SRCDIR/LICENSE.txt $DOCDIR
+install -m 644 $SRCDIR/LGPL.txt $DOCDIR
+ln -fs lib $TMPDIR/pkg/opt/$PACKAGE_NAME/lib32
+cd $TMPDIR/pkg
+tar cfj ../$PACKAGE_NAME-$VERSION-cygwin.tar.bz2 *
+cd $__PWD
+mv $TMPDIR/*.tar.bz2 .
+
+exit 0
diff --git a/release/makedpkg b/release/makedpkg
new file mode 100644
index 0000000..f565464
--- /dev/null
+++ b/release/makedpkg
@@ -0,0 +1,94 @@
+#!/bin/sh
+
+set -u
+set -e
+trap onexit INT
+trap onexit TERM
+trap onexit EXIT
+
+TMPDIR=
+
+onexit()
+{
+	if [ ! "$TMPDIR" = "" ]; then
+		sudo rm -rf $TMPDIR
+	fi
+}
+
+usage()
+{
+	echo "$0 <package name> <version> <build> <DEB architecture> <source dir.>"
+	exit 1
+}
+
+makedeb()
+{
+	SUPPLEMENT=$1
+	DIRNAME=$PACKAGE_NAME
+
+	if [ $SUPPLEMENT = 1 ]; then
+		PACKAGE_NAME=$PACKAGE_NAME\32
+		DEBARCH=amd64
+		__LIB=lib32
+	else
+		__LIB=lib
+	fi
+
+	umask 022
+	rm -f $PACKAGE_NAME\_$DEBARCH.deb
+	TMPDIR=`mktemp -d /tmp/$PACKAGE_NAME-build.XXXXXX`
+	mkdir $TMPDIR/DEBIAN
+	(cat $SRCDIR/release/deb-control.tmpl | sed s/{__PKGNAME}/$PACKAGE_NAME/g \
+		| sed s/{__VERSION}/$VERSION/g | sed s/{__BUILD}/$BUILD/g \
+		| sed s/{__ARCH}/$DEBARCH/g > $TMPDIR/DEBIAN/control)
+
+	make install prefix=$TMPDIR/opt/$DIRNAME libdir=$TMPDIR/opt/$DIRNAME/$__LIB mandir=$TMPDIR/opt/$DIRNAME/man
+	rm -f $TMPDIR/opt/$DIRNAME/$__LIB/*.la
+
+	if [ $SUPPLEMENT = 1 ]; then
+		rm -rf $TMPDIR/opt/$DIRNAME/include
+		rm -rf $TMPDIR/opt/$DIRNAME/man
+		rm -rf $TMPDIR/opt/$DIRNAME/bin
+	fi
+
+	mkdir -p $TMPDIR/usr/$__LIB
+	mv $TMPDIR/opt/$DIRNAME/$__LIB/libturbojpeg.* $TMPDIR/usr/$__LIB
+	ln -fs /usr/$__LIB/libturbojpeg.a $TMPDIR/opt/$DIRNAME/$__LIB/
+	/sbin/ldconfig -n $TMPDIR/opt/$DIRNAME/$__LIB
+	/sbin/ldconfig -n $TMPDIR/usr/$__LIB
+
+	if [ ! $SUPPLEMENT = 1 ]; then
+		if [ "$DEBARCH" = "amd64" ]; then
+			ln -fs lib $TMPDIR/opt/$DIRNAME/lib64
+		else
+			if [ ! "$__LIB" = "lib32" ]; then
+				ln -fs lib $TMPDIR/opt/$DIRNAME/lib32
+			fi
+		fi
+		mkdir -p $TMPDIR/usr/include
+		mv $TMPDIR/opt/$DIRNAME/include/turbojpeg.h $TMPDIR/usr/include
+		ln -fs /usr/include/turbojpeg.h $TMPDIR/opt/$DIRNAME/include/
+		mkdir -p $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+		install -m 644 $SRCDIR/LICENSE.txt $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+		install -m 644 $SRCDIR/LGPL.txt $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+		install -m 644 $SRCDIR/README-turbo.txt $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+		install -m 644 $SRCDIR/README $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+		install -m 644 $SRCDIR/libjpeg.txt $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+		install -m 644 $SRCDIR/usage.txt $TMPDIR/usr/share/doc/$DIRNAME-$VERSION
+	fi
+
+	sudo chown -Rh root:root $TMPDIR/*
+	dpkg -b $TMPDIR $PACKAGE_NAME\_$DEBARCH.deb
+}
+
+if [ $# -lt 5 ]; then usage $0; fi
+PACKAGE_NAME=$1
+VERSION=$2
+BUILD=$3
+DEBARCH=$4
+SRCDIR=$5
+
+makedeb 0
+if [ "$DEBARCH" = "i386" ]; then makedeb 1; fi
+
+exit
diff --git a/release/makemacpkg.in b/release/makemacpkg.in
new file mode 100644
index 0000000..b5a7422
--- /dev/null
+++ b/release/makemacpkg.in
@@ -0,0 +1,162 @@
+#!/bin/sh
+
+set -u
+set -e
+trap onexit INT
+trap onexit TERM
+trap onexit EXIT
+
+TMPDIR=
+
+onexit()
+{
+	if [ ! "$TMPDIR" = "" ]; then
+		sudo rm -rf $TMPDIR
+	fi
+}
+
+usage()
+{
+	echo "$0 [universal [32-bit build dir]]"
+	exit 1
+}
+
+UNIVERSAL=0
+
+PACKAGE_NAME=@PACKAGE_NAME@
+VERSION=@VERSION@
+BUILD=@BUILD@
+SRCDIR=@srcdir@
+BUILDDIR32=@srcdir@/osxx86
+if [ $# -gt 0 ]; then
+	if [ "$1" = "universal" ]; then
+		UNIVERSAL=1
+		if [ $# -gt 1 ]; then BUILDDIR32=$2; fi
+	fi
+fi
+PACKAGEMAKER=/Developer/Applications/Utilities/PackageMaker.app/Contents/MacOS/PackageMaker
+
+if [ -f $PACKAGE_NAME.dmg ]; then
+	rm -f $PACKAGE_NAME.dmg
+fi
+
+umask 022
+TMPDIR=`mktemp -d /tmp/$PACKAGE_NAME-build.XXXXXX`
+PKGROOT=$TMPDIR/pkg/Package_Root
+mkdir -p $PKGROOT
+mkdir -p $PKGROOT/opt/$PACKAGE_NAME/bin
+make install DESTDIR=$PKGROOT mandir=/opt/$PACKAGE_NAME/man
+rm -f $PKGROOT/opt/$PACKAGE_NAME/lib/*.la
+mkdir -p $PKGROOT/usr/lib
+mv $PKGROOT/opt/$PACKAGE_NAME/lib/libturbojpeg.* $PKGROOT/usr/lib
+mkdir -p $PKGROOT/usr/include
+mv $PKGROOT/opt/$PACKAGE_NAME/include/turbojpeg.h $PKGROOT/usr/include
+
+if [ $UNIVERSAL = 1 ]; then
+	if [ ! -d $BUILDDIR32 ]; then
+		echo ERROR: 32-bit build directory $BUILDDIR32 does not exist
+		exit 1
+	fi
+	if [ ! -f $BUILDDIR32/Makefile ]; then
+		echo ERROR: 32-bit build directory $BUILDDIR32 is not configured
+		exit 1
+	fi
+	mkdir -p $TMPDIR/dist.x86
+	pushd $BUILDDIR32
+	make install DESTDIR=$TMPDIR/dist.x86
+	popd
+	if [ ! -h $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib -a \
+		! -h $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib ]; then
+		lipo -create \
+			-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib \
+			-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib \
+			-output $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib
+	elif [ ! -h $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@.dylib -a \
+		! -h $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@.dylib ]; then
+		lipo -create \
+			-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@.dylib \
+			-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@.dylib \
+			-output $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@.dylib
+	fi
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libjpeg.a \
+		-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.a \
+		-output $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.a
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libturbojpeg.dylib \
+		-arch x86_64 $PKGROOT/usr/lib/libturbojpeg.dylib \
+		-output $PKGROOT/usr/lib/libturbojpeg.dylib
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/lib/libturbojpeg.a \
+		-arch x86_64 $PKGROOT/usr/lib/libturbojpeg.a \
+		-output $PKGROOT/usr/lib/libturbojpeg.a
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/bin/cjpeg \
+		-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/bin/cjpeg \
+		-output $PKGROOT/opt/$PACKAGE_NAME/bin/cjpeg
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/bin/djpeg \
+		-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/bin/djpeg \
+		-output $PKGROOT/opt/$PACKAGE_NAME/bin/djpeg
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/bin/jpegtran \
+		-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/bin/jpegtran \
+		-output $PKGROOT/opt/$PACKAGE_NAME/bin/jpegtran
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/bin/rdjpgcom \
+		-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/bin/rdjpgcom \
+		-output $PKGROOT/opt/$PACKAGE_NAME/bin/rdjpgcom
+	lipo -create \
+		-arch i386 $TMPDIR/dist.x86/opt/$PACKAGE_NAME/bin/wrjpgcom \
+		-arch x86_64 $PKGROOT/opt/$PACKAGE_NAME/bin/wrjpgcom \
+		-output $PKGROOT/opt/$PACKAGE_NAME/bin/wrjpgcom
+
+fi
+
+install_name_tool -id /opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib $PKGROOT/opt/$PACKAGE_NAME/lib/libjpeg.@SO_MAJOR_VERSION@.dylib
+install_name_tool -id libturbojpeg.dylib $PKGROOT/usr/lib/libturbojpeg.dylib
+
+ln -fs /usr/include/turbojpeg.h $PKGROOT/opt/$PACKAGE_NAME/include/
+ln -fs /usr/lib/libturbojpeg.a $PKGROOT/opt/$PACKAGE_NAME/lib/
+ln -fs lib $PKGROOT/opt/$PACKAGE_NAME/lib32
+ln -fs lib $PKGROOT/opt/$PACKAGE_NAME/lib64
+
+mkdir -p $PKGROOT/Library/Documentation/$PACKAGE_NAME
+chmod 1775 $PKGROOT/Library
+chmod 775 $PKGROOT/Library/Documentation 
+mkdir -p $TMPDIR/pkg/Resources 
+
+(cat $SRCDIR/release/Description.plist.tmpl | sed s/{__VERSION}/$VERSION/g \
+	| sed s/{__APPNAME}/$PACKAGE_NAME/g \
+	> $TMPDIR/pkg/Description.plist) 
+(cat $SRCDIR/release/Info.plist.tmpl | sed s/{__VERSION}/$VERSION/g	\
+	| sed s/{__BUILD}/$BUILD/g > $TMPDIR/pkg/Info.plist) 
+(cat $SRCDIR/release/uninstall.sh.tmpl \
+	| sed s/{__APPNAME}/$PACKAGE_NAME/g \
+	> $PKGROOT/opt/$PACKAGE_NAME/bin/uninstall) 
+chmod 755 $PKGROOT/opt/$PACKAGE_NAME/bin/uninstall
+
+install -m 644 $SRCDIR/LICENSE.txt $PKGROOT/Library/Documentation/$PACKAGE_NAME/LICENSE.txt 
+install -m 644 $SRCDIR/LGPL.txt $PKGROOT/Library/Documentation/$PACKAGE_NAME/LGPL.txt 
+install -m 644 $SRCDIR/README-turbo.txt $PKGROOT/Library/Documentation/$PACKAGE_NAME/README-turbo.txt 
+install -m 644 $SRCDIR/README $PKGROOT/Library/Documentation/$PACKAGE_NAME/README 
+install -m 644 $SRCDIR/libjpeg.txt $PKGROOT/Library/Documentation/$PACKAGE_NAME/libjpeg.txt 
+install -m 644 $SRCDIR/usage.txt $PKGROOT/Library/Documentation/$PACKAGE_NAME/usage.txt 
+
+sudo chown -R root:admin $PKGROOT 
+sudo chown -R root:0 $PKGROOT/usr 
+cp $SRCDIR/release/License.rtf $SRCDIR/release/Welcome.rtf $SRCDIR/release/ReadMe.rtf $TMPDIR/pkg/Resources/ 
+
+mkdir $TMPDIR/dmg
+$PACKAGEMAKER -build -v -p $TMPDIR/dmg/$PACKAGE_NAME.pkg \
+	-f $PKGROOT -r $TMPDIR/pkg/Resources \
+	-i $TMPDIR/pkg/Info.plist -d $TMPDIR/pkg/Description.plist 
+install -m 644 $SRCDIR/release/uninstall.applescript $TMPDIR 
+sudo osacompile -t APPL -o "$TMPDIR/dmg/Uninstall $PACKAGE_NAME.app" $TMPDIR/uninstall.applescript 
+sudo chown -R $USER "$TMPDIR/dmg/Uninstall $PACKAGE_NAME.app" 
+hdiutil create -fs HFS+ -volname $PACKAGE_NAME-$VERSION \
+	-srcfolder "$TMPDIR/dmg" \
+	$TMPDIR/$PACKAGE_NAME.dmg 
+cp $TMPDIR/$PACKAGE_NAME.dmg . 
+
+exit
diff --git a/release/makesunpkg.in b/release/makesunpkg.in
new file mode 100644
index 0000000..5180225
--- /dev/null
+++ b/release/makesunpkg.in
@@ -0,0 +1,138 @@
+#!/bin/sh
+
+set -u
+set -e
+trap onexit INT
+trap onexit TERM
+trap onexit EXIT
+
+TMPDIR=
+
+onexit()
+{
+	if [ ! "$TMPDIR" = "" ]; then
+		rm -rf $TMPDIR
+	fi
+}
+
+usage()
+{
+	echo "$0 [combined [32-bit build dir.]]"
+	exit 1
+}
+
+COMBINED=0
+
+PACKAGE_NAME=@PACKAGE_NAME@
+VERSION=@VERSION@
+BUILD=@BUILD@
+PKGARCH=@DEBARCH@
+SRCDIR=@srcdir@
+BUILDDIR32=@srcdir@/solx86
+if [ $# -gt 0 ]; then
+	if [ "$1" = "combined" ]; then
+		COMBINED=1
+		if [ $# -gt 1 ]; then	BUILDDIR32=$2; fi
+	fi
+fi
+
+umask 022
+TMPDIR=`mktemp -d /tmp/$PACKAGE_NAME-build.XXXXXX`
+rm -f $PACKAGE_NAME.pkg.bz2
+cp $SRCDIR/release/copyright $TMPDIR
+touch $TMPDIR/depend
+cat $SRCDIR/release/pkginfo.tmpl | sed s/{__VERSION}/$VERSION/g \
+	| sed s/{__BUILD}/$BUILD/g | sed s/{__APPNAME}/$PACKAGE_NAME/g \
+	| sed s/{__PKGNAME}/$PACKAGE_NAME/g > $TMPDIR/pkginfo 
+
+if [ "$PKGARCH" = "i386" ]; then
+	__LIB=lib
+else
+	__LIB=lib/$PKGARCH
+fi
+
+if [ $COMBINED = 1 ]; then
+	if [ ! -d $BUILDDIR32 ]; then
+		echo ERROR: 32-bit build directory $BUILDDIR32 does not exist
+		exit 1
+	fi
+	if [ ! -f $BUILDDIR32/Makefile ]; then
+		echo ERROR: 32-bit build directory $BUILDDIR32 is not configured
+		exit 1
+	fi
+	PWD=`pwd`
+	cd $BUILDDIR32
+	make install DESTDIR=$TMPDIR mandir=/opt/$PACKAGE_NAME/man AM_MAKEFLAGS="mandir=/opt/$PACKAGE_NAME/man"
+	cd $PWD
+fi
+# This mess is to work around a bug in /usr/ccs/bin/make
+make install DESTDIR=$TMPDIR libdir=/opt/$PACKAGE_NAME/$__LIB mandir=/opt/$PACKAGE_NAME/man AM_MAKEFLAGS="libdir=/opt/$PACKAGE_NAME/$__LIB mandir=/opt/$PACKAGE_NAME/man"
+rm -f $TMPDIR/opt/$PACKAGE_NAME/$__LIB/*.la
+mkdir -p $TMPDIR/opt/$PACKAGE_NAME/doc
+cp $SRCDIR/LICENSE.txt $SRCDIR/LGPL.txt $SRCDIR/README-turbo.txt $SRCDIR/README $SRCDIR/libjpeg.txt $SRCDIR/usage.txt $TMPDIR/opt/$PACKAGE_NAME/doc
+chmod 644 $TMPDIR/opt/$PACKAGE_NAME/doc/*
+
+cat >$TMPDIR/proto <<EOF
+i copyright
+i depend
+i pkginfo
+d none $PACKAGE_NAME 0755 root bin
+d none $PACKAGE_NAME/bin 0755 root bin
+f none $PACKAGE_NAME/bin/cjpeg 0755 root bin
+f none $PACKAGE_NAME/bin/djpeg 0755 root bin
+f none $PACKAGE_NAME/bin/jpegtran 0755 root bin
+f none $PACKAGE_NAME/bin/rdjpgcom 0755 root bin
+f none $PACKAGE_NAME/bin/wrjpgcom 0755 root bin
+d none $PACKAGE_NAME/lib 0755 root bin
+EOF
+if [ $COMBINED = 1 ]; then
+cat >>$TMPDIR/proto <<EOF
+f none $PACKAGE_NAME/lib/libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@ 0755 root bin
+s none $PACKAGE_NAME/lib/libjpeg.so.@SO_MAJOR_VERSION@=libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@
+s none $PACKAGE_NAME/lib/libjpeg.so=libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@
+f none $PACKAGE_NAME/lib/libjpeg.a 0644 root bin
+f none $PACKAGE_NAME/lib/libturbojpeg.so 0755 root bin
+f none $PACKAGE_NAME/lib/libturbojpeg.a 0644 root bin
+EOF
+fi
+if [ "${__LIB}" != "lib" ]; then
+echo d none $PACKAGE_NAME/${__LIB} 0755 root bin >>$TMPDIR/proto
+echo s none $PACKAGE_NAME/lib64=${__LIB} >>$TMPDIR/proto
+echo s none $PACKAGE_NAME/lib/64=$PKGARCH >>$TMPDIR/proto
+fi
+cat >>$TMPDIR/proto <<EOF
+f none $PACKAGE_NAME/${__LIB}/libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@ 0755 root bin
+s none $PACKAGE_NAME/${__LIB}/libjpeg.so.@SO_MAJOR_VERSION@=libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@
+s none $PACKAGE_NAME/${__LIB}/libjpeg.so=libjpeg.so.@SO_MAJOR_VERSION@.0.@SO_MINOR_VERSION@
+f none $PACKAGE_NAME/${__LIB}/libjpeg.a 0644 root bin
+f none $PACKAGE_NAME/${__LIB}/libturbojpeg.so 0755 root bin
+f none $PACKAGE_NAME/${__LIB}/libturbojpeg.a 0644 root bin
+s none $PACKAGE_NAME/lib32=lib
+d none $PACKAGE_NAME/man 0755 root bin
+d none $PACKAGE_NAME/man/man1 0755 root bin
+d none $PACKAGE_NAME/man/man1/cjpeg.1 0644 root bin
+d none $PACKAGE_NAME/man/man1/djpeg.1 0644 root bin
+d none $PACKAGE_NAME/man/man1/jpegtran.1 0644 root bin
+d none $PACKAGE_NAME/man/man1/rdjpgcom.1 0644 root bin
+d none $PACKAGE_NAME/man/man1/wrjpgcom.1 0644 root bin
+d none $PACKAGE_NAME/include 0755 root bin
+f none $PACKAGE_NAME/include/jconfig.h 0644 root bin
+f none $PACKAGE_NAME/include/jerror.h 0644 root bin
+f none $PACKAGE_NAME/include/jmorecfg.h 0644 root bin
+f none $PACKAGE_NAME/include/jpeglib.h 0644 root bin
+f none $PACKAGE_NAME/include/turbojpeg.h 0644 root bin
+d none $PACKAGE_NAME/doc 0755 root bin
+f none $PACKAGE_NAME/doc/LGPL.txt 0644 root bin
+f none $PACKAGE_NAME/doc/libjpeg.txt 0644 root bin
+f none $PACKAGE_NAME/doc/LICENSE.txt 0644 root bin
+f none $PACKAGE_NAME/doc/README 0644 root bin
+f none $PACKAGE_NAME/doc/README-turbo.txt 0644 root bin
+f none $PACKAGE_NAME/doc/usage.txt 0644 root bin
+EOF
+
+pkgmk -o -r $TMPDIR/opt -d $TMPDIR -a i386 -f $TMPDIR/proto
+pkgtrans -s $TMPDIR $TMPDIR/$PACKAGE_NAME.pkg $PACKAGE_NAME
+bzip2 $TMPDIR/$PACKAGE_NAME.pkg
+cp $TMPDIR/$PACKAGE_NAME.pkg.bz2 . 
+
+exit
diff --git a/release/pkginfo.tmpl b/release/pkginfo.tmpl
new file mode 100644
index 0000000..68bee88
--- /dev/null
+++ b/release/pkginfo.tmpl
@@ -0,0 +1,16 @@
+ARCH=i386
+PKG={__PKGNAME}
+NAME={__APPNAME} SDK and run time libraries
+VERSION={__VERSION},REV={__BUILD}
+SUNW_PKGVERS=1.0
+DESC=A SIMD-accelerated JPEG codec which provides both the libjpeg and TurboJPEG APIs
+VENDOR=The libjpeg-turbo Project
+HOTLINE=http://libjpeg-turbo.VirtualGL.org
+EMAIL=information@virtualgl.org
+MAXINST=1
+CATEGORY=application
+BASEDIR=/opt
+CLASSES=none
+SUNW_PKG_ALLZONES=true
+SUNW_PKG_HOLLOW=false
+SUNW_PKG_THISZONE=false
diff --git a/release/uninstall.applescript b/release/uninstall.applescript
new file mode 100644
index 0000000..002f681
--- /dev/null
+++ b/release/uninstall.applescript
@@ -0,0 +1,45 @@
+-- Copyright (C)2010 D. R. Commander
+-- Copyright (C)2009 Sun Microsystems, Inc.
+--
+-- This library is free software and may be redistributed and/or modified under
+-- the terms of the wxWindows Library License, Version 3.1 or (at your option)
+-- any later version.  The full license is in the LICENSE.txt file included
+-- with this distribution.
+--
+-- This library is distributed in the hope that it will be useful,
+-- but WITHOUT ANY WARRANTY; without even the implied warranty of
+-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+-- wxWindows Library License for more details.
+
+set disk to (path to startup disk) as string
+set tmpfile to path to temporary items from user domain
+set tmpfile to POSIX path of tmpfile & "libjpeg-turbo_uninstall.log"
+set uninstaller to disk & "opt:libjpeg-turbo:bin:uninstall"
+set uninstaller to "sh " & POSIX path of uninstaller & " 2>&1 >" & tmpfile
+set success to 0
+
+display dialog "You are about to uninstall libjpeg-turbo.  Proceed?" buttons {"Yes", "No"} default button "No"
+
+if button returned of result is "Yes" then
+	try
+		do shell script (uninstaller) with administrator privileges
+		set success to 1
+	on error errstr number errnum
+		if errnum is -128 then
+			display dialog "Uninstall aborted." buttons {"OK"}
+		else if errnum is 255 then
+			set errmsg to "The uninstall script could not remove some of the files or directories installed by the libjpeg-turbo package.  Consult:" & return & return & tmpfile & return & return & "for more details."
+			display dialog errmsg buttons {"OK"} default button "OK" with icon caution
+		else if errnum is 127 then
+			display dialog "Could not find the libjpeg-turbo uninstall script.  The libjpeg-turbo package may have already been uninstalled." buttons {"OK"} default button "OK" with icon stop
+		else
+			set errmsg to "ERROR " & errnum & ": " & errstr
+			display dialog errmsg buttons {"OK"} default button "OK" with icon stop
+		end if
+	end try
+	if success is 1 then
+		display dialog "libjpeg-turbo has been successfully uninstalled." buttons {"OK"}
+	end if
+else
+	display dialog "Uninstall aborted." buttons {"OK"}
+end if
diff --git a/release/uninstall.sh.tmpl b/release/uninstall.sh.tmpl
new file mode 100644
index 0000000..ed8846b
--- /dev/null
+++ b/release/uninstall.sh.tmpl
@@ -0,0 +1,77 @@
+# Copyright (C)2009-2010 D. R. Commander
+# Copyright (C)2009 Sun Microsystems, Inc.
+#
+# This library is free software and may be redistributed and/or modified under
+# the terms of the wxWindows Library License, Version 3.1 or (at your option)
+# any later version.  The full license is in the LICENSE.txt file included
+# with this distribution.
+#
+# This library is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# wxWindows Library License for more details.
+
+#!/bin/sh
+
+error()
+{
+	echo $1
+	exit -1
+}
+
+if [ ! "`id -u`" = "0" ]; then
+	error "This command must be run as root"
+fi
+
+PKGNAME={__APPNAME}
+MACPKGNAME=com.$PKGNAME.$PKGNAME
+RCPT=/Library/Receipts/$PKGNAME.pkg
+
+LSBOM=
+if [ -d $RCPT ]; then
+	LSBOM='lsbom -s -f -l '$RCPT'/Contents/Archive.bom'
+else
+	LSBOM='pkgutil --files '$MACPKGNAME
+fi
+
+echo Removing files ...
+$LSBOM >/dev/null || error "Could not list package contents"
+RETCODE=0
+PWD=`pwd`
+cd /
+$LSBOM | while read line; do
+	if [ ! -d "$line" ]; then rm "$line" 2>&1 || RETCODE=-1; fi
+done
+cd $PWD
+
+echo Removing directories ...
+rmdir /opt/$PKGNAME/bin 2>&1 || RETCODE=-1
+if [ -d /opt/$PKGNAME/bin ]; then
+	rmdir /opt/$PKGNAME/bin 2>&1 || RETCODE=-1
+fi
+if [ -d /opt/$PKGNAME/lib ]; then
+	rmdir /opt/$PKGNAME/lib 2>&1 || RETCODE=-1
+fi
+if [ -d /opt/$PKGNAME/include ]; then
+	rmdir /opt/$PKGNAME/include 2>&1 || RETCODE=-1
+fi
+rm /opt/$PKGNAME/lib32 2>&1 || RETCODE=-1
+rm /opt/$PKGNAME/lib64 2>&1 || RETCODE=-1
+if [ -d /opt/$PKGNAME/man/man1 ]; then
+	rmdir /opt/$PKGNAME/man/man1 2>&1 || RETCODE=-1
+fi
+if [ -d /opt/$PKGNAME/man ]; then
+	rmdir /opt/$PKGNAME/man 2>&1 || RETCODE=-1
+fi
+rmdir /opt/$PKGNAME 2>&1 || RETCODE=-1
+rmdir /Library/Documentation/$PKGNAME 2>&1 || RETCODE=-1
+
+if [ -d $RCPT ]; then
+	echo Removing package receipt $RCPT ...
+	rm -r $RCPT 2>&1 || RETCODE=-1
+else
+	echo Forgetting package $MACPKGNAME
+	pkgutil --forget $MACPKGNAME
+fi
+
+exit $RETCODE
diff --git a/rrtimer.h b/rrtimer.h
new file mode 100644
index 0000000..4db5e37
--- /dev/null
+++ b/rrtimer.h
@@ -0,0 +1,114 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005 Sun Microsystems, Inc.
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#ifndef __RRTIMER_H__
+#define __RRTIMER_H__
+
+#ifdef __cplusplus
+
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <sys/time.h>
+#endif
+
+class rrtimer
+{
+	public:
+
+		rrtimer(void) : t1(0.0)
+		{
+			#ifdef _WIN32
+			highres=false;  tick=0.001;
+			LARGE_INTEGER Frequency;
+			if(QueryPerformanceFrequency(&Frequency)!=0)
+			{
+				tick=(double)1.0/(double)(Frequency.QuadPart);
+				highres=true;
+			}
+			#endif
+		}
+
+		void start(void)
+		{
+			t1=time();
+		}
+
+		double time(void)
+		{
+			#ifdef _WIN32
+			if(highres)
+			{
+				LARGE_INTEGER Time;
+				QueryPerformanceCounter(&Time);
+				return((double)(Time.QuadPart)*tick);
+			}
+			else
+				return((double)GetTickCount()*tick);
+			#else
+			struct timeval __tv;
+			gettimeofday(&__tv, (struct timezone *)NULL);
+			return((double)(__tv.tv_sec)+(double)(__tv.tv_usec)*0.000001);
+			#endif
+		}
+
+		double elapsed(void)
+		{
+			return time()-t1;
+		}
+
+	private:
+
+		#ifdef _WIN32
+		bool highres;  double tick;
+		#endif
+		double t1;
+};
+
+#endif  // __cplusplus
+
+#ifdef _WIN32
+
+#include <windows.h>
+
+__inline double rrtime(void)
+{
+	LARGE_INTEGER Frequency, Time;
+	if(QueryPerformanceFrequency(&Frequency)!=0)
+	{
+		QueryPerformanceCounter(&Time);
+		return (double)Time.QuadPart/(double)Frequency.QuadPart;
+	}
+	else return (double)GetTickCount()*0.001;
+}
+
+#else
+
+#include <sys/time.h>
+
+#ifdef sun
+#define __inline inline
+#endif
+
+static __inline double rrtime(void)
+{
+	struct timeval __tv;
+	gettimeofday(&__tv, (struct timezone *)NULL);
+	return((double)__tv.tv_sec+(double)__tv.tv_usec*0.000001);
+}
+
+#endif
+
+#endif
+
diff --git a/rrutil.h b/rrutil.h
new file mode 100644
index 0000000..4b61dbf
--- /dev/null
+++ b/rrutil.h
@@ -0,0 +1,82 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005 Sun Microsystems, Inc.
+ * Copyright (C)2010 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#ifndef __RRUTIL_H__
+#define __RRUTIL_H__
+
+#ifdef _WIN32
+	#include <windows.h>
+	#define sleep(t) Sleep((t)*1000)
+	#define usleep(t) Sleep((t)/1000)
+#else
+	#include <unistd.h>
+	#define stricmp strcasecmp
+	#define strnicmp strncasecmp
+#endif
+
+#ifndef min
+ #define min(a,b) ((a)<(b)?(a):(b))
+#endif
+
+#ifndef max
+ #define max(a,b) ((a)>(b)?(a):(b))
+#endif
+
+#define pow2(i) (1<<(i))
+#define isPow2(x) (((x)&(x-1))==0)
+
+#ifdef sgi
+#define _SC_NPROCESSORS_CONF _SC_NPROC_CONF
+#endif
+
+#ifdef sun
+#define __inline inline
+#endif
+
+static __inline int numprocs(void)
+{
+	#ifdef _WIN32
+	DWORD_PTR ProcAff, SysAff, i;  int count=0;
+	if(!GetProcessAffinityMask(GetCurrentProcess(), &ProcAff, &SysAff)) return(1);
+	for(i=0; i<sizeof(long*)*8; i++) if(ProcAff&(1LL<<i)) count++;
+	return(count);
+	#elif defined (__APPLE__)
+	return(1);
+	#else
+	long count=1;
+	if((count=sysconf(_SC_NPROCESSORS_CONF))!=-1) return((int)count);
+	else return(1);
+	#endif
+}
+
+#define byteswap(i) ( \
+	(((i) & 0xff000000) >> 24) | \
+	(((i) & 0x00ff0000) >>  8) | \
+	(((i) & 0x0000ff00) <<  8) | \
+	(((i) & 0x000000ff) << 24) )
+
+#define byteswap16(i) ( \
+	(((i) & 0xff00) >> 8) | \
+	(((i) & 0x00ff) << 8) )
+
+static __inline int littleendian(void)
+{
+	unsigned int value=1;
+	unsigned char *ptr=(unsigned char *)(&value);
+	if(ptr[0]==1 && ptr[3]==0) return 1;
+	else return 0;
+}
+
+#endif
diff --git a/sharedlib/CMakeLists.txt b/sharedlib/CMakeLists.txt
new file mode 100755
index 0000000..369a3ed
--- /dev/null
+++ b/sharedlib/CMakeLists.txt
@@ -0,0 +1,60 @@
+# Anything that must be linked against the shared C library on Windows must
+# be built in this subdirectory, because CMake doesn't allow us to override
+# the compiler flags for each build type except at directory scope.  Note
+# to CMake developers:  Add a COMPILE_FLAGS_<CONFIG> target property, or
+# better yet, provide a friendly way of configuring a Windows target to use the
+# static C library.
+
+if(MSVC)
+  # Build all configurations against shared C library
+  foreach(var CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
+    CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO
+    CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+    CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+    if(${var} MATCHES "/MT")
+      string(REGEX REPLACE "/MT" "/MD" ${var} "${${var}}")
+    endif()
+  endforeach()
+endif()
+
+foreach(src ${JPEG_SOURCES})
+  set(JPEG_SRCS ${JPEG_SRCS} ${CMAKE_SOURCE_DIR}/${src})
+endforeach()
+
+if(WITH_SIMD)
+  # This tells CMake that the "source" files haven't been generated yet
+  set_source_files_properties(${SIMD_OBJS} PROPERTIES GENERATED 1)
+endif()
+
+add_library(jpeg SHARED ${JPEG_SRCS} ${SIMD_OBJS}
+  ${CMAKE_SOURCE_DIR}/win/jpeg${DLL_VERSION}.def)
+set_target_properties(jpeg PROPERTIES SOVERSION ${DLL_VERSION}
+  VERSION ${FULLVERSION})
+if(MSVC)
+  set_target_properties(jpeg PROPERTIES SUFFIX ${DLL_VERSION}.dll)
+elseif(MINGW OR CYGWIN)
+  set_target_properties(jpeg PROPERTIES SUFFIX -${DLL_VERSION}.dll)
+endif(MSVC)
+if(WITH_SIMD)
+  add_dependencies(jpeg simd)
+endif()
+
+add_executable(cjpeg ../cjpeg.c ../cdjpeg.c ../rdbmp.c ../rdgif.c ../rdppm.c
+  ../rdswitch.c ../rdtarga.c)
+set_property(TARGET cjpeg PROPERTY COMPILE_FLAGS
+  "-DBMP_SUPPORTED -DGIF_SUPPORTED -DPPM_SUPPORTED -DTARGA_SUPPORTED")
+target_link_libraries(cjpeg jpeg)
+
+add_executable(djpeg ../djpeg.c ../cdjpeg.c ../rdcolmap.c ../rdswitch.c
+  ../wrbmp.c ../wrgif.c ../wrppm.c ../wrtarga.c)
+set_property(TARGET djpeg PROPERTY COMPILE_FLAGS
+  "-DBMP_SUPPORTED -DGIF_SUPPORTED -DPPM_SUPPORTED -DTARGA_SUPPORTED")
+target_link_libraries(djpeg jpeg)
+
+add_executable(jpegtran ../jpegtran.c ../cdjpeg.c ../rdswitch.c ../transupp.c)
+target_link_libraries(jpegtran jpeg)
+
+install(TARGETS jpeg cjpeg djpeg jpegtran
+  ARCHIVE DESTINATION lib
+  LIBRARY DESTINATION lib
+  RUNTIME DESTINATION bin)
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt
new file mode 100755
index 0000000..5be325e
--- /dev/null
+++ b/simd/CMakeLists.txt
@@ -0,0 +1,62 @@
+if(NOT DEFINED NASM)
+  set(NASM nasm)
+endif()
+
+if(SIMD_X86_64)
+  set(NAFLAGS -fwin64 -DWIN64 -D__x86_64__ -I${CMAKE_SOURCE_DIR}/win/
+    -I${CMAKE_CURRENT_SOURCE_DIR}/)
+else()
+  set(NAFLAGS -fwin32 -DWIN32 -I${CMAKE_SOURCE_DIR}/win/
+    -I${CMAKE_CURRENT_SOURCE_DIR}/)
+endif()
+
+if(MSVC)
+  set(NAFLAGS ${NAFLAGS} -DMSVC)
+endif()
+
+# This only works if building from the command line.  There is currently no way
+# to set a variable's value based on the build type when using the MSVC IDE.
+if(CMAKE_BUILD_TYPE STREQUAL "Debug"
+  OR CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
+  set(NAFLAGS ${NAFLAGS} -g)
+endif()
+
+if(SIMD_X86_64)
+  set(SIMD_BASENAMES jfsseflt-64 jccolss2-64 jdcolss2-64 jcsamss2-64
+    jdsamss2-64 jdmerss2-64 jcqnts2i-64 jfss2fst-64 jfss2int-64 jiss2red-64
+    jiss2int-64 jiss2fst-64 jcqnts2f-64 jiss2flt-64)
+  message(STATUS "Building x86_64 SIMD extensions")
+else()
+  set(SIMD_BASENAMES jsimdcpu jccolmmx jdcolmmx jcsammmx jdsammmx jdmermmx
+    jcqntmmx jfmmxfst jfmmxint jimmxred jimmxint jimmxfst jcqnt3dn jf3dnflt
+    ji3dnflt jcqntsse jfsseflt jisseflt jccolss2 jdcolss2 jcsamss2 jdsamss2
+    jdmerss2 jcqnts2i jfss2fst jfss2int jiss2red jiss2int jiss2fst jcqnts2f
+    jiss2flt)
+  message(STATUS "Building i386 SIMD extensions")
+endif()
+
+if(MSVC_IDE)
+  set(OBJDIR "${CMAKE_CURRENT_BINARY_DIR}/$(OutDir)")
+else()
+  set(OBJDIR ${CMAKE_CURRENT_BINARY_DIR})
+endif()
+
+foreach(file ${SIMD_BASENAMES})
+  set(DEPFILE "")
+  set(SIMD_SRC ${CMAKE_CURRENT_SOURCE_DIR}/${file}.asm)
+  if(${file} MATCHES col)
+    set(DEPFILE ${SIMD_SRC})
+    string(REGEX REPLACE "col" "clr" DEPFILE ${DEPFILE})
+  endif()
+  if(${file} MATCHES mer)
+    set(DEPFILE ${SIMD_SRC})
+    string(REGEX REPLACE "mer" "mrg" DEPFILE ${DEPFILE})
+  endif()
+  set(SIMD_OBJ ${OBJDIR}/${file}.obj)
+  add_custom_command(OUTPUT ${SIMD_OBJ} DEPENDS ${SIMD_SRC} ${DEPFILE} *.inc
+    COMMAND ${NASM} ${NAFLAGS} ${SIMD_SRC} -o${SIMD_OBJ})
+  set(SIMD_OBJS ${SIMD_OBJS} ${SIMD_OBJ})
+endforeach()
+
+set(SIMD_OBJS ${SIMD_OBJS} PARENT_SCOPE)
+add_custom_target(simd DEPENDS ${SIMD_OBJS})
diff --git a/simd/Makefile.am b/simd/Makefile.am
new file mode 100644
index 0000000..81c23af
--- /dev/null
+++ b/simd/Makefile.am
@@ -0,0 +1,58 @@
+noinst_LTLIBRARIES = libsimd.la
+
+BUILT_SOURCES = jsimdcfg.inc
+
+EXTRA_DIST = nasm_lt.sh jcclrmmx.asm jcclrss2.asm jdclrmmx.asm jdclrss2.asm \
+	jdmrgmmx.asm jdmrgss2.asm jcclrss2-64.asm jdclrss2-64.asm \
+	jdmrgss2-64.asm CMakeLists.txt
+
+if SIMD_X86_64
+
+libsimd_la_SOURCES = jsimd_x86_64.c \
+	jsimd.h jsimdcfg.inc.h \
+	jsimdext.inc jcolsamp.inc jdct.inc \
+	jfsseflt-64.asm \
+	jccolss2-64.asm jdcolss2-64.asm \
+	jcsamss2-64.asm jdsamss2-64.asm jdmerss2-64.asm \
+	jcqnts2i-64.asm jfss2fst-64.asm jfss2int-64.asm \
+	jiss2red-64.asm jiss2int-64.asm jiss2fst-64.asm \
+	jcqnts2f-64.asm jiss2flt-64.asm
+
+jccolss2-64.lo: jcclrss2-64.asm
+jdcolss2-64.lo: jdclrss2-64.asm
+jdmerss2-64.lo: jdmrgss2-64.asm
+endif
+
+if SIMD_I386
+
+libsimd_la_SOURCES = jsimd_i386.c \
+	jsimd.h jsimdcfg.inc.h \
+	jsimdext.inc jcolsamp.inc jdct.inc \
+	jsimdcpu.asm \
+	jccolmmx.asm jdcolmmx.asm \
+	jcsammmx.asm jdsammmx.asm jdmermmx.asm \
+	jcqntmmx.asm jfmmxfst.asm jfmmxint.asm \
+	jimmxred.asm jimmxint.asm jimmxfst.asm \
+	jcqnt3dn.asm jf3dnflt.asm ji3dnflt.asm \
+	jcqntsse.asm jfsseflt.asm jisseflt.asm \
+	jccolss2.asm jdcolss2.asm \
+	jcsamss2.asm jdsamss2.asm jdmerss2.asm \
+	jcqnts2i.asm jfss2fst.asm jfss2int.asm \
+	jiss2red.asm jiss2int.asm jiss2fst.asm \
+	jcqnts2f.asm jiss2flt.asm
+
+jccolmmx.lo: jcclrmmx.asm
+jccolss2.lo: jcclrss2.asm
+jdcolmmx.lo: jdclrmmx.asm
+jdcolss2.lo: jdclrss2.asm
+jdmermmx.lo: jdmrgmmx.asm
+jdmerss2.lo: jdmrgss2.asm
+endif
+
+AM_CPPFLAGS = -I$(top_srcdir) 
+
+.asm.lo:
+	$(LIBTOOL) --mode=compile --tag NASM $(srcdir)/nasm_lt.sh $(NASM) $(NAFLAGS) -I$(srcdir) $< -o $@
+
+jsimdcfg.inc: $(srcdir)/jsimdcfg.inc.h ../jpeglib.h ../jconfig.h ../jmorecfg.h
+	$(CPP) -I$(top_builddir) -I$(top_builddir)/simd $(srcdir)/jsimdcfg.inc.h | $(EGREP) "^[\;%]|^\ %" | sed 's%_cpp_protection_%%' | sed 's@% define@%define@g' > $@
diff --git a/simd/jcclrmmx.asm b/simd/jcclrmmx.asm
new file mode 100644
index 0000000..b6b8912
--- /dev/null
+++ b/simd/jcclrmmx.asm
@@ -0,0 +1,479 @@
+;
+; jcclrmmx.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_mmx (JDIMENSION img_width,
+;                           JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+;                           JDIMENSION output_row, int num_rows);
+;
+
+%define img_width(b)	(b)+8			; JDIMENSION img_width
+%define input_buf(b)	(b)+12		; JSAMPARRAY input_buf
+%define output_buf(b)	(b)+16		; JSAMPIMAGE output_buf
+%define output_row(b)	(b)+20		; JDIMENSION output_row
+%define num_rows(b)	(b)+24		; int num_rows
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		8
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_rgb_ycc_convert_mmx)
+
+EXTN(jsimd_rgb_ycc_convert_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	ecx, JDIMENSION [img_width(eax)]	; num_cols
+	test	ecx,ecx
+	jz	near .return
+
+	push	ecx
+
+	mov	esi, JSAMPIMAGE [output_buf(eax)]
+	mov	ecx, JDIMENSION [output_row(eax)]
+	mov	edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
+	lea	edi, [edi+ecx*SIZEOF_JSAMPROW]
+	lea	ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+	lea	edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+	pop	ecx
+
+	mov	esi, JSAMPARRAY [input_buf(eax)]
+	mov	eax, INT [num_rows(eax)]
+	test	eax,eax
+	jle	near .return
+	alignx	16,7
+.rowloop:
+	pushpic	eax
+	push	edx
+	push	ebx
+	push	edi
+	push	esi
+	push	ecx			; col
+
+	mov	esi, JSAMPROW [esi]	; inptr
+	mov	edi, JSAMPROW [edi]	; outptr0
+	mov	ebx, JSAMPROW [ebx]	; outptr1
+	mov	edx, JSAMPROW [edx]	; outptr2
+	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
+
+	cmp	ecx, byte SIZEOF_MMWORD
+	jae	short .columnloop
+	alignx	16,7
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+.column_ld1:
+	push	eax
+	push	edx
+	lea	ecx,[ecx+ecx*2]		; imul ecx,RGB_PIXELSIZE
+	test	cl, SIZEOF_BYTE
+	jz	short .column_ld2
+	sub	ecx, byte SIZEOF_BYTE
+	xor	eax,eax
+	mov	al, BYTE [esi+ecx]
+.column_ld2:
+	test	cl, SIZEOF_WORD
+	jz	short .column_ld4
+	sub	ecx, byte SIZEOF_WORD
+	xor	edx,edx
+	mov	dx, WORD [esi+ecx]
+	shl	eax, WORD_BIT
+	or	eax,edx
+.column_ld4:
+	movd	mmA,eax
+	pop	edx
+	pop	eax
+	test	cl, SIZEOF_DWORD
+	jz	short .column_ld8
+	sub	ecx, byte SIZEOF_DWORD
+	movd	mmG, DWORD [esi+ecx]
+	psllq	mmA, DWORD_BIT
+	por	mmA,mmG
+.column_ld8:
+	test	cl, SIZEOF_MMWORD
+	jz	short .column_ld16
+	movq	mmG,mmA
+	movq	mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+	mov	ecx, SIZEOF_MMWORD
+	jmp	short .rgb_ycc_cnv
+.column_ld16:
+	test	cl, 2*SIZEOF_MMWORD
+	mov	ecx, SIZEOF_MMWORD
+	jz	short .rgb_ycc_cnv
+	movq	mmF,mmA
+	movq	mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mmG, MMWORD [esi+1*SIZEOF_MMWORD]
+	jmp	short .rgb_ycc_cnv
+	alignx	16,7
+
+.columnloop:
+	movq	mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mmG, MMWORD [esi+1*SIZEOF_MMWORD]
+	movq	mmF, MMWORD [esi+2*SIZEOF_MMWORD]
+
+.rgb_ycc_cnv:
+	; mmA=(00 10 20 01 11 21 02 12)
+	; mmG=(22 03 13 23 04 14 24 05)
+	; mmF=(15 25 06 16 26 07 17 27)
+
+	movq      mmD,mmA
+	psllq     mmA,4*BYTE_BIT	; mmA=(-- -- -- -- 00 10 20 01)
+	psrlq     mmD,4*BYTE_BIT	; mmD=(11 21 02 12 -- -- -- --)
+
+	punpckhbw mmA,mmG		; mmA=(00 04 10 14 20 24 01 05)
+	psllq     mmG,4*BYTE_BIT	; mmG=(-- -- -- -- 22 03 13 23)
+
+	punpcklbw mmD,mmF		; mmD=(11 15 21 25 02 06 12 16)
+	punpckhbw mmG,mmF		; mmG=(22 26 03 07 13 17 23 27)
+
+	movq      mmE,mmA
+	psllq     mmA,4*BYTE_BIT	; mmA=(-- -- -- -- 00 04 10 14)
+	psrlq     mmE,4*BYTE_BIT	; mmE=(20 24 01 05 -- -- -- --)
+
+	punpckhbw mmA,mmD		; mmA=(00 02 04 06 10 12 14 16)
+	psllq     mmD,4*BYTE_BIT	; mmD=(-- -- -- -- 11 15 21 25)
+
+	punpcklbw mmE,mmG		; mmE=(20 22 24 26 01 03 05 07)
+	punpckhbw mmD,mmG		; mmD=(11 13 15 17 21 23 25 27)
+
+	pxor      mmH,mmH
+
+	movq      mmC,mmA
+	punpcklbw mmA,mmH		; mmA=(00 02 04 06)
+	punpckhbw mmC,mmH		; mmC=(10 12 14 16)
+
+	movq      mmB,mmE
+	punpcklbw mmE,mmH		; mmE=(20 22 24 26)
+	punpckhbw mmB,mmH		; mmB=(01 03 05 07)
+
+	movq      mmF,mmD
+	punpcklbw mmD,mmH		; mmD=(11 13 15 17)
+	punpckhbw mmF,mmH		; mmF=(21 23 25 27)
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+	test	cl, SIZEOF_MMWORD/8
+	jz	short .column_ld2
+	sub	ecx, byte SIZEOF_MMWORD/8
+	movd	mmA, DWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+	test	cl, SIZEOF_MMWORD/4
+	jz	short .column_ld4
+	sub	ecx, byte SIZEOF_MMWORD/4
+	movq	mmF,mmA
+	movq	mmA, MMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld4:
+	test	cl, SIZEOF_MMWORD/2
+	mov	ecx, SIZEOF_MMWORD
+	jz	short .rgb_ycc_cnv
+	movq	mmD,mmA
+	movq	mmC,mmF
+	movq	mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mmF, MMWORD [esi+1*SIZEOF_MMWORD]
+	jmp	short .rgb_ycc_cnv
+	alignx	16,7
+
+.columnloop:
+	movq	mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mmF, MMWORD [esi+1*SIZEOF_MMWORD]
+	movq	mmD, MMWORD [esi+2*SIZEOF_MMWORD]
+	movq	mmC, MMWORD [esi+3*SIZEOF_MMWORD]
+
+.rgb_ycc_cnv:
+	; mmA=(00 10 20 30 01 11 21 31)
+	; mmF=(02 12 22 32 03 13 23 33)
+	; mmD=(04 14 24 34 05 15 25 35)
+	; mmC=(06 16 26 36 07 17 27 37)
+
+	movq      mmB,mmA
+	punpcklbw mmA,mmF		; mmA=(00 02 10 12 20 22 30 32)
+	punpckhbw mmB,mmF		; mmB=(01 03 11 13 21 23 31 33)
+
+	movq      mmG,mmD
+	punpcklbw mmD,mmC		; mmD=(04 06 14 16 24 26 34 36)
+	punpckhbw mmG,mmC		; mmG=(05 07 15 17 25 27 35 37)
+
+	movq      mmE,mmA
+	punpcklwd mmA,mmD		; mmA=(00 02 04 06 10 12 14 16)
+	punpckhwd mmE,mmD		; mmE=(20 22 24 26 30 32 34 36)
+
+	movq      mmH,mmB
+	punpcklwd mmB,mmG		; mmB=(01 03 05 07 11 13 15 17)
+	punpckhwd mmH,mmG		; mmH=(21 23 25 27 31 33 35 37)
+
+	pxor      mmF,mmF
+
+	movq      mmC,mmA
+	punpcklbw mmA,mmF		; mmA=(00 02 04 06)
+	punpckhbw mmC,mmF		; mmC=(10 12 14 16)
+
+	movq      mmD,mmB
+	punpcklbw mmB,mmF		; mmB=(01 03 05 07)
+	punpckhbw mmD,mmF		; mmD=(11 13 15 17)
+
+	movq      mmG,mmE
+	punpcklbw mmE,mmF		; mmE=(20 22 24 26)
+	punpckhbw mmG,mmF		; mmG=(30 32 34 36)
+
+	punpcklbw mmF,mmH
+	punpckhbw mmH,mmH
+	psrlw     mmF,BYTE_BIT		; mmF=(21 23 25 27)
+	psrlw     mmH,BYTE_BIT		; mmH=(31 33 35 37)
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+	; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE
+	; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO
+
+	; (Original)
+	; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+	;
+	; (This implementation)
+	; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+	movq      MMWORD [wk(0)], mm0	; wk(0)=RE
+	movq      MMWORD [wk(1)], mm1	; wk(1)=RO
+	movq      MMWORD [wk(2)], mm4	; wk(2)=BE
+	movq      MMWORD [wk(3)], mm5	; wk(3)=BO
+
+	movq      mm6,mm1
+	punpcklwd mm1,mm3
+	punpckhwd mm6,mm3
+	movq      mm7,mm1
+	movq      mm4,mm6
+	pmaddwd   mm1,[GOTOFF(eax,PW_F0299_F0337)] ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+	pmaddwd   mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+	pmaddwd   mm7,[GOTOFF(eax,PW_MF016_MF033)] ; mm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+	pmaddwd   mm4,[GOTOFF(eax,PW_MF016_MF033)] ; mm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+	movq      MMWORD [wk(4)], mm1	; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+	movq      MMWORD [wk(5)], mm6	; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+	pxor      mm1,mm1
+	pxor      mm6,mm6
+	punpcklwd mm1,mm5		; mm1=BOL
+	punpckhwd mm6,mm5		; mm6=BOH
+	psrld     mm1,1			; mm1=BOL*FIX(0.500)
+	psrld     mm6,1			; mm6=BOH*FIX(0.500)
+
+	movq      mm5,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm5=[PD_ONEHALFM1_CJ]
+
+	paddd     mm7,mm1
+	paddd     mm4,mm6
+	paddd     mm7,mm5
+	paddd     mm4,mm5
+	psrld     mm7,SCALEBITS		; mm7=CbOL
+	psrld     mm4,SCALEBITS		; mm4=CbOH
+	packssdw  mm7,mm4		; mm7=CbO
+
+	movq      mm1, MMWORD [wk(2)]	; mm1=BE
+
+	movq      mm6,mm0
+	punpcklwd mm0,mm2
+	punpckhwd mm6,mm2
+	movq      mm5,mm0
+	movq      mm4,mm6
+	pmaddwd   mm0,[GOTOFF(eax,PW_F0299_F0337)] ; mm0=REL*FIX(0.299)+GEL*FIX(0.337)
+	pmaddwd   mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=REH*FIX(0.299)+GEH*FIX(0.337)
+	pmaddwd   mm5,[GOTOFF(eax,PW_MF016_MF033)] ; mm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+	pmaddwd   mm4,[GOTOFF(eax,PW_MF016_MF033)] ; mm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+	movq      MMWORD [wk(6)], mm0	; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+	movq      MMWORD [wk(7)], mm6	; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+	pxor      mm0,mm0
+	pxor      mm6,mm6
+	punpcklwd mm0,mm1		; mm0=BEL
+	punpckhwd mm6,mm1		; mm6=BEH
+	psrld     mm0,1			; mm0=BEL*FIX(0.500)
+	psrld     mm6,1			; mm6=BEH*FIX(0.500)
+
+	movq      mm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm1=[PD_ONEHALFM1_CJ]
+
+	paddd     mm5,mm0
+	paddd     mm4,mm6
+	paddd     mm5,mm1
+	paddd     mm4,mm1
+	psrld     mm5,SCALEBITS		; mm5=CbEL
+	psrld     mm4,SCALEBITS		; mm4=CbEH
+	packssdw  mm5,mm4		; mm5=CbE
+
+	psllw     mm7,BYTE_BIT
+	por       mm5,mm7		; mm5=Cb
+	movq      MMWORD [ebx], mm5	; Save Cb
+
+	movq      mm0, MMWORD [wk(3)]	; mm0=BO
+	movq      mm6, MMWORD [wk(2)]	; mm6=BE
+	movq      mm1, MMWORD [wk(1)]	; mm1=RO
+
+	movq      mm4,mm0
+	punpcklwd mm0,mm3
+	punpckhwd mm4,mm3
+	movq      mm7,mm0
+	movq      mm5,mm4
+	pmaddwd   mm0,[GOTOFF(eax,PW_F0114_F0250)] ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+	pmaddwd   mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+	pmaddwd   mm7,[GOTOFF(eax,PW_MF008_MF041)] ; mm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+	pmaddwd   mm5,[GOTOFF(eax,PW_MF008_MF041)] ; mm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+	movq      mm3,[GOTOFF(eax,PD_ONEHALF)]	; mm3=[PD_ONEHALF]
+
+	paddd     mm0, MMWORD [wk(4)]
+	paddd     mm4, MMWORD [wk(5)]
+	paddd     mm0,mm3
+	paddd     mm4,mm3
+	psrld     mm0,SCALEBITS		; mm0=YOL
+	psrld     mm4,SCALEBITS		; mm4=YOH
+	packssdw  mm0,mm4		; mm0=YO
+
+	pxor      mm3,mm3
+	pxor      mm4,mm4
+	punpcklwd mm3,mm1		; mm3=ROL
+	punpckhwd mm4,mm1		; mm4=ROH
+	psrld     mm3,1			; mm3=ROL*FIX(0.500)
+	psrld     mm4,1			; mm4=ROH*FIX(0.500)
+
+	movq      mm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm1=[PD_ONEHALFM1_CJ]
+
+	paddd     mm7,mm3
+	paddd     mm5,mm4
+	paddd     mm7,mm1
+	paddd     mm5,mm1
+	psrld     mm7,SCALEBITS		; mm7=CrOL
+	psrld     mm5,SCALEBITS		; mm5=CrOH
+	packssdw  mm7,mm5		; mm7=CrO
+
+	movq      mm3, MMWORD [wk(0)]	; mm3=RE
+
+	movq      mm4,mm6
+	punpcklwd mm6,mm2
+	punpckhwd mm4,mm2
+	movq      mm1,mm6
+	movq      mm5,mm4
+	pmaddwd   mm6,[GOTOFF(eax,PW_F0114_F0250)] ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+	pmaddwd   mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+	pmaddwd   mm1,[GOTOFF(eax,PW_MF008_MF041)] ; mm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+	pmaddwd   mm5,[GOTOFF(eax,PW_MF008_MF041)] ; mm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+	movq      mm2,[GOTOFF(eax,PD_ONEHALF)]	; mm2=[PD_ONEHALF]
+
+	paddd     mm6, MMWORD [wk(6)]
+	paddd     mm4, MMWORD [wk(7)]
+	paddd     mm6,mm2
+	paddd     mm4,mm2
+	psrld     mm6,SCALEBITS		; mm6=YEL
+	psrld     mm4,SCALEBITS		; mm4=YEH
+	packssdw  mm6,mm4		; mm6=YE
+
+	psllw     mm0,BYTE_BIT
+	por       mm6,mm0		; mm6=Y
+	movq      MMWORD [edi], mm6	; Save Y
+
+	pxor      mm2,mm2
+	pxor      mm4,mm4
+	punpcklwd mm2,mm3		; mm2=REL
+	punpckhwd mm4,mm3		; mm4=REH
+	psrld     mm2,1			; mm2=REL*FIX(0.500)
+	psrld     mm4,1			; mm4=REH*FIX(0.500)
+
+	movq      mm0,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm0=[PD_ONEHALFM1_CJ]
+
+	paddd     mm1,mm2
+	paddd     mm5,mm4
+	paddd     mm1,mm0
+	paddd     mm5,mm0
+	psrld     mm1,SCALEBITS		; mm1=CrEL
+	psrld     mm5,SCALEBITS		; mm5=CrEH
+	packssdw  mm1,mm5		; mm1=CrE
+
+	psllw     mm7,BYTE_BIT
+	por       mm1,mm7		; mm1=Cr
+	movq      MMWORD [edx], mm1	; Save Cr
+
+	sub	ecx, byte SIZEOF_MMWORD
+	add	esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD	; inptr
+	add	edi, byte SIZEOF_MMWORD			; outptr0
+	add	ebx, byte SIZEOF_MMWORD			; outptr1
+	add	edx, byte SIZEOF_MMWORD			; outptr2
+	cmp	ecx, byte SIZEOF_MMWORD
+	jae	near .columnloop
+	test	ecx,ecx
+	jnz	near .column_ld1
+
+	pop	ecx			; col
+	pop	esi
+	pop	edi
+	pop	ebx
+	pop	edx
+	poppic	eax
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_buf
+	add	edi, byte SIZEOF_JSAMPROW
+	add	ebx, byte SIZEOF_JSAMPROW
+	add	edx, byte SIZEOF_JSAMPROW
+	dec	eax				; num_rows
+	jg	near .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcclrss2-64.asm b/simd/jcclrss2-64.asm
new file mode 100644
index 0000000..8ca47aa
--- /dev/null
+++ b/simd/jcclrss2-64.asm
@@ -0,0 +1,487 @@
+;
+; jcclrss2-64.asm - colorspace conversion (64-bit SSE2)
+;
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; Copyright (C) 2009, D. R. Commander.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_sse2 (JDIMENSION img_width,
+;                             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+;                             JDIMENSION output_row, int num_rows);
+;
+
+; r10 = JDIMENSION img_width
+; r11 = JSAMPARRAY input_buf
+; r12 = JSAMPIMAGE output_buf
+; r13 = JDIMENSION output_row
+; r14 = int num_rows
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		8
+
+	align	16
+
+	global	EXTN(jsimd_rgb_ycc_convert_sse2)
+
+EXTN(jsimd_rgb_ycc_convert_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+	push	rbx
+
+	mov	rcx, r10
+	test	rcx,rcx
+	jz	near .return
+
+	push	rcx
+
+	mov rsi, r12
+	mov rcx, r13
+	mov	rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+	mov	rbx, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
+	mov	rdx, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
+	lea	rdi, [rdi+rcx*SIZEOF_JSAMPROW]
+	lea	rbx, [rbx+rcx*SIZEOF_JSAMPROW]
+	lea	rdx, [rdx+rcx*SIZEOF_JSAMPROW]
+
+	pop	rcx
+
+	mov rsi, r11
+	mov	eax, r14d
+	test	rax,rax
+	jle	near .return
+.rowloop:
+	push	rdx
+	push	rbx
+	push	rdi
+	push	rsi
+	push	rcx			; col
+
+	mov	rsi, JSAMPROW [rsi]	; inptr
+	mov	rdi, JSAMPROW [rdi]	; outptr0
+	mov	rbx, JSAMPROW [rbx]	; outptr1
+	mov	rdx, JSAMPROW [rdx]	; outptr2
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jae	near .columnloop
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+.column_ld1:
+	push	rax
+	push	rdx
+	lea	rcx,[rcx+rcx*2]		; imul ecx,RGB_PIXELSIZE
+	test	cl, SIZEOF_BYTE
+	jz	short .column_ld2
+	sub	rcx, byte SIZEOF_BYTE
+	movzx	rax, BYTE [rsi+rcx]
+.column_ld2:
+	test	cl, SIZEOF_WORD
+	jz	short .column_ld4
+	sub	rcx, byte SIZEOF_WORD
+	movzx	rdx, WORD [rsi+rcx]
+	shl	rax, WORD_BIT
+	or	rax,rdx
+.column_ld4:
+	movd	xmmA,eax
+	pop	rdx
+	pop	rax
+	test	cl, SIZEOF_DWORD
+	jz	short .column_ld8
+	sub	rcx, byte SIZEOF_DWORD
+	movd	xmmF, XMM_DWORD [rsi+rcx]
+	pslldq	xmmA, SIZEOF_DWORD
+	por	xmmA,xmmF
+.column_ld8:
+	test	cl, SIZEOF_MMWORD
+	jz	short .column_ld16
+	sub	rcx, byte SIZEOF_MMWORD
+	movq	xmmB, XMM_MMWORD [rsi+rcx]
+	pslldq	xmmA, SIZEOF_MMWORD
+	por	xmmA,xmmB
+.column_ld16:
+	test	cl, SIZEOF_XMMWORD
+	jz	short .column_ld32
+	movdqa	xmmF,xmmA
+	movdqu	xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	mov	rcx, SIZEOF_XMMWORD
+	jmp	short .rgb_ycc_cnv
+.column_ld32:
+	test	cl, 2*SIZEOF_XMMWORD
+	mov	rcx, SIZEOF_XMMWORD
+	jz	short .rgb_ycc_cnv
+	movdqa	xmmB,xmmA
+	movdqu	xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqu	xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+	jmp	short .rgb_ycc_cnv
+
+.columnloop:
+	movdqu	xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqu	xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+	movdqu	xmmB, XMMWORD [rsi+2*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+	; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+	; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+	movdqa    xmmG,xmmA
+	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
+	psrldq    xmmG,8	; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
+
+	punpckhbw xmmA,xmmF	; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
+	pslldq    xmmF,8	; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
+
+	punpcklbw xmmG,xmmB	; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
+	punpckhbw xmmF,xmmB	; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
+
+	movdqa    xmmD,xmmA
+	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
+	psrldq    xmmD,8	; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
+
+	punpckhbw xmmA,xmmG	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
+	pslldq    xmmG,8	; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
+
+	punpcklbw xmmD,xmmF	; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
+	punpckhbw xmmG,xmmF	; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
+
+	movdqa    xmmE,xmmA
+	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
+	psrldq    xmmE,8	; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
+
+	punpckhbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+	pslldq    xmmD,8	; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
+
+	punpcklbw xmmE,xmmG	; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
+	punpckhbw xmmD,xmmG	; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
+
+	pxor      xmmH,xmmH
+
+	movdqa    xmmC,xmmA
+	punpcklbw xmmA,xmmH	; xmmA=(00 02 04 06 08 0A 0C 0E)
+	punpckhbw xmmC,xmmH	; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+	movdqa    xmmB,xmmE
+	punpcklbw xmmE,xmmH	; xmmE=(20 22 24 26 28 2A 2C 2E)
+	punpckhbw xmmB,xmmH	; xmmB=(01 03 05 07 09 0B 0D 0F)
+
+	movdqa    xmmF,xmmD
+	punpcklbw xmmD,xmmH	; xmmD=(11 13 15 17 19 1B 1D 1F)
+	punpckhbw xmmF,xmmH	; xmmF=(21 23 25 27 29 2B 2D 2F)
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+	test	cl, SIZEOF_XMMWORD/16
+	jz	short .column_ld2
+	sub	rcx, byte SIZEOF_XMMWORD/16
+	movd	xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld2:
+	test	cl, SIZEOF_XMMWORD/8
+	jz	short .column_ld4
+	sub	rcx, byte SIZEOF_XMMWORD/8
+	movq	xmmE, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE]
+	pslldq	xmmA, SIZEOF_MMWORD
+	por	xmmA,xmmE
+.column_ld4:
+	test	cl, SIZEOF_XMMWORD/4
+	jz	short .column_ld8
+	sub	rcx, byte SIZEOF_XMMWORD/4
+	movdqa	xmmE,xmmA
+	movdqu	xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE]
+.column_ld8:
+	test	cl, SIZEOF_XMMWORD/2
+	mov	rcx, SIZEOF_XMMWORD
+	jz	short .rgb_ycc_cnv
+	movdqa	xmmF,xmmA
+	movdqa	xmmH,xmmE
+	movdqu	xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqu	xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+	jmp	short .rgb_ycc_cnv
+
+.columnloop:
+	movdqu	xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqu	xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+	movdqu	xmmF, XMMWORD [rsi+2*SIZEOF_XMMWORD]
+	movdqu	xmmH, XMMWORD [rsi+3*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+	; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+	; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+	movdqa    xmmD,xmmA
+	punpcklbw xmmA,xmmE	; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
+	punpckhbw xmmD,xmmE	; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
+
+	movdqa    xmmC,xmmF
+	punpcklbw xmmF,xmmH	; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
+	punpckhbw xmmC,xmmH	; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
+
+	movdqa    xmmB,xmmA
+	punpcklwd xmmA,xmmF	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
+	punpckhwd xmmB,xmmF	; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
+
+	movdqa    xmmG,xmmD
+	punpcklwd xmmD,xmmC	; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
+	punpckhwd xmmG,xmmC	; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
+
+	movdqa    xmmE,xmmA
+	punpcklbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+	punpckhbw xmmE,xmmD	; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
+
+	movdqa    xmmH,xmmB
+	punpcklbw xmmB,xmmG	; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
+	punpckhbw xmmH,xmmG	; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
+
+	pxor      xmmF,xmmF
+
+	movdqa    xmmC,xmmA
+	punpcklbw xmmA,xmmF	; xmmA=(00 02 04 06 08 0A 0C 0E)
+	punpckhbw xmmC,xmmF	; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+	movdqa    xmmD,xmmB
+	punpcklbw xmmB,xmmF	; xmmB=(01 03 05 07 09 0B 0D 0F)
+	punpckhbw xmmD,xmmF	; xmmD=(11 13 15 17 19 1B 1D 1F)
+
+	movdqa    xmmG,xmmE
+	punpcklbw xmmE,xmmF	; xmmE=(20 22 24 26 28 2A 2C 2E)
+	punpckhbw xmmG,xmmF	; xmmG=(30 32 34 36 38 3A 3C 3E)
+
+	punpcklbw xmmF,xmmH
+	punpckhbw xmmH,xmmH
+	psrlw     xmmF,BYTE_BIT	; xmmF=(21 23 25 27 29 2B 2D 2F)
+	psrlw     xmmH,BYTE_BIT	; xmmH=(31 33 35 37 39 3B 3D 3F)
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+	; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
+	; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
+
+	; (Original)
+	; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+	;
+	; (This implementation)
+	; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+	movdqa    XMMWORD [wk(0)], xmm0	; wk(0)=RE
+	movdqa    XMMWORD [wk(1)], xmm1	; wk(1)=RO
+	movdqa    XMMWORD [wk(2)], xmm4	; wk(2)=BE
+	movdqa    XMMWORD [wk(3)], xmm5	; wk(3)=BO
+
+	movdqa    xmm6,xmm1
+	punpcklwd xmm1,xmm3
+	punpckhwd xmm6,xmm3
+	movdqa    xmm7,xmm1
+	movdqa    xmm4,xmm6
+	pmaddwd   xmm1,[rel PW_F0299_F0337] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+	pmaddwd   xmm6,[rel PW_F0299_F0337] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+	pmaddwd   xmm7,[rel PW_MF016_MF033] ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+	pmaddwd   xmm4,[rel PW_MF016_MF033] ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+	movdqa    XMMWORD [wk(4)], xmm1	; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+	movdqa    XMMWORD [wk(5)], xmm6	; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+	pxor      xmm1,xmm1
+	pxor      xmm6,xmm6
+	punpcklwd xmm1,xmm5		; xmm1=BOL
+	punpckhwd xmm6,xmm5		; xmm6=BOH
+	psrld     xmm1,1		; xmm1=BOL*FIX(0.500)
+	psrld     xmm6,1		; xmm6=BOH*FIX(0.500)
+
+	movdqa    xmm5,[rel PD_ONEHALFM1_CJ] ; xmm5=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm7,xmm1
+	paddd     xmm4,xmm6
+	paddd     xmm7,xmm5
+	paddd     xmm4,xmm5
+	psrld     xmm7,SCALEBITS	; xmm7=CbOL
+	psrld     xmm4,SCALEBITS	; xmm4=CbOH
+	packssdw  xmm7,xmm4		; xmm7=CbO
+
+	movdqa    xmm1, XMMWORD [wk(2)]	; xmm1=BE
+
+	movdqa    xmm6,xmm0
+	punpcklwd xmm0,xmm2
+	punpckhwd xmm6,xmm2
+	movdqa    xmm5,xmm0
+	movdqa    xmm4,xmm6
+	pmaddwd   xmm0,[rel PW_F0299_F0337] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
+	pmaddwd   xmm6,[rel PW_F0299_F0337] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
+	pmaddwd   xmm5,[rel PW_MF016_MF033] ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+	pmaddwd   xmm4,[rel PW_MF016_MF033] ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+	movdqa    XMMWORD [wk(6)], xmm0	; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+	movdqa    XMMWORD [wk(7)], xmm6	; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+	pxor      xmm0,xmm0
+	pxor      xmm6,xmm6
+	punpcklwd xmm0,xmm1		; xmm0=BEL
+	punpckhwd xmm6,xmm1		; xmm6=BEH
+	psrld     xmm0,1		; xmm0=BEL*FIX(0.500)
+	psrld     xmm6,1		; xmm6=BEH*FIX(0.500)
+
+	movdqa    xmm1,[rel PD_ONEHALFM1_CJ] ; xmm1=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm5,xmm0
+	paddd     xmm4,xmm6
+	paddd     xmm5,xmm1
+	paddd     xmm4,xmm1
+	psrld     xmm5,SCALEBITS	; xmm5=CbEL
+	psrld     xmm4,SCALEBITS	; xmm4=CbEH
+	packssdw  xmm5,xmm4		; xmm5=CbE
+
+	psllw     xmm7,BYTE_BIT
+	por       xmm5,xmm7		; xmm5=Cb
+	movdqa    XMMWORD [rbx], xmm5	; Save Cb
+
+	movdqa    xmm0, XMMWORD [wk(3)]	; xmm0=BO
+	movdqa    xmm6, XMMWORD [wk(2)]	; xmm6=BE
+	movdqa    xmm1, XMMWORD [wk(1)]	; xmm1=RO
+
+	movdqa    xmm4,xmm0
+	punpcklwd xmm0,xmm3
+	punpckhwd xmm4,xmm3
+	movdqa    xmm7,xmm0
+	movdqa    xmm5,xmm4
+	pmaddwd   xmm0,[rel PW_F0114_F0250] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+	pmaddwd   xmm4,[rel PW_F0114_F0250] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+	pmaddwd   xmm7,[rel PW_MF008_MF041] ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+	pmaddwd   xmm5,[rel PW_MF008_MF041] ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+	movdqa    xmm3,[rel PD_ONEHALF]	; xmm3=[PD_ONEHALF]
+
+	paddd     xmm0, XMMWORD [wk(4)]
+	paddd     xmm4, XMMWORD [wk(5)]
+	paddd     xmm0,xmm3
+	paddd     xmm4,xmm3
+	psrld     xmm0,SCALEBITS	; xmm0=YOL
+	psrld     xmm4,SCALEBITS	; xmm4=YOH
+	packssdw  xmm0,xmm4		; xmm0=YO
+
+	pxor      xmm3,xmm3
+	pxor      xmm4,xmm4
+	punpcklwd xmm3,xmm1		; xmm3=ROL
+	punpckhwd xmm4,xmm1		; xmm4=ROH
+	psrld     xmm3,1		; xmm3=ROL*FIX(0.500)
+	psrld     xmm4,1		; xmm4=ROH*FIX(0.500)
+
+	movdqa    xmm1,[rel PD_ONEHALFM1_CJ] ; xmm1=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm7,xmm3
+	paddd     xmm5,xmm4
+	paddd     xmm7,xmm1
+	paddd     xmm5,xmm1
+	psrld     xmm7,SCALEBITS	; xmm7=CrOL
+	psrld     xmm5,SCALEBITS	; xmm5=CrOH
+	packssdw  xmm7,xmm5		; xmm7=CrO
+
+	movdqa    xmm3, XMMWORD [wk(0)]	; xmm3=RE
+
+	movdqa    xmm4,xmm6
+	punpcklwd xmm6,xmm2
+	punpckhwd xmm4,xmm2
+	movdqa    xmm1,xmm6
+	movdqa    xmm5,xmm4
+	pmaddwd   xmm6,[rel PW_F0114_F0250] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+	pmaddwd   xmm4,[rel PW_F0114_F0250] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+	pmaddwd   xmm1,[rel PW_MF008_MF041] ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+	pmaddwd   xmm5,[rel PW_MF008_MF041] ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+	movdqa    xmm2,[rel PD_ONEHALF]	; xmm2=[PD_ONEHALF]
+
+	paddd     xmm6, XMMWORD [wk(6)]
+	paddd     xmm4, XMMWORD [wk(7)]
+	paddd     xmm6,xmm2
+	paddd     xmm4,xmm2
+	psrld     xmm6,SCALEBITS	; xmm6=YEL
+	psrld     xmm4,SCALEBITS	; xmm4=YEH
+	packssdw  xmm6,xmm4		; xmm6=YE
+
+	psllw     xmm0,BYTE_BIT
+	por       xmm6,xmm0		; xmm6=Y
+	movdqa    XMMWORD [rdi], xmm6	; Save Y
+
+	pxor      xmm2,xmm2
+	pxor      xmm4,xmm4
+	punpcklwd xmm2,xmm3		; xmm2=REL
+	punpckhwd xmm4,xmm3		; xmm4=REH
+	psrld     xmm2,1		; xmm2=REL*FIX(0.500)
+	psrld     xmm4,1		; xmm4=REH*FIX(0.500)
+
+	movdqa    xmm0,[rel PD_ONEHALFM1_CJ] ; xmm0=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm1,xmm2
+	paddd     xmm5,xmm4
+	paddd     xmm1,xmm0
+	paddd     xmm5,xmm0
+	psrld     xmm1,SCALEBITS	; xmm1=CrEL
+	psrld     xmm5,SCALEBITS	; xmm5=CrEH
+	packssdw  xmm1,xmm5		; xmm1=CrE
+
+	psllw     xmm7,BYTE_BIT
+	por       xmm1,xmm7		; xmm1=Cr
+	movdqa    XMMWORD [rdx], xmm1	; Save Cr
+
+	sub	rcx, byte SIZEOF_XMMWORD
+	add	rsi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; inptr
+	add	rdi, byte SIZEOF_XMMWORD		; outptr0
+	add	rbx, byte SIZEOF_XMMWORD		; outptr1
+	add	rdx, byte SIZEOF_XMMWORD		; outptr2
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jae	near .columnloop
+	test	rcx,rcx
+	jnz	near .column_ld1
+
+	pop	rcx			; col
+	pop	rsi
+	pop	rdi
+	pop	rbx
+	pop	rdx
+
+	add	rsi, byte SIZEOF_JSAMPROW	; input_buf
+	add	rdi, byte SIZEOF_JSAMPROW
+	add	rbx, byte SIZEOF_JSAMPROW
+	add	rdx, byte SIZEOF_JSAMPROW
+	dec	rax				; num_rows
+	jg	near .rowloop
+
+.return:
+	pop	rbx
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcclrss2.asm b/simd/jcclrss2.asm
new file mode 100644
index 0000000..8def718
--- /dev/null
+++ b/simd/jcclrss2.asm
@@ -0,0 +1,505 @@
+;
+; jcclrss2.asm - colorspace conversion (SSE2)
+;
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_sse2 (JDIMENSION img_width,
+;                             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+;                             JDIMENSION output_row, int num_rows);
+;
+
+%define img_width(b)	(b)+8			; JDIMENSION img_width
+%define input_buf(b)	(b)+12		; JSAMPARRAY input_buf
+%define output_buf(b)	(b)+16		; JSAMPIMAGE output_buf
+%define output_row(b)	(b)+20		; JDIMENSION output_row
+%define num_rows(b)	(b)+24		; int num_rows
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		8
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+
+	global	EXTN(jsimd_rgb_ycc_convert_sse2)
+
+EXTN(jsimd_rgb_ycc_convert_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	ecx, JDIMENSION [img_width(eax)]
+	test	ecx,ecx
+	jz	near .return
+
+	push	ecx
+
+	mov	esi, JSAMPIMAGE [output_buf(eax)]
+	mov	ecx, JDIMENSION [output_row(eax)]
+	mov	edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
+	lea	edi, [edi+ecx*SIZEOF_JSAMPROW]
+	lea	ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+	lea	edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+	pop	ecx
+
+	mov	esi, JSAMPARRAY [input_buf(eax)]
+	mov	eax, INT [num_rows(eax)]
+	test	eax,eax
+	jle	near .return
+	alignx	16,7
+.rowloop:
+	pushpic	eax
+	push	edx
+	push	ebx
+	push	edi
+	push	esi
+	push	ecx			; col
+
+	mov	esi, JSAMPROW [esi]	; inptr
+	mov	edi, JSAMPROW [edi]	; outptr0
+	mov	ebx, JSAMPROW [ebx]	; outptr1
+	mov	edx, JSAMPROW [edx]	; outptr2
+	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jae	near .columnloop
+	alignx	16,7
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+.column_ld1:
+	push	eax
+	push	edx
+	lea	ecx,[ecx+ecx*2]		; imul ecx,RGB_PIXELSIZE
+	test	cl, SIZEOF_BYTE
+	jz	short .column_ld2
+	sub	ecx, byte SIZEOF_BYTE
+	movzx	eax, BYTE [esi+ecx]
+.column_ld2:
+	test	cl, SIZEOF_WORD
+	jz	short .column_ld4
+	sub	ecx, byte SIZEOF_WORD
+	movzx	edx, WORD [esi+ecx]
+	shl	eax, WORD_BIT
+	or	eax,edx
+.column_ld4:
+	movd	xmmA,eax
+	pop	edx
+	pop	eax
+	test	cl, SIZEOF_DWORD
+	jz	short .column_ld8
+	sub	ecx, byte SIZEOF_DWORD
+	movd	xmmF, XMM_DWORD [esi+ecx]
+	pslldq	xmmA, SIZEOF_DWORD
+	por	xmmA,xmmF
+.column_ld8:
+	test	cl, SIZEOF_MMWORD
+	jz	short .column_ld16
+	sub	ecx, byte SIZEOF_MMWORD
+	movq	xmmB, XMM_MMWORD [esi+ecx]
+	pslldq	xmmA, SIZEOF_MMWORD
+	por	xmmA,xmmB
+.column_ld16:
+	test	cl, SIZEOF_XMMWORD
+	jz	short .column_ld32
+	movdqa	xmmF,xmmA
+	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	mov	ecx, SIZEOF_XMMWORD
+	jmp	short .rgb_ycc_cnv
+.column_ld32:
+	test	cl, 2*SIZEOF_XMMWORD
+	mov	ecx, SIZEOF_XMMWORD
+	jz	short .rgb_ycc_cnv
+	movdqa	xmmB,xmmA
+	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqu	xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
+	jmp	short .rgb_ycc_cnv
+	alignx	16,7
+
+.columnloop:
+	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqu	xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
+	movdqu	xmmB, XMMWORD [esi+2*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+	; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+	; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+	movdqa    xmmG,xmmA
+	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
+	psrldq    xmmG,8	; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
+
+	punpckhbw xmmA,xmmF	; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
+	pslldq    xmmF,8	; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
+
+	punpcklbw xmmG,xmmB	; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
+	punpckhbw xmmF,xmmB	; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
+
+	movdqa    xmmD,xmmA
+	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
+	psrldq    xmmD,8	; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
+
+	punpckhbw xmmA,xmmG	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
+	pslldq    xmmG,8	; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
+
+	punpcklbw xmmD,xmmF	; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
+	punpckhbw xmmG,xmmF	; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
+
+	movdqa    xmmE,xmmA
+	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
+	psrldq    xmmE,8	; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
+
+	punpckhbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+	pslldq    xmmD,8	; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
+
+	punpcklbw xmmE,xmmG	; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
+	punpckhbw xmmD,xmmG	; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
+
+	pxor      xmmH,xmmH
+
+	movdqa    xmmC,xmmA
+	punpcklbw xmmA,xmmH	; xmmA=(00 02 04 06 08 0A 0C 0E)
+	punpckhbw xmmC,xmmH	; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+	movdqa    xmmB,xmmE
+	punpcklbw xmmE,xmmH	; xmmE=(20 22 24 26 28 2A 2C 2E)
+	punpckhbw xmmB,xmmH	; xmmB=(01 03 05 07 09 0B 0D 0F)
+
+	movdqa    xmmF,xmmD
+	punpcklbw xmmD,xmmH	; xmmD=(11 13 15 17 19 1B 1D 1F)
+	punpckhbw xmmF,xmmH	; xmmF=(21 23 25 27 29 2B 2D 2F)
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+	test	cl, SIZEOF_XMMWORD/16
+	jz	short .column_ld2
+	sub	ecx, byte SIZEOF_XMMWORD/16
+	movd	xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+	test	cl, SIZEOF_XMMWORD/8
+	jz	short .column_ld4
+	sub	ecx, byte SIZEOF_XMMWORD/8
+	movq	xmmE, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
+	pslldq	xmmA, SIZEOF_MMWORD
+	por	xmmA,xmmE
+.column_ld4:
+	test	cl, SIZEOF_XMMWORD/4
+	jz	short .column_ld8
+	sub	ecx, byte SIZEOF_XMMWORD/4
+	movdqa	xmmE,xmmA
+	movdqu	xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld8:
+	test	cl, SIZEOF_XMMWORD/2
+	mov	ecx, SIZEOF_XMMWORD
+	jz	short .rgb_ycc_cnv
+	movdqa	xmmF,xmmA
+	movdqa	xmmH,xmmE
+	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqu	xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
+	jmp	short .rgb_ycc_cnv
+	alignx	16,7
+
+.columnloop:
+	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqu	xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
+	movdqu	xmmF, XMMWORD [esi+2*SIZEOF_XMMWORD]
+	movdqu	xmmH, XMMWORD [esi+3*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+	; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+	; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+	movdqa    xmmD,xmmA
+	punpcklbw xmmA,xmmE	; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
+	punpckhbw xmmD,xmmE	; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
+
+	movdqa    xmmC,xmmF
+	punpcklbw xmmF,xmmH	; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
+	punpckhbw xmmC,xmmH	; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
+
+	movdqa    xmmB,xmmA
+	punpcklwd xmmA,xmmF	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
+	punpckhwd xmmB,xmmF	; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
+
+	movdqa    xmmG,xmmD
+	punpcklwd xmmD,xmmC	; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
+	punpckhwd xmmG,xmmC	; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
+
+	movdqa    xmmE,xmmA
+	punpcklbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+	punpckhbw xmmE,xmmD	; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
+
+	movdqa    xmmH,xmmB
+	punpcklbw xmmB,xmmG	; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
+	punpckhbw xmmH,xmmG	; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
+
+	pxor      xmmF,xmmF
+
+	movdqa    xmmC,xmmA
+	punpcklbw xmmA,xmmF	; xmmA=(00 02 04 06 08 0A 0C 0E)
+	punpckhbw xmmC,xmmF	; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+	movdqa    xmmD,xmmB
+	punpcklbw xmmB,xmmF	; xmmB=(01 03 05 07 09 0B 0D 0F)
+	punpckhbw xmmD,xmmF	; xmmD=(11 13 15 17 19 1B 1D 1F)
+
+	movdqa    xmmG,xmmE
+	punpcklbw xmmE,xmmF	; xmmE=(20 22 24 26 28 2A 2C 2E)
+	punpckhbw xmmG,xmmF	; xmmG=(30 32 34 36 38 3A 3C 3E)
+
+	punpcklbw xmmF,xmmH
+	punpckhbw xmmH,xmmH
+	psrlw     xmmF,BYTE_BIT	; xmmF=(21 23 25 27 29 2B 2D 2F)
+	psrlw     xmmH,BYTE_BIT	; xmmH=(31 33 35 37 39 3B 3D 3F)
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+	; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
+	; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
+
+	; (Original)
+	; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+	;
+	; (This implementation)
+	; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+	movdqa    XMMWORD [wk(0)], xmm0	; wk(0)=RE
+	movdqa    XMMWORD [wk(1)], xmm1	; wk(1)=RO
+	movdqa    XMMWORD [wk(2)], xmm4	; wk(2)=BE
+	movdqa    XMMWORD [wk(3)], xmm5	; wk(3)=BO
+
+	movdqa    xmm6,xmm1
+	punpcklwd xmm1,xmm3
+	punpckhwd xmm6,xmm3
+	movdqa    xmm7,xmm1
+	movdqa    xmm4,xmm6
+	pmaddwd   xmm1,[GOTOFF(eax,PW_F0299_F0337)] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+	pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+	pmaddwd   xmm7,[GOTOFF(eax,PW_MF016_MF033)] ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+	pmaddwd   xmm4,[GOTOFF(eax,PW_MF016_MF033)] ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+	movdqa    XMMWORD [wk(4)], xmm1	; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+	movdqa    XMMWORD [wk(5)], xmm6	; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+	pxor      xmm1,xmm1
+	pxor      xmm6,xmm6
+	punpcklwd xmm1,xmm5		; xmm1=BOL
+	punpckhwd xmm6,xmm5		; xmm6=BOH
+	psrld     xmm1,1		; xmm1=BOL*FIX(0.500)
+	psrld     xmm6,1		; xmm6=BOH*FIX(0.500)
+
+	movdqa    xmm5,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm5=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm7,xmm1
+	paddd     xmm4,xmm6
+	paddd     xmm7,xmm5
+	paddd     xmm4,xmm5
+	psrld     xmm7,SCALEBITS	; xmm7=CbOL
+	psrld     xmm4,SCALEBITS	; xmm4=CbOH
+	packssdw  xmm7,xmm4		; xmm7=CbO
+
+	movdqa    xmm1, XMMWORD [wk(2)]	; xmm1=BE
+
+	movdqa    xmm6,xmm0
+	punpcklwd xmm0,xmm2
+	punpckhwd xmm6,xmm2
+	movdqa    xmm5,xmm0
+	movdqa    xmm4,xmm6
+	pmaddwd   xmm0,[GOTOFF(eax,PW_F0299_F0337)] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
+	pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
+	pmaddwd   xmm5,[GOTOFF(eax,PW_MF016_MF033)] ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+	pmaddwd   xmm4,[GOTOFF(eax,PW_MF016_MF033)] ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+	movdqa    XMMWORD [wk(6)], xmm0	; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+	movdqa    XMMWORD [wk(7)], xmm6	; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+	pxor      xmm0,xmm0
+	pxor      xmm6,xmm6
+	punpcklwd xmm0,xmm1		; xmm0=BEL
+	punpckhwd xmm6,xmm1		; xmm6=BEH
+	psrld     xmm0,1		; xmm0=BEL*FIX(0.500)
+	psrld     xmm6,1		; xmm6=BEH*FIX(0.500)
+
+	movdqa    xmm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm1=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm5,xmm0
+	paddd     xmm4,xmm6
+	paddd     xmm5,xmm1
+	paddd     xmm4,xmm1
+	psrld     xmm5,SCALEBITS	; xmm5=CbEL
+	psrld     xmm4,SCALEBITS	; xmm4=CbEH
+	packssdw  xmm5,xmm4		; xmm5=CbE
+
+	psllw     xmm7,BYTE_BIT
+	por       xmm5,xmm7		; xmm5=Cb
+	movdqa    XMMWORD [ebx], xmm5	; Save Cb
+
+	movdqa    xmm0, XMMWORD [wk(3)]	; xmm0=BO
+	movdqa    xmm6, XMMWORD [wk(2)]	; xmm6=BE
+	movdqa    xmm1, XMMWORD [wk(1)]	; xmm1=RO
+
+	movdqa    xmm4,xmm0
+	punpcklwd xmm0,xmm3
+	punpckhwd xmm4,xmm3
+	movdqa    xmm7,xmm0
+	movdqa    xmm5,xmm4
+	pmaddwd   xmm0,[GOTOFF(eax,PW_F0114_F0250)] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+	pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+	pmaddwd   xmm7,[GOTOFF(eax,PW_MF008_MF041)] ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+	pmaddwd   xmm5,[GOTOFF(eax,PW_MF008_MF041)] ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+	movdqa    xmm3,[GOTOFF(eax,PD_ONEHALF)]	; xmm3=[PD_ONEHALF]
+
+	paddd     xmm0, XMMWORD [wk(4)]
+	paddd     xmm4, XMMWORD [wk(5)]
+	paddd     xmm0,xmm3
+	paddd     xmm4,xmm3
+	psrld     xmm0,SCALEBITS	; xmm0=YOL
+	psrld     xmm4,SCALEBITS	; xmm4=YOH
+	packssdw  xmm0,xmm4		; xmm0=YO
+
+	pxor      xmm3,xmm3
+	pxor      xmm4,xmm4
+	punpcklwd xmm3,xmm1		; xmm3=ROL
+	punpckhwd xmm4,xmm1		; xmm4=ROH
+	psrld     xmm3,1		; xmm3=ROL*FIX(0.500)
+	psrld     xmm4,1		; xmm4=ROH*FIX(0.500)
+
+	movdqa    xmm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm1=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm7,xmm3
+	paddd     xmm5,xmm4
+	paddd     xmm7,xmm1
+	paddd     xmm5,xmm1
+	psrld     xmm7,SCALEBITS	; xmm7=CrOL
+	psrld     xmm5,SCALEBITS	; xmm5=CrOH
+	packssdw  xmm7,xmm5		; xmm7=CrO
+
+	movdqa    xmm3, XMMWORD [wk(0)]	; xmm3=RE
+
+	movdqa    xmm4,xmm6
+	punpcklwd xmm6,xmm2
+	punpckhwd xmm4,xmm2
+	movdqa    xmm1,xmm6
+	movdqa    xmm5,xmm4
+	pmaddwd   xmm6,[GOTOFF(eax,PW_F0114_F0250)] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+	pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+	pmaddwd   xmm1,[GOTOFF(eax,PW_MF008_MF041)] ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+	pmaddwd   xmm5,[GOTOFF(eax,PW_MF008_MF041)] ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+	movdqa    xmm2,[GOTOFF(eax,PD_ONEHALF)]	; xmm2=[PD_ONEHALF]
+
+	paddd     xmm6, XMMWORD [wk(6)]
+	paddd     xmm4, XMMWORD [wk(7)]
+	paddd     xmm6,xmm2
+	paddd     xmm4,xmm2
+	psrld     xmm6,SCALEBITS	; xmm6=YEL
+	psrld     xmm4,SCALEBITS	; xmm4=YEH
+	packssdw  xmm6,xmm4		; xmm6=YE
+
+	psllw     xmm0,BYTE_BIT
+	por       xmm6,xmm0		; xmm6=Y
+	movdqa    XMMWORD [edi], xmm6	; Save Y
+
+	pxor      xmm2,xmm2
+	pxor      xmm4,xmm4
+	punpcklwd xmm2,xmm3		; xmm2=REL
+	punpckhwd xmm4,xmm3		; xmm4=REH
+	psrld     xmm2,1		; xmm2=REL*FIX(0.500)
+	psrld     xmm4,1		; xmm4=REH*FIX(0.500)
+
+	movdqa    xmm0,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm0=[PD_ONEHALFM1_CJ]
+
+	paddd     xmm1,xmm2
+	paddd     xmm5,xmm4
+	paddd     xmm1,xmm0
+	paddd     xmm5,xmm0
+	psrld     xmm1,SCALEBITS	; xmm1=CrEL
+	psrld     xmm5,SCALEBITS	; xmm5=CrEH
+	packssdw  xmm1,xmm5		; xmm1=CrE
+
+	psllw     xmm7,BYTE_BIT
+	por       xmm1,xmm7		; xmm1=Cr
+	movdqa    XMMWORD [edx], xmm1	; Save Cr
+
+	sub	ecx, byte SIZEOF_XMMWORD
+	add	esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; inptr
+	add	edi, byte SIZEOF_XMMWORD		; outptr0
+	add	ebx, byte SIZEOF_XMMWORD		; outptr1
+	add	edx, byte SIZEOF_XMMWORD		; outptr2
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jae	near .columnloop
+	test	ecx,ecx
+	jnz	near .column_ld1
+
+	pop	ecx			; col
+	pop	esi
+	pop	edi
+	pop	ebx
+	pop	edx
+	poppic	eax
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_buf
+	add	edi, byte SIZEOF_JSAMPROW
+	add	ebx, byte SIZEOF_JSAMPROW
+	add	edx, byte SIZEOF_JSAMPROW
+	dec	eax				; num_rows
+	jg	near .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jccolmmx.asm b/simd/jccolmmx.asm
new file mode 100644
index 0000000..5e7f3be
--- /dev/null
+++ b/simd/jccolmmx.asm
@@ -0,0 +1,120 @@
+;
+; jccolmmx.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_081	equ	 5329			; FIX(0.08131)
+F_0_114	equ	 7471			; FIX(0.11400)
+F_0_168	equ	11059			; FIX(0.16874)
+F_0_250	equ	16384			; FIX(0.25000)
+F_0_299	equ	19595			; FIX(0.29900)
+F_0_331	equ	21709			; FIX(0.33126)
+F_0_418	equ	27439			; FIX(0.41869)
+F_0_587	equ	38470			; FIX(0.58700)
+F_0_337	equ	(F_0_587 - F_0_250)	; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_rgb_ycc_convert_mmx)
+
+EXTN(jconst_rgb_ycc_convert_mmx):
+
+PW_F0299_F0337	times 2 dw  F_0_299, F_0_337
+PW_F0114_F0250	times 2 dw  F_0_114, F_0_250
+PW_MF016_MF033	times 2 dw -F_0_168,-F_0_331
+PW_MF008_MF041	times 2 dw -F_0_081,-F_0_418
+PD_ONEHALFM1_CJ	times 2 dd  (1 << (SCALEBITS-1)) - 1 + (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF	times 2 dd  (1 << (SCALEBITS-1))
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jcclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_rgb_ycc_convert_mmx jsimd_extrgb_ycc_convert_mmx
+%include "jcclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_mmx jsimd_extrgbx_ycc_convert_mmx
+%include "jcclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_rgb_ycc_convert_mmx jsimd_extbgr_ycc_convert_mmx
+%include "jcclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_mmx jsimd_extbgrx_ycc_convert_mmx
+%include "jcclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_mmx jsimd_extxbgr_ycc_convert_mmx
+%include "jcclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_mmx jsimd_extxrgb_ycc_convert_mmx
+%include "jcclrmmx.asm"
diff --git a/simd/jccolss2-64.asm b/simd/jccolss2-64.asm
new file mode 100644
index 0000000..64ee0ba
--- /dev/null
+++ b/simd/jccolss2-64.asm
@@ -0,0 +1,117 @@
+;
+; jccolss2-64.asm - colorspace conversion (64-bit SSE2)
+;
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; Copyright (C) 2009, D. R. Commander.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_081	equ	 5329			; FIX(0.08131)
+F_0_114	equ	 7471			; FIX(0.11400)
+F_0_168	equ	11059			; FIX(0.16874)
+F_0_250	equ	16384			; FIX(0.25000)
+F_0_299	equ	19595			; FIX(0.29900)
+F_0_331	equ	21709			; FIX(0.33126)
+F_0_418	equ	27439			; FIX(0.41869)
+F_0_587	equ	38470			; FIX(0.58700)
+F_0_337	equ	(F_0_587 - F_0_250)	; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_rgb_ycc_convert_sse2)
+
+EXTN(jconst_rgb_ycc_convert_sse2):
+
+PW_F0299_F0337	times 4 dw  F_0_299, F_0_337
+PW_F0114_F0250	times 4 dw  F_0_114, F_0_250
+PW_MF016_MF033	times 4 dw -F_0_168,-F_0_331
+PW_MF008_MF041	times 4 dw -F_0_081,-F_0_418
+PD_ONEHALFM1_CJ	times 4 dd  (1 << (SCALEBITS-1)) - 1 + (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF	times 4 dd  (1 << (SCALEBITS-1))
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jcclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extrgb_ycc_convert_sse2
+%include "jcclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extrgbx_ycc_convert_sse2
+%include "jcclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extbgr_ycc_convert_sse2
+%include "jcclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extbgrx_ycc_convert_sse2
+%include "jcclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extxbgr_ycc_convert_sse2
+%include "jcclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extxrgb_ycc_convert_sse2
+%include "jcclrss2-64.asm"
diff --git a/simd/jccolss2.asm b/simd/jccolss2.asm
new file mode 100644
index 0000000..8d1f734
--- /dev/null
+++ b/simd/jccolss2.asm
@@ -0,0 +1,117 @@
+;
+; jccolss2.asm - colorspace conversion (SSE2)
+;
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; Copyright (C) 2009, D. R. Commander.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_081	equ	 5329			; FIX(0.08131)
+F_0_114	equ	 7471			; FIX(0.11400)
+F_0_168	equ	11059			; FIX(0.16874)
+F_0_250	equ	16384			; FIX(0.25000)
+F_0_299	equ	19595			; FIX(0.29900)
+F_0_331	equ	21709			; FIX(0.33126)
+F_0_418	equ	27439			; FIX(0.41869)
+F_0_587	equ	38470			; FIX(0.58700)
+F_0_337	equ	(F_0_587 - F_0_250)	; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_rgb_ycc_convert_sse2)
+
+EXTN(jconst_rgb_ycc_convert_sse2):
+
+PW_F0299_F0337	times 4 dw  F_0_299, F_0_337
+PW_F0114_F0250	times 4 dw  F_0_114, F_0_250
+PW_MF016_MF033	times 4 dw -F_0_168,-F_0_331
+PW_MF008_MF041	times 4 dw -F_0_081,-F_0_418
+PD_ONEHALFM1_CJ	times 4 dd  (1 << (SCALEBITS-1)) - 1 + (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF	times 4 dd  (1 << (SCALEBITS-1))
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jcclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extrgb_ycc_convert_sse2
+%include "jcclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extrgbx_ycc_convert_sse2
+%include "jcclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extbgr_ycc_convert_sse2
+%include "jcclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extbgrx_ycc_convert_sse2
+%include "jcclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extxbgr_ycc_convert_sse2
+%include "jcclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_rgb_ycc_convert_sse2 jsimd_extxrgb_ycc_convert_sse2
+%include "jcclrss2.asm"
diff --git a/simd/jcolsamp.inc b/simd/jcolsamp.inc
new file mode 100644
index 0000000..79751b7
--- /dev/null
+++ b/simd/jcolsamp.inc
@@ -0,0 +1,105 @@
+;
+; jcolsamp.inc - private declarations for color conversion & up/downsampling
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; [TAB8]
+
+; --------------------------------------------------------------------------
+
+; pseudo-resisters to make ordering of RGB configurable
+;
+%if RGB_RED == 0
+%define  mmA  mm0
+%define  mmB  mm1
+%define xmmA xmm0
+%define xmmB xmm1
+%elif RGB_GREEN == 0
+%define  mmA  mm2
+%define  mmB  mm3
+%define xmmA xmm2
+%define xmmB xmm3
+%elif RGB_BLUE == 0
+%define  mmA  mm4
+%define  mmB  mm5
+%define xmmA xmm4
+%define xmmB xmm5
+%else
+%define  mmA  mm6
+%define  mmB  mm7
+%define xmmA xmm6
+%define xmmB xmm7
+%endif
+
+%if RGB_RED == 1
+%define  mmC  mm0
+%define  mmD  mm1
+%define xmmC xmm0
+%define xmmD xmm1
+%elif RGB_GREEN == 1
+%define  mmC  mm2
+%define  mmD  mm3
+%define xmmC xmm2
+%define xmmD xmm3
+%elif RGB_BLUE == 1
+%define  mmC  mm4
+%define  mmD  mm5
+%define xmmC xmm4
+%define xmmD xmm5
+%else
+%define  mmC  mm6
+%define  mmD  mm7
+%define xmmC xmm6
+%define xmmD xmm7
+%endif
+
+%if RGB_RED == 2
+%define  mmE  mm0
+%define  mmF  mm1
+%define xmmE xmm0
+%define xmmF xmm1
+%elif RGB_GREEN == 2
+%define  mmE  mm2
+%define  mmF  mm3
+%define xmmE xmm2
+%define xmmF xmm3
+%elif RGB_BLUE == 2
+%define  mmE  mm4
+%define  mmF  mm5
+%define xmmE xmm4
+%define xmmF xmm5
+%else
+%define  mmE  mm6
+%define  mmF  mm7
+%define xmmE xmm6
+%define xmmF xmm7
+%endif
+
+%if RGB_RED == 3
+%define  mmG  mm0
+%define  mmH  mm1
+%define xmmG xmm0
+%define xmmH xmm1
+%elif RGB_GREEN == 3
+%define  mmG  mm2
+%define  mmH  mm3
+%define xmmG xmm2
+%define xmmH xmm3
+%elif RGB_BLUE == 3
+%define  mmG  mm4
+%define  mmH  mm5
+%define xmmG xmm4
+%define xmmH xmm5
+%else
+%define  mmG  mm6
+%define  mmH  mm7
+%define xmmG xmm6
+%define xmmH xmm7
+%endif
+
+; --------------------------------------------------------------------------
diff --git a/simd/jcqnt3dn.asm b/simd/jcqnt3dn.asm
new file mode 100644
index 0000000..182c869
--- /dev/null
+++ b/simd/jcqnt3dn.asm
@@ -0,0 +1,233 @@
+;
+; jcqnt3dn.asm - sample data conversion and quantization (3DNow! & MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_float_3dnow (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                             FAST_FLOAT * workspace);
+;
+
+%define sample_data	ebp+8		; JSAMPARRAY sample_data
+%define start_col	ebp+12		; JDIMENSION start_col
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_float_3dnow)
+
+EXTN(jsimd_convsamp_float_3dnow):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	pcmpeqw  mm7,mm7
+	psllw    mm7,7
+	packsswb mm7,mm7		; mm7 = PB_CENTERJSAMPLE (0x808080..)
+
+	mov	esi, JSAMPARRAY [sample_data]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [start_col]
+	mov	edi, POINTER [workspace]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/2
+	alignx	16,7
+.convloop:
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE]
+	movq	mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE]
+
+	psubb	mm0,mm7				; mm0=(01234567)
+	psubb	mm1,mm7				; mm1=(89ABCDEF)
+
+	punpcklbw mm2,mm0			; mm2=(*0*1*2*3)
+	punpckhbw mm0,mm0			; mm0=(*4*5*6*7)
+	punpcklbw mm3,mm1			; mm3=(*8*9*A*B)
+	punpckhbw mm1,mm1			; mm1=(*C*D*E*F)
+
+	punpcklwd mm4,mm2			; mm4=(***0***1)
+	punpckhwd mm2,mm2			; mm2=(***2***3)
+	punpcklwd mm5,mm0			; mm5=(***4***5)
+	punpckhwd mm0,mm0			; mm0=(***6***7)
+
+	psrad	mm4,(DWORD_BIT-BYTE_BIT)	; mm4=(01)
+	psrad	mm2,(DWORD_BIT-BYTE_BIT)	; mm2=(23)
+	pi2fd	mm4,mm4
+	pi2fd	mm2,mm2
+	psrad	mm5,(DWORD_BIT-BYTE_BIT)	; mm5=(45)
+	psrad	mm0,(DWORD_BIT-BYTE_BIT)	; mm0=(67)
+	pi2fd	mm5,mm5
+	pi2fd	mm0,mm0
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm4
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm2
+	movq	MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm5
+	movq	MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
+
+	punpcklwd mm6,mm3			; mm6=(***8***9)
+	punpckhwd mm3,mm3			; mm3=(***A***B)
+	punpcklwd mm4,mm1			; mm4=(***C***D)
+	punpckhwd mm1,mm1			; mm1=(***E***F)
+
+	psrad	mm6,(DWORD_BIT-BYTE_BIT)	; mm6=(89)
+	psrad	mm3,(DWORD_BIT-BYTE_BIT)	; mm3=(AB)
+	pi2fd	mm6,mm6
+	pi2fd	mm3,mm3
+	psrad	mm4,(DWORD_BIT-BYTE_BIT)	; mm4=(CD)
+	psrad	mm1,(DWORD_BIT-BYTE_BIT)	; mm1=(EF)
+	pi2fd	mm4,mm4
+	pi2fd	mm1,mm1
+
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm6
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm3
+	movq	MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm4
+	movq	MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm1
+
+	add	esi, byte 2*SIZEOF_JSAMPROW
+	add	edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .convloop
+
+	femms		; empty MMX/3DNow! state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; GLOBAL(void)
+; jsimd_quantize_float_3dnow (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+;                             FAST_FLOAT * workspace);
+;
+
+%define coef_block	ebp+8		; JCOEFPTR coef_block
+%define divisors	ebp+12		; FAST_FLOAT * divisors
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_float_3dnow)
+
+EXTN(jsimd_quantize_float_3dnow):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov       eax, 0x4B400000	; (float)0x00C00000 (rndint_magic)
+	movd      mm7,eax
+	punpckldq mm7,mm7		; mm7={12582912.0F 12582912.0F}
+
+	mov	esi, POINTER [workspace]
+	mov	edx, POINTER [divisors]
+	mov	edi, JCOEFPTR [coef_block]
+	mov	eax, DCTSIZE2/16
+	alignx	16,7
+.quantloop:
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm1, MMWORD [MMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)]
+	pfmul	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	pfmul	mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm2, MMWORD [MMBLOCK(0,2,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(0,3,esi,SIZEOF_FAST_FLOAT)]
+	pfmul	mm2, MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)]
+	pfmul	mm3, MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)]
+
+	pfadd	mm0,mm7			; mm0=(00 ** 01 **)
+	pfadd	mm1,mm7			; mm1=(02 ** 03 **)
+	pfadd	mm2,mm7			; mm0=(04 ** 05 **)
+	pfadd	mm3,mm7			; mm1=(06 ** 07 **)
+
+	movq      mm4,mm0
+	punpcklwd mm0,mm1		; mm0=(00 02 ** **)
+	punpckhwd mm4,mm1		; mm4=(01 03 ** **)
+	movq      mm5,mm2
+	punpcklwd mm2,mm3		; mm2=(04 06 ** **)
+	punpckhwd mm5,mm3		; mm5=(05 07 ** **)
+
+	punpcklwd mm0,mm4		; mm0=(00 01 02 03)
+	punpcklwd mm2,mm5		; mm2=(04 05 06 07)
+
+	movq	mm6, MMWORD [MMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm1, MMWORD [MMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)]
+	pfmul	mm6, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	pfmul	mm1, MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(1,2,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm4, MMWORD [MMBLOCK(1,3,esi,SIZEOF_FAST_FLOAT)]
+	pfmul	mm3, MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)]
+	pfmul	mm4, MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)]
+
+	pfadd	mm6,mm7			; mm0=(10 ** 11 **)
+	pfadd	mm1,mm7			; mm4=(12 ** 13 **)
+	pfadd	mm3,mm7			; mm0=(14 ** 15 **)
+	pfadd	mm4,mm7			; mm4=(16 ** 17 **)
+
+	movq      mm5,mm6
+	punpcklwd mm6,mm1		; mm6=(10 12 ** **)
+	punpckhwd mm5,mm1		; mm5=(11 13 ** **)
+	movq      mm1,mm3
+	punpcklwd mm3,mm4		; mm3=(14 16 ** **)
+	punpckhwd mm1,mm4		; mm1=(15 17 ** **)
+
+	punpcklwd mm6,mm5		; mm6=(10 11 12 13)
+	punpcklwd mm3,mm1		; mm3=(14 15 16 17)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm6
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3
+
+	add	esi, byte 16*SIZEOF_FAST_FLOAT
+	add	edx, byte 16*SIZEOF_FAST_FLOAT
+	add	edi, byte 16*SIZEOF_JCOEF
+	dec	eax
+	jnz	near .quantloop
+
+	femms		; empty MMX/3DNow! state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcqntmmx.asm b/simd/jcqntmmx.asm
new file mode 100644
index 0000000..08b08b7
--- /dev/null
+++ b/simd/jcqntmmx.asm
@@ -0,0 +1,274 @@
+;
+; jcqntmmx.asm - sample data conversion and quantization (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_mmx (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                     DCTELEM * workspace);
+;
+
+%define sample_data	ebp+8		; JSAMPARRAY sample_data
+%define start_col	ebp+12		; JDIMENSION start_col
+%define workspace	ebp+16		; DCTELEM * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_mmx)
+
+EXTN(jsimd_convsamp_mmx):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	pxor	mm6,mm6			; mm6=(all 0's)
+	pcmpeqw	mm7,mm7
+	psllw	mm7,7			; mm7={0xFF80 0xFF80 0xFF80 0xFF80}
+
+	mov	esi, JSAMPARRAY [sample_data]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [start_col]
+	mov	edi, POINTER [workspace]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.convloop:
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE]	; mm0=(01234567)
+	movq	mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE]	; mm1=(89ABCDEF)
+
+	mov	ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	mm2, MMWORD [ebx+eax*SIZEOF_JSAMPLE]	; mm2=(GHIJKLMN)
+	movq	mm3, MMWORD [edx+eax*SIZEOF_JSAMPLE]	; mm3=(OPQRSTUV)
+
+	movq      mm4,mm0
+	punpcklbw mm0,mm6		; mm0=(0123)
+	punpckhbw mm4,mm6		; mm4=(4567)
+	movq      mm5,mm1
+	punpcklbw mm1,mm6		; mm1=(89AB)
+	punpckhbw mm5,mm6		; mm5=(CDEF)
+
+	paddw	mm0,mm7
+	paddw	mm4,mm7
+	paddw	mm1,mm7
+	paddw	mm5,mm7
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm4
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_DCTELEM)], mm1
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_DCTELEM)], mm5
+
+	movq      mm0,mm2
+	punpcklbw mm2,mm6		; mm2=(GHIJ)
+	punpckhbw mm0,mm6		; mm0=(KLMN)
+	movq      mm4,mm3
+	punpcklbw mm3,mm6		; mm3=(OPQR)
+	punpckhbw mm4,mm6		; mm4=(STUV)
+
+	paddw	mm2,mm7
+	paddw	mm0,mm7
+	paddw	mm3,mm7
+	paddw	mm4,mm7
+
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_DCTELEM)], mm2
+	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_DCTELEM)], mm0
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_DCTELEM)], mm3
+	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_DCTELEM)], mm4
+
+	add	esi, byte 4*SIZEOF_JSAMPROW
+	add	edi, byte 4*DCTSIZE*SIZEOF_DCTELEM
+	dec	ecx
+	jnz	short .convloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; This implementation is based on an algorithm described in
+;   "How to optimize for the Pentium family of microprocessors"
+;   (http://www.agner.org/assem/).
+;
+; GLOBAL(void)
+; jsimd_quantize_mmx (JCOEFPTR coef_block, DCTELEM * divisors,
+;                     DCTELEM * workspace);
+;
+
+%define RECIPROCAL(m,n,b) MMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
+%define CORRECTION(m,n,b) MMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
+%define SCALE(m,n,b)      MMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
+%define SHIFT(m,n,b)      MMBLOCK(DCTSIZE*3+(m),(n),(b),SIZEOF_DCTELEM)
+
+%define coef_block	ebp+8		; JCOEFPTR coef_block
+%define divisors	ebp+12		; DCTELEM * divisors
+%define workspace	ebp+16		; DCTELEM * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_mmx)
+
+EXTN(jsimd_quantize_mmx):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	esi, POINTER [workspace]
+	mov	edx, POINTER [divisors]
+	mov	edi, JCOEFPTR [coef_block]
+	mov	ah, 2
+	alignx	16,7
+.quantloop1:
+	mov	al, DCTSIZE2/8/2
+	alignx	16,7
+.quantloop2:
+	movq	mm2, MMWORD [MMBLOCK(0,0,esi,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(0,1,esi,SIZEOF_DCTELEM)]
+
+	movq	mm0,mm2
+	movq	mm1,mm3
+
+	psraw	mm2,(WORD_BIT-1)  ; -1 if value < 0, 0 otherwise
+	psraw	mm3,(WORD_BIT-1)
+
+	pxor	mm0,mm2   ; val = -val
+	pxor	mm1,mm3
+	psubw	mm0,mm2
+	psubw	mm1,mm3
+
+	;
+	; MMX is an annoyingly crappy instruction set. It has two
+	; misfeatures that are causing problems here:
+	;
+	; - All multiplications are signed.
+	;
+	; - The second operand for the shifts is not treated as packed.
+	;
+	;
+	; We work around the first problem by implementing this algorithm:
+	;
+	; unsigned long unsigned_multiply(unsigned short x, unsigned short y)
+	; {
+	;   enum { SHORT_BIT = 16 };
+	;   signed short sx = (signed short) x;
+	;   signed short sy = (signed short) y;
+	;   signed long sz;
+	; 
+	;   sz = (long) sx * (long) sy;     /* signed multiply */
+	; 
+	;   if (sx < 0) sz += (long) sy << SHORT_BIT;
+	;   if (sy < 0) sz += (long) sx << SHORT_BIT;
+	; 
+	;   return (unsigned long) sz;
+	; }
+	;
+	; (note that a negative sx adds _sy_ and vice versa)
+	;
+	; For the second problem, we replace the shift by a multiplication.
+	; Unfortunately that means we have to deal with the signed issue again.
+	;
+
+	paddw	mm0, MMWORD [CORRECTION(0,0,edx)]   ; correction + roundfactor
+	paddw	mm1, MMWORD [CORRECTION(0,1,edx)]
+
+	movq	mm4,mm0   ; store current value for later
+	movq	mm5,mm1
+	pmulhw	mm0, MMWORD [RECIPROCAL(0,0,edx)]   ; reciprocal
+	pmulhw	mm1, MMWORD [RECIPROCAL(0,1,edx)]
+	paddw	mm0,mm4		; reciprocal is always negative (MSB=1),
+	paddw	mm1,mm5   ; so we always need to add the initial value
+	                ; (input value is never negative as we
+	                ; inverted it at the start of this routine)
+
+	; here it gets a bit tricky as both scale
+	; and mm0/mm1 can be negative
+	movq	mm6, MMWORD [SCALE(0,0,edx)]	; scale
+	movq	mm7, MMWORD [SCALE(0,1,edx)]
+	movq	mm4,mm0
+	movq	mm5,mm1
+	pmulhw	mm0,mm6
+	pmulhw	mm1,mm7
+
+	psraw	mm6,(WORD_BIT-1)    ; determine if scale is negative
+	psraw	mm7,(WORD_BIT-1)
+
+	pand	mm6,mm4             ; and add input if it is
+	pand	mm7,mm5
+	paddw	mm0,mm6
+	paddw	mm1,mm7
+
+	psraw	mm4,(WORD_BIT-1)    ; then check if negative input 
+	psraw	mm5,(WORD_BIT-1)
+
+	pand	mm4, MMWORD [SCALE(0,0,edx)]	; and add scale if it is
+	pand	mm5, MMWORD [SCALE(0,1,edx)]
+	paddw	mm0,mm4
+	paddw	mm1,mm5
+
+	pxor	mm0,mm2   ; val = -val
+	pxor	mm1,mm3
+	psubw	mm0,mm2
+	psubw	mm1,mm3
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm1
+
+	add	esi, byte 8*SIZEOF_DCTELEM
+	add	edx, byte 8*SIZEOF_DCTELEM
+	add	edi, byte 8*SIZEOF_JCOEF
+	dec	al
+	jnz	near .quantloop2
+	dec	ah
+	jnz	near .quantloop1	; to avoid branch misprediction
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcqnts2f-64.asm b/simd/jcqnts2f-64.asm
new file mode 100644
index 0000000..d0efa1b
--- /dev/null
+++ b/simd/jcqnts2f-64.asm
@@ -0,0 +1,158 @@
+;
+; jcqnts2f-64.asm - sample data conversion and quantization (64-bit SSE & SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                            FAST_FLOAT * workspace);
+;
+
+; r10 = JSAMPARRAY sample_data
+; r11 = JDIMENSION start_col
+; r12 = FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_float_sse2)
+
+EXTN(jsimd_convsamp_float_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+	push	rbx
+
+	pcmpeqw  xmm7,xmm7
+	psllw    xmm7,7
+	packsswb xmm7,xmm7		; xmm7 = PB_CENTERJSAMPLE (0x808080..)
+
+	mov rsi, r10
+	mov	rax, r11
+	mov rdi, r12
+	mov	rcx, DCTSIZE/2
+.convloop:
+	mov	rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]
+	movq	xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]
+
+	psubb	xmm0,xmm7			; xmm0=(01234567)
+	psubb	xmm1,xmm7			; xmm1=(89ABCDEF)
+
+	punpcklbw xmm0,xmm0			; xmm0=(*0*1*2*3*4*5*6*7)
+	punpcklbw xmm1,xmm1			; xmm1=(*8*9*A*B*C*D*E*F)
+
+	punpcklwd xmm2,xmm0			; xmm2=(***0***1***2***3)
+	punpckhwd xmm0,xmm0			; xmm0=(***4***5***6***7)
+	punpcklwd xmm3,xmm1			; xmm3=(***8***9***A***B)
+	punpckhwd xmm1,xmm1			; xmm1=(***C***D***E***F)
+
+	psrad     xmm2,(DWORD_BIT-BYTE_BIT)	; xmm2=(0123)
+	psrad     xmm0,(DWORD_BIT-BYTE_BIT)	; xmm0=(4567)
+	cvtdq2ps  xmm2,xmm2			; xmm2=(0123)
+	cvtdq2ps  xmm0,xmm0			; xmm0=(4567)
+	psrad     xmm3,(DWORD_BIT-BYTE_BIT)	; xmm3=(89AB)
+	psrad     xmm1,(DWORD_BIT-BYTE_BIT)	; xmm1=(CDEF)
+	cvtdq2ps  xmm3,xmm3			; xmm3=(89AB)
+	cvtdq2ps  xmm1,xmm1			; xmm1=(CDEF)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
+
+	add	rsi, byte 2*SIZEOF_JSAMPROW
+	add	rdi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	rcx
+	jnz	short .convloop
+
+	pop	rbx
+	uncollect_args
+	pop	rbp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; GLOBAL(void)
+; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+;                         FAST_FLOAT * workspace);
+;
+
+; r10 = JCOEFPTR coef_block
+; r11 = FAST_FLOAT * divisors
+; r12 = FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_float_sse2)
+
+EXTN(jsimd_quantize_float_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+
+	mov rsi, r12
+	mov rdx, r11
+	mov rdi, r10
+	mov	rax, DCTSIZE2/16
+.quantloop:
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,rsi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,rsi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
+
+	cvtps2dq xmm0,xmm0
+	cvtps2dq xmm1,xmm1
+	cvtps2dq xmm2,xmm2
+	cvtps2dq xmm3,xmm3
+
+	packssdw xmm0,xmm1
+	packssdw xmm2,xmm3
+
+	movdqa	XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_JCOEF)], xmm0
+	movdqa	XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_JCOEF)], xmm2
+
+	add	rsi, byte 16*SIZEOF_FAST_FLOAT
+	add	rdx, byte 16*SIZEOF_FAST_FLOAT
+	add	rdi, byte 16*SIZEOF_JCOEF
+	dec	rax
+	jnz	short .quantloop
+
+	uncollect_args
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcqnts2f.asm b/simd/jcqnts2f.asm
new file mode 100644
index 0000000..d80ae5d
--- /dev/null
+++ b/simd/jcqnts2f.asm
@@ -0,0 +1,171 @@
+;
+; jcqnts2f.asm - sample data conversion and quantization (SSE & SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                            FAST_FLOAT * workspace);
+;
+
+%define sample_data	ebp+8		; JSAMPARRAY sample_data
+%define start_col	ebp+12		; JDIMENSION start_col
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_float_sse2)
+
+EXTN(jsimd_convsamp_float_sse2):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	pcmpeqw  xmm7,xmm7
+	psllw    xmm7,7
+	packsswb xmm7,xmm7		; xmm7 = PB_CENTERJSAMPLE (0x808080..)
+
+	mov	esi, JSAMPARRAY [sample_data]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [start_col]
+	mov	edi, POINTER [workspace]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/2
+	alignx	16,7
+.convloop:
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	xmm0, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE]
+	movq	xmm1, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE]
+
+	psubb	xmm0,xmm7			; xmm0=(01234567)
+	psubb	xmm1,xmm7			; xmm1=(89ABCDEF)
+
+	punpcklbw xmm0,xmm0			; xmm0=(*0*1*2*3*4*5*6*7)
+	punpcklbw xmm1,xmm1			; xmm1=(*8*9*A*B*C*D*E*F)
+
+	punpcklwd xmm2,xmm0			; xmm2=(***0***1***2***3)
+	punpckhwd xmm0,xmm0			; xmm0=(***4***5***6***7)
+	punpcklwd xmm3,xmm1			; xmm3=(***8***9***A***B)
+	punpckhwd xmm1,xmm1			; xmm1=(***C***D***E***F)
+
+	psrad     xmm2,(DWORD_BIT-BYTE_BIT)	; xmm2=(0123)
+	psrad     xmm0,(DWORD_BIT-BYTE_BIT)	; xmm0=(4567)
+	cvtdq2ps  xmm2,xmm2			; xmm2=(0123)
+	cvtdq2ps  xmm0,xmm0			; xmm0=(4567)
+	psrad     xmm3,(DWORD_BIT-BYTE_BIT)	; xmm3=(89AB)
+	psrad     xmm1,(DWORD_BIT-BYTE_BIT)	; xmm1=(CDEF)
+	cvtdq2ps  xmm3,xmm3			; xmm3=(89AB)
+	cvtdq2ps  xmm1,xmm1			; xmm1=(CDEF)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
+
+	add	esi, byte 2*SIZEOF_JSAMPROW
+	add	edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	short .convloop
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; GLOBAL(void)
+; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+;                         FAST_FLOAT * workspace);
+;
+
+%define coef_block	ebp+8		; JCOEFPTR coef_block
+%define divisors	ebp+12		; FAST_FLOAT * divisors
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_float_sse2)
+
+EXTN(jsimd_quantize_float_sse2):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	esi, POINTER [workspace]
+	mov	edx, POINTER [divisors]
+	mov	edi, JCOEFPTR [coef_block]
+	mov	eax, DCTSIZE2/16
+	alignx	16,7
+.quantloop:
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+
+	cvtps2dq xmm0,xmm0
+	cvtps2dq xmm1,xmm1
+	cvtps2dq xmm2,xmm2
+	cvtps2dq xmm3,xmm3
+
+	packssdw xmm0,xmm1
+	packssdw xmm2,xmm3
+
+	movdqa	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_JCOEF)], xmm0
+	movdqa	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_JCOEF)], xmm2
+
+	add	esi, byte 16*SIZEOF_FAST_FLOAT
+	add	edx, byte 16*SIZEOF_FAST_FLOAT
+	add	edi, byte 16*SIZEOF_JCOEF
+	dec	eax
+	jnz	short .quantloop
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcqnts2i-64.asm b/simd/jcqnts2i-64.asm
new file mode 100644
index 0000000..cc33d59
--- /dev/null
+++ b/simd/jcqnts2i-64.asm
@@ -0,0 +1,187 @@
+;
+; jcqnts2i-64.asm - sample data conversion and quantization (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                      DCTELEM * workspace);
+;
+
+; r10 = JSAMPARRAY sample_data
+; r11 = JDIMENSION start_col
+; r12 = DCTELEM * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_sse2)
+
+EXTN(jsimd_convsamp_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+	push	rbx
+
+	pxor	xmm6,xmm6		; xmm6=(all 0's)
+	pcmpeqw	xmm7,xmm7
+	psllw	xmm7,7			; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+	mov rsi, r10
+	mov rax, r11
+	mov rdi, r12
+	mov	rcx, DCTSIZE/4
+.convloop:
+	mov	rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]	; xmm0=(01234567)
+	movq	xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]	; xmm1=(89ABCDEF)
+
+	mov	rbx, JSAMPROW [rsi+2*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	rdx, JSAMPROW [rsi+3*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	xmm2, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]	; xmm2=(GHIJKLMN)
+	movq	xmm3, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]	; xmm3=(OPQRSTUV)
+
+	punpcklbw xmm0,xmm6		; xmm0=(01234567)
+	punpcklbw xmm1,xmm6		; xmm1=(89ABCDEF)
+	paddw     xmm0,xmm7
+	paddw     xmm1,xmm7
+	punpcklbw xmm2,xmm6		; xmm2=(GHIJKLMN)
+	punpcklbw xmm3,xmm6		; xmm3=(OPQRSTUV)
+	paddw     xmm2,xmm7
+	paddw     xmm3,xmm7
+
+	movdqa	XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
+	movdqa	XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
+	movdqa	XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
+	movdqa	XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
+
+	add	rsi, byte 4*SIZEOF_JSAMPROW
+	add	rdi, byte 4*DCTSIZE*SIZEOF_DCTELEM
+	dec	rcx
+	jnz	short .convloop
+
+	pop	rbx
+	uncollect_args
+	pop	rbp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; This implementation is based on an algorithm described in
+;   "How to optimize for the Pentium family of microprocessors"
+;   (http://www.agner.org/assem/).
+;
+; GLOBAL(void)
+; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors,
+;                      DCTELEM * workspace);
+;
+
+%define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
+%define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
+%define SCALE(m,n,b)      XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
+
+; r10 = JCOEFPTR coef_block
+; r11 = DCTELEM * divisors
+; r12 = DCTELEM * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_sse2)
+
+EXTN(jsimd_quantize_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+
+	mov rsi, r12
+	mov rdx, r11
+	mov rdi, r10
+	mov	rax, DCTSIZE2/32
+.quantloop:
+	movdqa	xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_DCTELEM)]
+	movdqa	xmm5, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_DCTELEM)]
+	movdqa	xmm6, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_DCTELEM)]
+	movdqa	xmm7, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_DCTELEM)]
+	movdqa	xmm0,xmm4
+	movdqa	xmm1,xmm5
+	movdqa	xmm2,xmm6
+	movdqa	xmm3,xmm7
+	psraw	xmm4,(WORD_BIT-1)
+	psraw	xmm5,(WORD_BIT-1)
+	psraw	xmm6,(WORD_BIT-1)
+	psraw	xmm7,(WORD_BIT-1)
+	pxor	xmm0,xmm4
+	pxor	xmm1,xmm5
+	pxor	xmm2,xmm6
+	pxor	xmm3,xmm7
+	psubw	xmm0,xmm4		; if (xmm0 < 0) xmm0 = -xmm0;
+	psubw	xmm1,xmm5		; if (xmm1 < 0) xmm1 = -xmm1;
+	psubw	xmm2,xmm6		; if (xmm2 < 0) xmm2 = -xmm2;
+	psubw	xmm3,xmm7		; if (xmm3 < 0) xmm3 = -xmm3;
+
+	paddw	xmm0, XMMWORD [CORRECTION(0,0,rdx)]  ; correction + roundfactor
+	paddw	xmm1, XMMWORD [CORRECTION(1,0,rdx)]
+	paddw	xmm2, XMMWORD [CORRECTION(2,0,rdx)]
+	paddw	xmm3, XMMWORD [CORRECTION(3,0,rdx)]
+	pmulhuw	xmm0, XMMWORD [RECIPROCAL(0,0,rdx)]  ; reciprocal
+	pmulhuw	xmm1, XMMWORD [RECIPROCAL(1,0,rdx)]
+	pmulhuw	xmm2, XMMWORD [RECIPROCAL(2,0,rdx)]
+	pmulhuw	xmm3, XMMWORD [RECIPROCAL(3,0,rdx)]
+	pmulhuw	xmm0, XMMWORD [SCALE(0,0,rdx)]	; scale
+	pmulhuw	xmm1, XMMWORD [SCALE(1,0,rdx)]
+	pmulhuw	xmm2, XMMWORD [SCALE(2,0,rdx)]
+	pmulhuw	xmm3, XMMWORD [SCALE(3,0,rdx)]
+
+	pxor	xmm0,xmm4
+	pxor	xmm1,xmm5
+	pxor	xmm2,xmm6
+	pxor	xmm3,xmm7
+	psubw	xmm0,xmm4
+	psubw	xmm1,xmm5
+	psubw	xmm2,xmm6
+	psubw	xmm3,xmm7
+	movdqa	XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
+	movdqa	XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
+	movdqa	XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
+	movdqa	XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
+
+	add	rsi, byte 32*SIZEOF_DCTELEM
+	add	rdx, byte 32*SIZEOF_DCTELEM
+	add	rdi, byte 32*SIZEOF_JCOEF
+	dec	rax
+	jnz	near .quantloop
+
+	uncollect_args
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcqnts2i.asm b/simd/jcqnts2i.asm
new file mode 100644
index 0000000..0864d6e
--- /dev/null
+++ b/simd/jcqnts2i.asm
@@ -0,0 +1,200 @@
+;
+; jcqnts2i.asm - sample data conversion and quantization (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                      DCTELEM * workspace);
+;
+
+%define sample_data	ebp+8		; JSAMPARRAY sample_data
+%define start_col	ebp+12		; JDIMENSION start_col
+%define workspace	ebp+16		; DCTELEM * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_sse2)
+
+EXTN(jsimd_convsamp_sse2):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	pxor	xmm6,xmm6		; xmm6=(all 0's)
+	pcmpeqw	xmm7,xmm7
+	psllw	xmm7,7			; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+	mov	esi, JSAMPARRAY [sample_data]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [start_col]
+	mov	edi, POINTER [workspace]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.convloop:
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	xmm0, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE]	; xmm0=(01234567)
+	movq	xmm1, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE]	; xmm1=(89ABCDEF)
+
+	mov	ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	xmm2, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE]	; xmm2=(GHIJKLMN)
+	movq	xmm3, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE]	; xmm3=(OPQRSTUV)
+
+	punpcklbw xmm0,xmm6		; xmm0=(01234567)
+	punpcklbw xmm1,xmm6		; xmm1=(89ABCDEF)
+	paddw     xmm0,xmm7
+	paddw     xmm1,xmm7
+	punpcklbw xmm2,xmm6		; xmm2=(GHIJKLMN)
+	punpcklbw xmm3,xmm6		; xmm3=(OPQRSTUV)
+	paddw     xmm2,xmm7
+	paddw     xmm3,xmm7
+
+	movdqa	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_DCTELEM)], xmm0
+	movdqa	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_DCTELEM)], xmm1
+	movdqa	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_DCTELEM)], xmm2
+	movdqa	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_DCTELEM)], xmm3
+
+	add	esi, byte 4*SIZEOF_JSAMPROW
+	add	edi, byte 4*DCTSIZE*SIZEOF_DCTELEM
+	dec	ecx
+	jnz	short .convloop
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; This implementation is based on an algorithm described in
+;   "How to optimize for the Pentium family of microprocessors"
+;   (http://www.agner.org/assem/).
+;
+; GLOBAL(void)
+; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors,
+;                      DCTELEM * workspace);
+;
+
+%define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
+%define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
+%define SCALE(m,n,b)      XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
+
+%define coef_block	ebp+8		; JCOEFPTR coef_block
+%define divisors	ebp+12		; DCTELEM * divisors
+%define workspace	ebp+16		; DCTELEM * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_sse2)
+
+EXTN(jsimd_quantize_sse2):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	esi, POINTER [workspace]
+	mov	edx, POINTER [divisors]
+	mov	edi, JCOEFPTR [coef_block]
+	mov	eax, DCTSIZE2/32
+	alignx	16,7
+.quantloop:
+	movdqa	xmm4, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_DCTELEM)]
+	movdqa	xmm5, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_DCTELEM)]
+	movdqa	xmm6, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_DCTELEM)]
+	movdqa	xmm7, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_DCTELEM)]
+	movdqa	xmm0,xmm4
+	movdqa	xmm1,xmm5
+	movdqa	xmm2,xmm6
+	movdqa	xmm3,xmm7
+	psraw	xmm4,(WORD_BIT-1)
+	psraw	xmm5,(WORD_BIT-1)
+	psraw	xmm6,(WORD_BIT-1)
+	psraw	xmm7,(WORD_BIT-1)
+	pxor	xmm0,xmm4
+	pxor	xmm1,xmm5
+	pxor	xmm2,xmm6
+	pxor	xmm3,xmm7
+	psubw	xmm0,xmm4		; if (xmm0 < 0) xmm0 = -xmm0;
+	psubw	xmm1,xmm5		; if (xmm1 < 0) xmm1 = -xmm1;
+	psubw	xmm2,xmm6		; if (xmm2 < 0) xmm2 = -xmm2;
+	psubw	xmm3,xmm7		; if (xmm3 < 0) xmm3 = -xmm3;
+
+	paddw	xmm0, XMMWORD [CORRECTION(0,0,edx)]  ; correction + roundfactor
+	paddw	xmm1, XMMWORD [CORRECTION(1,0,edx)]
+	paddw	xmm2, XMMWORD [CORRECTION(2,0,edx)]
+	paddw	xmm3, XMMWORD [CORRECTION(3,0,edx)]
+	pmulhuw	xmm0, XMMWORD [RECIPROCAL(0,0,edx)]  ; reciprocal
+	pmulhuw	xmm1, XMMWORD [RECIPROCAL(1,0,edx)]
+	pmulhuw	xmm2, XMMWORD [RECIPROCAL(2,0,edx)]
+	pmulhuw	xmm3, XMMWORD [RECIPROCAL(3,0,edx)]
+	pmulhuw	xmm0, XMMWORD [SCALE(0,0,edx)]	; scale
+	pmulhuw	xmm1, XMMWORD [SCALE(1,0,edx)]
+	pmulhuw	xmm2, XMMWORD [SCALE(2,0,edx)]
+	pmulhuw	xmm3, XMMWORD [SCALE(3,0,edx)]
+
+	pxor	xmm0,xmm4
+	pxor	xmm1,xmm5
+	pxor	xmm2,xmm6
+	pxor	xmm3,xmm7
+	psubw	xmm0,xmm4
+	psubw	xmm1,xmm5
+	psubw	xmm2,xmm6
+	psubw	xmm3,xmm7
+	movdqa	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_DCTELEM)], xmm0
+	movdqa	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_DCTELEM)], xmm1
+	movdqa	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_DCTELEM)], xmm2
+	movdqa	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_DCTELEM)], xmm3
+
+	add	esi, byte 32*SIZEOF_DCTELEM
+	add	edx, byte 32*SIZEOF_DCTELEM
+	add	edi, byte 32*SIZEOF_JCOEF
+	dec	eax
+	jnz	near .quantloop
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcqntsse.asm b/simd/jcqntsse.asm
new file mode 100644
index 0000000..3065eca
--- /dev/null
+++ b/simd/jcqntsse.asm
@@ -0,0 +1,211 @@
+;
+; jcqntsse.asm - sample data conversion and quantization (SSE & MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_float_sse (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                           FAST_FLOAT * workspace);
+;
+
+%define sample_data	ebp+8		; JSAMPARRAY sample_data
+%define start_col	ebp+12		; JDIMENSION start_col
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_float_sse)
+
+EXTN(jsimd_convsamp_float_sse):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	pcmpeqw  mm7,mm7
+	psllw    mm7,7
+	packsswb mm7,mm7		; mm7 = PB_CENTERJSAMPLE (0x808080..)
+
+	mov	esi, JSAMPARRAY [sample_data]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [start_col]
+	mov	edi, POINTER [workspace]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/2
+	alignx	16,7
+.convloop:
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE]
+	movq	mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE]
+
+	psubb	mm0,mm7				; mm0=(01234567)
+	psubb	mm1,mm7				; mm1=(89ABCDEF)
+
+	punpcklbw mm2,mm0			; mm2=(*0*1*2*3)
+	punpckhbw mm0,mm0			; mm0=(*4*5*6*7)
+	punpcklbw mm3,mm1			; mm3=(*8*9*A*B)
+	punpckhbw mm1,mm1			; mm1=(*C*D*E*F)
+
+	punpcklwd mm4,mm2			; mm4=(***0***1)
+	punpckhwd mm2,mm2			; mm2=(***2***3)
+	punpcklwd mm5,mm0			; mm5=(***4***5)
+	punpckhwd mm0,mm0			; mm0=(***6***7)
+
+	psrad     mm4,(DWORD_BIT-BYTE_BIT)	; mm4=(01)
+	psrad     mm2,(DWORD_BIT-BYTE_BIT)	; mm2=(23)
+	cvtpi2ps  xmm0,mm4			; xmm0=(01**)
+	cvtpi2ps  xmm1,mm2			; xmm1=(23**)
+	psrad     mm5,(DWORD_BIT-BYTE_BIT)	; mm5=(45)
+	psrad     mm0,(DWORD_BIT-BYTE_BIT)	; mm0=(67)
+	cvtpi2ps  xmm2,mm5			; xmm2=(45**)
+	cvtpi2ps  xmm3,mm0			; xmm3=(67**)
+
+	punpcklwd mm6,mm3			; mm6=(***8***9)
+	punpckhwd mm3,mm3			; mm3=(***A***B)
+	punpcklwd mm4,mm1			; mm4=(***C***D)
+	punpckhwd mm1,mm1			; mm1=(***E***F)
+
+	psrad     mm6,(DWORD_BIT-BYTE_BIT)	; mm6=(89)
+	psrad     mm3,(DWORD_BIT-BYTE_BIT)	; mm3=(AB)
+	cvtpi2ps  xmm4,mm6			; xmm4=(89**)
+	cvtpi2ps  xmm5,mm3			; xmm5=(AB**)
+	psrad     mm4,(DWORD_BIT-BYTE_BIT)	; mm4=(CD)
+	psrad     mm1,(DWORD_BIT-BYTE_BIT)	; mm1=(EF)
+	cvtpi2ps  xmm6,mm4			; xmm6=(CD**)
+	cvtpi2ps  xmm7,mm1			; xmm7=(EF**)
+
+	movlhps   xmm0,xmm1			; xmm0=(0123)
+	movlhps   xmm2,xmm3			; xmm2=(4567)
+	movlhps   xmm4,xmm5			; xmm4=(89AB)
+	movlhps   xmm6,xmm7			; xmm6=(CDEF)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm4
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
+
+	add	esi, byte 2*SIZEOF_JSAMPROW
+	add	edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .convloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; GLOBAL(void)
+; jsimd_quantize_float_sse (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+;                           FAST_FLOAT * workspace);
+;
+
+%define coef_block	ebp+8		; JCOEFPTR coef_block
+%define divisors	ebp+12		; FAST_FLOAT * divisors
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_float_sse)
+
+EXTN(jsimd_quantize_float_sse):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	esi, POINTER [workspace]
+	mov	edx, POINTER [divisors]
+	mov	edi, JCOEFPTR [coef_block]
+	mov	eax, DCTSIZE2/16
+	alignx	16,7
+.quantloop:
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+
+	movhlps  xmm4,xmm0
+	movhlps  xmm5,xmm1
+
+	cvtps2pi mm0,xmm0
+	cvtps2pi mm1,xmm1
+	cvtps2pi mm4,xmm4
+	cvtps2pi mm5,xmm5
+
+	movhlps  xmm6,xmm2
+	movhlps  xmm7,xmm3
+
+	cvtps2pi mm2,xmm2
+	cvtps2pi mm3,xmm3
+	cvtps2pi mm6,xmm6
+	cvtps2pi mm7,xmm7
+
+	packssdw mm0,mm4
+	packssdw mm1,mm5
+	packssdw mm2,mm6
+	packssdw mm3,mm7
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3
+
+	add	esi, byte 16*SIZEOF_FAST_FLOAT
+	add	edx, byte 16*SIZEOF_FAST_FLOAT
+	add	edi, byte 16*SIZEOF_JCOEF
+	dec	eax
+	jnz	short .quantloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcsammmx.asm b/simd/jcsammmx.asm
new file mode 100644
index 0000000..9e43b2f
--- /dev/null
+++ b/simd/jcsammmx.asm
@@ -0,0 +1,324 @@
+;
+; jcsammmx.asm - downsampling (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_mmx (JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+;                            JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+%define img_width(b)	(b)+8			; JDIMENSION image_width
+%define max_v_samp(b)	(b)+12		; int max_v_samp_factor
+%define v_samp(b)			(b)+16		; JDIMENSION v_samp_factor
+%define width_blks(b)	(b)+20		; JDIMENSION width_blocks
+%define input_data(b)	(b)+24		; JSAMPARRAY input_data
+%define output_data(b)	(b)+28	; JSAMPARRAY output_data
+
+	align	16
+	global	EXTN(jsimd_h2v1_downsample_mmx)
+
+EXTN(jsimd_h2v1_downsample_mmx):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	ecx, JDIMENSION [width_blks(ebp)]
+	shl	ecx,3			; imul ecx,DCTSIZE (ecx = output_cols)
+	jz	near .return
+
+	mov	edx, JDIMENSION [img_width(ebp)]
+
+	; -- expand_right_edge
+
+	push	ecx
+	shl	ecx,1				; output_cols * 2
+	sub	ecx,edx
+	jle	short .expand_end
+
+	mov	eax, INT [max_v_samp(ebp)]
+	test	eax,eax
+	jle	short .expand_end
+
+	cld
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	alignx	16,7
+.expandloop:
+	push	eax
+	push	ecx
+
+	mov	edi, JSAMPROW [esi]
+	add	edi,edx
+	mov	al, JSAMPLE [edi-1]
+
+	rep stosb
+
+	pop	ecx
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW
+	dec	eax
+	jg	short .expandloop
+
+.expand_end:
+	pop	ecx				; output_cols
+
+	; -- h2v1_downsample
+
+	mov	eax, JDIMENSION [v_samp(ebp)]	; rowctr
+	test	eax,eax
+	jle	near .return
+
+	mov       edx, 0x00010000	; bias pattern
+	movd      mm7,edx
+	pcmpeqw   mm6,mm6
+	punpckldq mm7,mm7		; mm7={0, 1, 0, 1}
+	psrlw     mm6,BYTE_BIT		; mm6={0xFF 0x00 0xFF 0x00 ..}
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, JSAMPARRAY [output_data(ebp)]	; output_data
+	alignx	16,7
+.rowloop:
+	push	ecx
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]		; inptr
+	mov	edi, JSAMPROW [edi]		; outptr
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mm1, MMWORD [esi+1*SIZEOF_MMWORD]
+	movq	mm2,mm0
+	movq	mm3,mm1
+
+	pand	mm0,mm6
+	psrlw	mm2,BYTE_BIT
+	pand	mm1,mm6
+	psrlw	mm3,BYTE_BIT
+
+	paddw	mm0,mm2
+	paddw	mm1,mm3
+	paddw	mm0,mm7
+	paddw	mm1,mm7
+	psrlw	mm0,1
+	psrlw	mm1,1
+
+	packuswb mm0,mm1
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm0
+
+	add	esi, byte 2*SIZEOF_MMWORD	; inptr
+	add	edi, byte 1*SIZEOF_MMWORD	; outptr
+	sub	ecx, byte SIZEOF_MMWORD		; outcol
+	jnz	short .columnloop
+
+	pop	esi
+	pop	edi
+	pop	ecx
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_data
+	add	edi, byte SIZEOF_JSAMPROW	; output_data
+	dec	eax				; rowctr
+	jg	short .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_mmx (JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+;                            JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+%define img_width(b)	(b)+8			; JDIMENSION image_width
+%define max_v_samp(b)	(b)+12		; int max_v_samp_factor
+%define v_samp(b)			(b)+16		; JDIMENSION v_samp_factor
+%define width_blks(b)	(b)+20		; JDIMENSION width_blocks
+%define input_data(b)	(b)+24		; JSAMPARRAY input_data
+%define output_data(b)	(b)+28	; JSAMPARRAY output_data
+
+	align	16
+	global	EXTN(jsimd_h2v2_downsample_mmx)
+
+EXTN(jsimd_h2v2_downsample_mmx):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	ecx, JDIMENSION [width_blks(ebp)]
+	shl	ecx,3			; imul ecx,DCTSIZE (ecx = output_cols)
+	jz	near .return
+
+	mov	edx, JDIMENSION [img_width(ebp)]
+
+	; -- expand_right_edge
+
+	push	ecx
+	shl	ecx,1				; output_cols * 2
+	sub	ecx,edx
+	jle	short .expand_end
+
+	mov	eax, INT [max_v_samp(ebp)]
+	test	eax,eax
+	jle	short .expand_end
+
+	cld
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	alignx	16,7
+.expandloop:
+	push	eax
+	push	ecx
+
+	mov	edi, JSAMPROW [esi]
+	add	edi,edx
+	mov	al, JSAMPLE [edi-1]
+
+	rep stosb
+
+	pop	ecx
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW
+	dec	eax
+	jg	short .expandloop
+
+.expand_end:
+	pop	ecx				; output_cols
+
+	; -- h2v2_downsample
+
+	mov	eax, JDIMENSION [v_samp(ebp)]	; rowctr
+	test	eax,eax
+	jle	near .return
+
+	mov       edx, 0x00020001	; bias pattern
+	movd      mm7,edx
+	pcmpeqw   mm6,mm6
+	punpckldq mm7,mm7		; mm7={1, 2, 1, 2}
+	psrlw     mm6,BYTE_BIT		; mm6={0xFF 0x00 0xFF 0x00 ..}
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, JSAMPARRAY [output_data(ebp)]	; output_data
+	alignx	16,7
+.rowloop:
+	push	ecx
+	push	edi
+	push	esi
+
+	mov	edx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; inptr0
+	mov	esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; inptr1
+	mov	edi, JSAMPROW [edi]			; outptr
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [edx+0*SIZEOF_MMWORD]
+	movq	mm1, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mm2, MMWORD [edx+1*SIZEOF_MMWORD]
+	movq	mm3, MMWORD [esi+1*SIZEOF_MMWORD]
+
+	movq	mm4,mm0
+	movq	mm5,mm1
+	pand	mm0,mm6
+	psrlw	mm4,BYTE_BIT
+	pand	mm1,mm6
+	psrlw	mm5,BYTE_BIT
+	paddw	mm0,mm4
+	paddw	mm1,mm5
+
+	movq	mm4,mm2
+	movq	mm5,mm3
+	pand	mm2,mm6
+	psrlw	mm4,BYTE_BIT
+	pand	mm3,mm6
+	psrlw	mm5,BYTE_BIT
+	paddw	mm2,mm4
+	paddw	mm3,mm5
+
+	paddw	mm0,mm1
+	paddw	mm2,mm3
+	paddw	mm0,mm7
+	paddw	mm2,mm7
+	psrlw	mm0,2
+	psrlw	mm2,2
+
+	packuswb mm0,mm2
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm0
+
+	add	edx, byte 2*SIZEOF_MMWORD	; inptr0
+	add	esi, byte 2*SIZEOF_MMWORD	; inptr1
+	add	edi, byte 1*SIZEOF_MMWORD	; outptr
+	sub	ecx, byte SIZEOF_MMWORD		; outcol
+	jnz	near .columnloop
+
+	pop	esi
+	pop	edi
+	pop	ecx
+
+	add	esi, byte 2*SIZEOF_JSAMPROW	; input_data
+	add	edi, byte 1*SIZEOF_JSAMPROW	; output_data
+	dec	eax				; rowctr
+	jg	near .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcsamss2-64.asm b/simd/jcsamss2-64.asm
new file mode 100644
index 0000000..6a16dc5
--- /dev/null
+++ b/simd/jcsamss2-64.asm
@@ -0,0 +1,330 @@
+;
+; jcsamss2-64.asm - downsampling (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor,
+;                             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+;                             JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+; r10 = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12 = JDIMENSION v_samp_factor
+; r13 = JDIMENSION width_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+	align	16
+	global	EXTN(jsimd_h2v1_downsample_sse2)
+
+EXTN(jsimd_h2v1_downsample_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+
+	mov rcx, r13
+	shl	rcx,3			; imul rcx,DCTSIZE (rcx = output_cols)
+	jz	near .return
+
+	mov rdx, r10
+
+	; -- expand_right_edge
+
+	push	rcx
+	shl	rcx,1				; output_cols * 2
+	sub	rcx,rdx
+	jle	short .expand_end
+
+	mov	rax, r11
+	test	rax,rax
+	jle	short .expand_end
+
+	cld
+	mov	rsi, r14	; input_data
+.expandloop:
+	push	rax
+	push	rcx
+
+	mov	rdi, JSAMPROW [rsi]
+	add	rdi,rdx
+	mov	al, JSAMPLE [rdi-1]
+
+	rep stosb
+
+	pop	rcx
+	pop	rax
+
+	add	rsi, byte SIZEOF_JSAMPROW
+	dec	rax
+	jg	short .expandloop
+
+.expand_end:
+	pop	rcx				; output_cols
+
+	; -- h2v1_downsample
+
+	mov	rax, r12	; rowctr
+	test	eax,eax
+	jle	near .return
+
+	mov	rdx, 0x00010000		; bias pattern
+	movd	xmm7,edx
+	pcmpeqw	xmm6,xmm6
+	pshufd	xmm7,xmm7,0x00		; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+	psrlw	xmm6,BYTE_BIT		; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+	mov	rsi, r14	; input_data
+	mov	rdi, r15	; output_data
+.rowloop:
+	push	rcx
+	push	rdi
+	push	rsi
+
+	mov	rsi, JSAMPROW [rsi]		; inptr
+	mov rdi, JSAMPROW [rdi]		; outptr
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jae	short .columnloop
+
+.columnloop_r8:
+	movdqa	xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	pxor	xmm1,xmm1
+	mov	rcx, SIZEOF_XMMWORD
+	jmp	short .downsample
+
+.columnloop:
+	movdqa	xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqa	xmm1, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+.downsample:
+	movdqa	xmm2,xmm0
+	movdqa	xmm3,xmm1
+
+	pand	xmm0,xmm6
+	psrlw	xmm2,BYTE_BIT
+	pand	xmm1,xmm6
+	psrlw	xmm3,BYTE_BIT
+
+	paddw	xmm0,xmm2
+	paddw	xmm1,xmm3
+	paddw	xmm0,xmm7
+	paddw	xmm1,xmm7
+	psrlw	xmm0,1
+	psrlw	xmm1,1
+
+	packuswb xmm0,xmm1
+
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+
+	sub	rcx, byte SIZEOF_XMMWORD	; outcol
+	add	rsi, byte 2*SIZEOF_XMMWORD	; inptr
+	add	rdi, byte 1*SIZEOF_XMMWORD	; outptr
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jae	short .columnloop
+	test	rcx,rcx
+	jnz	short .columnloop_r8
+
+	pop	rsi
+	pop	rdi
+	pop	rcx
+
+	add	rsi, byte SIZEOF_JSAMPROW	; input_data
+	add	rdi, byte SIZEOF_JSAMPROW	; output_data
+	dec	rax				; rowctr
+	jg	near .rowloop
+
+.return:
+	uncollect_args
+	pop	rbp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor,
+;                             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+;                             JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+; r10 = JDIMENSION image_width
+; r11 = int max_v_samp_factor
+; r12 = JDIMENSION v_samp_factor
+; r13 = JDIMENSION width_blocks
+; r14 = JSAMPARRAY input_data
+; r15 = JSAMPARRAY output_data
+
+	align	16
+	global	EXTN(jsimd_h2v2_downsample_sse2)
+
+EXTN(jsimd_h2v2_downsample_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+
+	mov	rcx, r13
+	shl	rcx,3			; imul rcx,DCTSIZE (rcx = output_cols)
+	jz	near .return
+
+	mov	rdx, r10
+
+	; -- expand_right_edge
+
+	push	rcx
+	shl	rcx,1				; output_cols * 2
+	sub	rcx,rdx
+	jle	short .expand_end
+
+	mov	rax, r11
+	test	rax,rax
+	jle	short .expand_end
+
+	cld
+	mov	rsi, r14	; input_data
+.expandloop:
+	push	rax
+	push	rcx
+
+	mov	rdi, JSAMPROW [rsi]
+	add	rdi,rdx
+	mov	al, JSAMPLE [rdi-1]
+
+	rep stosb
+
+	pop	rcx
+	pop	rax
+
+	add	rsi, byte SIZEOF_JSAMPROW
+	dec	rax
+	jg	short .expandloop
+
+.expand_end:
+	pop	rcx				; output_cols
+
+	; -- h2v2_downsample
+
+	mov	rax, r12	; rowctr
+	test	rax,rax
+	jle	near .return
+
+	mov	rdx, 0x00020001		; bias pattern
+	movd	xmm7,edx
+	pcmpeqw	xmm6,xmm6
+	pshufd	xmm7,xmm7,0x00		; xmm7={1, 2, 1, 2, 1, 2, 1, 2}
+	psrlw	xmm6,BYTE_BIT		; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+	mov	rsi, r14	; input_data
+	mov	rdi, r15	; output_data
+.rowloop:
+	push	rcx
+	push	rdi
+	push	rsi
+
+	mov	rdx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]	; inptr0
+	mov	rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]	; inptr1
+	mov	rdi, JSAMPROW [rdi]			; outptr
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jae	short .columnloop
+
+.columnloop_r8:
+	movdqa	xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+	movdqa	xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	pxor	xmm2,xmm2
+	pxor	xmm3,xmm3
+	mov	rcx, SIZEOF_XMMWORD
+	jmp	short .downsample
+
+.columnloop:
+	movdqa	xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+	movdqa	xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqa	xmm2, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+	movdqa	xmm3, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+.downsample:
+	movdqa	xmm4,xmm0
+	movdqa	xmm5,xmm1
+	pand	xmm0,xmm6
+	psrlw	xmm4,BYTE_BIT
+	pand	xmm1,xmm6
+	psrlw	xmm5,BYTE_BIT
+	paddw	xmm0,xmm4
+	paddw	xmm1,xmm5
+
+	movdqa	xmm4,xmm2
+	movdqa	xmm5,xmm3
+	pand	xmm2,xmm6
+	psrlw	xmm4,BYTE_BIT
+	pand	xmm3,xmm6
+	psrlw	xmm5,BYTE_BIT
+	paddw	xmm2,xmm4
+	paddw	xmm3,xmm5
+
+	paddw	xmm0,xmm1
+	paddw	xmm2,xmm3
+	paddw	xmm0,xmm7
+	paddw	xmm2,xmm7
+	psrlw	xmm0,2
+	psrlw	xmm2,2
+
+	packuswb xmm0,xmm2
+
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+
+	sub	rcx, byte SIZEOF_XMMWORD	; outcol
+	add	rdx, byte 2*SIZEOF_XMMWORD	; inptr0
+	add	rsi, byte 2*SIZEOF_XMMWORD	; inptr1
+	add	rdi, byte 1*SIZEOF_XMMWORD	; outptr
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jae	near .columnloop
+	test	rcx,rcx
+	jnz	near .columnloop_r8
+
+	pop	rsi
+	pop	rdi
+	pop	rcx
+
+	add	rsi, byte 2*SIZEOF_JSAMPROW	; input_data
+	add	rdi, byte 1*SIZEOF_JSAMPROW	; output_data
+	dec	rax				; rowctr
+	jg	near .rowloop
+
+.return:
+	uncollect_args
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jcsamss2.asm b/simd/jcsamss2.asm
new file mode 100644
index 0000000..818e911
--- /dev/null
+++ b/simd/jcsamss2.asm
@@ -0,0 +1,351 @@
+;
+; jcsamss2.asm - downsampling (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor,
+;                             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+;                             JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+%define img_width(b)	(b)+8			; JDIMENSION image_width
+%define max_v_samp(b)	(b)+12		; int max_v_samp_factor
+%define v_samp(b)			(b)+16		; JDIMENSION v_samp_factor
+%define width_blks(b)	(b)+20		; JDIMENSION width_blocks
+%define input_data(b)	(b)+24		; JSAMPARRAY input_data
+%define output_data(b)	(b)+28		; JSAMPARRAY output_data
+
+	align	16
+	global	EXTN(jsimd_h2v1_downsample_sse2)
+
+EXTN(jsimd_h2v1_downsample_sse2):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	ecx, JDIMENSION [width_blks(ebp)]
+	shl	ecx,3			; imul ecx,DCTSIZE (ecx = output_cols)
+	jz	near .return
+
+	mov	edx, JDIMENSION [img_width(ebp)]
+
+	; -- expand_right_edge
+
+	push	ecx
+	shl	ecx,1				; output_cols * 2
+	sub	ecx,edx
+	jle	short .expand_end
+
+	mov	eax, INT [max_v_samp(ebp)]
+	test	eax,eax
+	jle	short .expand_end
+
+	cld
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	alignx	16,7
+.expandloop:
+	push	eax
+	push	ecx
+
+	mov	edi, JSAMPROW [esi]
+	add	edi,edx
+	mov	al, JSAMPLE [edi-1]
+
+	rep stosb
+
+	pop	ecx
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW
+	dec	eax
+	jg	short .expandloop
+
+.expand_end:
+	pop	ecx				; output_cols
+
+	; -- h2v1_downsample
+
+	mov	eax, JDIMENSION [v_samp(ebp)]	; rowctr
+	test	eax,eax
+	jle	near .return
+
+	mov	edx, 0x00010000		; bias pattern
+	movd	xmm7,edx
+	pcmpeqw	xmm6,xmm6
+	pshufd	xmm7,xmm7,0x00		; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+	psrlw	xmm6,BYTE_BIT		; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, JSAMPARRAY [output_data(ebp)]	; output_data
+	alignx	16,7
+.rowloop:
+	push	ecx
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]		; inptr
+	mov	edi, JSAMPROW [edi]		; outptr
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jae	short .columnloop
+	alignx	16,7
+
+.columnloop_r8:
+	movdqa	xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	pxor	xmm1,xmm1
+	mov	ecx, SIZEOF_XMMWORD
+	jmp	short .downsample
+	alignx	16,7
+
+.columnloop:
+	movdqa	xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqa	xmm1, XMMWORD [esi+1*SIZEOF_XMMWORD]
+
+.downsample:
+	movdqa	xmm2,xmm0
+	movdqa	xmm3,xmm1
+
+	pand	xmm0,xmm6
+	psrlw	xmm2,BYTE_BIT
+	pand	xmm1,xmm6
+	psrlw	xmm3,BYTE_BIT
+
+	paddw	xmm0,xmm2
+	paddw	xmm1,xmm3
+	paddw	xmm0,xmm7
+	paddw	xmm1,xmm7
+	psrlw	xmm0,1
+	psrlw	xmm1,1
+
+	packuswb xmm0,xmm1
+
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
+
+	sub	ecx, byte SIZEOF_XMMWORD	; outcol
+	add	esi, byte 2*SIZEOF_XMMWORD	; inptr
+	add	edi, byte 1*SIZEOF_XMMWORD	; outptr
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jae	short .columnloop
+	test	ecx,ecx
+	jnz	short .columnloop_r8
+
+	pop	esi
+	pop	edi
+	pop	ecx
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_data
+	add	edi, byte SIZEOF_JSAMPROW	; output_data
+	dec	eax				; rowctr
+	jg	near .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor,
+;                             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+;                             JSAMPARRAY input_data, JSAMPARRAY output_data);
+;
+
+%define img_width(b)	(b)+8			; JDIMENSION image_width
+%define max_v_samp(b)	(b)+12		; int max_v_samp_factor
+%define v_samp(b)			(b)+16		; JDIMENSION v_samp_factor
+%define width_blks(b)	(b)+20		; JDIMENSION width_blocks
+%define input_data(b)	(b)+24		; JSAMPARRAY input_data
+%define output_data(b)	(b)+28	; JSAMPARRAY output_data
+
+	align	16
+	global	EXTN(jsimd_h2v2_downsample_sse2)
+
+EXTN(jsimd_h2v2_downsample_sse2):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	ecx, JDIMENSION [width_blks(ebp)]
+	shl	ecx,3			; imul ecx,DCTSIZE (ecx = output_cols)
+	jz	near .return
+
+	mov	edx, JDIMENSION [img_width(ebp)]
+
+	; -- expand_right_edge
+
+	push	ecx
+	shl	ecx,1				; output_cols * 2
+	sub	ecx,edx
+	jle	short .expand_end
+
+	mov	eax, INT [max_v_samp(ebp)]
+	test	eax,eax
+	jle	short .expand_end
+
+	cld
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	alignx	16,7
+.expandloop:
+	push	eax
+	push	ecx
+
+	mov	edi, JSAMPROW [esi]
+	add	edi,edx
+	mov	al, JSAMPLE [edi-1]
+
+	rep stosb
+
+	pop	ecx
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW
+	dec	eax
+	jg	short .expandloop
+
+.expand_end:
+	pop	ecx				; output_cols
+
+	; -- h2v2_downsample
+
+	mov	eax, JDIMENSION [v_samp(ebp)]	; rowctr
+	test	eax,eax
+	jle	near .return
+
+	mov	edx, 0x00020001		; bias pattern
+	movd	xmm7,edx
+	pcmpeqw	xmm6,xmm6
+	pshufd	xmm7,xmm7,0x00		; xmm7={1, 2, 1, 2, 1, 2, 1, 2}
+	psrlw	xmm6,BYTE_BIT		; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, JSAMPARRAY [output_data(ebp)]	; output_data
+	alignx	16,7
+.rowloop:
+	push	ecx
+	push	edi
+	push	esi
+
+	mov	edx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; inptr0
+	mov	esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; inptr1
+	mov	edi, JSAMPROW [edi]			; outptr
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jae	short .columnloop
+	alignx	16,7
+
+.columnloop_r8:
+	movdqa	xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD]
+	movdqa	xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	pxor	xmm2,xmm2
+	pxor	xmm3,xmm3
+	mov	ecx, SIZEOF_XMMWORD
+	jmp	short .downsample
+	alignx	16,7
+
+.columnloop:
+	movdqa	xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD]
+	movdqa	xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqa	xmm2, XMMWORD [edx+1*SIZEOF_XMMWORD]
+	movdqa	xmm3, XMMWORD [esi+1*SIZEOF_XMMWORD]
+
+.downsample:
+	movdqa	xmm4,xmm0
+	movdqa	xmm5,xmm1
+	pand	xmm0,xmm6
+	psrlw	xmm4,BYTE_BIT
+	pand	xmm1,xmm6
+	psrlw	xmm5,BYTE_BIT
+	paddw	xmm0,xmm4
+	paddw	xmm1,xmm5
+
+	movdqa	xmm4,xmm2
+	movdqa	xmm5,xmm3
+	pand	xmm2,xmm6
+	psrlw	xmm4,BYTE_BIT
+	pand	xmm3,xmm6
+	psrlw	xmm5,BYTE_BIT
+	paddw	xmm2,xmm4
+	paddw	xmm3,xmm5
+
+	paddw	xmm0,xmm1
+	paddw	xmm2,xmm3
+	paddw	xmm0,xmm7
+	paddw	xmm2,xmm7
+	psrlw	xmm0,2
+	psrlw	xmm2,2
+
+	packuswb xmm0,xmm2
+
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
+
+	sub	ecx, byte SIZEOF_XMMWORD	; outcol
+	add	edx, byte 2*SIZEOF_XMMWORD	; inptr0
+	add	esi, byte 2*SIZEOF_XMMWORD	; inptr1
+	add	edi, byte 1*SIZEOF_XMMWORD	; outptr
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jae	near .columnloop
+	test	ecx,ecx
+	jnz	near .columnloop_r8
+
+	pop	esi
+	pop	edi
+	pop	ecx
+
+	add	esi, byte 2*SIZEOF_JSAMPROW	; input_data
+	add	edi, byte 1*SIZEOF_JSAMPROW	; output_data
+	dec	eax				; rowctr
+	jg	near .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdclrmmx.asm b/simd/jdclrmmx.asm
new file mode 100644
index 0000000..79772e0
--- /dev/null
+++ b/simd/jdclrmmx.asm
@@ -0,0 +1,407 @@
+;
+; jdclrmmx.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_mmx (JDIMENSION out_width,
+;                            JSAMPIMAGE input_buf, JDIMENSION input_row,
+;                            JSAMPARRAY output_buf, int num_rows)
+;
+
+%define out_width(b)	(b)+8			; JDIMENSION out_width
+%define input_buf(b)	(b)+12		; JSAMPIMAGE input_buf
+%define input_row(b)	(b)+16		; JDIMENSION input_row
+%define output_buf(b)	(b)+20		; JSAMPARRAY output_buf
+%define num_rows(b)	(b)+24		; int num_rows
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_ycc_rgb_convert_mmx)
+
+EXTN(jsimd_ycc_rgb_convert_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	ecx, JDIMENSION [out_width(eax)]	; num_cols
+	test	ecx,ecx
+	jz	near .return
+
+	push	ecx
+
+	mov	edi, JSAMPIMAGE [input_buf(eax)]
+	mov	ecx, JDIMENSION [input_row(eax)]
+	mov	esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+	lea	esi, [esi+ecx*SIZEOF_JSAMPROW]
+	lea	ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+	lea	edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+	pop	ecx
+
+	mov	edi, JSAMPARRAY [output_buf(eax)]
+	mov	eax, INT [num_rows(eax)]
+	test	eax,eax
+	jle	near .return
+	alignx	16,7
+.rowloop:
+	push	eax
+	push	edi
+	push	edx
+	push	ebx
+	push	esi
+	push	ecx			; col
+
+	mov	esi, JSAMPROW [esi]	; inptr0
+	mov	ebx, JSAMPROW [ebx]	; inptr1
+	mov	edx, JSAMPROW [edx]	; inptr2
+	mov	edi, JSAMPROW [edi]	; outptr
+	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
+	alignx	16,7
+.columnloop:
+
+	movq	mm5, MMWORD [ebx]	; mm5=Cb(01234567)
+	movq	mm1, MMWORD [edx]	; mm1=Cr(01234567)
+
+	pcmpeqw	mm4,mm4
+	pcmpeqw	mm7,mm7
+	psrlw	mm4,BYTE_BIT
+	psllw	mm7,7			; mm7={0xFF80 0xFF80 0xFF80 0xFF80}
+	movq	mm0,mm4			; mm0=mm4={0xFF 0x00 0xFF 0x00 ..}
+
+	pand	mm4,mm5			; mm4=Cb(0246)=CbE
+	psrlw	mm5,BYTE_BIT		; mm5=Cb(1357)=CbO
+	pand	mm0,mm1			; mm0=Cr(0246)=CrE
+	psrlw	mm1,BYTE_BIT		; mm1=Cr(1357)=CrO
+
+	paddw	mm4,mm7
+	paddw	mm5,mm7
+	paddw	mm0,mm7
+	paddw	mm1,mm7
+
+	; (Original)
+	; R = Y                + 1.40200 * Cr
+	; G = Y - 0.34414 * Cb - 0.71414 * Cr
+	; B = Y + 1.77200 * Cb
+	;
+	; (This implementation)
+	; R = Y                + 0.40200 * Cr + Cr
+	; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+	; B = Y - 0.22800 * Cb + Cb + Cb
+
+	movq	mm2,mm4			; mm2=CbE
+	movq	mm3,mm5			; mm3=CbO
+	paddw	mm4,mm4			; mm4=2*CbE
+	paddw	mm5,mm5			; mm5=2*CbO
+	movq	mm6,mm0			; mm6=CrE
+	movq	mm7,mm1			; mm7=CrO
+	paddw	mm0,mm0			; mm0=2*CrE
+	paddw	mm1,mm1			; mm1=2*CrO
+
+	pmulhw	mm4,[GOTOFF(eax,PW_MF0228)]	; mm4=(2*CbE * -FIX(0.22800))
+	pmulhw	mm5,[GOTOFF(eax,PW_MF0228)]	; mm5=(2*CbO * -FIX(0.22800))
+	pmulhw	mm0,[GOTOFF(eax,PW_F0402)]	; mm0=(2*CrE * FIX(0.40200))
+	pmulhw	mm1,[GOTOFF(eax,PW_F0402)]	; mm1=(2*CrO * FIX(0.40200))
+
+	paddw	mm4,[GOTOFF(eax,PW_ONE)]
+	paddw	mm5,[GOTOFF(eax,PW_ONE)]
+	psraw	mm4,1			; mm4=(CbE * -FIX(0.22800))
+	psraw	mm5,1			; mm5=(CbO * -FIX(0.22800))
+	paddw	mm0,[GOTOFF(eax,PW_ONE)]
+	paddw	mm1,[GOTOFF(eax,PW_ONE)]
+	psraw	mm0,1			; mm0=(CrE * FIX(0.40200))
+	psraw	mm1,1			; mm1=(CrO * FIX(0.40200))
+
+	paddw	mm4,mm2
+	paddw	mm5,mm3
+	paddw	mm4,mm2			; mm4=(CbE * FIX(1.77200))=(B-Y)E
+	paddw	mm5,mm3			; mm5=(CbO * FIX(1.77200))=(B-Y)O
+	paddw	mm0,mm6			; mm0=(CrE * FIX(1.40200))=(R-Y)E
+	paddw	mm1,mm7			; mm1=(CrO * FIX(1.40200))=(R-Y)O
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=(B-Y)E
+	movq	MMWORD [wk(1)], mm5	; wk(1)=(B-Y)O
+
+	movq      mm4,mm2
+	movq      mm5,mm3
+	punpcklwd mm2,mm6
+	punpckhwd mm4,mm6
+	pmaddwd   mm2,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   mm4,[GOTOFF(eax,PW_MF0344_F0285)]
+	punpcklwd mm3,mm7
+	punpckhwd mm5,mm7
+	pmaddwd   mm3,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   mm5,[GOTOFF(eax,PW_MF0344_F0285)]
+
+	paddd     mm2,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     mm4,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     mm2,SCALEBITS
+	psrad     mm4,SCALEBITS
+	paddd     mm3,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     mm5,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     mm3,SCALEBITS
+	psrad     mm5,SCALEBITS
+
+	packssdw  mm2,mm4	; mm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+	packssdw  mm3,mm5	; mm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+	psubw     mm2,mm6	; mm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+	psubw     mm3,mm7	; mm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+	movq      mm5, MMWORD [esi]	; mm5=Y(01234567)
+
+	pcmpeqw   mm4,mm4
+	psrlw     mm4,BYTE_BIT		; mm4={0xFF 0x00 0xFF 0x00 ..}
+	pand      mm4,mm5		; mm4=Y(0246)=YE
+	psrlw     mm5,BYTE_BIT		; mm5=Y(1357)=YO
+
+	paddw     mm0,mm4		; mm0=((R-Y)E+YE)=RE=(R0 R2 R4 R6)
+	paddw     mm1,mm5		; mm1=((R-Y)O+YO)=RO=(R1 R3 R5 R7)
+	packuswb  mm0,mm0		; mm0=(R0 R2 R4 R6 ** ** ** **)
+	packuswb  mm1,mm1		; mm1=(R1 R3 R5 R7 ** ** ** **)
+
+	paddw     mm2,mm4		; mm2=((G-Y)E+YE)=GE=(G0 G2 G4 G6)
+	paddw     mm3,mm5		; mm3=((G-Y)O+YO)=GO=(G1 G3 G5 G7)
+	packuswb  mm2,mm2		; mm2=(G0 G2 G4 G6 ** ** ** **)
+	packuswb  mm3,mm3		; mm3=(G1 G3 G5 G7 ** ** ** **)
+
+	paddw     mm4, MMWORD [wk(0)]	; mm4=(YE+(B-Y)E)=BE=(B0 B2 B4 B6)
+	paddw     mm5, MMWORD [wk(1)]	; mm5=(YO+(B-Y)O)=BO=(B1 B3 B5 B7)
+	packuswb  mm4,mm4		; mm4=(B0 B2 B4 B6 ** ** ** **)
+	packuswb  mm5,mm5		; mm5=(B1 B3 B5 B7 ** ** ** **)
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+	; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
+	; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
+	; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
+	; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **)
+
+	punpcklbw mmA,mmC		; mmA=(00 10 02 12 04 14 06 16)
+	punpcklbw mmE,mmB		; mmE=(20 01 22 03 24 05 26 07)
+	punpcklbw mmD,mmF		; mmD=(11 21 13 23 15 25 17 27)
+
+	movq      mmG,mmA
+	movq      mmH,mmA
+	punpcklwd mmA,mmE		; mmA=(00 10 20 01 02 12 22 03)
+	punpckhwd mmG,mmE		; mmG=(04 14 24 05 06 16 26 07)
+
+	psrlq     mmH,2*BYTE_BIT	; mmH=(02 12 04 14 06 16 -- --)
+	psrlq     mmE,2*BYTE_BIT	; mmE=(22 03 24 05 26 07 -- --)
+
+	movq      mmC,mmD
+	movq      mmB,mmD
+	punpcklwd mmD,mmH		; mmD=(11 21 02 12 13 23 04 14)
+	punpckhwd mmC,mmH		; mmC=(15 25 06 16 17 27 -- --)
+
+	psrlq     mmB,2*BYTE_BIT	; mmB=(13 23 15 25 17 27 -- --)
+
+	movq      mmF,mmE
+	punpcklwd mmE,mmB		; mmE=(22 03 13 23 24 05 15 25)
+	punpckhwd mmF,mmB		; mmF=(26 07 17 27 -- -- -- --)
+
+	punpckldq mmA,mmD		; mmA=(00 10 20 01 11 21 02 12)
+	punpckldq mmE,mmG		; mmE=(22 03 13 23 04 14 24 05)
+	punpckldq mmC,mmF		; mmC=(15 25 06 16 26 07 17 27)
+
+	cmp	ecx, byte SIZEOF_MMWORD
+	jb	short .column_st16
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmE
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mmC
+
+	sub	ecx, byte SIZEOF_MMWORD
+	jz	short .nextrow
+
+	add	esi, byte SIZEOF_MMWORD			; inptr0
+	add	ebx, byte SIZEOF_MMWORD			; inptr1
+	add	edx, byte SIZEOF_MMWORD			; inptr2
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD	; outptr
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st16:
+	lea	ecx, [ecx+ecx*2]	; imul ecx, RGB_PIXELSIZE
+	cmp	ecx, byte 2*SIZEOF_MMWORD
+	jb	short .column_st8
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmE
+	movq	mmA,mmC
+	sub	ecx, byte 2*SIZEOF_MMWORD
+	add	edi, byte 2*SIZEOF_MMWORD
+	jmp	short .column_st4
+.column_st8:
+	cmp	ecx, byte SIZEOF_MMWORD
+	jb	short .column_st4
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	mmA,mmE
+	sub	ecx, byte SIZEOF_MMWORD
+	add	edi, byte SIZEOF_MMWORD
+.column_st4:
+	movd	eax,mmA
+	cmp	ecx, byte SIZEOF_DWORD
+	jb	short .column_st2
+	mov	DWORD [edi+0*SIZEOF_DWORD], eax
+	psrlq	mmA,DWORD_BIT
+	movd	eax,mmA
+	sub	ecx, byte SIZEOF_DWORD
+	add	edi, byte SIZEOF_DWORD
+.column_st2:
+	cmp	ecx, byte SIZEOF_WORD
+	jb	short .column_st1
+	mov	WORD [edi+0*SIZEOF_WORD], ax
+	shr	eax,WORD_BIT
+	sub	ecx, byte SIZEOF_WORD
+	add	edi, byte SIZEOF_WORD
+.column_st1:
+	cmp	ecx, byte SIZEOF_BYTE
+	jb	short .nextrow
+	mov	BYTE [edi+0*SIZEOF_BYTE], al
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+	pcmpeqb   mm6,mm6		; mm6=(X0 X2 X4 X6 ** ** ** **)
+	pcmpeqb   mm7,mm7		; mm7=(X1 X3 X5 X7 ** ** ** **)
+%else
+	pxor      mm6,mm6		; mm6=(X0 X2 X4 X6 ** ** ** **)
+	pxor      mm7,mm7		; mm7=(X1 X3 X5 X7 ** ** ** **)
+%endif
+	; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
+	; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
+	; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
+	; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **)
+
+	punpcklbw mmA,mmC		; mmA=(00 10 02 12 04 14 06 16)
+	punpcklbw mmE,mmG		; mmE=(20 30 22 32 24 34 26 36)
+	punpcklbw mmB,mmD		; mmB=(01 11 03 13 05 15 07 17)
+	punpcklbw mmF,mmH		; mmF=(21 31 23 33 25 35 27 37)
+
+	movq      mmC,mmA
+	punpcklwd mmA,mmE		; mmA=(00 10 20 30 02 12 22 32)
+	punpckhwd mmC,mmE		; mmC=(04 14 24 34 06 16 26 36)
+	movq      mmG,mmB
+	punpcklwd mmB,mmF		; mmB=(01 11 21 31 03 13 23 33)
+	punpckhwd mmG,mmF		; mmG=(05 15 25 35 07 17 27 37)
+
+	movq      mmD,mmA
+	punpckldq mmA,mmB		; mmA=(00 10 20 30 01 11 21 31)
+	punpckhdq mmD,mmB		; mmD=(02 12 22 32 03 13 23 33)
+	movq      mmH,mmC
+	punpckldq mmC,mmG		; mmC=(04 14 24 34 05 15 25 35)
+	punpckhdq mmH,mmG		; mmH=(06 16 26 36 07 17 27 37)
+
+	cmp	ecx, byte SIZEOF_MMWORD
+	jb	short .column_st16
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmD
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mmC
+	movq	MMWORD [edi+3*SIZEOF_MMWORD], mmH
+
+	sub	ecx, byte SIZEOF_MMWORD
+	jz	short .nextrow
+
+	add	esi, byte SIZEOF_MMWORD			; inptr0
+	add	ebx, byte SIZEOF_MMWORD			; inptr1
+	add	edx, byte SIZEOF_MMWORD			; inptr2
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD	; outptr
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st16:
+	cmp	ecx, byte SIZEOF_MMWORD/2
+	jb	short .column_st8
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmD
+	movq	mmA,mmC
+	movq	mmD,mmH
+	sub	ecx, byte SIZEOF_MMWORD/2
+	add	edi, byte 2*SIZEOF_MMWORD
+.column_st8:
+	cmp	ecx, byte SIZEOF_MMWORD/4
+	jb	short .column_st4
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	mmA,mmD
+	sub	ecx, byte SIZEOF_MMWORD/4
+	add	edi, byte 1*SIZEOF_MMWORD
+.column_st4:
+	cmp	ecx, byte SIZEOF_MMWORD/8
+	jb	short .nextrow
+	movd	DWORD [edi+0*SIZEOF_DWORD], mmA
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+	alignx	16,7
+
+.nextrow:
+	pop	ecx
+	pop	esi
+	pop	ebx
+	pop	edx
+	pop	edi
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW
+	add	ebx, byte SIZEOF_JSAMPROW
+	add	edx, byte SIZEOF_JSAMPROW
+	add	edi, byte SIZEOF_JSAMPROW	; output_buf
+	dec	eax				; num_rows
+	jg	near .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdclrss2-64.asm b/simd/jdclrss2-64.asm
new file mode 100644
index 0000000..4282bd2
--- /dev/null
+++ b/simd/jdclrss2-64.asm
@@ -0,0 +1,487 @@
+;
+; jdclrss2-64.asm - colorspace conversion (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+				
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_sse2 (JDIMENSION out_width,
+;                             JSAMPIMAGE input_buf, JDIMENSION input_row,
+;                             JSAMPARRAY output_buf, int num_rows)
+;
+
+; r10 = JDIMENSION out_width
+; r11 = JSAMPIMAGE input_buf
+; r12 = JDIMENSION input_row
+; r13 = JSAMPARRAY output_buf
+; r14 = int num_rows
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_ycc_rgb_convert_sse2)
+
+EXTN(jsimd_ycc_rgb_convert_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+	push	rbx
+
+	mov	rcx, r10	; num_cols
+	test	rcx,rcx
+	jz	near .return
+
+	push	rcx
+
+	mov	rdi, r11
+	mov	rcx, r12
+	mov	rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+	mov	rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+	mov	rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+	lea	rsi, [rsi+rcx*SIZEOF_JSAMPROW]
+	lea	rbx, [rbx+rcx*SIZEOF_JSAMPROW]
+	lea	rdx, [rdx+rcx*SIZEOF_JSAMPROW]
+
+	pop	rcx
+
+	mov	rdi, r13
+	mov	eax, r14d
+	test	rax,rax
+	jle	near .return
+.rowloop:
+	push	rax
+	push	rdi
+	push	rdx
+	push	rbx
+	push	rsi
+	push	rcx			; col
+
+	mov	rsi, JSAMPROW [rsi]	; inptr0
+	mov	rbx, JSAMPROW [rbx]	; inptr1
+	mov	rdx, JSAMPROW [rdx]	; inptr2
+	mov	rdi, JSAMPROW [rdi]	; outptr
+.columnloop:
+
+	movdqa	xmm5, XMMWORD [rbx]	; xmm5=Cb(0123456789ABCDEF)
+	movdqa	xmm1, XMMWORD [rdx]	; xmm1=Cr(0123456789ABCDEF)
+
+	pcmpeqw	xmm4,xmm4
+	pcmpeqw	xmm7,xmm7
+	psrlw	xmm4,BYTE_BIT
+	psllw	xmm7,7			; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+	movdqa	xmm0,xmm4		; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
+
+	pand	xmm4,xmm5		; xmm4=Cb(02468ACE)=CbE
+	psrlw	xmm5,BYTE_BIT		; xmm5=Cb(13579BDF)=CbO
+	pand	xmm0,xmm1		; xmm0=Cr(02468ACE)=CrE
+	psrlw	xmm1,BYTE_BIT		; xmm1=Cr(13579BDF)=CrO
+
+	paddw	xmm4,xmm7
+	paddw	xmm5,xmm7
+	paddw	xmm0,xmm7
+	paddw	xmm1,xmm7
+
+	; (Original)
+	; R = Y                + 1.40200 * Cr
+	; G = Y - 0.34414 * Cb - 0.71414 * Cr
+	; B = Y + 1.77200 * Cb
+	;
+	; (This implementation)
+	; R = Y                + 0.40200 * Cr + Cr
+	; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+	; B = Y - 0.22800 * Cb + Cb + Cb
+
+	movdqa	xmm2,xmm4		; xmm2=CbE
+	movdqa	xmm3,xmm5		; xmm3=CbO
+	paddw	xmm4,xmm4		; xmm4=2*CbE
+	paddw	xmm5,xmm5		; xmm5=2*CbO
+	movdqa	xmm6,xmm0		; xmm6=CrE
+	movdqa	xmm7,xmm1		; xmm7=CrO
+	paddw	xmm0,xmm0		; xmm0=2*CrE
+	paddw	xmm1,xmm1		; xmm1=2*CrO
+
+	pmulhw	xmm4,[rel PW_MF0228]	; xmm4=(2*CbE * -FIX(0.22800))
+	pmulhw	xmm5,[rel PW_MF0228]	; xmm5=(2*CbO * -FIX(0.22800))
+	pmulhw	xmm0,[rel PW_F0402]	; xmm0=(2*CrE * FIX(0.40200))
+	pmulhw	xmm1,[rel PW_F0402]	; xmm1=(2*CrO * FIX(0.40200))
+
+	paddw	xmm4,[rel PW_ONE]
+	paddw	xmm5,[rel PW_ONE]
+	psraw	xmm4,1			; xmm4=(CbE * -FIX(0.22800))
+	psraw	xmm5,1			; xmm5=(CbO * -FIX(0.22800))
+	paddw	xmm0,[rel PW_ONE]
+	paddw	xmm1,[rel PW_ONE]
+	psraw	xmm0,1			; xmm0=(CrE * FIX(0.40200))
+	psraw	xmm1,1			; xmm1=(CrO * FIX(0.40200))
+
+	paddw	xmm4,xmm2
+	paddw	xmm5,xmm3
+	paddw	xmm4,xmm2		; xmm4=(CbE * FIX(1.77200))=(B-Y)E
+	paddw	xmm5,xmm3		; xmm5=(CbO * FIX(1.77200))=(B-Y)O
+	paddw	xmm0,xmm6		; xmm0=(CrE * FIX(1.40200))=(R-Y)E
+	paddw	xmm1,xmm7		; xmm1=(CrO * FIX(1.40200))=(R-Y)O
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=(B-Y)E
+	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(B-Y)O
+
+	movdqa    xmm4,xmm2
+	movdqa    xmm5,xmm3
+	punpcklwd xmm2,xmm6
+	punpckhwd xmm4,xmm6
+	pmaddwd   xmm2,[rel PW_MF0344_F0285]
+	pmaddwd   xmm4,[rel PW_MF0344_F0285]
+	punpcklwd xmm3,xmm7
+	punpckhwd xmm5,xmm7
+	pmaddwd   xmm3,[rel PW_MF0344_F0285]
+	pmaddwd   xmm5,[rel PW_MF0344_F0285]
+
+	paddd     xmm2,[rel PD_ONEHALF]
+	paddd     xmm4,[rel PD_ONEHALF]
+	psrad     xmm2,SCALEBITS
+	psrad     xmm4,SCALEBITS
+	paddd     xmm3,[rel PD_ONEHALF]
+	paddd     xmm5,[rel PD_ONEHALF]
+	psrad     xmm3,SCALEBITS
+	psrad     xmm5,SCALEBITS
+
+	packssdw  xmm2,xmm4	; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+	packssdw  xmm3,xmm5	; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+	psubw     xmm2,xmm6	; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+	psubw     xmm3,xmm7	; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+	movdqa    xmm5, XMMWORD [rsi]	; xmm5=Y(0123456789ABCDEF)
+
+	pcmpeqw   xmm4,xmm4
+	psrlw     xmm4,BYTE_BIT		; xmm4={0xFF 0x00 0xFF 0x00 ..}
+	pand      xmm4,xmm5		; xmm4=Y(02468ACE)=YE
+	psrlw     xmm5,BYTE_BIT		; xmm5=Y(13579BDF)=YO
+
+	paddw     xmm0,xmm4		; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
+	paddw     xmm1,xmm5		; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
+	packuswb  xmm0,xmm0		; xmm0=R(02468ACE********)
+	packuswb  xmm1,xmm1		; xmm1=R(13579BDF********)
+
+	paddw     xmm2,xmm4		; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
+	paddw     xmm3,xmm5		; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
+	packuswb  xmm2,xmm2		; xmm2=G(02468ACE********)
+	packuswb  xmm3,xmm3		; xmm3=G(13579BDF********)
+
+	paddw     xmm4, XMMWORD [wk(0)]	; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
+	paddw     xmm5, XMMWORD [wk(1)]	; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
+	packuswb  xmm4,xmm4		; xmm4=B(02468ACE********)
+	packuswb  xmm5,xmm5		; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmB	; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+	punpcklbw xmmD,xmmF	; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+	movdqa    xmmG,xmmA
+	movdqa    xmmH,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+	punpckhwd xmmG,xmmE	; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+	psrldq    xmmH,2	; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+	psrldq    xmmE,2	; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+	movdqa    xmmC,xmmD
+	movdqa    xmmB,xmmD
+	punpcklwd xmmD,xmmH	; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+	punpckhwd xmmC,xmmH	; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+	psrldq    xmmB,2	; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+	movdqa    xmmF,xmmE
+	punpcklwd xmmE,xmmB	; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+	punpckhwd xmmF,xmmB	; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+	pshufd    xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+	movdqa    xmmB,xmmE
+	punpckldq xmmA,xmmD	; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+	punpckldq xmmE,xmmH	; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+	punpckhdq xmmD,xmmB	; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+	pshufd    xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+	movdqa    xmmB,xmmF
+	punpckldq xmmG,xmmC	; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+	punpckldq xmmF,xmmH	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+	punpckhdq xmmC,xmmB	; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+	punpcklqdq xmmA,xmmE	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+	punpcklqdq xmmD,xmmG	; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+	punpcklqdq xmmF,xmmC	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	rdi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
+	add	rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmH,xmmH			; xmmH=(all 1's)
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmF,xmmH			; movntdqu XMMWORD [rdi], xmmF
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	rcx, byte SIZEOF_XMMWORD
+	jz	near .nextrow
+
+	add	rsi, byte SIZEOF_XMMWORD	; inptr0
+	add	rbx, byte SIZEOF_XMMWORD	; inptr1
+	add	rdx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+
+.column_st32:
+	pcmpeqb	xmmH,xmmH			; xmmH=(all 1's)
+	lea	rcx, [rcx+rcx*2]		; imul ecx, RGB_PIXELSIZE
+	cmp	rcx, byte 2*SIZEOF_XMMWORD
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmF
+	sub	rcx, byte 2*SIZEOF_XMMWORD
+	jmp	short .column_st15
+.column_st16:
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	rcx, byte SIZEOF_XMMWORD
+.column_st15:
+	mov	rax,rcx
+	xor	rcx, byte 0x0F
+	shl	rcx, 2
+	movd	xmmB,ecx
+	psrlq	xmmH,4
+	pcmpeqb	xmmE,xmmE
+	psrlq	xmmH,xmmB
+	psrlq	xmmE,xmmB
+	punpcklbw xmmE,xmmH
+	; ----------------
+	mov	rcx,rdi
+	and	rcx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	add	rax,rcx
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	rdi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	rcx, 3			; pslldq xmmA,ecx & pslldq xmmE,rcx
+	movdqa	xmmG,xmmA
+	movdqa	xmmC,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmD,ecx
+	sub	rcx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmF,ecx
+	psllq	xmmA,xmmF
+	psllq	xmmE,xmmF
+	jmp	short .adj0
+.adj1:	neg	ecx
+	movd	xmmF,ecx
+	psrlq	xmmA,xmmF
+	psrlq	xmmE,xmmF
+	psllq	xmmG,xmmD
+	psllq	xmmC,xmmD
+	por	xmmA,xmmG
+	por	xmmE,xmmC
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+	pcmpeqb   xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pcmpeqb   xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%else
+	pxor      xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pxor      xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%endif
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmG	; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+	punpcklbw xmmB,xmmD	; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+	punpcklbw xmmF,xmmH	; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+	movdqa    xmmC,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+	punpckhwd xmmC,xmmE	; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+	movdqa    xmmG,xmmB
+	punpcklwd xmmB,xmmF	; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+	punpckhwd xmmG,xmmF	; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+	movdqa    xmmD,xmmA
+	punpckldq xmmA,xmmB	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+	punpckhdq xmmD,xmmB	; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+	movdqa    xmmH,xmmC
+	punpckldq xmmC,xmmG	; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+	punpckhdq xmmH,xmmG	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	rdi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
+	movntdq	XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
+	add	rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmE,xmmE			; xmmE=(all 1's)
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmC,xmmE			; movntdqu XMMWORD [rdi], xmmC
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmH,xmmE			; movntdqu XMMWORD [rdi], xmmH
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	rcx, byte SIZEOF_XMMWORD
+	jz	near .nextrow
+
+	add	rsi, byte SIZEOF_XMMWORD	; inptr0
+	add	rbx, byte SIZEOF_XMMWORD	; inptr1
+	add	rdx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+
+.column_st32:
+	pcmpeqb	xmmE,xmmE			; xmmE=(all 1's)
+	cmp	rcx, byte SIZEOF_XMMWORD/2
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmC
+	movdqa	xmmD,xmmH
+	sub	rcx, byte SIZEOF_XMMWORD/2
+.column_st16:
+	cmp	rcx, byte SIZEOF_XMMWORD/4
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	rcx, byte SIZEOF_XMMWORD/4
+.column_st15:
+	cmp	rcx, byte SIZEOF_XMMWORD/16
+	jb	near .nextrow
+	mov	rax,rcx
+	xor	rcx, byte 0x03
+	inc	rcx
+	shl	rcx, 4
+	movd	xmmF,ecx
+	psrlq	xmmE,xmmF
+	punpcklbw xmmE,xmmE
+	; ----------------
+	mov	rcx,rdi
+	and	rcx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	lea	rax, [rcx+rax*4]	; RGB_PIXELSIZE
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	rdi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	rcx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmB,xmmA
+	movdqa	xmmG,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmC,ecx
+	sub	rcx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmH,ecx
+	psllq	xmmA,xmmH
+	psllq	xmmE,xmmH
+	jmp	short .adj0
+.adj1:	neg	rcx
+	movd	xmmH,ecx
+	psrlq	xmmA,xmmH
+	psrlq	xmmE,xmmH
+	psllq	xmmB,xmmC
+	psllq	xmmG,xmmC
+	por	xmmA,xmmB
+	por	xmmE,xmmG
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+.nextrow:
+	pop	rcx
+	pop	rsi
+	pop	rbx
+	pop	rdx
+	pop	rdi
+	pop	rax
+
+	add	rsi, byte SIZEOF_JSAMPROW
+	add	rbx, byte SIZEOF_JSAMPROW
+	add	rdx, byte SIZEOF_JSAMPROW
+	add	rdi, byte SIZEOF_JSAMPROW	; output_buf
+	dec	rax				; num_rows
+	jg	near .rowloop
+
+	sfence		; flush the write buffer
+
+.return:
+	pop	rbx
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdclrss2.asm b/simd/jdclrss2.asm
new file mode 100644
index 0000000..865fa82
--- /dev/null
+++ b/simd/jdclrss2.asm
@@ -0,0 +1,505 @@
+;
+; jdclrss2.asm - colorspace conversion (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+				
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_sse2 (JDIMENSION out_width,
+;                             JSAMPIMAGE input_buf, JDIMENSION input_row,
+;                             JSAMPARRAY output_buf, int num_rows)
+;
+
+%define out_width(b)	(b)+8			; JDIMENSION out_width
+%define input_buf(b)	(b)+12		; JSAMPIMAGE input_buf
+%define input_row(b)	(b)+16		; JDIMENSION input_row
+%define output_buf(b)	(b)+20		; JSAMPARRAY output_buf
+%define num_rows(b)	(b)+24		; int num_rows
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_ycc_rgb_convert_sse2)
+
+EXTN(jsimd_ycc_rgb_convert_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	ecx, JDIMENSION [out_width(eax)]	; num_cols
+	test	ecx,ecx
+	jz	near .return
+
+	push	ecx
+
+	mov	edi, JSAMPIMAGE [input_buf(eax)]
+	mov	ecx, JDIMENSION [input_row(eax)]
+	mov	esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+	lea	esi, [esi+ecx*SIZEOF_JSAMPROW]
+	lea	ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+	lea	edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+	pop	ecx
+
+	mov	edi, JSAMPARRAY [output_buf(eax)]
+	mov	eax, INT [num_rows(eax)]
+	test	eax,eax
+	jle	near .return
+	alignx	16,7
+.rowloop:
+	push	eax
+	push	edi
+	push	edx
+	push	ebx
+	push	esi
+	push	ecx			; col
+
+	mov	esi, JSAMPROW [esi]	; inptr0
+	mov	ebx, JSAMPROW [ebx]	; inptr1
+	mov	edx, JSAMPROW [edx]	; inptr2
+	mov	edi, JSAMPROW [edi]	; outptr
+	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
+	alignx	16,7
+.columnloop:
+
+	movdqa	xmm5, XMMWORD [ebx]	; xmm5=Cb(0123456789ABCDEF)
+	movdqa	xmm1, XMMWORD [edx]	; xmm1=Cr(0123456789ABCDEF)
+
+	pcmpeqw	xmm4,xmm4
+	pcmpeqw	xmm7,xmm7
+	psrlw	xmm4,BYTE_BIT
+	psllw	xmm7,7			; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+	movdqa	xmm0,xmm4		; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
+
+	pand	xmm4,xmm5		; xmm4=Cb(02468ACE)=CbE
+	psrlw	xmm5,BYTE_BIT		; xmm5=Cb(13579BDF)=CbO
+	pand	xmm0,xmm1		; xmm0=Cr(02468ACE)=CrE
+	psrlw	xmm1,BYTE_BIT		; xmm1=Cr(13579BDF)=CrO
+
+	paddw	xmm4,xmm7
+	paddw	xmm5,xmm7
+	paddw	xmm0,xmm7
+	paddw	xmm1,xmm7
+
+	; (Original)
+	; R = Y                + 1.40200 * Cr
+	; G = Y - 0.34414 * Cb - 0.71414 * Cr
+	; B = Y + 1.77200 * Cb
+	;
+	; (This implementation)
+	; R = Y                + 0.40200 * Cr + Cr
+	; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+	; B = Y - 0.22800 * Cb + Cb + Cb
+
+	movdqa	xmm2,xmm4		; xmm2=CbE
+	movdqa	xmm3,xmm5		; xmm3=CbO
+	paddw	xmm4,xmm4		; xmm4=2*CbE
+	paddw	xmm5,xmm5		; xmm5=2*CbO
+	movdqa	xmm6,xmm0		; xmm6=CrE
+	movdqa	xmm7,xmm1		; xmm7=CrO
+	paddw	xmm0,xmm0		; xmm0=2*CrE
+	paddw	xmm1,xmm1		; xmm1=2*CrO
+
+	pmulhw	xmm4,[GOTOFF(eax,PW_MF0228)]	; xmm4=(2*CbE * -FIX(0.22800))
+	pmulhw	xmm5,[GOTOFF(eax,PW_MF0228)]	; xmm5=(2*CbO * -FIX(0.22800))
+	pmulhw	xmm0,[GOTOFF(eax,PW_F0402)]	; xmm0=(2*CrE * FIX(0.40200))
+	pmulhw	xmm1,[GOTOFF(eax,PW_F0402)]	; xmm1=(2*CrO * FIX(0.40200))
+
+	paddw	xmm4,[GOTOFF(eax,PW_ONE)]
+	paddw	xmm5,[GOTOFF(eax,PW_ONE)]
+	psraw	xmm4,1			; xmm4=(CbE * -FIX(0.22800))
+	psraw	xmm5,1			; xmm5=(CbO * -FIX(0.22800))
+	paddw	xmm0,[GOTOFF(eax,PW_ONE)]
+	paddw	xmm1,[GOTOFF(eax,PW_ONE)]
+	psraw	xmm0,1			; xmm0=(CrE * FIX(0.40200))
+	psraw	xmm1,1			; xmm1=(CrO * FIX(0.40200))
+
+	paddw	xmm4,xmm2
+	paddw	xmm5,xmm3
+	paddw	xmm4,xmm2		; xmm4=(CbE * FIX(1.77200))=(B-Y)E
+	paddw	xmm5,xmm3		; xmm5=(CbO * FIX(1.77200))=(B-Y)O
+	paddw	xmm0,xmm6		; xmm0=(CrE * FIX(1.40200))=(R-Y)E
+	paddw	xmm1,xmm7		; xmm1=(CrO * FIX(1.40200))=(R-Y)O
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=(B-Y)E
+	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(B-Y)O
+
+	movdqa    xmm4,xmm2
+	movdqa    xmm5,xmm3
+	punpcklwd xmm2,xmm6
+	punpckhwd xmm4,xmm6
+	pmaddwd   xmm2,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   xmm4,[GOTOFF(eax,PW_MF0344_F0285)]
+	punpcklwd xmm3,xmm7
+	punpckhwd xmm5,xmm7
+	pmaddwd   xmm3,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   xmm5,[GOTOFF(eax,PW_MF0344_F0285)]
+
+	paddd     xmm2,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     xmm4,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     xmm2,SCALEBITS
+	psrad     xmm4,SCALEBITS
+	paddd     xmm3,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     xmm5,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     xmm3,SCALEBITS
+	psrad     xmm5,SCALEBITS
+
+	packssdw  xmm2,xmm4	; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+	packssdw  xmm3,xmm5	; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+	psubw     xmm2,xmm6	; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+	psubw     xmm3,xmm7	; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+	movdqa    xmm5, XMMWORD [esi]	; xmm5=Y(0123456789ABCDEF)
+
+	pcmpeqw   xmm4,xmm4
+	psrlw     xmm4,BYTE_BIT		; xmm4={0xFF 0x00 0xFF 0x00 ..}
+	pand      xmm4,xmm5		; xmm4=Y(02468ACE)=YE
+	psrlw     xmm5,BYTE_BIT		; xmm5=Y(13579BDF)=YO
+
+	paddw     xmm0,xmm4		; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
+	paddw     xmm1,xmm5		; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
+	packuswb  xmm0,xmm0		; xmm0=R(02468ACE********)
+	packuswb  xmm1,xmm1		; xmm1=R(13579BDF********)
+
+	paddw     xmm2,xmm4		; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
+	paddw     xmm3,xmm5		; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
+	packuswb  xmm2,xmm2		; xmm2=G(02468ACE********)
+	packuswb  xmm3,xmm3		; xmm3=G(13579BDF********)
+
+	paddw     xmm4, XMMWORD [wk(0)]	; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
+	paddw     xmm5, XMMWORD [wk(1)]	; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
+	packuswb  xmm4,xmm4		; xmm4=B(02468ACE********)
+	packuswb  xmm5,xmm5		; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmB	; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+	punpcklbw xmmD,xmmF	; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+	movdqa    xmmG,xmmA
+	movdqa    xmmH,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+	punpckhwd xmmG,xmmE	; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+	psrldq    xmmH,2	; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+	psrldq    xmmE,2	; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+	movdqa    xmmC,xmmD
+	movdqa    xmmB,xmmD
+	punpcklwd xmmD,xmmH	; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+	punpckhwd xmmC,xmmH	; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+	psrldq    xmmB,2	; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+	movdqa    xmmF,xmmE
+	punpcklwd xmmE,xmmB	; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+	punpckhwd xmmF,xmmB	; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+	pshufd    xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+	movdqa    xmmB,xmmE
+	punpckldq xmmA,xmmD	; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+	punpckldq xmmE,xmmH	; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+	punpckhdq xmmD,xmmB	; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+	pshufd    xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+	movdqa    xmmB,xmmF
+	punpckldq xmmG,xmmC	; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+	punpckldq xmmF,xmmH	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+	punpckhdq xmmC,xmmB	; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+	punpcklqdq xmmA,xmmE	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+	punpcklqdq xmmD,xmmG	; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+	punpcklqdq xmmF,xmmC	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	edi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmH,xmmH			; xmmH=(all 1's)
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmF,xmmH			; movntdqu XMMWORD [edi], xmmF
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	ecx, byte SIZEOF_XMMWORD
+	jz	near .nextrow
+
+	add	esi, byte SIZEOF_XMMWORD	; inptr0
+	add	ebx, byte SIZEOF_XMMWORD	; inptr1
+	add	edx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st32:
+	pcmpeqb	xmmH,xmmH			; xmmH=(all 1's)
+	lea	ecx, [ecx+ecx*2]		; imul ecx, RGB_PIXELSIZE
+	cmp	ecx, byte 2*SIZEOF_XMMWORD
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmF
+	sub	ecx, byte 2*SIZEOF_XMMWORD
+	jmp	short .column_st15
+.column_st16:
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	ecx, byte SIZEOF_XMMWORD
+.column_st15:
+	mov	eax,ecx
+	xor	ecx, byte 0x0F
+	shl	ecx, 2
+	movd	xmmB,ecx
+	psrlq	xmmH,4
+	pcmpeqb	xmmE,xmmE
+	psrlq	xmmH,xmmB
+	psrlq	xmmE,xmmB
+	punpcklbw xmmE,xmmH
+	; ----------------
+	mov	ecx,edi
+	and	ecx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	add	eax,ecx
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	edi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	ecx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmG,xmmA
+	movdqa	xmmC,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmD,ecx
+	sub	ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmF,ecx
+	psllq	xmmA,xmmF
+	psllq	xmmE,xmmF
+	jmp	short .adj0
+.adj1:	neg	ecx
+	movd	xmmF,ecx
+	psrlq	xmmA,xmmF
+	psrlq	xmmE,xmmF
+	psllq	xmmG,xmmD
+	psllq	xmmC,xmmD
+	por	xmmA,xmmG
+	por	xmmE,xmmC
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+	pcmpeqb   xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pcmpeqb   xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%else
+	pxor      xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pxor      xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%endif
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmG	; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+	punpcklbw xmmB,xmmD	; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+	punpcklbw xmmF,xmmH	; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+	movdqa    xmmC,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+	punpckhwd xmmC,xmmE	; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+	movdqa    xmmG,xmmB
+	punpcklwd xmmB,xmmF	; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+	punpckhwd xmmG,xmmF	; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+	movdqa    xmmD,xmmA
+	punpckldq xmmA,xmmB	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+	punpckhdq xmmD,xmmB	; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+	movdqa    xmmH,xmmC
+	punpckldq xmmC,xmmG	; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+	punpckhdq xmmH,xmmG	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	edi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
+	movntdq	XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmE,xmmE			; xmmE=(all 1's)
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmC,xmmE			; movntdqu XMMWORD [edi], xmmC
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmH,xmmE			; movntdqu XMMWORD [edi], xmmH
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	ecx, byte SIZEOF_XMMWORD
+	jz	near .nextrow
+
+	add	esi, byte SIZEOF_XMMWORD	; inptr0
+	add	ebx, byte SIZEOF_XMMWORD	; inptr1
+	add	edx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st32:
+	pcmpeqb	xmmE,xmmE			; xmmE=(all 1's)
+	cmp	ecx, byte SIZEOF_XMMWORD/2
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmC
+	movdqa	xmmD,xmmH
+	sub	ecx, byte SIZEOF_XMMWORD/2
+.column_st16:
+	cmp	ecx, byte SIZEOF_XMMWORD/4
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	ecx, byte SIZEOF_XMMWORD/4
+.column_st15:
+	cmp	ecx, byte SIZEOF_XMMWORD/16
+	jb	short .nextrow
+	mov	eax,ecx
+	xor	ecx, byte 0x03
+	inc	ecx
+	shl	ecx, 4
+	movd	xmmF,ecx
+	psrlq	xmmE,xmmF
+	punpcklbw xmmE,xmmE
+	; ----------------
+	mov	ecx,edi
+	and	ecx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	lea	eax, [ecx+eax*4]	; RGB_PIXELSIZE
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	edi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	ecx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmB,xmmA
+	movdqa	xmmG,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmC,ecx
+	sub	ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmH,ecx
+	psllq	xmmA,xmmH
+	psllq	xmmE,xmmH
+	jmp	short .adj0
+.adj1:	neg	ecx
+	movd	xmmH,ecx
+	psrlq	xmmA,xmmH
+	psrlq	xmmE,xmmH
+	psllq	xmmB,xmmC
+	psllq	xmmG,xmmC
+	por	xmmA,xmmB
+	por	xmmE,xmmG
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+	alignx	16,7
+
+.nextrow:
+	pop	ecx
+	pop	esi
+	pop	ebx
+	pop	edx
+	pop	edi
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW
+	add	ebx, byte SIZEOF_JSAMPROW
+	add	edx, byte SIZEOF_JSAMPROW
+	add	edi, byte SIZEOF_JSAMPROW	; output_buf
+	dec	eax				; num_rows
+	jg	near .rowloop
+
+	sfence		; flush the write buffer
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdcolmmx.asm b/simd/jdcolmmx.asm
new file mode 100644
index 0000000..58775e8
--- /dev/null
+++ b/simd/jdcolmmx.asm
@@ -0,0 +1,117 @@
+;
+; jdcolmmx.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_344	equ	 22554			; FIX(0.34414)
+F_0_714	equ	 46802			; FIX(0.71414)
+F_1_402	equ	 91881			; FIX(1.40200)
+F_1_772	equ	116130			; FIX(1.77200)
+F_0_402	equ	(F_1_402 - 65536)	; FIX(1.40200) - FIX(1)
+F_0_285	equ	( 65536 - F_0_714)	; FIX(1) - FIX(0.71414)
+F_0_228	equ	(131072 - F_1_772)	; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_ycc_rgb_convert_mmx)
+
+EXTN(jconst_ycc_rgb_convert_mmx):
+
+PW_F0402	times 4 dw  F_0_402
+PW_MF0228	times 4 dw -F_0_228
+PW_MF0344_F0285	times 2 dw -F_0_344, F_0_285
+PW_ONE		times 4 dw  1
+PD_ONEHALF	times 2 dd  1 << (SCALEBITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jdclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_ycc_rgb_convert_mmx jsimd_ycc_extrgb_convert_mmx
+%include "jdclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_mmx jsimd_ycc_extrgbx_convert_mmx
+%include "jdclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_ycc_rgb_convert_mmx jsimd_ycc_extbgr_convert_mmx
+%include "jdclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_mmx jsimd_ycc_extbgrx_convert_mmx
+%include "jdclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_mmx jsimd_ycc_extxbgr_convert_mmx
+%include "jdclrmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_mmx jsimd_ycc_extxrgb_convert_mmx
+%include "jdclrmmx.asm"
diff --git a/simd/jdcolss2-64.asm b/simd/jdcolss2-64.asm
new file mode 100644
index 0000000..2e97d59
--- /dev/null
+++ b/simd/jdcolss2-64.asm
@@ -0,0 +1,117 @@
+;
+; jdcolss2-64.asm - colorspace conversion (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_344	equ	 22554			; FIX(0.34414)
+F_0_714	equ	 46802			; FIX(0.71414)
+F_1_402	equ	 91881			; FIX(1.40200)
+F_1_772	equ	116130			; FIX(1.77200)
+F_0_402	equ	(F_1_402 - 65536)	; FIX(1.40200) - FIX(1)
+F_0_285	equ	( 65536 - F_0_714)	; FIX(1) - FIX(0.71414)
+F_0_228	equ	(131072 - F_1_772)	; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_ycc_rgb_convert_sse2)
+
+EXTN(jconst_ycc_rgb_convert_sse2):
+
+PW_F0402	times 8 dw  F_0_402
+PW_MF0228	times 8 dw -F_0_228
+PW_MF0344_F0285	times 4 dw -F_0_344, F_0_285
+PW_ONE		times 8 dw  1
+PD_ONEHALF	times 4 dd  1 << (SCALEBITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jdclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extrgb_convert_sse2
+%include "jdclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extrgbx_convert_sse2
+%include "jdclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extbgr_convert_sse2
+%include "jdclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extbgrx_convert_sse2
+%include "jdclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extxbgr_convert_sse2
+%include "jdclrss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extxrgb_convert_sse2
+%include "jdclrss2-64.asm"
diff --git a/simd/jdcolss2.asm b/simd/jdcolss2.asm
new file mode 100644
index 0000000..7ae985d
--- /dev/null
+++ b/simd/jdcolss2.asm
@@ -0,0 +1,117 @@
+;
+; jdcolss2.asm - colorspace conversion (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_344	equ	 22554			; FIX(0.34414)
+F_0_714	equ	 46802			; FIX(0.71414)
+F_1_402	equ	 91881			; FIX(1.40200)
+F_1_772	equ	116130			; FIX(1.77200)
+F_0_402	equ	(F_1_402 - 65536)	; FIX(1.40200) - FIX(1)
+F_0_285	equ	( 65536 - F_0_714)	; FIX(1) - FIX(0.71414)
+F_0_228	equ	(131072 - F_1_772)	; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_ycc_rgb_convert_sse2)
+
+EXTN(jconst_ycc_rgb_convert_sse2):
+
+PW_F0402	times 8 dw  F_0_402
+PW_MF0228	times 8 dw -F_0_228
+PW_MF0344_F0285	times 4 dw -F_0_344, F_0_285
+PW_ONE		times 8 dw  1
+PD_ONEHALF	times 4 dd  1 << (SCALEBITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jdclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extrgb_convert_sse2
+%include "jdclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extrgbx_convert_sse2
+%include "jdclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extbgr_convert_sse2
+%include "jdclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extbgrx_convert_sse2
+%include "jdclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extxbgr_convert_sse2
+%include "jdclrss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_ycc_rgb_convert_sse2 jsimd_ycc_extxrgb_convert_sse2
+%include "jdclrss2.asm"
diff --git a/simd/jdct.inc b/simd/jdct.inc
new file mode 100644
index 0000000..cc62704
--- /dev/null
+++ b/simd/jdct.inc
@@ -0,0 +1,28 @@
+;
+; jdct.inc - private declarations for forward & reverse DCT subsystems
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; [TAB8]
+
+; Each IDCT routine is responsible for range-limiting its results and
+; converting them to unsigned form (0..MAXJSAMPLE).  The raw outputs could
+; be quite far out of range if the input data is corrupt, so a bulletproof
+; range-limiting step is required.  We use a mask-and-table-lookup method
+; to do the combined operations quickly.
+;
+%define RANGE_MASK  (MAXJSAMPLE * 4 + 3)  ; 2 bits wider than legal samples
+
+%define ROW(n,b,s)		((b)+(n)*(s))
+%define COL(n,b,s)		((b)+(n)*(s)*DCTSIZE)
+
+%define DWBLOCK(m,n,b,s)	((b)+(m)*DCTSIZE*(s)+(n)*SIZEOF_DWORD)
+%define MMBLOCK(m,n,b,s)	((b)+(m)*DCTSIZE*(s)+(n)*SIZEOF_MMWORD)
+%define XMMBLOCK(m,n,b,s)	((b)+(m)*DCTSIZE*(s)+(n)*SIZEOF_XMMWORD)
+
+; --------------------------------------------------------------------------
diff --git a/simd/jdmermmx.asm b/simd/jdmermmx.asm
new file mode 100644
index 0000000..fd587fb
--- /dev/null
+++ b/simd/jdmermmx.asm
@@ -0,0 +1,123 @@
+;
+; jdmermmx.asm - merged upsampling/color conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_344	equ	 22554			; FIX(0.34414)
+F_0_714	equ	 46802			; FIX(0.71414)
+F_1_402	equ	 91881			; FIX(1.40200)
+F_1_772	equ	116130			; FIX(1.77200)
+F_0_402	equ	(F_1_402 - 65536)	; FIX(1.40200) - FIX(1)
+F_0_285	equ	( 65536 - F_0_714)	; FIX(1) - FIX(0.71414)
+F_0_228	equ	(131072 - F_1_772)	; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_merged_upsample_mmx)
+
+EXTN(jconst_merged_upsample_mmx):
+
+PW_F0402	times 4 dw  F_0_402
+PW_MF0228	times 4 dw -F_0_228
+PW_MF0344_F0285	times 2 dw -F_0_344, F_0_285
+PW_ONE		times 4 dw  1
+PD_ONEHALF	times 2 dd  1 << (SCALEBITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jdmrgmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_extrgb_merged_upsample_mmx
+%define jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_extrgb_merged_upsample_mmx
+%include "jdmrgmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_extrgbx_merged_upsample_mmx
+%define jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_extrgbx_merged_upsample_mmx
+%include "jdmrgmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_extbgr_merged_upsample_mmx
+%define jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_extbgr_merged_upsample_mmx
+%include "jdmrgmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_extbgrx_merged_upsample_mmx
+%define jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_extbgrx_merged_upsample_mmx
+%include "jdmrgmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_extxbgr_merged_upsample_mmx
+%define jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_extxbgr_merged_upsample_mmx
+%include "jdmrgmmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_extxrgb_merged_upsample_mmx
+%define jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_extxrgb_merged_upsample_mmx
+%include "jdmrgmmx.asm"
diff --git a/simd/jdmerss2-64.asm b/simd/jdmerss2-64.asm
new file mode 100644
index 0000000..1f0b10f
--- /dev/null
+++ b/simd/jdmerss2-64.asm
@@ -0,0 +1,123 @@
+;
+; jdmerss2-64.asm - merged upsampling/color conversion (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_344	equ	 22554			; FIX(0.34414)
+F_0_714	equ	 46802			; FIX(0.71414)
+F_1_402	equ	 91881			; FIX(1.40200)
+F_1_772	equ	116130			; FIX(1.77200)
+F_0_402	equ	(F_1_402 - 65536)	; FIX(1.40200) - FIX(1)
+F_0_285	equ	( 65536 - F_0_714)	; FIX(1) - FIX(0.71414)
+F_0_228	equ	(131072 - F_1_772)	; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_merged_upsample_sse2)
+
+EXTN(jconst_merged_upsample_sse2):
+
+PW_F0402	times 8 dw  F_0_402
+PW_MF0228	times 8 dw -F_0_228
+PW_MF0344_F0285	times 4 dw -F_0_344, F_0_285
+PW_ONE		times 8 dw  1
+PD_ONEHALF	times 4 dd  1 << (SCALEBITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jdmrgss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extrgb_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extrgb_merged_upsample_sse2
+%include "jdmrgss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extrgbx_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extrgbx_merged_upsample_sse2
+%include "jdmrgss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extbgr_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extbgr_merged_upsample_sse2
+%include "jdmrgss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extbgrx_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extbgrx_merged_upsample_sse2
+%include "jdmrgss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extxbgr_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extxbgr_merged_upsample_sse2
+%include "jdmrgss2-64.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extxrgb_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extxrgb_merged_upsample_sse2
+%include "jdmrgss2-64.asm"
diff --git a/simd/jdmerss2.asm b/simd/jdmerss2.asm
new file mode 100644
index 0000000..2294e0d
--- /dev/null
+++ b/simd/jdmerss2.asm
@@ -0,0 +1,123 @@
+;
+; jdmerss2.asm - merged upsampling/color conversion (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS	16
+
+F_0_344	equ	 22554			; FIX(0.34414)
+F_0_714	equ	 46802			; FIX(0.71414)
+F_1_402	equ	 91881			; FIX(1.40200)
+F_1_772	equ	116130			; FIX(1.77200)
+F_0_402	equ	(F_1_402 - 65536)	; FIX(1.40200) - FIX(1)
+F_0_285	equ	( 65536 - F_0_714)	; FIX(1) - FIX(0.71414)
+F_0_228	equ	(131072 - F_1_772)	; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_merged_upsample_sse2)
+
+EXTN(jconst_merged_upsample_sse2):
+
+PW_F0402	times 8 dw  F_0_402
+PW_MF0228	times 8 dw -F_0_228
+PW_MF0344_F0285	times 4 dw -F_0_344, F_0_285
+PW_ONE		times 8 dw  1
+PD_ONEHALF	times 4 dd  1 << (SCALEBITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+%include "jdmrgss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extrgb_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extrgb_merged_upsample_sse2
+%include "jdmrgss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extrgbx_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extrgbx_merged_upsample_sse2
+%include "jdmrgss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 3
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extbgr_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extbgr_merged_upsample_sse2
+%include "jdmrgss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 2
+%define RGB_GREEN 1
+%define RGB_BLUE 0
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extbgrx_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extbgrx_merged_upsample_sse2
+%include "jdmrgss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 3
+%define RGB_GREEN 2
+%define RGB_BLUE 1
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extxbgr_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extxbgr_merged_upsample_sse2
+%include "jdmrgss2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED 1
+%define RGB_GREEN 2
+%define RGB_BLUE 3
+%define RGB_PIXELSIZE 4
+%define jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_extxrgb_merged_upsample_sse2
+%define jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_extxrgb_merged_upsample_sse2
+%include "jdmrgss2.asm"
diff --git a/simd/jdmrgmmx.asm b/simd/jdmrgmmx.asm
new file mode 100644
index 0000000..b5777a3
--- /dev/null
+++ b/simd/jdmrgmmx.asm
@@ -0,0 +1,466 @@
+;
+; jdmrgmmx.asm - merged upsampling/color conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v1_merged_upsample_mmx (JDIMENSION output_width,
+;                                 JSAMPIMAGE input_buf,
+;                                 JDIMENSION in_row_group_ctr,
+;                                 JSAMPARRAY output_buf);
+;
+
+%define output_width(b)	(b)+8			; JDIMENSION output_width
+%define input_buf(b)		(b)+12		; JSAMPIMAGE input_buf
+%define in_row_group_ctr(b)	(b)+16		; JDIMENSION in_row_group_ctr
+%define output_buf(b)		(b)+20		; JSAMPARRAY output_buf
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		3
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_merged_upsample_mmx)
+
+EXTN(jsimd_h2v1_merged_upsample_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	ecx, JDIMENSION [output_width(eax)]	; col
+	test	ecx,ecx
+	jz	near .return
+
+	push	ecx
+
+	mov	edi, JSAMPIMAGE [input_buf(eax)]
+	mov	ecx, JDIMENSION [in_row_group_ctr(eax)]
+	mov	esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+	mov	edi, JSAMPARRAY [output_buf(eax)]
+	mov	esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW]		; inptr0
+	mov	ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW]		; inptr1
+	mov	edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW]		; inptr2
+	mov	edi, JSAMPROW [edi]				; outptr
+
+	pop	ecx			; col
+
+	alignx	16,7
+.columnloop:
+	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
+
+	movq      mm6, MMWORD [ebx]	; mm6=Cb(01234567)
+	movq      mm7, MMWORD [edx]	; mm7=Cr(01234567)
+
+	pxor      mm1,mm1		; mm1=(all 0's)
+	pcmpeqw   mm3,mm3
+	psllw     mm3,7			; mm3={0xFF80 0xFF80 0xFF80 0xFF80}
+
+	movq      mm4,mm6
+	punpckhbw mm6,mm1		; mm6=Cb(4567)=CbH
+	punpcklbw mm4,mm1		; mm4=Cb(0123)=CbL
+	movq      mm0,mm7
+	punpckhbw mm7,mm1		; mm7=Cr(4567)=CrH
+	punpcklbw mm0,mm1		; mm0=Cr(0123)=CrL
+
+	paddw     mm6,mm3
+	paddw     mm4,mm3
+	paddw     mm7,mm3
+	paddw     mm0,mm3
+
+	; (Original)
+	; R = Y                + 1.40200 * Cr
+	; G = Y - 0.34414 * Cb - 0.71414 * Cr
+	; B = Y + 1.77200 * Cb
+	;
+	; (This implementation)
+	; R = Y                + 0.40200 * Cr + Cr
+	; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+	; B = Y - 0.22800 * Cb + Cb + Cb
+
+	movq	mm5,mm6			; mm5=CbH
+	movq	mm2,mm4			; mm2=CbL
+	paddw	mm6,mm6			; mm6=2*CbH
+	paddw	mm4,mm4			; mm4=2*CbL
+	movq	mm1,mm7			; mm1=CrH
+	movq	mm3,mm0			; mm3=CrL
+	paddw	mm7,mm7			; mm7=2*CrH
+	paddw	mm0,mm0			; mm0=2*CrL
+
+	pmulhw	mm6,[GOTOFF(eax,PW_MF0228)]	; mm6=(2*CbH * -FIX(0.22800))
+	pmulhw	mm4,[GOTOFF(eax,PW_MF0228)]	; mm4=(2*CbL * -FIX(0.22800))
+	pmulhw	mm7,[GOTOFF(eax,PW_F0402)]	; mm7=(2*CrH * FIX(0.40200))
+	pmulhw	mm0,[GOTOFF(eax,PW_F0402)]	; mm0=(2*CrL * FIX(0.40200))
+
+	paddw	mm6,[GOTOFF(eax,PW_ONE)]
+	paddw	mm4,[GOTOFF(eax,PW_ONE)]
+	psraw	mm6,1			; mm6=(CbH * -FIX(0.22800))
+	psraw	mm4,1			; mm4=(CbL * -FIX(0.22800))
+	paddw	mm7,[GOTOFF(eax,PW_ONE)]
+	paddw	mm0,[GOTOFF(eax,PW_ONE)]
+	psraw	mm7,1			; mm7=(CrH * FIX(0.40200))
+	psraw	mm0,1			; mm0=(CrL * FIX(0.40200))
+
+	paddw	mm6,mm5
+	paddw	mm4,mm2
+	paddw	mm6,mm5			; mm6=(CbH * FIX(1.77200))=(B-Y)H
+	paddw	mm4,mm2			; mm4=(CbL * FIX(1.77200))=(B-Y)L
+	paddw	mm7,mm1			; mm7=(CrH * FIX(1.40200))=(R-Y)H
+	paddw	mm0,mm3			; mm0=(CrL * FIX(1.40200))=(R-Y)L
+
+	movq	MMWORD [wk(0)], mm6	; wk(0)=(B-Y)H
+	movq	MMWORD [wk(1)], mm7	; wk(1)=(R-Y)H
+
+	movq      mm6,mm5
+	movq      mm7,mm2
+	punpcklwd mm5,mm1
+	punpckhwd mm6,mm1
+	pmaddwd   mm5,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   mm6,[GOTOFF(eax,PW_MF0344_F0285)]
+	punpcklwd mm2,mm3
+	punpckhwd mm7,mm3
+	pmaddwd   mm2,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   mm7,[GOTOFF(eax,PW_MF0344_F0285)]
+
+	paddd     mm5,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     mm6,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     mm5,SCALEBITS
+	psrad     mm6,SCALEBITS
+	paddd     mm2,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     mm7,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     mm2,SCALEBITS
+	psrad     mm7,SCALEBITS
+
+	packssdw  mm5,mm6	; mm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
+	packssdw  mm2,mm7	; mm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
+	psubw     mm5,mm1	; mm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
+	psubw     mm2,mm3	; mm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
+
+	movq	MMWORD [wk(2)], mm5	; wk(2)=(G-Y)H
+
+	mov	al,2			; Yctr
+	jmp	short .Yloop_1st
+	alignx	16,7
+
+.Yloop_2nd:
+	movq	mm0, MMWORD [wk(1)]	; mm0=(R-Y)H
+	movq	mm2, MMWORD [wk(2)]	; mm2=(G-Y)H
+	movq	mm4, MMWORD [wk(0)]	; mm4=(B-Y)H
+	alignx	16,7
+
+.Yloop_1st:
+	movq	mm7, MMWORD [esi]	; mm7=Y(01234567)
+
+	pcmpeqw	mm6,mm6
+	psrlw	mm6,BYTE_BIT		; mm6={0xFF 0x00 0xFF 0x00 ..}
+	pand	mm6,mm7			; mm6=Y(0246)=YE
+	psrlw	mm7,BYTE_BIT		; mm7=Y(1357)=YO
+
+	movq	mm1,mm0			; mm1=mm0=(R-Y)(L/H)
+	movq	mm3,mm2			; mm3=mm2=(G-Y)(L/H)
+	movq	mm5,mm4			; mm5=mm4=(B-Y)(L/H)
+
+	paddw     mm0,mm6		; mm0=((R-Y)+YE)=RE=(R0 R2 R4 R6)
+	paddw     mm1,mm7		; mm1=((R-Y)+YO)=RO=(R1 R3 R5 R7)
+	packuswb  mm0,mm0		; mm0=(R0 R2 R4 R6 ** ** ** **)
+	packuswb  mm1,mm1		; mm1=(R1 R3 R5 R7 ** ** ** **)
+
+	paddw     mm2,mm6		; mm2=((G-Y)+YE)=GE=(G0 G2 G4 G6)
+	paddw     mm3,mm7		; mm3=((G-Y)+YO)=GO=(G1 G3 G5 G7)
+	packuswb  mm2,mm2		; mm2=(G0 G2 G4 G6 ** ** ** **)
+	packuswb  mm3,mm3		; mm3=(G1 G3 G5 G7 ** ** ** **)
+
+	paddw     mm4,mm6		; mm4=((B-Y)+YE)=BE=(B0 B2 B4 B6)
+	paddw     mm5,mm7		; mm5=((B-Y)+YO)=BO=(B1 B3 B5 B7)
+	packuswb  mm4,mm4		; mm4=(B0 B2 B4 B6 ** ** ** **)
+	packuswb  mm5,mm5		; mm5=(B1 B3 B5 B7 ** ** ** **)
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+	; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
+	; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
+	; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
+	; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **)
+
+	punpcklbw mmA,mmC		; mmA=(00 10 02 12 04 14 06 16)
+	punpcklbw mmE,mmB		; mmE=(20 01 22 03 24 05 26 07)
+	punpcklbw mmD,mmF		; mmD=(11 21 13 23 15 25 17 27)
+
+	movq      mmG,mmA
+	movq      mmH,mmA
+	punpcklwd mmA,mmE		; mmA=(00 10 20 01 02 12 22 03)
+	punpckhwd mmG,mmE		; mmG=(04 14 24 05 06 16 26 07)
+
+	psrlq     mmH,2*BYTE_BIT	; mmH=(02 12 04 14 06 16 -- --)
+	psrlq     mmE,2*BYTE_BIT	; mmE=(22 03 24 05 26 07 -- --)
+
+	movq      mmC,mmD
+	movq      mmB,mmD
+	punpcklwd mmD,mmH		; mmD=(11 21 02 12 13 23 04 14)
+	punpckhwd mmC,mmH		; mmC=(15 25 06 16 17 27 -- --)
+
+	psrlq     mmB,2*BYTE_BIT	; mmB=(13 23 15 25 17 27 -- --)
+
+	movq      mmF,mmE
+	punpcklwd mmE,mmB		; mmE=(22 03 13 23 24 05 15 25)
+	punpckhwd mmF,mmB		; mmF=(26 07 17 27 -- -- -- --)
+
+	punpckldq mmA,mmD		; mmA=(00 10 20 01 11 21 02 12)
+	punpckldq mmE,mmG		; mmE=(22 03 13 23 04 14 24 05)
+	punpckldq mmC,mmF		; mmC=(15 25 06 16 26 07 17 27)
+
+	cmp	ecx, byte SIZEOF_MMWORD
+	jb	short .column_st16
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmE
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mmC
+
+	sub	ecx, byte SIZEOF_MMWORD
+	jz	near .endcolumn
+
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD	; outptr
+	add	esi, byte SIZEOF_MMWORD			; inptr0
+	dec	al			; Yctr
+	jnz	near .Yloop_2nd
+
+	add	ebx, byte SIZEOF_MMWORD			; inptr1
+	add	edx, byte SIZEOF_MMWORD			; inptr2
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st16:
+	lea	ecx, [ecx+ecx*2]	; imul ecx, RGB_PIXELSIZE
+	cmp	ecx, byte 2*SIZEOF_MMWORD
+	jb	short .column_st8
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmE
+	movq	mmA,mmC
+	sub	ecx, byte 2*SIZEOF_MMWORD
+	add	edi, byte 2*SIZEOF_MMWORD
+	jmp	short .column_st4
+.column_st8:
+	cmp	ecx, byte SIZEOF_MMWORD
+	jb	short .column_st4
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	mmA,mmE
+	sub	ecx, byte SIZEOF_MMWORD
+	add	edi, byte SIZEOF_MMWORD
+.column_st4:
+	movd	eax,mmA
+	cmp	ecx, byte SIZEOF_DWORD
+	jb	short .column_st2
+	mov	DWORD [edi+0*SIZEOF_DWORD], eax
+	psrlq	mmA,DWORD_BIT
+	movd	eax,mmA
+	sub	ecx, byte SIZEOF_DWORD
+	add	edi, byte SIZEOF_DWORD
+.column_st2:
+	cmp	ecx, byte SIZEOF_WORD
+	jb	short .column_st1
+	mov	WORD [edi+0*SIZEOF_WORD], ax
+	shr	eax,WORD_BIT
+	sub	ecx, byte SIZEOF_WORD
+	add	edi, byte SIZEOF_WORD
+.column_st1:
+	cmp	ecx, byte SIZEOF_BYTE
+	jb	short .endcolumn
+	mov	BYTE [edi+0*SIZEOF_BYTE], al
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+	pcmpeqb   mm6,mm6		; mm6=(X0 X2 X4 X6 ** ** ** **)
+	pcmpeqb   mm7,mm7		; mm7=(X1 X3 X5 X7 ** ** ** **)
+%else
+	pxor      mm6,mm6		; mm6=(X0 X2 X4 X6 ** ** ** **)
+	pxor      mm7,mm7		; mm7=(X1 X3 X5 X7 ** ** ** **)
+%endif
+	; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
+	; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
+	; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
+	; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **)
+
+	punpcklbw mmA,mmC		; mmA=(00 10 02 12 04 14 06 16)
+	punpcklbw mmE,mmG		; mmE=(20 30 22 32 24 34 26 36)
+	punpcklbw mmB,mmD		; mmB=(01 11 03 13 05 15 07 17)
+	punpcklbw mmF,mmH		; mmF=(21 31 23 33 25 35 27 37)
+
+	movq      mmC,mmA
+	punpcklwd mmA,mmE		; mmA=(00 10 20 30 02 12 22 32)
+	punpckhwd mmC,mmE		; mmC=(04 14 24 34 06 16 26 36)
+	movq      mmG,mmB
+	punpcklwd mmB,mmF		; mmB=(01 11 21 31 03 13 23 33)
+	punpckhwd mmG,mmF		; mmG=(05 15 25 35 07 17 27 37)
+
+	movq      mmD,mmA
+	punpckldq mmA,mmB		; mmA=(00 10 20 30 01 11 21 31)
+	punpckhdq mmD,mmB		; mmD=(02 12 22 32 03 13 23 33)
+	movq      mmH,mmC
+	punpckldq mmC,mmG		; mmC=(04 14 24 34 05 15 25 35)
+	punpckhdq mmH,mmG		; mmH=(06 16 26 36 07 17 27 37)
+
+	cmp	ecx, byte SIZEOF_MMWORD
+	jb	short .column_st16
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmD
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mmC
+	movq	MMWORD [edi+3*SIZEOF_MMWORD], mmH
+
+	sub	ecx, byte SIZEOF_MMWORD
+	jz	short .endcolumn
+
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD	; outptr
+	add	esi, byte SIZEOF_MMWORD			; inptr0
+	dec	al			; Yctr
+	jnz	near .Yloop_2nd
+
+	add	ebx, byte SIZEOF_MMWORD			; inptr1
+	add	edx, byte SIZEOF_MMWORD			; inptr2
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st16:
+	cmp	ecx, byte SIZEOF_MMWORD/2
+	jb	short .column_st8
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mmD
+	movq	mmA,mmC
+	movq	mmD,mmH
+	sub	ecx, byte SIZEOF_MMWORD/2
+	add	edi, byte 2*SIZEOF_MMWORD
+.column_st8:
+	cmp	ecx, byte SIZEOF_MMWORD/4
+	jb	short .column_st4
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mmA
+	movq	mmA,mmD
+	sub	ecx, byte SIZEOF_MMWORD/4
+	add	edi, byte 1*SIZEOF_MMWORD
+.column_st4:
+	cmp	ecx, byte SIZEOF_MMWORD/8
+	jb	short .endcolumn
+	movd	DWORD [edi+0*SIZEOF_DWORD], mmA
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+.endcolumn:
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v2_merged_upsample_mmx (JDIMENSION output_width,
+;                                 JSAMPIMAGE input_buf,
+;                                 JDIMENSION in_row_group_ctr,
+;                                 JSAMPARRAY output_buf);
+;
+
+%define output_width(b)	(b)+8			; JDIMENSION output_width
+%define input_buf(b)		(b)+12		; JSAMPIMAGE input_buf
+%define in_row_group_ctr(b)	(b)+16		; JDIMENSION in_row_group_ctr
+%define output_buf(b)		(b)+20		; JSAMPARRAY output_buf
+
+	align	16
+	global	EXTN(jsimd_h2v2_merged_upsample_mmx)
+
+EXTN(jsimd_h2v2_merged_upsample_mmx):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	eax, JDIMENSION [output_width(ebp)]
+
+	mov	edi, JSAMPIMAGE [input_buf(ebp)]
+	mov	ecx, JDIMENSION [in_row_group_ctr(ebp)]
+	mov	esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+	mov	edi, JSAMPARRAY [output_buf(ebp)]
+	lea	esi, [esi+ecx*SIZEOF_JSAMPROW]
+
+	push	edx			; inptr2
+	push	ebx			; inptr1
+	push	esi			; inptr00
+	mov	ebx,esp
+
+	push	edi			; output_buf (outptr0)
+	push	ecx			; in_row_group_ctr
+	push	ebx			; input_buf
+	push	eax			; output_width
+
+	call	near EXTN(jsimd_h2v1_merged_upsample_mmx)
+
+	add	esi, byte SIZEOF_JSAMPROW	; inptr01
+	add	edi, byte SIZEOF_JSAMPROW	; outptr1
+	mov	POINTER [ebx+0*SIZEOF_POINTER], esi
+	mov	POINTER [ebx-1*SIZEOF_POINTER], edi
+
+	call	near EXTN(jsimd_h2v1_merged_upsample_mmx)
+
+	add	esp, byte 7*SIZEOF_DWORD
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdmrgss2-64.asm b/simd/jdmrgss2-64.asm
new file mode 100644
index 0000000..121bb82
--- /dev/null
+++ b/simd/jdmrgss2-64.asm
@@ -0,0 +1,584 @@
+;
+; jdmrgss2-64.asm - merged upsampling/color conversion (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ for
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+				
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v1_merged_upsample_sse2 (JDIMENSION output_width,
+;                                  JSAMPIMAGE input_buf,
+;                                  JDIMENSION in_row_group_ctr,
+;                                  JSAMPARRAY output_buf);
+;
+
+; r10 = JDIMENSION output_width
+; r11 = JSAMPIMAGE input_buf
+; r12 = JDIMENSION in_row_group_ctr
+; r13 = JSAMPARRAY output_buf
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		3
+
+	align	16
+	global	EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+EXTN(jsimd_h2v1_merged_upsample_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+	push	rbx
+
+	mov	rcx, r10	; col
+	test	rcx,rcx
+	jz	near .return
+
+	push	rcx
+
+	mov	rdi, r11
+	mov	rcx, r12
+	mov	rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+	mov	rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+	mov	rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+	mov	rdi, r13
+	mov	rsi, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW]		; inptr0
+	mov	rbx, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW]		; inptr1
+	mov	rdx, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW]		; inptr2
+	mov	rdi, JSAMPROW [rdi]				; outptr
+
+	pop	rcx			; col
+
+.columnloop:
+
+	movdqa    xmm6, XMMWORD [rbx]	; xmm6=Cb(0123456789ABCDEF)
+	movdqa    xmm7, XMMWORD [rdx]	; xmm7=Cr(0123456789ABCDEF)
+
+	pxor      xmm1,xmm1		; xmm1=(all 0's)
+	pcmpeqw   xmm3,xmm3
+	psllw     xmm3,7		; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+	movdqa    xmm4,xmm6
+	punpckhbw xmm6,xmm1		; xmm6=Cb(89ABCDEF)=CbH
+	punpcklbw xmm4,xmm1		; xmm4=Cb(01234567)=CbL
+	movdqa    xmm0,xmm7
+	punpckhbw xmm7,xmm1		; xmm7=Cr(89ABCDEF)=CrH
+	punpcklbw xmm0,xmm1		; xmm0=Cr(01234567)=CrL
+
+	paddw     xmm6,xmm3
+	paddw     xmm4,xmm3
+	paddw     xmm7,xmm3
+	paddw     xmm0,xmm3
+
+	; (Original)
+	; R = Y                + 1.40200 * Cr
+	; G = Y - 0.34414 * Cb - 0.71414 * Cr
+	; B = Y + 1.77200 * Cb
+	;
+	; (This implementation)
+	; R = Y                + 0.40200 * Cr + Cr
+	; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+	; B = Y - 0.22800 * Cb + Cb + Cb
+
+	movdqa	xmm5,xmm6		; xmm5=CbH
+	movdqa	xmm2,xmm4		; xmm2=CbL
+	paddw	xmm6,xmm6		; xmm6=2*CbH
+	paddw	xmm4,xmm4		; xmm4=2*CbL
+	movdqa	xmm1,xmm7		; xmm1=CrH
+	movdqa	xmm3,xmm0		; xmm3=CrL
+	paddw	xmm7,xmm7		; xmm7=2*CrH
+	paddw	xmm0,xmm0		; xmm0=2*CrL
+
+	pmulhw	xmm6,[rel PW_MF0228]	; xmm6=(2*CbH * -FIX(0.22800))
+	pmulhw	xmm4,[rel PW_MF0228]	; xmm4=(2*CbL * -FIX(0.22800))
+	pmulhw	xmm7,[rel PW_F0402]	; xmm7=(2*CrH * FIX(0.40200))
+	pmulhw	xmm0,[rel PW_F0402]	; xmm0=(2*CrL * FIX(0.40200))
+
+	paddw	xmm6,[rel PW_ONE]
+	paddw	xmm4,[rel PW_ONE]
+	psraw	xmm6,1			; xmm6=(CbH * -FIX(0.22800))
+	psraw	xmm4,1			; xmm4=(CbL * -FIX(0.22800))
+	paddw	xmm7,[rel PW_ONE]
+	paddw	xmm0,[rel PW_ONE]
+	psraw	xmm7,1			; xmm7=(CrH * FIX(0.40200))
+	psraw	xmm0,1			; xmm0=(CrL * FIX(0.40200))
+
+	paddw	xmm6,xmm5
+	paddw	xmm4,xmm2
+	paddw	xmm6,xmm5		; xmm6=(CbH * FIX(1.77200))=(B-Y)H
+	paddw	xmm4,xmm2		; xmm4=(CbL * FIX(1.77200))=(B-Y)L
+	paddw	xmm7,xmm1		; xmm7=(CrH * FIX(1.40200))=(R-Y)H
+	paddw	xmm0,xmm3		; xmm0=(CrL * FIX(1.40200))=(R-Y)L
+
+	movdqa	XMMWORD [wk(0)], xmm6	; wk(0)=(B-Y)H
+	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=(R-Y)H
+
+	movdqa    xmm6,xmm5
+	movdqa    xmm7,xmm2
+	punpcklwd xmm5,xmm1
+	punpckhwd xmm6,xmm1
+	pmaddwd   xmm5,[rel PW_MF0344_F0285]
+	pmaddwd   xmm6,[rel PW_MF0344_F0285]
+	punpcklwd xmm2,xmm3
+	punpckhwd xmm7,xmm3
+	pmaddwd   xmm2,[rel PW_MF0344_F0285]
+	pmaddwd   xmm7,[rel PW_MF0344_F0285]
+
+	paddd     xmm5,[rel PD_ONEHALF]
+	paddd     xmm6,[rel PD_ONEHALF]
+	psrad     xmm5,SCALEBITS
+	psrad     xmm6,SCALEBITS
+	paddd     xmm2,[rel PD_ONEHALF]
+	paddd     xmm7,[rel PD_ONEHALF]
+	psrad     xmm2,SCALEBITS
+	psrad     xmm7,SCALEBITS
+
+	packssdw  xmm5,xmm6	; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
+	packssdw  xmm2,xmm7	; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
+	psubw     xmm5,xmm1	; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
+	psubw     xmm2,xmm3	; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
+
+	movdqa	XMMWORD [wk(2)], xmm5	; wk(2)=(G-Y)H
+
+	mov	al,2			; Yctr
+	jmp	short .Yloop_1st
+
+.Yloop_2nd:
+	movdqa	xmm0, XMMWORD [wk(1)]	; xmm0=(R-Y)H
+	movdqa	xmm2, XMMWORD [wk(2)]	; xmm2=(G-Y)H
+	movdqa	xmm4, XMMWORD [wk(0)]	; xmm4=(B-Y)H
+
+.Yloop_1st:
+	movdqa	xmm7, XMMWORD [rsi]	; xmm7=Y(0123456789ABCDEF)
+
+	pcmpeqw	xmm6,xmm6
+	psrlw	xmm6,BYTE_BIT		; xmm6={0xFF 0x00 0xFF 0x00 ..}
+	pand	xmm6,xmm7		; xmm6=Y(02468ACE)=YE
+	psrlw	xmm7,BYTE_BIT		; xmm7=Y(13579BDF)=YO
+
+	movdqa	xmm1,xmm0		; xmm1=xmm0=(R-Y)(L/H)
+	movdqa	xmm3,xmm2		; xmm3=xmm2=(G-Y)(L/H)
+	movdqa	xmm5,xmm4		; xmm5=xmm4=(B-Y)(L/H)
+
+	paddw     xmm0,xmm6		; xmm0=((R-Y)+YE)=RE=R(02468ACE)
+	paddw     xmm1,xmm7		; xmm1=((R-Y)+YO)=RO=R(13579BDF)
+	packuswb  xmm0,xmm0		; xmm0=R(02468ACE********)
+	packuswb  xmm1,xmm1		; xmm1=R(13579BDF********)
+
+	paddw     xmm2,xmm6		; xmm2=((G-Y)+YE)=GE=G(02468ACE)
+	paddw     xmm3,xmm7		; xmm3=((G-Y)+YO)=GO=G(13579BDF)
+	packuswb  xmm2,xmm2		; xmm2=G(02468ACE********)
+	packuswb  xmm3,xmm3		; xmm3=G(13579BDF********)
+
+	paddw     xmm4,xmm6		; xmm4=((B-Y)+YE)=BE=B(02468ACE)
+	paddw     xmm5,xmm7		; xmm5=((B-Y)+YO)=BO=B(13579BDF)
+	packuswb  xmm4,xmm4		; xmm4=B(02468ACE********)
+	packuswb  xmm5,xmm5		; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmB	; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+	punpcklbw xmmD,xmmF	; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+	movdqa    xmmG,xmmA
+	movdqa    xmmH,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+	punpckhwd xmmG,xmmE	; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+	psrldq    xmmH,2	; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+	psrldq    xmmE,2	; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+	movdqa    xmmC,xmmD
+	movdqa    xmmB,xmmD
+	punpcklwd xmmD,xmmH	; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+	punpckhwd xmmC,xmmH	; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+	psrldq    xmmB,2	; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+	movdqa    xmmF,xmmE
+	punpcklwd xmmE,xmmB	; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+	punpckhwd xmmF,xmmB	; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+	pshufd    xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+	movdqa    xmmB,xmmE
+	punpckldq xmmA,xmmD	; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+	punpckldq xmmE,xmmH	; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+	punpckhdq xmmD,xmmB	; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+	pshufd    xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+	movdqa    xmmB,xmmF
+	punpckldq xmmG,xmmC	; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+	punpckldq xmmF,xmmH	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+	punpckhdq xmmC,xmmB	; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+	punpcklqdq xmmA,xmmE	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+	punpcklqdq xmmD,xmmG	; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+	punpcklqdq xmmF,xmmC	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	rdi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
+	add	rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmH,xmmH			; xmmH=(all 1's)
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmF,xmmH			; movntdqu XMMWORD [rdi], xmmF
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	rcx, byte SIZEOF_XMMWORD
+	jz	near .endcolumn
+
+	add	rsi, byte SIZEOF_XMMWORD	; inptr0
+	dec	al			; Yctr
+	jnz	near .Yloop_2nd
+
+	add	rbx, byte SIZEOF_XMMWORD	; inptr1
+	add	rdx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+
+.column_st32:
+	pcmpeqb	xmmH,xmmH			; xmmH=(all 1's)
+	lea	rcx, [rcx+rcx*2]		; imul ecx, RGB_PIXELSIZE
+	cmp	rcx, byte 2*SIZEOF_XMMWORD
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmF
+	sub	rcx, byte 2*SIZEOF_XMMWORD
+	jmp	short .column_st15
+.column_st16:
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	rcx, byte SIZEOF_XMMWORD
+.column_st15:
+	mov	rax,rcx
+	xor	rcx, byte 0x0F
+	shl	rcx, 2
+	movd	xmmB,ecx
+	psrlq	xmmH,4
+	pcmpeqb	xmmE,xmmE
+	psrlq	xmmH,xmmB
+	psrlq	xmmE,xmmB
+	punpcklbw xmmE,xmmH
+	; ----------------
+	mov	rcx,rdi
+	and	rcx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	add	rax,rcx
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	rdi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	rcx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmG,xmmA
+	movdqa	xmmC,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmD,ecx
+	sub	rcx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmF,ecx
+	psllq	xmmA,xmmF
+	psllq	xmmE,xmmF
+	jmp	short .adj0
+.adj1:	neg	rcx
+	movd	xmmF,ecx
+	psrlq	xmmA,xmmF
+	psrlq	xmmE,xmmF
+	psllq	xmmG,xmmD
+	psllq	xmmC,xmmD
+	por	xmmA,xmmG
+	por	xmmE,xmmC
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+	pcmpeqb   xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pcmpeqb   xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%else
+	pxor      xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pxor      xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%endif
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmG	; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+	punpcklbw xmmB,xmmD	; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+	punpcklbw xmmF,xmmH	; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+	movdqa    xmmC,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+	punpckhwd xmmC,xmmE	; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+	movdqa    xmmG,xmmB
+	punpcklwd xmmB,xmmF	; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+	punpckhwd xmmG,xmmF	; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+	movdqa    xmmD,xmmA
+	punpckldq xmmA,xmmB	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+	punpckhdq xmmD,xmmB	; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+	movdqa    xmmH,xmmC
+	punpckldq xmmC,xmmG	; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+	punpckhdq xmmH,xmmG	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+	cmp	rcx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	rdi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
+	movntdq	XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
+	add	rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmE,xmmE			; xmmE=(all 1's)
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmC,xmmE			; movntdqu XMMWORD [rdi], xmmC
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmH,xmmE			; movntdqu XMMWORD [rdi], xmmH
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	rcx, byte SIZEOF_XMMWORD
+	jz	near .endcolumn
+
+	add	rsi, byte SIZEOF_XMMWORD	; inptr0
+	dec	al			; Yctr
+	jnz	near .Yloop_2nd
+
+	add	rbx, byte SIZEOF_XMMWORD	; inptr1
+	add	rdx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+
+.column_st32:
+	pcmpeqb	xmmE,xmmE			; xmmE=(all 1's)
+	cmp	rcx, byte SIZEOF_XMMWORD/2
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [rdi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [rdi], xmmD
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmC
+	movdqa	xmmD,xmmH
+	sub	rcx, byte SIZEOF_XMMWORD/2
+.column_st16:
+	cmp	rcx, byte SIZEOF_XMMWORD/4
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	rdi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	rcx, byte SIZEOF_XMMWORD/4
+.column_st15:
+	cmp	rcx, byte SIZEOF_XMMWORD/16
+	jb	near .endcolumn
+	mov	rax,rcx
+	xor	rcx, byte 0x03
+	inc	rcx
+	shl	rcx, 4
+	movd	xmmF,ecx
+	psrlq	xmmE,xmmF
+	punpcklbw xmmE,xmmE
+	; ----------------
+	mov	rcx,rdi
+	and	rcx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	lea	rax, [rcx+rax*4]	; RGB_PIXELSIZE
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	rdi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	rcx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmB,xmmA
+	movdqa	xmmG,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmC,ecx
+	sub	rcx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmH,ecx
+	psllq	xmmA,xmmH
+	psllq	xmmE,xmmH
+	jmp	short .adj0
+.adj1:	neg	rcx
+	movd	xmmH,ecx
+	psrlq	xmmA,xmmH
+	psrlq	xmmE,xmmH
+	psllq	xmmB,xmmC
+	psllq	xmmG,xmmC
+	por	xmmA,xmmB
+	por	xmmE,xmmG
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+.endcolumn:
+	sfence		; flush the write buffer
+
+.return:
+	pop	rbx
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v2_merged_upsample_sse2 (JDIMENSION output_width,
+;                                  JSAMPIMAGE input_buf,
+;                                  JDIMENSION in_row_group_ctr,
+;                                  JSAMPARRAY output_buf);
+;
+
+; r10 = JDIMENSION output_width
+; r11 = JSAMPIMAGE input_buf
+; r12 = JDIMENSION in_row_group_ctr
+; r13 = JSAMPARRAY output_buf
+
+	align	16
+	global	EXTN(jsimd_h2v2_merged_upsample_sse2)
+
+EXTN(jsimd_h2v2_merged_upsample_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+	push	rbx
+
+	mov	rax, r10
+
+	mov	rdi, r11
+	mov	rcx, r12
+	mov	rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+	mov	rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+	mov	rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+	mov	rdi, r13
+	lea	rsi, [rsi+rcx*SIZEOF_JSAMPROW]
+
+	push	rdx			; inptr2
+	push	rbx			; inptr1
+	push	rsi			; inptr00
+	mov	rbx,rsp
+
+	push	rdi
+	push	rcx
+	push	rax
+
+	%ifdef WIN64
+	mov r8, rcx
+	mov r9, rdi
+	mov rcx, rax
+	mov rdx, rbx
+	%else
+	mov rdx, rcx
+	mov rcx, rdi
+	mov	rdi, rax
+	mov rsi, rbx
+	%endif
+
+	call	EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+	pop rax
+	pop rcx
+	pop rdi
+	pop rsi
+	pop rbx
+	pop rdx
+
+	add	rdi, byte SIZEOF_JSAMPROW	; outptr1
+	add	rsi, byte SIZEOF_JSAMPROW	; inptr01
+
+	push	rdx			; inptr2
+	push	rbx			; inptr1
+	push	rsi			; inptr00
+	mov	rbx,rsp
+
+	push	rdi
+	push	rcx
+	push	rax
+
+	%ifdef WIN64
+	mov r8, rcx
+	mov r9, rdi
+	mov rcx, rax
+	mov rdx, rbx
+	%else
+	mov rdx, rcx
+	mov rcx, rdi
+	mov	rdi, rax
+	mov rsi, rbx
+	%endif
+
+	call	EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+	pop rax
+	pop rcx
+	pop rdi
+	pop rsi
+	pop rbx
+	pop rdx
+
+	pop	rbx
+	uncollect_args
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdmrgss2.asm b/simd/jdmrgss2.asm
new file mode 100644
index 0000000..99b7eb9
--- /dev/null
+++ b/simd/jdmrgss2.asm
@@ -0,0 +1,564 @@
+;
+; jdmrgss2.asm - merged upsampling/color conversion (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jcolsamp.inc"
+				
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v1_merged_upsample_sse2 (JDIMENSION output_width,
+;                                  JSAMPIMAGE input_buf,
+;                                  JDIMENSION in_row_group_ctr,
+;                                  JSAMPARRAY output_buf);
+;
+
+%define output_width(b)	(b)+8			; JDIMENSION output_width
+%define input_buf(b)		(b)+12		; JSAMPIMAGE input_buf
+%define in_row_group_ctr(b)	(b)+16		; JDIMENSION in_row_group_ctr
+%define output_buf(b)		(b)+20		; JSAMPARRAY output_buf
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		3
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+EXTN(jsimd_h2v1_merged_upsample_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	ecx, JDIMENSION [output_width(eax)]	; col
+	test	ecx,ecx
+	jz	near .return
+
+	push	ecx
+
+	mov	edi, JSAMPIMAGE [input_buf(eax)]
+	mov	ecx, JDIMENSION [in_row_group_ctr(eax)]
+	mov	esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+	mov	edi, JSAMPARRAY [output_buf(eax)]
+	mov	esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW]		; inptr0
+	mov	ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW]		; inptr1
+	mov	edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW]		; inptr2
+	mov	edi, JSAMPROW [edi]				; outptr
+
+	pop	ecx			; col
+
+	alignx	16,7
+.columnloop:
+	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
+
+	movdqa    xmm6, XMMWORD [ebx]	; xmm6=Cb(0123456789ABCDEF)
+	movdqa    xmm7, XMMWORD [edx]	; xmm7=Cr(0123456789ABCDEF)
+
+	pxor      xmm1,xmm1		; xmm1=(all 0's)
+	pcmpeqw   xmm3,xmm3
+	psllw     xmm3,7		; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+	movdqa    xmm4,xmm6
+	punpckhbw xmm6,xmm1		; xmm6=Cb(89ABCDEF)=CbH
+	punpcklbw xmm4,xmm1		; xmm4=Cb(01234567)=CbL
+	movdqa    xmm0,xmm7
+	punpckhbw xmm7,xmm1		; xmm7=Cr(89ABCDEF)=CrH
+	punpcklbw xmm0,xmm1		; xmm0=Cr(01234567)=CrL
+
+	paddw     xmm6,xmm3
+	paddw     xmm4,xmm3
+	paddw     xmm7,xmm3
+	paddw     xmm0,xmm3
+
+	; (Original)
+	; R = Y                + 1.40200 * Cr
+	; G = Y - 0.34414 * Cb - 0.71414 * Cr
+	; B = Y + 1.77200 * Cb
+	;
+	; (This implementation)
+	; R = Y                + 0.40200 * Cr + Cr
+	; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+	; B = Y - 0.22800 * Cb + Cb + Cb
+
+	movdqa	xmm5,xmm6		; xmm5=CbH
+	movdqa	xmm2,xmm4		; xmm2=CbL
+	paddw	xmm6,xmm6		; xmm6=2*CbH
+	paddw	xmm4,xmm4		; xmm4=2*CbL
+	movdqa	xmm1,xmm7		; xmm1=CrH
+	movdqa	xmm3,xmm0		; xmm3=CrL
+	paddw	xmm7,xmm7		; xmm7=2*CrH
+	paddw	xmm0,xmm0		; xmm0=2*CrL
+
+	pmulhw	xmm6,[GOTOFF(eax,PW_MF0228)]	; xmm6=(2*CbH * -FIX(0.22800))
+	pmulhw	xmm4,[GOTOFF(eax,PW_MF0228)]	; xmm4=(2*CbL * -FIX(0.22800))
+	pmulhw	xmm7,[GOTOFF(eax,PW_F0402)]	; xmm7=(2*CrH * FIX(0.40200))
+	pmulhw	xmm0,[GOTOFF(eax,PW_F0402)]	; xmm0=(2*CrL * FIX(0.40200))
+
+	paddw	xmm6,[GOTOFF(eax,PW_ONE)]
+	paddw	xmm4,[GOTOFF(eax,PW_ONE)]
+	psraw	xmm6,1			; xmm6=(CbH * -FIX(0.22800))
+	psraw	xmm4,1			; xmm4=(CbL * -FIX(0.22800))
+	paddw	xmm7,[GOTOFF(eax,PW_ONE)]
+	paddw	xmm0,[GOTOFF(eax,PW_ONE)]
+	psraw	xmm7,1			; xmm7=(CrH * FIX(0.40200))
+	psraw	xmm0,1			; xmm0=(CrL * FIX(0.40200))
+
+	paddw	xmm6,xmm5
+	paddw	xmm4,xmm2
+	paddw	xmm6,xmm5		; xmm6=(CbH * FIX(1.77200))=(B-Y)H
+	paddw	xmm4,xmm2		; xmm4=(CbL * FIX(1.77200))=(B-Y)L
+	paddw	xmm7,xmm1		; xmm7=(CrH * FIX(1.40200))=(R-Y)H
+	paddw	xmm0,xmm3		; xmm0=(CrL * FIX(1.40200))=(R-Y)L
+
+	movdqa	XMMWORD [wk(0)], xmm6	; wk(0)=(B-Y)H
+	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=(R-Y)H
+
+	movdqa    xmm6,xmm5
+	movdqa    xmm7,xmm2
+	punpcklwd xmm5,xmm1
+	punpckhwd xmm6,xmm1
+	pmaddwd   xmm5,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   xmm6,[GOTOFF(eax,PW_MF0344_F0285)]
+	punpcklwd xmm2,xmm3
+	punpckhwd xmm7,xmm3
+	pmaddwd   xmm2,[GOTOFF(eax,PW_MF0344_F0285)]
+	pmaddwd   xmm7,[GOTOFF(eax,PW_MF0344_F0285)]
+
+	paddd     xmm5,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     xmm6,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     xmm5,SCALEBITS
+	psrad     xmm6,SCALEBITS
+	paddd     xmm2,[GOTOFF(eax,PD_ONEHALF)]
+	paddd     xmm7,[GOTOFF(eax,PD_ONEHALF)]
+	psrad     xmm2,SCALEBITS
+	psrad     xmm7,SCALEBITS
+
+	packssdw  xmm5,xmm6	; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
+	packssdw  xmm2,xmm7	; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
+	psubw     xmm5,xmm1	; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
+	psubw     xmm2,xmm3	; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
+
+	movdqa	XMMWORD [wk(2)], xmm5	; wk(2)=(G-Y)H
+
+	mov	al,2			; Yctr
+	jmp	short .Yloop_1st
+	alignx	16,7
+
+.Yloop_2nd:
+	movdqa	xmm0, XMMWORD [wk(1)]	; xmm0=(R-Y)H
+	movdqa	xmm2, XMMWORD [wk(2)]	; xmm2=(G-Y)H
+	movdqa	xmm4, XMMWORD [wk(0)]	; xmm4=(B-Y)H
+	alignx	16,7
+
+.Yloop_1st:
+	movdqa	xmm7, XMMWORD [esi]	; xmm7=Y(0123456789ABCDEF)
+
+	pcmpeqw	xmm6,xmm6
+	psrlw	xmm6,BYTE_BIT		; xmm6={0xFF 0x00 0xFF 0x00 ..}
+	pand	xmm6,xmm7		; xmm6=Y(02468ACE)=YE
+	psrlw	xmm7,BYTE_BIT		; xmm7=Y(13579BDF)=YO
+
+	movdqa	xmm1,xmm0		; xmm1=xmm0=(R-Y)(L/H)
+	movdqa	xmm3,xmm2		; xmm3=xmm2=(G-Y)(L/H)
+	movdqa	xmm5,xmm4		; xmm5=xmm4=(B-Y)(L/H)
+
+	paddw     xmm0,xmm6		; xmm0=((R-Y)+YE)=RE=R(02468ACE)
+	paddw     xmm1,xmm7		; xmm1=((R-Y)+YO)=RO=R(13579BDF)
+	packuswb  xmm0,xmm0		; xmm0=R(02468ACE********)
+	packuswb  xmm1,xmm1		; xmm1=R(13579BDF********)
+
+	paddw     xmm2,xmm6		; xmm2=((G-Y)+YE)=GE=G(02468ACE)
+	paddw     xmm3,xmm7		; xmm3=((G-Y)+YO)=GO=G(13579BDF)
+	packuswb  xmm2,xmm2		; xmm2=G(02468ACE********)
+	packuswb  xmm3,xmm3		; xmm3=G(13579BDF********)
+
+	paddw     xmm4,xmm6		; xmm4=((B-Y)+YE)=BE=B(02468ACE)
+	paddw     xmm5,xmm7		; xmm5=((B-Y)+YO)=BO=B(13579BDF)
+	packuswb  xmm4,xmm4		; xmm4=B(02468ACE********)
+	packuswb  xmm5,xmm5		; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3 ; ---------------
+
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmB	; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+	punpcklbw xmmD,xmmF	; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+	movdqa    xmmG,xmmA
+	movdqa    xmmH,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+	punpckhwd xmmG,xmmE	; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+	psrldq    xmmH,2	; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+	psrldq    xmmE,2	; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+	movdqa    xmmC,xmmD
+	movdqa    xmmB,xmmD
+	punpcklwd xmmD,xmmH	; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+	punpckhwd xmmC,xmmH	; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+	psrldq    xmmB,2	; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+	movdqa    xmmF,xmmE
+	punpcklwd xmmE,xmmB	; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+	punpckhwd xmmF,xmmB	; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+	pshufd    xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+	movdqa    xmmB,xmmE
+	punpckldq xmmA,xmmD	; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+	punpckldq xmmE,xmmH	; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+	punpckhdq xmmD,xmmB	; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+	pshufd    xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+	movdqa    xmmB,xmmF
+	punpckldq xmmG,xmmC	; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+	punpckldq xmmF,xmmH	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+	punpckhdq xmmC,xmmB	; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+	punpcklqdq xmmA,xmmE	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+	punpcklqdq xmmD,xmmG	; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+	punpcklqdq xmmF,xmmC	; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	edi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmH,xmmH			; xmmH=(all 1's)
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmF,xmmH			; movntdqu XMMWORD [edi], xmmF
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	ecx, byte SIZEOF_XMMWORD
+	jz	near .endcolumn
+
+	add	esi, byte SIZEOF_XMMWORD	; inptr0
+	dec	al			; Yctr
+	jnz	near .Yloop_2nd
+
+	add	ebx, byte SIZEOF_XMMWORD	; inptr1
+	add	edx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st32:
+	pcmpeqb	xmmH,xmmH			; xmmH=(all 1's)
+	lea	ecx, [ecx+ecx*2]		; imul ecx, RGB_PIXELSIZE
+	cmp	ecx, byte 2*SIZEOF_XMMWORD
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmH			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmF
+	sub	ecx, byte 2*SIZEOF_XMMWORD
+	jmp	short .column_st15
+.column_st16:
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmH			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	ecx, byte SIZEOF_XMMWORD
+.column_st15:
+	mov	eax,ecx
+	xor	ecx, byte 0x0F
+	shl	ecx, 2
+	movd	xmmB,ecx
+	psrlq	xmmH,4
+	pcmpeqb	xmmE,xmmE
+	psrlq	xmmH,xmmB
+	psrlq	xmmE,xmmB
+	punpcklbw xmmE,xmmH
+	; ----------------
+	mov	ecx,edi
+	and	ecx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	add	eax,ecx
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	edi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	ecx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmG,xmmA
+	movdqa	xmmC,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmD,ecx
+	sub	ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmF,ecx
+	psllq	xmmA,xmmF
+	psllq	xmmE,xmmF
+	jmp	short .adj0
+.adj1:	neg	ecx
+	movd	xmmF,ecx
+	psrlq	xmmA,xmmF
+	psrlq	xmmE,xmmF
+	psllq	xmmG,xmmD
+	psllq	xmmC,xmmD
+	por	xmmA,xmmG
+	por	xmmE,xmmC
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+
+%else ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+	pcmpeqb   xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pcmpeqb   xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%else
+	pxor      xmm6,xmm6		; xmm6=XE=X(02468ACE********)
+	pxor      xmm7,xmm7		; xmm7=XO=X(13579BDF********)
+%endif
+	; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+	; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+	; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+	; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+	punpcklbw xmmA,xmmC	; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+	punpcklbw xmmE,xmmG	; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+	punpcklbw xmmB,xmmD	; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+	punpcklbw xmmF,xmmH	; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+	movdqa    xmmC,xmmA
+	punpcklwd xmmA,xmmE	; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+	punpckhwd xmmC,xmmE	; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+	movdqa    xmmG,xmmB
+	punpcklwd xmmB,xmmF	; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+	punpckhwd xmmG,xmmF	; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+	movdqa    xmmD,xmmA
+	punpckldq xmmA,xmmB	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+	punpckhdq xmmD,xmmB	; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+	movdqa    xmmH,xmmC
+	punpckldq xmmC,xmmG	; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+	punpckhdq xmmH,xmmG	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+	cmp	ecx, byte SIZEOF_XMMWORD
+	jb	short .column_st32
+
+	test	edi, SIZEOF_XMMWORD-1
+	jnz	short .out1
+	; --(aligned)-------------------
+	movntdq	XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+	movntdq	XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+	movntdq	XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
+	movntdq	XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
+	add	edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; outptr
+	jmp	short .out0
+.out1:	; --(unaligned)-----------------
+	pcmpeqb    xmmE,xmmE			; xmmE=(all 1's)
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmC,xmmE			; movntdqu XMMWORD [edi], xmmC
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmH,xmmE			; movntdqu XMMWORD [edi], xmmH
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+.out0:
+	sub	ecx, byte SIZEOF_XMMWORD
+	jz	near .endcolumn
+
+	add	esi, byte SIZEOF_XMMWORD	; inptr0
+	dec	al			; Yctr
+	jnz	near .Yloop_2nd
+
+	add	ebx, byte SIZEOF_XMMWORD	; inptr1
+	add	edx, byte SIZEOF_XMMWORD	; inptr2
+	jmp	near .columnloop
+	alignx	16,7
+
+.column_st32:
+	pcmpeqb	xmmE,xmmE			; xmmE=(all 1's)
+	cmp	ecx, byte SIZEOF_XMMWORD/2
+	jb	short .column_st16
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	maskmovdqu xmmD,xmmE			; movntdqu XMMWORD [edi], xmmD
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmC
+	movdqa	xmmD,xmmH
+	sub	ecx, byte SIZEOF_XMMWORD/2
+.column_st16:
+	cmp	ecx, byte SIZEOF_XMMWORD/4
+	jb	short .column_st15
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+	add	edi, byte SIZEOF_XMMWORD	; outptr
+	movdqa	xmmA,xmmD
+	sub	ecx, byte SIZEOF_XMMWORD/4
+.column_st15:
+	cmp	ecx, byte SIZEOF_XMMWORD/16
+	jb	short .endcolumn
+	mov	eax,ecx
+	xor	ecx, byte 0x03
+	inc	ecx
+	shl	ecx, 4
+	movd	xmmF,ecx
+	psrlq	xmmE,xmmF
+	punpcklbw xmmE,xmmE
+	; ----------------
+	mov	ecx,edi
+	and	ecx, byte SIZEOF_XMMWORD-1
+	jz	short .adj0
+	lea	eax, [ecx+eax*4]	; RGB_PIXELSIZE
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	short .adj0
+	and	edi, byte (-SIZEOF_XMMWORD)	; align to 16-byte boundary
+	shl	ecx, 3			; pslldq xmmA,ecx & pslldq xmmE,ecx
+	movdqa	xmmB,xmmA
+	movdqa	xmmG,xmmE
+	pslldq	xmmA, SIZEOF_XMMWORD/2
+	pslldq	xmmE, SIZEOF_XMMWORD/2
+	movd	xmmC,ecx
+	sub	ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
+	jb	short .adj1
+	movd	xmmH,ecx
+	psllq	xmmA,xmmH
+	psllq	xmmE,xmmH
+	jmp	short .adj0
+.adj1:	neg	ecx
+	movd	xmmH,ecx
+	psrlq	xmmA,xmmH
+	psrlq	xmmE,xmmH
+	psllq	xmmB,xmmC
+	psllq	xmmG,xmmC
+	por	xmmA,xmmB
+	por	xmmE,xmmG
+.adj0:	; ----------------
+	maskmovdqu xmmA,xmmE			; movntdqu XMMWORD [edi], xmmA
+
+%endif ; RGB_PIXELSIZE ; ---------------
+
+.endcolumn:
+	sfence		; flush the write buffer
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+;
+; GLOBAL(void)
+; jsimd_h2v2_merged_upsample_sse2 (JDIMENSION output_width,
+;                                  JSAMPIMAGE input_buf,
+;                                  JDIMENSION in_row_group_ctr,
+;                                  JSAMPARRAY output_buf);
+;
+
+%define output_width(b)	(b)+8			; JDIMENSION output_width
+%define input_buf(b)		(b)+12		; JSAMPIMAGE input_buf
+%define in_row_group_ctr(b)	(b)+16		; JDIMENSION in_row_group_ctr
+%define output_buf(b)		(b)+20		; JSAMPARRAY output_buf
+
+	align	16
+	global	EXTN(jsimd_h2v2_merged_upsample_sse2)
+
+EXTN(jsimd_h2v2_merged_upsample_sse2):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	eax, POINTER [output_width(ebp)]
+
+	mov	edi, JSAMPIMAGE [input_buf(ebp)]
+	mov	ecx, JDIMENSION [in_row_group_ctr(ebp)]
+	mov	esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+	mov	ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+	mov	edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+	mov	edi, JSAMPARRAY [output_buf(ebp)]
+	lea	esi, [esi+ecx*SIZEOF_JSAMPROW]
+
+	push	edx			; inptr2
+	push	ebx			; inptr1
+	push	esi			; inptr00
+	mov	ebx,esp
+
+	push	edi			; output_buf (outptr0)
+	push	ecx			; in_row_group_ctr
+	push	ebx			; input_buf
+	push	eax			; output_width
+
+	call	near EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+	add	esi, byte SIZEOF_JSAMPROW	; inptr01
+	add	edi, byte SIZEOF_JSAMPROW	; outptr1
+	mov	POINTER [ebx+0*SIZEOF_POINTER], esi
+	mov	POINTER [ebx-1*SIZEOF_POINTER], edi
+
+	call	near EXTN(jsimd_h2v1_merged_upsample_sse2)
+
+	add	esp, byte 7*SIZEOF_DWORD
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdsammmx.asm b/simd/jdsammmx.asm
new file mode 100644
index 0000000..c09e5b9
--- /dev/null
+++ b/simd/jdsammmx.asm
@@ -0,0 +1,737 @@
+;
+; jdsammmx.asm - upsampling (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fancy_upsample_mmx)
+
+EXTN(jconst_fancy_upsample_mmx):
+
+PW_ONE		times 4 dw  1
+PW_TWO		times 4 dw  2
+PW_THREE	times 4 dw  3
+PW_SEVEN	times 4 dw  7
+PW_EIGHT	times 4 dw  8
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+;
+; The upsampling algorithm is linear interpolation between pixel centers,
+; also known as a "triangle filter".  This is a good compromise between
+; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+; of the way between input pixel centers.
+;
+; GLOBAL(void)
+; jsimd_h2v1_fancy_upsample_mmx (int max_v_samp_factor,
+;                                JDIMENSION downsampled_width,
+;                                JSAMPARRAY input_data,
+;                                JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define downsamp_width(b)	(b)+12	; JDIMENSION downsampled_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_fancy_upsample_mmx)
+
+EXTN(jsimd_h2v1_fancy_upsample_mmx):
+	push	ebp
+	mov	ebp,esp
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	mov	eax, JDIMENSION [downsamp_width(ebp)]  ; colctr
+	test	eax,eax
+	jz	near .return
+
+	mov	ecx, INT [max_v_samp(ebp)]	; rowctr
+	test	ecx,ecx
+	jz	near .return
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, POINTER [output_data_ptr(ebp)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	eax			; colctr
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]	; inptr
+	mov	edi, JSAMPROW [edi]	; outptr
+
+	test	eax, SIZEOF_MMWORD-1
+	jz	short .skip
+	mov	dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl	; insert a dummy sample
+.skip:
+	pxor	mm0,mm0			; mm0=(all 0's)
+	pcmpeqb	mm7,mm7
+	psrlq	mm7,(SIZEOF_MMWORD-1)*BYTE_BIT
+	pand	mm7, MMWORD [esi+0*SIZEOF_MMWORD]
+
+	add	eax, byte SIZEOF_MMWORD-1
+	and	eax, byte -SIZEOF_MMWORD
+	cmp	eax, byte SIZEOF_MMWORD
+	ja	short .columnloop
+	alignx	16,7
+
+.columnloop_last:
+	pcmpeqb	mm6,mm6
+	psllq	mm6,(SIZEOF_MMWORD-1)*BYTE_BIT
+	pand	mm6, MMWORD [esi+0*SIZEOF_MMWORD]
+	jmp	short .upsample
+	alignx	16,7
+
+.columnloop:
+	movq	mm6, MMWORD [esi+1*SIZEOF_MMWORD]
+	psllq	mm6,(SIZEOF_MMWORD-1)*BYTE_BIT
+
+.upsample:
+	movq	mm1, MMWORD [esi+0*SIZEOF_MMWORD]
+	movq	mm2,mm1
+	movq	mm3,mm1			; mm1=( 0 1 2 3 4 5 6 7)
+	psllq	mm2,BYTE_BIT		; mm2=( - 0 1 2 3 4 5 6)
+	psrlq	mm3,BYTE_BIT		; mm3=( 1 2 3 4 5 6 7 -)
+
+	por	mm2,mm7			; mm2=(-1 0 1 2 3 4 5 6)
+	por	mm3,mm6			; mm3=( 1 2 3 4 5 6 7 8)
+
+	movq	mm7,mm1
+	psrlq	mm7,(SIZEOF_MMWORD-1)*BYTE_BIT	; mm7=( 7 - - - - - - -)
+
+	movq      mm4,mm1
+	punpcklbw mm1,mm0		; mm1=( 0 1 2 3)
+	punpckhbw mm4,mm0		; mm4=( 4 5 6 7)
+	movq      mm5,mm2
+	punpcklbw mm2,mm0		; mm2=(-1 0 1 2)
+	punpckhbw mm5,mm0		; mm5=( 3 4 5 6)
+	movq      mm6,mm3
+	punpcklbw mm3,mm0		; mm3=( 1 2 3 4)
+	punpckhbw mm6,mm0		; mm6=( 5 6 7 8)
+
+	pmullw	mm1,[GOTOFF(ebx,PW_THREE)]
+	pmullw	mm4,[GOTOFF(ebx,PW_THREE)]
+	paddw	mm2,[GOTOFF(ebx,PW_ONE)]
+	paddw	mm5,[GOTOFF(ebx,PW_ONE)]
+	paddw	mm3,[GOTOFF(ebx,PW_TWO)]
+	paddw	mm6,[GOTOFF(ebx,PW_TWO)]
+
+	paddw	mm2,mm1
+	paddw	mm5,mm4
+	psrlw	mm2,2			; mm2=OutLE=( 0  2  4  6)
+	psrlw	mm5,2			; mm5=OutHE=( 8 10 12 14)
+	paddw	mm3,mm1
+	paddw	mm6,mm4
+	psrlw	mm3,2			; mm3=OutLO=( 1  3  5  7)
+	psrlw	mm6,2			; mm6=OutHO=( 9 11 13 15)
+
+	psllw	mm3,BYTE_BIT
+	psllw	mm6,BYTE_BIT
+	por	mm2,mm3			; mm2=OutL=( 0  1  2  3  4  5  6  7)
+	por	mm5,mm6			; mm5=OutH=( 8  9 10 11 12 13 14 15)
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm2
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mm5
+
+	sub	eax, byte SIZEOF_MMWORD
+	add	esi, byte 1*SIZEOF_MMWORD	; inptr
+	add	edi, byte 2*SIZEOF_MMWORD	; outptr
+	cmp	eax, byte SIZEOF_MMWORD
+	ja	near .columnloop
+	test	eax,eax
+	jnz	near .columnloop_last
+
+	pop	esi
+	pop	edi
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_data
+	add	edi, byte SIZEOF_JSAMPROW	; output_data
+	dec	ecx				; rowctr
+	jg	near .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+; Again a triangle filter; see comments for h2v1 case, above.
+;
+; GLOBAL(void)
+; jsimd_h2v2_fancy_upsample_mmx (int max_v_samp_factor,
+;                                JDIMENSION downsampled_width,
+;                                JSAMPARRAY input_data,
+;                                JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define downsamp_width(b)	(b)+12	; JDIMENSION downsampled_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		4
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_h2v2_fancy_upsample_mmx)
+
+EXTN(jsimd_h2v2_fancy_upsample_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	edx,eax				; edx = original ebp
+	mov	eax, JDIMENSION [downsamp_width(edx)]  ; colctr
+	test	eax,eax
+	jz	near .return
+
+	mov	ecx, INT [max_v_samp(edx)]	; rowctr
+	test	ecx,ecx
+	jz	near .return
+
+	mov	esi, JSAMPARRAY [input_data(edx)]	; input_data
+	mov	edi, POINTER [output_data_ptr(edx)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	eax					; colctr
+	push	ecx
+	push	edi
+	push	esi
+
+	mov	ecx, JSAMPROW [esi-1*SIZEOF_JSAMPROW]	; inptr1(above)
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; inptr0
+	mov	esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; inptr1(below)
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]	; outptr0
+	mov	edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]	; outptr1
+
+	test	eax, SIZEOF_MMWORD-1
+	jz	short .skip
+	push	edx
+	mov	dl, JSAMPLE [ecx+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [ecx+eax*SIZEOF_JSAMPLE], dl
+	mov	dl, JSAMPLE [ebx+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [ebx+eax*SIZEOF_JSAMPLE], dl
+	mov	dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl	; insert a dummy sample
+	pop	edx
+.skip:
+	; -- process the first column block
+
+	movq	mm0, MMWORD [ebx+0*SIZEOF_MMWORD]	; mm0=row[ 0][0]
+	movq	mm1, MMWORD [ecx+0*SIZEOF_MMWORD]	; mm1=row[-1][0]
+	movq	mm2, MMWORD [esi+0*SIZEOF_MMWORD]	; mm2=row[+1][0]
+
+	pushpic	ebx
+	movpic	ebx, POINTER [gotptr]	; load GOT address
+
+	pxor      mm3,mm3		; mm3=(all 0's)
+	movq      mm4,mm0
+	punpcklbw mm0,mm3		; mm0=row[ 0][0]( 0 1 2 3)
+	punpckhbw mm4,mm3		; mm4=row[ 0][0]( 4 5 6 7)
+	movq      mm5,mm1
+	punpcklbw mm1,mm3		; mm1=row[-1][0]( 0 1 2 3)
+	punpckhbw mm5,mm3		; mm5=row[-1][0]( 4 5 6 7)
+	movq      mm6,mm2
+	punpcklbw mm2,mm3		; mm2=row[+1][0]( 0 1 2 3)
+	punpckhbw mm6,mm3		; mm6=row[+1][0]( 4 5 6 7)
+
+	pmullw	mm0,[GOTOFF(ebx,PW_THREE)]
+	pmullw	mm4,[GOTOFF(ebx,PW_THREE)]
+
+	pcmpeqb	mm7,mm7
+	psrlq	mm7,(SIZEOF_MMWORD-2)*BYTE_BIT
+
+	paddw	mm1,mm0			; mm1=Int0L=( 0 1 2 3)
+	paddw	mm5,mm4			; mm5=Int0H=( 4 5 6 7)
+	paddw	mm2,mm0			; mm2=Int1L=( 0 1 2 3)
+	paddw	mm6,mm4			; mm6=Int1H=( 4 5 6 7)
+
+	movq	MMWORD [edx+0*SIZEOF_MMWORD], mm1	; temporarily save
+	movq	MMWORD [edx+1*SIZEOF_MMWORD], mm5	; the intermediate data
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm2
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mm6
+
+	pand	mm1,mm7			; mm1=( 0 - - -)
+	pand	mm2,mm7			; mm2=( 0 - - -)
+
+	movq	MMWORD [wk(0)], mm1
+	movq	MMWORD [wk(1)], mm2
+
+	poppic	ebx
+
+	add	eax, byte SIZEOF_MMWORD-1
+	and	eax, byte -SIZEOF_MMWORD
+	cmp	eax, byte SIZEOF_MMWORD
+	ja	short .columnloop
+	alignx	16,7
+
+.columnloop_last:
+	; -- process the last column block
+
+	pushpic	ebx
+	movpic	ebx, POINTER [gotptr]	; load GOT address
+
+	pcmpeqb	mm1,mm1
+	psllq	mm1,(SIZEOF_MMWORD-2)*BYTE_BIT
+	movq	mm2,mm1
+
+	pand	mm1, MMWORD [edx+1*SIZEOF_MMWORD]	; mm1=( - - - 7)
+	pand	mm2, MMWORD [edi+1*SIZEOF_MMWORD]	; mm2=( - - - 7)
+
+	movq	MMWORD [wk(2)], mm1
+	movq	MMWORD [wk(3)], mm2
+
+	jmp	short .upsample
+	alignx	16,7
+
+.columnloop:
+	; -- process the next column block
+
+	movq	mm0, MMWORD [ebx+1*SIZEOF_MMWORD]	; mm0=row[ 0][1]
+	movq	mm1, MMWORD [ecx+1*SIZEOF_MMWORD]	; mm1=row[-1][1]
+	movq	mm2, MMWORD [esi+1*SIZEOF_MMWORD]	; mm2=row[+1][1]
+
+	pushpic	ebx
+	movpic	ebx, POINTER [gotptr]	; load GOT address
+
+	pxor      mm3,mm3		; mm3=(all 0's)
+	movq      mm4,mm0
+	punpcklbw mm0,mm3		; mm0=row[ 0][1]( 0 1 2 3)
+	punpckhbw mm4,mm3		; mm4=row[ 0][1]( 4 5 6 7)
+	movq      mm5,mm1
+	punpcklbw mm1,mm3		; mm1=row[-1][1]( 0 1 2 3)
+	punpckhbw mm5,mm3		; mm5=row[-1][1]( 4 5 6 7)
+	movq      mm6,mm2
+	punpcklbw mm2,mm3		; mm2=row[+1][1]( 0 1 2 3)
+	punpckhbw mm6,mm3		; mm6=row[+1][1]( 4 5 6 7)
+
+	pmullw	mm0,[GOTOFF(ebx,PW_THREE)]
+	pmullw	mm4,[GOTOFF(ebx,PW_THREE)]
+
+	paddw	mm1,mm0			; mm1=Int0L=( 0 1 2 3)
+	paddw	mm5,mm4			; mm5=Int0H=( 4 5 6 7)
+	paddw	mm2,mm0			; mm2=Int1L=( 0 1 2 3)
+	paddw	mm6,mm4			; mm6=Int1H=( 4 5 6 7)
+
+	movq	MMWORD [edx+2*SIZEOF_MMWORD], mm1	; temporarily save
+	movq	MMWORD [edx+3*SIZEOF_MMWORD], mm5	; the intermediate data
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mm2
+	movq	MMWORD [edi+3*SIZEOF_MMWORD], mm6
+
+	psllq	mm1,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm1=( - - - 0)
+	psllq	mm2,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm2=( - - - 0)
+
+	movq	MMWORD [wk(2)], mm1
+	movq	MMWORD [wk(3)], mm2
+
+.upsample:
+	; -- process the upper row
+
+	movq	mm7, MMWORD [edx+0*SIZEOF_MMWORD]	; mm7=Int0L=( 0 1 2 3)
+	movq	mm3, MMWORD [edx+1*SIZEOF_MMWORD]	; mm3=Int0H=( 4 5 6 7)
+
+	movq	mm0,mm7
+	movq	mm4,mm3
+	psrlq	mm0,2*BYTE_BIT			; mm0=( 1 2 3 -)
+	psllq	mm4,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm4=( - - - 4)
+	movq	mm5,mm7
+	movq	mm6,mm3
+	psrlq	mm5,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm5=( 3 - - -)
+	psllq	mm6,2*BYTE_BIT			; mm6=( - 4 5 6)
+
+	por	mm0,mm4				; mm0=( 1 2 3 4)
+	por	mm5,mm6				; mm5=( 3 4 5 6)
+
+	movq	mm1,mm7
+	movq	mm2,mm3
+	psllq	mm1,2*BYTE_BIT			; mm1=( - 0 1 2)
+	psrlq	mm2,2*BYTE_BIT			; mm2=( 5 6 7 -)
+	movq	mm4,mm3
+	psrlq	mm4,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm4=( 7 - - -)
+
+	por	mm1, MMWORD [wk(0)]		; mm1=(-1 0 1 2)
+	por	mm2, MMWORD [wk(2)]		; mm2=( 5 6 7 8)
+
+	movq	MMWORD [wk(0)], mm4
+
+	pmullw	mm7,[GOTOFF(ebx,PW_THREE)]
+	pmullw	mm3,[GOTOFF(ebx,PW_THREE)]
+	paddw	mm1,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	mm5,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	mm0,[GOTOFF(ebx,PW_SEVEN)]
+	paddw	mm2,[GOTOFF(ebx,PW_SEVEN)]
+
+	paddw	mm1,mm7
+	paddw	mm5,mm3
+	psrlw	mm1,4			; mm1=Out0LE=( 0  2  4  6)
+	psrlw	mm5,4			; mm5=Out0HE=( 8 10 12 14)
+	paddw	mm0,mm7
+	paddw	mm2,mm3
+	psrlw	mm0,4			; mm0=Out0LO=( 1  3  5  7)
+	psrlw	mm2,4			; mm2=Out0HO=( 9 11 13 15)
+
+	psllw	mm0,BYTE_BIT
+	psllw	mm2,BYTE_BIT
+	por	mm1,mm0			; mm1=Out0L=( 0  1  2  3  4  5  6  7)
+	por	mm5,mm2			; mm5=Out0H=( 8  9 10 11 12 13 14 15)
+
+	movq	MMWORD [edx+0*SIZEOF_MMWORD], mm1
+	movq	MMWORD [edx+1*SIZEOF_MMWORD], mm5
+
+	; -- process the lower row
+
+	movq	mm6, MMWORD [edi+0*SIZEOF_MMWORD]	; mm6=Int1L=( 0 1 2 3)
+	movq	mm4, MMWORD [edi+1*SIZEOF_MMWORD]	; mm4=Int1H=( 4 5 6 7)
+
+	movq	mm7,mm6
+	movq	mm3,mm4
+	psrlq	mm7,2*BYTE_BIT			; mm7=( 1 2 3 -)
+	psllq	mm3,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm3=( - - - 4)
+	movq	mm0,mm6
+	movq	mm2,mm4
+	psrlq	mm0,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm0=( 3 - - -)
+	psllq	mm2,2*BYTE_BIT			; mm2=( - 4 5 6)
+
+	por	mm7,mm3				; mm7=( 1 2 3 4)
+	por	mm0,mm2				; mm0=( 3 4 5 6)
+
+	movq	mm1,mm6
+	movq	mm5,mm4
+	psllq	mm1,2*BYTE_BIT			; mm1=( - 0 1 2)
+	psrlq	mm5,2*BYTE_BIT			; mm5=( 5 6 7 -)
+	movq	mm3,mm4
+	psrlq	mm3,(SIZEOF_MMWORD-2)*BYTE_BIT	; mm3=( 7 - - -)
+
+	por	mm1, MMWORD [wk(1)]		; mm1=(-1 0 1 2)
+	por	mm5, MMWORD [wk(3)]		; mm5=( 5 6 7 8)
+
+	movq	MMWORD [wk(1)], mm3
+
+	pmullw	mm6,[GOTOFF(ebx,PW_THREE)]
+	pmullw	mm4,[GOTOFF(ebx,PW_THREE)]
+	paddw	mm1,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	mm0,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	mm7,[GOTOFF(ebx,PW_SEVEN)]
+	paddw	mm5,[GOTOFF(ebx,PW_SEVEN)]
+
+	paddw	mm1,mm6
+	paddw	mm0,mm4
+	psrlw	mm1,4			; mm1=Out1LE=( 0  2  4  6)
+	psrlw	mm0,4			; mm0=Out1HE=( 8 10 12 14)
+	paddw	mm7,mm6
+	paddw	mm5,mm4
+	psrlw	mm7,4			; mm7=Out1LO=( 1  3  5  7)
+	psrlw	mm5,4			; mm5=Out1HO=( 9 11 13 15)
+
+	psllw	mm7,BYTE_BIT
+	psllw	mm5,BYTE_BIT
+	por	mm1,mm7			; mm1=Out1L=( 0  1  2  3  4  5  6  7)
+	por	mm0,mm5			; mm0=Out1H=( 8  9 10 11 12 13 14 15)
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm1
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mm0
+
+	poppic	ebx
+
+	sub	eax, byte SIZEOF_MMWORD
+	add	ecx, byte 1*SIZEOF_MMWORD	; inptr1(above)
+	add	ebx, byte 1*SIZEOF_MMWORD	; inptr0
+	add	esi, byte 1*SIZEOF_MMWORD	; inptr1(below)
+	add	edx, byte 2*SIZEOF_MMWORD	; outptr0
+	add	edi, byte 2*SIZEOF_MMWORD	; outptr1
+	cmp	eax, byte SIZEOF_MMWORD
+	ja	near .columnloop
+	test	eax,eax
+	jnz	near .columnloop_last
+
+	pop	esi
+	pop	edi
+	pop	ecx
+	pop	eax
+
+	add	esi, byte 1*SIZEOF_JSAMPROW	; input_data
+	add	edi, byte 2*SIZEOF_JSAMPROW	; output_data
+	sub	ecx, byte 2			; rowctr
+	jg	near .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v1_upsample_mmx (int max_v_samp_factor,
+;                          JDIMENSION output_width,
+;                          JSAMPARRAY input_data,
+;                          JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define output_width(b)	(b)+12		; JDIMENSION output_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_upsample_mmx)
+
+EXTN(jsimd_h2v1_upsample_mmx):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	edx, JDIMENSION [output_width(ebp)]
+	add	edx, byte (2*SIZEOF_MMWORD)-1
+	and	edx, byte -(2*SIZEOF_MMWORD)
+	jz	short .return
+
+	mov	ecx, INT [max_v_samp(ebp)]	; rowctr
+	test	ecx,ecx
+	jz	short .return
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, POINTER [output_data_ptr(ebp)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]		; inptr
+	mov	edi, JSAMPROW [edi]		; outptr
+	mov	eax,edx				; colctr
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [esi+0*SIZEOF_MMWORD]
+
+	movq      mm1,mm0
+	punpcklbw mm0,mm0
+	punpckhbw mm1,mm1
+
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm0
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mm1
+
+	sub	eax, byte 2*SIZEOF_MMWORD
+	jz	short .nextrow
+
+	movq	mm2, MMWORD [esi+1*SIZEOF_MMWORD]
+
+	movq      mm3,mm2
+	punpcklbw mm2,mm2
+	punpckhbw mm3,mm3
+
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mm2
+	movq	MMWORD [edi+3*SIZEOF_MMWORD], mm3
+
+	sub	eax, byte 2*SIZEOF_MMWORD
+	jz	short .nextrow
+
+	add	esi, byte 2*SIZEOF_MMWORD	; inptr
+	add	edi, byte 4*SIZEOF_MMWORD	; outptr
+	jmp	short .columnloop
+	alignx	16,7
+
+.nextrow:
+	pop	esi
+	pop	edi
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_data
+	add	edi, byte SIZEOF_JSAMPROW	; output_data
+	dec	ecx				; rowctr
+	jg	short .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v2_upsample_mmx (int max_v_samp_factor,
+;                          JDIMENSION output_width,
+;                          JSAMPARRAY input_data,
+;                          JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define output_width(b)	(b)+12		; JDIMENSION output_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v2_upsample_mmx)
+
+EXTN(jsimd_h2v2_upsample_mmx):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	edx, JDIMENSION [output_width(ebp)]
+	add	edx, byte (2*SIZEOF_MMWORD)-1
+	and	edx, byte -(2*SIZEOF_MMWORD)
+	jz	near .return
+
+	mov	ecx, INT [max_v_samp(ebp)]	; rowctr
+	test	ecx,ecx
+	jz	short .return
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, POINTER [output_data_ptr(ebp)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]			; inptr
+	mov	ebx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]	; outptr0
+	mov	edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]	; outptr1
+	mov	eax,edx					; colctr
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [esi+0*SIZEOF_MMWORD]
+
+	movq      mm1,mm0
+	punpcklbw mm0,mm0
+	punpckhbw mm1,mm1
+
+	movq	MMWORD [ebx+0*SIZEOF_MMWORD], mm0
+	movq	MMWORD [ebx+1*SIZEOF_MMWORD], mm1
+	movq	MMWORD [edi+0*SIZEOF_MMWORD], mm0
+	movq	MMWORD [edi+1*SIZEOF_MMWORD], mm1
+
+	sub	eax, byte 2*SIZEOF_MMWORD
+	jz	short .nextrow
+
+	movq	mm2, MMWORD [esi+1*SIZEOF_MMWORD]
+
+	movq      mm3,mm2
+	punpcklbw mm2,mm2
+	punpckhbw mm3,mm3
+
+	movq	MMWORD [ebx+2*SIZEOF_MMWORD], mm2
+	movq	MMWORD [ebx+3*SIZEOF_MMWORD], mm3
+	movq	MMWORD [edi+2*SIZEOF_MMWORD], mm2
+	movq	MMWORD [edi+3*SIZEOF_MMWORD], mm3
+
+	sub	eax, byte 2*SIZEOF_MMWORD
+	jz	short .nextrow
+
+	add	esi, byte 2*SIZEOF_MMWORD	; inptr
+	add	ebx, byte 4*SIZEOF_MMWORD	; outptr0
+	add	edi, byte 4*SIZEOF_MMWORD	; outptr1
+	jmp	short .columnloop
+	alignx	16,7
+
+.nextrow:
+	pop	esi
+	pop	edi
+
+	add	esi, byte 1*SIZEOF_JSAMPROW	; input_data
+	add	edi, byte 2*SIZEOF_JSAMPROW	; output_data
+	sub	ecx, byte 2			; rowctr
+	jg	short .rowloop
+
+	emms		; empty MMX state
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdsamss2-64.asm b/simd/jdsamss2-64.asm
new file mode 100644
index 0000000..f36c156
--- /dev/null
+++ b/simd/jdsamss2-64.asm
@@ -0,0 +1,671 @@
+;
+; jdsamss2-64.asm - upsampling (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fancy_upsample_sse2)
+
+EXTN(jconst_fancy_upsample_sse2):
+
+PW_ONE		times 8 dw  1
+PW_TWO		times 8 dw  2
+PW_THREE	times 8 dw  3
+PW_SEVEN	times 8 dw  7
+PW_EIGHT	times 8 dw  8
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+;
+; The upsampling algorithm is linear interpolation between pixel centers,
+; also known as a "triangle filter".  This is a good compromise between
+; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+; of the way between input pixel centers.
+;
+; GLOBAL(void)
+; jsimd_h2v1_fancy_upsample_sse2 (int max_v_samp_factor,
+;                                 JDIMENSION downsampled_width,
+;                                 JSAMPARRAY input_data,
+;                                 JSAMPARRAY * output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11 = JDIMENSION downsampled_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_fancy_upsample_sse2)
+
+EXTN(jsimd_h2v1_fancy_upsample_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+
+	mov	rax, r11  ; colctr
+	test	rax,rax
+	jz	near .return
+
+	mov	rcx, r10	; rowctr
+	test	rcx,rcx
+	jz	near .return
+
+	mov	rsi, r12	; input_data
+	mov	rdi, r13
+	mov	rdi, JSAMPARRAY [rdi]			; output_data
+.rowloop:
+	push	rax			; colctr
+	push	rdi
+	push	rsi
+
+	mov	rsi, JSAMPROW [rsi]	; inptr
+	mov	rdi, JSAMPROW [rdi]	; outptr
+
+	test	rax, SIZEOF_XMMWORD-1
+	jz	short .skip
+	mov	dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl	; insert a dummy sample
+.skip:
+	pxor	xmm0,xmm0		; xmm0=(all 0's)
+	pcmpeqb	xmm7,xmm7
+	psrldq	xmm7,(SIZEOF_XMMWORD-1)
+	pand	xmm7, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+
+	add	rax, byte SIZEOF_XMMWORD-1
+	and	rax, byte -SIZEOF_XMMWORD
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	short .columnloop
+
+.columnloop_last:
+	pcmpeqb	xmm6,xmm6
+	pslldq	xmm6,(SIZEOF_XMMWORD-1)
+	pand	xmm6, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	jmp	short .upsample
+
+.columnloop:
+	movdqa	xmm6, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+	pslldq	xmm6,(SIZEOF_XMMWORD-1)
+
+.upsample:
+	movdqa	xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+	movdqa	xmm2,xmm1
+	movdqa	xmm3,xmm1		; xmm1=( 0  1  2 ... 13 14 15)
+	pslldq	xmm2,1			; xmm2=(--  0  1 ... 12 13 14)
+	psrldq	xmm3,1			; xmm3=( 1  2  3 ... 14 15 --)
+
+	por	xmm2,xmm7		; xmm2=(-1  0  1 ... 12 13 14)
+	por	xmm3,xmm6		; xmm3=( 1  2  3 ... 14 15 16)
+
+	movdqa	xmm7,xmm1
+	psrldq	xmm7,(SIZEOF_XMMWORD-1)	; xmm7=(15 -- -- ... -- -- --)
+
+	movdqa    xmm4,xmm1
+	punpcklbw xmm1,xmm0		; xmm1=( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm4,xmm0		; xmm4=( 8  9 10 11 12 13 14 15)
+	movdqa    xmm5,xmm2
+	punpcklbw xmm2,xmm0		; xmm2=(-1  0  1  2  3  4  5  6)
+	punpckhbw xmm5,xmm0		; xmm5=( 7  8  9 10 11 12 13 14)
+	movdqa    xmm6,xmm3
+	punpcklbw xmm3,xmm0		; xmm3=( 1  2  3  4  5  6  7  8)
+	punpckhbw xmm6,xmm0		; xmm6=( 9 10 11 12 13 14 15 16)
+
+	pmullw	xmm1,[rel PW_THREE]
+	pmullw	xmm4,[rel PW_THREE]
+	paddw	xmm2,[rel PW_ONE]
+	paddw	xmm5,[rel PW_ONE]
+	paddw	xmm3,[rel PW_TWO]
+	paddw	xmm6,[rel PW_TWO]
+
+	paddw	xmm2,xmm1
+	paddw	xmm5,xmm4
+	psrlw	xmm2,2			; xmm2=OutLE=( 0  2  4  6  8 10 12 14)
+	psrlw	xmm5,2			; xmm5=OutHE=(16 18 20 22 24 26 28 30)
+	paddw	xmm3,xmm1
+	paddw	xmm6,xmm4
+	psrlw	xmm3,2			; xmm3=OutLO=( 1  3  5  7  9 11 13 15)
+	psrlw	xmm6,2			; xmm6=OutHO=(17 19 21 23 25 27 29 31)
+
+	psllw	xmm3,BYTE_BIT
+	psllw	xmm6,BYTE_BIT
+	por	xmm2,xmm3		; xmm2=OutL=( 0  1  2 ... 13 14 15)
+	por	xmm5,xmm6		; xmm5=OutH=(16 17 18 ... 29 30 31)
+
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm5
+
+	sub	rax, byte SIZEOF_XMMWORD
+	add	rsi, byte 1*SIZEOF_XMMWORD	; inptr
+	add	rdi, byte 2*SIZEOF_XMMWORD	; outptr
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	near .columnloop
+	test	eax,eax
+	jnz	near .columnloop_last
+
+	pop	rsi
+	pop	rdi
+	pop	rax
+
+	add	rsi, byte SIZEOF_JSAMPROW	; input_data
+	add	rdi, byte SIZEOF_JSAMPROW	; output_data
+	dec	rcx				; rowctr
+	jg	near .rowloop
+
+.return:
+	uncollect_args
+	pop	rbp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+; Again a triangle filter; see comments for h2v1 case, above.
+;
+; GLOBAL(void)
+; jsimd_h2v2_fancy_upsample_sse2 (int max_v_samp_factor,
+;                                 JDIMENSION downsampled_width,
+;                                 JSAMPARRAY input_data,
+;                                 JSAMPARRAY * output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11 = JDIMENSION downsampled_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY * output_data_ptr
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		4
+
+	align	16
+	global	EXTN(jsimd_h2v2_fancy_upsample_sse2)
+
+EXTN(jsimd_h2v2_fancy_upsample_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+	push	rbx
+
+	mov	rax, r11  ; colctr
+	test	rax,rax
+	jz	near .return
+
+	mov	rcx, r10	; rowctr
+	test	rcx,rcx
+	jz	near .return
+
+	mov	rsi, r12	; input_data
+	mov	rdi, r13
+	mov	rdi, JSAMPARRAY [rdi]			; output_data
+.rowloop:
+	push	rax					; colctr
+	push	rcx
+	push	rdi
+	push	rsi
+
+	mov	rcx, JSAMPROW [rsi-1*SIZEOF_JSAMPROW]	; inptr1(above)
+	mov	rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]	; inptr0
+	mov	rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]	; inptr1(below)
+	mov	rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]	; outptr0
+	mov	rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]	; outptr1
+
+	test	rax, SIZEOF_XMMWORD-1
+	jz	short .skip
+	push	rdx
+	mov	dl, JSAMPLE [rcx+(rax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [rcx+rax*SIZEOF_JSAMPLE], dl
+	mov	dl, JSAMPLE [rbx+(rax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [rbx+rax*SIZEOF_JSAMPLE], dl
+	mov	dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl	; insert a dummy sample
+	pop	rdx
+.skip:
+	; -- process the first column block
+
+	movdqa	xmm0, XMMWORD [rbx+0*SIZEOF_XMMWORD]	; xmm0=row[ 0][0]
+	movdqa	xmm1, XMMWORD [rcx+0*SIZEOF_XMMWORD]	; xmm1=row[-1][0]
+	movdqa	xmm2, XMMWORD [rsi+0*SIZEOF_XMMWORD]	; xmm2=row[+1][0]
+
+	pxor      xmm3,xmm3		; xmm3=(all 0's)
+	movdqa    xmm4,xmm0
+	punpcklbw xmm0,xmm3		; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm4,xmm3		; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm5,xmm1
+	punpcklbw xmm1,xmm3		; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm5,xmm3		; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm6,xmm2
+	punpcklbw xmm2,xmm3		; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm6,xmm3		; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
+
+	pmullw	xmm0,[rel PW_THREE]
+	pmullw	xmm4,[rel PW_THREE]
+
+	pcmpeqb	xmm7,xmm7
+	psrldq	xmm7,(SIZEOF_XMMWORD-2)
+
+	paddw	xmm1,xmm0		; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm5,xmm4		; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
+	paddw	xmm2,xmm0		; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm6,xmm4		; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
+
+	movdqa	XMMWORD [rdx+0*SIZEOF_XMMWORD], xmm1	; temporarily save
+	movdqa	XMMWORD [rdx+1*SIZEOF_XMMWORD], xmm5	; the intermediate data
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm6
+
+	pand	xmm1,xmm7		; xmm1=( 0 -- -- -- -- -- -- --)
+	pand	xmm2,xmm7		; xmm2=( 0 -- -- -- -- -- -- --)
+
+	movdqa	XMMWORD [wk(0)], xmm1
+	movdqa	XMMWORD [wk(1)], xmm2
+
+	add	rax, byte SIZEOF_XMMWORD-1
+	and	rax, byte -SIZEOF_XMMWORD
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	short .columnloop
+
+.columnloop_last:
+	; -- process the last column block
+
+	pcmpeqb	xmm1,xmm1
+	pslldq	xmm1,(SIZEOF_XMMWORD-2)
+	movdqa	xmm2,xmm1
+
+	pand	xmm1, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+	pand	xmm2, XMMWORD [rdi+1*SIZEOF_XMMWORD]
+
+	movdqa	XMMWORD [wk(2)], xmm1	; xmm1=(-- -- -- -- -- -- -- 15)
+	movdqa	XMMWORD [wk(3)], xmm2	; xmm2=(-- -- -- -- -- -- -- 15)
+
+	jmp	near .upsample
+
+.columnloop:
+	; -- process the next column block
+
+	movdqa	xmm0, XMMWORD [rbx+1*SIZEOF_XMMWORD]	; xmm0=row[ 0][1]
+	movdqa	xmm1, XMMWORD [rcx+1*SIZEOF_XMMWORD]	; xmm1=row[-1][1]
+	movdqa	xmm2, XMMWORD [rsi+1*SIZEOF_XMMWORD]	; xmm2=row[+1][1]
+
+	pxor      xmm3,xmm3		; xmm3=(all 0's)
+	movdqa    xmm4,xmm0
+	punpcklbw xmm0,xmm3		; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm4,xmm3		; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm5,xmm1
+	punpcklbw xmm1,xmm3		; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm5,xmm3		; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm6,xmm2
+	punpcklbw xmm2,xmm3		; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm6,xmm3		; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
+
+	pmullw	xmm0,[rel PW_THREE]
+	pmullw	xmm4,[rel PW_THREE]
+
+	paddw	xmm1,xmm0		; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm5,xmm4		; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
+	paddw	xmm2,xmm0		; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm6,xmm4		; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
+
+	movdqa	XMMWORD [rdx+2*SIZEOF_XMMWORD], xmm1	; temporarily save
+	movdqa	XMMWORD [rdx+3*SIZEOF_XMMWORD], xmm5	; the intermediate data
+	movdqa	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [rdi+3*SIZEOF_XMMWORD], xmm6
+
+	pslldq	xmm1,(SIZEOF_XMMWORD-2)	; xmm1=(-- -- -- -- -- -- --  0)
+	pslldq	xmm2,(SIZEOF_XMMWORD-2)	; xmm2=(-- -- -- -- -- -- --  0)
+
+	movdqa	XMMWORD [wk(2)], xmm1
+	movdqa	XMMWORD [wk(3)], xmm2
+
+.upsample:
+	; -- process the upper row
+
+	movdqa	xmm7, XMMWORD [rdx+0*SIZEOF_XMMWORD]
+	movdqa	xmm3, XMMWORD [rdx+1*SIZEOF_XMMWORD]
+
+	movdqa	xmm0,xmm7		; xmm7=Int0L=( 0  1  2  3  4  5  6  7)
+	movdqa	xmm4,xmm3		; xmm3=Int0H=( 8  9 10 11 12 13 14 15)
+	psrldq	xmm0,2			; xmm0=( 1  2  3  4  5  6  7 --)
+	pslldq	xmm4,(SIZEOF_XMMWORD-2)	; xmm4=(-- -- -- -- -- -- --  8)
+	movdqa	xmm5,xmm7
+	movdqa	xmm6,xmm3
+	psrldq	xmm5,(SIZEOF_XMMWORD-2)	; xmm5=( 7 -- -- -- -- -- -- --)
+	pslldq	xmm6,2			; xmm6=(--  8  9 10 11 12 13 14)
+
+	por	xmm0,xmm4		; xmm0=( 1  2  3  4  5  6  7  8)
+	por	xmm5,xmm6		; xmm5=( 7  8  9 10 11 12 13 14)
+
+	movdqa	xmm1,xmm7
+	movdqa	xmm2,xmm3
+	pslldq	xmm1,2			; xmm1=(--  0  1  2  3  4  5  6)
+	psrldq	xmm2,2			; xmm2=( 9 10 11 12 13 14 15 --)
+	movdqa	xmm4,xmm3
+	psrldq	xmm4,(SIZEOF_XMMWORD-2)	; xmm4=(15 -- -- -- -- -- -- --)
+
+	por	xmm1, XMMWORD [wk(0)]	; xmm1=(-1  0  1  2  3  4  5  6)
+	por	xmm2, XMMWORD [wk(2)]	; xmm2=( 9 10 11 12 13 14 15 16)
+
+	movdqa	XMMWORD [wk(0)], xmm4
+
+	pmullw	xmm7,[rel PW_THREE]
+	pmullw	xmm3,[rel PW_THREE]
+	paddw	xmm1,[rel PW_EIGHT]
+	paddw	xmm5,[rel PW_EIGHT]
+	paddw	xmm0,[rel PW_SEVEN]
+	paddw	xmm2,[rel PW_SEVEN]
+
+	paddw	xmm1,xmm7
+	paddw	xmm5,xmm3
+	psrlw	xmm1,4			; xmm1=Out0LE=( 0  2  4  6  8 10 12 14)
+	psrlw	xmm5,4			; xmm5=Out0HE=(16 18 20 22 24 26 28 30)
+	paddw	xmm0,xmm7
+	paddw	xmm2,xmm3
+	psrlw	xmm0,4			; xmm0=Out0LO=( 1  3  5  7  9 11 13 15)
+	psrlw	xmm2,4			; xmm2=Out0HO=(17 19 21 23 25 27 29 31)
+
+	psllw	xmm0,BYTE_BIT
+	psllw	xmm2,BYTE_BIT
+	por	xmm1,xmm0		; xmm1=Out0L=( 0  1  2 ... 13 14 15)
+	por	xmm5,xmm2		; xmm5=Out0H=(16 17 18 ... 29 30 31)
+
+	movdqa	XMMWORD [rdx+0*SIZEOF_XMMWORD], xmm1
+	movdqa	XMMWORD [rdx+1*SIZEOF_XMMWORD], xmm5
+
+	; -- process the lower row
+
+	movdqa	xmm6, XMMWORD [rdi+0*SIZEOF_XMMWORD]
+	movdqa	xmm4, XMMWORD [rdi+1*SIZEOF_XMMWORD]
+
+	movdqa	xmm7,xmm6		; xmm6=Int1L=( 0  1  2  3  4  5  6  7)
+	movdqa	xmm3,xmm4		; xmm4=Int1H=( 8  9 10 11 12 13 14 15)
+	psrldq	xmm7,2			; xmm7=( 1  2  3  4  5  6  7 --)
+	pslldq	xmm3,(SIZEOF_XMMWORD-2)	; xmm3=(-- -- -- -- -- -- --  8)
+	movdqa	xmm0,xmm6
+	movdqa	xmm2,xmm4
+	psrldq	xmm0,(SIZEOF_XMMWORD-2)	; xmm0=( 7 -- -- -- -- -- -- --)
+	pslldq	xmm2,2			; xmm2=(--  8  9 10 11 12 13 14)
+
+	por	xmm7,xmm3		; xmm7=( 1  2  3  4  5  6  7  8)
+	por	xmm0,xmm2		; xmm0=( 7  8  9 10 11 12 13 14)
+
+	movdqa	xmm1,xmm6
+	movdqa	xmm5,xmm4
+	pslldq	xmm1,2			; xmm1=(--  0  1  2  3  4  5  6)
+	psrldq	xmm5,2			; xmm5=( 9 10 11 12 13 14 15 --)
+	movdqa	xmm3,xmm4
+	psrldq	xmm3,(SIZEOF_XMMWORD-2)	; xmm3=(15 -- -- -- -- -- -- --)
+
+	por	xmm1, XMMWORD [wk(1)]	; xmm1=(-1  0  1  2  3  4  5  6)
+	por	xmm5, XMMWORD [wk(3)]	; xmm5=( 9 10 11 12 13 14 15 16)
+
+	movdqa	XMMWORD [wk(1)], xmm3
+
+	pmullw	xmm6,[rel PW_THREE]
+	pmullw	xmm4,[rel PW_THREE]
+	paddw	xmm1,[rel PW_EIGHT]
+	paddw	xmm0,[rel PW_EIGHT]
+	paddw	xmm7,[rel PW_SEVEN]
+	paddw	xmm5,[rel PW_SEVEN]
+
+	paddw	xmm1,xmm6
+	paddw	xmm0,xmm4
+	psrlw	xmm1,4			; xmm1=Out1LE=( 0  2  4  6  8 10 12 14)
+	psrlw	xmm0,4			; xmm0=Out1HE=(16 18 20 22 24 26 28 30)
+	paddw	xmm7,xmm6
+	paddw	xmm5,xmm4
+	psrlw	xmm7,4			; xmm7=Out1LO=( 1  3  5  7  9 11 13 15)
+	psrlw	xmm5,4			; xmm5=Out1HO=(17 19 21 23 25 27 29 31)
+
+	psllw	xmm7,BYTE_BIT
+	psllw	xmm5,BYTE_BIT
+	por	xmm1,xmm7		; xmm1=Out1L=( 0  1  2 ... 13 14 15)
+	por	xmm0,xmm5		; xmm0=Out1H=(16 17 18 ... 29 30 31)
+
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm1
+	movdqa	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm0
+
+	sub	rax, byte SIZEOF_XMMWORD
+	add	rcx, byte 1*SIZEOF_XMMWORD	; inptr1(above)
+	add	rbx, byte 1*SIZEOF_XMMWORD	; inptr0
+	add	rsi, byte 1*SIZEOF_XMMWORD	; inptr1(below)
+	add	rdx, byte 2*SIZEOF_XMMWORD	; outptr0
+	add	rdi, byte 2*SIZEOF_XMMWORD	; outptr1
+	cmp	rax, byte SIZEOF_XMMWORD
+	ja	near .columnloop
+	test	rax,rax
+	jnz	near .columnloop_last
+
+	pop	rsi
+	pop	rdi
+	pop	rcx
+	pop	rax
+
+	add	rsi, byte 1*SIZEOF_JSAMPROW	; input_data
+	add	rdi, byte 2*SIZEOF_JSAMPROW	; output_data
+	sub	rcx, byte 2			; rowctr
+	jg	near .rowloop
+
+.return:
+	pop	rbx
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v1_upsample_sse2 (int max_v_samp_factor,
+;                           JDIMENSION output_width,
+;                           JSAMPARRAY input_data,
+;                           JSAMPARRAY * output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11 = JDIMENSION output_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_upsample_sse2)
+
+EXTN(jsimd_h2v1_upsample_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+
+	mov	rdx, r11
+	add	rdx, byte (2*SIZEOF_XMMWORD)-1
+	and	rdx, byte -(2*SIZEOF_XMMWORD)
+	jz	near .return
+
+	mov	rcx, r10	; rowctr
+	test	rcx,rcx
+	jz	short .return
+
+	mov	rsi, r12 ; input_data
+	mov	rdi, r13
+	mov	rdi, JSAMPARRAY [rdi]			; output_data
+.rowloop:
+	push	rdi
+	push	rsi
+
+	mov	rsi, JSAMPROW [rsi]		; inptr
+	mov	rdi, JSAMPROW [rdi]		; outptr
+	mov	rax,rdx				; colctr
+.columnloop:
+
+	movdqa	xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+
+	movdqa    xmm1,xmm0
+	punpcklbw xmm0,xmm0
+	punpckhbw xmm1,xmm1
+
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+	movdqa	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
+
+	sub	rax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	movdqa	xmm2, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+	movdqa    xmm3,xmm2
+	punpcklbw xmm2,xmm2
+	punpckhbw xmm3,xmm3
+
+	movdqa	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [rdi+3*SIZEOF_XMMWORD], xmm3
+
+	sub	rax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	add	rsi, byte 2*SIZEOF_XMMWORD	; inptr
+	add	rdi, byte 4*SIZEOF_XMMWORD	; outptr
+	jmp	short .columnloop
+
+.nextrow:
+	pop	rsi
+	pop	rdi
+
+	add	rsi, byte SIZEOF_JSAMPROW	; input_data
+	add	rdi, byte SIZEOF_JSAMPROW	; output_data
+	dec	rcx				; rowctr
+	jg	short .rowloop
+
+.return:
+	uncollect_args
+	pop	rbp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v2_upsample_sse2 (nt max_v_samp_factor,
+;                           JDIMENSION output_width,
+;                           JSAMPARRAY input_data,
+;                           JSAMPARRAY * output_data_ptr);
+;
+
+; r10 = int max_v_samp_factor
+; r11 = JDIMENSION output_width
+; r12 = JSAMPARRAY input_data
+; r13 = JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v2_upsample_sse2)
+
+EXTN(jsimd_h2v2_upsample_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+	push	rbx
+
+	mov	rdx, r11
+	add	rdx, byte (2*SIZEOF_XMMWORD)-1
+	and	rdx, byte -(2*SIZEOF_XMMWORD)
+	jz	near .return
+
+	mov	rcx, r10	; rowctr
+	test	rcx,rcx
+	jz	near .return
+
+	mov	rsi, r12	; input_data
+	mov	rdi, r13
+	mov	rdi, JSAMPARRAY [rdi]			; output_data
+.rowloop:
+	push	rdi
+	push	rsi
+
+	mov	rsi, JSAMPROW [rsi]			; inptr
+	mov	rbx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]	; outptr0
+	mov	rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]	; outptr1
+	mov	rax,rdx					; colctr
+.columnloop:
+
+	movdqa	xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
+
+	movdqa    xmm1,xmm0
+	punpcklbw xmm0,xmm0
+	punpckhbw xmm1,xmm1
+
+	movdqa	XMMWORD [rbx+0*SIZEOF_XMMWORD], xmm0
+	movdqa	XMMWORD [rbx+1*SIZEOF_XMMWORD], xmm1
+	movdqa	XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
+	movdqa	XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
+
+	sub	rax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	movdqa	xmm2, XMMWORD [rsi+1*SIZEOF_XMMWORD]
+
+	movdqa    xmm3,xmm2
+	punpcklbw xmm2,xmm2
+	punpckhbw xmm3,xmm3
+
+	movdqa	XMMWORD [rbx+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [rbx+3*SIZEOF_XMMWORD], xmm3
+	movdqa	XMMWORD [rdi+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [rdi+3*SIZEOF_XMMWORD], xmm3
+
+	sub	rax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	add	rsi, byte 2*SIZEOF_XMMWORD	; inptr
+	add	rbx, byte 4*SIZEOF_XMMWORD	; outptr0
+	add	rdi, byte 4*SIZEOF_XMMWORD	; outptr1
+	jmp	short .columnloop
+
+.nextrow:
+	pop	rsi
+	pop	rdi
+
+	add	rsi, byte 1*SIZEOF_JSAMPROW	; input_data
+	add	rdi, byte 2*SIZEOF_JSAMPROW	; output_data
+	sub	rcx, byte 2			; rowctr
+	jg	near .rowloop
+
+.return:
+	pop	rbx
+	uncollect_args
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jdsamss2.asm b/simd/jdsamss2.asm
new file mode 100644
index 0000000..b5c863b
--- /dev/null
+++ b/simd/jdsamss2.asm
@@ -0,0 +1,729 @@
+;
+; jdsamss2.asm - upsampling (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fancy_upsample_sse2)
+
+EXTN(jconst_fancy_upsample_sse2):
+
+PW_ONE		times 8 dw  1
+PW_TWO		times 8 dw  2
+PW_THREE	times 8 dw  3
+PW_SEVEN	times 8 dw  7
+PW_EIGHT	times 8 dw  8
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+;
+; The upsampling algorithm is linear interpolation between pixel centers,
+; also known as a "triangle filter".  This is a good compromise between
+; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+; of the way between input pixel centers.
+;
+; GLOBAL(void)
+; jsimd_h2v1_fancy_upsample_sse2 (int max_v_samp_factor,
+;                                 JDIMENSION downsampled_width,
+;                                 JSAMPARRAY input_data,
+;                                 JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define downsamp_width(b)	(b)+12	; JDIMENSION downsampled_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_fancy_upsample_sse2)
+
+EXTN(jsimd_h2v1_fancy_upsample_sse2):
+	push	ebp
+	mov	ebp,esp
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	mov	eax, JDIMENSION [downsamp_width(ebp)]  ; colctr
+	test	eax,eax
+	jz	near .return
+
+	mov	ecx, INT [max_v_samp(ebp)]	; rowctr
+	test	ecx,ecx
+	jz	near .return
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, POINTER [output_data_ptr(ebp)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	eax			; colctr
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]	; inptr
+	mov	edi, JSAMPROW [edi]	; outptr
+
+	test	eax, SIZEOF_XMMWORD-1
+	jz	short .skip
+	mov	dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl	; insert a dummy sample
+.skip:
+	pxor	xmm0,xmm0		; xmm0=(all 0's)
+	pcmpeqb	xmm7,xmm7
+	psrldq	xmm7,(SIZEOF_XMMWORD-1)
+	pand	xmm7, XMMWORD [esi+0*SIZEOF_XMMWORD]
+
+	add	eax, byte SIZEOF_XMMWORD-1
+	and	eax, byte -SIZEOF_XMMWORD
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	short .columnloop
+	alignx	16,7
+
+.columnloop_last:
+	pcmpeqb	xmm6,xmm6
+	pslldq	xmm6,(SIZEOF_XMMWORD-1)
+	pand	xmm6, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	jmp	short .upsample
+	alignx	16,7
+
+.columnloop:
+	movdqa	xmm6, XMMWORD [esi+1*SIZEOF_XMMWORD]
+	pslldq	xmm6,(SIZEOF_XMMWORD-1)
+
+.upsample:
+	movdqa	xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
+	movdqa	xmm2,xmm1
+	movdqa	xmm3,xmm1		; xmm1=( 0  1  2 ... 13 14 15)
+	pslldq	xmm2,1			; xmm2=(--  0  1 ... 12 13 14)
+	psrldq	xmm3,1			; xmm3=( 1  2  3 ... 14 15 --)
+
+	por	xmm2,xmm7		; xmm2=(-1  0  1 ... 12 13 14)
+	por	xmm3,xmm6		; xmm3=( 1  2  3 ... 14 15 16)
+
+	movdqa	xmm7,xmm1
+	psrldq	xmm7,(SIZEOF_XMMWORD-1)	; xmm7=(15 -- -- ... -- -- --)
+
+	movdqa    xmm4,xmm1
+	punpcklbw xmm1,xmm0		; xmm1=( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm4,xmm0		; xmm4=( 8  9 10 11 12 13 14 15)
+	movdqa    xmm5,xmm2
+	punpcklbw xmm2,xmm0		; xmm2=(-1  0  1  2  3  4  5  6)
+	punpckhbw xmm5,xmm0		; xmm5=( 7  8  9 10 11 12 13 14)
+	movdqa    xmm6,xmm3
+	punpcklbw xmm3,xmm0		; xmm3=( 1  2  3  4  5  6  7  8)
+	punpckhbw xmm6,xmm0		; xmm6=( 9 10 11 12 13 14 15 16)
+
+	pmullw	xmm1,[GOTOFF(ebx,PW_THREE)]
+	pmullw	xmm4,[GOTOFF(ebx,PW_THREE)]
+	paddw	xmm2,[GOTOFF(ebx,PW_ONE)]
+	paddw	xmm5,[GOTOFF(ebx,PW_ONE)]
+	paddw	xmm3,[GOTOFF(ebx,PW_TWO)]
+	paddw	xmm6,[GOTOFF(ebx,PW_TWO)]
+
+	paddw	xmm2,xmm1
+	paddw	xmm5,xmm4
+	psrlw	xmm2,2			; xmm2=OutLE=( 0  2  4  6  8 10 12 14)
+	psrlw	xmm5,2			; xmm5=OutHE=(16 18 20 22 24 26 28 30)
+	paddw	xmm3,xmm1
+	paddw	xmm6,xmm4
+	psrlw	xmm3,2			; xmm3=OutLO=( 1  3  5  7  9 11 13 15)
+	psrlw	xmm6,2			; xmm6=OutHO=(17 19 21 23 25 27 29 31)
+
+	psllw	xmm3,BYTE_BIT
+	psllw	xmm6,BYTE_BIT
+	por	xmm2,xmm3		; xmm2=OutL=( 0  1  2 ... 13 14 15)
+	por	xmm5,xmm6		; xmm5=OutH=(16 17 18 ... 29 30 31)
+
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [edi+1*SIZEOF_XMMWORD], xmm5
+
+	sub	eax, byte SIZEOF_XMMWORD
+	add	esi, byte 1*SIZEOF_XMMWORD	; inptr
+	add	edi, byte 2*SIZEOF_XMMWORD	; outptr
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	near .columnloop
+	test	eax,eax
+	jnz	near .columnloop_last
+
+	pop	esi
+	pop	edi
+	pop	eax
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_data
+	add	edi, byte SIZEOF_JSAMPROW	; output_data
+	dec	ecx				; rowctr
+	jg	near .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+; Again a triangle filter; see comments for h2v1 case, above.
+;
+; GLOBAL(void)
+; jsimd_h2v2_fancy_upsample_sse2 (int max_v_samp_factor,
+;                                 JDIMENSION downsampled_width,
+;                                 JSAMPARRAY input_data,
+;                                 JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define downsamp_width(b)	(b)+12	; JDIMENSION downsampled_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		4
+%define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
+
+	align	16
+	global	EXTN(jsimd_h2v2_fancy_upsample_sse2)
+
+EXTN(jsimd_h2v2_fancy_upsample_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	eax		; make a room for GOT address
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx			; get GOT address
+	movpic	POINTER [gotptr], ebx	; save GOT address
+
+	mov	edx,eax				; edx = original ebp
+	mov	eax, JDIMENSION [downsamp_width(edx)]  ; colctr
+	test	eax,eax
+	jz	near .return
+
+	mov	ecx, INT [max_v_samp(edx)]	; rowctr
+	test	ecx,ecx
+	jz	near .return
+
+	mov	esi, JSAMPARRAY [input_data(edx)]	; input_data
+	mov	edi, POINTER [output_data_ptr(edx)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	eax					; colctr
+	push	ecx
+	push	edi
+	push	esi
+
+	mov	ecx, JSAMPROW [esi-1*SIZEOF_JSAMPROW]	; inptr1(above)
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; inptr0
+	mov	esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; inptr1(below)
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]	; outptr0
+	mov	edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]	; outptr1
+
+	test	eax, SIZEOF_XMMWORD-1
+	jz	short .skip
+	push	edx
+	mov	dl, JSAMPLE [ecx+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [ecx+eax*SIZEOF_JSAMPLE], dl
+	mov	dl, JSAMPLE [ebx+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [ebx+eax*SIZEOF_JSAMPLE], dl
+	mov	dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
+	mov	JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl	; insert a dummy sample
+	pop	edx
+.skip:
+	; -- process the first column block
+
+	movdqa	xmm0, XMMWORD [ebx+0*SIZEOF_XMMWORD]	; xmm0=row[ 0][0]
+	movdqa	xmm1, XMMWORD [ecx+0*SIZEOF_XMMWORD]	; xmm1=row[-1][0]
+	movdqa	xmm2, XMMWORD [esi+0*SIZEOF_XMMWORD]	; xmm2=row[+1][0]
+
+	pushpic	ebx
+	movpic	ebx, POINTER [gotptr]	; load GOT address
+
+	pxor      xmm3,xmm3		; xmm3=(all 0's)
+	movdqa    xmm4,xmm0
+	punpcklbw xmm0,xmm3		; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm4,xmm3		; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm5,xmm1
+	punpcklbw xmm1,xmm3		; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm5,xmm3		; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm6,xmm2
+	punpcklbw xmm2,xmm3		; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm6,xmm3		; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
+
+	pmullw	xmm0,[GOTOFF(ebx,PW_THREE)]
+	pmullw	xmm4,[GOTOFF(ebx,PW_THREE)]
+
+	pcmpeqb	xmm7,xmm7
+	psrldq	xmm7,(SIZEOF_XMMWORD-2)
+
+	paddw	xmm1,xmm0		; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm5,xmm4		; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
+	paddw	xmm2,xmm0		; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm6,xmm4		; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
+
+	movdqa	XMMWORD [edx+0*SIZEOF_XMMWORD], xmm1	; temporarily save
+	movdqa	XMMWORD [edx+1*SIZEOF_XMMWORD], xmm5	; the intermediate data
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [edi+1*SIZEOF_XMMWORD], xmm6
+
+	pand	xmm1,xmm7		; xmm1=( 0 -- -- -- -- -- -- --)
+	pand	xmm2,xmm7		; xmm2=( 0 -- -- -- -- -- -- --)
+
+	movdqa	XMMWORD [wk(0)], xmm1
+	movdqa	XMMWORD [wk(1)], xmm2
+
+	poppic	ebx
+
+	add	eax, byte SIZEOF_XMMWORD-1
+	and	eax, byte -SIZEOF_XMMWORD
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	short .columnloop
+	alignx	16,7
+
+.columnloop_last:
+	; -- process the last column block
+
+	pushpic	ebx
+	movpic	ebx, POINTER [gotptr]	; load GOT address
+
+	pcmpeqb	xmm1,xmm1
+	pslldq	xmm1,(SIZEOF_XMMWORD-2)
+	movdqa	xmm2,xmm1
+
+	pand	xmm1, XMMWORD [edx+1*SIZEOF_XMMWORD]
+	pand	xmm2, XMMWORD [edi+1*SIZEOF_XMMWORD]
+
+	movdqa	XMMWORD [wk(2)], xmm1	; xmm1=(-- -- -- -- -- -- -- 15)
+	movdqa	XMMWORD [wk(3)], xmm2	; xmm2=(-- -- -- -- -- -- -- 15)
+
+	jmp	near .upsample
+	alignx	16,7
+
+.columnloop:
+	; -- process the next column block
+
+	movdqa	xmm0, XMMWORD [ebx+1*SIZEOF_XMMWORD]	; xmm0=row[ 0][1]
+	movdqa	xmm1, XMMWORD [ecx+1*SIZEOF_XMMWORD]	; xmm1=row[-1][1]
+	movdqa	xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]	; xmm2=row[+1][1]
+
+	pushpic	ebx
+	movpic	ebx, POINTER [gotptr]	; load GOT address
+
+	pxor      xmm3,xmm3		; xmm3=(all 0's)
+	movdqa    xmm4,xmm0
+	punpcklbw xmm0,xmm3		; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm4,xmm3		; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm5,xmm1
+	punpcklbw xmm1,xmm3		; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm5,xmm3		; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
+	movdqa    xmm6,xmm2
+	punpcklbw xmm2,xmm3		; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
+	punpckhbw xmm6,xmm3		; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
+
+	pmullw	xmm0,[GOTOFF(ebx,PW_THREE)]
+	pmullw	xmm4,[GOTOFF(ebx,PW_THREE)]
+
+	paddw	xmm1,xmm0		; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm5,xmm4		; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
+	paddw	xmm2,xmm0		; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
+	paddw	xmm6,xmm4		; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
+
+	movdqa	XMMWORD [edx+2*SIZEOF_XMMWORD], xmm1	; temporarily save
+	movdqa	XMMWORD [edx+3*SIZEOF_XMMWORD], xmm5	; the intermediate data
+	movdqa	XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [edi+3*SIZEOF_XMMWORD], xmm6
+
+	pslldq	xmm1,(SIZEOF_XMMWORD-2)	; xmm1=(-- -- -- -- -- -- --  0)
+	pslldq	xmm2,(SIZEOF_XMMWORD-2)	; xmm2=(-- -- -- -- -- -- --  0)
+
+	movdqa	XMMWORD [wk(2)], xmm1
+	movdqa	XMMWORD [wk(3)], xmm2
+
+.upsample:
+	; -- process the upper row
+
+	movdqa	xmm7, XMMWORD [edx+0*SIZEOF_XMMWORD]
+	movdqa	xmm3, XMMWORD [edx+1*SIZEOF_XMMWORD]
+
+	movdqa	xmm0,xmm7		; xmm7=Int0L=( 0  1  2  3  4  5  6  7)
+	movdqa	xmm4,xmm3		; xmm3=Int0H=( 8  9 10 11 12 13 14 15)
+	psrldq	xmm0,2			; xmm0=( 1  2  3  4  5  6  7 --)
+	pslldq	xmm4,(SIZEOF_XMMWORD-2)	; xmm4=(-- -- -- -- -- -- --  8)
+	movdqa	xmm5,xmm7
+	movdqa	xmm6,xmm3
+	psrldq	xmm5,(SIZEOF_XMMWORD-2)	; xmm5=( 7 -- -- -- -- -- -- --)
+	pslldq	xmm6,2			; xmm6=(--  8  9 10 11 12 13 14)
+
+	por	xmm0,xmm4		; xmm0=( 1  2  3  4  5  6  7  8)
+	por	xmm5,xmm6		; xmm5=( 7  8  9 10 11 12 13 14)
+
+	movdqa	xmm1,xmm7
+	movdqa	xmm2,xmm3
+	pslldq	xmm1,2			; xmm1=(--  0  1  2  3  4  5  6)
+	psrldq	xmm2,2			; xmm2=( 9 10 11 12 13 14 15 --)
+	movdqa	xmm4,xmm3
+	psrldq	xmm4,(SIZEOF_XMMWORD-2)	; xmm4=(15 -- -- -- -- -- -- --)
+
+	por	xmm1, XMMWORD [wk(0)]	; xmm1=(-1  0  1  2  3  4  5  6)
+	por	xmm2, XMMWORD [wk(2)]	; xmm2=( 9 10 11 12 13 14 15 16)
+
+	movdqa	XMMWORD [wk(0)], xmm4
+
+	pmullw	xmm7,[GOTOFF(ebx,PW_THREE)]
+	pmullw	xmm3,[GOTOFF(ebx,PW_THREE)]
+	paddw	xmm1,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	xmm5,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	xmm0,[GOTOFF(ebx,PW_SEVEN)]
+	paddw	xmm2,[GOTOFF(ebx,PW_SEVEN)]
+
+	paddw	xmm1,xmm7
+	paddw	xmm5,xmm3
+	psrlw	xmm1,4			; xmm1=Out0LE=( 0  2  4  6  8 10 12 14)
+	psrlw	xmm5,4			; xmm5=Out0HE=(16 18 20 22 24 26 28 30)
+	paddw	xmm0,xmm7
+	paddw	xmm2,xmm3
+	psrlw	xmm0,4			; xmm0=Out0LO=( 1  3  5  7  9 11 13 15)
+	psrlw	xmm2,4			; xmm2=Out0HO=(17 19 21 23 25 27 29 31)
+
+	psllw	xmm0,BYTE_BIT
+	psllw	xmm2,BYTE_BIT
+	por	xmm1,xmm0		; xmm1=Out0L=( 0  1  2 ... 13 14 15)
+	por	xmm5,xmm2		; xmm5=Out0H=(16 17 18 ... 29 30 31)
+
+	movdqa	XMMWORD [edx+0*SIZEOF_XMMWORD], xmm1
+	movdqa	XMMWORD [edx+1*SIZEOF_XMMWORD], xmm5
+
+	; -- process the lower row
+
+	movdqa	xmm6, XMMWORD [edi+0*SIZEOF_XMMWORD]
+	movdqa	xmm4, XMMWORD [edi+1*SIZEOF_XMMWORD]
+
+	movdqa	xmm7,xmm6		; xmm6=Int1L=( 0  1  2  3  4  5  6  7)
+	movdqa	xmm3,xmm4		; xmm4=Int1H=( 8  9 10 11 12 13 14 15)
+	psrldq	xmm7,2			; xmm7=( 1  2  3  4  5  6  7 --)
+	pslldq	xmm3,(SIZEOF_XMMWORD-2)	; xmm3=(-- -- -- -- -- -- --  8)
+	movdqa	xmm0,xmm6
+	movdqa	xmm2,xmm4
+	psrldq	xmm0,(SIZEOF_XMMWORD-2)	; xmm0=( 7 -- -- -- -- -- -- --)
+	pslldq	xmm2,2			; xmm2=(--  8  9 10 11 12 13 14)
+
+	por	xmm7,xmm3		; xmm7=( 1  2  3  4  5  6  7  8)
+	por	xmm0,xmm2		; xmm0=( 7  8  9 10 11 12 13 14)
+
+	movdqa	xmm1,xmm6
+	movdqa	xmm5,xmm4
+	pslldq	xmm1,2			; xmm1=(--  0  1  2  3  4  5  6)
+	psrldq	xmm5,2			; xmm5=( 9 10 11 12 13 14 15 --)
+	movdqa	xmm3,xmm4
+	psrldq	xmm3,(SIZEOF_XMMWORD-2)	; xmm3=(15 -- -- -- -- -- -- --)
+
+	por	xmm1, XMMWORD [wk(1)]	; xmm1=(-1  0  1  2  3  4  5  6)
+	por	xmm5, XMMWORD [wk(3)]	; xmm5=( 9 10 11 12 13 14 15 16)
+
+	movdqa	XMMWORD [wk(1)], xmm3
+
+	pmullw	xmm6,[GOTOFF(ebx,PW_THREE)]
+	pmullw	xmm4,[GOTOFF(ebx,PW_THREE)]
+	paddw	xmm1,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	xmm0,[GOTOFF(ebx,PW_EIGHT)]
+	paddw	xmm7,[GOTOFF(ebx,PW_SEVEN)]
+	paddw	xmm5,[GOTOFF(ebx,PW_SEVEN)]
+
+	paddw	xmm1,xmm6
+	paddw	xmm0,xmm4
+	psrlw	xmm1,4			; xmm1=Out1LE=( 0  2  4  6  8 10 12 14)
+	psrlw	xmm0,4			; xmm0=Out1HE=(16 18 20 22 24 26 28 30)
+	paddw	xmm7,xmm6
+	paddw	xmm5,xmm4
+	psrlw	xmm7,4			; xmm7=Out1LO=( 1  3  5  7  9 11 13 15)
+	psrlw	xmm5,4			; xmm5=Out1HO=(17 19 21 23 25 27 29 31)
+
+	psllw	xmm7,BYTE_BIT
+	psllw	xmm5,BYTE_BIT
+	por	xmm1,xmm7		; xmm1=Out1L=( 0  1  2 ... 13 14 15)
+	por	xmm0,xmm5		; xmm0=Out1H=(16 17 18 ... 29 30 31)
+
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm1
+	movdqa	XMMWORD [edi+1*SIZEOF_XMMWORD], xmm0
+
+	poppic	ebx
+
+	sub	eax, byte SIZEOF_XMMWORD
+	add	ecx, byte 1*SIZEOF_XMMWORD	; inptr1(above)
+	add	ebx, byte 1*SIZEOF_XMMWORD	; inptr0
+	add	esi, byte 1*SIZEOF_XMMWORD	; inptr1(below)
+	add	edx, byte 2*SIZEOF_XMMWORD	; outptr0
+	add	edi, byte 2*SIZEOF_XMMWORD	; outptr1
+	cmp	eax, byte SIZEOF_XMMWORD
+	ja	near .columnloop
+	test	eax,eax
+	jnz	near .columnloop_last
+
+	pop	esi
+	pop	edi
+	pop	ecx
+	pop	eax
+
+	add	esi, byte 1*SIZEOF_JSAMPROW	; input_data
+	add	edi, byte 2*SIZEOF_JSAMPROW	; output_data
+	sub	ecx, byte 2			; rowctr
+	jg	near .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v1_upsample_sse2 (int max_v_samp_factor,
+;                           JDIMENSION output_width,
+;                           JSAMPARRAY input_data,
+;                           JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define output_width(b)	(b)+12		; JDIMENSION output_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v1_upsample_sse2)
+
+EXTN(jsimd_h2v1_upsample_sse2):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	edx, JDIMENSION [output_width(ebp)]
+	add	edx, byte (2*SIZEOF_XMMWORD)-1
+	and	edx, byte -(2*SIZEOF_XMMWORD)
+	jz	short .return
+
+	mov	ecx, INT [max_v_samp(ebp)]	; rowctr
+	test	ecx,ecx
+	jz	short .return
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, POINTER [output_data_ptr(ebp)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]		; inptr
+	mov	edi, JSAMPROW [edi]		; outptr
+	mov	eax,edx				; colctr
+	alignx	16,7
+.columnloop:
+
+	movdqa	xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
+
+	movdqa    xmm1,xmm0
+	punpcklbw xmm0,xmm0
+	punpckhbw xmm1,xmm1
+
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
+	movdqa	XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1
+
+	sub	eax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	movdqa	xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]
+
+	movdqa    xmm3,xmm2
+	punpcklbw xmm2,xmm2
+	punpckhbw xmm3,xmm3
+
+	movdqa	XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [edi+3*SIZEOF_XMMWORD], xmm3
+
+	sub	eax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	add	esi, byte 2*SIZEOF_XMMWORD	; inptr
+	add	edi, byte 4*SIZEOF_XMMWORD	; outptr
+	jmp	short .columnloop
+	alignx	16,7
+
+.nextrow:
+	pop	esi
+	pop	edi
+
+	add	esi, byte SIZEOF_JSAMPROW	; input_data
+	add	edi, byte SIZEOF_JSAMPROW	; output_data
+	dec	ecx				; rowctr
+	jg	short .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
+; --------------------------------------------------------------------------
+;
+; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+; It's still a box filter.
+;
+; GLOBAL(void)
+; jsimd_h2v2_upsample_sse2 (nt max_v_samp_factor,
+;                           JDIMENSION output_width,
+;                           JSAMPARRAY input_data,
+;                           JSAMPARRAY * output_data_ptr);
+;
+
+%define max_v_samp(b)		(b)+8			; int max_v_samp_factor
+%define output_width(b)	(b)+12		; JDIMENSION output_width
+%define input_data(b)		(b)+16		; JSAMPARRAY input_data
+%define output_data_ptr(b)	(b)+20		; JSAMPARRAY * output_data_ptr
+
+	align	16
+	global	EXTN(jsimd_h2v2_upsample_sse2)
+
+EXTN(jsimd_h2v2_upsample_sse2):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	edx, JDIMENSION [output_width(ebp)]
+	add	edx, byte (2*SIZEOF_XMMWORD)-1
+	and	edx, byte -(2*SIZEOF_XMMWORD)
+	jz	near .return
+
+	mov	ecx, INT [max_v_samp(ebp)]	; rowctr
+	test	ecx,ecx
+	jz	near .return
+
+	mov	esi, JSAMPARRAY [input_data(ebp)]	; input_data
+	mov	edi, POINTER [output_data_ptr(ebp)]
+	mov	edi, JSAMPARRAY [edi]			; output_data
+	alignx	16,7
+.rowloop:
+	push	edi
+	push	esi
+
+	mov	esi, JSAMPROW [esi]			; inptr
+	mov	ebx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]	; outptr0
+	mov	edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]	; outptr1
+	mov	eax,edx					; colctr
+	alignx	16,7
+.columnloop:
+
+	movdqa	xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
+
+	movdqa    xmm1,xmm0
+	punpcklbw xmm0,xmm0
+	punpckhbw xmm1,xmm1
+
+	movdqa	XMMWORD [ebx+0*SIZEOF_XMMWORD], xmm0
+	movdqa	XMMWORD [ebx+1*SIZEOF_XMMWORD], xmm1
+	movdqa	XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
+	movdqa	XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1
+
+	sub	eax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	movdqa	xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]
+
+	movdqa    xmm3,xmm2
+	punpcklbw xmm2,xmm2
+	punpckhbw xmm3,xmm3
+
+	movdqa	XMMWORD [ebx+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [ebx+3*SIZEOF_XMMWORD], xmm3
+	movdqa	XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
+	movdqa	XMMWORD [edi+3*SIZEOF_XMMWORD], xmm3
+
+	sub	eax, byte 2*SIZEOF_XMMWORD
+	jz	short .nextrow
+
+	add	esi, byte 2*SIZEOF_XMMWORD	; inptr
+	add	ebx, byte 4*SIZEOF_XMMWORD	; outptr0
+	add	edi, byte 4*SIZEOF_XMMWORD	; outptr1
+	jmp	short .columnloop
+	alignx	16,7
+
+.nextrow:
+	pop	esi
+	pop	edi
+
+	add	esi, byte 1*SIZEOF_JSAMPROW	; input_data
+	add	edi, byte 2*SIZEOF_JSAMPROW	; output_data
+	sub	ecx, byte 2			; rowctr
+	jg	short .rowloop
+
+.return:
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jf3dnflt.asm b/simd/jf3dnflt.asm
new file mode 100644
index 0000000..542672d
--- /dev/null
+++ b/simd/jf3dnflt.asm
@@ -0,0 +1,320 @@
+;
+; jf3dnflt.asm - floating-point FDCT (3DNow!)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the forward DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_float_3dnow)
+
+EXTN(jconst_fdct_float_3dnow):
+
+PD_0_382	times 2 dd  0.382683432365089771728460
+PD_0_707	times 2 dd  0.707106781186547524400844
+PD_0_541	times 2 dd  0.541196100146196984399723
+PD_1_306	times 2 dd  1.306562964876376527856643
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_float_3dnow (FAST_FLOAT * data)
+;
+
+%define data(b)		(b)+8		; FAST_FLOAT * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_float_3dnow)
+
+EXTN(jsimd_fdct_float_3dnow):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
+	mov	ecx, DCTSIZE/2
+	alignx	16,7
+.rowloop:
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm2, MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)]
+
+	; mm0=(00 01), mm1=(10 11), mm2=(06 07), mm3=(16 17)
+
+	movq      mm4,mm0		; transpose coefficients
+	punpckldq mm0,mm1		; mm0=(00 10)=data0
+	punpckhdq mm4,mm1		; mm4=(01 11)=data1
+	movq      mm5,mm2		; transpose coefficients
+	punpckldq mm2,mm3		; mm2=(06 16)=data6
+	punpckhdq mm5,mm3		; mm5=(07 17)=data7
+
+	movq	mm6,mm4
+	movq	mm7,mm0
+	pfsub	mm4,mm2			; mm4=data1-data6=tmp6
+	pfsub	mm0,mm5			; mm0=data0-data7=tmp7
+	pfadd	mm6,mm2			; mm6=data1+data6=tmp1
+	pfadd	mm7,mm5			; mm7=data0+data7=tmp0
+
+	movq	mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm2, MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm5, MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)]
+
+	; mm1=(02 03), mm3=(12 13), mm2=(04 05), mm5=(14 15)
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=tmp6
+	movq	MMWORD [wk(1)], mm0	; wk(1)=tmp7
+
+	movq      mm4,mm1		; transpose coefficients
+	punpckldq mm1,mm3		; mm1=(02 12)=data2
+	punpckhdq mm4,mm3		; mm4=(03 13)=data3
+	movq      mm0,mm2		; transpose coefficients
+	punpckldq mm2,mm5		; mm2=(04 14)=data4
+	punpckhdq mm0,mm5		; mm0=(05 15)=data5
+
+	movq	mm3,mm4
+	movq	mm5,mm1
+	pfadd	mm4,mm2			; mm4=data3+data4=tmp3
+	pfadd	mm1,mm0			; mm1=data2+data5=tmp2
+	pfsub	mm3,mm2			; mm3=data3-data4=tmp4
+	pfsub	mm5,mm0			; mm5=data2-data5=tmp5
+
+	; -- Even part
+
+	movq	mm2,mm7
+	movq	mm0,mm6
+	pfsub	mm7,mm4			; mm7=tmp13
+	pfsub	mm6,mm1			; mm6=tmp12
+	pfadd	mm2,mm4			; mm2=tmp10
+	pfadd	mm0,mm1			; mm0=tmp11
+
+	pfadd	mm6,mm7
+	pfmul	mm6,[GOTOFF(ebx,PD_0_707)] ; mm6=z1
+
+	movq	mm4,mm2
+	movq	mm1,mm7
+	pfsub	mm2,mm0			; mm2=data4
+	pfsub	mm7,mm6			; mm7=data6
+	pfadd	mm4,mm0			; mm4=data0
+	pfadd	mm1,mm6			; mm1=data2
+
+	movq	MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)], mm2
+	movq	MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)], mm7
+	movq	MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4
+	movq	MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], mm1
+
+	; -- Odd part
+
+	movq	mm0, MMWORD [wk(0)]	; mm0=tmp6
+	movq	mm6, MMWORD [wk(1)]	; mm6=tmp7
+
+	pfadd	mm3,mm5			; mm3=tmp10
+	pfadd	mm5,mm0			; mm5=tmp11
+	pfadd	mm0,mm6			; mm0=tmp12, mm6=tmp7
+
+	pfmul	mm5,[GOTOFF(ebx,PD_0_707)] ; mm5=z3
+
+	movq	mm2,mm3			; mm2=tmp10
+	pfsub	mm3,mm0
+	pfmul	mm3,[GOTOFF(ebx,PD_0_382)] ; mm3=z5
+	pfmul	mm2,[GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610)
+	pfmul	mm0,[GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296)
+	pfadd	mm2,mm3			; mm2=z2
+	pfadd	mm0,mm3			; mm0=z4
+
+	movq	mm7,mm6
+	pfsub	mm6,mm5			; mm6=z13
+	pfadd	mm7,mm5			; mm7=z11
+
+	movq	mm4,mm6
+	movq	mm1,mm7
+	pfsub	mm6,mm2			; mm6=data3
+	pfsub	mm7,mm0			; mm7=data7
+	pfadd	mm4,mm2			; mm4=data5
+	pfadd	mm1,mm0			; mm1=data1
+
+	movq	MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], mm6
+	movq	MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)], mm7
+	movq	MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)], mm4
+	movq	MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1
+
+	add	edx, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .rowloop
+
+	; ---- Pass 2: process columns.
+
+	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
+	mov	ecx, DCTSIZE/2
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
+
+	; mm0=(00 10), mm1=(01 11), mm2=(60 70), mm3=(61 71)
+
+	movq      mm4,mm0		; transpose coefficients
+	punpckldq mm0,mm1		; mm0=(00 01)=data0
+	punpckhdq mm4,mm1		; mm4=(10 11)=data1
+	movq      mm5,mm2		; transpose coefficients
+	punpckldq mm2,mm3		; mm2=(60 61)=data6
+	punpckhdq mm5,mm3		; mm5=(70 71)=data7
+
+	movq	mm6,mm4
+	movq	mm7,mm0
+	pfsub	mm4,mm2			; mm4=data1-data6=tmp6
+	pfsub	mm0,mm5			; mm0=data0-data7=tmp7
+	pfadd	mm6,mm2			; mm6=data1+data6=tmp1
+	pfadd	mm7,mm5			; mm7=data0+data7=tmp0
+
+	movq	mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
+	movq	mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
+
+	; mm1=(20 30), mm3=(21 31), mm2=(40 50), mm5=(41 51)
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=tmp6
+	movq	MMWORD [wk(1)], mm0	; wk(1)=tmp7
+
+	movq      mm4,mm1		; transpose coefficients
+	punpckldq mm1,mm3		; mm1=(20 21)=data2
+	punpckhdq mm4,mm3		; mm4=(30 31)=data3
+	movq      mm0,mm2		; transpose coefficients
+	punpckldq mm2,mm5		; mm2=(40 41)=data4
+	punpckhdq mm0,mm5		; mm0=(50 51)=data5
+
+	movq	mm3,mm4
+	movq	mm5,mm1
+	pfadd	mm4,mm2			; mm4=data3+data4=tmp3
+	pfadd	mm1,mm0			; mm1=data2+data5=tmp2
+	pfsub	mm3,mm2			; mm3=data3-data4=tmp4
+	pfsub	mm5,mm0			; mm5=data2-data5=tmp5
+
+	; -- Even part
+
+	movq	mm2,mm7
+	movq	mm0,mm6
+	pfsub	mm7,mm4			; mm7=tmp13
+	pfsub	mm6,mm1			; mm6=tmp12
+	pfadd	mm2,mm4			; mm2=tmp10
+	pfadd	mm0,mm1			; mm0=tmp11
+
+	pfadd	mm6,mm7
+	pfmul	mm6,[GOTOFF(ebx,PD_0_707)] ; mm6=z1
+
+	movq	mm4,mm2
+	movq	mm1,mm7
+	pfsub	mm2,mm0			; mm2=data4
+	pfsub	mm7,mm6			; mm7=data6
+	pfadd	mm4,mm0			; mm4=data0
+	pfadd	mm1,mm6			; mm1=data2
+
+	movq	MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], mm2
+	movq	MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], mm7
+	movq	MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4
+	movq	MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], mm1
+
+	; -- Odd part
+
+	movq	mm0, MMWORD [wk(0)]	; mm0=tmp6
+	movq	mm6, MMWORD [wk(1)]	; mm6=tmp7
+
+	pfadd	mm3,mm5			; mm3=tmp10
+	pfadd	mm5,mm0			; mm5=tmp11
+	pfadd	mm0,mm6			; mm0=tmp12, mm6=tmp7
+
+	pfmul	mm5,[GOTOFF(ebx,PD_0_707)] ; mm5=z3
+
+	movq	mm2,mm3			; mm2=tmp10
+	pfsub	mm3,mm0
+	pfmul	mm3,[GOTOFF(ebx,PD_0_382)] ; mm3=z5
+	pfmul	mm2,[GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610)
+	pfmul	mm0,[GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296)
+	pfadd	mm2,mm3			; mm2=z2
+	pfadd	mm0,mm3			; mm0=z4
+
+	movq	mm7,mm6
+	pfsub	mm6,mm5			; mm6=z13
+	pfadd	mm7,mm5			; mm7=z11
+
+	movq	mm4,mm6
+	movq	mm1,mm7
+	pfsub	mm6,mm2			; mm6=data3
+	pfsub	mm7,mm0			; mm7=data7
+	pfadd	mm4,mm2			; mm4=data5
+	pfadd	mm1,mm0			; mm1=data1
+
+	movq	MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], mm6
+	movq	MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], mm7
+	movq	MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], mm4
+	movq	MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1
+
+	add	edx, byte 2*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .columnloop
+
+	femms		; empty MMX/3DNow! state
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfmmxfst.asm b/simd/jfmmxfst.asm
new file mode 100644
index 0000000..0647242
--- /dev/null
+++ b/simd/jfmmxfst.asm
@@ -0,0 +1,397 @@
+;
+; jfmmxfst.asm - fast integer FDCT (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the forward DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
+; for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	8	; 14 is also OK.
+
+%if CONST_BITS == 8
+F_0_382	equ	 98		; FIX(0.382683433)
+F_0_541	equ	139		; FIX(0.541196100)
+F_0_707	equ	181		; FIX(0.707106781)
+F_1_306	equ	334		; FIX(1.306562965)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_382	equ	DESCALE( 410903207,30-CONST_BITS)	; FIX(0.382683433)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_707	equ	DESCALE( 759250124,30-CONST_BITS)	; FIX(0.707106781)
+F_1_306	equ	DESCALE(1402911301,30-CONST_BITS)	; FIX(1.306562965)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS   2
+%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+	alignz	16
+	global	EXTN(jconst_fdct_ifast_mmx)
+
+EXTN(jconst_fdct_ifast_mmx):
+
+PW_F0707	times 4 dw  F_0_707 << CONST_SHIFT
+PW_F0382	times 4 dw  F_0_382 << CONST_SHIFT
+PW_F0541	times 4 dw  F_0_541 << CONST_SHIFT
+PW_F1306	times 4 dw  F_1_306 << CONST_SHIFT
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_ifast_mmx (DCTELEM * data)
+;
+
+%define data(b)		(b)+8		; DCTELEM * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_ifast_mmx)
+
+EXTN(jsimd_fdct_ifast_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.rowloop:
+
+	movq	mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
+	movq	mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)]
+
+	; mm0=(20 21 22 23), mm2=(24 25 26 27)
+	; mm1=(30 31 32 33), mm3=(34 35 36 37)
+
+	movq      mm4,mm0		; transpose coefficients(phase 1)
+	punpcklwd mm0,mm1		; mm0=(20 30 21 31)
+	punpckhwd mm4,mm1		; mm4=(22 32 23 33)
+	movq      mm5,mm2		; transpose coefficients(phase 1)
+	punpcklwd mm2,mm3		; mm2=(24 34 25 35)
+	punpckhwd mm5,mm3		; mm5=(26 36 27 37)
+
+	movq	mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
+	movq	mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)]
+
+	; mm6=(00 01 02 03), mm1=(04 05 06 07)
+	; mm7=(10 11 12 13), mm3=(14 15 16 17)
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=(22 32 23 33)
+	movq	MMWORD [wk(1)], mm2	; wk(1)=(24 34 25 35)
+
+	movq      mm4,mm6		; transpose coefficients(phase 1)
+	punpcklwd mm6,mm7		; mm6=(00 10 01 11)
+	punpckhwd mm4,mm7		; mm4=(02 12 03 13)
+	movq      mm2,mm1		; transpose coefficients(phase 1)
+	punpcklwd mm1,mm3		; mm1=(04 14 05 15)
+	punpckhwd mm2,mm3		; mm2=(06 16 07 17)
+
+	movq      mm7,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm0		; mm6=(00 10 20 30)=data0
+	punpckhdq mm7,mm0		; mm7=(01 11 21 31)=data1
+	movq      mm3,mm2		; transpose coefficients(phase 2)
+	punpckldq mm2,mm5		; mm2=(06 16 26 36)=data6
+	punpckhdq mm3,mm5		; mm3=(07 17 27 37)=data7
+
+	movq	mm0,mm7
+	movq	mm5,mm6
+	psubw	mm7,mm2			; mm7=data1-data6=tmp6
+	psubw	mm6,mm3			; mm6=data0-data7=tmp7
+	paddw	mm0,mm2			; mm0=data1+data6=tmp1
+	paddw	mm5,mm3			; mm5=data0+data7=tmp0
+
+	movq	mm2, MMWORD [wk(0)]	; mm2=(22 32 23 33)
+	movq	mm3, MMWORD [wk(1)]	; mm3=(24 34 25 35)
+	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp6
+	movq	MMWORD [wk(1)], mm6	; wk(1)=tmp7
+
+	movq      mm7,mm4		; transpose coefficients(phase 2)
+	punpckldq mm4,mm2		; mm4=(02 12 22 32)=data2
+	punpckhdq mm7,mm2		; mm7=(03 13 23 33)=data3
+	movq      mm6,mm1		; transpose coefficients(phase 2)
+	punpckldq mm1,mm3		; mm1=(04 14 24 34)=data4
+	punpckhdq mm6,mm3		; mm6=(05 15 25 35)=data5
+
+	movq	mm2,mm7
+	movq	mm3,mm4
+	paddw	mm7,mm1			; mm7=data3+data4=tmp3
+	paddw	mm4,mm6			; mm4=data2+data5=tmp2
+	psubw	mm2,mm1			; mm2=data3-data4=tmp4
+	psubw	mm3,mm6			; mm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movq	mm1,mm5
+	movq	mm6,mm0
+	psubw	mm5,mm7			; mm5=tmp13
+	psubw	mm0,mm4			; mm0=tmp12
+	paddw	mm1,mm7			; mm1=tmp10
+	paddw	mm6,mm4			; mm6=tmp11
+
+	paddw	mm0,mm5
+	psllw	mm0,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm0,[GOTOFF(ebx,PW_F0707)] ; mm0=z1
+
+	movq	mm7,mm1
+	movq	mm4,mm5
+	psubw	mm1,mm6			; mm1=data4
+	psubw	mm5,mm0			; mm5=data6
+	paddw	mm7,mm6			; mm7=data0
+	paddw	mm4,mm0			; mm4=data2
+
+	movq	MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm1
+	movq	MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm5
+	movq	MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7
+	movq	MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
+
+	; -- Odd part
+
+	movq	mm6, MMWORD [wk(0)]	; mm6=tmp6
+	movq	mm0, MMWORD [wk(1)]	; mm0=tmp7
+
+	paddw	mm2,mm3			; mm2=tmp10
+	paddw	mm3,mm6			; mm3=tmp11
+	paddw	mm6,mm0			; mm6=tmp12, mm0=tmp7
+
+	psllw	mm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	mm6,PRE_MULTIPLY_SCALE_BITS
+
+	psllw	mm3,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm3,[GOTOFF(ebx,PW_F0707)] ; mm3=z3
+
+	movq	mm1,mm2			; mm1=tmp10
+	psubw	mm2,mm6
+	pmulhw	mm2,[GOTOFF(ebx,PW_F0382)] ; mm2=z5
+	pmulhw	mm1,[GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610)
+	pmulhw	mm6,[GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296)
+	paddw	mm1,mm2			; mm1=z2
+	paddw	mm6,mm2			; mm6=z4
+
+	movq	mm5,mm0
+	psubw	mm0,mm3			; mm0=z13
+	paddw	mm5,mm3			; mm5=z11
+
+	movq	mm7,mm0
+	movq	mm4,mm5
+	psubw	mm0,mm1			; mm0=data3
+	psubw	mm5,mm6			; mm5=data7
+	paddw	mm7,mm1			; mm7=data5
+	paddw	mm4,mm6			; mm4=data1
+
+	movq	MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0
+	movq	MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm5
+	movq	MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm7
+	movq	MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4
+
+	add	edx, byte 4*DCTSIZE*SIZEOF_DCTELEM
+	dec	ecx
+	jnz	near .rowloop
+
+	; ---- Pass 2: process columns.
+
+	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
+	movq	mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
+
+	; mm0=(02 12 22 32), mm2=(42 52 62 72)
+	; mm1=(03 13 23 33), mm3=(43 53 63 73)
+
+	movq      mm4,mm0		; transpose coefficients(phase 1)
+	punpcklwd mm0,mm1		; mm0=(02 03 12 13)
+	punpckhwd mm4,mm1		; mm4=(22 23 32 33)
+	movq      mm5,mm2		; transpose coefficients(phase 1)
+	punpcklwd mm2,mm3		; mm2=(42 43 52 53)
+	punpckhwd mm5,mm3		; mm5=(62 63 72 73)
+
+	movq	mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
+	movq	mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
+
+	; mm6=(00 10 20 30), mm1=(40 50 60 70)
+	; mm7=(01 11 21 31), mm3=(41 51 61 71)
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=(22 23 32 33)
+	movq	MMWORD [wk(1)], mm2	; wk(1)=(42 43 52 53)
+
+	movq      mm4,mm6		; transpose coefficients(phase 1)
+	punpcklwd mm6,mm7		; mm6=(00 01 10 11)
+	punpckhwd mm4,mm7		; mm4=(20 21 30 31)
+	movq      mm2,mm1		; transpose coefficients(phase 1)
+	punpcklwd mm1,mm3		; mm1=(40 41 50 51)
+	punpckhwd mm2,mm3		; mm2=(60 61 70 71)
+
+	movq      mm7,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm0		; mm6=(00 01 02 03)=data0
+	punpckhdq mm7,mm0		; mm7=(10 11 12 13)=data1
+	movq      mm3,mm2		; transpose coefficients(phase 2)
+	punpckldq mm2,mm5		; mm2=(60 61 62 63)=data6
+	punpckhdq mm3,mm5		; mm3=(70 71 72 73)=data7
+
+	movq	mm0,mm7
+	movq	mm5,mm6
+	psubw	mm7,mm2			; mm7=data1-data6=tmp6
+	psubw	mm6,mm3			; mm6=data0-data7=tmp7
+	paddw	mm0,mm2			; mm0=data1+data6=tmp1
+	paddw	mm5,mm3			; mm5=data0+data7=tmp0
+
+	movq	mm2, MMWORD [wk(0)]	; mm2=(22 23 32 33)
+	movq	mm3, MMWORD [wk(1)]	; mm3=(42 43 52 53)
+	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp6
+	movq	MMWORD [wk(1)], mm6	; wk(1)=tmp7
+
+	movq      mm7,mm4		; transpose coefficients(phase 2)
+	punpckldq mm4,mm2		; mm4=(20 21 22 23)=data2
+	punpckhdq mm7,mm2		; mm7=(30 31 32 33)=data3
+	movq      mm6,mm1		; transpose coefficients(phase 2)
+	punpckldq mm1,mm3		; mm1=(40 41 42 43)=data4
+	punpckhdq mm6,mm3		; mm6=(50 51 52 53)=data5
+
+	movq	mm2,mm7
+	movq	mm3,mm4
+	paddw	mm7,mm1			; mm7=data3+data4=tmp3
+	paddw	mm4,mm6			; mm4=data2+data5=tmp2
+	psubw	mm2,mm1			; mm2=data3-data4=tmp4
+	psubw	mm3,mm6			; mm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movq	mm1,mm5
+	movq	mm6,mm0
+	psubw	mm5,mm7			; mm5=tmp13
+	psubw	mm0,mm4			; mm0=tmp12
+	paddw	mm1,mm7			; mm1=tmp10
+	paddw	mm6,mm4			; mm6=tmp11
+
+	paddw	mm0,mm5
+	psllw	mm0,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm0,[GOTOFF(ebx,PW_F0707)] ; mm0=z1
+
+	movq	mm7,mm1
+	movq	mm4,mm5
+	psubw	mm1,mm6			; mm1=data4
+	psubw	mm5,mm0			; mm5=data6
+	paddw	mm7,mm6			; mm7=data0
+	paddw	mm4,mm0			; mm4=data2
+
+	movq	MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm1
+	movq	MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm5
+	movq	MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7
+	movq	MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
+
+	; -- Odd part
+
+	movq	mm6, MMWORD [wk(0)]	; mm6=tmp6
+	movq	mm0, MMWORD [wk(1)]	; mm0=tmp7
+
+	paddw	mm2,mm3			; mm2=tmp10
+	paddw	mm3,mm6			; mm3=tmp11
+	paddw	mm6,mm0			; mm6=tmp12, mm0=tmp7
+
+	psllw	mm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	mm6,PRE_MULTIPLY_SCALE_BITS
+
+	psllw	mm3,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm3,[GOTOFF(ebx,PW_F0707)] ; mm3=z3
+
+	movq	mm1,mm2			; mm1=tmp10
+	psubw	mm2,mm6
+	pmulhw	mm2,[GOTOFF(ebx,PW_F0382)] ; mm2=z5
+	pmulhw	mm1,[GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610)
+	pmulhw	mm6,[GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296)
+	paddw	mm1,mm2			; mm1=z2
+	paddw	mm6,mm2			; mm6=z4
+
+	movq	mm5,mm0
+	psubw	mm0,mm3			; mm0=z13
+	paddw	mm5,mm3			; mm5=z11
+
+	movq	mm7,mm0
+	movq	mm4,mm5
+	psubw	mm0,mm1			; mm0=data3
+	psubw	mm5,mm6			; mm5=data7
+	paddw	mm7,mm1			; mm7=data5
+	paddw	mm4,mm6			; mm4=data1
+
+	movq	MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0
+	movq	MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm5
+	movq	MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm7
+	movq	MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4
+
+	add	edx, byte 4*SIZEOF_DCTELEM
+	dec	ecx
+	jnz	near .columnloop
+
+	emms		; empty MMX state
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfmmxint.asm b/simd/jfmmxint.asm
new file mode 100644
index 0000000..a7e73f7
--- /dev/null
+++ b/simd/jfmmxint.asm
@@ -0,0 +1,622 @@
+;
+; jfmmxint.asm - accurate integer FDCT (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slow-but-accurate integer implementation of the
+; forward DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jfdctint.c; see the jfdctint.c for
+; more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
+%define DESCALE_P2	(CONST_BITS+PASS1_BITS)
+
+%if CONST_BITS == 13
+F_0_298	equ	 2446		; FIX(0.298631336)
+F_0_390	equ	 3196		; FIX(0.390180644)
+F_0_541	equ	 4433		; FIX(0.541196100)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_175	equ	 9633		; FIX(1.175875602)
+F_1_501	equ	12299		; FIX(1.501321110)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_1_961	equ	16069		; FIX(1.961570560)
+F_2_053	equ	16819		; FIX(2.053119869)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_072	equ	25172		; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
+F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
+F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
+F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_islow_mmx)
+
+EXTN(jconst_fdct_islow_mmx):
+
+PW_F130_F054	times 2 dw  (F_0_541+F_0_765), F_0_541
+PW_F054_MF130	times 2 dw  F_0_541, (F_0_541-F_1_847)
+PW_MF078_F117	times 2 dw  (F_1_175-F_1_961), F_1_175
+PW_F117_F078	times 2 dw  F_1_175, (F_1_175-F_0_390)
+PW_MF060_MF089	times 2 dw  (F_0_298-F_0_899),-F_0_899
+PW_MF089_F060	times 2 dw -F_0_899, (F_1_501-F_0_899)
+PW_MF050_MF256	times 2 dw  (F_2_053-F_2_562),-F_2_562
+PW_MF256_F050	times 2 dw -F_2_562, (F_3_072-F_2_562)
+PD_DESCALE_P1	times 2 dd  1 << (DESCALE_P1-1)
+PD_DESCALE_P2	times 2 dd  1 << (DESCALE_P2-1)
+PW_DESCALE_P2X	times 4 dw  1 << (PASS1_BITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_islow_mmx (DCTELEM * data)
+;
+
+%define data(b)		(b)+8		; DCTELEM * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_islow_mmx)
+
+EXTN(jsimd_fdct_islow_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.rowloop:
+
+	movq	mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
+	movq	mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)]
+
+	; mm0=(20 21 22 23), mm2=(24 25 26 27)
+	; mm1=(30 31 32 33), mm3=(34 35 36 37)
+
+	movq      mm4,mm0		; transpose coefficients(phase 1)
+	punpcklwd mm0,mm1		; mm0=(20 30 21 31)
+	punpckhwd mm4,mm1		; mm4=(22 32 23 33)
+	movq      mm5,mm2		; transpose coefficients(phase 1)
+	punpcklwd mm2,mm3		; mm2=(24 34 25 35)
+	punpckhwd mm5,mm3		; mm5=(26 36 27 37)
+
+	movq	mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
+	movq	mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)]
+
+	; mm6=(00 01 02 03), mm1=(04 05 06 07)
+	; mm7=(10 11 12 13), mm3=(14 15 16 17)
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=(22 32 23 33)
+	movq	MMWORD [wk(1)], mm2	; wk(1)=(24 34 25 35)
+
+	movq      mm4,mm6		; transpose coefficients(phase 1)
+	punpcklwd mm6,mm7		; mm6=(00 10 01 11)
+	punpckhwd mm4,mm7		; mm4=(02 12 03 13)
+	movq      mm2,mm1		; transpose coefficients(phase 1)
+	punpcklwd mm1,mm3		; mm1=(04 14 05 15)
+	punpckhwd mm2,mm3		; mm2=(06 16 07 17)
+
+	movq      mm7,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm0		; mm6=(00 10 20 30)=data0
+	punpckhdq mm7,mm0		; mm7=(01 11 21 31)=data1
+	movq      mm3,mm2		; transpose coefficients(phase 2)
+	punpckldq mm2,mm5		; mm2=(06 16 26 36)=data6
+	punpckhdq mm3,mm5		; mm3=(07 17 27 37)=data7
+
+	movq	mm0,mm7
+	movq	mm5,mm6
+	psubw	mm7,mm2			; mm7=data1-data6=tmp6
+	psubw	mm6,mm3			; mm6=data0-data7=tmp7
+	paddw	mm0,mm2			; mm0=data1+data6=tmp1
+	paddw	mm5,mm3			; mm5=data0+data7=tmp0
+
+	movq	mm2, MMWORD [wk(0)]	; mm2=(22 32 23 33)
+	movq	mm3, MMWORD [wk(1)]	; mm3=(24 34 25 35)
+	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp6
+	movq	MMWORD [wk(1)], mm6	; wk(1)=tmp7
+
+	movq      mm7,mm4		; transpose coefficients(phase 2)
+	punpckldq mm4,mm2		; mm4=(02 12 22 32)=data2
+	punpckhdq mm7,mm2		; mm7=(03 13 23 33)=data3
+	movq      mm6,mm1		; transpose coefficients(phase 2)
+	punpckldq mm1,mm3		; mm1=(04 14 24 34)=data4
+	punpckhdq mm6,mm3		; mm6=(05 15 25 35)=data5
+
+	movq	mm2,mm7
+	movq	mm3,mm4
+	paddw	mm7,mm1			; mm7=data3+data4=tmp3
+	paddw	mm4,mm6			; mm4=data2+data5=tmp2
+	psubw	mm2,mm1			; mm2=data3-data4=tmp4
+	psubw	mm3,mm6			; mm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movq	mm1,mm5
+	movq	mm6,mm0
+	paddw	mm5,mm7			; mm5=tmp10
+	paddw	mm0,mm4			; mm0=tmp11
+	psubw	mm1,mm7			; mm1=tmp13
+	psubw	mm6,mm4			; mm6=tmp12
+
+	movq	mm7,mm5
+	paddw	mm5,mm0			; mm5=tmp10+tmp11
+	psubw	mm7,mm0			; mm7=tmp10-tmp11
+
+	psllw	mm5,PASS1_BITS		; mm5=data0
+	psllw	mm7,PASS1_BITS		; mm7=data4
+
+	movq	MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm5
+	movq	MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm7
+
+	; (Original)
+	; z1 = (tmp12 + tmp13) * 0.541196100;
+	; data2 = z1 + tmp13 * 0.765366865;
+	; data6 = z1 + tmp12 * -1.847759065;
+	;
+	; (This implementation)
+	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+	movq      mm4,mm1		; mm1=tmp13
+	movq      mm0,mm1
+	punpcklwd mm4,mm6		; mm6=tmp12
+	punpckhwd mm0,mm6
+	movq      mm1,mm4
+	movq      mm6,mm0
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F130_F054)]	; mm4=data2L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_F130_F054)]	; mm0=data2H
+	pmaddwd   mm1,[GOTOFF(ebx,PW_F054_MF130)]	; mm1=data6L
+	pmaddwd   mm6,[GOTOFF(ebx,PW_F054_MF130)]	; mm6=data6H
+
+	paddd	mm4,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	mm0,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	mm4,DESCALE_P1
+	psrad	mm0,DESCALE_P1
+	paddd	mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	mm6,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	mm1,DESCALE_P1
+	psrad	mm6,DESCALE_P1
+
+	packssdw  mm4,mm0		; mm4=data2
+	packssdw  mm1,mm6		; mm1=data6
+
+	movq	MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
+	movq	MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm1
+
+	; -- Odd part
+
+	movq	mm5, MMWORD [wk(0)]	; mm5=tmp6
+	movq	mm7, MMWORD [wk(1)]	; mm7=tmp7
+
+	movq	mm0,mm2			; mm2=tmp4
+	movq	mm6,mm3			; mm3=tmp5
+	paddw	mm0,mm5			; mm0=z3
+	paddw	mm6,mm7			; mm6=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movq      mm4,mm0
+	movq      mm1,mm0
+	punpcklwd mm4,mm6
+	punpckhwd mm1,mm6
+	movq      mm0,mm4
+	movq      mm6,mm1
+	pmaddwd   mm4,[GOTOFF(ebx,PW_MF078_F117)]	; mm4=z3L
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF078_F117)]	; mm1=z3H
+	pmaddwd   mm0,[GOTOFF(ebx,PW_F117_F078)]	; mm0=z4L
+	pmaddwd   mm6,[GOTOFF(ebx,PW_F117_F078)]	; mm6=z4H
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=z3L
+	movq	MMWORD [wk(1)], mm1	; wk(1)=z3H
+
+	; (Original)
+	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+	;
+	; (This implementation)
+	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+	movq      mm4,mm2
+	movq      mm1,mm2
+	punpcklwd mm4,mm7
+	punpckhwd mm1,mm7
+	movq      mm2,mm4
+	movq      mm7,mm1
+	pmaddwd   mm4,[GOTOFF(ebx,PW_MF060_MF089)]	; mm4=tmp4L
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF060_MF089)]	; mm1=tmp4H
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF089_F060)]	; mm2=tmp7L
+	pmaddwd   mm7,[GOTOFF(ebx,PW_MF089_F060)]	; mm7=tmp7H
+
+	paddd	mm4, MMWORD [wk(0)]	; mm4=data7L
+	paddd	mm1, MMWORD [wk(1)]	; mm1=data7H
+	paddd	mm2,mm0			; mm2=data1L
+	paddd	mm7,mm6			; mm7=data1H
+
+	paddd	mm4,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	mm4,DESCALE_P1
+	psrad	mm1,DESCALE_P1
+	paddd	mm2,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	mm7,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	mm2,DESCALE_P1
+	psrad	mm7,DESCALE_P1
+
+	packssdw  mm4,mm1		; mm4=data7
+	packssdw  mm2,mm7		; mm2=data1
+
+	movq	MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm4
+	movq	MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm2
+
+	movq      mm1,mm3
+	movq      mm7,mm3
+	punpcklwd mm1,mm5
+	punpckhwd mm7,mm5
+	movq      mm3,mm1
+	movq      mm5,mm7
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF050_MF256)]	; mm1=tmp5L
+	pmaddwd   mm7,[GOTOFF(ebx,PW_MF050_MF256)]	; mm7=tmp5H
+	pmaddwd   mm3,[GOTOFF(ebx,PW_MF256_F050)]	; mm3=tmp6L
+	pmaddwd   mm5,[GOTOFF(ebx,PW_MF256_F050)]	; mm5=tmp6H
+
+	paddd	mm1,mm0			; mm1=data5L
+	paddd	mm7,mm6			; mm7=data5H
+	paddd	mm3, MMWORD [wk(0)]	; mm3=data3L
+	paddd	mm5, MMWORD [wk(1)]	; mm5=data3H
+
+	paddd	mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	mm7,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	mm1,DESCALE_P1
+	psrad	mm7,DESCALE_P1
+	paddd	mm3,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	mm5,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	mm3,DESCALE_P1
+	psrad	mm5,DESCALE_P1
+
+	packssdw  mm1,mm7		; mm1=data5
+	packssdw  mm3,mm5		; mm3=data3
+
+	movq	MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm1
+	movq	MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm3
+
+	add	edx, byte 4*DCTSIZE*SIZEOF_DCTELEM
+	dec	ecx
+	jnz	near .rowloop
+
+	; ---- Pass 2: process columns.
+
+	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.columnloop:
+
+	movq	mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
+	movq	mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
+
+	; mm0=(02 12 22 32), mm2=(42 52 62 72)
+	; mm1=(03 13 23 33), mm3=(43 53 63 73)
+
+	movq      mm4,mm0		; transpose coefficients(phase 1)
+	punpcklwd mm0,mm1		; mm0=(02 03 12 13)
+	punpckhwd mm4,mm1		; mm4=(22 23 32 33)
+	movq      mm5,mm2		; transpose coefficients(phase 1)
+	punpcklwd mm2,mm3		; mm2=(42 43 52 53)
+	punpckhwd mm5,mm3		; mm5=(62 63 72 73)
+
+	movq	mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
+	movq	mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
+	movq	mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
+	movq	mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
+
+	; mm6=(00 10 20 30), mm1=(40 50 60 70)
+	; mm7=(01 11 21 31), mm3=(41 51 61 71)
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=(22 23 32 33)
+	movq	MMWORD [wk(1)], mm2	; wk(1)=(42 43 52 53)
+
+	movq      mm4,mm6		; transpose coefficients(phase 1)
+	punpcklwd mm6,mm7		; mm6=(00 01 10 11)
+	punpckhwd mm4,mm7		; mm4=(20 21 30 31)
+	movq      mm2,mm1		; transpose coefficients(phase 1)
+	punpcklwd mm1,mm3		; mm1=(40 41 50 51)
+	punpckhwd mm2,mm3		; mm2=(60 61 70 71)
+
+	movq      mm7,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm0		; mm6=(00 01 02 03)=data0
+	punpckhdq mm7,mm0		; mm7=(10 11 12 13)=data1
+	movq      mm3,mm2		; transpose coefficients(phase 2)
+	punpckldq mm2,mm5		; mm2=(60 61 62 63)=data6
+	punpckhdq mm3,mm5		; mm3=(70 71 72 73)=data7
+
+	movq	mm0,mm7
+	movq	mm5,mm6
+	psubw	mm7,mm2			; mm7=data1-data6=tmp6
+	psubw	mm6,mm3			; mm6=data0-data7=tmp7
+	paddw	mm0,mm2			; mm0=data1+data6=tmp1
+	paddw	mm5,mm3			; mm5=data0+data7=tmp0
+
+	movq	mm2, MMWORD [wk(0)]	; mm2=(22 23 32 33)
+	movq	mm3, MMWORD [wk(1)]	; mm3=(42 43 52 53)
+	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp6
+	movq	MMWORD [wk(1)], mm6	; wk(1)=tmp7
+
+	movq      mm7,mm4		; transpose coefficients(phase 2)
+	punpckldq mm4,mm2		; mm4=(20 21 22 23)=data2
+	punpckhdq mm7,mm2		; mm7=(30 31 32 33)=data3
+	movq      mm6,mm1		; transpose coefficients(phase 2)
+	punpckldq mm1,mm3		; mm1=(40 41 42 43)=data4
+	punpckhdq mm6,mm3		; mm6=(50 51 52 53)=data5
+
+	movq	mm2,mm7
+	movq	mm3,mm4
+	paddw	mm7,mm1			; mm7=data3+data4=tmp3
+	paddw	mm4,mm6			; mm4=data2+data5=tmp2
+	psubw	mm2,mm1			; mm2=data3-data4=tmp4
+	psubw	mm3,mm6			; mm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movq	mm1,mm5
+	movq	mm6,mm0
+	paddw	mm5,mm7			; mm5=tmp10
+	paddw	mm0,mm4			; mm0=tmp11
+	psubw	mm1,mm7			; mm1=tmp13
+	psubw	mm6,mm4			; mm6=tmp12
+
+	movq	mm7,mm5
+	paddw	mm5,mm0			; mm5=tmp10+tmp11
+	psubw	mm7,mm0			; mm7=tmp10-tmp11
+
+	paddw	mm5,[GOTOFF(ebx,PW_DESCALE_P2X)]
+	paddw	mm7,[GOTOFF(ebx,PW_DESCALE_P2X)]
+	psraw	mm5,PASS1_BITS		; mm5=data0
+	psraw	mm7,PASS1_BITS		; mm7=data4
+
+	movq	MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm5
+	movq	MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm7
+
+	; (Original)
+	; z1 = (tmp12 + tmp13) * 0.541196100;
+	; data2 = z1 + tmp13 * 0.765366865;
+	; data6 = z1 + tmp12 * -1.847759065;
+	;
+	; (This implementation)
+	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+	movq      mm4,mm1		; mm1=tmp13
+	movq      mm0,mm1
+	punpcklwd mm4,mm6		; mm6=tmp12
+	punpckhwd mm0,mm6
+	movq      mm1,mm4
+	movq      mm6,mm0
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F130_F054)]	; mm4=data2L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_F130_F054)]	; mm0=data2H
+	pmaddwd   mm1,[GOTOFF(ebx,PW_F054_MF130)]	; mm1=data6L
+	pmaddwd   mm6,[GOTOFF(ebx,PW_F054_MF130)]	; mm6=data6H
+
+	paddd	mm4,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	mm0,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	mm4,DESCALE_P2
+	psrad	mm0,DESCALE_P2
+	paddd	mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	mm6,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	mm1,DESCALE_P2
+	psrad	mm6,DESCALE_P2
+
+	packssdw  mm4,mm0		; mm4=data2
+	packssdw  mm1,mm6		; mm1=data6
+
+	movq	MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
+	movq	MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm1
+
+	; -- Odd part
+
+	movq	mm5, MMWORD [wk(0)]	; mm5=tmp6
+	movq	mm7, MMWORD [wk(1)]	; mm7=tmp7
+
+	movq	mm0,mm2			; mm2=tmp4
+	movq	mm6,mm3			; mm3=tmp5
+	paddw	mm0,mm5			; mm0=z3
+	paddw	mm6,mm7			; mm6=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movq      mm4,mm0
+	movq      mm1,mm0
+	punpcklwd mm4,mm6
+	punpckhwd mm1,mm6
+	movq      mm0,mm4
+	movq      mm6,mm1
+	pmaddwd   mm4,[GOTOFF(ebx,PW_MF078_F117)]	; mm4=z3L
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF078_F117)]	; mm1=z3H
+	pmaddwd   mm0,[GOTOFF(ebx,PW_F117_F078)]	; mm0=z4L
+	pmaddwd   mm6,[GOTOFF(ebx,PW_F117_F078)]	; mm6=z4H
+
+	movq	MMWORD [wk(0)], mm4	; wk(0)=z3L
+	movq	MMWORD [wk(1)], mm1	; wk(1)=z3H
+
+	; (Original)
+	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+	;
+	; (This implementation)
+	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+	movq      mm4,mm2
+	movq      mm1,mm2
+	punpcklwd mm4,mm7
+	punpckhwd mm1,mm7
+	movq      mm2,mm4
+	movq      mm7,mm1
+	pmaddwd   mm4,[GOTOFF(ebx,PW_MF060_MF089)]	; mm4=tmp4L
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF060_MF089)]	; mm1=tmp4H
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF089_F060)]	; mm2=tmp7L
+	pmaddwd   mm7,[GOTOFF(ebx,PW_MF089_F060)]	; mm7=tmp7H
+
+	paddd	mm4, MMWORD [wk(0)]	; mm4=data7L
+	paddd	mm1, MMWORD [wk(1)]	; mm1=data7H
+	paddd	mm2,mm0			; mm2=data1L
+	paddd	mm7,mm6			; mm7=data1H
+
+	paddd	mm4,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	mm4,DESCALE_P2
+	psrad	mm1,DESCALE_P2
+	paddd	mm2,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	mm7,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	mm2,DESCALE_P2
+	psrad	mm7,DESCALE_P2
+
+	packssdw  mm4,mm1		; mm4=data7
+	packssdw  mm2,mm7		; mm2=data1
+
+	movq	MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm4
+	movq	MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm2
+
+	movq      mm1,mm3
+	movq      mm7,mm3
+	punpcklwd mm1,mm5
+	punpckhwd mm7,mm5
+	movq      mm3,mm1
+	movq      mm5,mm7
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF050_MF256)]	; mm1=tmp5L
+	pmaddwd   mm7,[GOTOFF(ebx,PW_MF050_MF256)]	; mm7=tmp5H
+	pmaddwd   mm3,[GOTOFF(ebx,PW_MF256_F050)]	; mm3=tmp6L
+	pmaddwd   mm5,[GOTOFF(ebx,PW_MF256_F050)]	; mm5=tmp6H
+
+	paddd	mm1,mm0			; mm1=data5L
+	paddd	mm7,mm6			; mm7=data5H
+	paddd	mm3, MMWORD [wk(0)]	; mm3=data3L
+	paddd	mm5, MMWORD [wk(1)]	; mm5=data3H
+
+	paddd	mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	mm7,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	mm1,DESCALE_P2
+	psrad	mm7,DESCALE_P2
+	paddd	mm3,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	mm5,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	mm3,DESCALE_P2
+	psrad	mm5,DESCALE_P2
+
+	packssdw  mm1,mm7		; mm1=data5
+	packssdw  mm3,mm5		; mm3=data3
+
+	movq	MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm1
+	movq	MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm3
+
+	add	edx, byte 4*SIZEOF_DCTELEM
+	dec	ecx
+	jnz	near .columnloop
+
+	emms		; empty MMX state
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfss2fst-64.asm b/simd/jfss2fst-64.asm
new file mode 100644
index 0000000..6953caf
--- /dev/null
+++ b/simd/jfss2fst-64.asm
@@ -0,0 +1,392 @@
+;
+; jfss2fst-64.asm - fast integer FDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the forward DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
+; for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	8	; 14 is also OK.
+
+%if CONST_BITS == 8
+F_0_382	equ	 98		; FIX(0.382683433)
+F_0_541	equ	139		; FIX(0.541196100)
+F_0_707	equ	181		; FIX(0.707106781)
+F_1_306	equ	334		; FIX(1.306562965)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_382	equ	DESCALE( 410903207,30-CONST_BITS)	; FIX(0.382683433)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_707	equ	DESCALE( 759250124,30-CONST_BITS)	; FIX(0.707106781)
+F_1_306	equ	DESCALE(1402911301,30-CONST_BITS)	; FIX(1.306562965)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS   2
+%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+	alignz	16
+	global	EXTN(jconst_fdct_ifast_sse2)
+
+EXTN(jconst_fdct_ifast_sse2):
+
+PW_F0707	times 8 dw  F_0_707 << CONST_SHIFT
+PW_F0382	times 8 dw  F_0_382 << CONST_SHIFT
+PW_F0541	times 8 dw  F_0_541 << CONST_SHIFT
+PW_F1306	times 8 dw  F_1_306 << CONST_SHIFT
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_ifast_sse2 (DCTELEM * data)
+;
+
+; r10 = DCTELEM * data
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_ifast_sse2)
+
+EXTN(jsimd_fdct_ifast_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+
+	; ---- Pass 1: process rows.
+
+	mov	rdx, r10	; (DCTELEM *)
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
+
+	; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
+	; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
+
+	movdqa    xmm4,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm1		; xmm0=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm4,xmm1		; xmm4=(04 14 05 15 06 16 07 17)
+	movdqa    xmm5,xmm2		; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm3		; xmm2=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm5,xmm3		; xmm5=(24 34 25 35 26 36 27 37)
+
+	movdqa	xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
+
+	; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
+	; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(20 30 21 31 22 32 23 33)
+	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(24 34 25 35 26 36 27 37)
+
+	movdqa    xmm2,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm7		; xmm6=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm2,xmm7		; xmm2=(44 54 45 55 46 56 47 57)
+	movdqa    xmm5,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm3		; xmm1=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm5,xmm3		; xmm5=(64 74 65 75 66 76 67 77)
+
+	movdqa    xmm7,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm1		; xmm6=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm7,xmm1		; xmm7=(42 52 62 72 43 53 63 73)
+	movdqa    xmm3,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm5		; xmm2=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm3,xmm5		; xmm3=(46 56 66 76 47 57 67 77)
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(20 30 21 31 22 32 23 33)
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(24 34 25 35 26 36 27 37)
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=(42 52 62 72 43 53 63 73)
+	movdqa	XMMWORD [wk(1)], xmm2	; wk(1)=(44 54 64 74 45 55 65 75)
+
+	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm1		; xmm0=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm7,xmm1		; xmm7=(02 12 22 32 03 13 23 33)
+	movdqa    xmm2,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm5		; xmm4=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm2,xmm5		; xmm2=(06 16 26 36 07 17 27 37)
+
+	movdqa     xmm1,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm6		; xmm0=(00 10 20 30 40 50 60 70)=data0
+	punpckhqdq xmm1,xmm6		; xmm1=(01 11 21 31 41 51 61 71)=data1
+	movdqa     xmm5,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm3		; xmm2=(06 16 26 36 46 56 66 76)=data6
+	punpckhqdq xmm5,xmm3		; xmm5=(07 17 27 37 47 57 67 77)=data7
+
+	movdqa	xmm6,xmm1
+	movdqa	xmm3,xmm0
+	psubw	xmm1,xmm2		; xmm1=data1-data6=tmp6
+	psubw	xmm0,xmm5		; xmm0=data0-data7=tmp7
+	paddw	xmm6,xmm2		; xmm6=data1+data6=tmp1
+	paddw	xmm3,xmm5		; xmm3=data0+data7=tmp0
+
+	movdqa	xmm2, XMMWORD [wk(0)]	; xmm2=(42 52 62 72 43 53 63 73)
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(44 54 64 74 45 55 65 75)
+	movdqa	XMMWORD [wk(0)], xmm1	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp7
+
+	movdqa     xmm1,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm2		; xmm7=(02 12 22 32 42 52 62 72)=data2
+	punpckhqdq xmm1,xmm2		; xmm1=(03 13 23 33 43 53 63 73)=data3
+	movdqa     xmm0,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm5		; xmm4=(04 14 24 34 44 54 64 74)=data4
+	punpckhqdq xmm0,xmm5		; xmm0=(05 15 25 35 45 55 65 75)=data5
+
+	movdqa	xmm2,xmm1
+	movdqa	xmm5,xmm7
+	paddw	xmm1,xmm4		; xmm1=data3+data4=tmp3
+	paddw	xmm7,xmm0		; xmm7=data2+data5=tmp2
+	psubw	xmm2,xmm4		; xmm2=data3-data4=tmp4
+	psubw	xmm5,xmm0		; xmm5=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm0,xmm6
+	psubw	xmm3,xmm1		; xmm3=tmp13
+	psubw	xmm6,xmm7		; xmm6=tmp12
+	paddw	xmm4,xmm1		; xmm4=tmp10
+	paddw	xmm0,xmm7		; xmm0=tmp11
+
+	paddw	xmm6,xmm3
+	psllw	xmm6,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm6,[rel PW_F0707] ; xmm6=z1
+
+	movdqa	xmm1,xmm4
+	movdqa	xmm7,xmm3
+	psubw	xmm4,xmm0		; xmm4=data4
+	psubw	xmm3,xmm6		; xmm3=data6
+	paddw	xmm1,xmm0		; xmm1=data0
+	paddw	xmm7,xmm6		; xmm7=data2
+
+	movdqa	xmm0, XMMWORD [wk(0)]	; xmm0=tmp6
+	movdqa	xmm6, XMMWORD [wk(1)]	; xmm6=tmp7
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=data4
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=data6
+
+	; -- Odd part
+
+	paddw	xmm2,xmm5		; xmm2=tmp10
+	paddw	xmm5,xmm0		; xmm5=tmp11
+	paddw	xmm0,xmm6		; xmm0=tmp12, xmm6=tmp7
+
+	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
+
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm5,[rel PW_F0707] ; xmm5=z3
+
+	movdqa	xmm4,xmm2		; xmm4=tmp10
+	psubw	xmm2,xmm0
+	pmulhw	xmm2,[rel PW_F0382] ; xmm2=z5
+	pmulhw	xmm4,[rel PW_F0541] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
+	pmulhw	xmm0,[rel PW_F1306] ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
+	paddw	xmm4,xmm2		; xmm4=z2
+	paddw	xmm0,xmm2		; xmm0=z4
+
+	movdqa	xmm3,xmm6
+	psubw	xmm6,xmm5		; xmm6=z13
+	paddw	xmm3,xmm5		; xmm3=z11
+
+	movdqa	xmm2,xmm6
+	movdqa	xmm5,xmm3
+	psubw	xmm6,xmm4		; xmm6=data3
+	psubw	xmm3,xmm0		; xmm3=data7
+	paddw	xmm2,xmm4		; xmm2=data5
+	paddw	xmm5,xmm0		; xmm5=data1
+
+	; ---- Pass 2: process columns.
+
+	; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
+	; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
+
+	movdqa    xmm4,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm5		; xmm1=(00 01 10 11 20 21 30 31)
+	punpckhwd xmm4,xmm5		; xmm4=(40 41 50 51 60 61 70 71)
+	movdqa    xmm0,xmm7		; transpose coefficients(phase 1)
+	punpcklwd xmm7,xmm6		; xmm7=(02 03 12 13 22 23 32 33)
+	punpckhwd xmm0,xmm6		; xmm0=(42 43 52 53 62 63 72 73)
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=col4
+	movdqa	xmm6, XMMWORD [wk(1)]	; xmm6=col6
+
+	; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
+	; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
+
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=(02 03 12 13 22 23 32 33)
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(42 43 52 53 62 63 72 73)
+
+	movdqa    xmm7,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm2		; xmm5=(04 05 14 15 24 25 34 35)
+	punpckhwd xmm7,xmm2		; xmm7=(44 45 54 55 64 65 74 75)
+	movdqa    xmm0,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm3		; xmm6=(06 07 16 17 26 27 36 37)
+	punpckhwd xmm0,xmm3		; xmm0=(46 47 56 57 66 67 76 77)
+
+	movdqa    xmm2,xmm5		; transpose coefficients(phase 2)
+	punpckldq xmm5,xmm6		; xmm5=(04 05 06 07 14 15 16 17)
+	punpckhdq xmm2,xmm6		; xmm2=(24 25 26 27 34 35 36 37)
+	movdqa    xmm3,xmm7		; transpose coefficients(phase 2)
+	punpckldq xmm7,xmm0		; xmm7=(44 45 46 47 54 55 56 57)
+	punpckhdq xmm3,xmm0		; xmm3=(64 65 66 67 74 75 76 77)
+
+	movdqa	xmm6, XMMWORD [wk(0)]	; xmm6=(02 03 12 13 22 23 32 33)
+	movdqa	xmm0, XMMWORD [wk(1)]	; xmm0=(42 43 52 53 62 63 72 73)
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(24 25 26 27 34 35 36 37)
+	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=(44 45 46 47 54 55 56 57)
+
+	movdqa    xmm2,xmm1		; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm6		; xmm1=(00 01 02 03 10 11 12 13)
+	punpckhdq xmm2,xmm6		; xmm2=(20 21 22 23 30 31 32 33)
+	movdqa    xmm7,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm0		; xmm4=(40 41 42 43 50 51 52 53)
+	punpckhdq xmm7,xmm0		; xmm7=(60 61 62 63 70 71 72 73)
+
+	movdqa     xmm6,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm5		; xmm1=(00 01 02 03 04 05 06 07)=data0
+	punpckhqdq xmm6,xmm5		; xmm6=(10 11 12 13 14 15 16 17)=data1
+	movdqa     xmm0,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm3		; xmm7=(60 61 62 63 64 65 66 67)=data6
+	punpckhqdq xmm0,xmm3		; xmm0=(70 71 72 73 74 75 76 77)=data7
+
+	movdqa	xmm5,xmm6
+	movdqa	xmm3,xmm1
+	psubw	xmm6,xmm7		; xmm6=data1-data6=tmp6
+	psubw	xmm1,xmm0		; xmm1=data0-data7=tmp7
+	paddw	xmm5,xmm7		; xmm5=data1+data6=tmp1
+	paddw	xmm3,xmm0		; xmm3=data0+data7=tmp0
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=(24 25 26 27 34 35 36 37)
+	movdqa	xmm0, XMMWORD [wk(1)]	; xmm0=(44 45 46 47 54 55 56 57)
+	movdqa	XMMWORD [wk(0)], xmm6	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm1	; wk(1)=tmp7
+
+	movdqa     xmm6,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm7		; xmm2=(20 21 22 23 24 25 26 27)=data2
+	punpckhqdq xmm6,xmm7		; xmm6=(30 31 32 33 34 35 36 37)=data3
+	movdqa     xmm1,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm0		; xmm4=(40 41 42 43 44 45 46 47)=data4
+	punpckhqdq xmm1,xmm0		; xmm1=(50 51 52 53 54 55 56 57)=data5
+
+	movdqa	xmm7,xmm6
+	movdqa	xmm0,xmm2
+	paddw	xmm6,xmm4		; xmm6=data3+data4=tmp3
+	paddw	xmm2,xmm1		; xmm2=data2+data5=tmp2
+	psubw	xmm7,xmm4		; xmm7=data3-data4=tmp4
+	psubw	xmm0,xmm1		; xmm0=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm1,xmm5
+	psubw	xmm3,xmm6		; xmm3=tmp13
+	psubw	xmm5,xmm2		; xmm5=tmp12
+	paddw	xmm4,xmm6		; xmm4=tmp10
+	paddw	xmm1,xmm2		; xmm1=tmp11
+
+	paddw	xmm5,xmm3
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm5,[rel PW_F0707] ; xmm5=z1
+
+	movdqa	xmm6,xmm4
+	movdqa	xmm2,xmm3
+	psubw	xmm4,xmm1		; xmm4=data4
+	psubw	xmm3,xmm5		; xmm3=data6
+	paddw	xmm6,xmm1		; xmm6=data0
+	paddw	xmm2,xmm5		; xmm2=data2
+
+	movdqa	XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm4
+	movdqa	XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm3
+	movdqa	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm6
+	movdqa	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm2
+
+	; -- Odd part
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=tmp6
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp7
+
+	paddw	xmm7,xmm0		; xmm7=tmp10
+	paddw	xmm0,xmm1		; xmm0=tmp11
+	paddw	xmm1,xmm5		; xmm1=tmp12, xmm5=tmp7
+
+	psllw	xmm7,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm1,PRE_MULTIPLY_SCALE_BITS
+
+	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm0,[rel PW_F0707] ; xmm0=z3
+
+	movdqa	xmm4,xmm7		; xmm4=tmp10
+	psubw	xmm7,xmm1
+	pmulhw	xmm7,[rel PW_F0382] ; xmm7=z5
+	pmulhw	xmm4,[rel PW_F0541] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
+	pmulhw	xmm1,[rel PW_F1306] ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
+	paddw	xmm4,xmm7		; xmm4=z2
+	paddw	xmm1,xmm7		; xmm1=z4
+
+	movdqa	xmm3,xmm5
+	psubw	xmm5,xmm0		; xmm5=z13
+	paddw	xmm3,xmm0		; xmm3=z11
+
+	movdqa	xmm6,xmm5
+	movdqa	xmm2,xmm3
+	psubw	xmm5,xmm4		; xmm5=data3
+	psubw	xmm3,xmm1		; xmm3=data7
+	paddw	xmm6,xmm4		; xmm6=data5
+	paddw	xmm2,xmm1		; xmm2=data1
+
+	movdqa	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm5
+	movdqa	XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm3
+	movdqa	XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm6
+	movdqa	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm2
+
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfss2fst.asm b/simd/jfss2fst.asm
new file mode 100644
index 0000000..73fc9e5
--- /dev/null
+++ b/simd/jfss2fst.asm
@@ -0,0 +1,404 @@
+;
+; jfss2fst.asm - fast integer FDCT (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the forward DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
+; for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	8	; 14 is also OK.
+
+%if CONST_BITS == 8
+F_0_382	equ	 98		; FIX(0.382683433)
+F_0_541	equ	139		; FIX(0.541196100)
+F_0_707	equ	181		; FIX(0.707106781)
+F_1_306	equ	334		; FIX(1.306562965)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_382	equ	DESCALE( 410903207,30-CONST_BITS)	; FIX(0.382683433)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_707	equ	DESCALE( 759250124,30-CONST_BITS)	; FIX(0.707106781)
+F_1_306	equ	DESCALE(1402911301,30-CONST_BITS)	; FIX(1.306562965)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS   2
+%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+	alignz	16
+	global	EXTN(jconst_fdct_ifast_sse2)
+
+EXTN(jconst_fdct_ifast_sse2):
+
+PW_F0707	times 8 dw  F_0_707 << CONST_SHIFT
+PW_F0382	times 8 dw  F_0_382 << CONST_SHIFT
+PW_F0541	times 8 dw  F_0_541 << CONST_SHIFT
+PW_F1306	times 8 dw  F_1_306 << CONST_SHIFT
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_ifast_sse2 (DCTELEM * data)
+;
+
+%define data(b)		(b)+8		; DCTELEM * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_ifast_sse2)
+
+EXTN(jsimd_fdct_ifast_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
+
+	; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
+	; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
+
+	movdqa    xmm4,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm1		; xmm0=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm4,xmm1		; xmm4=(04 14 05 15 06 16 07 17)
+	movdqa    xmm5,xmm2		; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm3		; xmm2=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm5,xmm3		; xmm5=(24 34 25 35 26 36 27 37)
+
+	movdqa	xmm6, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm7, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
+
+	; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
+	; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(20 30 21 31 22 32 23 33)
+	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(24 34 25 35 26 36 27 37)
+
+	movdqa    xmm2,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm7		; xmm6=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm2,xmm7		; xmm2=(44 54 45 55 46 56 47 57)
+	movdqa    xmm5,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm3		; xmm1=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm5,xmm3		; xmm5=(64 74 65 75 66 76 67 77)
+
+	movdqa    xmm7,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm1		; xmm6=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm7,xmm1		; xmm7=(42 52 62 72 43 53 63 73)
+	movdqa    xmm3,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm5		; xmm2=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm3,xmm5		; xmm3=(46 56 66 76 47 57 67 77)
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(20 30 21 31 22 32 23 33)
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(24 34 25 35 26 36 27 37)
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=(42 52 62 72 43 53 63 73)
+	movdqa	XMMWORD [wk(1)], xmm2	; wk(1)=(44 54 64 74 45 55 65 75)
+
+	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm1		; xmm0=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm7,xmm1		; xmm7=(02 12 22 32 03 13 23 33)
+	movdqa    xmm2,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm5		; xmm4=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm2,xmm5		; xmm2=(06 16 26 36 07 17 27 37)
+
+	movdqa     xmm1,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm6		; xmm0=(00 10 20 30 40 50 60 70)=data0
+	punpckhqdq xmm1,xmm6		; xmm1=(01 11 21 31 41 51 61 71)=data1
+	movdqa     xmm5,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm3		; xmm2=(06 16 26 36 46 56 66 76)=data6
+	punpckhqdq xmm5,xmm3		; xmm5=(07 17 27 37 47 57 67 77)=data7
+
+	movdqa	xmm6,xmm1
+	movdqa	xmm3,xmm0
+	psubw	xmm1,xmm2		; xmm1=data1-data6=tmp6
+	psubw	xmm0,xmm5		; xmm0=data0-data7=tmp7
+	paddw	xmm6,xmm2		; xmm6=data1+data6=tmp1
+	paddw	xmm3,xmm5		; xmm3=data0+data7=tmp0
+
+	movdqa	xmm2, XMMWORD [wk(0)]	; xmm2=(42 52 62 72 43 53 63 73)
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(44 54 64 74 45 55 65 75)
+	movdqa	XMMWORD [wk(0)], xmm1	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp7
+
+	movdqa     xmm1,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm2		; xmm7=(02 12 22 32 42 52 62 72)=data2
+	punpckhqdq xmm1,xmm2		; xmm1=(03 13 23 33 43 53 63 73)=data3
+	movdqa     xmm0,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm5		; xmm4=(04 14 24 34 44 54 64 74)=data4
+	punpckhqdq xmm0,xmm5		; xmm0=(05 15 25 35 45 55 65 75)=data5
+
+	movdqa	xmm2,xmm1
+	movdqa	xmm5,xmm7
+	paddw	xmm1,xmm4		; xmm1=data3+data4=tmp3
+	paddw	xmm7,xmm0		; xmm7=data2+data5=tmp2
+	psubw	xmm2,xmm4		; xmm2=data3-data4=tmp4
+	psubw	xmm5,xmm0		; xmm5=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm0,xmm6
+	psubw	xmm3,xmm1		; xmm3=tmp13
+	psubw	xmm6,xmm7		; xmm6=tmp12
+	paddw	xmm4,xmm1		; xmm4=tmp10
+	paddw	xmm0,xmm7		; xmm0=tmp11
+
+	paddw	xmm6,xmm3
+	psllw	xmm6,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm6,[GOTOFF(ebx,PW_F0707)] ; xmm6=z1
+
+	movdqa	xmm1,xmm4
+	movdqa	xmm7,xmm3
+	psubw	xmm4,xmm0		; xmm4=data4
+	psubw	xmm3,xmm6		; xmm3=data6
+	paddw	xmm1,xmm0		; xmm1=data0
+	paddw	xmm7,xmm6		; xmm7=data2
+
+	movdqa	xmm0, XMMWORD [wk(0)]	; xmm0=tmp6
+	movdqa	xmm6, XMMWORD [wk(1)]	; xmm6=tmp7
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=data4
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=data6
+
+	; -- Odd part
+
+	paddw	xmm2,xmm5		; xmm2=tmp10
+	paddw	xmm5,xmm0		; xmm5=tmp11
+	paddw	xmm0,xmm6		; xmm0=tmp12, xmm6=tmp7
+
+	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
+
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm5,[GOTOFF(ebx,PW_F0707)] ; xmm5=z3
+
+	movdqa	xmm4,xmm2		; xmm4=tmp10
+	psubw	xmm2,xmm0
+	pmulhw	xmm2,[GOTOFF(ebx,PW_F0382)] ; xmm2=z5
+	pmulhw	xmm4,[GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
+	pmulhw	xmm0,[GOTOFF(ebx,PW_F1306)] ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
+	paddw	xmm4,xmm2		; xmm4=z2
+	paddw	xmm0,xmm2		; xmm0=z4
+
+	movdqa	xmm3,xmm6
+	psubw	xmm6,xmm5		; xmm6=z13
+	paddw	xmm3,xmm5		; xmm3=z11
+
+	movdqa	xmm2,xmm6
+	movdqa	xmm5,xmm3
+	psubw	xmm6,xmm4		; xmm6=data3
+	psubw	xmm3,xmm0		; xmm3=data7
+	paddw	xmm2,xmm4		; xmm2=data5
+	paddw	xmm5,xmm0		; xmm5=data1
+
+	; ---- Pass 2: process columns.
+
+;	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+
+	; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
+	; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
+
+	movdqa    xmm4,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm5		; xmm1=(00 01 10 11 20 21 30 31)
+	punpckhwd xmm4,xmm5		; xmm4=(40 41 50 51 60 61 70 71)
+	movdqa    xmm0,xmm7		; transpose coefficients(phase 1)
+	punpcklwd xmm7,xmm6		; xmm7=(02 03 12 13 22 23 32 33)
+	punpckhwd xmm0,xmm6		; xmm0=(42 43 52 53 62 63 72 73)
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=col4
+	movdqa	xmm6, XMMWORD [wk(1)]	; xmm6=col6
+
+	; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
+	; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
+
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=(02 03 12 13 22 23 32 33)
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(42 43 52 53 62 63 72 73)
+
+	movdqa    xmm7,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm2		; xmm5=(04 05 14 15 24 25 34 35)
+	punpckhwd xmm7,xmm2		; xmm7=(44 45 54 55 64 65 74 75)
+	movdqa    xmm0,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm3		; xmm6=(06 07 16 17 26 27 36 37)
+	punpckhwd xmm0,xmm3		; xmm0=(46 47 56 57 66 67 76 77)
+
+	movdqa    xmm2,xmm5		; transpose coefficients(phase 2)
+	punpckldq xmm5,xmm6		; xmm5=(04 05 06 07 14 15 16 17)
+	punpckhdq xmm2,xmm6		; xmm2=(24 25 26 27 34 35 36 37)
+	movdqa    xmm3,xmm7		; transpose coefficients(phase 2)
+	punpckldq xmm7,xmm0		; xmm7=(44 45 46 47 54 55 56 57)
+	punpckhdq xmm3,xmm0		; xmm3=(64 65 66 67 74 75 76 77)
+
+	movdqa	xmm6, XMMWORD [wk(0)]	; xmm6=(02 03 12 13 22 23 32 33)
+	movdqa	xmm0, XMMWORD [wk(1)]	; xmm0=(42 43 52 53 62 63 72 73)
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(24 25 26 27 34 35 36 37)
+	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=(44 45 46 47 54 55 56 57)
+
+	movdqa    xmm2,xmm1		; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm6		; xmm1=(00 01 02 03 10 11 12 13)
+	punpckhdq xmm2,xmm6		; xmm2=(20 21 22 23 30 31 32 33)
+	movdqa    xmm7,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm0		; xmm4=(40 41 42 43 50 51 52 53)
+	punpckhdq xmm7,xmm0		; xmm7=(60 61 62 63 70 71 72 73)
+
+	movdqa     xmm6,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm5		; xmm1=(00 01 02 03 04 05 06 07)=data0
+	punpckhqdq xmm6,xmm5		; xmm6=(10 11 12 13 14 15 16 17)=data1
+	movdqa     xmm0,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm3		; xmm7=(60 61 62 63 64 65 66 67)=data6
+	punpckhqdq xmm0,xmm3		; xmm0=(70 71 72 73 74 75 76 77)=data7
+
+	movdqa	xmm5,xmm6
+	movdqa	xmm3,xmm1
+	psubw	xmm6,xmm7		; xmm6=data1-data6=tmp6
+	psubw	xmm1,xmm0		; xmm1=data0-data7=tmp7
+	paddw	xmm5,xmm7		; xmm5=data1+data6=tmp1
+	paddw	xmm3,xmm0		; xmm3=data0+data7=tmp0
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=(24 25 26 27 34 35 36 37)
+	movdqa	xmm0, XMMWORD [wk(1)]	; xmm0=(44 45 46 47 54 55 56 57)
+	movdqa	XMMWORD [wk(0)], xmm6	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm1	; wk(1)=tmp7
+
+	movdqa     xmm6,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm7		; xmm2=(20 21 22 23 24 25 26 27)=data2
+	punpckhqdq xmm6,xmm7		; xmm6=(30 31 32 33 34 35 36 37)=data3
+	movdqa     xmm1,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm0		; xmm4=(40 41 42 43 44 45 46 47)=data4
+	punpckhqdq xmm1,xmm0		; xmm1=(50 51 52 53 54 55 56 57)=data5
+
+	movdqa	xmm7,xmm6
+	movdqa	xmm0,xmm2
+	paddw	xmm6,xmm4		; xmm6=data3+data4=tmp3
+	paddw	xmm2,xmm1		; xmm2=data2+data5=tmp2
+	psubw	xmm7,xmm4		; xmm7=data3-data4=tmp4
+	psubw	xmm0,xmm1		; xmm0=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm1,xmm5
+	psubw	xmm3,xmm6		; xmm3=tmp13
+	psubw	xmm5,xmm2		; xmm5=tmp12
+	paddw	xmm4,xmm6		; xmm4=tmp10
+	paddw	xmm1,xmm2		; xmm1=tmp11
+
+	paddw	xmm5,xmm3
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm5,[GOTOFF(ebx,PW_F0707)] ; xmm5=z1
+
+	movdqa	xmm6,xmm4
+	movdqa	xmm2,xmm3
+	psubw	xmm4,xmm1		; xmm4=data4
+	psubw	xmm3,xmm5		; xmm3=data6
+	paddw	xmm6,xmm1		; xmm6=data0
+	paddw	xmm2,xmm5		; xmm2=data2
+
+	movdqa	XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)], xmm4
+	movdqa	XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)], xmm3
+	movdqa	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)], xmm6
+	movdqa	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)], xmm2
+
+	; -- Odd part
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=tmp6
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp7
+
+	paddw	xmm7,xmm0		; xmm7=tmp10
+	paddw	xmm0,xmm1		; xmm0=tmp11
+	paddw	xmm1,xmm5		; xmm1=tmp12, xmm5=tmp7
+
+	psllw	xmm7,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm1,PRE_MULTIPLY_SCALE_BITS
+
+	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm0,[GOTOFF(ebx,PW_F0707)] ; xmm0=z3
+
+	movdqa	xmm4,xmm7		; xmm4=tmp10
+	psubw	xmm7,xmm1
+	pmulhw	xmm7,[GOTOFF(ebx,PW_F0382)] ; xmm7=z5
+	pmulhw	xmm4,[GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
+	pmulhw	xmm1,[GOTOFF(ebx,PW_F1306)] ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
+	paddw	xmm4,xmm7		; xmm4=z2
+	paddw	xmm1,xmm7		; xmm1=z4
+
+	movdqa	xmm3,xmm5
+	psubw	xmm5,xmm0		; xmm5=z13
+	paddw	xmm3,xmm0		; xmm3=z11
+
+	movdqa	xmm6,xmm5
+	movdqa	xmm2,xmm3
+	psubw	xmm5,xmm4		; xmm5=data3
+	psubw	xmm3,xmm1		; xmm3=data7
+	paddw	xmm6,xmm4		; xmm6=data5
+	paddw	xmm2,xmm1		; xmm2=data1
+
+	movdqa	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)], xmm5
+	movdqa	XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)], xmm3
+	movdqa	XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)], xmm6
+	movdqa	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)], xmm2
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfss2int-64.asm b/simd/jfss2int-64.asm
new file mode 100644
index 0000000..bd1bd45
--- /dev/null
+++ b/simd/jfss2int-64.asm
@@ -0,0 +1,622 @@
+;
+; jfss2int-64.asm - accurate integer FDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slow-but-accurate integer implementation of the
+; forward DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jfdctint.c; see the jfdctint.c for
+; more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
+%define DESCALE_P2	(CONST_BITS+PASS1_BITS)
+
+%if CONST_BITS == 13
+F_0_298	equ	 2446		; FIX(0.298631336)
+F_0_390	equ	 3196		; FIX(0.390180644)
+F_0_541	equ	 4433		; FIX(0.541196100)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_175	equ	 9633		; FIX(1.175875602)
+F_1_501	equ	12299		; FIX(1.501321110)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_1_961	equ	16069		; FIX(1.961570560)
+F_2_053	equ	16819		; FIX(2.053119869)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_072	equ	25172		; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
+F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
+F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
+F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_islow_sse2)
+
+EXTN(jconst_fdct_islow_sse2):
+
+PW_F130_F054	times 4 dw  (F_0_541+F_0_765), F_0_541
+PW_F054_MF130	times 4 dw  F_0_541, (F_0_541-F_1_847)
+PW_MF078_F117	times 4 dw  (F_1_175-F_1_961), F_1_175
+PW_F117_F078	times 4 dw  F_1_175, (F_1_175-F_0_390)
+PW_MF060_MF089	times 4 dw  (F_0_298-F_0_899),-F_0_899
+PW_MF089_F060	times 4 dw -F_0_899, (F_1_501-F_0_899)
+PW_MF050_MF256	times 4 dw  (F_2_053-F_2_562),-F_2_562
+PW_MF256_F050	times 4 dw -F_2_562, (F_3_072-F_2_562)
+PD_DESCALE_P1	times 4 dd  1 << (DESCALE_P1-1)
+PD_DESCALE_P2	times 4 dd  1 << (DESCALE_P2-1)
+PW_DESCALE_P2X	times 8 dw  1 << (PASS1_BITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_islow_sse2 (DCTELEM * data)
+;
+
+; r10 = DCTELEM * data
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		6
+
+	align	16
+	global	EXTN(jsimd_fdct_islow_sse2)
+
+EXTN(jsimd_fdct_islow_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+
+	; ---- Pass 1: process rows.
+
+	mov	rdx, r10	; (DCTELEM *)
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
+
+	; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
+	; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
+
+	movdqa    xmm4,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm1		; xmm0=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm4,xmm1		; xmm4=(04 14 05 15 06 16 07 17)
+	movdqa    xmm5,xmm2		; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm3		; xmm2=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm5,xmm3		; xmm5=(24 34 25 35 26 36 27 37)
+
+	movdqa	xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
+
+	; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
+	; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(20 30 21 31 22 32 23 33)
+	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(24 34 25 35 26 36 27 37)
+
+	movdqa    xmm2,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm7		; xmm6=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm2,xmm7		; xmm2=(44 54 45 55 46 56 47 57)
+	movdqa    xmm5,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm3		; xmm1=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm5,xmm3		; xmm5=(64 74 65 75 66 76 67 77)
+
+	movdqa    xmm7,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm1		; xmm6=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm7,xmm1		; xmm7=(42 52 62 72 43 53 63 73)
+	movdqa    xmm3,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm5		; xmm2=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm3,xmm5		; xmm3=(46 56 66 76 47 57 67 77)
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(20 30 21 31 22 32 23 33)
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(24 34 25 35 26 36 27 37)
+	movdqa	XMMWORD [wk(2)], xmm7	; wk(2)=(42 52 62 72 43 53 63 73)
+	movdqa	XMMWORD [wk(3)], xmm2	; wk(3)=(44 54 64 74 45 55 65 75)
+
+	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm1		; xmm0=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm7,xmm1		; xmm7=(02 12 22 32 03 13 23 33)
+	movdqa    xmm2,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm5		; xmm4=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm2,xmm5		; xmm2=(06 16 26 36 07 17 27 37)
+
+	movdqa     xmm1,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm6		; xmm0=(00 10 20 30 40 50 60 70)=data0
+	punpckhqdq xmm1,xmm6		; xmm1=(01 11 21 31 41 51 61 71)=data1
+	movdqa     xmm5,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm3		; xmm2=(06 16 26 36 46 56 66 76)=data6
+	punpckhqdq xmm5,xmm3		; xmm5=(07 17 27 37 47 57 67 77)=data7
+
+	movdqa	xmm6,xmm1
+	movdqa	xmm3,xmm0
+	psubw	xmm1,xmm2		; xmm1=data1-data6=tmp6
+	psubw	xmm0,xmm5		; xmm0=data0-data7=tmp7
+	paddw	xmm6,xmm2		; xmm6=data1+data6=tmp1
+	paddw	xmm3,xmm5		; xmm3=data0+data7=tmp0
+
+	movdqa	xmm2, XMMWORD [wk(2)]	; xmm2=(42 52 62 72 43 53 63 73)
+	movdqa	xmm5, XMMWORD [wk(3)]	; xmm5=(44 54 64 74 45 55 65 75)
+	movdqa	XMMWORD [wk(0)], xmm1	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp7
+
+	movdqa     xmm1,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm2		; xmm7=(02 12 22 32 42 52 62 72)=data2
+	punpckhqdq xmm1,xmm2		; xmm1=(03 13 23 33 43 53 63 73)=data3
+	movdqa     xmm0,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm5		; xmm4=(04 14 24 34 44 54 64 74)=data4
+	punpckhqdq xmm0,xmm5		; xmm0=(05 15 25 35 45 55 65 75)=data5
+
+	movdqa	xmm2,xmm1
+	movdqa	xmm5,xmm7
+	paddw	xmm1,xmm4		; xmm1=data3+data4=tmp3
+	paddw	xmm7,xmm0		; xmm7=data2+data5=tmp2
+	psubw	xmm2,xmm4		; xmm2=data3-data4=tmp4
+	psubw	xmm5,xmm0		; xmm5=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm0,xmm6
+	paddw	xmm3,xmm1		; xmm3=tmp10
+	paddw	xmm6,xmm7		; xmm6=tmp11
+	psubw	xmm4,xmm1		; xmm4=tmp13
+	psubw	xmm0,xmm7		; xmm0=tmp12
+
+	movdqa	xmm1,xmm3
+	paddw	xmm3,xmm6		; xmm3=tmp10+tmp11
+	psubw	xmm1,xmm6		; xmm1=tmp10-tmp11
+
+	psllw	xmm3,PASS1_BITS		; xmm3=data0
+	psllw	xmm1,PASS1_BITS		; xmm1=data4
+
+	movdqa	XMMWORD [wk(2)], xmm3	; wk(2)=data0
+	movdqa	XMMWORD [wk(3)], xmm1	; wk(3)=data4
+
+	; (Original)
+	; z1 = (tmp12 + tmp13) * 0.541196100;
+	; data2 = z1 + tmp13 * 0.765366865;
+	; data6 = z1 + tmp12 * -1.847759065;
+	;
+	; (This implementation)
+	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+	movdqa    xmm7,xmm4		; xmm4=tmp13
+	movdqa    xmm6,xmm4
+	punpcklwd xmm7,xmm0		; xmm0=tmp12
+	punpckhwd xmm6,xmm0
+	movdqa    xmm4,xmm7
+	movdqa    xmm0,xmm6
+	pmaddwd   xmm7,[rel PW_F130_F054]	; xmm7=data2L
+	pmaddwd   xmm6,[rel PW_F130_F054]	; xmm6=data2H
+	pmaddwd   xmm4,[rel PW_F054_MF130]	; xmm4=data6L
+	pmaddwd   xmm0,[rel PW_F054_MF130]	; xmm0=data6H
+
+	paddd	xmm7,[rel PD_DESCALE_P1]
+	paddd	xmm6,[rel PD_DESCALE_P1]
+	psrad	xmm7,DESCALE_P1
+	psrad	xmm6,DESCALE_P1
+	paddd	xmm4,[rel PD_DESCALE_P1]
+	paddd	xmm0,[rel PD_DESCALE_P1]
+	psrad	xmm4,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+
+	packssdw  xmm7,xmm6		; xmm7=data2
+	packssdw  xmm4,xmm0		; xmm4=data6
+
+	movdqa	XMMWORD [wk(4)], xmm7	; wk(4)=data2
+	movdqa	XMMWORD [wk(5)], xmm4	; wk(5)=data6
+
+	; -- Odd part
+
+	movdqa	xmm3, XMMWORD [wk(0)]	; xmm3=tmp6
+	movdqa	xmm1, XMMWORD [wk(1)]	; xmm1=tmp7
+
+	movdqa	xmm6,xmm2		; xmm2=tmp4
+	movdqa	xmm0,xmm5		; xmm5=tmp5
+	paddw	xmm6,xmm3		; xmm6=z3
+	paddw	xmm0,xmm1		; xmm0=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm7,xmm6
+	movdqa    xmm4,xmm6
+	punpcklwd xmm7,xmm0
+	punpckhwd xmm4,xmm0
+	movdqa    xmm6,xmm7
+	movdqa    xmm0,xmm4
+	pmaddwd   xmm7,[rel PW_MF078_F117]	; xmm7=z3L
+	pmaddwd   xmm4,[rel PW_MF078_F117]	; xmm4=z3H
+	pmaddwd   xmm6,[rel PW_F117_F078]	; xmm6=z4L
+	pmaddwd   xmm0,[rel PW_F117_F078]	; xmm0=z4H
+
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=z3L
+	movdqa	XMMWORD [wk(1)], xmm4	; wk(1)=z3H
+
+	; (Original)
+	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+	;
+	; (This implementation)
+	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+	movdqa    xmm7,xmm2
+	movdqa    xmm4,xmm2
+	punpcklwd xmm7,xmm1
+	punpckhwd xmm4,xmm1
+	movdqa    xmm2,xmm7
+	movdqa    xmm1,xmm4
+	pmaddwd   xmm7,[rel PW_MF060_MF089]	; xmm7=tmp4L
+	pmaddwd   xmm4,[rel PW_MF060_MF089]	; xmm4=tmp4H
+	pmaddwd   xmm2,[rel PW_MF089_F060]	; xmm2=tmp7L
+	pmaddwd   xmm1,[rel PW_MF089_F060]	; xmm1=tmp7H
+
+	paddd	xmm7, XMMWORD [wk(0)]	; xmm7=data7L
+	paddd	xmm4, XMMWORD [wk(1)]	; xmm4=data7H
+	paddd	xmm2,xmm6		; xmm2=data1L
+	paddd	xmm1,xmm0		; xmm1=data1H
+
+	paddd	xmm7,[rel PD_DESCALE_P1]
+	paddd	xmm4,[rel PD_DESCALE_P1]
+	psrad	xmm7,DESCALE_P1
+	psrad	xmm4,DESCALE_P1
+	paddd	xmm2,[rel PD_DESCALE_P1]
+	paddd	xmm1,[rel PD_DESCALE_P1]
+	psrad	xmm2,DESCALE_P1
+	psrad	xmm1,DESCALE_P1
+
+	packssdw  xmm7,xmm4		; xmm7=data7
+	packssdw  xmm2,xmm1		; xmm2=data1
+
+	movdqa    xmm4,xmm5
+	movdqa    xmm1,xmm5
+	punpcklwd xmm4,xmm3
+	punpckhwd xmm1,xmm3
+	movdqa    xmm5,xmm4
+	movdqa    xmm3,xmm1
+	pmaddwd   xmm4,[rel PW_MF050_MF256]	; xmm4=tmp5L
+	pmaddwd   xmm1,[rel PW_MF050_MF256]	; xmm1=tmp5H
+	pmaddwd   xmm5,[rel PW_MF256_F050]	; xmm5=tmp6L
+	pmaddwd   xmm3,[rel PW_MF256_F050]	; xmm3=tmp6H
+
+	paddd	xmm4,xmm6		; xmm4=data5L
+	paddd	xmm1,xmm0		; xmm1=data5H
+	paddd	xmm5, XMMWORD [wk(0)]	; xmm5=data3L
+	paddd	xmm3, XMMWORD [wk(1)]	; xmm3=data3H
+
+	paddd	xmm4,[rel PD_DESCALE_P1]
+	paddd	xmm1,[rel PD_DESCALE_P1]
+	psrad	xmm4,DESCALE_P1
+	psrad	xmm1,DESCALE_P1
+	paddd	xmm5,[rel PD_DESCALE_P1]
+	paddd	xmm3,[rel PD_DESCALE_P1]
+	psrad	xmm5,DESCALE_P1
+	psrad	xmm3,DESCALE_P1
+
+	packssdw  xmm4,xmm1		; xmm4=data5
+	packssdw  xmm5,xmm3		; xmm5=data3
+
+	; ---- Pass 2: process columns.
+
+	movdqa	xmm6, XMMWORD [wk(2)]	; xmm6=col0
+	movdqa	xmm0, XMMWORD [wk(4)]	; xmm0=col2
+
+	; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
+	; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
+
+	movdqa    xmm1,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm2		; xmm6=(00 01 10 11 20 21 30 31)
+	punpckhwd xmm1,xmm2		; xmm1=(40 41 50 51 60 61 70 71)
+	movdqa    xmm3,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm5		; xmm0=(02 03 12 13 22 23 32 33)
+	punpckhwd xmm3,xmm5		; xmm3=(42 43 52 53 62 63 72 73)
+
+	movdqa	xmm2, XMMWORD [wk(3)]	; xmm2=col4
+	movdqa	xmm5, XMMWORD [wk(5)]	; xmm5=col6
+
+	; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
+	; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
+
+	movdqa	XMMWORD [wk(0)], xmm0	; wk(0)=(02 03 12 13 22 23 32 33)
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=(42 43 52 53 62 63 72 73)
+
+	movdqa    xmm0,xmm2		; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm4		; xmm2=(04 05 14 15 24 25 34 35)
+	punpckhwd xmm0,xmm4		; xmm0=(44 45 54 55 64 65 74 75)
+	movdqa    xmm3,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm7		; xmm5=(06 07 16 17 26 27 36 37)
+	punpckhwd xmm3,xmm7		; xmm3=(46 47 56 57 66 67 76 77)
+
+	movdqa    xmm4,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm5		; xmm2=(04 05 06 07 14 15 16 17)
+	punpckhdq xmm4,xmm5		; xmm4=(24 25 26 27 34 35 36 37)
+	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm3		; xmm0=(44 45 46 47 54 55 56 57)
+	punpckhdq xmm7,xmm3		; xmm7=(64 65 66 67 74 75 76 77)
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=(02 03 12 13 22 23 32 33)
+	movdqa	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53 62 63 72 73)
+	movdqa	XMMWORD [wk(2)], xmm4	; wk(2)=(24 25 26 27 34 35 36 37)
+	movdqa	XMMWORD [wk(3)], xmm0	; wk(3)=(44 45 46 47 54 55 56 57)
+
+	movdqa    xmm4,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm5		; xmm6=(00 01 02 03 10 11 12 13)
+	punpckhdq xmm4,xmm5		; xmm4=(20 21 22 23 30 31 32 33)
+	movdqa    xmm0,xmm1		; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm3		; xmm1=(40 41 42 43 50 51 52 53)
+	punpckhdq xmm0,xmm3		; xmm0=(60 61 62 63 70 71 72 73)
+
+	movdqa     xmm5,xmm6		; transpose coefficients(phase 3)
+	punpcklqdq xmm6,xmm2		; xmm6=(00 01 02 03 04 05 06 07)=data0
+	punpckhqdq xmm5,xmm2		; xmm5=(10 11 12 13 14 15 16 17)=data1
+	movdqa     xmm3,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm7		; xmm0=(60 61 62 63 64 65 66 67)=data6
+	punpckhqdq xmm3,xmm7		; xmm3=(70 71 72 73 74 75 76 77)=data7
+
+	movdqa	xmm2,xmm5
+	movdqa	xmm7,xmm6
+	psubw	xmm5,xmm0		; xmm5=data1-data6=tmp6
+	psubw	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	paddw	xmm2,xmm0		; xmm2=data1+data6=tmp1
+	paddw	xmm7,xmm3		; xmm7=data0+data7=tmp0
+
+	movdqa	xmm0, XMMWORD [wk(2)]	; xmm0=(24 25 26 27 34 35 36 37)
+	movdqa	xmm3, XMMWORD [wk(3)]	; xmm3=(44 45 46 47 54 55 56 57)
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movdqa     xmm5,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm0		; xmm4=(20 21 22 23 24 25 26 27)=data2
+	punpckhqdq xmm5,xmm0		; xmm5=(30 31 32 33 34 35 36 37)=data3
+	movdqa     xmm6,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm3		; xmm1=(40 41 42 43 44 45 46 47)=data4
+	punpckhqdq xmm6,xmm3		; xmm6=(50 51 52 53 54 55 56 57)=data5
+
+	movdqa	xmm0,xmm5
+	movdqa	xmm3,xmm4
+	paddw	xmm5,xmm1		; xmm5=data3+data4=tmp3
+	paddw	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	psubw	xmm0,xmm1		; xmm0=data3-data4=tmp4
+	psubw	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm1,xmm7
+	movdqa	xmm6,xmm2
+	paddw	xmm7,xmm5		; xmm7=tmp10
+	paddw	xmm2,xmm4		; xmm2=tmp11
+	psubw	xmm1,xmm5		; xmm1=tmp13
+	psubw	xmm6,xmm4		; xmm6=tmp12
+
+	movdqa	xmm5,xmm7
+	paddw	xmm7,xmm2		; xmm7=tmp10+tmp11
+	psubw	xmm5,xmm2		; xmm5=tmp10-tmp11
+
+	paddw	xmm7,[rel PW_DESCALE_P2X]
+	paddw	xmm5,[rel PW_DESCALE_P2X]
+	psraw	xmm7,PASS1_BITS		; xmm7=data0
+	psraw	xmm5,PASS1_BITS		; xmm5=data4
+
+	movdqa	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7
+	movdqa	XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5
+
+	; (Original)
+	; z1 = (tmp12 + tmp13) * 0.541196100;
+	; data2 = z1 + tmp13 * 0.765366865;
+	; data6 = z1 + tmp12 * -1.847759065;
+	;
+	; (This implementation)
+	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+	movdqa    xmm4,xmm1		; xmm1=tmp13
+	movdqa    xmm2,xmm1
+	punpcklwd xmm4,xmm6		; xmm6=tmp12
+	punpckhwd xmm2,xmm6
+	movdqa    xmm1,xmm4
+	movdqa    xmm6,xmm2
+	pmaddwd   xmm4,[rel PW_F130_F054]	; xmm4=data2L
+	pmaddwd   xmm2,[rel PW_F130_F054]	; xmm2=data2H
+	pmaddwd   xmm1,[rel PW_F054_MF130]	; xmm1=data6L
+	pmaddwd   xmm6,[rel PW_F054_MF130]	; xmm6=data6H
+
+	paddd	xmm4,[rel PD_DESCALE_P2]
+	paddd	xmm2,[rel PD_DESCALE_P2]
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm2,DESCALE_P2
+	paddd	xmm1,[rel PD_DESCALE_P2]
+	paddd	xmm6,[rel PD_DESCALE_P2]
+	psrad	xmm1,DESCALE_P2
+	psrad	xmm6,DESCALE_P2
+
+	packssdw  xmm4,xmm2		; xmm4=data2
+	packssdw  xmm1,xmm6		; xmm1=data6
+
+	movdqa	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4
+	movdqa	XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1
+
+	; -- Odd part
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp6
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp7
+
+	movdqa	xmm2,xmm0		; xmm0=tmp4
+	movdqa	xmm6,xmm3		; xmm3=tmp5
+	paddw	xmm2,xmm7		; xmm2=z3
+	paddw	xmm6,xmm5		; xmm6=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm4,xmm2
+	movdqa    xmm1,xmm2
+	punpcklwd xmm4,xmm6
+	punpckhwd xmm1,xmm6
+	movdqa    xmm2,xmm4
+	movdqa    xmm6,xmm1
+	pmaddwd   xmm4,[rel PW_MF078_F117]	; xmm4=z3L
+	pmaddwd   xmm1,[rel PW_MF078_F117]	; xmm1=z3H
+	pmaddwd   xmm2,[rel PW_F117_F078]	; xmm2=z4L
+	pmaddwd   xmm6,[rel PW_F117_F078]	; xmm6=z4H
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=z3L
+	movdqa	XMMWORD [wk(1)], xmm1	; wk(1)=z3H
+
+	; (Original)
+	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+	;
+	; (This implementation)
+	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+	movdqa    xmm4,xmm0
+	movdqa    xmm1,xmm0
+	punpcklwd xmm4,xmm5
+	punpckhwd xmm1,xmm5
+	movdqa    xmm0,xmm4
+	movdqa    xmm5,xmm1
+	pmaddwd   xmm4,[rel PW_MF060_MF089]	; xmm4=tmp4L
+	pmaddwd   xmm1,[rel PW_MF060_MF089]	; xmm1=tmp4H
+	pmaddwd   xmm0,[rel PW_MF089_F060]	; xmm0=tmp7L
+	pmaddwd   xmm5,[rel PW_MF089_F060]	; xmm5=tmp7H
+
+	paddd	xmm4, XMMWORD [wk(0)]	; xmm4=data7L
+	paddd	xmm1, XMMWORD [wk(1)]	; xmm1=data7H
+	paddd	xmm0,xmm2		; xmm0=data1L
+	paddd	xmm5,xmm6		; xmm5=data1H
+
+	paddd	xmm4,[rel PD_DESCALE_P2]
+	paddd	xmm1,[rel PD_DESCALE_P2]
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm1,DESCALE_P2
+	paddd	xmm0,[rel PD_DESCALE_P2]
+	paddd	xmm5,[rel PD_DESCALE_P2]
+	psrad	xmm0,DESCALE_P2
+	psrad	xmm5,DESCALE_P2
+
+	packssdw  xmm4,xmm1		; xmm4=data7
+	packssdw  xmm0,xmm5		; xmm0=data1
+
+	movdqa	XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4
+	movdqa	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0
+
+	movdqa    xmm1,xmm3
+	movdqa    xmm5,xmm3
+	punpcklwd xmm1,xmm7
+	punpckhwd xmm5,xmm7
+	movdqa    xmm3,xmm1
+	movdqa    xmm7,xmm5
+	pmaddwd   xmm1,[rel PW_MF050_MF256]	; xmm1=tmp5L
+	pmaddwd   xmm5,[rel PW_MF050_MF256]	; xmm5=tmp5H
+	pmaddwd   xmm3,[rel PW_MF256_F050]	; xmm3=tmp6L
+	pmaddwd   xmm7,[rel PW_MF256_F050]	; xmm7=tmp6H
+
+	paddd	xmm1,xmm2		; xmm1=data5L
+	paddd	xmm5,xmm6		; xmm5=data5H
+	paddd	xmm3, XMMWORD [wk(0)]	; xmm3=data3L
+	paddd	xmm7, XMMWORD [wk(1)]	; xmm7=data3H
+
+	paddd	xmm1,[rel PD_DESCALE_P2]
+	paddd	xmm5,[rel PD_DESCALE_P2]
+	psrad	xmm1,DESCALE_P2
+	psrad	xmm5,DESCALE_P2
+	paddd	xmm3,[rel PD_DESCALE_P2]
+	paddd	xmm7,[rel PD_DESCALE_P2]
+	psrad	xmm3,DESCALE_P2
+	psrad	xmm7,DESCALE_P2
+
+	packssdw  xmm1,xmm5		; xmm1=data5
+	packssdw  xmm3,xmm7		; xmm3=data3
+
+	movdqa	XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1
+	movdqa	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3
+
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfss2int.asm b/simd/jfss2int.asm
new file mode 100644
index 0000000..5e3f2aa
--- /dev/null
+++ b/simd/jfss2int.asm
@@ -0,0 +1,634 @@
+;
+; jfss2int.asm - accurate integer FDCT (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slow-but-accurate integer implementation of the
+; forward DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jfdctint.c; see the jfdctint.c for
+; more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
+%define DESCALE_P2	(CONST_BITS+PASS1_BITS)
+
+%if CONST_BITS == 13
+F_0_298	equ	 2446		; FIX(0.298631336)
+F_0_390	equ	 3196		; FIX(0.390180644)
+F_0_541	equ	 4433		; FIX(0.541196100)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_175	equ	 9633		; FIX(1.175875602)
+F_1_501	equ	12299		; FIX(1.501321110)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_1_961	equ	16069		; FIX(1.961570560)
+F_2_053	equ	16819		; FIX(2.053119869)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_072	equ	25172		; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
+F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
+F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
+F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_islow_sse2)
+
+EXTN(jconst_fdct_islow_sse2):
+
+PW_F130_F054	times 4 dw  (F_0_541+F_0_765), F_0_541
+PW_F054_MF130	times 4 dw  F_0_541, (F_0_541-F_1_847)
+PW_MF078_F117	times 4 dw  (F_1_175-F_1_961), F_1_175
+PW_F117_F078	times 4 dw  F_1_175, (F_1_175-F_0_390)
+PW_MF060_MF089	times 4 dw  (F_0_298-F_0_899),-F_0_899
+PW_MF089_F060	times 4 dw -F_0_899, (F_1_501-F_0_899)
+PW_MF050_MF256	times 4 dw  (F_2_053-F_2_562),-F_2_562
+PW_MF256_F050	times 4 dw -F_2_562, (F_3_072-F_2_562)
+PD_DESCALE_P1	times 4 dd  1 << (DESCALE_P1-1)
+PD_DESCALE_P2	times 4 dd  1 << (DESCALE_P2-1)
+PW_DESCALE_P2X	times 8 dw  1 << (PASS1_BITS-1)
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_islow_sse2 (DCTELEM * data)
+;
+
+%define data(b)		(b)+8		; DCTELEM * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		6
+
+	align	16
+	global	EXTN(jsimd_fdct_islow_sse2)
+
+EXTN(jsimd_fdct_islow_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
+
+	; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
+	; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
+
+	movdqa    xmm4,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm1		; xmm0=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm4,xmm1		; xmm4=(04 14 05 15 06 16 07 17)
+	movdqa    xmm5,xmm2		; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm3		; xmm2=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm5,xmm3		; xmm5=(24 34 25 35 26 36 27 37)
+
+	movdqa	xmm6, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm7, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
+
+	; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
+	; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(20 30 21 31 22 32 23 33)
+	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(24 34 25 35 26 36 27 37)
+
+	movdqa    xmm2,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm7		; xmm6=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm2,xmm7		; xmm2=(44 54 45 55 46 56 47 57)
+	movdqa    xmm5,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm3		; xmm1=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm5,xmm3		; xmm5=(64 74 65 75 66 76 67 77)
+
+	movdqa    xmm7,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm1		; xmm6=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm7,xmm1		; xmm7=(42 52 62 72 43 53 63 73)
+	movdqa    xmm3,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm5		; xmm2=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm3,xmm5		; xmm3=(46 56 66 76 47 57 67 77)
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(20 30 21 31 22 32 23 33)
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(24 34 25 35 26 36 27 37)
+	movdqa	XMMWORD [wk(2)], xmm7	; wk(2)=(42 52 62 72 43 53 63 73)
+	movdqa	XMMWORD [wk(3)], xmm2	; wk(3)=(44 54 64 74 45 55 65 75)
+
+	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm1		; xmm0=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm7,xmm1		; xmm7=(02 12 22 32 03 13 23 33)
+	movdqa    xmm2,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm5		; xmm4=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm2,xmm5		; xmm2=(06 16 26 36 07 17 27 37)
+
+	movdqa     xmm1,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm6		; xmm0=(00 10 20 30 40 50 60 70)=data0
+	punpckhqdq xmm1,xmm6		; xmm1=(01 11 21 31 41 51 61 71)=data1
+	movdqa     xmm5,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm3		; xmm2=(06 16 26 36 46 56 66 76)=data6
+	punpckhqdq xmm5,xmm3		; xmm5=(07 17 27 37 47 57 67 77)=data7
+
+	movdqa	xmm6,xmm1
+	movdqa	xmm3,xmm0
+	psubw	xmm1,xmm2		; xmm1=data1-data6=tmp6
+	psubw	xmm0,xmm5		; xmm0=data0-data7=tmp7
+	paddw	xmm6,xmm2		; xmm6=data1+data6=tmp1
+	paddw	xmm3,xmm5		; xmm3=data0+data7=tmp0
+
+	movdqa	xmm2, XMMWORD [wk(2)]	; xmm2=(42 52 62 72 43 53 63 73)
+	movdqa	xmm5, XMMWORD [wk(3)]	; xmm5=(44 54 64 74 45 55 65 75)
+	movdqa	XMMWORD [wk(0)], xmm1	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp7
+
+	movdqa     xmm1,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm2		; xmm7=(02 12 22 32 42 52 62 72)=data2
+	punpckhqdq xmm1,xmm2		; xmm1=(03 13 23 33 43 53 63 73)=data3
+	movdqa     xmm0,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm5		; xmm4=(04 14 24 34 44 54 64 74)=data4
+	punpckhqdq xmm0,xmm5		; xmm0=(05 15 25 35 45 55 65 75)=data5
+
+	movdqa	xmm2,xmm1
+	movdqa	xmm5,xmm7
+	paddw	xmm1,xmm4		; xmm1=data3+data4=tmp3
+	paddw	xmm7,xmm0		; xmm7=data2+data5=tmp2
+	psubw	xmm2,xmm4		; xmm2=data3-data4=tmp4
+	psubw	xmm5,xmm0		; xmm5=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm0,xmm6
+	paddw	xmm3,xmm1		; xmm3=tmp10
+	paddw	xmm6,xmm7		; xmm6=tmp11
+	psubw	xmm4,xmm1		; xmm4=tmp13
+	psubw	xmm0,xmm7		; xmm0=tmp12
+
+	movdqa	xmm1,xmm3
+	paddw	xmm3,xmm6		; xmm3=tmp10+tmp11
+	psubw	xmm1,xmm6		; xmm1=tmp10-tmp11
+
+	psllw	xmm3,PASS1_BITS		; xmm3=data0
+	psllw	xmm1,PASS1_BITS		; xmm1=data4
+
+	movdqa	XMMWORD [wk(2)], xmm3	; wk(2)=data0
+	movdqa	XMMWORD [wk(3)], xmm1	; wk(3)=data4
+
+	; (Original)
+	; z1 = (tmp12 + tmp13) * 0.541196100;
+	; data2 = z1 + tmp13 * 0.765366865;
+	; data6 = z1 + tmp12 * -1.847759065;
+	;
+	; (This implementation)
+	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+	movdqa    xmm7,xmm4		; xmm4=tmp13
+	movdqa    xmm6,xmm4
+	punpcklwd xmm7,xmm0		; xmm0=tmp12
+	punpckhwd xmm6,xmm0
+	movdqa    xmm4,xmm7
+	movdqa    xmm0,xmm6
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_F130_F054)]	; xmm7=data2L
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_F130_F054)]	; xmm6=data2H
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_F054_MF130)]	; xmm4=data6L
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_F054_MF130)]	; xmm0=data6H
+
+	paddd	xmm7,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	xmm6,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	xmm7,DESCALE_P1
+	psrad	xmm6,DESCALE_P1
+	paddd	xmm4,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	xmm0,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	xmm4,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+
+	packssdw  xmm7,xmm6		; xmm7=data2
+	packssdw  xmm4,xmm0		; xmm4=data6
+
+	movdqa	XMMWORD [wk(4)], xmm7	; wk(4)=data2
+	movdqa	XMMWORD [wk(5)], xmm4	; wk(5)=data6
+
+	; -- Odd part
+
+	movdqa	xmm3, XMMWORD [wk(0)]	; xmm3=tmp6
+	movdqa	xmm1, XMMWORD [wk(1)]	; xmm1=tmp7
+
+	movdqa	xmm6,xmm2		; xmm2=tmp4
+	movdqa	xmm0,xmm5		; xmm5=tmp5
+	paddw	xmm6,xmm3		; xmm6=z3
+	paddw	xmm0,xmm1		; xmm0=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm7,xmm6
+	movdqa    xmm4,xmm6
+	punpcklwd xmm7,xmm0
+	punpckhwd xmm4,xmm0
+	movdqa    xmm6,xmm7
+	movdqa    xmm0,xmm4
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF078_F117)]	; xmm7=z3L
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_MF078_F117)]	; xmm4=z3H
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_F117_F078)]	; xmm6=z4L
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_F117_F078)]	; xmm0=z4H
+
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=z3L
+	movdqa	XMMWORD [wk(1)], xmm4	; wk(1)=z3H
+
+	; (Original)
+	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+	;
+	; (This implementation)
+	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+	movdqa    xmm7,xmm2
+	movdqa    xmm4,xmm2
+	punpcklwd xmm7,xmm1
+	punpckhwd xmm4,xmm1
+	movdqa    xmm2,xmm7
+	movdqa    xmm1,xmm4
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm7=tmp4L
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm4=tmp4H
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_MF089_F060)]	; xmm2=tmp7L
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF089_F060)]	; xmm1=tmp7H
+
+	paddd	xmm7, XMMWORD [wk(0)]	; xmm7=data7L
+	paddd	xmm4, XMMWORD [wk(1)]	; xmm4=data7H
+	paddd	xmm2,xmm6		; xmm2=data1L
+	paddd	xmm1,xmm0		; xmm1=data1H
+
+	paddd	xmm7,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	xmm4,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	xmm7,DESCALE_P1
+	psrad	xmm4,DESCALE_P1
+	paddd	xmm2,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	xmm1,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	xmm2,DESCALE_P1
+	psrad	xmm1,DESCALE_P1
+
+	packssdw  xmm7,xmm4		; xmm7=data7
+	packssdw  xmm2,xmm1		; xmm2=data1
+
+	movdqa    xmm4,xmm5
+	movdqa    xmm1,xmm5
+	punpcklwd xmm4,xmm3
+	punpckhwd xmm1,xmm3
+	movdqa    xmm5,xmm4
+	movdqa    xmm3,xmm1
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm4=tmp5L
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm1=tmp5H
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_MF256_F050)]	; xmm5=tmp6L
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_MF256_F050)]	; xmm3=tmp6H
+
+	paddd	xmm4,xmm6		; xmm4=data5L
+	paddd	xmm1,xmm0		; xmm1=data5H
+	paddd	xmm5, XMMWORD [wk(0)]	; xmm5=data3L
+	paddd	xmm3, XMMWORD [wk(1)]	; xmm3=data3H
+
+	paddd	xmm4,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	xmm1,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	xmm4,DESCALE_P1
+	psrad	xmm1,DESCALE_P1
+	paddd	xmm5,[GOTOFF(ebx,PD_DESCALE_P1)]
+	paddd	xmm3,[GOTOFF(ebx,PD_DESCALE_P1)]
+	psrad	xmm5,DESCALE_P1
+	psrad	xmm3,DESCALE_P1
+
+	packssdw  xmm4,xmm1		; xmm4=data5
+	packssdw  xmm5,xmm3		; xmm5=data3
+
+	; ---- Pass 2: process columns.
+
+;	mov	edx, POINTER [data(eax)]	; (DCTELEM *)
+
+	movdqa	xmm6, XMMWORD [wk(2)]	; xmm6=col0
+	movdqa	xmm0, XMMWORD [wk(4)]	; xmm0=col2
+
+	; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
+	; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
+
+	movdqa    xmm1,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm2		; xmm6=(00 01 10 11 20 21 30 31)
+	punpckhwd xmm1,xmm2		; xmm1=(40 41 50 51 60 61 70 71)
+	movdqa    xmm3,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm5		; xmm0=(02 03 12 13 22 23 32 33)
+	punpckhwd xmm3,xmm5		; xmm3=(42 43 52 53 62 63 72 73)
+
+	movdqa	xmm2, XMMWORD [wk(3)]	; xmm2=col4
+	movdqa	xmm5, XMMWORD [wk(5)]	; xmm5=col6
+
+	; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
+	; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
+
+	movdqa	XMMWORD [wk(0)], xmm0	; wk(0)=(02 03 12 13 22 23 32 33)
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=(42 43 52 53 62 63 72 73)
+
+	movdqa    xmm0,xmm2		; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm4		; xmm2=(04 05 14 15 24 25 34 35)
+	punpckhwd xmm0,xmm4		; xmm0=(44 45 54 55 64 65 74 75)
+	movdqa    xmm3,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm7		; xmm5=(06 07 16 17 26 27 36 37)
+	punpckhwd xmm3,xmm7		; xmm3=(46 47 56 57 66 67 76 77)
+
+	movdqa    xmm4,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm5		; xmm2=(04 05 06 07 14 15 16 17)
+	punpckhdq xmm4,xmm5		; xmm4=(24 25 26 27 34 35 36 37)
+	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm3		; xmm0=(44 45 46 47 54 55 56 57)
+	punpckhdq xmm7,xmm3		; xmm7=(64 65 66 67 74 75 76 77)
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=(02 03 12 13 22 23 32 33)
+	movdqa	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53 62 63 72 73)
+	movdqa	XMMWORD [wk(2)], xmm4	; wk(2)=(24 25 26 27 34 35 36 37)
+	movdqa	XMMWORD [wk(3)], xmm0	; wk(3)=(44 45 46 47 54 55 56 57)
+
+	movdqa    xmm4,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm5		; xmm6=(00 01 02 03 10 11 12 13)
+	punpckhdq xmm4,xmm5		; xmm4=(20 21 22 23 30 31 32 33)
+	movdqa    xmm0,xmm1		; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm3		; xmm1=(40 41 42 43 50 51 52 53)
+	punpckhdq xmm0,xmm3		; xmm0=(60 61 62 63 70 71 72 73)
+
+	movdqa     xmm5,xmm6		; transpose coefficients(phase 3)
+	punpcklqdq xmm6,xmm2		; xmm6=(00 01 02 03 04 05 06 07)=data0
+	punpckhqdq xmm5,xmm2		; xmm5=(10 11 12 13 14 15 16 17)=data1
+	movdqa     xmm3,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm7		; xmm0=(60 61 62 63 64 65 66 67)=data6
+	punpckhqdq xmm3,xmm7		; xmm3=(70 71 72 73 74 75 76 77)=data7
+
+	movdqa	xmm2,xmm5
+	movdqa	xmm7,xmm6
+	psubw	xmm5,xmm0		; xmm5=data1-data6=tmp6
+	psubw	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	paddw	xmm2,xmm0		; xmm2=data1+data6=tmp1
+	paddw	xmm7,xmm3		; xmm7=data0+data7=tmp0
+
+	movdqa	xmm0, XMMWORD [wk(2)]	; xmm0=(24 25 26 27 34 35 36 37)
+	movdqa	xmm3, XMMWORD [wk(3)]	; xmm3=(44 45 46 47 54 55 56 57)
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=tmp6
+	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movdqa     xmm5,xmm4		; transpose coefficients(phase 3)
+	punpcklqdq xmm4,xmm0		; xmm4=(20 21 22 23 24 25 26 27)=data2
+	punpckhqdq xmm5,xmm0		; xmm5=(30 31 32 33 34 35 36 37)=data3
+	movdqa     xmm6,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm3		; xmm1=(40 41 42 43 44 45 46 47)=data4
+	punpckhqdq xmm6,xmm3		; xmm6=(50 51 52 53 54 55 56 57)=data5
+
+	movdqa	xmm0,xmm5
+	movdqa	xmm3,xmm4
+	paddw	xmm5,xmm1		; xmm5=data3+data4=tmp3
+	paddw	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	psubw	xmm0,xmm1		; xmm0=data3-data4=tmp4
+	psubw	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movdqa	xmm1,xmm7
+	movdqa	xmm6,xmm2
+	paddw	xmm7,xmm5		; xmm7=tmp10
+	paddw	xmm2,xmm4		; xmm2=tmp11
+	psubw	xmm1,xmm5		; xmm1=tmp13
+	psubw	xmm6,xmm4		; xmm6=tmp12
+
+	movdqa	xmm5,xmm7
+	paddw	xmm7,xmm2		; xmm7=tmp10+tmp11
+	psubw	xmm5,xmm2		; xmm5=tmp10-tmp11
+
+	paddw	xmm7,[GOTOFF(ebx,PW_DESCALE_P2X)]
+	paddw	xmm5,[GOTOFF(ebx,PW_DESCALE_P2X)]
+	psraw	xmm7,PASS1_BITS		; xmm7=data0
+	psraw	xmm5,PASS1_BITS		; xmm5=data4
+
+	movdqa	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)], xmm7
+	movdqa	XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)], xmm5
+
+	; (Original)
+	; z1 = (tmp12 + tmp13) * 0.541196100;
+	; data2 = z1 + tmp13 * 0.765366865;
+	; data6 = z1 + tmp12 * -1.847759065;
+	;
+	; (This implementation)
+	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
+	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
+
+	movdqa    xmm4,xmm1		; xmm1=tmp13
+	movdqa    xmm2,xmm1
+	punpcklwd xmm4,xmm6		; xmm6=tmp12
+	punpckhwd xmm2,xmm6
+	movdqa    xmm1,xmm4
+	movdqa    xmm6,xmm2
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_F130_F054)]	; xmm4=data2L
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_F130_F054)]	; xmm2=data2H
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_F054_MF130)]	; xmm1=data6L
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_F054_MF130)]	; xmm6=data6H
+
+	paddd	xmm4,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	xmm2,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm2,DESCALE_P2
+	paddd	xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	xmm6,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	xmm1,DESCALE_P2
+	psrad	xmm6,DESCALE_P2
+
+	packssdw  xmm4,xmm2		; xmm4=data2
+	packssdw  xmm1,xmm6		; xmm1=data6
+
+	movdqa	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)], xmm4
+	movdqa	XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)], xmm1
+
+	; -- Odd part
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp6
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp7
+
+	movdqa	xmm2,xmm0		; xmm0=tmp4
+	movdqa	xmm6,xmm3		; xmm3=tmp5
+	paddw	xmm2,xmm7		; xmm2=z3
+	paddw	xmm6,xmm5		; xmm6=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm4,xmm2
+	movdqa    xmm1,xmm2
+	punpcklwd xmm4,xmm6
+	punpckhwd xmm1,xmm6
+	movdqa    xmm2,xmm4
+	movdqa    xmm6,xmm1
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_MF078_F117)]	; xmm4=z3L
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF078_F117)]	; xmm1=z3H
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_F117_F078)]	; xmm2=z4L
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_F117_F078)]	; xmm6=z4H
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=z3L
+	movdqa	XMMWORD [wk(1)], xmm1	; wk(1)=z3H
+
+	; (Original)
+	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
+	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
+	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
+	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
+	;
+	; (This implementation)
+	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
+	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
+	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
+	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
+	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
+	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
+
+	movdqa    xmm4,xmm0
+	movdqa    xmm1,xmm0
+	punpcklwd xmm4,xmm5
+	punpckhwd xmm1,xmm5
+	movdqa    xmm0,xmm4
+	movdqa    xmm5,xmm1
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm4=tmp4L
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm1=tmp4H
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF089_F060)]	; xmm0=tmp7L
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_MF089_F060)]	; xmm5=tmp7H
+
+	paddd	xmm4, XMMWORD [wk(0)]	; xmm4=data7L
+	paddd	xmm1, XMMWORD [wk(1)]	; xmm1=data7H
+	paddd	xmm0,xmm2		; xmm0=data1L
+	paddd	xmm5,xmm6		; xmm5=data1H
+
+	paddd	xmm4,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm1,DESCALE_P2
+	paddd	xmm0,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	xmm5,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	xmm0,DESCALE_P2
+	psrad	xmm5,DESCALE_P2
+
+	packssdw  xmm4,xmm1		; xmm4=data7
+	packssdw  xmm0,xmm5		; xmm0=data1
+
+	movdqa	XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)], xmm4
+	movdqa	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)], xmm0
+
+	movdqa    xmm1,xmm3
+	movdqa    xmm5,xmm3
+	punpcklwd xmm1,xmm7
+	punpckhwd xmm5,xmm7
+	movdqa    xmm3,xmm1
+	movdqa    xmm7,xmm5
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm1=tmp5L
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm5=tmp5H
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_MF256_F050)]	; xmm3=tmp6L
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF256_F050)]	; xmm7=tmp6H
+
+	paddd	xmm1,xmm2		; xmm1=data5L
+	paddd	xmm5,xmm6		; xmm5=data5H
+	paddd	xmm3, XMMWORD [wk(0)]	; xmm3=data3L
+	paddd	xmm7, XMMWORD [wk(1)]	; xmm7=data3H
+
+	paddd	xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	xmm5,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	xmm1,DESCALE_P2
+	psrad	xmm5,DESCALE_P2
+	paddd	xmm3,[GOTOFF(ebx,PD_DESCALE_P2)]
+	paddd	xmm7,[GOTOFF(ebx,PD_DESCALE_P2)]
+	psrad	xmm3,DESCALE_P2
+	psrad	xmm7,DESCALE_P2
+
+	packssdw  xmm1,xmm5		; xmm1=data5
+	packssdw  xmm3,xmm7		; xmm3=data3
+
+	movdqa	XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)], xmm1
+	movdqa	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)], xmm3
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfsseflt-64.asm b/simd/jfsseflt-64.asm
new file mode 100644
index 0000000..07245d2
--- /dev/null
+++ b/simd/jfsseflt-64.asm
@@ -0,0 +1,358 @@
+;
+; jfsseflt-64.asm - floating-point FDCT (64-bit SSE)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the forward DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_float_sse)
+
+EXTN(jconst_fdct_float_sse):
+
+PD_0_382	times 4 dd  0.382683432365089771728460
+PD_0_707	times 4 dd  0.707106781186547524400844
+PD_0_541	times 4 dd  0.541196100146196984399723
+PD_1_306	times 4 dd  1.306562964876376527856643
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_float_sse (FAST_FLOAT * data)
+;
+
+; r10 = FAST_FLOAT * data
+
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_float_sse)
+
+EXTN(jsimd_fdct_float_sse):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+
+	; ---- Pass 1: process rows.
+
+	mov	rdx, r10	; (FAST_FLOAT *)
+	mov	rcx, DCTSIZE/4
+.rowloop:
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)]
+
+	; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
+	; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
+
+	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
+	unpcklps xmm0,xmm1		; xmm0=(20 30 21 31)
+	unpckhps xmm4,xmm1		; xmm4=(22 32 23 33)
+	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
+	unpcklps xmm2,xmm3		; xmm2=(24 34 25 35)
+	unpckhps xmm5,xmm3		; xmm5=(26 36 27 37)
+
+	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
+
+	; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
+	; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
+
+	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 32 23 33)
+	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(24 34 25 35)
+
+	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
+	unpcklps xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps xmm4,xmm7		; xmm4=(02 12 03 13)
+	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
+	unpcklps xmm1,xmm3		; xmm1=(04 14 05 15)
+	unpckhps xmm2,xmm3		; xmm2=(06 16 07 17)
+
+	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm0		; xmm6=(00 10 20 30)=data0
+	unpckhps2 xmm7,xmm0		; xmm7=(01 11 21 31)=data1
+	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
+	unpcklps2 xmm2,xmm5		; xmm2=(06 16 26 36)=data6
+	unpckhps2 xmm3,xmm5		; xmm3=(07 17 27 37)=data7
+
+	movaps	xmm0,xmm7
+	movaps	xmm5,xmm6
+	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
+	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
+	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
+
+	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 32 23 33)
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(24 34 25 35)
+	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
+	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(02 12 22 32)=data2
+	unpckhps2 xmm7,xmm2		; xmm7=(03 13 23 33)=data3
+	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm3		; xmm1=(04 14 24 34)=data4
+	unpckhps2 xmm6,xmm3		; xmm6=(05 15 25 35)=data5
+
+	movaps	xmm2,xmm7
+	movaps	xmm3,xmm4
+	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
+	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
+	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movaps	xmm1,xmm5
+	movaps	xmm6,xmm0
+	subps	xmm5,xmm7		; xmm5=tmp13
+	subps	xmm0,xmm4		; xmm0=tmp12
+	addps	xmm1,xmm7		; xmm1=tmp10
+	addps	xmm6,xmm4		; xmm6=tmp11
+
+	addps	xmm0,xmm5
+	mulps	xmm0,[rel PD_0_707] ; xmm0=z1
+
+	movaps	xmm7,xmm1
+	movaps	xmm4,xmm5
+	subps	xmm1,xmm6		; xmm1=data4
+	subps	xmm5,xmm0		; xmm5=data6
+	addps	xmm7,xmm6		; xmm7=data0
+	addps	xmm4,xmm0		; xmm4=data2
+
+	movaps	XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+	; -- Odd part
+
+	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
+	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
+
+	addps	xmm2,xmm3		; xmm2=tmp10
+	addps	xmm3,xmm6		; xmm3=tmp11
+	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
+
+	mulps	xmm3,[rel PD_0_707] ; xmm3=z3
+
+	movaps	xmm1,xmm2		; xmm1=tmp10
+	subps	xmm2,xmm6
+	mulps	xmm2,[rel PD_0_382] ; xmm2=z5
+	mulps	xmm1,[rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+	mulps	xmm6,[rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+	addps	xmm1,xmm2		; xmm1=z2
+	addps	xmm6,xmm2		; xmm6=z4
+
+	movaps	xmm5,xmm0
+	subps	xmm0,xmm3		; xmm0=z13
+	addps	xmm5,xmm3		; xmm5=z11
+
+	movaps	xmm7,xmm0
+	movaps	xmm4,xmm5
+	subps	xmm0,xmm1		; xmm0=data3
+	subps	xmm5,xmm6		; xmm5=data7
+	addps	xmm7,xmm1		; xmm7=data5
+	addps	xmm4,xmm6		; xmm4=data1
+
+	movaps	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+	add	rdx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	rcx
+	jnz	near .rowloop
+
+	; ---- Pass 2: process columns.
+
+	mov	rdx, r10	; (FAST_FLOAT *)
+	mov	rcx, DCTSIZE/4
+.columnloop:
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)]
+
+	; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
+	; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
+
+	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
+	unpcklps xmm0,xmm1		; xmm0=(02 03 12 13)
+	unpckhps xmm4,xmm1		; xmm4=(22 23 32 33)
+	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
+	unpcklps xmm2,xmm3		; xmm2=(42 43 52 53)
+	unpckhps xmm5,xmm3		; xmm5=(62 63 72 73)
+
+	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)]
+
+	; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
+	; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
+
+	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 23 32 33)
+	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(42 43 52 53)
+
+	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
+	unpcklps xmm6,xmm7		; xmm6=(00 01 10 11)
+	unpckhps xmm4,xmm7		; xmm4=(20 21 30 31)
+	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
+	unpcklps xmm1,xmm3		; xmm1=(40 41 50 51)
+	unpckhps xmm2,xmm3		; xmm2=(60 61 70 71)
+
+	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm0		; xmm6=(00 01 02 03)=data0
+	unpckhps2 xmm7,xmm0		; xmm7=(10 11 12 13)=data1
+	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
+	unpcklps2 xmm2,xmm5		; xmm2=(60 61 62 63)=data6
+	unpckhps2 xmm3,xmm5		; xmm3=(70 71 72 73)=data7
+
+	movaps	xmm0,xmm7
+	movaps	xmm5,xmm6
+	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
+	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
+	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
+
+	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 23 32 33)
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53)
+	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
+	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(20 21 22 23)=data2
+	unpckhps2 xmm7,xmm2		; xmm7=(30 31 32 33)=data3
+	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm3		; xmm1=(40 41 42 43)=data4
+	unpckhps2 xmm6,xmm3		; xmm6=(50 51 52 53)=data5
+
+	movaps	xmm2,xmm7
+	movaps	xmm3,xmm4
+	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
+	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
+	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movaps	xmm1,xmm5
+	movaps	xmm6,xmm0
+	subps	xmm5,xmm7		; xmm5=tmp13
+	subps	xmm0,xmm4		; xmm0=tmp12
+	addps	xmm1,xmm7		; xmm1=tmp10
+	addps	xmm6,xmm4		; xmm6=tmp11
+
+	addps	xmm0,xmm5
+	mulps	xmm0,[rel PD_0_707] ; xmm0=z1
+
+	movaps	xmm7,xmm1
+	movaps	xmm4,xmm5
+	subps	xmm1,xmm6		; xmm1=data4
+	subps	xmm5,xmm0		; xmm5=data6
+	addps	xmm7,xmm6		; xmm7=data0
+	addps	xmm4,xmm0		; xmm4=data2
+
+	movaps	XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+	; -- Odd part
+
+	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
+	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
+
+	addps	xmm2,xmm3		; xmm2=tmp10
+	addps	xmm3,xmm6		; xmm3=tmp11
+	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
+
+	mulps	xmm3,[rel PD_0_707] ; xmm3=z3
+
+	movaps	xmm1,xmm2		; xmm1=tmp10
+	subps	xmm2,xmm6
+	mulps	xmm2,[rel PD_0_382] ; xmm2=z5
+	mulps	xmm1,[rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+	mulps	xmm6,[rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+	addps	xmm1,xmm2		; xmm1=z2
+	addps	xmm6,xmm2		; xmm6=z4
+
+	movaps	xmm5,xmm0
+	subps	xmm0,xmm3		; xmm0=z13
+	addps	xmm5,xmm3		; xmm5=z11
+
+	movaps	xmm7,xmm0
+	movaps	xmm4,xmm5
+	subps	xmm0,xmm1		; xmm0=data3
+	subps	xmm5,xmm6		; xmm5=data7
+	addps	xmm7,xmm1		; xmm7=data5
+	addps	xmm4,xmm6		; xmm4=data1
+
+	movaps	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
+
+	add	rdx, byte 4*SIZEOF_FAST_FLOAT
+	dec	rcx
+	jnz	near .columnloop
+
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jfsseflt.asm b/simd/jfsseflt.asm
new file mode 100644
index 0000000..bc54ccc
--- /dev/null
+++ b/simd/jfsseflt.asm
@@ -0,0 +1,370 @@
+;
+; jfsseflt.asm - floating-point FDCT (SSE)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the forward DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_float_sse)
+
+EXTN(jconst_fdct_float_sse):
+
+PD_0_382	times 4 dd  0.382683432365089771728460
+PD_0_707	times 4 dd  0.707106781186547524400844
+PD_0_541	times 4 dd  0.541196100146196984399723
+PD_1_306	times 4 dd  1.306562964876376527856643
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_float_sse (FAST_FLOAT * data)
+;
+
+%define data(b)		(b)+8		; FAST_FLOAT * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_float_sse)
+
+EXTN(jsimd_fdct_float_sse):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.rowloop:
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
+	; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
+
+	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
+	unpcklps xmm0,xmm1		; xmm0=(20 30 21 31)
+	unpckhps xmm4,xmm1		; xmm4=(22 32 23 33)
+	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
+	unpcklps xmm2,xmm3		; xmm2=(24 34 25 35)
+	unpckhps xmm5,xmm3		; xmm5=(26 36 27 37)
+
+	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
+	; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
+
+	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 32 23 33)
+	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(24 34 25 35)
+
+	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
+	unpcklps xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps xmm4,xmm7		; xmm4=(02 12 03 13)
+	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
+	unpcklps xmm1,xmm3		; xmm1=(04 14 05 15)
+	unpckhps xmm2,xmm3		; xmm2=(06 16 07 17)
+
+	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm0		; xmm6=(00 10 20 30)=data0
+	unpckhps2 xmm7,xmm0		; xmm7=(01 11 21 31)=data1
+	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
+	unpcklps2 xmm2,xmm5		; xmm2=(06 16 26 36)=data6
+	unpckhps2 xmm3,xmm5		; xmm3=(07 17 27 37)=data7
+
+	movaps	xmm0,xmm7
+	movaps	xmm5,xmm6
+	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
+	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
+	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
+
+	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 32 23 33)
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(24 34 25 35)
+	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
+	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(02 12 22 32)=data2
+	unpckhps2 xmm7,xmm2		; xmm7=(03 13 23 33)=data3
+	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm3		; xmm1=(04 14 24 34)=data4
+	unpckhps2 xmm6,xmm3		; xmm6=(05 15 25 35)=data5
+
+	movaps	xmm2,xmm7
+	movaps	xmm3,xmm4
+	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
+	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
+	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movaps	xmm1,xmm5
+	movaps	xmm6,xmm0
+	subps	xmm5,xmm7		; xmm5=tmp13
+	subps	xmm0,xmm4		; xmm0=tmp12
+	addps	xmm1,xmm7		; xmm1=tmp10
+	addps	xmm6,xmm4		; xmm6=tmp11
+
+	addps	xmm0,xmm5
+	mulps	xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
+
+	movaps	xmm7,xmm1
+	movaps	xmm4,xmm5
+	subps	xmm1,xmm6		; xmm1=data4
+	subps	xmm5,xmm0		; xmm5=data6
+	addps	xmm7,xmm6		; xmm7=data0
+	addps	xmm4,xmm0		; xmm4=data2
+
+	movaps	XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	; -- Odd part
+
+	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
+	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
+
+	addps	xmm2,xmm3		; xmm2=tmp10
+	addps	xmm3,xmm6		; xmm3=tmp11
+	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
+
+	mulps	xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
+
+	movaps	xmm1,xmm2		; xmm1=tmp10
+	subps	xmm2,xmm6
+	mulps	xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
+	mulps	xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+	mulps	xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+	addps	xmm1,xmm2		; xmm1=z2
+	addps	xmm6,xmm2		; xmm6=z4
+
+	movaps	xmm5,xmm0
+	subps	xmm0,xmm3		; xmm0=z13
+	addps	xmm5,xmm3		; xmm5=z11
+
+	movaps	xmm7,xmm0
+	movaps	xmm4,xmm5
+	subps	xmm0,xmm1		; xmm0=data3
+	subps	xmm5,xmm6		; xmm5=data7
+	addps	xmm7,xmm1		; xmm7=data5
+	addps	xmm4,xmm6		; xmm4=data1
+
+	movaps	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	add	edx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .rowloop
+
+	; ---- Pass 2: process columns.
+
+	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.columnloop:
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
+	; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
+
+	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
+	unpcklps xmm0,xmm1		; xmm0=(02 03 12 13)
+	unpckhps xmm4,xmm1		; xmm4=(22 23 32 33)
+	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
+	unpcklps xmm2,xmm3		; xmm2=(42 43 52 53)
+	unpckhps xmm5,xmm3		; xmm5=(62 63 72 73)
+
+	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
+	; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
+
+	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 23 32 33)
+	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(42 43 52 53)
+
+	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
+	unpcklps xmm6,xmm7		; xmm6=(00 01 10 11)
+	unpckhps xmm4,xmm7		; xmm4=(20 21 30 31)
+	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
+	unpcklps xmm1,xmm3		; xmm1=(40 41 50 51)
+	unpckhps xmm2,xmm3		; xmm2=(60 61 70 71)
+
+	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm0		; xmm6=(00 01 02 03)=data0
+	unpckhps2 xmm7,xmm0		; xmm7=(10 11 12 13)=data1
+	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
+	unpcklps2 xmm2,xmm5		; xmm2=(60 61 62 63)=data6
+	unpckhps2 xmm3,xmm5		; xmm3=(70 71 72 73)=data7
+
+	movaps	xmm0,xmm7
+	movaps	xmm5,xmm6
+	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
+	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
+	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
+
+	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 23 32 33)
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53)
+	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
+	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(20 21 22 23)=data2
+	unpckhps2 xmm7,xmm2		; xmm7=(30 31 32 33)=data3
+	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm3		; xmm1=(40 41 42 43)=data4
+	unpckhps2 xmm6,xmm3		; xmm6=(50 51 52 53)=data5
+
+	movaps	xmm2,xmm7
+	movaps	xmm3,xmm4
+	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
+	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
+	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movaps	xmm1,xmm5
+	movaps	xmm6,xmm0
+	subps	xmm5,xmm7		; xmm5=tmp13
+	subps	xmm0,xmm4		; xmm0=tmp12
+	addps	xmm1,xmm7		; xmm1=tmp10
+	addps	xmm6,xmm4		; xmm6=tmp11
+
+	addps	xmm0,xmm5
+	mulps	xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
+
+	movaps	xmm7,xmm1
+	movaps	xmm4,xmm5
+	subps	xmm1,xmm6		; xmm1=data4
+	subps	xmm5,xmm0		; xmm5=data6
+	addps	xmm7,xmm6		; xmm7=data0
+	addps	xmm4,xmm0		; xmm4=data2
+
+	movaps	XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	; -- Odd part
+
+	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
+	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
+
+	addps	xmm2,xmm3		; xmm2=tmp10
+	addps	xmm3,xmm6		; xmm3=tmp11
+	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
+
+	mulps	xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
+
+	movaps	xmm1,xmm2		; xmm1=tmp10
+	subps	xmm2,xmm6
+	mulps	xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
+	mulps	xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+	mulps	xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+	addps	xmm1,xmm2		; xmm1=z2
+	addps	xmm6,xmm2		; xmm6=z4
+
+	movaps	xmm5,xmm0
+	subps	xmm0,xmm3		; xmm0=z13
+	addps	xmm5,xmm3		; xmm5=z11
+
+	movaps	xmm7,xmm0
+	movaps	xmm4,xmm5
+	subps	xmm0,xmm1		; xmm0=data3
+	subps	xmm5,xmm6		; xmm5=data7
+	addps	xmm7,xmm1		; xmm7=data5
+	addps	xmm4,xmm6		; xmm4=data1
+
+	movaps	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	add	edx, byte 4*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .columnloop
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/ji3dnflt.asm b/simd/ji3dnflt.asm
new file mode 100644
index 0000000..dc2076f
--- /dev/null
+++ b/simd/ji3dnflt.asm
@@ -0,0 +1,452 @@
+;
+; ji3dnflt.asm - floating-point IDCT (3DNow! & MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the inverse DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jidctflt.c; see the jidctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_float_3dnow)
+
+EXTN(jconst_idct_float_3dnow):
+
+PD_1_414	times 2 dd  1.414213562373095048801689
+PD_1_847	times 2 dd  1.847759065022573512256366
+PD_1_082	times 2 dd  1.082392200292393968799446
+PD_2_613	times 2 dd  2.613125929752753055713286
+PD_RNDINT_MAGIC	times 2 dd  100663296.0	; (float)(0x00C00000 << 3)
+PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_float_3dnow (void * dct_table, JCOEFPTR coef_block,
+;                         JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
+					; FAST_FLOAT workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_float_3dnow)
+
+EXTN(jsimd_idct_float_3dnow):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; FAST_FLOAT * wsptr
+	mov	ecx, DCTSIZE/2				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_FLOAT_3DNOW
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	short .columnDCT
+
+	pushpic	ebx		; save GOT address
+	mov	ebx, DWORD [DWBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	mov	eax, DWORD [DWBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	or	ebx, DWORD [DWBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	or	ebx, DWORD [DWBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	or	eax,ebx
+	poppic	ebx		; restore GOT address
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movd      mm0, DWORD [DWBLOCK(0,0,esi,SIZEOF_JCOEF)]
+
+	punpcklwd mm0,mm0
+	psrad     mm0,(DWORD_BIT-WORD_BIT)
+	pi2fd     mm0,mm0
+
+	pfmul     mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movq      mm1,mm0
+	punpckldq mm0,mm0
+	punpckhdq mm1,mm1
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm0
+	movq	MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm0
+	movq	MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm1
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm1
+	movq	MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm1
+	movq	MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm1
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movd      mm0, DWORD [DWBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movd      mm1, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movd      mm2, DWORD [DWBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movd      mm3, DWORD [DWBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	punpcklwd mm0,mm0
+	punpcklwd mm1,mm1
+	psrad     mm0,(DWORD_BIT-WORD_BIT)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)
+	pi2fd     mm0,mm0
+	pi2fd     mm1,mm1
+
+	pfmul     mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	pfmul     mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	punpcklwd mm2,mm2
+	punpcklwd mm3,mm3
+	psrad     mm2,(DWORD_BIT-WORD_BIT)
+	psrad     mm3,(DWORD_BIT-WORD_BIT)
+	pi2fd     mm2,mm2
+	pi2fd     mm3,mm3
+
+	pfmul     mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	pfmul     mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movq	mm4,mm0
+	movq	mm5,mm1
+	pfsub	mm0,mm2			; mm0=tmp11
+	pfsub	mm1,mm3
+	pfadd	mm4,mm2			; mm4=tmp10
+	pfadd	mm5,mm3			; mm5=tmp13
+
+	pfmul	mm1,[GOTOFF(ebx,PD_1_414)]
+	pfsub	mm1,mm5			; mm1=tmp12
+
+	movq	mm6,mm4
+	movq	mm7,mm0
+	pfsub	mm4,mm5			; mm4=tmp3
+	pfsub	mm0,mm1			; mm0=tmp2
+	pfadd	mm6,mm5			; mm6=tmp0
+	pfadd	mm7,mm1			; mm7=tmp1
+
+	movq	MMWORD [wk(1)], mm4	; tmp3
+	movq	MMWORD [wk(0)], mm0	; tmp2
+
+	; -- Odd part
+
+	movd      mm2, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movd      mm3, DWORD [DWBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movd      mm5, DWORD [DWBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movd      mm1, DWORD [DWBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	punpcklwd mm2,mm2
+	punpcklwd mm3,mm3
+	psrad     mm2,(DWORD_BIT-WORD_BIT)
+	psrad     mm3,(DWORD_BIT-WORD_BIT)
+	pi2fd     mm2,mm2
+	pi2fd     mm3,mm3
+
+	pfmul     mm2, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	pfmul     mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	punpcklwd mm5,mm5
+	punpcklwd mm1,mm1
+	psrad     mm5,(DWORD_BIT-WORD_BIT)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)
+	pi2fd     mm5,mm5
+	pi2fd     mm1,mm1
+
+	pfmul     mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	pfmul     mm1, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movq	mm4,mm2
+	movq	mm0,mm5
+	pfadd	mm2,mm1			; mm2=z11
+	pfadd	mm5,mm3			; mm5=z13
+	pfsub	mm4,mm1			; mm4=z12
+	pfsub	mm0,mm3			; mm0=z10
+
+	movq	mm1,mm2
+	pfsub	mm2,mm5
+	pfadd	mm1,mm5			; mm1=tmp7
+
+	pfmul	mm2,[GOTOFF(ebx,PD_1_414)]	; mm2=tmp11
+
+	movq	mm3,mm0
+	pfadd	mm0,mm4
+	pfmul	mm0,[GOTOFF(ebx,PD_1_847)]	; mm0=z5
+	pfmul	mm3,[GOTOFF(ebx,PD_2_613)]	; mm3=(z10 * 2.613125930)
+	pfmul	mm4,[GOTOFF(ebx,PD_1_082)]	; mm4=(z12 * 1.082392200)
+	pfsubr	mm3,mm0			; mm3=tmp12
+	pfsub	mm4,mm0			; mm4=tmp10
+
+	; -- Final output stage
+
+	pfsub	mm3,mm1			; mm3=tmp6
+	movq	mm5,mm6
+	movq	mm0,mm7
+	pfadd	mm6,mm1			; mm6=data0=(00 01)
+	pfadd	mm7,mm3			; mm7=data1=(10 11)
+	pfsub	mm5,mm1			; mm5=data7=(70 71)
+	pfsub	mm0,mm3			; mm0=data6=(60 61)
+	pfsub	mm2,mm3			; mm2=tmp5
+
+	movq      mm1,mm6		; transpose coefficients
+	punpckldq mm6,mm7		; mm6=(00 10)
+	punpckhdq mm1,mm7		; mm1=(01 11)
+	movq      mm3,mm0		; transpose coefficients
+	punpckldq mm0,mm5		; mm0=(60 70)
+	punpckhdq mm3,mm5		; mm3=(61 71)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm6
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm1
+	movq	MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
+	movq	MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm3
+
+	movq	mm7, MMWORD [wk(0)]	; mm7=tmp2
+	movq	mm5, MMWORD [wk(1)]	; mm5=tmp3
+
+	pfadd	mm4,mm2			; mm4=tmp4
+	movq	mm6,mm7
+	movq	mm1,mm5
+	pfadd	mm7,mm2			; mm7=data2=(20 21)
+	pfadd	mm5,mm4			; mm5=data4=(40 41)
+	pfsub	mm6,mm2			; mm6=data5=(50 51)
+	pfsub	mm1,mm4			; mm1=data3=(30 31)
+
+	movq      mm0,mm7		; transpose coefficients
+	punpckldq mm7,mm1		; mm7=(20 30)
+	punpckhdq mm0,mm1		; mm0=(21 31)
+	movq      mm3,mm5		; transpose coefficients
+	punpckldq mm5,mm6		; mm5=(40 50)
+	punpckhdq mm3,mm6		; mm3=(41 51)
+
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm7
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm0
+	movq	MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm5
+	movq	MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm3
+
+.nextcolumn:
+	add	esi, byte 2*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 2*SIZEOF_FLOAT_MULT_TYPE	; quantptr
+	add	edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; -- Prefetch the next coefficient block
+
+	prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
+	prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
+	prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
+	prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; FAST_FLOAT * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+	mov	ecx, DCTSIZE/2				; ctr
+	alignx	16,7
+.rowloop:
+
+	; -- Even part
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movq	mm4,mm0
+	movq	mm5,mm1
+	pfsub	mm0,mm2			; mm0=tmp11
+	pfsub	mm1,mm3
+	pfadd	mm4,mm2			; mm4=tmp10
+	pfadd	mm5,mm3			; mm5=tmp13
+
+	pfmul	mm1,[GOTOFF(ebx,PD_1_414)]
+	pfsub	mm1,mm5			; mm1=tmp12
+
+	movq	mm6,mm4
+	movq	mm7,mm0
+	pfsub	mm4,mm5			; mm4=tmp3
+	pfsub	mm0,mm1			; mm0=tmp2
+	pfadd	mm6,mm5			; mm6=tmp0
+	pfadd	mm7,mm1			; mm7=tmp1
+
+	movq	MMWORD [wk(1)], mm4	; tmp3
+	movq	MMWORD [wk(0)], mm0	; tmp2
+
+	; -- Odd part
+
+	movq	mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
+	movq	mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movq	mm4,mm2
+	movq	mm0,mm5
+	pfadd	mm2,mm1			; mm2=z11
+	pfadd	mm5,mm3			; mm5=z13
+	pfsub	mm4,mm1			; mm4=z12
+	pfsub	mm0,mm3			; mm0=z10
+
+	movq	mm1,mm2
+	pfsub	mm2,mm5
+	pfadd	mm1,mm5			; mm1=tmp7
+
+	pfmul	mm2,[GOTOFF(ebx,PD_1_414)]	; mm2=tmp11
+
+	movq	mm3,mm0
+	pfadd	mm0,mm4
+	pfmul	mm0,[GOTOFF(ebx,PD_1_847)]	; mm0=z5
+	pfmul	mm3,[GOTOFF(ebx,PD_2_613)]	; mm3=(z10 * 2.613125930)
+	pfmul	mm4,[GOTOFF(ebx,PD_1_082)]	; mm4=(z12 * 1.082392200)
+	pfsubr	mm3,mm0			; mm3=tmp12
+	pfsub	mm4,mm0			; mm4=tmp10
+
+	; -- Final output stage
+
+	pfsub	mm3,mm1			; mm3=tmp6
+	movq	mm5,mm6
+	movq	mm0,mm7
+	pfadd	mm6,mm1			; mm6=data0=(00 10)
+	pfadd	mm7,mm3			; mm7=data1=(01 11)
+	pfsub	mm5,mm1			; mm5=data7=(07 17)
+	pfsub	mm0,mm3			; mm0=data6=(06 16)
+	pfsub	mm2,mm3			; mm2=tmp5
+
+	movq	mm1,[GOTOFF(ebx,PD_RNDINT_MAGIC)]	; mm1=[PD_RNDINT_MAGIC]
+	pcmpeqd	mm3,mm3
+	psrld	mm3,WORD_BIT		; mm3={0xFFFF 0x0000 0xFFFF 0x0000}
+
+	pfadd	mm6,mm1			; mm6=roundint(data0/8)=(00 ** 10 **)
+	pfadd	mm7,mm1			; mm7=roundint(data1/8)=(01 ** 11 **)
+	pfadd	mm0,mm1			; mm0=roundint(data6/8)=(06 ** 16 **)
+	pfadd	mm5,mm1			; mm5=roundint(data7/8)=(07 ** 17 **)
+
+	pand	mm6,mm3			; mm6=(00 -- 10 --)
+	pslld	mm7,WORD_BIT		; mm7=(-- 01 -- 11)
+	pand	mm0,mm3			; mm0=(06 -- 16 --)
+	pslld	mm5,WORD_BIT		; mm5=(-- 07 -- 17)
+	por	mm6,mm7			; mm6=(00 01 10 11)
+	por	mm0,mm5			; mm0=(06 07 16 17)
+
+	movq	mm1, MMWORD [wk(0)]	; mm1=tmp2
+	movq	mm3, MMWORD [wk(1)]	; mm3=tmp3
+
+	pfadd	mm4,mm2			; mm4=tmp4
+	movq	mm7,mm1
+	movq	mm5,mm3
+	pfadd	mm1,mm2			; mm1=data2=(02 12)
+	pfadd	mm3,mm4			; mm3=data4=(04 14)
+	pfsub	mm7,mm2			; mm7=data5=(05 15)
+	pfsub	mm5,mm4			; mm5=data3=(03 13)
+
+	movq	mm2,[GOTOFF(ebx,PD_RNDINT_MAGIC)]	; mm2=[PD_RNDINT_MAGIC]
+	pcmpeqd	mm4,mm4
+	psrld	mm4,WORD_BIT		; mm4={0xFFFF 0x0000 0xFFFF 0x0000}
+
+	pfadd	mm3,mm2			; mm3=roundint(data4/8)=(04 ** 14 **)
+	pfadd	mm7,mm2			; mm7=roundint(data5/8)=(05 ** 15 **)
+	pfadd	mm1,mm2			; mm1=roundint(data2/8)=(02 ** 12 **)
+	pfadd	mm5,mm2			; mm5=roundint(data3/8)=(03 ** 13 **)
+
+	pand	mm3,mm4			; mm3=(04 -- 14 --)
+	pslld	mm7,WORD_BIT		; mm7=(-- 05 -- 15)
+	pand	mm1,mm4			; mm1=(02 -- 12 --)
+	pslld	mm5,WORD_BIT		; mm5=(-- 03 -- 13)
+	por	mm3,mm7			; mm3=(04 05 14 15)
+	por	mm1,mm5			; mm1=(02 03 12 13)
+
+	movq      mm2,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm2=[PB_CENTERJSAMP]
+
+	packsswb  mm6,mm3		; mm6=(00 01 10 11 04 05 14 15)
+	packsswb  mm1,mm0		; mm1=(02 03 12 13 06 07 16 17)
+	paddb     mm6,mm2
+	paddb     mm1,mm2
+
+	movq      mm4,mm6		; transpose coefficients(phase 2)
+	punpcklwd mm6,mm1		; mm6=(00 01 02 03 10 11 12 13)
+	punpckhwd mm4,mm1		; mm4=(04 05 06 07 14 15 16 17)
+
+	movq      mm7,mm6		; transpose coefficients(phase 3)
+	punpckldq mm6,mm4		; mm6=(00 01 02 03 04 05 06 07)
+	punpckhdq mm7,mm4		; mm7=(10 11 12 13 14 15 16 17)
+
+	pushpic	ebx			; save GOT address
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm6
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm7
+
+	poppic	ebx			; restore GOT address
+
+	add	esi, byte 2*SIZEOF_FAST_FLOAT	; wsptr
+	add	edi, byte 2*SIZEOF_JSAMPROW
+	dec	ecx				; ctr
+	jnz	near .rowloop
+
+	femms		; empty MMX/3DNow! state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jimmxfst.asm b/simd/jimmxfst.asm
new file mode 100644
index 0000000..3b05572
--- /dev/null
+++ b/simd/jimmxfst.asm
@@ -0,0 +1,500 @@
+;
+; jimmxfst.asm - fast integer IDCT (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the inverse DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jidctfst.c; see the jidctfst.c
+; for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	8	; 14 is also OK.
+%define PASS1_BITS	2
+
+%if IFAST_SCALE_BITS != PASS1_BITS
+%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
+%endif
+
+%if CONST_BITS == 8
+F_1_082	equ	277		; FIX(1.082392200)
+F_1_414	equ	362		; FIX(1.414213562)
+F_1_847	equ	473		; FIX(1.847759065)
+F_2_613	equ	669		; FIX(2.613125930)
+F_1_613	equ	(F_2_613 - 256)	; FIX(2.613125930) - FIX(1)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define	DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_1_082	equ	DESCALE(1162209775,30-CONST_BITS)	; FIX(1.082392200)
+F_1_414	equ	DESCALE(1518500249,30-CONST_BITS)	; FIX(1.414213562)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_2_613	equ	DESCALE(2805822602,30-CONST_BITS)	; FIX(2.613125930)
+F_1_613	equ	(F_2_613 - (1 << CONST_BITS))	; FIX(2.613125930) - FIX(1)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS   2
+%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+	alignz	16
+	global	EXTN(jconst_idct_ifast_mmx)
+
+EXTN(jconst_idct_ifast_mmx):
+
+PW_F1414	times 4 dw  F_1_414 << CONST_SHIFT
+PW_F1847	times 4 dw  F_1_847 << CONST_SHIFT
+PW_MF1613	times 4 dw -F_1_613 << CONST_SHIFT
+PW_F1082	times 4 dw  F_1_082 << CONST_SHIFT
+PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_ifast_mmx (void * dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_JCOEF
+					; JCOEF workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_ifast_mmx)
+
+EXTN(jsimd_idct_ifast_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; JCOEF * wsptr
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_IFAST_MMX
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	short .columnDCT
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	mm1,mm0
+	packsswb mm1,mm1
+	movd	eax,mm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movq      mm2,mm0		; mm0=in0=(00 01 02 03)
+	punpcklwd mm0,mm0		; mm0=(00 00 01 01)
+	punpckhwd mm2,mm2		; mm2=(02 02 03 03)
+
+	movq      mm1,mm0
+	punpckldq mm0,mm0		; mm0=(00 00 00 00)
+	punpckhdq mm1,mm1		; mm1=(01 01 01 01)
+	movq      mm3,mm2
+	punpckldq mm2,mm2		; mm2=(02 02 02 02)
+	punpckhdq mm3,mm3		; mm3=(03 03 03 03)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
+	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movq	mm4,mm0
+	movq	mm5,mm1
+	psubw	mm0,mm2			; mm0=tmp11
+	psubw	mm1,mm3
+	paddw	mm4,mm2			; mm4=tmp10
+	paddw	mm5,mm3			; mm5=tmp13
+
+	psllw	mm1,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm1,[GOTOFF(ebx,PW_F1414)]
+	psubw	mm1,mm5			; mm1=tmp12
+
+	movq	mm6,mm4
+	movq	mm7,mm0
+	psubw	mm4,mm5			; mm4=tmp3
+	psubw	mm0,mm1			; mm0=tmp2
+	paddw	mm6,mm5			; mm6=tmp0
+	paddw	mm7,mm1			; mm7=tmp1
+
+	movq	MMWORD [wk(1)], mm4	; wk(1)=tmp3
+	movq	MMWORD [wk(0)], mm0	; wk(0)=tmp2
+
+	; -- Odd part
+
+	movq	mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm2, MMWORD [MMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	movq	mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	mm1, MMWORD [MMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movq	mm4,mm2
+	movq	mm0,mm5
+	psubw	mm2,mm1			; mm2=z12
+	psubw	mm5,mm3			; mm5=z10
+	paddw	mm4,mm1			; mm4=z11
+	paddw	mm0,mm3			; mm0=z13
+
+	movq	mm1,mm5			; mm1=z10(unscaled)
+	psllw	mm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	mm5,PRE_MULTIPLY_SCALE_BITS
+
+	movq	mm3,mm4
+	psubw	mm4,mm0
+	paddw	mm3,mm0			; mm3=tmp7
+
+	psllw	mm4,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm4,[GOTOFF(ebx,PW_F1414)]	; mm4=tmp11
+
+	; To avoid overflow...
+	;
+	; (Original)
+	; tmp12 = -2.613125930 * z10 + z5;
+	;
+	; (This implementation)
+	; tmp12 = (-1.613125930 - 1) * z10 + z5;
+	;       = -1.613125930 * z10 - z10 + z5;
+
+	movq	mm0,mm5
+	paddw	mm5,mm2
+	pmulhw	mm5,[GOTOFF(ebx,PW_F1847)]	; mm5=z5
+	pmulhw	mm0,[GOTOFF(ebx,PW_MF1613)]
+	pmulhw	mm2,[GOTOFF(ebx,PW_F1082)]
+	psubw	mm0,mm1
+	psubw	mm2,mm5			; mm2=tmp10
+	paddw	mm0,mm5			; mm0=tmp12
+
+	; -- Final output stage
+
+	psubw	mm0,mm3			; mm0=tmp6
+	movq	mm1,mm6
+	movq	mm5,mm7
+	paddw	mm6,mm3			; mm6=data0=(00 01 02 03)
+	paddw	mm7,mm0			; mm7=data1=(10 11 12 13)
+	psubw	mm1,mm3			; mm1=data7=(70 71 72 73)
+	psubw	mm5,mm0			; mm5=data6=(60 61 62 63)
+	psubw	mm4,mm0			; mm4=tmp5
+
+	movq      mm3,mm6		; transpose coefficients(phase 1)
+	punpcklwd mm6,mm7		; mm6=(00 10 01 11)
+	punpckhwd mm3,mm7		; mm3=(02 12 03 13)
+	movq      mm0,mm5		; transpose coefficients(phase 1)
+	punpcklwd mm5,mm1		; mm5=(60 70 61 71)
+	punpckhwd mm0,mm1		; mm0=(62 72 63 73)
+
+	movq	mm7, MMWORD [wk(0)]	; mm7=tmp2
+	movq	mm1, MMWORD [wk(1)]	; mm1=tmp3
+
+	movq	MMWORD [wk(0)], mm5	; wk(0)=(60 70 61 71)
+	movq	MMWORD [wk(1)], mm0	; wk(1)=(62 72 63 73)
+
+	paddw	mm2,mm4			; mm2=tmp4
+	movq	mm5,mm7
+	movq	mm0,mm1
+	paddw	mm7,mm4			; mm7=data2=(20 21 22 23)
+	paddw	mm1,mm2			; mm1=data4=(40 41 42 43)
+	psubw	mm5,mm4			; mm5=data5=(50 51 52 53)
+	psubw	mm0,mm2			; mm0=data3=(30 31 32 33)
+
+	movq      mm4,mm7		; transpose coefficients(phase 1)
+	punpcklwd mm7,mm0		; mm7=(20 30 21 31)
+	punpckhwd mm4,mm0		; mm4=(22 32 23 33)
+	movq      mm2,mm1		; transpose coefficients(phase 1)
+	punpcklwd mm1,mm5		; mm1=(40 50 41 51)
+	punpckhwd mm2,mm5		; mm2=(42 52 43 53)
+
+	movq      mm0,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm7		; mm6=(00 10 20 30)
+	punpckhdq mm0,mm7		; mm0=(01 11 21 31)
+	movq      mm5,mm3		; transpose coefficients(phase 2)
+	punpckldq mm3,mm4		; mm3=(02 12 22 32)
+	punpckhdq mm5,mm4		; mm5=(03 13 23 33)
+
+	movq	mm7, MMWORD [wk(0)]	; mm7=(60 70 61 71)
+	movq	mm4, MMWORD [wk(1)]	; mm4=(62 72 63 73)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm6
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm3
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5
+
+	movq      mm6,mm1		; transpose coefficients(phase 2)
+	punpckldq mm1,mm7		; mm1=(40 50 60 70)
+	punpckhdq mm6,mm7		; mm6=(41 51 61 71)
+	movq      mm0,mm2		; transpose coefficients(phase 2)
+	punpckldq mm2,mm4		; mm2=(42 52 62 72)
+	punpckhdq mm0,mm4		; mm0=(43 53 63 73)
+
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm6
+	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm0
+
+.nextcolumn:
+	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 4*SIZEOF_IFAST_MULT_TYPE	; quantptr
+	add	edi, byte 4*DCTSIZE*SIZEOF_JCOEF	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; JCOEF * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.rowloop:
+
+	; -- Even part
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	movq	mm4,mm0
+	movq	mm5,mm1
+	psubw	mm0,mm2			; mm0=tmp11
+	psubw	mm1,mm3
+	paddw	mm4,mm2			; mm4=tmp10
+	paddw	mm5,mm3			; mm5=tmp13
+
+	psllw	mm1,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm1,[GOTOFF(ebx,PW_F1414)]
+	psubw	mm1,mm5			; mm1=tmp12
+
+	movq	mm6,mm4
+	movq	mm7,mm0
+	psubw	mm4,mm5			; mm4=tmp3
+	psubw	mm0,mm1			; mm0=tmp2
+	paddw	mm6,mm5			; mm6=tmp0
+	paddw	mm7,mm1			; mm7=tmp1
+
+	movq	MMWORD [wk(1)], mm4	; wk(1)=tmp3
+	movq	MMWORD [wk(0)], mm0	; wk(0)=tmp2
+
+	; -- Odd part
+
+	movq	mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq	mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	movq	mm4,mm2
+	movq	mm0,mm5
+	psubw	mm2,mm1			; mm2=z12
+	psubw	mm5,mm3			; mm5=z10
+	paddw	mm4,mm1			; mm4=z11
+	paddw	mm0,mm3			; mm0=z13
+
+	movq	mm1,mm5			; mm1=z10(unscaled)
+	psllw	mm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	mm5,PRE_MULTIPLY_SCALE_BITS
+
+	movq	mm3,mm4
+	psubw	mm4,mm0
+	paddw	mm3,mm0			; mm3=tmp7
+
+	psllw	mm4,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	mm4,[GOTOFF(ebx,PW_F1414)]	; mm4=tmp11
+
+	; To avoid overflow...
+	;
+	; (Original)
+	; tmp12 = -2.613125930 * z10 + z5;
+	;
+	; (This implementation)
+	; tmp12 = (-1.613125930 - 1) * z10 + z5;
+	;       = -1.613125930 * z10 - z10 + z5;
+
+	movq	mm0,mm5
+	paddw	mm5,mm2
+	pmulhw	mm5,[GOTOFF(ebx,PW_F1847)]	; mm5=z5
+	pmulhw	mm0,[GOTOFF(ebx,PW_MF1613)]
+	pmulhw	mm2,[GOTOFF(ebx,PW_F1082)]
+	psubw	mm0,mm1
+	psubw	mm2,mm5			; mm2=tmp10
+	paddw	mm0,mm5			; mm0=tmp12
+
+	; -- Final output stage
+
+	psubw	mm0,mm3			; mm0=tmp6
+	movq	mm1,mm6
+	movq	mm5,mm7
+	paddw	mm6,mm3			; mm6=data0=(00 10 20 30)
+	paddw	mm7,mm0			; mm7=data1=(01 11 21 31)
+	psraw	mm6,(PASS1_BITS+3)	; descale
+	psraw	mm7,(PASS1_BITS+3)	; descale
+	psubw	mm1,mm3			; mm1=data7=(07 17 27 37)
+	psubw	mm5,mm0			; mm5=data6=(06 16 26 36)
+	psraw	mm1,(PASS1_BITS+3)	; descale
+	psraw	mm5,(PASS1_BITS+3)	; descale
+	psubw	mm4,mm0			; mm4=tmp5
+
+	packsswb  mm6,mm5		; mm6=(00 10 20 30 06 16 26 36)
+	packsswb  mm7,mm1		; mm7=(01 11 21 31 07 17 27 37)
+
+	movq	mm3, MMWORD [wk(0)]	; mm3=tmp2
+	movq	mm0, MMWORD [wk(1)]	; mm0=tmp3
+
+	paddw	mm2,mm4			; mm2=tmp4
+	movq	mm5,mm3
+	movq	mm1,mm0
+	paddw	mm3,mm4			; mm3=data2=(02 12 22 32)
+	paddw	mm0,mm2			; mm0=data4=(04 14 24 34)
+	psraw	mm3,(PASS1_BITS+3)	; descale
+	psraw	mm0,(PASS1_BITS+3)	; descale
+	psubw	mm5,mm4			; mm5=data5=(05 15 25 35)
+	psubw	mm1,mm2			; mm1=data3=(03 13 23 33)
+	psraw	mm5,(PASS1_BITS+3)	; descale
+	psraw	mm1,(PASS1_BITS+3)	; descale
+
+	movq      mm4,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm4=[PB_CENTERJSAMP]
+
+	packsswb  mm3,mm0		; mm3=(02 12 22 32 04 14 24 34)
+	packsswb  mm1,mm5		; mm1=(03 13 23 33 05 15 25 35)
+
+	paddb     mm6,mm4
+	paddb     mm7,mm4
+	paddb     mm3,mm4
+	paddb     mm1,mm4
+
+	movq      mm2,mm6		; transpose coefficients(phase 1)
+	punpcklbw mm6,mm7		; mm6=(00 01 10 11 20 21 30 31)
+	punpckhbw mm2,mm7		; mm2=(06 07 16 17 26 27 36 37)
+	movq      mm0,mm3		; transpose coefficients(phase 1)
+	punpcklbw mm3,mm1		; mm3=(02 03 12 13 22 23 32 33)
+	punpckhbw mm0,mm1		; mm0=(04 05 14 15 24 25 34 35)
+
+	movq      mm5,mm6		; transpose coefficients(phase 2)
+	punpcklwd mm6,mm3		; mm6=(00 01 02 03 10 11 12 13)
+	punpckhwd mm5,mm3		; mm5=(20 21 22 23 30 31 32 33)
+	movq      mm4,mm0		; transpose coefficients(phase 2)
+	punpcklwd mm0,mm2		; mm0=(04 05 06 07 14 15 16 17)
+	punpckhwd mm4,mm2		; mm4=(24 25 26 27 34 35 36 37)
+
+	movq      mm7,mm6		; transpose coefficients(phase 3)
+	punpckldq mm6,mm0		; mm6=(00 01 02 03 04 05 06 07)
+	punpckhdq mm7,mm0		; mm7=(10 11 12 13 14 15 16 17)
+	movq      mm1,mm5		; transpose coefficients(phase 3)
+	punpckldq mm5,mm4		; mm5=(20 21 22 23 24 25 26 27)
+	punpckhdq mm1,mm4		; mm1=(30 31 32 33 34 35 36 37)
+
+	pushpic	ebx			; save GOT address
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm6
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm7
+	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
+
+	poppic	ebx			; restore GOT address
+
+	add	esi, byte 4*SIZEOF_JCOEF	; wsptr
+	add	edi, byte 4*SIZEOF_JSAMPROW
+	dec	ecx				; ctr
+	jnz	near .rowloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jimmxint.asm b/simd/jimmxint.asm
new file mode 100644
index 0000000..7b52fae
--- /dev/null
+++ b/simd/jimmxint.asm
@@ -0,0 +1,852 @@
+;
+; jimmxint.asm - accurate integer IDCT (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slow-but-accurate integer implementation of the
+; inverse DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jidctint.c; see the jidctint.c for
+; more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
+%define DESCALE_P2	(CONST_BITS+PASS1_BITS+3)
+
+%if CONST_BITS == 13
+F_0_298	equ	 2446		; FIX(0.298631336)
+F_0_390	equ	 3196		; FIX(0.390180644)
+F_0_541	equ	 4433		; FIX(0.541196100)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_175	equ	 9633		; FIX(1.175875602)
+F_1_501	equ	12299		; FIX(1.501321110)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_1_961	equ	16069		; FIX(1.961570560)
+F_2_053	equ	16819		; FIX(2.053119869)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_072	equ	25172		; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
+F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
+F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
+F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_islow_mmx)
+
+EXTN(jconst_idct_islow_mmx):
+
+PW_F130_F054	times 2 dw  (F_0_541+F_0_765), F_0_541
+PW_F054_MF130	times 2 dw  F_0_541, (F_0_541-F_1_847)
+PW_MF078_F117	times 2 dw  (F_1_175-F_1_961), F_1_175
+PW_F117_F078	times 2 dw  F_1_175, (F_1_175-F_0_390)
+PW_MF060_MF089	times 2 dw  (F_0_298-F_0_899),-F_0_899
+PW_MF089_F060	times 2 dw -F_0_899, (F_1_501-F_0_899)
+PW_MF050_MF256	times 2 dw  (F_2_053-F_2_562),-F_2_562
+PW_MF256_F050	times 2 dw -F_2_562, (F_3_072-F_2_562)
+PD_DESCALE_P1	times 2 dd  1 << (DESCALE_P1-1)
+PD_DESCALE_P2	times 2 dd  1 << (DESCALE_P2-1)
+PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_islow_mmx (void * dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		12
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_JCOEF
+					; JCOEF workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_islow_mmx)
+
+EXTN(jsimd_idct_islow_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; JCOEF * wsptr
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_ISLOW_MMX
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	short .columnDCT
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	mm1,mm0
+	packsswb mm1,mm1
+	movd	eax,mm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	psllw	mm0,PASS1_BITS
+
+	movq      mm2,mm0		; mm0=in0=(00 01 02 03)
+	punpcklwd mm0,mm0		; mm0=(00 00 01 01)
+	punpckhwd mm2,mm2		; mm2=(02 02 03 03)
+
+	movq      mm1,mm0
+	punpckldq mm0,mm0		; mm0=(00 00 00 00)
+	punpckhdq mm1,mm1		; mm1=(01 01 01 01)
+	movq      mm3,mm2
+	punpckldq mm2,mm2		; mm2=(02 02 02 02)
+	punpckhdq mm3,mm3		; mm3=(03 03 03 03)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
+	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; (Original)
+	; z1 = (z2 + z3) * 0.541196100;
+	; tmp2 = z1 + z3 * -1.847759065;
+	; tmp3 = z1 + z2 * 0.765366865;
+	;
+	; (This implementation)
+	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+	movq      mm4,mm1		; mm1=in2=z2
+	movq      mm5,mm1
+	punpcklwd mm4,mm3		; mm3=in6=z3
+	punpckhwd mm5,mm3
+	movq      mm1,mm4
+	movq      mm3,mm5
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F130_F054)]	; mm4=tmp3L
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F130_F054)]	; mm5=tmp3H
+	pmaddwd   mm1,[GOTOFF(ebx,PW_F054_MF130)]	; mm1=tmp2L
+	pmaddwd   mm3,[GOTOFF(ebx,PW_F054_MF130)]	; mm3=tmp2H
+
+	movq      mm6,mm0
+	paddw     mm0,mm2		; mm0=in0+in4
+	psubw     mm6,mm2		; mm6=in0-in4
+
+	pxor      mm7,mm7
+	pxor      mm2,mm2
+	punpcklwd mm7,mm0		; mm7=tmp0L
+	punpckhwd mm2,mm0		; mm2=tmp0H
+	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
+	psrad     mm2,(16-CONST_BITS)	; psrad mm2,16 & pslld mm2,CONST_BITS
+
+	movq	mm0,mm7
+	paddd	mm7,mm4			; mm7=tmp10L
+	psubd	mm0,mm4			; mm0=tmp13L
+	movq	mm4,mm2
+	paddd	mm2,mm5			; mm2=tmp10H
+	psubd	mm4,mm5			; mm4=tmp13H
+
+	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp10L
+	movq	MMWORD [wk(1)], mm2	; wk(1)=tmp10H
+	movq	MMWORD [wk(2)], mm0	; wk(2)=tmp13L
+	movq	MMWORD [wk(3)], mm4	; wk(3)=tmp13H
+
+	pxor      mm5,mm5
+	pxor      mm7,mm7
+	punpcklwd mm5,mm6		; mm5=tmp1L
+	punpckhwd mm7,mm6		; mm7=tmp1H
+	psrad     mm5,(16-CONST_BITS)	; psrad mm5,16 & pslld mm5,CONST_BITS
+	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
+
+	movq	mm2,mm5
+	paddd	mm5,mm1			; mm5=tmp11L
+	psubd	mm2,mm1			; mm2=tmp12L
+	movq	mm0,mm7
+	paddd	mm7,mm3			; mm7=tmp11H
+	psubd	mm0,mm3			; mm0=tmp12H
+
+	movq	MMWORD [wk(4)], mm5	; wk(4)=tmp11L
+	movq	MMWORD [wk(5)], mm7	; wk(5)=tmp11H
+	movq	MMWORD [wk(6)], mm2	; wk(6)=tmp12L
+	movq	MMWORD [wk(7)], mm0	; wk(7)=tmp12H
+
+	; -- Odd part
+
+	movq	mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm4, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm6, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movq	mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm1, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movq	mm5,mm6
+	movq	mm7,mm4
+	paddw	mm5,mm3			; mm5=z3
+	paddw	mm7,mm1			; mm7=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movq      mm2,mm5
+	movq      mm0,mm5
+	punpcklwd mm2,mm7
+	punpckhwd mm0,mm7
+	movq      mm5,mm2
+	movq      mm7,mm0
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF078_F117)]	; mm2=z3L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_MF078_F117)]	; mm0=z3H
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F117_F078)]	; mm5=z4L
+	pmaddwd   mm7,[GOTOFF(ebx,PW_F117_F078)]	; mm7=z4H
+
+	movq	MMWORD [wk(10)], mm2	; wk(10)=z3L
+	movq	MMWORD [wk(11)], mm0	; wk(11)=z3H
+
+	; (Original)
+	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+	;
+	; (This implementation)
+	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+	; tmp0 += z3;  tmp1 += z4;
+	; tmp2 += z3;  tmp3 += z4;
+
+	movq      mm2,mm3
+	movq      mm0,mm3
+	punpcklwd mm2,mm4
+	punpckhwd mm0,mm4
+	movq      mm3,mm2
+	movq      mm4,mm0
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF060_MF089)]	; mm2=tmp0L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_MF060_MF089)]	; mm0=tmp0H
+	pmaddwd   mm3,[GOTOFF(ebx,PW_MF089_F060)]	; mm3=tmp3L
+	pmaddwd   mm4,[GOTOFF(ebx,PW_MF089_F060)]	; mm4=tmp3H
+
+	paddd	mm2, MMWORD [wk(10)]	; mm2=tmp0L
+	paddd	mm0, MMWORD [wk(11)]	; mm0=tmp0H
+	paddd	mm3,mm5			; mm3=tmp3L
+	paddd	mm4,mm7			; mm4=tmp3H
+
+	movq	MMWORD [wk(8)], mm2	; wk(8)=tmp0L
+	movq	MMWORD [wk(9)], mm0	; wk(9)=tmp0H
+
+	movq      mm2,mm1
+	movq      mm0,mm1
+	punpcklwd mm2,mm6
+	punpckhwd mm0,mm6
+	movq      mm1,mm2
+	movq      mm6,mm0
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF050_MF256)]	; mm2=tmp1L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_MF050_MF256)]	; mm0=tmp1H
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF256_F050)]	; mm1=tmp2L
+	pmaddwd   mm6,[GOTOFF(ebx,PW_MF256_F050)]	; mm6=tmp2H
+
+	paddd	mm2,mm5			; mm2=tmp1L
+	paddd	mm0,mm7			; mm0=tmp1H
+	paddd	mm1, MMWORD [wk(10)]	; mm1=tmp2L
+	paddd	mm6, MMWORD [wk(11)]	; mm6=tmp2H
+
+	movq	MMWORD [wk(10)], mm2	; wk(10)=tmp1L
+	movq	MMWORD [wk(11)], mm0	; wk(11)=tmp1H
+
+	; -- Final output stage
+
+	movq	mm5, MMWORD [wk(0)]	; mm5=tmp10L
+	movq	mm7, MMWORD [wk(1)]	; mm7=tmp10H
+
+	movq	mm2,mm5
+	movq	mm0,mm7
+	paddd	mm5,mm3			; mm5=data0L
+	paddd	mm7,mm4			; mm7=data0H
+	psubd	mm2,mm3			; mm2=data7L
+	psubd	mm0,mm4			; mm0=data7H
+
+	movq	mm3,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm3=[PD_DESCALE_P1]
+
+	paddd	mm5,mm3
+	paddd	mm7,mm3
+	psrad	mm5,DESCALE_P1
+	psrad	mm7,DESCALE_P1
+	paddd	mm2,mm3
+	paddd	mm0,mm3
+	psrad	mm2,DESCALE_P1
+	psrad	mm0,DESCALE_P1
+
+	packssdw  mm5,mm7		; mm5=data0=(00 01 02 03)
+	packssdw  mm2,mm0		; mm2=data7=(70 71 72 73)
+
+	movq	mm4, MMWORD [wk(4)]	; mm4=tmp11L
+	movq	mm3, MMWORD [wk(5)]	; mm3=tmp11H
+
+	movq	mm7,mm4
+	movq	mm0,mm3
+	paddd	mm4,mm1			; mm4=data1L
+	paddd	mm3,mm6			; mm3=data1H
+	psubd	mm7,mm1			; mm7=data6L
+	psubd	mm0,mm6			; mm0=data6H
+
+	movq	mm1,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm1=[PD_DESCALE_P1]
+
+	paddd	mm4,mm1
+	paddd	mm3,mm1
+	psrad	mm4,DESCALE_P1
+	psrad	mm3,DESCALE_P1
+	paddd	mm7,mm1
+	paddd	mm0,mm1
+	psrad	mm7,DESCALE_P1
+	psrad	mm0,DESCALE_P1
+
+	packssdw  mm4,mm3		; mm4=data1=(10 11 12 13)
+	packssdw  mm7,mm0		; mm7=data6=(60 61 62 63)
+
+	movq      mm6,mm5		; transpose coefficients(phase 1)
+	punpcklwd mm5,mm4		; mm5=(00 10 01 11)
+	punpckhwd mm6,mm4		; mm6=(02 12 03 13)
+	movq      mm1,mm7		; transpose coefficients(phase 1)
+	punpcklwd mm7,mm2		; mm7=(60 70 61 71)
+	punpckhwd mm1,mm2		; mm1=(62 72 63 73)
+
+	movq	mm3, MMWORD [wk(6)]	; mm3=tmp12L
+	movq	mm0, MMWORD [wk(7)]	; mm0=tmp12H
+	movq	mm4, MMWORD [wk(10)]	; mm4=tmp1L
+	movq	mm2, MMWORD [wk(11)]	; mm2=tmp1H
+
+	movq	MMWORD [wk(0)], mm5	; wk(0)=(00 10 01 11)
+	movq	MMWORD [wk(1)], mm6	; wk(1)=(02 12 03 13)
+	movq	MMWORD [wk(4)], mm7	; wk(4)=(60 70 61 71)
+	movq	MMWORD [wk(5)], mm1	; wk(5)=(62 72 63 73)
+
+	movq	mm5,mm3
+	movq	mm6,mm0
+	paddd	mm3,mm4			; mm3=data2L
+	paddd	mm0,mm2			; mm0=data2H
+	psubd	mm5,mm4			; mm5=data5L
+	psubd	mm6,mm2			; mm6=data5H
+
+	movq	mm7,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm7=[PD_DESCALE_P1]
+
+	paddd	mm3,mm7
+	paddd	mm0,mm7
+	psrad	mm3,DESCALE_P1
+	psrad	mm0,DESCALE_P1
+	paddd	mm5,mm7
+	paddd	mm6,mm7
+	psrad	mm5,DESCALE_P1
+	psrad	mm6,DESCALE_P1
+
+	packssdw  mm3,mm0		; mm3=data2=(20 21 22 23)
+	packssdw  mm5,mm6		; mm5=data5=(50 51 52 53)
+
+	movq	mm1, MMWORD [wk(2)]	; mm1=tmp13L
+	movq	mm4, MMWORD [wk(3)]	; mm4=tmp13H
+	movq	mm2, MMWORD [wk(8)]	; mm2=tmp0L
+	movq	mm7, MMWORD [wk(9)]	; mm7=tmp0H
+
+	movq	mm0,mm1
+	movq	mm6,mm4
+	paddd	mm1,mm2			; mm1=data3L
+	paddd	mm4,mm7			; mm4=data3H
+	psubd	mm0,mm2			; mm0=data4L
+	psubd	mm6,mm7			; mm6=data4H
+
+	movq	mm2,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm2=[PD_DESCALE_P1]
+
+	paddd	mm1,mm2
+	paddd	mm4,mm2
+	psrad	mm1,DESCALE_P1
+	psrad	mm4,DESCALE_P1
+	paddd	mm0,mm2
+	paddd	mm6,mm2
+	psrad	mm0,DESCALE_P1
+	psrad	mm6,DESCALE_P1
+
+	packssdw  mm1,mm4		; mm1=data3=(30 31 32 33)
+	packssdw  mm0,mm6		; mm0=data4=(40 41 42 43)
+
+	movq	mm7, MMWORD [wk(0)]	; mm7=(00 10 01 11)
+	movq	mm2, MMWORD [wk(1)]	; mm2=(02 12 03 13)
+
+	movq      mm4,mm3		; transpose coefficients(phase 1)
+	punpcklwd mm3,mm1		; mm3=(20 30 21 31)
+	punpckhwd mm4,mm1		; mm4=(22 32 23 33)
+	movq      mm6,mm0		; transpose coefficients(phase 1)
+	punpcklwd mm0,mm5		; mm0=(40 50 41 51)
+	punpckhwd mm6,mm5		; mm6=(42 52 43 53)
+
+	movq      mm1,mm7		; transpose coefficients(phase 2)
+	punpckldq mm7,mm3		; mm7=(00 10 20 30)
+	punpckhdq mm1,mm3		; mm1=(01 11 21 31)
+	movq      mm5,mm2		; transpose coefficients(phase 2)
+	punpckldq mm2,mm4		; mm2=(02 12 22 32)
+	punpckhdq mm5,mm4		; mm5=(03 13 23 33)
+
+	movq	mm3, MMWORD [wk(4)]	; mm3=(60 70 61 71)
+	movq	mm4, MMWORD [wk(5)]	; mm4=(62 72 63 73)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm7
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5
+
+	movq      mm7,mm0		; transpose coefficients(phase 2)
+	punpckldq mm0,mm3		; mm0=(40 50 60 70)
+	punpckhdq mm7,mm3		; mm7=(41 51 61 71)
+	movq      mm1,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm4		; mm6=(42 52 62 72)
+	punpckhdq mm1,mm4		; mm1=(43 53 63 73)
+
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm7
+	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm6
+	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm1
+
+.nextcolumn:
+	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 4*SIZEOF_ISLOW_MULT_TYPE	; quantptr
+	add	edi, byte 4*DCTSIZE*SIZEOF_JCOEF	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; JCOEF * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.rowloop:
+
+	; -- Even part
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	; (Original)
+	; z1 = (z2 + z3) * 0.541196100;
+	; tmp2 = z1 + z3 * -1.847759065;
+	; tmp3 = z1 + z2 * 0.765366865;
+	;
+	; (This implementation)
+	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+	movq      mm4,mm1		; mm1=in2=z2
+	movq      mm5,mm1
+	punpcklwd mm4,mm3		; mm3=in6=z3
+	punpckhwd mm5,mm3
+	movq      mm1,mm4
+	movq      mm3,mm5
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F130_F054)]	; mm4=tmp3L
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F130_F054)]	; mm5=tmp3H
+	pmaddwd   mm1,[GOTOFF(ebx,PW_F054_MF130)]	; mm1=tmp2L
+	pmaddwd   mm3,[GOTOFF(ebx,PW_F054_MF130)]	; mm3=tmp2H
+
+	movq      mm6,mm0
+	paddw     mm0,mm2		; mm0=in0+in4
+	psubw     mm6,mm2		; mm6=in0-in4
+
+	pxor      mm7,mm7
+	pxor      mm2,mm2
+	punpcklwd mm7,mm0		; mm7=tmp0L
+	punpckhwd mm2,mm0		; mm2=tmp0H
+	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
+	psrad     mm2,(16-CONST_BITS)	; psrad mm2,16 & pslld mm2,CONST_BITS
+
+	movq	mm0,mm7
+	paddd	mm7,mm4			; mm7=tmp10L
+	psubd	mm0,mm4			; mm0=tmp13L
+	movq	mm4,mm2
+	paddd	mm2,mm5			; mm2=tmp10H
+	psubd	mm4,mm5			; mm4=tmp13H
+
+	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp10L
+	movq	MMWORD [wk(1)], mm2	; wk(1)=tmp10H
+	movq	MMWORD [wk(2)], mm0	; wk(2)=tmp13L
+	movq	MMWORD [wk(3)], mm4	; wk(3)=tmp13H
+
+	pxor      mm5,mm5
+	pxor      mm7,mm7
+	punpcklwd mm5,mm6		; mm5=tmp1L
+	punpckhwd mm7,mm6		; mm7=tmp1H
+	psrad     mm5,(16-CONST_BITS)	; psrad mm5,16 & pslld mm5,CONST_BITS
+	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
+
+	movq	mm2,mm5
+	paddd	mm5,mm1			; mm5=tmp11L
+	psubd	mm2,mm1			; mm2=tmp12L
+	movq	mm0,mm7
+	paddd	mm7,mm3			; mm7=tmp11H
+	psubd	mm0,mm3			; mm0=tmp12H
+
+	movq	MMWORD [wk(4)], mm5	; wk(4)=tmp11L
+	movq	MMWORD [wk(5)], mm7	; wk(5)=tmp11H
+	movq	MMWORD [wk(6)], mm2	; wk(6)=tmp12L
+	movq	MMWORD [wk(7)], mm0	; wk(7)=tmp12H
+
+	; -- Odd part
+
+	movq	mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	movq	mm5,mm6
+	movq	mm7,mm4
+	paddw	mm5,mm3			; mm5=z3
+	paddw	mm7,mm1			; mm7=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movq      mm2,mm5
+	movq      mm0,mm5
+	punpcklwd mm2,mm7
+	punpckhwd mm0,mm7
+	movq      mm5,mm2
+	movq      mm7,mm0
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF078_F117)]	; mm2=z3L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_MF078_F117)]	; mm0=z3H
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F117_F078)]	; mm5=z4L
+	pmaddwd   mm7,[GOTOFF(ebx,PW_F117_F078)]	; mm7=z4H
+
+	movq	MMWORD [wk(10)], mm2	; wk(10)=z3L
+	movq	MMWORD [wk(11)], mm0	; wk(11)=z3H
+
+	; (Original)
+	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+	;
+	; (This implementation)
+	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+	; tmp0 += z3;  tmp1 += z4;
+	; tmp2 += z3;  tmp3 += z4;
+
+	movq      mm2,mm3
+	movq      mm0,mm3
+	punpcklwd mm2,mm4
+	punpckhwd mm0,mm4
+	movq      mm3,mm2
+	movq      mm4,mm0
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF060_MF089)]	; mm2=tmp0L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_MF060_MF089)]	; mm0=tmp0H
+	pmaddwd   mm3,[GOTOFF(ebx,PW_MF089_F060)]	; mm3=tmp3L
+	pmaddwd   mm4,[GOTOFF(ebx,PW_MF089_F060)]	; mm4=tmp3H
+
+	paddd	mm2, MMWORD [wk(10)]	; mm2=tmp0L
+	paddd	mm0, MMWORD [wk(11)]	; mm0=tmp0H
+	paddd	mm3,mm5			; mm3=tmp3L
+	paddd	mm4,mm7			; mm4=tmp3H
+
+	movq	MMWORD [wk(8)], mm2	; wk(8)=tmp0L
+	movq	MMWORD [wk(9)], mm0	; wk(9)=tmp0H
+
+	movq      mm2,mm1
+	movq      mm0,mm1
+	punpcklwd mm2,mm6
+	punpckhwd mm0,mm6
+	movq      mm1,mm2
+	movq      mm6,mm0
+	pmaddwd   mm2,[GOTOFF(ebx,PW_MF050_MF256)]	; mm2=tmp1L
+	pmaddwd   mm0,[GOTOFF(ebx,PW_MF050_MF256)]	; mm0=tmp1H
+	pmaddwd   mm1,[GOTOFF(ebx,PW_MF256_F050)]	; mm1=tmp2L
+	pmaddwd   mm6,[GOTOFF(ebx,PW_MF256_F050)]	; mm6=tmp2H
+
+	paddd	mm2,mm5			; mm2=tmp1L
+	paddd	mm0,mm7			; mm0=tmp1H
+	paddd	mm1, MMWORD [wk(10)]	; mm1=tmp2L
+	paddd	mm6, MMWORD [wk(11)]	; mm6=tmp2H
+
+	movq	MMWORD [wk(10)], mm2	; wk(10)=tmp1L
+	movq	MMWORD [wk(11)], mm0	; wk(11)=tmp1H
+
+	; -- Final output stage
+
+	movq	mm5, MMWORD [wk(0)]	; mm5=tmp10L
+	movq	mm7, MMWORD [wk(1)]	; mm7=tmp10H
+
+	movq	mm2,mm5
+	movq	mm0,mm7
+	paddd	mm5,mm3			; mm5=data0L
+	paddd	mm7,mm4			; mm7=data0H
+	psubd	mm2,mm3			; mm2=data7L
+	psubd	mm0,mm4			; mm0=data7H
+
+	movq	mm3,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm3=[PD_DESCALE_P2]
+
+	paddd	mm5,mm3
+	paddd	mm7,mm3
+	psrad	mm5,DESCALE_P2
+	psrad	mm7,DESCALE_P2
+	paddd	mm2,mm3
+	paddd	mm0,mm3
+	psrad	mm2,DESCALE_P2
+	psrad	mm0,DESCALE_P2
+
+	packssdw  mm5,mm7		; mm5=data0=(00 10 20 30)
+	packssdw  mm2,mm0		; mm2=data7=(07 17 27 37)
+
+	movq	mm4, MMWORD [wk(4)]	; mm4=tmp11L
+	movq	mm3, MMWORD [wk(5)]	; mm3=tmp11H
+
+	movq	mm7,mm4
+	movq	mm0,mm3
+	paddd	mm4,mm1			; mm4=data1L
+	paddd	mm3,mm6			; mm3=data1H
+	psubd	mm7,mm1			; mm7=data6L
+	psubd	mm0,mm6			; mm0=data6H
+
+	movq	mm1,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm1=[PD_DESCALE_P2]
+
+	paddd	mm4,mm1
+	paddd	mm3,mm1
+	psrad	mm4,DESCALE_P2
+	psrad	mm3,DESCALE_P2
+	paddd	mm7,mm1
+	paddd	mm0,mm1
+	psrad	mm7,DESCALE_P2
+	psrad	mm0,DESCALE_P2
+
+	packssdw  mm4,mm3		; mm4=data1=(01 11 21 31)
+	packssdw  mm7,mm0		; mm7=data6=(06 16 26 36)
+
+	packsswb  mm5,mm7		; mm5=(00 10 20 30 06 16 26 36)
+	packsswb  mm4,mm2		; mm4=(01 11 21 31 07 17 27 37)
+
+	movq	mm6, MMWORD [wk(6)]	; mm6=tmp12L
+	movq	mm1, MMWORD [wk(7)]	; mm1=tmp12H
+	movq	mm3, MMWORD [wk(10)]	; mm3=tmp1L
+	movq	mm0, MMWORD [wk(11)]	; mm0=tmp1H
+
+	movq	MMWORD [wk(0)], mm5	; wk(0)=(00 10 20 30 06 16 26 36)
+	movq	MMWORD [wk(1)], mm4	; wk(1)=(01 11 21 31 07 17 27 37)
+
+	movq	mm7,mm6
+	movq	mm2,mm1
+	paddd	mm6,mm3			; mm6=data2L
+	paddd	mm1,mm0			; mm1=data2H
+	psubd	mm7,mm3			; mm7=data5L
+	psubd	mm2,mm0			; mm2=data5H
+
+	movq	mm5,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm5=[PD_DESCALE_P2]
+
+	paddd	mm6,mm5
+	paddd	mm1,mm5
+	psrad	mm6,DESCALE_P2
+	psrad	mm1,DESCALE_P2
+	paddd	mm7,mm5
+	paddd	mm2,mm5
+	psrad	mm7,DESCALE_P2
+	psrad	mm2,DESCALE_P2
+
+	packssdw  mm6,mm1		; mm6=data2=(02 12 22 32)
+	packssdw  mm7,mm2		; mm7=data5=(05 15 25 35)
+
+	movq	mm4, MMWORD [wk(2)]	; mm4=tmp13L
+	movq	mm3, MMWORD [wk(3)]	; mm3=tmp13H
+	movq	mm0, MMWORD [wk(8)]	; mm0=tmp0L
+	movq	mm5, MMWORD [wk(9)]	; mm5=tmp0H
+
+	movq	mm1,mm4
+	movq	mm2,mm3
+	paddd	mm4,mm0			; mm4=data3L
+	paddd	mm3,mm5			; mm3=data3H
+	psubd	mm1,mm0			; mm1=data4L
+	psubd	mm2,mm5			; mm2=data4H
+
+	movq	mm0,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm0=[PD_DESCALE_P2]
+
+	paddd	mm4,mm0
+	paddd	mm3,mm0
+	psrad	mm4,DESCALE_P2
+	psrad	mm3,DESCALE_P2
+	paddd	mm1,mm0
+	paddd	mm2,mm0
+	psrad	mm1,DESCALE_P2
+	psrad	mm2,DESCALE_P2
+
+	movq      mm5,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm5=[PB_CENTERJSAMP]
+
+	packssdw  mm4,mm3		; mm4=data3=(03 13 23 33)
+	packssdw  mm1,mm2		; mm1=data4=(04 14 24 34)
+
+	movq      mm0, MMWORD [wk(0)]	; mm0=(00 10 20 30 06 16 26 36)
+	movq      mm3, MMWORD [wk(1)]	; mm3=(01 11 21 31 07 17 27 37)
+
+	packsswb  mm6,mm1		; mm6=(02 12 22 32 04 14 24 34)
+	packsswb  mm4,mm7		; mm4=(03 13 23 33 05 15 25 35)
+
+	paddb     mm0,mm5
+	paddb     mm3,mm5
+	paddb     mm6,mm5
+	paddb     mm4,mm5
+
+	movq      mm2,mm0		; transpose coefficients(phase 1)
+	punpcklbw mm0,mm3		; mm0=(00 01 10 11 20 21 30 31)
+	punpckhbw mm2,mm3		; mm2=(06 07 16 17 26 27 36 37)
+	movq      mm1,mm6		; transpose coefficients(phase 1)
+	punpcklbw mm6,mm4		; mm6=(02 03 12 13 22 23 32 33)
+	punpckhbw mm1,mm4		; mm1=(04 05 14 15 24 25 34 35)
+
+	movq      mm7,mm0		; transpose coefficients(phase 2)
+	punpcklwd mm0,mm6		; mm0=(00 01 02 03 10 11 12 13)
+	punpckhwd mm7,mm6		; mm7=(20 21 22 23 30 31 32 33)
+	movq      mm5,mm1		; transpose coefficients(phase 2)
+	punpcklwd mm1,mm2		; mm1=(04 05 06 07 14 15 16 17)
+	punpckhwd mm5,mm2		; mm5=(24 25 26 27 34 35 36 37)
+
+	movq      mm3,mm0		; transpose coefficients(phase 3)
+	punpckldq mm0,mm1		; mm0=(00 01 02 03 04 05 06 07)
+	punpckhdq mm3,mm1		; mm3=(10 11 12 13 14 15 16 17)
+	movq      mm4,mm7		; transpose coefficients(phase 3)
+	punpckldq mm7,mm5		; mm7=(20 21 22 23 24 25 26 27)
+	punpckhdq mm4,mm5		; mm4=(30 31 32 33 34 35 36 37)
+
+	pushpic	ebx			; save GOT address
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm3
+	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm7
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
+
+	poppic	ebx			; restore GOT address
+
+	add	esi, byte 4*SIZEOF_JCOEF	; wsptr
+	add	edi, byte 4*SIZEOF_JSAMPROW
+	dec	ecx				; ctr
+	jnz	near .rowloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jimmxred.asm b/simd/jimmxred.asm
new file mode 100644
index 0000000..a2b7103
--- /dev/null
+++ b/simd/jimmxred.asm
@@ -0,0 +1,706 @@
+;
+; jimmxred.asm - reduced-size IDCT (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains inverse-DCT routines that produce reduced-size
+; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
+; The following code is based directly on the IJG's original jidctred.c;
+; see the jidctred.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1_4	(CONST_BITS-PASS1_BITS+1)
+%define DESCALE_P2_4	(CONST_BITS+PASS1_BITS+3+1)
+%define DESCALE_P1_2	(CONST_BITS-PASS1_BITS+2)
+%define DESCALE_P2_2	(CONST_BITS+PASS1_BITS+3+2)
+
+%if CONST_BITS == 13
+F_0_211	equ	 1730		; FIX(0.211164243)
+F_0_509	equ	 4176		; FIX(0.509795579)
+F_0_601	equ	 4926		; FIX(0.601344887)
+F_0_720	equ	 5906		; FIX(0.720959822)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_850	equ	 6967		; FIX(0.850430095)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_061	equ	 8697		; FIX(1.061594337)
+F_1_272	equ	10426		; FIX(1.272758580)
+F_1_451	equ	11893		; FIX(1.451774981)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_2_172	equ	17799		; FIX(2.172734803)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_624	equ	29692		; FIX(3.624509785)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_211	equ	DESCALE( 226735879,30-CONST_BITS)	; FIX(0.211164243)
+F_0_509	equ	DESCALE( 547388834,30-CONST_BITS)	; FIX(0.509795579)
+F_0_601	equ	DESCALE( 645689155,30-CONST_BITS)	; FIX(0.601344887)
+F_0_720	equ	DESCALE( 774124714,30-CONST_BITS)	; FIX(0.720959822)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_850	equ	DESCALE( 913142361,30-CONST_BITS)	; FIX(0.850430095)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_061	equ	DESCALE(1139878239,30-CONST_BITS)	; FIX(1.061594337)
+F_1_272	equ	DESCALE(1366614119,30-CONST_BITS)	; FIX(1.272758580)
+F_1_451	equ	DESCALE(1558831516,30-CONST_BITS)	; FIX(1.451774981)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_2_172	equ	DESCALE(2332956230,30-CONST_BITS)	; FIX(2.172734803)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_624	equ	DESCALE(3891787747,30-CONST_BITS)	; FIX(3.624509785)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_red_mmx)
+
+EXTN(jconst_idct_red_mmx):
+
+PW_F184_MF076	times 2 dw  F_1_847,-F_0_765
+PW_F256_F089	times 2 dw  F_2_562, F_0_899
+PW_F106_MF217	times 2 dw  F_1_061,-F_2_172
+PW_MF060_MF050	times 2 dw -F_0_601,-F_0_509
+PW_F145_MF021	times 2 dw  F_1_451,-F_0_211
+PW_F362_MF127	times 2 dw  F_3_624,-F_1_272
+PW_F085_MF072	times 2 dw  F_0_850,-F_0_720
+PD_DESCALE_P1_4	times 2 dd  1 << (DESCALE_P1_4-1)
+PD_DESCALE_P2_4	times 2 dd  1 << (DESCALE_P2_4-1)
+PD_DESCALE_P1_2	times 2 dd  1 << (DESCALE_P1_2-1)
+PD_DESCALE_P2_2	times 2 dd  1 << (DESCALE_P2_2-1)
+PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 4x4 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_4x4_mmx (void * dct_table, JCOEFPTR coef_block,
+;                     JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_JCOEF
+					; JCOEF workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_4x4_mmx)
+
+EXTN(jsimd_idct_4x4_mmx):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; JCOEF * wsptr
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_4X4_MMX
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	short .columnDCT
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	mm0,mm1
+	packsswb mm0,mm0
+	movd	eax,mm0
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	psllw	mm0,PASS1_BITS
+
+	movq      mm2,mm0		; mm0=in0=(00 01 02 03)
+	punpcklwd mm0,mm0		; mm0=(00 00 01 01)
+	punpckhwd mm2,mm2		; mm2=(02 02 03 03)
+
+	movq      mm1,mm0
+	punpckldq mm0,mm0		; mm0=(00 00 00 00)
+	punpckhdq mm1,mm1		; mm1=(01 01 01 01)
+	movq      mm3,mm2
+	punpckldq mm2,mm2		; mm2=(02 02 02 02)
+	punpckhdq mm3,mm3		; mm3=(03 03 03 03)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Odd part
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movq	mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movq      mm4,mm0
+	movq      mm5,mm0
+	punpcklwd mm4,mm1
+	punpckhwd mm5,mm1
+	movq      mm0,mm4
+	movq      mm1,mm5
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F256_F089)]	; mm4=(tmp2L)
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F256_F089)]	; mm5=(tmp2H)
+	pmaddwd   mm0,[GOTOFF(ebx,PW_F106_MF217)]	; mm0=(tmp0L)
+	pmaddwd   mm1,[GOTOFF(ebx,PW_F106_MF217)]	; mm1=(tmp0H)
+
+	movq      mm6,mm2
+	movq      mm7,mm2
+	punpcklwd mm6,mm3
+	punpckhwd mm7,mm3
+	movq      mm2,mm6
+	movq      mm3,mm7
+	pmaddwd   mm6,[GOTOFF(ebx,PW_MF060_MF050)]	; mm6=(tmp2L)
+	pmaddwd   mm7,[GOTOFF(ebx,PW_MF060_MF050)]	; mm7=(tmp2H)
+	pmaddwd   mm2,[GOTOFF(ebx,PW_F145_MF021)]	; mm2=(tmp0L)
+	pmaddwd   mm3,[GOTOFF(ebx,PW_F145_MF021)]	; mm3=(tmp0H)
+
+	paddd	mm6,mm4			; mm6=tmp2L
+	paddd	mm7,mm5			; mm7=tmp2H
+	paddd	mm2,mm0			; mm2=tmp0L
+	paddd	mm3,mm1			; mm3=tmp0H
+
+	movq	MMWORD [wk(0)], mm2	; wk(0)=tmp0L
+	movq	MMWORD [wk(1)], mm3	; wk(1)=tmp0H
+
+	; -- Even part
+
+	movq	mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq	mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm4, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm5, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm0, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	pxor      mm1,mm1
+	pxor      mm2,mm2
+	punpcklwd mm1,mm4		; mm1=tmp0L
+	punpckhwd mm2,mm4		; mm2=tmp0H
+	psrad     mm1,(16-CONST_BITS-1)	; psrad mm1,16 & pslld mm1,CONST_BITS+1
+	psrad     mm2,(16-CONST_BITS-1)	; psrad mm2,16 & pslld mm2,CONST_BITS+1
+
+	movq      mm3,mm5		; mm5=in2=z2
+	punpcklwd mm5,mm0		; mm0=in6=z3
+	punpckhwd mm3,mm0
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F184_MF076)]	; mm5=tmp2L
+	pmaddwd   mm3,[GOTOFF(ebx,PW_F184_MF076)]	; mm3=tmp2H
+
+	movq	mm4,mm1
+	movq	mm0,mm2
+	paddd	mm1,mm5			; mm1=tmp10L
+	paddd	mm2,mm3			; mm2=tmp10H
+	psubd	mm4,mm5			; mm4=tmp12L
+	psubd	mm0,mm3			; mm0=tmp12H
+
+	; -- Final output stage
+
+	movq	mm5,mm1
+	movq	mm3,mm2
+	paddd	mm1,mm6			; mm1=data0L
+	paddd	mm2,mm7			; mm2=data0H
+	psubd	mm5,mm6			; mm5=data3L
+	psubd	mm3,mm7			; mm3=data3H
+
+	movq	mm6,[GOTOFF(ebx,PD_DESCALE_P1_4)]	; mm6=[PD_DESCALE_P1_4]
+
+	paddd	mm1,mm6
+	paddd	mm2,mm6
+	psrad	mm1,DESCALE_P1_4
+	psrad	mm2,DESCALE_P1_4
+	paddd	mm5,mm6
+	paddd	mm3,mm6
+	psrad	mm5,DESCALE_P1_4
+	psrad	mm3,DESCALE_P1_4
+
+	packssdw  mm1,mm2		; mm1=data0=(00 01 02 03)
+	packssdw  mm5,mm3		; mm5=data3=(30 31 32 33)
+
+	movq	mm7, MMWORD [wk(0)]	; mm7=tmp0L
+	movq	mm6, MMWORD [wk(1)]	; mm6=tmp0H
+
+	movq	mm2,mm4
+	movq	mm3,mm0
+	paddd	mm4,mm7			; mm4=data1L
+	paddd	mm0,mm6			; mm0=data1H
+	psubd	mm2,mm7			; mm2=data2L
+	psubd	mm3,mm6			; mm3=data2H
+
+	movq	mm7,[GOTOFF(ebx,PD_DESCALE_P1_4)]	; mm7=[PD_DESCALE_P1_4]
+
+	paddd	mm4,mm7
+	paddd	mm0,mm7
+	psrad	mm4,DESCALE_P1_4
+	psrad	mm0,DESCALE_P1_4
+	paddd	mm2,mm7
+	paddd	mm3,mm7
+	psrad	mm2,DESCALE_P1_4
+	psrad	mm3,DESCALE_P1_4
+
+	packssdw  mm4,mm0		; mm4=data1=(10 11 12 13)
+	packssdw  mm2,mm3		; mm2=data2=(20 21 22 23)
+
+	movq      mm6,mm1		; transpose coefficients(phase 1)
+	punpcklwd mm1,mm4		; mm1=(00 10 01 11)
+	punpckhwd mm6,mm4		; mm6=(02 12 03 13)
+	movq      mm7,mm2		; transpose coefficients(phase 1)
+	punpcklwd mm2,mm5		; mm2=(20 30 21 31)
+	punpckhwd mm7,mm5		; mm7=(22 32 23 33)
+
+	movq      mm0,mm1		; transpose coefficients(phase 2)
+	punpckldq mm1,mm2		; mm1=(00 10 20 30)
+	punpckhdq mm0,mm2		; mm0=(01 11 21 31)
+	movq      mm3,mm6		; transpose coefficients(phase 2)
+	punpckldq mm6,mm7		; mm6=(02 12 22 32)
+	punpckhdq mm3,mm7		; mm3=(03 13 23 33)
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm6
+	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
+
+.nextcolumn:
+	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 4*SIZEOF_ISLOW_MULT_TYPE	; quantptr
+	add	edi, byte 4*DCTSIZE*SIZEOF_JCOEF	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; JCOEF * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+
+	; -- Odd part
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq	mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	movq      mm4,mm0
+	movq      mm5,mm0
+	punpcklwd mm4,mm1
+	punpckhwd mm5,mm1
+	movq      mm0,mm4
+	movq      mm1,mm5
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F256_F089)]	; mm4=(tmp2L)
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F256_F089)]	; mm5=(tmp2H)
+	pmaddwd   mm0,[GOTOFF(ebx,PW_F106_MF217)]	; mm0=(tmp0L)
+	pmaddwd   mm1,[GOTOFF(ebx,PW_F106_MF217)]	; mm1=(tmp0H)
+
+	movq      mm6,mm2
+	movq      mm7,mm2
+	punpcklwd mm6,mm3
+	punpckhwd mm7,mm3
+	movq      mm2,mm6
+	movq      mm3,mm7
+	pmaddwd   mm6,[GOTOFF(ebx,PW_MF060_MF050)]	; mm6=(tmp2L)
+	pmaddwd   mm7,[GOTOFF(ebx,PW_MF060_MF050)]	; mm7=(tmp2H)
+	pmaddwd   mm2,[GOTOFF(ebx,PW_F145_MF021)]	; mm2=(tmp0L)
+	pmaddwd   mm3,[GOTOFF(ebx,PW_F145_MF021)]	; mm3=(tmp0H)
+
+	paddd	mm6,mm4			; mm6=tmp2L
+	paddd	mm7,mm5			; mm7=tmp2H
+	paddd	mm2,mm0			; mm2=tmp0L
+	paddd	mm3,mm1			; mm3=tmp0H
+
+	movq	MMWORD [wk(0)], mm2	; wk(0)=tmp0L
+	movq	MMWORD [wk(1)], mm3	; wk(1)=tmp0H
+
+	; -- Even part
+
+	movq	mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq	mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	pxor      mm1,mm1
+	pxor      mm2,mm2
+	punpcklwd mm1,mm4		; mm1=tmp0L
+	punpckhwd mm2,mm4		; mm2=tmp0H
+	psrad     mm1,(16-CONST_BITS-1)	; psrad mm1,16 & pslld mm1,CONST_BITS+1
+	psrad     mm2,(16-CONST_BITS-1)	; psrad mm2,16 & pslld mm2,CONST_BITS+1
+
+	movq      mm3,mm5		; mm5=in2=z2
+	punpcklwd mm5,mm0		; mm0=in6=z3
+	punpckhwd mm3,mm0
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F184_MF076)]	; mm5=tmp2L
+	pmaddwd   mm3,[GOTOFF(ebx,PW_F184_MF076)]	; mm3=tmp2H
+
+	movq	mm4,mm1
+	movq	mm0,mm2
+	paddd	mm1,mm5			; mm1=tmp10L
+	paddd	mm2,mm3			; mm2=tmp10H
+	psubd	mm4,mm5			; mm4=tmp12L
+	psubd	mm0,mm3			; mm0=tmp12H
+
+	; -- Final output stage
+
+	movq	mm5,mm1
+	movq	mm3,mm2
+	paddd	mm1,mm6			; mm1=data0L
+	paddd	mm2,mm7			; mm2=data0H
+	psubd	mm5,mm6			; mm5=data3L
+	psubd	mm3,mm7			; mm3=data3H
+
+	movq	mm6,[GOTOFF(ebx,PD_DESCALE_P2_4)]	; mm6=[PD_DESCALE_P2_4]
+
+	paddd	mm1,mm6
+	paddd	mm2,mm6
+	psrad	mm1,DESCALE_P2_4
+	psrad	mm2,DESCALE_P2_4
+	paddd	mm5,mm6
+	paddd	mm3,mm6
+	psrad	mm5,DESCALE_P2_4
+	psrad	mm3,DESCALE_P2_4
+
+	packssdw  mm1,mm2		; mm1=data0=(00 10 20 30)
+	packssdw  mm5,mm3		; mm5=data3=(03 13 23 33)
+
+	movq	mm7, MMWORD [wk(0)]	; mm7=tmp0L
+	movq	mm6, MMWORD [wk(1)]	; mm6=tmp0H
+
+	movq	mm2,mm4
+	movq	mm3,mm0
+	paddd	mm4,mm7			; mm4=data1L
+	paddd	mm0,mm6			; mm0=data1H
+	psubd	mm2,mm7			; mm2=data2L
+	psubd	mm3,mm6			; mm3=data2H
+
+	movq	mm7,[GOTOFF(ebx,PD_DESCALE_P2_4)]	; mm7=[PD_DESCALE_P2_4]
+
+	paddd	mm4,mm7
+	paddd	mm0,mm7
+	psrad	mm4,DESCALE_P2_4
+	psrad	mm0,DESCALE_P2_4
+	paddd	mm2,mm7
+	paddd	mm3,mm7
+	psrad	mm2,DESCALE_P2_4
+	psrad	mm3,DESCALE_P2_4
+
+	packssdw  mm4,mm0		; mm4=data1=(01 11 21 31)
+	packssdw  mm2,mm3		; mm2=data2=(02 12 22 32)
+
+	movq      mm6,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm6=[PB_CENTERJSAMP]
+
+	packsswb  mm1,mm2		; mm1=(00 10 20 30 02 12 22 32)
+	packsswb  mm4,mm5		; mm4=(01 11 21 31 03 13 23 33)
+	paddb     mm1,mm6
+	paddb     mm4,mm6
+
+	movq      mm7,mm1		; transpose coefficients(phase 1)
+	punpcklbw mm1,mm4		; mm1=(00 01 10 11 20 21 30 31)
+	punpckhbw mm7,mm4		; mm7=(02 03 12 13 22 23 32 33)
+
+	movq      mm0,mm1		; transpose coefficients(phase 2)
+	punpcklwd mm1,mm7		; mm1=(00 01 02 03 10 11 12 13)
+	punpckhwd mm0,mm7		; mm0=(20 21 22 23 30 31 32 33)
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	movd	DWORD [edx+eax*SIZEOF_JSAMPLE], mm1
+	movd	DWORD [esi+eax*SIZEOF_JSAMPLE], mm0
+
+	psrlq	mm1,4*BYTE_BIT
+	psrlq	mm0,4*BYTE_BIT
+
+	mov	edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movd	DWORD [edx+eax*SIZEOF_JSAMPLE], mm1
+	movd	DWORD [esi+eax*SIZEOF_JSAMPLE], mm0
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 2x2 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_2x2_mmx (void * dct_table, JCOEFPTR coef_block,
+;                     JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+	align	16
+	global	EXTN(jsimd_idct_2x2_mmx)
+
+EXTN(jsimd_idct_2x2_mmx):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input.
+
+	mov	edx, POINTER [dct_table(ebp)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(ebp)]		; inptr
+
+	; | input:                  | result:        |
+	; | 00 01 ** 03 ** 05 ** 07 |                |
+	; | 10 11 ** 13 ** 15 ** 17 |                |
+	; | ** ** ** ** ** ** ** ** |                |
+	; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
+	; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
+	; | 50 51 ** 53 ** 55 ** 57 |                |
+	; | ** ** ** ** ** ** ** ** |                |
+	; | 70 71 ** 73 ** 75 ** 77 |                |
+
+	; -- Odd part
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movq	mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; mm0=(10 11 ** 13), mm1=(30 31 ** 33)
+	; mm2=(50 51 ** 53), mm3=(70 71 ** 73)
+
+	pcmpeqd   mm7,mm7
+	pslld     mm7,WORD_BIT		; mm7={0x0000 0xFFFF 0x0000 0xFFFF}
+
+	movq      mm4,mm0		; mm4=(10 11 ** 13)
+	movq      mm5,mm2		; mm5=(50 51 ** 53)
+	punpcklwd mm4,mm1		; mm4=(10 30 11 31)
+	punpcklwd mm5,mm3		; mm5=(50 70 51 71)
+	pmaddwd   mm4,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F085_MF072)]
+
+	psrld	mm0,WORD_BIT		; mm0=(11 -- 13 --)
+	pand	mm1,mm7			; mm1=(-- 31 -- 33)
+	psrld	mm2,WORD_BIT		; mm2=(51 -- 53 --)
+	pand	mm3,mm7			; mm3=(-- 71 -- 73)
+	por	mm0,mm1			; mm0=(11 31 13 33)
+	por	mm2,mm3			; mm2=(51 71 53 73)
+	pmaddwd	mm0,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd	mm2,[GOTOFF(ebx,PW_F085_MF072)]
+
+	paddd	mm4,mm5			; mm4=tmp0[col0 col1]
+
+	movq	mm6, MMWORD [MMBLOCK(1,1,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(3,1,esi,SIZEOF_JCOEF)]
+	pmullw	mm6, MMWORD [MMBLOCK(1,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm1, MMWORD [MMBLOCK(3,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movq	mm3, MMWORD [MMBLOCK(5,1,esi,SIZEOF_JCOEF)]
+	movq	mm5, MMWORD [MMBLOCK(7,1,esi,SIZEOF_JCOEF)]
+	pmullw	mm3, MMWORD [MMBLOCK(5,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm5, MMWORD [MMBLOCK(7,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; mm6=(** 15 ** 17), mm1=(** 35 ** 37)
+	; mm3=(** 55 ** 57), mm5=(** 75 ** 77)
+
+	psrld	mm6,WORD_BIT		; mm6=(15 -- 17 --)
+	pand	mm1,mm7			; mm1=(-- 35 -- 37)
+	psrld	mm3,WORD_BIT		; mm3=(55 -- 57 --)
+	pand	mm5,mm7			; mm5=(-- 75 -- 77)
+	por	mm6,mm1			; mm6=(15 35 17 37)
+	por	mm3,mm5			; mm3=(55 75 57 77)
+	pmaddwd	mm6,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd	mm3,[GOTOFF(ebx,PW_F085_MF072)]
+
+	paddd	mm0,mm2			; mm0=tmp0[col1 col3]
+	paddd	mm6,mm3			; mm6=tmp0[col5 col7]
+
+	; -- Even part
+
+	movq	mm1, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq	mm5, MMWORD [MMBLOCK(0,1,esi,SIZEOF_JCOEF)]
+	pmullw	mm1, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	mm5, MMWORD [MMBLOCK(0,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; mm1=(00 01 ** 03), mm5=(** 05 ** 07)
+
+	movq	mm2,mm1				; mm2=(00 01 ** 03)
+	pslld	mm1,WORD_BIT			; mm1=(-- 00 -- **)
+	psrad	mm1,(WORD_BIT-CONST_BITS-2)	; mm1=tmp10[col0 ****]
+
+	pand	mm2,mm7				; mm2=(-- 01 -- 03)
+	pand	mm5,mm7				; mm5=(-- 05 -- 07)
+	psrad	mm2,(WORD_BIT-CONST_BITS-2)	; mm2=tmp10[col1 col3]
+	psrad	mm5,(WORD_BIT-CONST_BITS-2)	; mm5=tmp10[col5 col7]
+
+	; -- Final output stage
+
+	movq      mm3,mm1
+	paddd     mm1,mm4		; mm1=data0[col0 ****]=(A0 **)
+	psubd     mm3,mm4		; mm3=data1[col0 ****]=(B0 **)
+	punpckldq mm1,mm3		; mm1=(A0 B0)
+
+	movq	mm7,[GOTOFF(ebx,PD_DESCALE_P1_2)]	; mm7=[PD_DESCALE_P1_2]
+
+	movq	mm4,mm2
+	movq	mm3,mm5
+	paddd	mm2,mm0			; mm2=data0[col1 col3]=(A1 A3)
+	paddd	mm5,mm6			; mm5=data0[col5 col7]=(A5 A7)
+	psubd	mm4,mm0			; mm4=data1[col1 col3]=(B1 B3)
+	psubd	mm3,mm6			; mm3=data1[col5 col7]=(B5 B7)
+
+	paddd	mm1,mm7
+	psrad	mm1,DESCALE_P1_2
+
+	paddd	mm2,mm7
+	paddd	mm5,mm7
+	psrad	mm2,DESCALE_P1_2
+	psrad	mm5,DESCALE_P1_2
+	paddd	mm4,mm7
+	paddd	mm3,mm7
+	psrad	mm4,DESCALE_P1_2
+	psrad	mm3,DESCALE_P1_2
+
+	; ---- Pass 2: process rows, store into output array.
+
+	mov	edi, JSAMPARRAY [output_buf(ebp)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(ebp)]
+
+	; | input:| result:|
+	; | A0 B0 |        |
+	; | A1 B1 | C0 C1  |
+	; | A3 B3 | D0 D1  |
+	; | A5 B5 |        |
+	; | A7 B7 |        |
+
+	; -- Odd part
+
+	packssdw  mm2,mm4		; mm2=(A1 A3 B1 B3)
+	packssdw  mm5,mm3		; mm5=(A5 A7 B5 B7)
+	pmaddwd   mm2,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd   mm5,[GOTOFF(ebx,PW_F085_MF072)]
+
+	paddd     mm2,mm5		; mm2=tmp0[row0 row1]
+
+	; -- Even part
+
+	pslld     mm1,(CONST_BITS+2)	; mm1=tmp10[row0 row1]
+
+	; -- Final output stage
+
+	movq      mm0,[GOTOFF(ebx,PD_DESCALE_P2_2)]	; mm0=[PD_DESCALE_P2_2]
+
+	movq      mm6,mm1
+	paddd     mm1,mm2		; mm1=data0[row0 row1]=(C0 C1)
+	psubd     mm6,mm2		; mm6=data1[row0 row1]=(D0 D1)
+
+	paddd     mm1,mm0
+	paddd     mm6,mm0
+	psrad     mm1,DESCALE_P2_2
+	psrad     mm6,DESCALE_P2_2
+
+	movq      mm7,mm1		; transpose coefficients
+	punpckldq mm1,mm6		; mm1=(C0 D0)
+	punpckhdq mm7,mm6		; mm7=(C1 D1)
+
+	packssdw  mm1,mm7		; mm1=(C0 D0 C1 D1)
+	packsswb  mm1,mm1		; mm1=(C0 D0 C1 D1 C0 D0 C1 D1)
+	paddb     mm1,[GOTOFF(ebx,PB_CENTERJSAMP)]
+
+	movd	ecx,mm1
+	movd	ebx,mm1			; ebx=(C0 D0 C1 D1)
+	shr	ecx,2*BYTE_BIT		; ecx=(C1 D1 -- --)
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	mov	WORD [edx+eax*SIZEOF_JSAMPLE], bx
+	mov	WORD [esi+eax*SIZEOF_JSAMPLE], cx
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2flt-64.asm b/simd/jiss2flt-64.asm
new file mode 100644
index 0000000..6e7e6d4
--- /dev/null
+++ b/simd/jiss2flt-64.asm
@@ -0,0 +1,483 @@
+;
+; jiss2flt-64.asm - floating-point IDCT (64-bit SSE & SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the inverse DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jidctflt.c; see the jidctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_float_sse2)
+
+EXTN(jconst_idct_float_sse2):
+
+PD_1_414	times 4 dd  1.414213562373095048801689
+PD_1_847	times 4 dd  1.847759065022573512256366
+PD_1_082	times 4 dd  1.082392200292393968799446
+PD_M2_613	times 4 dd -2.613125929752753055713286
+PD_RNDINT_MAGIC	times 4 dd  100663296.0	; (float)(0x00C00000 << 3)
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_float_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                        JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = void * dct_table
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13 = JDIMENSION output_col
+
+%define original_rbp	rbp+0
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
+					; FAST_FLOAT workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_float_sse2)
+
+EXTN(jsimd_idct_float_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [workspace]
+	collect_args
+	push	rbx
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+	mov	rdx, r10	; quantptr
+	mov	rsi, r11		; inptr
+	lea	rdi, [workspace]			; FAST_FLOAT * wsptr
+	mov	rcx, DCTSIZE/4				; ctr
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
+	mov	eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movq	xmm1, XMM_MMWORD [MMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movq	xmm2, XMM_MMWORD [MMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	movq	xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	movq	xmm4, XMM_MMWORD [MMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+	movq	xmm5, XMM_MMWORD [MMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	movq	xmm6, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	movq	xmm7, XMM_MMWORD [MMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1,xmm2
+	por	xmm3,xmm4
+	por	xmm5,xmm6
+	por	xmm1,xmm3
+	por	xmm5,xmm7
+	por	xmm1,xmm5
+	packsswb xmm1,xmm1
+	movd	eax,xmm1
+	test	rax,rax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq      xmm0, XMM_MMWORD [MMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+
+	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
+	psrad     xmm0,(DWORD_BIT-WORD_BIT)	; xmm0=in0=(00 01 02 03)
+	cvtdq2ps  xmm0,xmm0			; xmm0=in0=(00 01 02 03)
+
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm1,xmm0
+	movaps	xmm2,xmm0
+	movaps	xmm3,xmm0
+
+	shufps	xmm0,xmm0,0x00			; xmm0=(00 00 00 00)
+	shufps	xmm1,xmm1,0x55			; xmm1=(01 01 01 01)
+	shufps	xmm2,xmm2,0xAA			; xmm2=(02 02 02 02)
+	shufps	xmm3,xmm3,0xFF			; xmm3=(03 03 03 03)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(2,1,rdi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
+	jmp	near .nextcolumn
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movq      xmm0, XMM_MMWORD [MMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	movq      xmm1, XMM_MMWORD [MMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	movq      xmm2, XMM_MMWORD [MMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+	movq      xmm3, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+
+	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
+	punpcklwd xmm1,xmm1		; xmm1=(20 20 21 21 22 22 23 23)
+	psrad     xmm0,(DWORD_BIT-WORD_BIT)	; xmm0=in0=(00 01 02 03)
+	psrad     xmm1,(DWORD_BIT-WORD_BIT)	; xmm1=in2=(20 21 22 23)
+	cvtdq2ps  xmm0,xmm0			; xmm0=in0=(00 01 02 03)
+	cvtdq2ps  xmm1,xmm1			; xmm1=in2=(20 21 22 23)
+
+	punpcklwd xmm2,xmm2		; xmm2=(40 40 41 41 42 42 43 43)
+	punpcklwd xmm3,xmm3		; xmm3=(60 60 61 61 62 62 63 63)
+	psrad     xmm2,(DWORD_BIT-WORD_BIT)	; xmm2=in4=(40 41 42 43)
+	psrad     xmm3,(DWORD_BIT-WORD_BIT)	; xmm3=in6=(60 61 62 63)
+	cvtdq2ps  xmm2,xmm2			; xmm2=in4=(40 41 42 43)
+	cvtdq2ps  xmm3,xmm3			; xmm3=in6=(60 61 62 63)
+
+	mulps     xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[rel PD_1_414]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movq      xmm2, XMM_MMWORD [MMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movq      xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	movq      xmm5, XMM_MMWORD [MMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	movq      xmm1, XMM_MMWORD [MMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+
+	punpcklwd xmm2,xmm2		; xmm2=(10 10 11 11 12 12 13 13)
+	punpcklwd xmm3,xmm3		; xmm3=(30 30 31 31 32 32 33 33)
+	psrad     xmm2,(DWORD_BIT-WORD_BIT)	; xmm2=in1=(10 11 12 13)
+	psrad     xmm3,(DWORD_BIT-WORD_BIT)	; xmm3=in3=(30 31 32 33)
+	cvtdq2ps  xmm2,xmm2			; xmm2=in1=(10 11 12 13)
+	cvtdq2ps  xmm3,xmm3			; xmm3=in3=(30 31 32 33)
+
+	punpcklwd xmm5,xmm5		; xmm5=(50 50 51 51 52 52 53 53)
+	punpcklwd xmm1,xmm1		; xmm1=(70 70 71 71 72 72 73 73)
+	psrad     xmm5,(DWORD_BIT-WORD_BIT)	; xmm5=in5=(50 51 52 53)
+	psrad     xmm1,(DWORD_BIT-WORD_BIT)	; xmm1=in7=(70 71 72 73)
+	cvtdq2ps  xmm5,xmm5			; xmm5=in5=(50 51 52 53)
+	cvtdq2ps  xmm1,xmm1			; xmm1=in7=(70 71 72 73)
+
+	mulps     xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[rel PD_1_414]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[rel PD_1_847]	; xmm0=z5
+	mulps	xmm3,[rel PD_M2_613]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[rel PD_1_082]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 01 02 03)
+	addps	xmm7,xmm3		; xmm7=data1=(10 11 12 13)
+	subps	xmm5,xmm1		; xmm5=data7=(70 71 72 73)
+	subps	xmm0,xmm3		; xmm0=data6=(60 61 62 63)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps    xmm1,xmm6		; transpose coefficients(phase 1)
+	unpcklps  xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps  xmm1,xmm7		; xmm1=(02 12 03 13)
+	movaps    xmm3,xmm0		; transpose coefficients(phase 1)
+	unpcklps  xmm0,xmm5		; xmm0=(60 70 61 71)
+	unpckhps  xmm3,xmm5		; xmm3=(62 72 63 73)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
+	movaps	xmm5, XMMWORD [wk(1)]	; xmm5=tmp3
+
+	movaps	XMMWORD [wk(0)], xmm0	; wk(0)=(60 70 61 71)
+	movaps	XMMWORD [wk(1)], xmm3	; wk(1)=(62 72 63 73)
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm0,xmm7
+	movaps	xmm3,xmm5
+	addps	xmm7,xmm2		; xmm7=data2=(20 21 22 23)
+	addps	xmm5,xmm4		; xmm5=data4=(40 41 42 43)
+	subps	xmm0,xmm2		; xmm0=data5=(50 51 52 53)
+	subps	xmm3,xmm4		; xmm3=data3=(30 31 32 33)
+
+	movaps    xmm2,xmm7		; transpose coefficients(phase 1)
+	unpcklps  xmm7,xmm3		; xmm7=(20 30 21 31)
+	unpckhps  xmm2,xmm3		; xmm2=(22 32 23 33)
+	movaps    xmm4,xmm5		; transpose coefficients(phase 1)
+	unpcklps  xmm5,xmm0		; xmm5=(40 50 41 51)
+	unpckhps  xmm4,xmm0		; xmm4=(42 52 43 53)
+
+	movaps    xmm3,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm7		; xmm6=(00 10 20 30)
+	unpckhps2 xmm3,xmm7		; xmm3=(01 11 21 31)
+	movaps    xmm0,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm2		; xmm1=(02 12 22 32)
+	unpckhps2 xmm0,xmm2		; xmm0=(03 13 23 33)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=(60 70 61 71)
+	movaps	xmm2, XMMWORD [wk(1)]	; xmm2=(62 72 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm0
+
+	movaps    xmm6,xmm5		; transpose coefficients(phase 2)
+	unpcklps2 xmm5,xmm7		; xmm5=(40 50 60 70)
+	unpckhps2 xmm6,xmm7		; xmm6=(41 51 61 71)
+	movaps    xmm3,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(42 52 62 72)
+	unpckhps2 xmm3,xmm2		; xmm3=(43 53 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(2,1,rdi,SIZEOF_FAST_FLOAT)], xmm4
+	movaps	XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
+
+.nextcolumn:
+	add	rsi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	rdx, byte 4*SIZEOF_FLOAT_MULT_TYPE	; quantptr
+	add	rdi,      4*DCTSIZE*SIZEOF_FAST_FLOAT	; wsptr
+	dec	rcx					; ctr
+	jnz	near .columnloop
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
+	prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
+	prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
+	prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	rax, [original_rbp]
+	lea	rsi, [workspace]			; FAST_FLOAT * wsptr
+	mov	rdi, r12	; (JSAMPROW *)
+	mov	rax, r13
+	mov	rcx, DCTSIZE/4				; ctr
+.rowloop:
+
+	; -- Even part
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[rel PD_1_414]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[rel PD_1_414]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[rel PD_1_847]	; xmm0=z5
+	mulps	xmm3,[rel PD_M2_613]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[rel PD_1_082]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 10 20 30)
+	addps	xmm7,xmm3		; xmm7=data1=(01 11 21 31)
+	subps	xmm5,xmm1		; xmm5=data7=(07 17 27 37)
+	subps	xmm0,xmm3		; xmm0=data6=(06 16 26 36)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps	xmm1,[rel PD_RNDINT_MAGIC]	; xmm1=[rel PD_RNDINT_MAGIC]
+	pcmpeqd	xmm3,xmm3
+	psrld	xmm3,WORD_BIT		; xmm3={0xFFFF 0x0000 0xFFFF 0x0000 ..}
+
+	addps	xmm6,xmm1	; xmm6=roundint(data0/8)=(00 ** 10 ** 20 ** 30 **)
+	addps	xmm7,xmm1	; xmm7=roundint(data1/8)=(01 ** 11 ** 21 ** 31 **)
+	addps	xmm0,xmm1	; xmm0=roundint(data6/8)=(06 ** 16 ** 26 ** 36 **)
+	addps	xmm5,xmm1	; xmm5=roundint(data7/8)=(07 ** 17 ** 27 ** 37 **)
+
+	pand	xmm6,xmm3		; xmm6=(00 -- 10 -- 20 -- 30 --)
+	pslld	xmm7,WORD_BIT		; xmm7=(-- 01 -- 11 -- 21 -- 31)
+	pand	xmm0,xmm3		; xmm0=(06 -- 16 -- 26 -- 36 --)
+	pslld	xmm5,WORD_BIT		; xmm5=(-- 07 -- 17 -- 27 -- 37)
+	por	xmm6,xmm7		; xmm6=(00 01 10 11 20 21 30 31)
+	por	xmm0,xmm5		; xmm0=(06 07 16 17 26 27 36 37)
+
+	movaps	xmm1, XMMWORD [wk(0)]	; xmm1=tmp2
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=tmp3
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm7,xmm1
+	movaps	xmm5,xmm3
+	addps	xmm1,xmm2		; xmm1=data2=(02 12 22 32)
+	addps	xmm3,xmm4		; xmm3=data4=(04 14 24 34)
+	subps	xmm7,xmm2		; xmm7=data5=(05 15 25 35)
+	subps	xmm5,xmm4		; xmm5=data3=(03 13 23 33)
+
+	movaps	xmm2,[rel PD_RNDINT_MAGIC]	; xmm2=[rel PD_RNDINT_MAGIC]
+	pcmpeqd	xmm4,xmm4
+	psrld	xmm4,WORD_BIT		; xmm4={0xFFFF 0x0000 0xFFFF 0x0000 ..}
+
+	addps	xmm3,xmm2	; xmm3=roundint(data4/8)=(04 ** 14 ** 24 ** 34 **)
+	addps	xmm7,xmm2	; xmm7=roundint(data5/8)=(05 ** 15 ** 25 ** 35 **)
+	addps	xmm1,xmm2	; xmm1=roundint(data2/8)=(02 ** 12 ** 22 ** 32 **)
+	addps	xmm5,xmm2	; xmm5=roundint(data3/8)=(03 ** 13 ** 23 ** 33 **)
+
+	pand	xmm3,xmm4		; xmm3=(04 -- 14 -- 24 -- 34 --)
+	pslld	xmm7,WORD_BIT		; xmm7=(-- 05 -- 15 -- 25 -- 35)
+	pand	xmm1,xmm4		; xmm1=(02 -- 12 -- 22 -- 32 --)
+	pslld	xmm5,WORD_BIT		; xmm5=(-- 03 -- 13 -- 23 -- 33)
+	por	xmm3,xmm7		; xmm3=(04 05 14 15 24 25 34 35)
+	por	xmm1,xmm5		; xmm1=(02 03 12 13 22 23 32 33)
+
+	movdqa    xmm2,[rel PB_CENTERJSAMP]	; xmm2=[rel PB_CENTERJSAMP]
+
+	packsswb  xmm6,xmm3	; xmm6=(00 01 10 11 20 21 30 31 04 05 14 15 24 25 34 35)
+	packsswb  xmm1,xmm0	; xmm1=(02 03 12 13 22 23 32 33 06 07 16 17 26 27 36 37)
+	paddb     xmm6,xmm2
+	paddb     xmm1,xmm2
+
+	movdqa    xmm4,xmm6	; transpose coefficients(phase 2)
+	punpcklwd xmm6,xmm1	; xmm6=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+	punpckhwd xmm4,xmm1	; xmm4=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+
+	movdqa    xmm7,xmm6	; transpose coefficients(phase 3)
+	punpckldq xmm6,xmm4	; xmm6=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+	punpckhdq xmm7,xmm4	; xmm7=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+
+	pshufd	xmm5,xmm6,0x4E	; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+	pshufd	xmm3,xmm7,0x4E	; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+
+	mov	rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+	mov	rbx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+	movq	XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm7
+	mov	rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+	mov	rbx, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
+	movq	XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm3
+
+	add	rsi, byte 4*SIZEOF_FAST_FLOAT	; wsptr
+	add	rdi, byte 4*SIZEOF_JSAMPROW
+	dec	rcx				; ctr
+	jnz	near .rowloop
+
+	pop	rbx
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2flt.asm b/simd/jiss2flt.asm
new file mode 100644
index 0000000..17bc363
--- /dev/null
+++ b/simd/jiss2flt.asm
@@ -0,0 +1,498 @@
+;
+; jiss2flt.asm - floating-point IDCT (SSE & SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the inverse DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jidctflt.c; see the jidctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_float_sse2)
+
+EXTN(jconst_idct_float_sse2):
+
+PD_1_414	times 4 dd  1.414213562373095048801689
+PD_1_847	times 4 dd  1.847759065022573512256366
+PD_1_082	times 4 dd  1.082392200292393968799446
+PD_M2_613	times 4 dd -2.613125929752753055713286
+PD_RNDINT_MAGIC	times 4 dd  100663296.0	; (float)(0x00C00000 << 3)
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_float_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                        JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
+					; FAST_FLOAT workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_float_sse2)
+
+EXTN(jsimd_idct_float_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; FAST_FLOAT * wsptr
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movq	xmm1, XMM_MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	xmm2, XMM_MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq	xmm3, XMM_MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq	xmm4, XMM_MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq	xmm5, XMM_MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq	xmm6, XMM_MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	movq	xmm7, XMM_MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	xmm1,xmm2
+	por	xmm3,xmm4
+	por	xmm5,xmm6
+	por	xmm1,xmm3
+	por	xmm5,xmm7
+	por	xmm1,xmm5
+	packsswb xmm1,xmm1
+	movd	eax,xmm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq      xmm0, XMM_MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+
+	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
+	psrad     xmm0,(DWORD_BIT-WORD_BIT)	; xmm0=in0=(00 01 02 03)
+	cvtdq2ps  xmm0,xmm0			; xmm0=in0=(00 01 02 03)
+
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm1,xmm0
+	movaps	xmm2,xmm0
+	movaps	xmm3,xmm0
+
+	shufps	xmm0,xmm0,0x00			; xmm0=(00 00 00 00)
+	shufps	xmm1,xmm1,0x55			; xmm1=(01 01 01 01)
+	shufps	xmm2,xmm2,0xAA			; xmm2=(02 02 02 02)
+	shufps	xmm3,xmm3,0xFF			; xmm3=(03 03 03 03)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movq      xmm0, XMM_MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq      xmm1, XMM_MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq      xmm2, XMM_MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq      xmm3, XMM_MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
+	punpcklwd xmm1,xmm1		; xmm1=(20 20 21 21 22 22 23 23)
+	psrad     xmm0,(DWORD_BIT-WORD_BIT)	; xmm0=in0=(00 01 02 03)
+	psrad     xmm1,(DWORD_BIT-WORD_BIT)	; xmm1=in2=(20 21 22 23)
+	cvtdq2ps  xmm0,xmm0			; xmm0=in0=(00 01 02 03)
+	cvtdq2ps  xmm1,xmm1			; xmm1=in2=(20 21 22 23)
+
+	punpcklwd xmm2,xmm2		; xmm2=(40 40 41 41 42 42 43 43)
+	punpcklwd xmm3,xmm3		; xmm3=(60 60 61 61 62 62 63 63)
+	psrad     xmm2,(DWORD_BIT-WORD_BIT)	; xmm2=in4=(40 41 42 43)
+	psrad     xmm3,(DWORD_BIT-WORD_BIT)	; xmm3=in6=(60 61 62 63)
+	cvtdq2ps  xmm2,xmm2			; xmm2=in4=(40 41 42 43)
+	cvtdq2ps  xmm3,xmm3			; xmm3=in6=(60 61 62 63)
+
+	mulps     xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[GOTOFF(ebx,PD_1_414)]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movq      xmm2, XMM_MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq      xmm3, XMM_MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq      xmm5, XMM_MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq      xmm1, XMM_MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	punpcklwd xmm2,xmm2		; xmm2=(10 10 11 11 12 12 13 13)
+	punpcklwd xmm3,xmm3		; xmm3=(30 30 31 31 32 32 33 33)
+	psrad     xmm2,(DWORD_BIT-WORD_BIT)	; xmm2=in1=(10 11 12 13)
+	psrad     xmm3,(DWORD_BIT-WORD_BIT)	; xmm3=in3=(30 31 32 33)
+	cvtdq2ps  xmm2,xmm2			; xmm2=in1=(10 11 12 13)
+	cvtdq2ps  xmm3,xmm3			; xmm3=in3=(30 31 32 33)
+
+	punpcklwd xmm5,xmm5		; xmm5=(50 50 51 51 52 52 53 53)
+	punpcklwd xmm1,xmm1		; xmm1=(70 70 71 71 72 72 73 73)
+	psrad     xmm5,(DWORD_BIT-WORD_BIT)	; xmm5=in5=(50 51 52 53)
+	psrad     xmm1,(DWORD_BIT-WORD_BIT)	; xmm1=in7=(70 71 72 73)
+	cvtdq2ps  xmm5,xmm5			; xmm5=in5=(50 51 52 53)
+	cvtdq2ps  xmm1,xmm1			; xmm1=in7=(70 71 72 73)
+
+	mulps     xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[GOTOFF(ebx,PD_1_414)]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[GOTOFF(ebx,PD_1_847)]	; xmm0=z5
+	mulps	xmm3,[GOTOFF(ebx,PD_M2_613)]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[GOTOFF(ebx,PD_1_082)]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 01 02 03)
+	addps	xmm7,xmm3		; xmm7=data1=(10 11 12 13)
+	subps	xmm5,xmm1		; xmm5=data7=(70 71 72 73)
+	subps	xmm0,xmm3		; xmm0=data6=(60 61 62 63)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps    xmm1,xmm6		; transpose coefficients(phase 1)
+	unpcklps  xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps  xmm1,xmm7		; xmm1=(02 12 03 13)
+	movaps    xmm3,xmm0		; transpose coefficients(phase 1)
+	unpcklps  xmm0,xmm5		; xmm0=(60 70 61 71)
+	unpckhps  xmm3,xmm5		; xmm3=(62 72 63 73)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
+	movaps	xmm5, XMMWORD [wk(1)]	; xmm5=tmp3
+
+	movaps	XMMWORD [wk(0)], xmm0	; wk(0)=(60 70 61 71)
+	movaps	XMMWORD [wk(1)], xmm3	; wk(1)=(62 72 63 73)
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm0,xmm7
+	movaps	xmm3,xmm5
+	addps	xmm7,xmm2		; xmm7=data2=(20 21 22 23)
+	addps	xmm5,xmm4		; xmm5=data4=(40 41 42 43)
+	subps	xmm0,xmm2		; xmm0=data5=(50 51 52 53)
+	subps	xmm3,xmm4		; xmm3=data3=(30 31 32 33)
+
+	movaps    xmm2,xmm7		; transpose coefficients(phase 1)
+	unpcklps  xmm7,xmm3		; xmm7=(20 30 21 31)
+	unpckhps  xmm2,xmm3		; xmm2=(22 32 23 33)
+	movaps    xmm4,xmm5		; transpose coefficients(phase 1)
+	unpcklps  xmm5,xmm0		; xmm5=(40 50 41 51)
+	unpckhps  xmm4,xmm0		; xmm4=(42 52 43 53)
+
+	movaps    xmm3,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm7		; xmm6=(00 10 20 30)
+	unpckhps2 xmm3,xmm7		; xmm3=(01 11 21 31)
+	movaps    xmm0,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm2		; xmm1=(02 12 22 32)
+	unpckhps2 xmm0,xmm2		; xmm0=(03 13 23 33)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=(60 70 61 71)
+	movaps	xmm2, XMMWORD [wk(1)]	; xmm2=(62 72 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+
+	movaps    xmm6,xmm5		; transpose coefficients(phase 2)
+	unpcklps2 xmm5,xmm7		; xmm5=(40 50 60 70)
+	unpckhps2 xmm6,xmm7		; xmm6=(41 51 61 71)
+	movaps    xmm3,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(42 52 62 72)
+	unpckhps2 xmm3,xmm2		; xmm3=(43 53 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
+	movaps	XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
+
+.nextcolumn:
+	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 4*SIZEOF_FLOAT_MULT_TYPE	; quantptr
+	add	edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; FAST_FLOAT * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.rowloop:
+
+	; -- Even part
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[GOTOFF(ebx,PD_1_414)]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[GOTOFF(ebx,PD_1_414)]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[GOTOFF(ebx,PD_1_847)]	; xmm0=z5
+	mulps	xmm3,[GOTOFF(ebx,PD_M2_613)]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[GOTOFF(ebx,PD_1_082)]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 10 20 30)
+	addps	xmm7,xmm3		; xmm7=data1=(01 11 21 31)
+	subps	xmm5,xmm1		; xmm5=data7=(07 17 27 37)
+	subps	xmm0,xmm3		; xmm0=data6=(06 16 26 36)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps	xmm1,[GOTOFF(ebx,PD_RNDINT_MAGIC)]	; xmm1=[PD_RNDINT_MAGIC]
+	pcmpeqd	xmm3,xmm3
+	psrld	xmm3,WORD_BIT		; xmm3={0xFFFF 0x0000 0xFFFF 0x0000 ..}
+
+	addps	xmm6,xmm1	; xmm6=roundint(data0/8)=(00 ** 10 ** 20 ** 30 **)
+	addps	xmm7,xmm1	; xmm7=roundint(data1/8)=(01 ** 11 ** 21 ** 31 **)
+	addps	xmm0,xmm1	; xmm0=roundint(data6/8)=(06 ** 16 ** 26 ** 36 **)
+	addps	xmm5,xmm1	; xmm5=roundint(data7/8)=(07 ** 17 ** 27 ** 37 **)
+
+	pand	xmm6,xmm3		; xmm6=(00 -- 10 -- 20 -- 30 --)
+	pslld	xmm7,WORD_BIT		; xmm7=(-- 01 -- 11 -- 21 -- 31)
+	pand	xmm0,xmm3		; xmm0=(06 -- 16 -- 26 -- 36 --)
+	pslld	xmm5,WORD_BIT		; xmm5=(-- 07 -- 17 -- 27 -- 37)
+	por	xmm6,xmm7		; xmm6=(00 01 10 11 20 21 30 31)
+	por	xmm0,xmm5		; xmm0=(06 07 16 17 26 27 36 37)
+
+	movaps	xmm1, XMMWORD [wk(0)]	; xmm1=tmp2
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=tmp3
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm7,xmm1
+	movaps	xmm5,xmm3
+	addps	xmm1,xmm2		; xmm1=data2=(02 12 22 32)
+	addps	xmm3,xmm4		; xmm3=data4=(04 14 24 34)
+	subps	xmm7,xmm2		; xmm7=data5=(05 15 25 35)
+	subps	xmm5,xmm4		; xmm5=data3=(03 13 23 33)
+
+	movaps	xmm2,[GOTOFF(ebx,PD_RNDINT_MAGIC)]	; xmm2=[PD_RNDINT_MAGIC]
+	pcmpeqd	xmm4,xmm4
+	psrld	xmm4,WORD_BIT		; xmm4={0xFFFF 0x0000 0xFFFF 0x0000 ..}
+
+	addps	xmm3,xmm2	; xmm3=roundint(data4/8)=(04 ** 14 ** 24 ** 34 **)
+	addps	xmm7,xmm2	; xmm7=roundint(data5/8)=(05 ** 15 ** 25 ** 35 **)
+	addps	xmm1,xmm2	; xmm1=roundint(data2/8)=(02 ** 12 ** 22 ** 32 **)
+	addps	xmm5,xmm2	; xmm5=roundint(data3/8)=(03 ** 13 ** 23 ** 33 **)
+
+	pand	xmm3,xmm4		; xmm3=(04 -- 14 -- 24 -- 34 --)
+	pslld	xmm7,WORD_BIT		; xmm7=(-- 05 -- 15 -- 25 -- 35)
+	pand	xmm1,xmm4		; xmm1=(02 -- 12 -- 22 -- 32 --)
+	pslld	xmm5,WORD_BIT		; xmm5=(-- 03 -- 13 -- 23 -- 33)
+	por	xmm3,xmm7		; xmm3=(04 05 14 15 24 25 34 35)
+	por	xmm1,xmm5		; xmm1=(02 03 12 13 22 23 32 33)
+
+	movdqa    xmm2,[GOTOFF(ebx,PB_CENTERJSAMP)]	; xmm2=[PB_CENTERJSAMP]
+
+	packsswb  xmm6,xmm3	; xmm6=(00 01 10 11 20 21 30 31 04 05 14 15 24 25 34 35)
+	packsswb  xmm1,xmm0	; xmm1=(02 03 12 13 22 23 32 33 06 07 16 17 26 27 36 37)
+	paddb     xmm6,xmm2
+	paddb     xmm1,xmm2
+
+	movdqa    xmm4,xmm6	; transpose coefficients(phase 2)
+	punpcklwd xmm6,xmm1	; xmm6=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+	punpckhwd xmm4,xmm1	; xmm4=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+
+	movdqa    xmm7,xmm6	; transpose coefficients(phase 3)
+	punpckldq xmm6,xmm4	; xmm6=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+	punpckhdq xmm7,xmm4	; xmm7=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+
+	pshufd	xmm5,xmm6,0x4E	; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+	pshufd	xmm3,xmm7,0x4E	; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+
+	pushpic	ebx			; save GOT address
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6
+	movq	XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE], xmm7
+	mov	edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm5
+	movq	XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE], xmm3
+
+	poppic	ebx			; restore GOT address
+
+	add	esi, byte 4*SIZEOF_FAST_FLOAT	; wsptr
+	add	edi, byte 4*SIZEOF_JSAMPROW
+	dec	ecx				; ctr
+	jnz	near .rowloop
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2fst-64.asm b/simd/jiss2fst-64.asm
new file mode 100644
index 0000000..0887505
--- /dev/null
+++ b/simd/jiss2fst-64.asm
@@ -0,0 +1,492 @@
+;
+; jiss2fst-64.asm - fast integer IDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/projecpt/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the inverse DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jidctfst.c; see the jidctfst.c
+; for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	8	; 14 is also OK.
+%define PASS1_BITS	2
+
+%if IFAST_SCALE_BITS != PASS1_BITS
+%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
+%endif
+
+%if CONST_BITS == 8
+F_1_082	equ	277		; FIX(1.082392200)
+F_1_414	equ	362		; FIX(1.414213562)
+F_1_847	equ	473		; FIX(1.847759065)
+F_2_613	equ	669		; FIX(2.613125930)
+F_1_613	equ	(F_2_613 - 256)	; FIX(2.613125930) - FIX(1)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define	DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_1_082	equ	DESCALE(1162209775,30-CONST_BITS)	; FIX(1.082392200)
+F_1_414	equ	DESCALE(1518500249,30-CONST_BITS)	; FIX(1.414213562)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_2_613	equ	DESCALE(2805822602,30-CONST_BITS)	; FIX(2.613125930)
+F_1_613	equ	(F_2_613 - (1 << CONST_BITS))	; FIX(2.613125930) - FIX(1)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS   2
+%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+	alignz	16
+	global	EXTN(jconst_idct_ifast_sse2)
+
+EXTN(jconst_idct_ifast_sse2):
+
+PW_F1414	times 8 dw  F_1_414 << CONST_SHIFT
+PW_F1847	times 8 dw  F_1_847 << CONST_SHIFT
+PW_MF1613	times 8 dw -F_1_613 << CONST_SHIFT
+PW_F1082	times 8 dw  F_1_082 << CONST_SHIFT
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_ifast_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = jpeg_component_info * compptr
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13 = JDIMENSION output_col
+
+%define original_rbp	rbp+0
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_idct_ifast_sse2)
+
+EXTN(jsimd_idct_ifast_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+
+	; ---- Pass 1: process columns from input.
+
+	mov	rdx, r10	; quantptr
+	mov	rsi, r11		; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
+	mov	eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1,xmm0
+	packsswb xmm1,xmm1
+	packsswb xmm1,xmm1
+	movd	eax,xmm1
+	test	rax,rax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movdqa    xmm7,xmm0		; xmm0=in0=(00 01 02 03 04 05 06 07)
+	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
+	punpckhwd xmm7,xmm7		; xmm7=(04 04 05 05 06 06 07 07)
+
+	pshufd	xmm6,xmm0,0x00		; xmm6=col0=(00 00 00 00 00 00 00 00)
+	pshufd	xmm2,xmm0,0x55		; xmm2=col1=(01 01 01 01 01 01 01 01)
+	pshufd	xmm5,xmm0,0xAA		; xmm5=col2=(02 02 02 02 02 02 02 02)
+	pshufd	xmm0,xmm0,0xFF		; xmm0=col3=(03 03 03 03 03 03 03 03)
+	pshufd	xmm1,xmm7,0x00		; xmm1=col4=(04 04 04 04 04 04 04 04)
+	pshufd	xmm4,xmm7,0x55		; xmm4=col5=(05 05 05 05 05 05 05 05)
+	pshufd	xmm3,xmm7,0xAA		; xmm3=col6=(06 06 06 06 06 06 06 06)
+	pshufd	xmm7,xmm7,0xFF		; xmm7=col7=(07 07 07 07 07 07 07 07)
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=col1
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=col3
+	jmp	near .column_end
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movdqa	xmm4,xmm0
+	movdqa	xmm5,xmm1
+	psubw	xmm0,xmm2		; xmm0=tmp11
+	psubw	xmm1,xmm3
+	paddw	xmm4,xmm2		; xmm4=tmp10
+	paddw	xmm5,xmm3		; xmm5=tmp13
+
+	psllw	xmm1,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm1,[rel PW_F1414]
+	psubw	xmm1,xmm5		; xmm1=tmp12
+
+	movdqa	xmm6,xmm4
+	movdqa	xmm7,xmm0
+	psubw	xmm4,xmm5		; xmm4=tmp3
+	psubw	xmm0,xmm1		; xmm0=tmp2
+	paddw	xmm6,xmm5		; xmm6=tmp0
+	paddw	xmm7,xmm1		; xmm7=tmp1
+
+	movdqa	XMMWORD [wk(1)], xmm4	; wk(1)=tmp3
+	movdqa	XMMWORD [wk(0)], xmm0	; wk(0)=tmp2
+
+	; -- Odd part
+
+	movdqa	xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+	movdqa	xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movdqa	xmm4,xmm2
+	movdqa	xmm0,xmm5
+	psubw	xmm2,xmm1		; xmm2=z12
+	psubw	xmm5,xmm3		; xmm5=z10
+	paddw	xmm4,xmm1		; xmm4=z11
+	paddw	xmm0,xmm3		; xmm0=z13
+
+	movdqa	xmm1,xmm5		; xmm1=z10(unscaled)
+	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+
+	movdqa	xmm3,xmm4
+	psubw	xmm4,xmm0
+	paddw	xmm3,xmm0		; xmm3=tmp7
+
+	psllw	xmm4,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm4,[rel PW_F1414]	; xmm4=tmp11
+
+	; To avoid overflow...
+	;
+	; (Original)
+	; tmp12 = -2.613125930 * z10 + z5;
+	;
+	; (This implementation)
+	; tmp12 = (-1.613125930 - 1) * z10 + z5;
+	;       = -1.613125930 * z10 - z10 + z5;
+
+	movdqa	xmm0,xmm5
+	paddw	xmm5,xmm2
+	pmulhw	xmm5,[rel PW_F1847]	; xmm5=z5
+	pmulhw	xmm0,[rel PW_MF1613]
+	pmulhw	xmm2,[rel PW_F1082]
+	psubw	xmm0,xmm1
+	psubw	xmm2,xmm5		; xmm2=tmp10
+	paddw	xmm0,xmm5		; xmm0=tmp12
+
+	; -- Final output stage
+
+	psubw	xmm0,xmm3		; xmm0=tmp6
+	movdqa	xmm1,xmm6
+	movdqa	xmm5,xmm7
+	paddw	xmm6,xmm3		; xmm6=data0=(00 01 02 03 04 05 06 07)
+	paddw	xmm7,xmm0		; xmm7=data1=(10 11 12 13 14 15 16 17)
+	psubw	xmm1,xmm3		; xmm1=data7=(70 71 72 73 74 75 76 77)
+	psubw	xmm5,xmm0		; xmm5=data6=(60 61 62 63 64 65 66 67)
+	psubw	xmm4,xmm0		; xmm4=tmp5
+
+	movdqa    xmm3,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm7		; xmm6=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm3,xmm7		; xmm3=(04 14 05 15 06 16 07 17)
+	movdqa    xmm0,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm1		; xmm5=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm0,xmm1		; xmm0=(64 74 65 75 66 76 67 77)
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
+	movdqa	xmm1, XMMWORD [wk(1)]	; xmm1=tmp3
+
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(60 70 61 71 62 72 63 73)
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(64 74 65 75 66 76 67 77)
+
+	paddw	xmm2,xmm4		; xmm2=tmp4
+	movdqa	xmm5,xmm7
+	movdqa	xmm0,xmm1
+	paddw	xmm7,xmm4		; xmm7=data2=(20 21 22 23 24 25 26 27)
+	paddw	xmm1,xmm2		; xmm1=data4=(40 41 42 43 44 45 46 47)
+	psubw	xmm5,xmm4		; xmm5=data5=(50 51 52 53 54 55 56 57)
+	psubw	xmm0,xmm2		; xmm0=data3=(30 31 32 33 34 35 36 37)
+
+	movdqa    xmm4,xmm7		; transpose coefficients(phase 1)
+	punpcklwd xmm7,xmm0		; xmm7=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm4,xmm0		; xmm4=(24 34 25 35 26 36 27 37)
+	movdqa    xmm2,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm5		; xmm1=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm2,xmm5		; xmm2=(44 54 45 55 46 56 47 57)
+
+	movdqa    xmm0,xmm3		; transpose coefficients(phase 2)
+	punpckldq xmm3,xmm4		; xmm3=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm0,xmm4		; xmm0=(06 16 26 36 07 17 27 37)
+	movdqa    xmm5,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm7		; xmm6=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm5,xmm7		; xmm5=(02 12 22 32 03 13 23 33)
+
+	movdqa	xmm4, XMMWORD [wk(0)]	; xmm4=(60 70 61 71 62 72 63 73)
+	movdqa	xmm7, XMMWORD [wk(1)]	; xmm7=(64 74 65 75 66 76 67 77)
+
+	movdqa	XMMWORD [wk(0)], xmm3	; wk(0)=(04 14 24 34 05 15 25 35)
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(06 16 26 36 07 17 27 37)
+
+	movdqa    xmm3,xmm1		; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm4		; xmm1=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm3,xmm4		; xmm3=(42 52 62 72 43 53 63 73)
+	movdqa    xmm0,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm7		; xmm2=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm0,xmm7		; xmm0=(46 56 66 76 47 57 67 77)
+
+	movdqa     xmm4,xmm6		; transpose coefficients(phase 3)
+	punpcklqdq xmm6,xmm1		; xmm6=col0=(00 10 20 30 40 50 60 70)
+	punpckhqdq xmm4,xmm1		; xmm4=col1=(01 11 21 31 41 51 61 71)
+	movdqa     xmm7,xmm5		; transpose coefficients(phase 3)
+	punpcklqdq xmm5,xmm3		; xmm5=col2=(02 12 22 32 42 52 62 72)
+	punpckhqdq xmm7,xmm3		; xmm7=col3=(03 13 23 33 43 53 63 73)
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(04 14 24 34 05 15 25 35)
+	movdqa	xmm3, XMMWORD [wk(1)]	; xmm3=(06 16 26 36 07 17 27 37)
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=col1
+	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=col3
+
+	movdqa     xmm4,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm2		; xmm1=col4=(04 14 24 34 44 54 64 74)
+	punpckhqdq xmm4,xmm2		; xmm4=col5=(05 15 25 35 45 55 65 75)
+	movdqa     xmm7,xmm3		; transpose coefficients(phase 3)
+	punpcklqdq xmm3,xmm0		; xmm3=col6=(06 16 26 36 46 56 66 76)
+	punpckhqdq xmm7,xmm0		; xmm7=col7=(07 17 27 37 47 57 67 77)
+.column_end:
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	rax, [original_rbp]
+	mov	rdi, r12	; (JSAMPROW *)
+	mov	rax, r13
+
+	; -- Even part
+
+	; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
+
+	movdqa	xmm2,xmm6
+	movdqa	xmm0,xmm5
+	psubw	xmm6,xmm1		; xmm6=tmp11
+	psubw	xmm5,xmm3
+	paddw	xmm2,xmm1		; xmm2=tmp10
+	paddw	xmm0,xmm3		; xmm0=tmp13
+
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm5,[rel PW_F1414]
+	psubw	xmm5,xmm0		; xmm5=tmp12
+
+	movdqa	xmm1,xmm2
+	movdqa	xmm3,xmm6
+	psubw	xmm2,xmm0		; xmm2=tmp3
+	psubw	xmm6,xmm5		; xmm6=tmp2
+	paddw	xmm1,xmm0		; xmm1=tmp0
+	paddw	xmm3,xmm5		; xmm3=tmp1
+
+	movdqa	xmm0, XMMWORD [wk(0)]	; xmm0=col1
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=col3
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=tmp3
+	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=tmp2
+
+	; -- Odd part
+
+	; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
+
+	movdqa	xmm2,xmm0
+	movdqa	xmm6,xmm4
+	psubw	xmm0,xmm7		; xmm0=z12
+	psubw	xmm4,xmm5		; xmm4=z10
+	paddw	xmm2,xmm7		; xmm2=z11
+	paddw	xmm6,xmm5		; xmm6=z13
+
+	movdqa	xmm7,xmm4		; xmm7=z10(unscaled)
+	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm4,PRE_MULTIPLY_SCALE_BITS
+
+	movdqa	xmm5,xmm2
+	psubw	xmm2,xmm6
+	paddw	xmm5,xmm6		; xmm5=tmp7
+
+	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm2,[rel PW_F1414]	; xmm2=tmp11
+
+	; To avoid overflow...
+	;
+	; (Original)
+	; tmp12 = -2.613125930 * z10 + z5;
+	;
+	; (This implementation)
+	; tmp12 = (-1.613125930 - 1) * z10 + z5;
+	;       = -1.613125930 * z10 - z10 + z5;
+
+	movdqa	xmm6,xmm4
+	paddw	xmm4,xmm0
+	pmulhw	xmm4,[rel PW_F1847]	; xmm4=z5
+	pmulhw	xmm6,[rel PW_MF1613]
+	pmulhw	xmm0,[rel PW_F1082]
+	psubw	xmm6,xmm7
+	psubw	xmm0,xmm4		; xmm0=tmp10
+	paddw	xmm6,xmm4		; xmm6=tmp12
+
+	; -- Final output stage
+
+	psubw	xmm6,xmm5		; xmm6=tmp6
+	movdqa	xmm7,xmm1
+	movdqa	xmm4,xmm3
+	paddw	xmm1,xmm5		; xmm1=data0=(00 10 20 30 40 50 60 70)
+	paddw	xmm3,xmm6		; xmm3=data1=(01 11 21 31 41 51 61 71)
+	psraw	xmm1,(PASS1_BITS+3)	; descale
+	psraw	xmm3,(PASS1_BITS+3)	; descale
+	psubw	xmm7,xmm5		; xmm7=data7=(07 17 27 37 47 57 67 77)
+	psubw	xmm4,xmm6		; xmm4=data6=(06 16 26 36 46 56 66 76)
+	psraw	xmm7,(PASS1_BITS+3)	; descale
+	psraw	xmm4,(PASS1_BITS+3)	; descale
+	psubw	xmm2,xmm6		; xmm2=tmp5
+
+	packsswb  xmm1,xmm4	; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	packsswb  xmm3,xmm7	; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp2
+	movdqa	xmm6, XMMWORD [wk(0)]	; xmm6=tmp3
+
+	paddw	xmm0,xmm2		; xmm0=tmp4
+	movdqa	xmm4,xmm5
+	movdqa	xmm7,xmm6
+	paddw	xmm5,xmm2		; xmm5=data2=(02 12 22 32 42 52 62 72)
+	paddw	xmm6,xmm0		; xmm6=data4=(04 14 24 34 44 54 64 74)
+	psraw	xmm5,(PASS1_BITS+3)	; descale
+	psraw	xmm6,(PASS1_BITS+3)	; descale
+	psubw	xmm4,xmm2		; xmm4=data5=(05 15 25 35 45 55 65 75)
+	psubw	xmm7,xmm0		; xmm7=data3=(03 13 23 33 43 53 63 73)
+	psraw	xmm4,(PASS1_BITS+3)	; descale
+	psraw	xmm7,(PASS1_BITS+3)	; descale
+
+	movdqa    xmm2,[rel PB_CENTERJSAMP]	; xmm2=[rel PB_CENTERJSAMP]
+
+	packsswb  xmm5,xmm6	; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
+	packsswb  xmm7,xmm4	; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
+
+	paddb     xmm1,xmm2
+	paddb     xmm3,xmm2
+	paddb     xmm5,xmm2
+	paddb     xmm7,xmm2
+
+	movdqa    xmm0,xmm1	; transpose coefficients(phase 1)
+	punpcklbw xmm1,xmm3	; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
+	punpckhbw xmm0,xmm3	; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
+	movdqa    xmm6,xmm5	; transpose coefficients(phase 1)
+	punpcklbw xmm5,xmm7	; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
+	punpckhbw xmm6,xmm7	; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
+
+	movdqa    xmm4,xmm1	; transpose coefficients(phase 2)
+	punpcklwd xmm1,xmm5	; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+	punpckhwd xmm4,xmm5	; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
+	movdqa    xmm2,xmm6	; transpose coefficients(phase 2)
+	punpcklwd xmm6,xmm0	; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+	punpckhwd xmm2,xmm0	; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
+
+	movdqa    xmm3,xmm1	; transpose coefficients(phase 3)
+	punpckldq xmm1,xmm6	; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+	punpckhdq xmm3,xmm6	; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+	movdqa    xmm7,xmm4	; transpose coefficients(phase 3)
+	punpckldq xmm4,xmm2	; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+	punpckhdq xmm7,xmm2	; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+	pshufd	xmm5,xmm1,0x4E	; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+	pshufd	xmm0,xmm3,0x4E	; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+	pshufd	xmm6,xmm4,0x4E	; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+	pshufd	xmm2,xmm7,0x4E	; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+	mov	rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+	mov	rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
+
+	mov	rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
+	mov	rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
+
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2fst.asm b/simd/jiss2fst.asm
new file mode 100644
index 0000000..b53664d
--- /dev/null
+++ b/simd/jiss2fst.asm
@@ -0,0 +1,502 @@
+;
+; jiss2fst.asm - fast integer IDCT (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a fast, not so accurate integer implementation of
+; the inverse DCT (Discrete Cosine Transform). The following code is
+; based directly on the IJG's original jidctfst.c; see the jidctfst.c
+; for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	8	; 14 is also OK.
+%define PASS1_BITS	2
+
+%if IFAST_SCALE_BITS != PASS1_BITS
+%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
+%endif
+
+%if CONST_BITS == 8
+F_1_082	equ	277		; FIX(1.082392200)
+F_1_414	equ	362		; FIX(1.414213562)
+F_1_847	equ	473		; FIX(1.847759065)
+F_2_613	equ	669		; FIX(2.613125930)
+F_1_613	equ	(F_2_613 - 256)	; FIX(2.613125930) - FIX(1)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define	DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_1_082	equ	DESCALE(1162209775,30-CONST_BITS)	; FIX(1.082392200)
+F_1_414	equ	DESCALE(1518500249,30-CONST_BITS)	; FIX(1.414213562)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_2_613	equ	DESCALE(2805822602,30-CONST_BITS)	; FIX(2.613125930)
+F_1_613	equ	(F_2_613 - (1 << CONST_BITS))	; FIX(2.613125930) - FIX(1)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
+; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
+
+%define PRE_MULTIPLY_SCALE_BITS   2
+%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
+
+	alignz	16
+	global	EXTN(jconst_idct_ifast_sse2)
+
+EXTN(jconst_idct_ifast_sse2):
+
+PW_F1414	times 8 dw  F_1_414 << CONST_SHIFT
+PW_F1847	times 8 dw  F_1_847 << CONST_SHIFT
+PW_MF1613	times 8 dw -F_1_613 << CONST_SHIFT
+PW_F1082	times 8 dw  F_1_082 << CONST_SHIFT
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_ifast_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_idct_ifast_sse2)
+
+EXTN(jsimd_idct_ifast_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	xmm1,xmm0
+	packsswb xmm1,xmm1
+	packsswb xmm1,xmm1
+	movd	eax,xmm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movdqa    xmm7,xmm0		; xmm0=in0=(00 01 02 03 04 05 06 07)
+	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
+	punpckhwd xmm7,xmm7		; xmm7=(04 04 05 05 06 06 07 07)
+
+	pshufd	xmm6,xmm0,0x00		; xmm6=col0=(00 00 00 00 00 00 00 00)
+	pshufd	xmm2,xmm0,0x55		; xmm2=col1=(01 01 01 01 01 01 01 01)
+	pshufd	xmm5,xmm0,0xAA		; xmm5=col2=(02 02 02 02 02 02 02 02)
+	pshufd	xmm0,xmm0,0xFF		; xmm0=col3=(03 03 03 03 03 03 03 03)
+	pshufd	xmm1,xmm7,0x00		; xmm1=col4=(04 04 04 04 04 04 04 04)
+	pshufd	xmm4,xmm7,0x55		; xmm4=col5=(05 05 05 05 05 05 05 05)
+	pshufd	xmm3,xmm7,0xAA		; xmm3=col6=(06 06 06 06 06 06 06 06)
+	pshufd	xmm7,xmm7,0xFF		; xmm7=col7=(07 07 07 07 07 07 07 07)
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=col1
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=col3
+	jmp	near .column_end
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movdqa	xmm4,xmm0
+	movdqa	xmm5,xmm1
+	psubw	xmm0,xmm2		; xmm0=tmp11
+	psubw	xmm1,xmm3
+	paddw	xmm4,xmm2		; xmm4=tmp10
+	paddw	xmm5,xmm3		; xmm5=tmp13
+
+	psllw	xmm1,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm1,[GOTOFF(ebx,PW_F1414)]
+	psubw	xmm1,xmm5		; xmm1=tmp12
+
+	movdqa	xmm6,xmm4
+	movdqa	xmm7,xmm0
+	psubw	xmm4,xmm5		; xmm4=tmp3
+	psubw	xmm0,xmm1		; xmm0=tmp2
+	paddw	xmm6,xmm5		; xmm6=tmp0
+	paddw	xmm7,xmm1		; xmm7=tmp1
+
+	movdqa	XMMWORD [wk(1)], xmm4	; wk(1)=tmp3
+	movdqa	XMMWORD [wk(0)], xmm0	; wk(0)=tmp2
+
+	; -- Odd part
+
+	movdqa	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	movdqa	xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)]
+
+	movdqa	xmm4,xmm2
+	movdqa	xmm0,xmm5
+	psubw	xmm2,xmm1		; xmm2=z12
+	psubw	xmm5,xmm3		; xmm5=z10
+	paddw	xmm4,xmm1		; xmm4=z11
+	paddw	xmm0,xmm3		; xmm0=z13
+
+	movdqa	xmm1,xmm5		; xmm1=z10(unscaled)
+	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+
+	movdqa	xmm3,xmm4
+	psubw	xmm4,xmm0
+	paddw	xmm3,xmm0		; xmm3=tmp7
+
+	psllw	xmm4,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm4,[GOTOFF(ebx,PW_F1414)]	; xmm4=tmp11
+
+	; To avoid overflow...
+	;
+	; (Original)
+	; tmp12 = -2.613125930 * z10 + z5;
+	;
+	; (This implementation)
+	; tmp12 = (-1.613125930 - 1) * z10 + z5;
+	;       = -1.613125930 * z10 - z10 + z5;
+
+	movdqa	xmm0,xmm5
+	paddw	xmm5,xmm2
+	pmulhw	xmm5,[GOTOFF(ebx,PW_F1847)]	; xmm5=z5
+	pmulhw	xmm0,[GOTOFF(ebx,PW_MF1613)]
+	pmulhw	xmm2,[GOTOFF(ebx,PW_F1082)]
+	psubw	xmm0,xmm1
+	psubw	xmm2,xmm5		; xmm2=tmp10
+	paddw	xmm0,xmm5		; xmm0=tmp12
+
+	; -- Final output stage
+
+	psubw	xmm0,xmm3		; xmm0=tmp6
+	movdqa	xmm1,xmm6
+	movdqa	xmm5,xmm7
+	paddw	xmm6,xmm3		; xmm6=data0=(00 01 02 03 04 05 06 07)
+	paddw	xmm7,xmm0		; xmm7=data1=(10 11 12 13 14 15 16 17)
+	psubw	xmm1,xmm3		; xmm1=data7=(70 71 72 73 74 75 76 77)
+	psubw	xmm5,xmm0		; xmm5=data6=(60 61 62 63 64 65 66 67)
+	psubw	xmm4,xmm0		; xmm4=tmp5
+
+	movdqa    xmm3,xmm6		; transpose coefficients(phase 1)
+	punpcklwd xmm6,xmm7		; xmm6=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm3,xmm7		; xmm3=(04 14 05 15 06 16 07 17)
+	movdqa    xmm0,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm1		; xmm5=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm0,xmm1		; xmm0=(64 74 65 75 66 76 67 77)
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
+	movdqa	xmm1, XMMWORD [wk(1)]	; xmm1=tmp3
+
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(60 70 61 71 62 72 63 73)
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(64 74 65 75 66 76 67 77)
+
+	paddw	xmm2,xmm4		; xmm2=tmp4
+	movdqa	xmm5,xmm7
+	movdqa	xmm0,xmm1
+	paddw	xmm7,xmm4		; xmm7=data2=(20 21 22 23 24 25 26 27)
+	paddw	xmm1,xmm2		; xmm1=data4=(40 41 42 43 44 45 46 47)
+	psubw	xmm5,xmm4		; xmm5=data5=(50 51 52 53 54 55 56 57)
+	psubw	xmm0,xmm2		; xmm0=data3=(30 31 32 33 34 35 36 37)
+
+	movdqa    xmm4,xmm7		; transpose coefficients(phase 1)
+	punpcklwd xmm7,xmm0		; xmm7=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm4,xmm0		; xmm4=(24 34 25 35 26 36 27 37)
+	movdqa    xmm2,xmm1		; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm5		; xmm1=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm2,xmm5		; xmm2=(44 54 45 55 46 56 47 57)
+
+	movdqa    xmm0,xmm3		; transpose coefficients(phase 2)
+	punpckldq xmm3,xmm4		; xmm3=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm0,xmm4		; xmm0=(06 16 26 36 07 17 27 37)
+	movdqa    xmm5,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm7		; xmm6=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm5,xmm7		; xmm5=(02 12 22 32 03 13 23 33)
+
+	movdqa	xmm4, XMMWORD [wk(0)]	; xmm4=(60 70 61 71 62 72 63 73)
+	movdqa	xmm7, XMMWORD [wk(1)]	; xmm7=(64 74 65 75 66 76 67 77)
+
+	movdqa	XMMWORD [wk(0)], xmm3	; wk(0)=(04 14 24 34 05 15 25 35)
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(06 16 26 36 07 17 27 37)
+
+	movdqa    xmm3,xmm1		; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm4		; xmm1=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm3,xmm4		; xmm3=(42 52 62 72 43 53 63 73)
+	movdqa    xmm0,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm7		; xmm2=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm0,xmm7		; xmm0=(46 56 66 76 47 57 67 77)
+
+	movdqa     xmm4,xmm6		; transpose coefficients(phase 3)
+	punpcklqdq xmm6,xmm1		; xmm6=col0=(00 10 20 30 40 50 60 70)
+	punpckhqdq xmm4,xmm1		; xmm4=col1=(01 11 21 31 41 51 61 71)
+	movdqa     xmm7,xmm5		; transpose coefficients(phase 3)
+	punpcklqdq xmm5,xmm3		; xmm5=col2=(02 12 22 32 42 52 62 72)
+	punpckhqdq xmm7,xmm3		; xmm7=col3=(03 13 23 33 43 53 63 73)
+
+	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(04 14 24 34 05 15 25 35)
+	movdqa	xmm3, XMMWORD [wk(1)]	; xmm3=(06 16 26 36 07 17 27 37)
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=col1
+	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=col3
+
+	movdqa     xmm4,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm2		; xmm1=col4=(04 14 24 34 44 54 64 74)
+	punpckhqdq xmm4,xmm2		; xmm4=col5=(05 15 25 35 45 55 65 75)
+	movdqa     xmm7,xmm3		; transpose coefficients(phase 3)
+	punpcklqdq xmm3,xmm0		; xmm3=col6=(06 16 26 36 46 56 66 76)
+	punpckhqdq xmm7,xmm0		; xmm7=col7=(07 17 27 37 47 57 67 77)
+.column_end:
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+
+	; -- Even part
+
+	; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
+
+	movdqa	xmm2,xmm6
+	movdqa	xmm0,xmm5
+	psubw	xmm6,xmm1		; xmm6=tmp11
+	psubw	xmm5,xmm3
+	paddw	xmm2,xmm1		; xmm2=tmp10
+	paddw	xmm0,xmm3		; xmm0=tmp13
+
+	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm5,[GOTOFF(ebx,PW_F1414)]
+	psubw	xmm5,xmm0		; xmm5=tmp12
+
+	movdqa	xmm1,xmm2
+	movdqa	xmm3,xmm6
+	psubw	xmm2,xmm0		; xmm2=tmp3
+	psubw	xmm6,xmm5		; xmm6=tmp2
+	paddw	xmm1,xmm0		; xmm1=tmp0
+	paddw	xmm3,xmm5		; xmm3=tmp1
+
+	movdqa	xmm0, XMMWORD [wk(0)]	; xmm0=col1
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=col3
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=tmp3
+	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=tmp2
+
+	; -- Odd part
+
+	; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
+
+	movdqa	xmm2,xmm0
+	movdqa	xmm6,xmm4
+	psubw	xmm0,xmm7		; xmm0=z12
+	psubw	xmm4,xmm5		; xmm4=z10
+	paddw	xmm2,xmm7		; xmm2=z11
+	paddw	xmm6,xmm5		; xmm6=z13
+
+	movdqa	xmm7,xmm4		; xmm7=z10(unscaled)
+	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
+	psllw	xmm4,PRE_MULTIPLY_SCALE_BITS
+
+	movdqa	xmm5,xmm2
+	psubw	xmm2,xmm6
+	paddw	xmm5,xmm6		; xmm5=tmp7
+
+	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
+	pmulhw	xmm2,[GOTOFF(ebx,PW_F1414)]	; xmm2=tmp11
+
+	; To avoid overflow...
+	;
+	; (Original)
+	; tmp12 = -2.613125930 * z10 + z5;
+	;
+	; (This implementation)
+	; tmp12 = (-1.613125930 - 1) * z10 + z5;
+	;       = -1.613125930 * z10 - z10 + z5;
+
+	movdqa	xmm6,xmm4
+	paddw	xmm4,xmm0
+	pmulhw	xmm4,[GOTOFF(ebx,PW_F1847)]	; xmm4=z5
+	pmulhw	xmm6,[GOTOFF(ebx,PW_MF1613)]
+	pmulhw	xmm0,[GOTOFF(ebx,PW_F1082)]
+	psubw	xmm6,xmm7
+	psubw	xmm0,xmm4		; xmm0=tmp10
+	paddw	xmm6,xmm4		; xmm6=tmp12
+
+	; -- Final output stage
+
+	psubw	xmm6,xmm5		; xmm6=tmp6
+	movdqa	xmm7,xmm1
+	movdqa	xmm4,xmm3
+	paddw	xmm1,xmm5		; xmm1=data0=(00 10 20 30 40 50 60 70)
+	paddw	xmm3,xmm6		; xmm3=data1=(01 11 21 31 41 51 61 71)
+	psraw	xmm1,(PASS1_BITS+3)	; descale
+	psraw	xmm3,(PASS1_BITS+3)	; descale
+	psubw	xmm7,xmm5		; xmm7=data7=(07 17 27 37 47 57 67 77)
+	psubw	xmm4,xmm6		; xmm4=data6=(06 16 26 36 46 56 66 76)
+	psraw	xmm7,(PASS1_BITS+3)	; descale
+	psraw	xmm4,(PASS1_BITS+3)	; descale
+	psubw	xmm2,xmm6		; xmm2=tmp5
+
+	packsswb  xmm1,xmm4	; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	packsswb  xmm3,xmm7	; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp2
+	movdqa	xmm6, XMMWORD [wk(0)]	; xmm6=tmp3
+
+	paddw	xmm0,xmm2		; xmm0=tmp4
+	movdqa	xmm4,xmm5
+	movdqa	xmm7,xmm6
+	paddw	xmm5,xmm2		; xmm5=data2=(02 12 22 32 42 52 62 72)
+	paddw	xmm6,xmm0		; xmm6=data4=(04 14 24 34 44 54 64 74)
+	psraw	xmm5,(PASS1_BITS+3)	; descale
+	psraw	xmm6,(PASS1_BITS+3)	; descale
+	psubw	xmm4,xmm2		; xmm4=data5=(05 15 25 35 45 55 65 75)
+	psubw	xmm7,xmm0		; xmm7=data3=(03 13 23 33 43 53 63 73)
+	psraw	xmm4,(PASS1_BITS+3)	; descale
+	psraw	xmm7,(PASS1_BITS+3)	; descale
+
+	movdqa    xmm2,[GOTOFF(ebx,PB_CENTERJSAMP)]	; xmm2=[PB_CENTERJSAMP]
+
+	packsswb  xmm5,xmm6	; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
+	packsswb  xmm7,xmm4	; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
+
+	paddb     xmm1,xmm2
+	paddb     xmm3,xmm2
+	paddb     xmm5,xmm2
+	paddb     xmm7,xmm2
+
+	movdqa    xmm0,xmm1	; transpose coefficients(phase 1)
+	punpcklbw xmm1,xmm3	; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
+	punpckhbw xmm0,xmm3	; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
+	movdqa    xmm6,xmm5	; transpose coefficients(phase 1)
+	punpcklbw xmm5,xmm7	; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
+	punpckhbw xmm6,xmm7	; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
+
+	movdqa    xmm4,xmm1	; transpose coefficients(phase 2)
+	punpcklwd xmm1,xmm5	; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+	punpckhwd xmm4,xmm5	; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
+	movdqa    xmm2,xmm6	; transpose coefficients(phase 2)
+	punpcklwd xmm6,xmm0	; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+	punpckhwd xmm2,xmm0	; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
+
+	movdqa    xmm3,xmm1	; transpose coefficients(phase 3)
+	punpckldq xmm1,xmm6	; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+	punpckhdq xmm3,xmm6	; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+	movdqa    xmm7,xmm4	; transpose coefficients(phase 3)
+	punpckldq xmm4,xmm2	; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+	punpckhdq xmm7,xmm2	; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+	pshufd	xmm5,xmm1,0x4E	; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+	pshufd	xmm0,xmm3,0x4E	; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+	pshufd	xmm6,xmm4,0x4E	; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+	pshufd	xmm2,xmm7,0x4E	; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm1
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
+	mov	edx, JSAMPROW [edi+4*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+6*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm7
+
+	mov	edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm5
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm0
+	mov	edx, JSAMPROW [edi+5*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+7*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm2
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2int-64.asm b/simd/jiss2int-64.asm
new file mode 100644
index 0000000..13764d6
--- /dev/null
+++ b/simd/jiss2int-64.asm
@@ -0,0 +1,848 @@
+;
+; jiss2int-64.asm - accurate integer IDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slow-but-accurate integer implementation of the
+; inverse DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jidctint.c; see the jidctint.c for
+; more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
+%define DESCALE_P2	(CONST_BITS+PASS1_BITS+3)
+
+%if CONST_BITS == 13
+F_0_298	equ	 2446		; FIX(0.298631336)
+F_0_390	equ	 3196		; FIX(0.390180644)
+F_0_541	equ	 4433		; FIX(0.541196100)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_175	equ	 9633		; FIX(1.175875602)
+F_1_501	equ	12299		; FIX(1.501321110)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_1_961	equ	16069		; FIX(1.961570560)
+F_2_053	equ	16819		; FIX(2.053119869)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_072	equ	25172		; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
+F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
+F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
+F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_islow_sse2)
+
+EXTN(jconst_idct_islow_sse2):
+
+PW_F130_F054	times 4 dw  (F_0_541+F_0_765), F_0_541
+PW_F054_MF130	times 4 dw  F_0_541, (F_0_541-F_1_847)
+PW_MF078_F117	times 4 dw  (F_1_175-F_1_961), F_1_175
+PW_F117_F078	times 4 dw  F_1_175, (F_1_175-F_0_390)
+PW_MF060_MF089	times 4 dw  (F_0_298-F_0_899),-F_0_899
+PW_MF089_F060	times 4 dw -F_0_899, (F_1_501-F_0_899)
+PW_MF050_MF256	times 4 dw  (F_2_053-F_2_562),-F_2_562
+PW_MF256_F050	times 4 dw -F_2_562, (F_3_072-F_2_562)
+PD_DESCALE_P1	times 4 dd  1 << (DESCALE_P1-1)
+PD_DESCALE_P2	times 4 dd  1 << (DESCALE_P2-1)
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_islow_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                        JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = jpeg_component_info * compptr
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13 = JDIMENSION output_col
+
+%define original_rbp	rbp+0
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		12
+
+	align	16
+	global	EXTN(jsimd_idct_islow_sse2)
+
+EXTN(jsimd_idct_islow_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+
+	; ---- Pass 1: process columns from input.
+
+	mov	rdx, r10	; quantptr
+	mov	rsi, r11		; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_ISLOW_SSE2
+	mov	eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1,xmm0
+	packsswb xmm1,xmm1
+	packsswb xmm1,xmm1
+	movd	eax,xmm1
+	test	rax,rax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movdqa	xmm5, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm5, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	psllw	xmm5,PASS1_BITS
+
+	movdqa    xmm4,xmm5		; xmm5=in0=(00 01 02 03 04 05 06 07)
+	punpcklwd xmm5,xmm5		; xmm5=(00 00 01 01 02 02 03 03)
+	punpckhwd xmm4,xmm4		; xmm4=(04 04 05 05 06 06 07 07)
+
+	pshufd	xmm7,xmm5,0x00		; xmm7=col0=(00 00 00 00 00 00 00 00)
+	pshufd	xmm6,xmm5,0x55		; xmm6=col1=(01 01 01 01 01 01 01 01)
+	pshufd	xmm1,xmm5,0xAA		; xmm1=col2=(02 02 02 02 02 02 02 02)
+	pshufd	xmm5,xmm5,0xFF		; xmm5=col3=(03 03 03 03 03 03 03 03)
+	pshufd	xmm0,xmm4,0x00		; xmm0=col4=(04 04 04 04 04 04 04 04)
+	pshufd	xmm3,xmm4,0x55		; xmm3=col5=(05 05 05 05 05 05 05 05)
+	pshufd	xmm2,xmm4,0xAA		; xmm2=col6=(06 06 06 06 06 06 06 06)
+	pshufd	xmm4,xmm4,0xFF		; xmm4=col7=(07 07 07 07 07 07 07 07)
+
+	movdqa	XMMWORD [wk(8)], xmm6	; wk(8)=col1
+	movdqa	XMMWORD [wk(9)], xmm5	; wk(9)=col3
+	movdqa	XMMWORD [wk(10)], xmm3	; wk(10)=col5
+	movdqa	XMMWORD [wk(11)], xmm4	; wk(11)=col7
+	jmp	near .column_end
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; (Original)
+	; z1 = (z2 + z3) * 0.541196100;
+	; tmp2 = z1 + z3 * -1.847759065;
+	; tmp3 = z1 + z2 * 0.765366865;
+	;
+	; (This implementation)
+	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+	movdqa    xmm4,xmm1		; xmm1=in2=z2
+	movdqa    xmm5,xmm1
+	punpcklwd xmm4,xmm3		; xmm3=in6=z3
+	punpckhwd xmm5,xmm3
+	movdqa    xmm1,xmm4
+	movdqa    xmm3,xmm5
+	pmaddwd   xmm4,[rel PW_F130_F054]	; xmm4=tmp3L
+	pmaddwd   xmm5,[rel PW_F130_F054]	; xmm5=tmp3H
+	pmaddwd   xmm1,[rel PW_F054_MF130]	; xmm1=tmp2L
+	pmaddwd   xmm3,[rel PW_F054_MF130]	; xmm3=tmp2H
+
+	movdqa    xmm6,xmm0
+	paddw     xmm0,xmm2		; xmm0=in0+in4
+	psubw     xmm6,xmm2		; xmm6=in0-in4
+
+	pxor      xmm7,xmm7
+	pxor      xmm2,xmm2
+	punpcklwd xmm7,xmm0		; xmm7=tmp0L
+	punpckhwd xmm2,xmm0		; xmm2=tmp0H
+	psrad     xmm7,(16-CONST_BITS)	; psrad xmm7,16 & pslld xmm7,CONST_BITS
+	psrad     xmm2,(16-CONST_BITS)	; psrad xmm2,16 & pslld xmm2,CONST_BITS
+
+	movdqa	xmm0,xmm7
+	paddd	xmm7,xmm4		; xmm7=tmp10L
+	psubd	xmm0,xmm4		; xmm0=tmp13L
+	movdqa	xmm4,xmm2
+	paddd	xmm2,xmm5		; xmm2=tmp10H
+	psubd	xmm4,xmm5		; xmm4=tmp13H
+
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=tmp10L
+	movdqa	XMMWORD [wk(1)], xmm2	; wk(1)=tmp10H
+	movdqa	XMMWORD [wk(2)], xmm0	; wk(2)=tmp13L
+	movdqa	XMMWORD [wk(3)], xmm4	; wk(3)=tmp13H
+
+	pxor      xmm5,xmm5
+	pxor      xmm7,xmm7
+	punpcklwd xmm5,xmm6		; xmm5=tmp1L
+	punpckhwd xmm7,xmm6		; xmm7=tmp1H
+	psrad     xmm5,(16-CONST_BITS)	; psrad xmm5,16 & pslld xmm5,CONST_BITS
+	psrad     xmm7,(16-CONST_BITS)	; psrad xmm7,16 & pslld xmm7,CONST_BITS
+
+	movdqa	xmm2,xmm5
+	paddd	xmm5,xmm1		; xmm5=tmp11L
+	psubd	xmm2,xmm1		; xmm2=tmp12L
+	movdqa	xmm0,xmm7
+	paddd	xmm7,xmm3		; xmm7=tmp11H
+	psubd	xmm0,xmm3		; xmm0=tmp12H
+
+	movdqa	XMMWORD [wk(4)], xmm5	; wk(4)=tmp11L
+	movdqa	XMMWORD [wk(5)], xmm7	; wk(5)=tmp11H
+	movdqa	XMMWORD [wk(6)], xmm2	; wk(6)=tmp12L
+	movdqa	XMMWORD [wk(7)], xmm0	; wk(7)=tmp12H
+
+	; -- Odd part
+
+	movdqa	xmm4, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm6, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm4, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm6, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movdqa	xmm5,xmm6
+	movdqa	xmm7,xmm4
+	paddw	xmm5,xmm3		; xmm5=z3
+	paddw	xmm7,xmm1		; xmm7=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm2,xmm5
+	movdqa    xmm0,xmm5
+	punpcklwd xmm2,xmm7
+	punpckhwd xmm0,xmm7
+	movdqa    xmm5,xmm2
+	movdqa    xmm7,xmm0
+	pmaddwd   xmm2,[rel PW_MF078_F117]	; xmm2=z3L
+	pmaddwd   xmm0,[rel PW_MF078_F117]	; xmm0=z3H
+	pmaddwd   xmm5,[rel PW_F117_F078]	; xmm5=z4L
+	pmaddwd   xmm7,[rel PW_F117_F078]	; xmm7=z4H
+
+	movdqa	XMMWORD [wk(10)], xmm2	; wk(10)=z3L
+	movdqa	XMMWORD [wk(11)], xmm0	; wk(11)=z3H
+
+	; (Original)
+	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+	;
+	; (This implementation)
+	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+	; tmp0 += z3;  tmp1 += z4;
+	; tmp2 += z3;  tmp3 += z4;
+
+	movdqa    xmm2,xmm3
+	movdqa    xmm0,xmm3
+	punpcklwd xmm2,xmm4
+	punpckhwd xmm0,xmm4
+	movdqa    xmm3,xmm2
+	movdqa    xmm4,xmm0
+	pmaddwd   xmm2,[rel PW_MF060_MF089]	; xmm2=tmp0L
+	pmaddwd   xmm0,[rel PW_MF060_MF089]	; xmm0=tmp0H
+	pmaddwd   xmm3,[rel PW_MF089_F060]	; xmm3=tmp3L
+	pmaddwd   xmm4,[rel PW_MF089_F060]	; xmm4=tmp3H
+
+	paddd	xmm2, XMMWORD [wk(10)]	; xmm2=tmp0L
+	paddd	xmm0, XMMWORD [wk(11)]	; xmm0=tmp0H
+	paddd	xmm3,xmm5		; xmm3=tmp3L
+	paddd	xmm4,xmm7		; xmm4=tmp3H
+
+	movdqa	XMMWORD [wk(8)], xmm2	; wk(8)=tmp0L
+	movdqa	XMMWORD [wk(9)], xmm0	; wk(9)=tmp0H
+
+	movdqa    xmm2,xmm1
+	movdqa    xmm0,xmm1
+	punpcklwd xmm2,xmm6
+	punpckhwd xmm0,xmm6
+	movdqa    xmm1,xmm2
+	movdqa    xmm6,xmm0
+	pmaddwd   xmm2,[rel PW_MF050_MF256]	; xmm2=tmp1L
+	pmaddwd   xmm0,[rel PW_MF050_MF256]	; xmm0=tmp1H
+	pmaddwd   xmm1,[rel PW_MF256_F050]	; xmm1=tmp2L
+	pmaddwd   xmm6,[rel PW_MF256_F050]	; xmm6=tmp2H
+
+	paddd	xmm2,xmm5		; xmm2=tmp1L
+	paddd	xmm0,xmm7		; xmm0=tmp1H
+	paddd	xmm1, XMMWORD [wk(10)]	; xmm1=tmp2L
+	paddd	xmm6, XMMWORD [wk(11)]	; xmm6=tmp2H
+
+	movdqa	XMMWORD [wk(10)], xmm2	; wk(10)=tmp1L
+	movdqa	XMMWORD [wk(11)], xmm0	; wk(11)=tmp1H
+
+	; -- Final output stage
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=tmp10L
+	movdqa	xmm7, XMMWORD [wk(1)]	; xmm7=tmp10H
+
+	movdqa	xmm2,xmm5
+	movdqa	xmm0,xmm7
+	paddd	xmm5,xmm3		; xmm5=data0L
+	paddd	xmm7,xmm4		; xmm7=data0H
+	psubd	xmm2,xmm3		; xmm2=data7L
+	psubd	xmm0,xmm4		; xmm0=data7H
+
+	movdqa	xmm3,[rel PD_DESCALE_P1]	; xmm3=[rel PD_DESCALE_P1]
+
+	paddd	xmm5,xmm3
+	paddd	xmm7,xmm3
+	psrad	xmm5,DESCALE_P1
+	psrad	xmm7,DESCALE_P1
+	paddd	xmm2,xmm3
+	paddd	xmm0,xmm3
+	psrad	xmm2,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+
+	packssdw  xmm5,xmm7		; xmm5=data0=(00 01 02 03 04 05 06 07)
+	packssdw  xmm2,xmm0		; xmm2=data7=(70 71 72 73 74 75 76 77)
+
+	movdqa	xmm4, XMMWORD [wk(4)]	; xmm4=tmp11L
+	movdqa	xmm3, XMMWORD [wk(5)]	; xmm3=tmp11H
+
+	movdqa	xmm7,xmm4
+	movdqa	xmm0,xmm3
+	paddd	xmm4,xmm1		; xmm4=data1L
+	paddd	xmm3,xmm6		; xmm3=data1H
+	psubd	xmm7,xmm1		; xmm7=data6L
+	psubd	xmm0,xmm6		; xmm0=data6H
+
+	movdqa	xmm1,[rel PD_DESCALE_P1]	; xmm1=[rel PD_DESCALE_P1]
+
+	paddd	xmm4,xmm1
+	paddd	xmm3,xmm1
+	psrad	xmm4,DESCALE_P1
+	psrad	xmm3,DESCALE_P1
+	paddd	xmm7,xmm1
+	paddd	xmm0,xmm1
+	psrad	xmm7,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+
+	packssdw  xmm4,xmm3		; xmm4=data1=(10 11 12 13 14 15 16 17)
+	packssdw  xmm7,xmm0		; xmm7=data6=(60 61 62 63 64 65 66 67)
+
+	movdqa    xmm6,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm4		; xmm5=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm6,xmm4		; xmm6=(04 14 05 15 06 16 07 17)
+	movdqa    xmm1,xmm7		; transpose coefficients(phase 1)
+	punpcklwd xmm7,xmm2		; xmm7=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm1,xmm2		; xmm1=(64 74 65 75 66 76 67 77)
+
+	movdqa	xmm3, XMMWORD [wk(6)]	; xmm3=tmp12L
+	movdqa	xmm0, XMMWORD [wk(7)]	; xmm0=tmp12H
+	movdqa	xmm4, XMMWORD [wk(10)]	; xmm4=tmp1L
+	movdqa	xmm2, XMMWORD [wk(11)]	; xmm2=tmp1H
+
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(00 10 01 11 02 12 03 13)
+	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=(04 14 05 15 06 16 07 17)
+	movdqa	XMMWORD [wk(4)], xmm7	; wk(4)=(60 70 61 71 62 72 63 73)
+	movdqa	XMMWORD [wk(5)], xmm1	; wk(5)=(64 74 65 75 66 76 67 77)
+
+	movdqa	xmm5,xmm3
+	movdqa	xmm6,xmm0
+	paddd	xmm3,xmm4		; xmm3=data2L
+	paddd	xmm0,xmm2		; xmm0=data2H
+	psubd	xmm5,xmm4		; xmm5=data5L
+	psubd	xmm6,xmm2		; xmm6=data5H
+
+	movdqa	xmm7,[rel PD_DESCALE_P1]	; xmm7=[rel PD_DESCALE_P1]
+
+	paddd	xmm3,xmm7
+	paddd	xmm0,xmm7
+	psrad	xmm3,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+	paddd	xmm5,xmm7
+	paddd	xmm6,xmm7
+	psrad	xmm5,DESCALE_P1
+	psrad	xmm6,DESCALE_P1
+
+	packssdw  xmm3,xmm0		; xmm3=data2=(20 21 22 23 24 25 26 27)
+	packssdw  xmm5,xmm6		; xmm5=data5=(50 51 52 53 54 55 56 57)
+
+	movdqa	xmm1, XMMWORD [wk(2)]	; xmm1=tmp13L
+	movdqa	xmm4, XMMWORD [wk(3)]	; xmm4=tmp13H
+	movdqa	xmm2, XMMWORD [wk(8)]	; xmm2=tmp0L
+	movdqa	xmm7, XMMWORD [wk(9)]	; xmm7=tmp0H
+
+	movdqa	xmm0,xmm1
+	movdqa	xmm6,xmm4
+	paddd	xmm1,xmm2		; xmm1=data3L
+	paddd	xmm4,xmm7		; xmm4=data3H
+	psubd	xmm0,xmm2		; xmm0=data4L
+	psubd	xmm6,xmm7		; xmm6=data4H
+
+	movdqa	xmm2,[rel PD_DESCALE_P1]	; xmm2=[rel PD_DESCALE_P1]
+
+	paddd	xmm1,xmm2
+	paddd	xmm4,xmm2
+	psrad	xmm1,DESCALE_P1
+	psrad	xmm4,DESCALE_P1
+	paddd	xmm0,xmm2
+	paddd	xmm6,xmm2
+	psrad	xmm0,DESCALE_P1
+	psrad	xmm6,DESCALE_P1
+
+	packssdw  xmm1,xmm4		; xmm1=data3=(30 31 32 33 34 35 36 37)
+	packssdw  xmm0,xmm6		; xmm0=data4=(40 41 42 43 44 45 46 47)
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=(00 10 01 11 02 12 03 13)
+	movdqa	xmm2, XMMWORD [wk(1)]	; xmm2=(04 14 05 15 06 16 07 17)
+
+	movdqa    xmm4,xmm3		; transpose coefficients(phase 1)
+	punpcklwd xmm3,xmm1		; xmm3=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm4,xmm1		; xmm4=(24 34 25 35 26 36 27 37)
+	movdqa    xmm6,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm5		; xmm0=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm6,xmm5		; xmm6=(44 54 45 55 46 56 47 57)
+
+	movdqa    xmm1,xmm7		; transpose coefficients(phase 2)
+	punpckldq xmm7,xmm3		; xmm7=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm1,xmm3		; xmm1=(02 12 22 32 03 13 23 33)
+	movdqa    xmm5,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm4		; xmm2=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm5,xmm4		; xmm5=(06 16 26 36 07 17 27 37)
+
+	movdqa	xmm3, XMMWORD [wk(4)]	; xmm3=(60 70 61 71 62 72 63 73)
+	movdqa	xmm4, XMMWORD [wk(5)]	; xmm4=(64 74 65 75 66 76 67 77)
+
+	movdqa	XMMWORD [wk(6)], xmm2	; wk(6)=(04 14 24 34 05 15 25 35)
+	movdqa	XMMWORD [wk(7)], xmm5	; wk(7)=(06 16 26 36 07 17 27 37)
+
+	movdqa    xmm2,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm3		; xmm0=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm2,xmm3		; xmm2=(42 52 62 72 43 53 63 73)
+	movdqa    xmm5,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm4		; xmm6=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm5,xmm4		; xmm5=(46 56 66 76 47 57 67 77)
+
+	movdqa     xmm3,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm0		; xmm7=col0=(00 10 20 30 40 50 60 70)
+	punpckhqdq xmm3,xmm0		; xmm3=col1=(01 11 21 31 41 51 61 71)
+	movdqa     xmm4,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm2		; xmm1=col2=(02 12 22 32 42 52 62 72)
+	punpckhqdq xmm4,xmm2		; xmm4=col3=(03 13 23 33 43 53 63 73)
+
+	movdqa	xmm0, XMMWORD [wk(6)]	; xmm0=(04 14 24 34 05 15 25 35)
+	movdqa	xmm2, XMMWORD [wk(7)]	; xmm2=(06 16 26 36 07 17 27 37)
+
+	movdqa	XMMWORD [wk(8)], xmm3	; wk(8)=col1
+	movdqa	XMMWORD [wk(9)], xmm4	; wk(9)=col3
+
+	movdqa     xmm3,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm6		; xmm0=col4=(04 14 24 34 44 54 64 74)
+	punpckhqdq xmm3,xmm6		; xmm3=col5=(05 15 25 35 45 55 65 75)
+	movdqa     xmm4,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm5		; xmm2=col6=(06 16 26 36 46 56 66 76)
+	punpckhqdq xmm4,xmm5		; xmm4=col7=(07 17 27 37 47 57 67 77)
+
+	movdqa	XMMWORD [wk(10)], xmm3	; wk(10)=col5
+	movdqa	XMMWORD [wk(11)], xmm4	; wk(11)=col7
+.column_end:
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	rax, [original_rbp]
+	mov	rdi, r12	; (JSAMPROW *)
+	mov	rax, r13
+
+	; -- Even part
+
+	; xmm7=col0, xmm1=col2, xmm0=col4, xmm2=col6
+
+	; (Original)
+	; z1 = (z2 + z3) * 0.541196100;
+	; tmp2 = z1 + z3 * -1.847759065;
+	; tmp3 = z1 + z2 * 0.765366865;
+	;
+	; (This implementation)
+	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+	movdqa    xmm6,xmm1		; xmm1=in2=z2
+	movdqa    xmm5,xmm1
+	punpcklwd xmm6,xmm2		; xmm2=in6=z3
+	punpckhwd xmm5,xmm2
+	movdqa    xmm1,xmm6
+	movdqa    xmm2,xmm5
+	pmaddwd   xmm6,[rel PW_F130_F054]	; xmm6=tmp3L
+	pmaddwd   xmm5,[rel PW_F130_F054]	; xmm5=tmp3H
+	pmaddwd   xmm1,[rel PW_F054_MF130]	; xmm1=tmp2L
+	pmaddwd   xmm2,[rel PW_F054_MF130]	; xmm2=tmp2H
+
+	movdqa    xmm3,xmm7
+	paddw     xmm7,xmm0		; xmm7=in0+in4
+	psubw     xmm3,xmm0		; xmm3=in0-in4
+
+	pxor      xmm4,xmm4
+	pxor      xmm0,xmm0
+	punpcklwd xmm4,xmm7		; xmm4=tmp0L
+	punpckhwd xmm0,xmm7		; xmm0=tmp0H
+	psrad     xmm4,(16-CONST_BITS)	; psrad xmm4,16 & pslld xmm4,CONST_BITS
+	psrad     xmm0,(16-CONST_BITS)	; psrad xmm0,16 & pslld xmm0,CONST_BITS
+
+	movdqa	xmm7,xmm4
+	paddd	xmm4,xmm6		; xmm4=tmp10L
+	psubd	xmm7,xmm6		; xmm7=tmp13L
+	movdqa	xmm6,xmm0
+	paddd	xmm0,xmm5		; xmm0=tmp10H
+	psubd	xmm6,xmm5		; xmm6=tmp13H
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=tmp10L
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp10H
+	movdqa	XMMWORD [wk(2)], xmm7	; wk(2)=tmp13L
+	movdqa	XMMWORD [wk(3)], xmm6	; wk(3)=tmp13H
+
+	pxor      xmm5,xmm5
+	pxor      xmm4,xmm4
+	punpcklwd xmm5,xmm3		; xmm5=tmp1L
+	punpckhwd xmm4,xmm3		; xmm4=tmp1H
+	psrad     xmm5,(16-CONST_BITS)	; psrad xmm5,16 & pslld xmm5,CONST_BITS
+	psrad     xmm4,(16-CONST_BITS)	; psrad xmm4,16 & pslld xmm4,CONST_BITS
+
+	movdqa	xmm0,xmm5
+	paddd	xmm5,xmm1		; xmm5=tmp11L
+	psubd	xmm0,xmm1		; xmm0=tmp12L
+	movdqa	xmm7,xmm4
+	paddd	xmm4,xmm2		; xmm4=tmp11H
+	psubd	xmm7,xmm2		; xmm7=tmp12H
+
+	movdqa	XMMWORD [wk(4)], xmm5	; wk(4)=tmp11L
+	movdqa	XMMWORD [wk(5)], xmm4	; wk(5)=tmp11H
+	movdqa	XMMWORD [wk(6)], xmm0	; wk(6)=tmp12L
+	movdqa	XMMWORD [wk(7)], xmm7	; wk(7)=tmp12H
+
+	; -- Odd part
+
+	movdqa	xmm6, XMMWORD [wk(9)]	; xmm6=col3
+	movdqa	xmm3, XMMWORD [wk(8)]	; xmm3=col1
+	movdqa	xmm1, XMMWORD [wk(11)]	; xmm1=col7
+	movdqa	xmm2, XMMWORD [wk(10)]	; xmm2=col5
+
+	movdqa	xmm5,xmm6
+	movdqa	xmm4,xmm3
+	paddw	xmm5,xmm1		; xmm5=z3
+	paddw	xmm4,xmm2		; xmm4=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm0,xmm5
+	movdqa    xmm7,xmm5
+	punpcklwd xmm0,xmm4
+	punpckhwd xmm7,xmm4
+	movdqa    xmm5,xmm0
+	movdqa    xmm4,xmm7
+	pmaddwd   xmm0,[rel PW_MF078_F117]	; xmm0=z3L
+	pmaddwd   xmm7,[rel PW_MF078_F117]	; xmm7=z3H
+	pmaddwd   xmm5,[rel PW_F117_F078]	; xmm5=z4L
+	pmaddwd   xmm4,[rel PW_F117_F078]	; xmm4=z4H
+
+	movdqa	XMMWORD [wk(10)], xmm0	; wk(10)=z3L
+	movdqa	XMMWORD [wk(11)], xmm7	; wk(11)=z3H
+
+	; (Original)
+	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+	;
+	; (This implementation)
+	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+	; tmp0 += z3;  tmp1 += z4;
+	; tmp2 += z3;  tmp3 += z4;
+
+	movdqa    xmm0,xmm1
+	movdqa    xmm7,xmm1
+	punpcklwd xmm0,xmm3
+	punpckhwd xmm7,xmm3
+	movdqa    xmm1,xmm0
+	movdqa    xmm3,xmm7
+	pmaddwd   xmm0,[rel PW_MF060_MF089]	; xmm0=tmp0L
+	pmaddwd   xmm7,[rel PW_MF060_MF089]	; xmm7=tmp0H
+	pmaddwd   xmm1,[rel PW_MF089_F060]	; xmm1=tmp3L
+	pmaddwd   xmm3,[rel PW_MF089_F060]	; xmm3=tmp3H
+
+	paddd	xmm0, XMMWORD [wk(10)]	; xmm0=tmp0L
+	paddd	xmm7, XMMWORD [wk(11)]	; xmm7=tmp0H
+	paddd	xmm1,xmm5		; xmm1=tmp3L
+	paddd	xmm3,xmm4		; xmm3=tmp3H
+
+	movdqa	XMMWORD [wk(8)], xmm0	; wk(8)=tmp0L
+	movdqa	XMMWORD [wk(9)], xmm7	; wk(9)=tmp0H
+
+	movdqa    xmm0,xmm2
+	movdqa    xmm7,xmm2
+	punpcklwd xmm0,xmm6
+	punpckhwd xmm7,xmm6
+	movdqa    xmm2,xmm0
+	movdqa    xmm6,xmm7
+	pmaddwd   xmm0,[rel PW_MF050_MF256]	; xmm0=tmp1L
+	pmaddwd   xmm7,[rel PW_MF050_MF256]	; xmm7=tmp1H
+	pmaddwd   xmm2,[rel PW_MF256_F050]	; xmm2=tmp2L
+	pmaddwd   xmm6,[rel PW_MF256_F050]	; xmm6=tmp2H
+
+	paddd	xmm0,xmm5		; xmm0=tmp1L
+	paddd	xmm7,xmm4		; xmm7=tmp1H
+	paddd	xmm2, XMMWORD [wk(10)]	; xmm2=tmp2L
+	paddd	xmm6, XMMWORD [wk(11)]	; xmm6=tmp2H
+
+	movdqa	XMMWORD [wk(10)], xmm0	; wk(10)=tmp1L
+	movdqa	XMMWORD [wk(11)], xmm7	; wk(11)=tmp1H
+
+	; -- Final output stage
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=tmp10L
+	movdqa	xmm4, XMMWORD [wk(1)]	; xmm4=tmp10H
+
+	movdqa	xmm0,xmm5
+	movdqa	xmm7,xmm4
+	paddd	xmm5,xmm1		; xmm5=data0L
+	paddd	xmm4,xmm3		; xmm4=data0H
+	psubd	xmm0,xmm1		; xmm0=data7L
+	psubd	xmm7,xmm3		; xmm7=data7H
+
+	movdqa	xmm1,[rel PD_DESCALE_P2]	; xmm1=[rel PD_DESCALE_P2]
+
+	paddd	xmm5,xmm1
+	paddd	xmm4,xmm1
+	psrad	xmm5,DESCALE_P2
+	psrad	xmm4,DESCALE_P2
+	paddd	xmm0,xmm1
+	paddd	xmm7,xmm1
+	psrad	xmm0,DESCALE_P2
+	psrad	xmm7,DESCALE_P2
+
+	packssdw  xmm5,xmm4		; xmm5=data0=(00 10 20 30 40 50 60 70)
+	packssdw  xmm0,xmm7		; xmm0=data7=(07 17 27 37 47 57 67 77)
+
+	movdqa	xmm3, XMMWORD [wk(4)]	; xmm3=tmp11L
+	movdqa	xmm1, XMMWORD [wk(5)]	; xmm1=tmp11H
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm7,xmm1
+	paddd	xmm3,xmm2		; xmm3=data1L
+	paddd	xmm1,xmm6		; xmm1=data1H
+	psubd	xmm4,xmm2		; xmm4=data6L
+	psubd	xmm7,xmm6		; xmm7=data6H
+
+	movdqa	xmm2,[rel PD_DESCALE_P2]	; xmm2=[rel PD_DESCALE_P2]
+
+	paddd	xmm3,xmm2
+	paddd	xmm1,xmm2
+	psrad	xmm3,DESCALE_P2
+	psrad	xmm1,DESCALE_P2
+	paddd	xmm4,xmm2
+	paddd	xmm7,xmm2
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm7,DESCALE_P2
+
+	packssdw  xmm3,xmm1		; xmm3=data1=(01 11 21 31 41 51 61 71)
+	packssdw  xmm4,xmm7		; xmm4=data6=(06 16 26 36 46 56 66 76)
+
+	packsswb  xmm5,xmm4		; xmm5=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	packsswb  xmm3,xmm0		; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	movdqa	xmm6, XMMWORD [wk(6)]	; xmm6=tmp12L
+	movdqa	xmm2, XMMWORD [wk(7)]	; xmm2=tmp12H
+	movdqa	xmm1, XMMWORD [wk(10)]	; xmm1=tmp1L
+	movdqa	xmm7, XMMWORD [wk(11)]	; xmm7=tmp1H
+
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	movdqa	xmm4,xmm6
+	movdqa	xmm0,xmm2
+	paddd	xmm6,xmm1		; xmm6=data2L
+	paddd	xmm2,xmm7		; xmm2=data2H
+	psubd	xmm4,xmm1		; xmm4=data5L
+	psubd	xmm0,xmm7		; xmm0=data5H
+
+	movdqa	xmm5,[rel PD_DESCALE_P2]	; xmm5=[rel PD_DESCALE_P2]
+
+	paddd	xmm6,xmm5
+	paddd	xmm2,xmm5
+	psrad	xmm6,DESCALE_P2
+	psrad	xmm2,DESCALE_P2
+	paddd	xmm4,xmm5
+	paddd	xmm0,xmm5
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm0,DESCALE_P2
+
+	packssdw  xmm6,xmm2		; xmm6=data2=(02 12 22 32 42 52 62 72)
+	packssdw  xmm4,xmm0		; xmm4=data5=(05 15 25 35 45 55 65 75)
+
+	movdqa	xmm3, XMMWORD [wk(2)]	; xmm3=tmp13L
+	movdqa	xmm1, XMMWORD [wk(3)]	; xmm1=tmp13H
+	movdqa	xmm7, XMMWORD [wk(8)]	; xmm7=tmp0L
+	movdqa	xmm5, XMMWORD [wk(9)]	; xmm5=tmp0H
+
+	movdqa	xmm2,xmm3
+	movdqa	xmm0,xmm1
+	paddd	xmm3,xmm7		; xmm3=data3L
+	paddd	xmm1,xmm5		; xmm1=data3H
+	psubd	xmm2,xmm7		; xmm2=data4L
+	psubd	xmm0,xmm5		; xmm0=data4H
+
+	movdqa	xmm7,[rel PD_DESCALE_P2]	; xmm7=[rel PD_DESCALE_P2]
+
+	paddd	xmm3,xmm7
+	paddd	xmm1,xmm7
+	psrad	xmm3,DESCALE_P2
+	psrad	xmm1,DESCALE_P2
+	paddd	xmm2,xmm7
+	paddd	xmm0,xmm7
+	psrad	xmm2,DESCALE_P2
+	psrad	xmm0,DESCALE_P2
+
+	movdqa    xmm5,[rel PB_CENTERJSAMP]	; xmm5=[rel PB_CENTERJSAMP]
+
+	packssdw  xmm3,xmm1		; xmm3=data3=(03 13 23 33 43 53 63 73)
+	packssdw  xmm2,xmm0		; xmm2=data4=(04 14 24 34 44 54 64 74)
+
+	movdqa    xmm7, XMMWORD [wk(0)]	; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	movdqa    xmm1, XMMWORD [wk(1)]	; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	packsswb  xmm6,xmm2		; xmm6=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
+	packsswb  xmm3,xmm4		; xmm3=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
+
+	paddb     xmm7,xmm5
+	paddb     xmm1,xmm5
+	paddb     xmm6,xmm5
+	paddb     xmm3,xmm5
+
+	movdqa    xmm0,xmm7	; transpose coefficients(phase 1)
+	punpcklbw xmm7,xmm1	; xmm7=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
+	punpckhbw xmm0,xmm1	; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
+	movdqa    xmm2,xmm6	; transpose coefficients(phase 1)
+	punpcklbw xmm6,xmm3	; xmm6=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
+	punpckhbw xmm2,xmm3	; xmm2=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
+
+	movdqa    xmm4,xmm7	; transpose coefficients(phase 2)
+	punpcklwd xmm7,xmm6	; xmm7=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+	punpckhwd xmm4,xmm6	; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
+	movdqa    xmm5,xmm2	; transpose coefficients(phase 2)
+	punpcklwd xmm2,xmm0	; xmm2=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+	punpckhwd xmm5,xmm0	; xmm5=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
+
+	movdqa    xmm1,xmm7	; transpose coefficients(phase 3)
+	punpckldq xmm7,xmm2	; xmm7=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+	punpckhdq xmm1,xmm2	; xmm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+	movdqa    xmm3,xmm4	; transpose coefficients(phase 3)
+	punpckldq xmm4,xmm5	; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+	punpckhdq xmm3,xmm5	; xmm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+	pshufd	xmm6,xmm7,0x4E	; xmm6=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+	pshufd	xmm0,xmm1,0x4E	; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+	pshufd	xmm2,xmm4,0x4E	; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+	pshufd	xmm5,xmm3,0x4E	; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+	mov	rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm7
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
+	mov	rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+
+	mov	rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
+	mov	rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
+	movq	XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
+
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2int.asm b/simd/jiss2int.asm
new file mode 100644
index 0000000..adf39fb
--- /dev/null
+++ b/simd/jiss2int.asm
@@ -0,0 +1,859 @@
+;
+; jiss2int.asm - accurate integer IDCT (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a slow-but-accurate integer implementation of the
+; inverse DCT (Discrete Cosine Transform). The following code is based
+; directly on the IJG's original jidctint.c; see the jidctint.c for
+; more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
+%define DESCALE_P2	(CONST_BITS+PASS1_BITS+3)
+
+%if CONST_BITS == 13
+F_0_298	equ	 2446		; FIX(0.298631336)
+F_0_390	equ	 3196		; FIX(0.390180644)
+F_0_541	equ	 4433		; FIX(0.541196100)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_175	equ	 9633		; FIX(1.175875602)
+F_1_501	equ	12299		; FIX(1.501321110)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_1_961	equ	16069		; FIX(1.961570560)
+F_2_053	equ	16819		; FIX(2.053119869)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_072	equ	25172		; FIX(3.072711026)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
+F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
+F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
+F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
+F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_islow_sse2)
+
+EXTN(jconst_idct_islow_sse2):
+
+PW_F130_F054	times 4 dw  (F_0_541+F_0_765), F_0_541
+PW_F054_MF130	times 4 dw  F_0_541, (F_0_541-F_1_847)
+PW_MF078_F117	times 4 dw  (F_1_175-F_1_961), F_1_175
+PW_F117_F078	times 4 dw  F_1_175, (F_1_175-F_0_390)
+PW_MF060_MF089	times 4 dw  (F_0_298-F_0_899),-F_0_899
+PW_MF089_F060	times 4 dw -F_0_899, (F_1_501-F_0_899)
+PW_MF050_MF256	times 4 dw  (F_2_053-F_2_562),-F_2_562
+PW_MF256_F050	times 4 dw -F_2_562, (F_3_072-F_2_562)
+PD_DESCALE_P1	times 4 dd  1 << (DESCALE_P1-1)
+PD_DESCALE_P2	times 4 dd  1 << (DESCALE_P2-1)
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_islow_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                        JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		12
+
+	align	16
+	global	EXTN(jsimd_idct_islow_sse2)
+
+EXTN(jsimd_idct_islow_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_ISLOW_SSE2
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	xmm1,xmm0
+	packsswb xmm1,xmm1
+	packsswb xmm1,xmm1
+	movd	eax,xmm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movdqa	xmm5, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm5, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	psllw	xmm5,PASS1_BITS
+
+	movdqa    xmm4,xmm5		; xmm5=in0=(00 01 02 03 04 05 06 07)
+	punpcklwd xmm5,xmm5		; xmm5=(00 00 01 01 02 02 03 03)
+	punpckhwd xmm4,xmm4		; xmm4=(04 04 05 05 06 06 07 07)
+
+	pshufd	xmm7,xmm5,0x00		; xmm7=col0=(00 00 00 00 00 00 00 00)
+	pshufd	xmm6,xmm5,0x55		; xmm6=col1=(01 01 01 01 01 01 01 01)
+	pshufd	xmm1,xmm5,0xAA		; xmm1=col2=(02 02 02 02 02 02 02 02)
+	pshufd	xmm5,xmm5,0xFF		; xmm5=col3=(03 03 03 03 03 03 03 03)
+	pshufd	xmm0,xmm4,0x00		; xmm0=col4=(04 04 04 04 04 04 04 04)
+	pshufd	xmm3,xmm4,0x55		; xmm3=col5=(05 05 05 05 05 05 05 05)
+	pshufd	xmm2,xmm4,0xAA		; xmm2=col6=(06 06 06 06 06 06 06 06)
+	pshufd	xmm4,xmm4,0xFF		; xmm4=col7=(07 07 07 07 07 07 07 07)
+
+	movdqa	XMMWORD [wk(8)], xmm6	; wk(8)=col1
+	movdqa	XMMWORD [wk(9)], xmm5	; wk(9)=col3
+	movdqa	XMMWORD [wk(10)], xmm3	; wk(10)=col5
+	movdqa	XMMWORD [wk(11)], xmm4	; wk(11)=col7
+	jmp	near .column_end
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; (Original)
+	; z1 = (z2 + z3) * 0.541196100;
+	; tmp2 = z1 + z3 * -1.847759065;
+	; tmp3 = z1 + z2 * 0.765366865;
+	;
+	; (This implementation)
+	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+	movdqa    xmm4,xmm1		; xmm1=in2=z2
+	movdqa    xmm5,xmm1
+	punpcklwd xmm4,xmm3		; xmm3=in6=z3
+	punpckhwd xmm5,xmm3
+	movdqa    xmm1,xmm4
+	movdqa    xmm3,xmm5
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_F130_F054)]	; xmm4=tmp3L
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F130_F054)]	; xmm5=tmp3H
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_F054_MF130)]	; xmm1=tmp2L
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_F054_MF130)]	; xmm3=tmp2H
+
+	movdqa    xmm6,xmm0
+	paddw     xmm0,xmm2		; xmm0=in0+in4
+	psubw     xmm6,xmm2		; xmm6=in0-in4
+
+	pxor      xmm7,xmm7
+	pxor      xmm2,xmm2
+	punpcklwd xmm7,xmm0		; xmm7=tmp0L
+	punpckhwd xmm2,xmm0		; xmm2=tmp0H
+	psrad     xmm7,(16-CONST_BITS)	; psrad xmm7,16 & pslld xmm7,CONST_BITS
+	psrad     xmm2,(16-CONST_BITS)	; psrad xmm2,16 & pslld xmm2,CONST_BITS
+
+	movdqa	xmm0,xmm7
+	paddd	xmm7,xmm4		; xmm7=tmp10L
+	psubd	xmm0,xmm4		; xmm0=tmp13L
+	movdqa	xmm4,xmm2
+	paddd	xmm2,xmm5		; xmm2=tmp10H
+	psubd	xmm4,xmm5		; xmm4=tmp13H
+
+	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=tmp10L
+	movdqa	XMMWORD [wk(1)], xmm2	; wk(1)=tmp10H
+	movdqa	XMMWORD [wk(2)], xmm0	; wk(2)=tmp13L
+	movdqa	XMMWORD [wk(3)], xmm4	; wk(3)=tmp13H
+
+	pxor      xmm5,xmm5
+	pxor      xmm7,xmm7
+	punpcklwd xmm5,xmm6		; xmm5=tmp1L
+	punpckhwd xmm7,xmm6		; xmm7=tmp1H
+	psrad     xmm5,(16-CONST_BITS)	; psrad xmm5,16 & pslld xmm5,CONST_BITS
+	psrad     xmm7,(16-CONST_BITS)	; psrad xmm7,16 & pslld xmm7,CONST_BITS
+
+	movdqa	xmm2,xmm5
+	paddd	xmm5,xmm1		; xmm5=tmp11L
+	psubd	xmm2,xmm1		; xmm2=tmp12L
+	movdqa	xmm0,xmm7
+	paddd	xmm7,xmm3		; xmm7=tmp11H
+	psubd	xmm0,xmm3		; xmm0=tmp12H
+
+	movdqa	XMMWORD [wk(4)], xmm5	; wk(4)=tmp11L
+	movdqa	XMMWORD [wk(5)], xmm7	; wk(5)=tmp11H
+	movdqa	XMMWORD [wk(6)], xmm2	; wk(6)=tmp12L
+	movdqa	XMMWORD [wk(7)], xmm0	; wk(7)=tmp12H
+
+	; -- Odd part
+
+	movdqa	xmm4, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm6, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm4, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm6, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movdqa	xmm5,xmm6
+	movdqa	xmm7,xmm4
+	paddw	xmm5,xmm3		; xmm5=z3
+	paddw	xmm7,xmm1		; xmm7=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm2,xmm5
+	movdqa    xmm0,xmm5
+	punpcklwd xmm2,xmm7
+	punpckhwd xmm0,xmm7
+	movdqa    xmm5,xmm2
+	movdqa    xmm7,xmm0
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_MF078_F117)]	; xmm2=z3L
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF078_F117)]	; xmm0=z3H
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F117_F078)]	; xmm5=z4L
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_F117_F078)]	; xmm7=z4H
+
+	movdqa	XMMWORD [wk(10)], xmm2	; wk(10)=z3L
+	movdqa	XMMWORD [wk(11)], xmm0	; wk(11)=z3H
+
+	; (Original)
+	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+	;
+	; (This implementation)
+	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+	; tmp0 += z3;  tmp1 += z4;
+	; tmp2 += z3;  tmp3 += z4;
+
+	movdqa    xmm2,xmm3
+	movdqa    xmm0,xmm3
+	punpcklwd xmm2,xmm4
+	punpckhwd xmm0,xmm4
+	movdqa    xmm3,xmm2
+	movdqa    xmm4,xmm0
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm2=tmp0L
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm0=tmp0H
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_MF089_F060)]	; xmm3=tmp3L
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_MF089_F060)]	; xmm4=tmp3H
+
+	paddd	xmm2, XMMWORD [wk(10)]	; xmm2=tmp0L
+	paddd	xmm0, XMMWORD [wk(11)]	; xmm0=tmp0H
+	paddd	xmm3,xmm5		; xmm3=tmp3L
+	paddd	xmm4,xmm7		; xmm4=tmp3H
+
+	movdqa	XMMWORD [wk(8)], xmm2	; wk(8)=tmp0L
+	movdqa	XMMWORD [wk(9)], xmm0	; wk(9)=tmp0H
+
+	movdqa    xmm2,xmm1
+	movdqa    xmm0,xmm1
+	punpcklwd xmm2,xmm6
+	punpckhwd xmm0,xmm6
+	movdqa    xmm1,xmm2
+	movdqa    xmm6,xmm0
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm2=tmp1L
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm0=tmp1H
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF256_F050)]	; xmm1=tmp2L
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_MF256_F050)]	; xmm6=tmp2H
+
+	paddd	xmm2,xmm5		; xmm2=tmp1L
+	paddd	xmm0,xmm7		; xmm0=tmp1H
+	paddd	xmm1, XMMWORD [wk(10)]	; xmm1=tmp2L
+	paddd	xmm6, XMMWORD [wk(11)]	; xmm6=tmp2H
+
+	movdqa	XMMWORD [wk(10)], xmm2	; wk(10)=tmp1L
+	movdqa	XMMWORD [wk(11)], xmm0	; wk(11)=tmp1H
+
+	; -- Final output stage
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=tmp10L
+	movdqa	xmm7, XMMWORD [wk(1)]	; xmm7=tmp10H
+
+	movdqa	xmm2,xmm5
+	movdqa	xmm0,xmm7
+	paddd	xmm5,xmm3		; xmm5=data0L
+	paddd	xmm7,xmm4		; xmm7=data0H
+	psubd	xmm2,xmm3		; xmm2=data7L
+	psubd	xmm0,xmm4		; xmm0=data7H
+
+	movdqa	xmm3,[GOTOFF(ebx,PD_DESCALE_P1)]	; xmm3=[PD_DESCALE_P1]
+
+	paddd	xmm5,xmm3
+	paddd	xmm7,xmm3
+	psrad	xmm5,DESCALE_P1
+	psrad	xmm7,DESCALE_P1
+	paddd	xmm2,xmm3
+	paddd	xmm0,xmm3
+	psrad	xmm2,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+
+	packssdw  xmm5,xmm7		; xmm5=data0=(00 01 02 03 04 05 06 07)
+	packssdw  xmm2,xmm0		; xmm2=data7=(70 71 72 73 74 75 76 77)
+
+	movdqa	xmm4, XMMWORD [wk(4)]	; xmm4=tmp11L
+	movdqa	xmm3, XMMWORD [wk(5)]	; xmm3=tmp11H
+
+	movdqa	xmm7,xmm4
+	movdqa	xmm0,xmm3
+	paddd	xmm4,xmm1		; xmm4=data1L
+	paddd	xmm3,xmm6		; xmm3=data1H
+	psubd	xmm7,xmm1		; xmm7=data6L
+	psubd	xmm0,xmm6		; xmm0=data6H
+
+	movdqa	xmm1,[GOTOFF(ebx,PD_DESCALE_P1)]	; xmm1=[PD_DESCALE_P1]
+
+	paddd	xmm4,xmm1
+	paddd	xmm3,xmm1
+	psrad	xmm4,DESCALE_P1
+	psrad	xmm3,DESCALE_P1
+	paddd	xmm7,xmm1
+	paddd	xmm0,xmm1
+	psrad	xmm7,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+
+	packssdw  xmm4,xmm3		; xmm4=data1=(10 11 12 13 14 15 16 17)
+	packssdw  xmm7,xmm0		; xmm7=data6=(60 61 62 63 64 65 66 67)
+
+	movdqa    xmm6,xmm5		; transpose coefficients(phase 1)
+	punpcklwd xmm5,xmm4		; xmm5=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm6,xmm4		; xmm6=(04 14 05 15 06 16 07 17)
+	movdqa    xmm1,xmm7		; transpose coefficients(phase 1)
+	punpcklwd xmm7,xmm2		; xmm7=(60 70 61 71 62 72 63 73)
+	punpckhwd xmm1,xmm2		; xmm1=(64 74 65 75 66 76 67 77)
+
+	movdqa	xmm3, XMMWORD [wk(6)]	; xmm3=tmp12L
+	movdqa	xmm0, XMMWORD [wk(7)]	; xmm0=tmp12H
+	movdqa	xmm4, XMMWORD [wk(10)]	; xmm4=tmp1L
+	movdqa	xmm2, XMMWORD [wk(11)]	; xmm2=tmp1H
+
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(00 10 01 11 02 12 03 13)
+	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=(04 14 05 15 06 16 07 17)
+	movdqa	XMMWORD [wk(4)], xmm7	; wk(4)=(60 70 61 71 62 72 63 73)
+	movdqa	XMMWORD [wk(5)], xmm1	; wk(5)=(64 74 65 75 66 76 67 77)
+
+	movdqa	xmm5,xmm3
+	movdqa	xmm6,xmm0
+	paddd	xmm3,xmm4		; xmm3=data2L
+	paddd	xmm0,xmm2		; xmm0=data2H
+	psubd	xmm5,xmm4		; xmm5=data5L
+	psubd	xmm6,xmm2		; xmm6=data5H
+
+	movdqa	xmm7,[GOTOFF(ebx,PD_DESCALE_P1)]	; xmm7=[PD_DESCALE_P1]
+
+	paddd	xmm3,xmm7
+	paddd	xmm0,xmm7
+	psrad	xmm3,DESCALE_P1
+	psrad	xmm0,DESCALE_P1
+	paddd	xmm5,xmm7
+	paddd	xmm6,xmm7
+	psrad	xmm5,DESCALE_P1
+	psrad	xmm6,DESCALE_P1
+
+	packssdw  xmm3,xmm0		; xmm3=data2=(20 21 22 23 24 25 26 27)
+	packssdw  xmm5,xmm6		; xmm5=data5=(50 51 52 53 54 55 56 57)
+
+	movdqa	xmm1, XMMWORD [wk(2)]	; xmm1=tmp13L
+	movdqa	xmm4, XMMWORD [wk(3)]	; xmm4=tmp13H
+	movdqa	xmm2, XMMWORD [wk(8)]	; xmm2=tmp0L
+	movdqa	xmm7, XMMWORD [wk(9)]	; xmm7=tmp0H
+
+	movdqa	xmm0,xmm1
+	movdqa	xmm6,xmm4
+	paddd	xmm1,xmm2		; xmm1=data3L
+	paddd	xmm4,xmm7		; xmm4=data3H
+	psubd	xmm0,xmm2		; xmm0=data4L
+	psubd	xmm6,xmm7		; xmm6=data4H
+
+	movdqa	xmm2,[GOTOFF(ebx,PD_DESCALE_P1)]	; xmm2=[PD_DESCALE_P1]
+
+	paddd	xmm1,xmm2
+	paddd	xmm4,xmm2
+	psrad	xmm1,DESCALE_P1
+	psrad	xmm4,DESCALE_P1
+	paddd	xmm0,xmm2
+	paddd	xmm6,xmm2
+	psrad	xmm0,DESCALE_P1
+	psrad	xmm6,DESCALE_P1
+
+	packssdw  xmm1,xmm4		; xmm1=data3=(30 31 32 33 34 35 36 37)
+	packssdw  xmm0,xmm6		; xmm0=data4=(40 41 42 43 44 45 46 47)
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=(00 10 01 11 02 12 03 13)
+	movdqa	xmm2, XMMWORD [wk(1)]	; xmm2=(04 14 05 15 06 16 07 17)
+
+	movdqa    xmm4,xmm3		; transpose coefficients(phase 1)
+	punpcklwd xmm3,xmm1		; xmm3=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm4,xmm1		; xmm4=(24 34 25 35 26 36 27 37)
+	movdqa    xmm6,xmm0		; transpose coefficients(phase 1)
+	punpcklwd xmm0,xmm5		; xmm0=(40 50 41 51 42 52 43 53)
+	punpckhwd xmm6,xmm5		; xmm6=(44 54 45 55 46 56 47 57)
+
+	movdqa    xmm1,xmm7		; transpose coefficients(phase 2)
+	punpckldq xmm7,xmm3		; xmm7=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm1,xmm3		; xmm1=(02 12 22 32 03 13 23 33)
+	movdqa    xmm5,xmm2		; transpose coefficients(phase 2)
+	punpckldq xmm2,xmm4		; xmm2=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm5,xmm4		; xmm5=(06 16 26 36 07 17 27 37)
+
+	movdqa	xmm3, XMMWORD [wk(4)]	; xmm3=(60 70 61 71 62 72 63 73)
+	movdqa	xmm4, XMMWORD [wk(5)]	; xmm4=(64 74 65 75 66 76 67 77)
+
+	movdqa	XMMWORD [wk(6)], xmm2	; wk(6)=(04 14 24 34 05 15 25 35)
+	movdqa	XMMWORD [wk(7)], xmm5	; wk(7)=(06 16 26 36 07 17 27 37)
+
+	movdqa    xmm2,xmm0		; transpose coefficients(phase 2)
+	punpckldq xmm0,xmm3		; xmm0=(40 50 60 70 41 51 61 71)
+	punpckhdq xmm2,xmm3		; xmm2=(42 52 62 72 43 53 63 73)
+	movdqa    xmm5,xmm6		; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm4		; xmm6=(44 54 64 74 45 55 65 75)
+	punpckhdq xmm5,xmm4		; xmm5=(46 56 66 76 47 57 67 77)
+
+	movdqa     xmm3,xmm7		; transpose coefficients(phase 3)
+	punpcklqdq xmm7,xmm0		; xmm7=col0=(00 10 20 30 40 50 60 70)
+	punpckhqdq xmm3,xmm0		; xmm3=col1=(01 11 21 31 41 51 61 71)
+	movdqa     xmm4,xmm1		; transpose coefficients(phase 3)
+	punpcklqdq xmm1,xmm2		; xmm1=col2=(02 12 22 32 42 52 62 72)
+	punpckhqdq xmm4,xmm2		; xmm4=col3=(03 13 23 33 43 53 63 73)
+
+	movdqa	xmm0, XMMWORD [wk(6)]	; xmm0=(04 14 24 34 05 15 25 35)
+	movdqa	xmm2, XMMWORD [wk(7)]	; xmm2=(06 16 26 36 07 17 27 37)
+
+	movdqa	XMMWORD [wk(8)], xmm3	; wk(8)=col1
+	movdqa	XMMWORD [wk(9)], xmm4	; wk(9)=col3
+
+	movdqa     xmm3,xmm0		; transpose coefficients(phase 3)
+	punpcklqdq xmm0,xmm6		; xmm0=col4=(04 14 24 34 44 54 64 74)
+	punpckhqdq xmm3,xmm6		; xmm3=col5=(05 15 25 35 45 55 65 75)
+	movdqa     xmm4,xmm2		; transpose coefficients(phase 3)
+	punpcklqdq xmm2,xmm5		; xmm2=col6=(06 16 26 36 46 56 66 76)
+	punpckhqdq xmm4,xmm5		; xmm4=col7=(07 17 27 37 47 57 67 77)
+
+	movdqa	XMMWORD [wk(10)], xmm3	; wk(10)=col5
+	movdqa	XMMWORD [wk(11)], xmm4	; wk(11)=col7
+.column_end:
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+
+	; -- Even part
+
+	; xmm7=col0, xmm1=col2, xmm0=col4, xmm2=col6
+
+	; (Original)
+	; z1 = (z2 + z3) * 0.541196100;
+	; tmp2 = z1 + z3 * -1.847759065;
+	; tmp3 = z1 + z2 * 0.765366865;
+	;
+	; (This implementation)
+	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
+	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
+
+	movdqa    xmm6,xmm1		; xmm1=in2=z2
+	movdqa    xmm5,xmm1
+	punpcklwd xmm6,xmm2		; xmm2=in6=z3
+	punpckhwd xmm5,xmm2
+	movdqa    xmm1,xmm6
+	movdqa    xmm2,xmm5
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_F130_F054)]	; xmm6=tmp3L
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F130_F054)]	; xmm5=tmp3H
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_F054_MF130)]	; xmm1=tmp2L
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_F054_MF130)]	; xmm2=tmp2H
+
+	movdqa    xmm3,xmm7
+	paddw     xmm7,xmm0		; xmm7=in0+in4
+	psubw     xmm3,xmm0		; xmm3=in0-in4
+
+	pxor      xmm4,xmm4
+	pxor      xmm0,xmm0
+	punpcklwd xmm4,xmm7		; xmm4=tmp0L
+	punpckhwd xmm0,xmm7		; xmm0=tmp0H
+	psrad     xmm4,(16-CONST_BITS)	; psrad xmm4,16 & pslld xmm4,CONST_BITS
+	psrad     xmm0,(16-CONST_BITS)	; psrad xmm0,16 & pslld xmm0,CONST_BITS
+
+	movdqa	xmm7,xmm4
+	paddd	xmm4,xmm6		; xmm4=tmp10L
+	psubd	xmm7,xmm6		; xmm7=tmp13L
+	movdqa	xmm6,xmm0
+	paddd	xmm0,xmm5		; xmm0=tmp10H
+	psubd	xmm6,xmm5		; xmm6=tmp13H
+
+	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=tmp10L
+	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp10H
+	movdqa	XMMWORD [wk(2)], xmm7	; wk(2)=tmp13L
+	movdqa	XMMWORD [wk(3)], xmm6	; wk(3)=tmp13H
+
+	pxor      xmm5,xmm5
+	pxor      xmm4,xmm4
+	punpcklwd xmm5,xmm3		; xmm5=tmp1L
+	punpckhwd xmm4,xmm3		; xmm4=tmp1H
+	psrad     xmm5,(16-CONST_BITS)	; psrad xmm5,16 & pslld xmm5,CONST_BITS
+	psrad     xmm4,(16-CONST_BITS)	; psrad xmm4,16 & pslld xmm4,CONST_BITS
+
+	movdqa	xmm0,xmm5
+	paddd	xmm5,xmm1		; xmm5=tmp11L
+	psubd	xmm0,xmm1		; xmm0=tmp12L
+	movdqa	xmm7,xmm4
+	paddd	xmm4,xmm2		; xmm4=tmp11H
+	psubd	xmm7,xmm2		; xmm7=tmp12H
+
+	movdqa	XMMWORD [wk(4)], xmm5	; wk(4)=tmp11L
+	movdqa	XMMWORD [wk(5)], xmm4	; wk(5)=tmp11H
+	movdqa	XMMWORD [wk(6)], xmm0	; wk(6)=tmp12L
+	movdqa	XMMWORD [wk(7)], xmm7	; wk(7)=tmp12H
+
+	; -- Odd part
+
+	movdqa	xmm6, XMMWORD [wk(9)]	; xmm6=col3
+	movdqa	xmm3, XMMWORD [wk(8)]	; xmm3=col1
+	movdqa	xmm1, XMMWORD [wk(11)]	; xmm1=col7
+	movdqa	xmm2, XMMWORD [wk(10)]	; xmm2=col5
+
+	movdqa	xmm5,xmm6
+	movdqa	xmm4,xmm3
+	paddw	xmm5,xmm1		; xmm5=z3
+	paddw	xmm4,xmm2		; xmm4=z4
+
+	; (Original)
+	; z5 = (z3 + z4) * 1.175875602;
+	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+	; z3 += z5;  z4 += z5;
+	;
+	; (This implementation)
+	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+
+	movdqa    xmm0,xmm5
+	movdqa    xmm7,xmm5
+	punpcklwd xmm0,xmm4
+	punpckhwd xmm7,xmm4
+	movdqa    xmm5,xmm0
+	movdqa    xmm4,xmm7
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF078_F117)]	; xmm0=z3L
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF078_F117)]	; xmm7=z3H
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F117_F078)]	; xmm5=z4L
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_F117_F078)]	; xmm4=z4H
+
+	movdqa	XMMWORD [wk(10)], xmm0	; wk(10)=z3L
+	movdqa	XMMWORD [wk(11)], xmm7	; wk(11)=z3H
+
+	; (Original)
+	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
+	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
+	;
+	; (This implementation)
+	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+	; tmp0 += z3;  tmp1 += z4;
+	; tmp2 += z3;  tmp3 += z4;
+
+	movdqa    xmm0,xmm1
+	movdqa    xmm7,xmm1
+	punpcklwd xmm0,xmm3
+	punpckhwd xmm7,xmm3
+	movdqa    xmm1,xmm0
+	movdqa    xmm3,xmm7
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm0=tmp0L
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF060_MF089)]	; xmm7=tmp0H
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_MF089_F060)]	; xmm1=tmp3L
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_MF089_F060)]	; xmm3=tmp3H
+
+	paddd	xmm0, XMMWORD [wk(10)]	; xmm0=tmp0L
+	paddd	xmm7, XMMWORD [wk(11)]	; xmm7=tmp0H
+	paddd	xmm1,xmm5		; xmm1=tmp3L
+	paddd	xmm3,xmm4		; xmm3=tmp3H
+
+	movdqa	XMMWORD [wk(8)], xmm0	; wk(8)=tmp0L
+	movdqa	XMMWORD [wk(9)], xmm7	; wk(9)=tmp0H
+
+	movdqa    xmm0,xmm2
+	movdqa    xmm7,xmm2
+	punpcklwd xmm0,xmm6
+	punpckhwd xmm7,xmm6
+	movdqa    xmm2,xmm0
+	movdqa    xmm6,xmm7
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm0=tmp1L
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF050_MF256)]	; xmm7=tmp1H
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_MF256_F050)]	; xmm2=tmp2L
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_MF256_F050)]	; xmm6=tmp2H
+
+	paddd	xmm0,xmm5		; xmm0=tmp1L
+	paddd	xmm7,xmm4		; xmm7=tmp1H
+	paddd	xmm2, XMMWORD [wk(10)]	; xmm2=tmp2L
+	paddd	xmm6, XMMWORD [wk(11)]	; xmm6=tmp2H
+
+	movdqa	XMMWORD [wk(10)], xmm0	; wk(10)=tmp1L
+	movdqa	XMMWORD [wk(11)], xmm7	; wk(11)=tmp1H
+
+	; -- Final output stage
+
+	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=tmp10L
+	movdqa	xmm4, XMMWORD [wk(1)]	; xmm4=tmp10H
+
+	movdqa	xmm0,xmm5
+	movdqa	xmm7,xmm4
+	paddd	xmm5,xmm1		; xmm5=data0L
+	paddd	xmm4,xmm3		; xmm4=data0H
+	psubd	xmm0,xmm1		; xmm0=data7L
+	psubd	xmm7,xmm3		; xmm7=data7H
+
+	movdqa	xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]	; xmm1=[PD_DESCALE_P2]
+
+	paddd	xmm5,xmm1
+	paddd	xmm4,xmm1
+	psrad	xmm5,DESCALE_P2
+	psrad	xmm4,DESCALE_P2
+	paddd	xmm0,xmm1
+	paddd	xmm7,xmm1
+	psrad	xmm0,DESCALE_P2
+	psrad	xmm7,DESCALE_P2
+
+	packssdw  xmm5,xmm4		; xmm5=data0=(00 10 20 30 40 50 60 70)
+	packssdw  xmm0,xmm7		; xmm0=data7=(07 17 27 37 47 57 67 77)
+
+	movdqa	xmm3, XMMWORD [wk(4)]	; xmm3=tmp11L
+	movdqa	xmm1, XMMWORD [wk(5)]	; xmm1=tmp11H
+
+	movdqa	xmm4,xmm3
+	movdqa	xmm7,xmm1
+	paddd	xmm3,xmm2		; xmm3=data1L
+	paddd	xmm1,xmm6		; xmm1=data1H
+	psubd	xmm4,xmm2		; xmm4=data6L
+	psubd	xmm7,xmm6		; xmm7=data6H
+
+	movdqa	xmm2,[GOTOFF(ebx,PD_DESCALE_P2)]	; xmm2=[PD_DESCALE_P2]
+
+	paddd	xmm3,xmm2
+	paddd	xmm1,xmm2
+	psrad	xmm3,DESCALE_P2
+	psrad	xmm1,DESCALE_P2
+	paddd	xmm4,xmm2
+	paddd	xmm7,xmm2
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm7,DESCALE_P2
+
+	packssdw  xmm3,xmm1		; xmm3=data1=(01 11 21 31 41 51 61 71)
+	packssdw  xmm4,xmm7		; xmm4=data6=(06 16 26 36 46 56 66 76)
+
+	packsswb  xmm5,xmm4		; xmm5=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	packsswb  xmm3,xmm0		; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	movdqa	xmm6, XMMWORD [wk(6)]	; xmm6=tmp12L
+	movdqa	xmm2, XMMWORD [wk(7)]	; xmm2=tmp12H
+	movdqa	xmm1, XMMWORD [wk(10)]	; xmm1=tmp1L
+	movdqa	xmm7, XMMWORD [wk(11)]	; xmm7=tmp1H
+
+	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	movdqa	xmm4,xmm6
+	movdqa	xmm0,xmm2
+	paddd	xmm6,xmm1		; xmm6=data2L
+	paddd	xmm2,xmm7		; xmm2=data2H
+	psubd	xmm4,xmm1		; xmm4=data5L
+	psubd	xmm0,xmm7		; xmm0=data5H
+
+	movdqa	xmm5,[GOTOFF(ebx,PD_DESCALE_P2)]	; xmm5=[PD_DESCALE_P2]
+
+	paddd	xmm6,xmm5
+	paddd	xmm2,xmm5
+	psrad	xmm6,DESCALE_P2
+	psrad	xmm2,DESCALE_P2
+	paddd	xmm4,xmm5
+	paddd	xmm0,xmm5
+	psrad	xmm4,DESCALE_P2
+	psrad	xmm0,DESCALE_P2
+
+	packssdw  xmm6,xmm2		; xmm6=data2=(02 12 22 32 42 52 62 72)
+	packssdw  xmm4,xmm0		; xmm4=data5=(05 15 25 35 45 55 65 75)
+
+	movdqa	xmm3, XMMWORD [wk(2)]	; xmm3=tmp13L
+	movdqa	xmm1, XMMWORD [wk(3)]	; xmm1=tmp13H
+	movdqa	xmm7, XMMWORD [wk(8)]	; xmm7=tmp0L
+	movdqa	xmm5, XMMWORD [wk(9)]	; xmm5=tmp0H
+
+	movdqa	xmm2,xmm3
+	movdqa	xmm0,xmm1
+	paddd	xmm3,xmm7		; xmm3=data3L
+	paddd	xmm1,xmm5		; xmm1=data3H
+	psubd	xmm2,xmm7		; xmm2=data4L
+	psubd	xmm0,xmm5		; xmm0=data4H
+
+	movdqa	xmm7,[GOTOFF(ebx,PD_DESCALE_P2)]	; xmm7=[PD_DESCALE_P2]
+
+	paddd	xmm3,xmm7
+	paddd	xmm1,xmm7
+	psrad	xmm3,DESCALE_P2
+	psrad	xmm1,DESCALE_P2
+	paddd	xmm2,xmm7
+	paddd	xmm0,xmm7
+	psrad	xmm2,DESCALE_P2
+	psrad	xmm0,DESCALE_P2
+
+	movdqa    xmm5,[GOTOFF(ebx,PB_CENTERJSAMP)]	; xmm5=[PB_CENTERJSAMP]
+
+	packssdw  xmm3,xmm1		; xmm3=data3=(03 13 23 33 43 53 63 73)
+	packssdw  xmm2,xmm0		; xmm2=data4=(04 14 24 34 44 54 64 74)
+
+	movdqa    xmm7, XMMWORD [wk(0)]	; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
+	movdqa    xmm1, XMMWORD [wk(1)]	; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
+
+	packsswb  xmm6,xmm2		; xmm6=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
+	packsswb  xmm3,xmm4		; xmm3=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
+
+	paddb     xmm7,xmm5
+	paddb     xmm1,xmm5
+	paddb     xmm6,xmm5
+	paddb     xmm3,xmm5
+
+	movdqa    xmm0,xmm7	; transpose coefficients(phase 1)
+	punpcklbw xmm7,xmm1	; xmm7=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
+	punpckhbw xmm0,xmm1	; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
+	movdqa    xmm2,xmm6	; transpose coefficients(phase 1)
+	punpcklbw xmm6,xmm3	; xmm6=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
+	punpckhbw xmm2,xmm3	; xmm2=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
+
+	movdqa    xmm4,xmm7	; transpose coefficients(phase 2)
+	punpcklwd xmm7,xmm6	; xmm7=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
+	punpckhwd xmm4,xmm6	; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
+	movdqa    xmm5,xmm2	; transpose coefficients(phase 2)
+	punpcklwd xmm2,xmm0	; xmm2=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
+	punpckhwd xmm5,xmm0	; xmm5=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
+
+	movdqa    xmm1,xmm7	; transpose coefficients(phase 3)
+	punpckldq xmm7,xmm2	; xmm7=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
+	punpckhdq xmm1,xmm2	; xmm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
+	movdqa    xmm3,xmm4	; transpose coefficients(phase 3)
+	punpckldq xmm4,xmm5	; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
+	punpckhdq xmm3,xmm5	; xmm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
+
+	pshufd	xmm6,xmm7,0x4E	; xmm6=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
+	pshufd	xmm0,xmm1,0x4E	; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
+	pshufd	xmm2,xmm4,0x4E	; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
+	pshufd	xmm5,xmm3,0x4E	; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm7
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm1
+	mov	edx, JSAMPROW [edi+4*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+6*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
+
+	mov	edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm0
+	mov	edx, JSAMPROW [edi+5*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+7*SIZEOF_JSAMPROW]
+	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm2
+	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm5
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2red-64.asm b/simd/jiss2red-64.asm
new file mode 100644
index 0000000..6807f17
--- /dev/null
+++ b/simd/jiss2red-64.asm
@@ -0,0 +1,576 @@
+;
+; jiss2red-64.asm - reduced-size IDCT (64-bit SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2009 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains inverse-DCT routines that produce reduced-size
+; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
+; The following code is based directly on the IJG's original jidctred.c;
+; see the jidctred.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1_4	(CONST_BITS-PASS1_BITS+1)
+%define DESCALE_P2_4	(CONST_BITS+PASS1_BITS+3+1)
+%define DESCALE_P1_2	(CONST_BITS-PASS1_BITS+2)
+%define DESCALE_P2_2	(CONST_BITS+PASS1_BITS+3+2)
+
+%if CONST_BITS == 13
+F_0_211	equ	 1730		; FIX(0.211164243)
+F_0_509	equ	 4176		; FIX(0.509795579)
+F_0_601	equ	 4926		; FIX(0.601344887)
+F_0_720	equ	 5906		; FIX(0.720959822)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_850	equ	 6967		; FIX(0.850430095)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_061	equ	 8697		; FIX(1.061594337)
+F_1_272	equ	10426		; FIX(1.272758580)
+F_1_451	equ	11893		; FIX(1.451774981)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_2_172	equ	17799		; FIX(2.172734803)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_624	equ	29692		; FIX(3.624509785)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_211	equ	DESCALE( 226735879,30-CONST_BITS)	; FIX(0.211164243)
+F_0_509	equ	DESCALE( 547388834,30-CONST_BITS)	; FIX(0.509795579)
+F_0_601	equ	DESCALE( 645689155,30-CONST_BITS)	; FIX(0.601344887)
+F_0_720	equ	DESCALE( 774124714,30-CONST_BITS)	; FIX(0.720959822)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_850	equ	DESCALE( 913142361,30-CONST_BITS)	; FIX(0.850430095)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_061	equ	DESCALE(1139878239,30-CONST_BITS)	; FIX(1.061594337)
+F_1_272	equ	DESCALE(1366614119,30-CONST_BITS)	; FIX(1.272758580)
+F_1_451	equ	DESCALE(1558831516,30-CONST_BITS)	; FIX(1.451774981)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_2_172	equ	DESCALE(2332956230,30-CONST_BITS)	; FIX(2.172734803)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_624	equ	DESCALE(3891787747,30-CONST_BITS)	; FIX(3.624509785)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_red_sse2)
+
+EXTN(jconst_idct_red_sse2):
+
+PW_F184_MF076	times 4 dw  F_1_847,-F_0_765
+PW_F256_F089	times 4 dw  F_2_562, F_0_899
+PW_F106_MF217	times 4 dw  F_1_061,-F_2_172
+PW_MF060_MF050	times 4 dw -F_0_601,-F_0_509
+PW_F145_MF021	times 4 dw  F_1_451,-F_0_211
+PW_F362_MF127	times 4 dw  F_3_624,-F_1_272
+PW_F085_MF072	times 4 dw  F_0_850,-F_0_720
+PD_DESCALE_P1_4	times 4 dd  1 << (DESCALE_P1_4-1)
+PD_DESCALE_P2_4	times 4 dd  1 << (DESCALE_P2_4-1)
+PD_DESCALE_P1_2	times 4 dd  1 << (DESCALE_P1_2-1)
+PD_DESCALE_P2_2	times 4 dd  1 << (DESCALE_P2_2-1)
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	64
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 4x4 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_4x4_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                      JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = void * dct_table
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13 = JDIMENSION output_col
+
+%define original_rbp	rbp+0
+%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_idct_4x4_sse2)
+
+EXTN(jsimd_idct_4x4_sse2):
+	push	rbp
+	mov	rax,rsp				; rax = original rbp
+	sub	rsp, byte 4
+	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[rsp],rax
+	mov	rbp,rsp				; rbp = aligned rbp
+	lea	rsp, [wk(0)]
+	collect_args
+
+	; ---- Pass 1: process columns from input.
+
+	mov	rdx, r10	; quantptr
+	mov	rsi, r11		; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
+	mov	eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	jnz	short .columnDCT
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	por	xmm0,xmm1
+	packsswb xmm0,xmm0
+	packsswb xmm0,xmm0
+	movd	eax,xmm0
+	test	rax,rax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	psllw	xmm0,PASS1_BITS
+
+	movdqa    xmm3,xmm0	; xmm0=in0=(00 01 02 03 04 05 06 07)
+	punpcklwd xmm0,xmm0	; xmm0=(00 00 01 01 02 02 03 03)
+	punpckhwd xmm3,xmm3	; xmm3=(04 04 05 05 06 06 07 07)
+
+	pshufd	xmm1,xmm0,0x50	; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
+	pshufd	xmm0,xmm0,0xFA	; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
+	pshufd	xmm6,xmm3,0x50	; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
+	pshufd	xmm3,xmm3,0xFA	; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
+
+	jmp	near .column_end
+%endif
+.columnDCT:
+
+	; -- Odd part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movdqa    xmm4,xmm0
+	movdqa    xmm5,xmm0
+	punpcklwd xmm4,xmm1
+	punpckhwd xmm5,xmm1
+	movdqa    xmm0,xmm4
+	movdqa    xmm1,xmm5
+	pmaddwd   xmm4,[rel PW_F256_F089]	; xmm4=(tmp2L)
+	pmaddwd   xmm5,[rel PW_F256_F089]	; xmm5=(tmp2H)
+	pmaddwd   xmm0,[rel PW_F106_MF217]	; xmm0=(tmp0L)
+	pmaddwd   xmm1,[rel PW_F106_MF217]	; xmm1=(tmp0H)
+
+	movdqa    xmm6,xmm2
+	movdqa    xmm7,xmm2
+	punpcklwd xmm6,xmm3
+	punpckhwd xmm7,xmm3
+	movdqa    xmm2,xmm6
+	movdqa    xmm3,xmm7
+	pmaddwd   xmm6,[rel PW_MF060_MF050]	; xmm6=(tmp2L)
+	pmaddwd   xmm7,[rel PW_MF060_MF050]	; xmm7=(tmp2H)
+	pmaddwd   xmm2,[rel PW_F145_MF021]	; xmm2=(tmp0L)
+	pmaddwd   xmm3,[rel PW_F145_MF021]	; xmm3=(tmp0H)
+
+	paddd	xmm6,xmm4		; xmm6=tmp2L
+	paddd	xmm7,xmm5		; xmm7=tmp2H
+	paddd	xmm2,xmm0		; xmm2=tmp0L
+	paddd	xmm3,xmm1		; xmm3=tmp0H
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=tmp0L
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=tmp0H
+
+	; -- Even part
+
+	movdqa	xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	pxor      xmm1,xmm1
+	pxor      xmm2,xmm2
+	punpcklwd xmm1,xmm4		; xmm1=tmp0L
+	punpckhwd xmm2,xmm4		; xmm2=tmp0H
+	psrad     xmm1,(16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
+	psrad     xmm2,(16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
+
+	movdqa    xmm3,xmm5		; xmm5=in2=z2
+	punpcklwd xmm5,xmm0		; xmm0=in6=z3
+	punpckhwd xmm3,xmm0
+	pmaddwd   xmm5,[rel PW_F184_MF076]	; xmm5=tmp2L
+	pmaddwd   xmm3,[rel PW_F184_MF076]	; xmm3=tmp2H
+
+	movdqa	xmm4,xmm1
+	movdqa	xmm0,xmm2
+	paddd	xmm1,xmm5		; xmm1=tmp10L
+	paddd	xmm2,xmm3		; xmm2=tmp10H
+	psubd	xmm4,xmm5		; xmm4=tmp12L
+	psubd	xmm0,xmm3		; xmm0=tmp12H
+
+	; -- Final output stage
+
+	movdqa	xmm5,xmm1
+	movdqa	xmm3,xmm2
+	paddd	xmm1,xmm6		; xmm1=data0L
+	paddd	xmm2,xmm7		; xmm2=data0H
+	psubd	xmm5,xmm6		; xmm5=data3L
+	psubd	xmm3,xmm7		; xmm3=data3H
+
+	movdqa	xmm6,[rel PD_DESCALE_P1_4]	; xmm6=[rel PD_DESCALE_P1_4]
+
+	paddd	xmm1,xmm6
+	paddd	xmm2,xmm6
+	psrad	xmm1,DESCALE_P1_4
+	psrad	xmm2,DESCALE_P1_4
+	paddd	xmm5,xmm6
+	paddd	xmm3,xmm6
+	psrad	xmm5,DESCALE_P1_4
+	psrad	xmm3,DESCALE_P1_4
+
+	packssdw  xmm1,xmm2		; xmm1=data0=(00 01 02 03 04 05 06 07)
+	packssdw  xmm5,xmm3		; xmm5=data3=(30 31 32 33 34 35 36 37)
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp0L
+	movdqa	xmm6, XMMWORD [wk(1)]	; xmm6=tmp0H
+
+	movdqa	xmm2,xmm4
+	movdqa	xmm3,xmm0
+	paddd	xmm4,xmm7		; xmm4=data1L
+	paddd	xmm0,xmm6		; xmm0=data1H
+	psubd	xmm2,xmm7		; xmm2=data2L
+	psubd	xmm3,xmm6		; xmm3=data2H
+
+	movdqa	xmm7,[rel PD_DESCALE_P1_4]	; xmm7=[rel PD_DESCALE_P1_4]
+
+	paddd	xmm4,xmm7
+	paddd	xmm0,xmm7
+	psrad	xmm4,DESCALE_P1_4
+	psrad	xmm0,DESCALE_P1_4
+	paddd	xmm2,xmm7
+	paddd	xmm3,xmm7
+	psrad	xmm2,DESCALE_P1_4
+	psrad	xmm3,DESCALE_P1_4
+
+	packssdw  xmm4,xmm0		; xmm4=data1=(10 11 12 13 14 15 16 17)
+	packssdw  xmm2,xmm3		; xmm2=data2=(20 21 22 23 24 25 26 27)
+
+	movdqa    xmm6,xmm1	; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm4	; xmm1=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm6,xmm4	; xmm6=(04 14 05 15 06 16 07 17)
+	movdqa    xmm7,xmm2	; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm5	; xmm2=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm7,xmm5	; xmm7=(24 34 25 35 26 36 27 37)
+
+	movdqa    xmm0,xmm1	; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm2	; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm0,xmm2	; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
+	movdqa    xmm3,xmm6	; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm7	; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm3,xmm7	; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
+.column_end:
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows, store into output array.
+
+	mov	rax, [original_rbp]
+	mov	rdi, r12	; (JSAMPROW *)
+	mov	rax, r13
+
+	; -- Even part
+
+	pxor      xmm4,xmm4
+	punpcklwd xmm4,xmm1		; xmm4=tmp0
+	psrad     xmm4,(16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
+
+	; -- Odd part
+
+	punpckhwd xmm1,xmm0
+	punpckhwd xmm6,xmm3
+	movdqa    xmm5,xmm1
+	movdqa    xmm2,xmm6
+	pmaddwd   xmm1,[rel PW_F256_F089]	; xmm1=(tmp2)
+	pmaddwd   xmm6,[rel PW_MF060_MF050]	; xmm6=(tmp2)
+	pmaddwd   xmm5,[rel PW_F106_MF217]	; xmm5=(tmp0)
+	pmaddwd   xmm2,[rel PW_F145_MF021]	; xmm2=(tmp0)
+
+	paddd     xmm6,xmm1		; xmm6=tmp2
+	paddd     xmm2,xmm5		; xmm2=tmp0
+
+	; -- Even part
+
+	punpcklwd xmm0,xmm3
+	pmaddwd   xmm0,[rel PW_F184_MF076]	; xmm0=tmp2
+
+	movdqa    xmm7,xmm4
+	paddd     xmm4,xmm0		; xmm4=tmp10
+	psubd     xmm7,xmm0		; xmm7=tmp12
+
+	; -- Final output stage
+
+	movdqa	xmm1,[rel PD_DESCALE_P2_4]	; xmm1=[rel PD_DESCALE_P2_4]
+
+	movdqa	xmm5,xmm4
+	movdqa	xmm3,xmm7
+	paddd	xmm4,xmm6		; xmm4=data0=(00 10 20 30)
+	paddd	xmm7,xmm2		; xmm7=data1=(01 11 21 31)
+	psubd	xmm5,xmm6		; xmm5=data3=(03 13 23 33)
+	psubd	xmm3,xmm2		; xmm3=data2=(02 12 22 32)
+
+	paddd	xmm4,xmm1
+	paddd	xmm7,xmm1
+	psrad	xmm4,DESCALE_P2_4
+	psrad	xmm7,DESCALE_P2_4
+	paddd	xmm5,xmm1
+	paddd	xmm3,xmm1
+	psrad	xmm5,DESCALE_P2_4
+	psrad	xmm3,DESCALE_P2_4
+
+	packssdw  xmm4,xmm3		; xmm4=(00 10 20 30 02 12 22 32)
+	packssdw  xmm7,xmm5		; xmm7=(01 11 21 31 03 13 23 33)
+
+	movdqa    xmm0,xmm4		; transpose coefficients(phase 1)
+	punpcklwd xmm4,xmm7		; xmm4=(00 01 10 11 20 21 30 31)
+	punpckhwd xmm0,xmm7		; xmm0=(02 03 12 13 22 23 32 33)
+
+	movdqa    xmm6,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm0		; xmm4=(00 01 02 03 10 11 12 13)
+	punpckhdq xmm6,xmm0		; xmm6=(20 21 22 23 30 31 32 33)
+
+	packsswb  xmm4,xmm6		; xmm4=(00 01 02 03 10 11 12 13 20 ..)
+	paddb     xmm4,[rel PB_CENTERJSAMP]
+
+	pshufd    xmm2,xmm4,0x39	; xmm2=(10 11 12 13 20 21 22 23 30 ..)
+	pshufd    xmm1,xmm4,0x4E	; xmm1=(20 21 22 23 30 31 32 33 00 ..)
+	pshufd    xmm3,xmm4,0x93	; xmm3=(30 31 32 33 00 01 02 03 10 ..)
+
+	mov	rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+	movd	XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
+	movd	XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
+	mov	rdx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+	movd	XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
+	movd	XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
+
+	uncollect_args
+	mov	rsp,rbp		; rsp <- aligned rbp
+	pop	rsp		; rsp <- original rbp
+	pop	rbp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 2x2 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_2x2_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                      JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+; r10 = void * dct_table
+; r11 = JCOEFPTR coef_block
+; r12 = JSAMPARRAY output_buf
+; r13 = JDIMENSION output_col
+
+	align	16
+	global	EXTN(jsimd_idct_2x2_sse2)
+
+EXTN(jsimd_idct_2x2_sse2):
+	push	rbp
+	mov	rax,rsp
+	mov	rbp,rsp
+	collect_args
+	push	rbx
+
+	; ---- Pass 1: process columns from input.
+
+	mov	rdx, r10	; quantptr
+	mov	rsi, r11		; inptr
+
+	; | input:                  | result:        |
+	; | 00 01 ** 03 ** 05 ** 07 |                |
+	; | 10 11 ** 13 ** 15 ** 17 |                |
+	; | ** ** ** ** ** ** ** ** |                |
+	; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
+	; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
+	; | 50 51 ** 53 ** 55 ** 57 |                |
+	; | ** ** ** ** ** ** ** ** |                |
+	; | 70 71 ** 73 ** 75 ** 77 |                |
+
+	; -- Odd part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
+	; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
+
+	pcmpeqd   xmm7,xmm7
+	pslld     xmm7,WORD_BIT		; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
+
+	movdqa    xmm4,xmm0		; xmm4=(10 11 ** 13 ** 15 ** 17)
+	movdqa    xmm5,xmm2		; xmm5=(50 51 ** 53 ** 55 ** 57)
+	punpcklwd xmm4,xmm1		; xmm4=(10 30 11 31 ** ** 13 33)
+	punpcklwd xmm5,xmm3		; xmm5=(50 70 51 71 ** ** 53 73)
+	pmaddwd   xmm4,[rel PW_F362_MF127]
+	pmaddwd   xmm5,[rel PW_F085_MF072]
+
+	psrld	xmm0,WORD_BIT		; xmm0=(11 -- 13 -- 15 -- 17 --)
+	pand	xmm1,xmm7		; xmm1=(-- 31 -- 33 -- 35 -- 37)
+	psrld	xmm2,WORD_BIT		; xmm2=(51 -- 53 -- 55 -- 57 --)
+	pand	xmm3,xmm7		; xmm3=(-- 71 -- 73 -- 75 -- 77)
+	por	xmm0,xmm1		; xmm0=(11 31 13 33 15 35 17 37)
+	por	xmm2,xmm3		; xmm2=(51 71 53 73 55 75 57 77)
+	pmaddwd	xmm0,[rel PW_F362_MF127]
+	pmaddwd	xmm2,[rel PW_F085_MF072]
+
+	paddd	xmm4,xmm5		; xmm4=tmp0[col0 col1 **** col3]
+	paddd	xmm0,xmm2		; xmm0=tmp0[col1 col3 col5 col7]
+
+	; -- Even part
+
+	movdqa	xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
+	pmullw	xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; xmm6=(00 01 ** 03 ** 05 ** 07)
+
+	movdqa	xmm1,xmm6		; xmm1=(00 01 ** 03 ** 05 ** 07)
+	pslld	xmm6,WORD_BIT		; xmm6=(-- 00 -- ** -- ** -- **)
+	pand	xmm1,xmm7		; xmm1=(-- 01 -- 03 -- 05 -- 07)
+	psrad	xmm6,(WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****]
+	psrad	xmm1,(WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7]
+
+	; -- Final output stage
+
+	movdqa	xmm3,xmm6
+	movdqa	xmm5,xmm1
+	paddd	xmm6,xmm4	; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
+	paddd	xmm1,xmm0	; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
+	psubd	xmm3,xmm4	; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
+	psubd	xmm5,xmm0	; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
+
+	movdqa	xmm2,[rel PD_DESCALE_P1_2]	; xmm2=[rel PD_DESCALE_P1_2]
+
+	punpckldq  xmm6,xmm3		; xmm6=(A0 B0 ** **)
+
+	movdqa     xmm7,xmm1
+	punpcklqdq xmm1,xmm5		; xmm1=(A1 A3 B1 B3)
+	punpckhqdq xmm7,xmm5		; xmm7=(A5 A7 B5 B7)
+
+	paddd	xmm6,xmm2
+	psrad	xmm6,DESCALE_P1_2
+
+	paddd	xmm1,xmm2
+	paddd	xmm7,xmm2
+	psrad	xmm1,DESCALE_P1_2
+	psrad	xmm7,DESCALE_P1_2
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows, store into output array.
+
+	mov	rdi, r12	; (JSAMPROW *)
+	mov	rax, r13
+
+	; | input:| result:|
+	; | A0 B0 |        |
+	; | A1 B1 | C0 C1  |
+	; | A3 B3 | D0 D1  |
+	; | A5 B5 |        |
+	; | A7 B7 |        |
+
+	; -- Odd part
+
+	packssdw  xmm1,xmm1		; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
+	packssdw  xmm7,xmm7		; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
+	pmaddwd   xmm1,[rel PW_F362_MF127]
+	pmaddwd   xmm7,[rel PW_F085_MF072]
+
+	paddd     xmm1,xmm7		; xmm1=tmp0[row0 row1 row0 row1]
+
+	; -- Even part
+
+	pslld     xmm6,(CONST_BITS+2)	; xmm6=tmp10[row0 row1 **** ****]
+
+	; -- Final output stage
+
+	movdqa    xmm4,xmm6
+	paddd     xmm6,xmm1	; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
+	psubd     xmm4,xmm1	; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
+
+	punpckldq xmm6,xmm4	; xmm6=(C0 D0 C1 D1)
+
+	paddd     xmm6,[rel PD_DESCALE_P2_2]
+	psrad     xmm6,DESCALE_P2_2
+
+	packssdw  xmm6,xmm6		; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
+	packsswb  xmm6,xmm6		; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
+	paddb     xmm6,[rel PB_CENTERJSAMP]
+
+	pextrw	ebx,xmm6,0x00		; ebx=(C0 D0 -- --)
+	pextrw	ecx,xmm6,0x01		; ecx=(C1 D1 -- --)
+
+	mov	rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+	mov	rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+	mov	WORD [rdx+rax*SIZEOF_JSAMPLE], bx
+	mov	WORD [rsi+rax*SIZEOF_JSAMPLE], cx
+
+	pop	rbx
+	uncollect_args
+	pop	rbp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jiss2red.asm b/simd/jiss2red.asm
new file mode 100644
index 0000000..238c61d
--- /dev/null
+++ b/simd/jiss2red.asm
@@ -0,0 +1,594 @@
+;
+; jiss2red.asm - reduced-size IDCT (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains inverse-DCT routines that produce reduced-size
+; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
+; The following code is based directly on the IJG's original jidctred.c;
+; see the jidctred.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%define CONST_BITS	13
+%define PASS1_BITS	2
+
+%define DESCALE_P1_4	(CONST_BITS-PASS1_BITS+1)
+%define DESCALE_P2_4	(CONST_BITS+PASS1_BITS+3+1)
+%define DESCALE_P1_2	(CONST_BITS-PASS1_BITS+2)
+%define DESCALE_P2_2	(CONST_BITS+PASS1_BITS+3+2)
+
+%if CONST_BITS == 13
+F_0_211	equ	 1730		; FIX(0.211164243)
+F_0_509	equ	 4176		; FIX(0.509795579)
+F_0_601	equ	 4926		; FIX(0.601344887)
+F_0_720	equ	 5906		; FIX(0.720959822)
+F_0_765	equ	 6270		; FIX(0.765366865)
+F_0_850	equ	 6967		; FIX(0.850430095)
+F_0_899	equ	 7373		; FIX(0.899976223)
+F_1_061	equ	 8697		; FIX(1.061594337)
+F_1_272	equ	10426		; FIX(1.272758580)
+F_1_451	equ	11893		; FIX(1.451774981)
+F_1_847	equ	15137		; FIX(1.847759065)
+F_2_172	equ	17799		; FIX(2.172734803)
+F_2_562	equ	20995		; FIX(2.562915447)
+F_3_624	equ	29692		; FIX(3.624509785)
+%else
+; NASM cannot do compile-time arithmetic on floating-point constants.
+%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
+F_0_211	equ	DESCALE( 226735879,30-CONST_BITS)	; FIX(0.211164243)
+F_0_509	equ	DESCALE( 547388834,30-CONST_BITS)	; FIX(0.509795579)
+F_0_601	equ	DESCALE( 645689155,30-CONST_BITS)	; FIX(0.601344887)
+F_0_720	equ	DESCALE( 774124714,30-CONST_BITS)	; FIX(0.720959822)
+F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
+F_0_850	equ	DESCALE( 913142361,30-CONST_BITS)	; FIX(0.850430095)
+F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
+F_1_061	equ	DESCALE(1139878239,30-CONST_BITS)	; FIX(1.061594337)
+F_1_272	equ	DESCALE(1366614119,30-CONST_BITS)	; FIX(1.272758580)
+F_1_451	equ	DESCALE(1558831516,30-CONST_BITS)	; FIX(1.451774981)
+F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
+F_2_172	equ	DESCALE(2332956230,30-CONST_BITS)	; FIX(2.172734803)
+F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
+F_3_624	equ	DESCALE(3891787747,30-CONST_BITS)	; FIX(3.624509785)
+%endif
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_red_sse2)
+
+EXTN(jconst_idct_red_sse2):
+
+PW_F184_MF076	times 4 dw  F_1_847,-F_0_765
+PW_F256_F089	times 4 dw  F_2_562, F_0_899
+PW_F106_MF217	times 4 dw  F_1_061,-F_2_172
+PW_MF060_MF050	times 4 dw -F_0_601,-F_0_509
+PW_F145_MF021	times 4 dw  F_1_451,-F_0_211
+PW_F362_MF127	times 4 dw  F_3_624,-F_1_272
+PW_F085_MF072	times 4 dw  F_0_850,-F_0_720
+PD_DESCALE_P1_4	times 4 dd  1 << (DESCALE_P1_4-1)
+PD_DESCALE_P2_4	times 4 dd  1 << (DESCALE_P2_4-1)
+PD_DESCALE_P1_2	times 4 dd  1 << (DESCALE_P1_2-1)
+PD_DESCALE_P2_2	times 4 dd  1 << (DESCALE_P2_2-1)
+PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 4x4 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_4x4_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                      JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_idct_4x4_sse2)
+
+EXTN(jsimd_idct_4x4_sse2):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+
+%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	short .columnDCT
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	xmm0, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	xmm0,xmm1
+	packsswb xmm0,xmm0
+	packsswb xmm0,xmm0
+	movd	eax,xmm0
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	psllw	xmm0,PASS1_BITS
+
+	movdqa    xmm3,xmm0	; xmm0=in0=(00 01 02 03 04 05 06 07)
+	punpcklwd xmm0,xmm0	; xmm0=(00 00 01 01 02 02 03 03)
+	punpckhwd xmm3,xmm3	; xmm3=(04 04 05 05 06 06 07 07)
+
+	pshufd	xmm1,xmm0,0x50	; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
+	pshufd	xmm0,xmm0,0xFA	; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
+	pshufd	xmm6,xmm3,0x50	; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
+	pshufd	xmm3,xmm3,0xFA	; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
+
+	jmp	near .column_end
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Odd part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	movdqa    xmm4,xmm0
+	movdqa    xmm5,xmm0
+	punpcklwd xmm4,xmm1
+	punpckhwd xmm5,xmm1
+	movdqa    xmm0,xmm4
+	movdqa    xmm1,xmm5
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_F256_F089)]	; xmm4=(tmp2L)
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F256_F089)]	; xmm5=(tmp2H)
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_F106_MF217)]	; xmm0=(tmp0L)
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_F106_MF217)]	; xmm1=(tmp0H)
+
+	movdqa    xmm6,xmm2
+	movdqa    xmm7,xmm2
+	punpcklwd xmm6,xmm3
+	punpckhwd xmm7,xmm3
+	movdqa    xmm2,xmm6
+	movdqa    xmm3,xmm7
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_MF060_MF050)]	; xmm6=(tmp2L)
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_MF060_MF050)]	; xmm7=(tmp2H)
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_F145_MF021)]	; xmm2=(tmp0L)
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_F145_MF021)]	; xmm3=(tmp0H)
+
+	paddd	xmm6,xmm4		; xmm6=tmp2L
+	paddd	xmm7,xmm5		; xmm7=tmp2H
+	paddd	xmm2,xmm0		; xmm2=tmp0L
+	paddd	xmm3,xmm1		; xmm3=tmp0H
+
+	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=tmp0L
+	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=tmp0H
+
+	; -- Even part
+
+	movdqa	xmm4, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm5, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm0, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm4, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm5, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	pxor      xmm1,xmm1
+	pxor      xmm2,xmm2
+	punpcklwd xmm1,xmm4		; xmm1=tmp0L
+	punpckhwd xmm2,xmm4		; xmm2=tmp0H
+	psrad     xmm1,(16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
+	psrad     xmm2,(16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
+
+	movdqa    xmm3,xmm5		; xmm5=in2=z2
+	punpcklwd xmm5,xmm0		; xmm0=in6=z3
+	punpckhwd xmm3,xmm0
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F184_MF076)]	; xmm5=tmp2L
+	pmaddwd   xmm3,[GOTOFF(ebx,PW_F184_MF076)]	; xmm3=tmp2H
+
+	movdqa	xmm4,xmm1
+	movdqa	xmm0,xmm2
+	paddd	xmm1,xmm5		; xmm1=tmp10L
+	paddd	xmm2,xmm3		; xmm2=tmp10H
+	psubd	xmm4,xmm5		; xmm4=tmp12L
+	psubd	xmm0,xmm3		; xmm0=tmp12H
+
+	; -- Final output stage
+
+	movdqa	xmm5,xmm1
+	movdqa	xmm3,xmm2
+	paddd	xmm1,xmm6		; xmm1=data0L
+	paddd	xmm2,xmm7		; xmm2=data0H
+	psubd	xmm5,xmm6		; xmm5=data3L
+	psubd	xmm3,xmm7		; xmm3=data3H
+
+	movdqa	xmm6,[GOTOFF(ebx,PD_DESCALE_P1_4)]	; xmm6=[PD_DESCALE_P1_4]
+
+	paddd	xmm1,xmm6
+	paddd	xmm2,xmm6
+	psrad	xmm1,DESCALE_P1_4
+	psrad	xmm2,DESCALE_P1_4
+	paddd	xmm5,xmm6
+	paddd	xmm3,xmm6
+	psrad	xmm5,DESCALE_P1_4
+	psrad	xmm3,DESCALE_P1_4
+
+	packssdw  xmm1,xmm2		; xmm1=data0=(00 01 02 03 04 05 06 07)
+	packssdw  xmm5,xmm3		; xmm5=data3=(30 31 32 33 34 35 36 37)
+
+	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp0L
+	movdqa	xmm6, XMMWORD [wk(1)]	; xmm6=tmp0H
+
+	movdqa	xmm2,xmm4
+	movdqa	xmm3,xmm0
+	paddd	xmm4,xmm7		; xmm4=data1L
+	paddd	xmm0,xmm6		; xmm0=data1H
+	psubd	xmm2,xmm7		; xmm2=data2L
+	psubd	xmm3,xmm6		; xmm3=data2H
+
+	movdqa	xmm7,[GOTOFF(ebx,PD_DESCALE_P1_4)]	; xmm7=[PD_DESCALE_P1_4]
+
+	paddd	xmm4,xmm7
+	paddd	xmm0,xmm7
+	psrad	xmm4,DESCALE_P1_4
+	psrad	xmm0,DESCALE_P1_4
+	paddd	xmm2,xmm7
+	paddd	xmm3,xmm7
+	psrad	xmm2,DESCALE_P1_4
+	psrad	xmm3,DESCALE_P1_4
+
+	packssdw  xmm4,xmm0		; xmm4=data1=(10 11 12 13 14 15 16 17)
+	packssdw  xmm2,xmm3		; xmm2=data2=(20 21 22 23 24 25 26 27)
+
+	movdqa    xmm6,xmm1	; transpose coefficients(phase 1)
+	punpcklwd xmm1,xmm4	; xmm1=(00 10 01 11 02 12 03 13)
+	punpckhwd xmm6,xmm4	; xmm6=(04 14 05 15 06 16 07 17)
+	movdqa    xmm7,xmm2	; transpose coefficients(phase 1)
+	punpcklwd xmm2,xmm5	; xmm2=(20 30 21 31 22 32 23 33)
+	punpckhwd xmm7,xmm5	; xmm7=(24 34 25 35 26 36 27 37)
+
+	movdqa    xmm0,xmm1	; transpose coefficients(phase 2)
+	punpckldq xmm1,xmm2	; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
+	punpckhdq xmm0,xmm2	; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
+	movdqa    xmm3,xmm6	; transpose coefficients(phase 2)
+	punpckldq xmm6,xmm7	; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
+	punpckhdq xmm3,xmm7	; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
+.column_end:
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows, store into output array.
+
+	mov	eax, [original_ebp]
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+
+	; -- Even part
+
+	pxor      xmm4,xmm4
+	punpcklwd xmm4,xmm1		; xmm4=tmp0
+	psrad     xmm4,(16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
+
+	; -- Odd part
+
+	punpckhwd xmm1,xmm0
+	punpckhwd xmm6,xmm3
+	movdqa    xmm5,xmm1
+	movdqa    xmm2,xmm6
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_F256_F089)]	; xmm1=(tmp2)
+	pmaddwd   xmm6,[GOTOFF(ebx,PW_MF060_MF050)]	; xmm6=(tmp2)
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F106_MF217)]	; xmm5=(tmp0)
+	pmaddwd   xmm2,[GOTOFF(ebx,PW_F145_MF021)]	; xmm2=(tmp0)
+
+	paddd     xmm6,xmm1		; xmm6=tmp2
+	paddd     xmm2,xmm5		; xmm2=tmp0
+
+	; -- Even part
+
+	punpcklwd xmm0,xmm3
+	pmaddwd   xmm0,[GOTOFF(ebx,PW_F184_MF076)]	; xmm0=tmp2
+
+	movdqa    xmm7,xmm4
+	paddd     xmm4,xmm0		; xmm4=tmp10
+	psubd     xmm7,xmm0		; xmm7=tmp12
+
+	; -- Final output stage
+
+	movdqa	xmm1,[GOTOFF(ebx,PD_DESCALE_P2_4)]	; xmm1=[PD_DESCALE_P2_4]
+
+	movdqa	xmm5,xmm4
+	movdqa	xmm3,xmm7
+	paddd	xmm4,xmm6		; xmm4=data0=(00 10 20 30)
+	paddd	xmm7,xmm2		; xmm7=data1=(01 11 21 31)
+	psubd	xmm5,xmm6		; xmm5=data3=(03 13 23 33)
+	psubd	xmm3,xmm2		; xmm3=data2=(02 12 22 32)
+
+	paddd	xmm4,xmm1
+	paddd	xmm7,xmm1
+	psrad	xmm4,DESCALE_P2_4
+	psrad	xmm7,DESCALE_P2_4
+	paddd	xmm5,xmm1
+	paddd	xmm3,xmm1
+	psrad	xmm5,DESCALE_P2_4
+	psrad	xmm3,DESCALE_P2_4
+
+	packssdw  xmm4,xmm3		; xmm4=(00 10 20 30 02 12 22 32)
+	packssdw  xmm7,xmm5		; xmm7=(01 11 21 31 03 13 23 33)
+
+	movdqa    xmm0,xmm4		; transpose coefficients(phase 1)
+	punpcklwd xmm4,xmm7		; xmm4=(00 01 10 11 20 21 30 31)
+	punpckhwd xmm0,xmm7		; xmm0=(02 03 12 13 22 23 32 33)
+
+	movdqa    xmm6,xmm4		; transpose coefficients(phase 2)
+	punpckldq xmm4,xmm0		; xmm4=(00 01 02 03 10 11 12 13)
+	punpckhdq xmm6,xmm0		; xmm6=(20 21 22 23 30 31 32 33)
+
+	packsswb  xmm4,xmm6		; xmm4=(00 01 02 03 10 11 12 13 20 ..)
+	paddb     xmm4,[GOTOFF(ebx,PB_CENTERJSAMP)]
+
+	pshufd    xmm2,xmm4,0x39	; xmm2=(10 11 12 13 20 21 22 23 30 ..)
+	pshufd    xmm1,xmm4,0x4E	; xmm1=(20 21 22 23 30 31 32 33 00 ..)
+	pshufd    xmm3,xmm4,0x93	; xmm3=(30 31 32 33 00 01 02 03 10 ..)
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	movd	XMM_DWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
+	movd	XMM_DWORD [esi+eax*SIZEOF_JSAMPLE], xmm2
+	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movd	XMM_DWORD [edx+eax*SIZEOF_JSAMPLE], xmm1
+	movd	XMM_DWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Perform dequantization and inverse DCT on one block of coefficients,
+; producing a reduced-size 2x2 output block.
+;
+; GLOBAL(void)
+; jsimd_idct_2x2_sse2 (void * dct_table, JCOEFPTR coef_block,
+;                      JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+	align	16
+	global	EXTN(jsimd_idct_2x2_sse2)
+
+EXTN(jsimd_idct_2x2_sse2):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input.
+
+	mov	edx, POINTER [dct_table(ebp)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(ebp)]		; inptr
+
+	; | input:                  | result:        |
+	; | 00 01 ** 03 ** 05 ** 07 |                |
+	; | 10 11 ** 13 ** 15 ** 17 |                |
+	; | ** ** ** ** ** ** ** ** |                |
+	; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
+	; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
+	; | 50 51 ** 53 ** 55 ** 57 |                |
+	; | ** ** ** ** ** ** ** ** |                |
+	; | 70 71 ** 73 ** 75 ** 77 |                |
+
+	; -- Odd part
+
+	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm1, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm0, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	movdqa	xmm2, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm2, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+	pmullw	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
+	; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
+
+	pcmpeqd   xmm7,xmm7
+	pslld     xmm7,WORD_BIT		; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
+
+	movdqa    xmm4,xmm0		; xmm4=(10 11 ** 13 ** 15 ** 17)
+	movdqa    xmm5,xmm2		; xmm5=(50 51 ** 53 ** 55 ** 57)
+	punpcklwd xmm4,xmm1		; xmm4=(10 30 11 31 ** ** 13 33)
+	punpcklwd xmm5,xmm3		; xmm5=(50 70 51 71 ** ** 53 73)
+	pmaddwd   xmm4,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd   xmm5,[GOTOFF(ebx,PW_F085_MF072)]
+
+	psrld	xmm0,WORD_BIT		; xmm0=(11 -- 13 -- 15 -- 17 --)
+	pand	xmm1,xmm7		; xmm1=(-- 31 -- 33 -- 35 -- 37)
+	psrld	xmm2,WORD_BIT		; xmm2=(51 -- 53 -- 55 -- 57 --)
+	pand	xmm3,xmm7		; xmm3=(-- 71 -- 73 -- 75 -- 77)
+	por	xmm0,xmm1		; xmm0=(11 31 13 33 15 35 17 37)
+	por	xmm2,xmm3		; xmm2=(51 71 53 73 55 75 57 77)
+	pmaddwd	xmm0,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd	xmm2,[GOTOFF(ebx,PW_F085_MF072)]
+
+	paddd	xmm4,xmm5		; xmm4=tmp0[col0 col1 **** col3]
+	paddd	xmm0,xmm2		; xmm0=tmp0[col1 col3 col5 col7]
+
+	; -- Even part
+
+	movdqa	xmm6, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	pmullw	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
+
+	; xmm6=(00 01 ** 03 ** 05 ** 07)
+
+	movdqa	xmm1,xmm6		; xmm1=(00 01 ** 03 ** 05 ** 07)
+	pslld	xmm6,WORD_BIT		; xmm6=(-- 00 -- ** -- ** -- **)
+	pand	xmm1,xmm7		; xmm1=(-- 01 -- 03 -- 05 -- 07)
+	psrad	xmm6,(WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****]
+	psrad	xmm1,(WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7]
+
+	; -- Final output stage
+
+	movdqa	xmm3,xmm6
+	movdqa	xmm5,xmm1
+	paddd	xmm6,xmm4	; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
+	paddd	xmm1,xmm0	; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
+	psubd	xmm3,xmm4	; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
+	psubd	xmm5,xmm0	; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
+
+	movdqa	xmm2,[GOTOFF(ebx,PD_DESCALE_P1_2)]	; xmm2=[PD_DESCALE_P1_2]
+
+	punpckldq  xmm6,xmm3		; xmm6=(A0 B0 ** **)
+
+	movdqa     xmm7,xmm1
+	punpcklqdq xmm1,xmm5		; xmm1=(A1 A3 B1 B3)
+	punpckhqdq xmm7,xmm5		; xmm7=(A5 A7 B5 B7)
+
+	paddd	xmm6,xmm2
+	psrad	xmm6,DESCALE_P1_2
+
+	paddd	xmm1,xmm2
+	paddd	xmm7,xmm2
+	psrad	xmm1,DESCALE_P1_2
+	psrad	xmm7,DESCALE_P1_2
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows, store into output array.
+
+	mov	edi, JSAMPARRAY [output_buf(ebp)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(ebp)]
+
+	; | input:| result:|
+	; | A0 B0 |        |
+	; | A1 B1 | C0 C1  |
+	; | A3 B3 | D0 D1  |
+	; | A5 B5 |        |
+	; | A7 B7 |        |
+
+	; -- Odd part
+
+	packssdw  xmm1,xmm1		; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
+	packssdw  xmm7,xmm7		; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
+	pmaddwd   xmm1,[GOTOFF(ebx,PW_F362_MF127)]
+	pmaddwd   xmm7,[GOTOFF(ebx,PW_F085_MF072)]
+
+	paddd     xmm1,xmm7		; xmm1=tmp0[row0 row1 row0 row1]
+
+	; -- Even part
+
+	pslld     xmm6,(CONST_BITS+2)	; xmm6=tmp10[row0 row1 **** ****]
+
+	; -- Final output stage
+
+	movdqa    xmm4,xmm6
+	paddd     xmm6,xmm1	; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
+	psubd     xmm4,xmm1	; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
+
+	punpckldq xmm6,xmm4	; xmm6=(C0 D0 C1 D1)
+
+	paddd     xmm6,[GOTOFF(ebx,PD_DESCALE_P2_2)]
+	psrad     xmm6,DESCALE_P2_2
+
+	packssdw  xmm6,xmm6		; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
+	packsswb  xmm6,xmm6		; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
+	paddb     xmm6,[GOTOFF(ebx,PB_CENTERJSAMP)]
+
+	pextrw	ebx,xmm6,0x00		; ebx=(C0 D0 -- --)
+	pextrw	ecx,xmm6,0x01		; ecx=(C1 D1 -- --)
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	mov	WORD [edx+eax*SIZEOF_JSAMPLE], bx
+	mov	WORD [esi+eax*SIZEOF_JSAMPLE], cx
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jisseflt.asm b/simd/jisseflt.asm
new file mode 100644
index 0000000..d6147c1
--- /dev/null
+++ b/simd/jisseflt.asm
@@ -0,0 +1,572 @@
+;
+; jisseflt.asm - floating-point IDCT (SSE & MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the inverse DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jidctflt.c; see the jidctflt.c for more details.
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+%include "jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_float_sse)
+
+EXTN(jconst_idct_float_sse):
+
+PD_1_414	times 4 dd  1.414213562373095048801689
+PD_1_847	times 4 dd  1.847759065022573512256366
+PD_1_082	times 4 dd  1.082392200292393968799446
+PD_M2_613	times 4 dd -2.613125929752753055713286
+PD_0_125	times 4 dd  0.125	; 1/8
+PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_float_sse (void * dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
+					; FAST_FLOAT workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_float_sse)
+
+EXTN(jsimd_idct_float_sse):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; FAST_FLOAT * wsptr
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	mm1,mm0
+	packsswb mm1,mm1
+	movd	eax,mm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq      mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+
+	punpckhwd mm1,mm0			; mm1=(** 02 ** 03)
+	punpcklwd mm0,mm0			; mm0=(00 00 01 01)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)	; mm1=in0H=(02 03)
+	psrad     mm0,(DWORD_BIT-WORD_BIT)	; mm0=in0L=(00 01)
+	cvtpi2ps  xmm3,mm1			; xmm3=(02 03 ** **)
+	cvtpi2ps  xmm0,mm0			; xmm0=(00 01 ** **)
+	movlhps   xmm0,xmm3			; xmm0=in0=(00 01 02 03)
+
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm1,xmm0
+	movaps	xmm2,xmm0
+	movaps	xmm3,xmm0
+
+	shufps	xmm0,xmm0,0x00			; xmm0=(00 00 00 00)
+	shufps	xmm1,xmm1,0x55			; xmm1=(01 01 01 01)
+	shufps	xmm2,xmm2,0xAA			; xmm2=(02 02 02 02)
+	shufps	xmm3,xmm3,0xFF			; xmm3=(03 03 03 03)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movq      mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq      mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq      mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq      mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	punpckhwd mm4,mm0			; mm4=(** 02 ** 03)
+	punpcklwd mm0,mm0			; mm0=(00 00 01 01)
+	punpckhwd mm5,mm1			; mm5=(** 22 ** 23)
+	punpcklwd mm1,mm1			; mm1=(20 20 21 21)
+
+	psrad     mm4,(DWORD_BIT-WORD_BIT)	; mm4=in0H=(02 03)
+	psrad     mm0,(DWORD_BIT-WORD_BIT)	; mm0=in0L=(00 01)
+	cvtpi2ps  xmm4,mm4			; xmm4=(02 03 ** **)
+	cvtpi2ps  xmm0,mm0			; xmm0=(00 01 ** **)
+	psrad     mm5,(DWORD_BIT-WORD_BIT)	; mm5=in2H=(22 23)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)	; mm1=in2L=(20 21)
+	cvtpi2ps  xmm5,mm5			; xmm5=(22 23 ** **)
+	cvtpi2ps  xmm1,mm1			; xmm1=(20 21 ** **)
+
+	punpckhwd mm6,mm2			; mm6=(** 42 ** 43)
+	punpcklwd mm2,mm2			; mm2=(40 40 41 41)
+	punpckhwd mm7,mm3			; mm7=(** 62 ** 63)
+	punpcklwd mm3,mm3			; mm3=(60 60 61 61)
+
+	psrad     mm6,(DWORD_BIT-WORD_BIT)	; mm6=in4H=(42 43)
+	psrad     mm2,(DWORD_BIT-WORD_BIT)	; mm2=in4L=(40 41)
+	cvtpi2ps  xmm6,mm6			; xmm6=(42 43 ** **)
+	cvtpi2ps  xmm2,mm2			; xmm2=(40 41 ** **)
+	psrad     mm7,(DWORD_BIT-WORD_BIT)	; mm7=in6H=(62 63)
+	psrad     mm3,(DWORD_BIT-WORD_BIT)	; mm3=in6L=(60 61)
+	cvtpi2ps  xmm7,mm7			; xmm7=(62 63 ** **)
+	cvtpi2ps  xmm3,mm3			; xmm3=(60 61 ** **)
+
+	movlhps   xmm0,xmm4			; xmm0=in0=(00 01 02 03)
+	movlhps   xmm1,xmm5			; xmm1=in2=(20 21 22 23)
+	mulps     xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movlhps   xmm2,xmm6			; xmm2=in4=(40 41 42 43)
+	movlhps   xmm3,xmm7			; xmm3=in6=(60 61 62 63)
+	mulps     xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[GOTOFF(ebx,PD_1_414)]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movq      mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq      mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq      mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq      mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	punpckhwd mm6,mm4			; mm6=(** 12 ** 13)
+	punpcklwd mm4,mm4			; mm4=(10 10 11 11)
+	punpckhwd mm2,mm0			; mm2=(** 32 ** 33)
+	punpcklwd mm0,mm0			; mm0=(30 30 31 31)
+
+	psrad     mm6,(DWORD_BIT-WORD_BIT)	; mm6=in1H=(12 13)
+	psrad     mm4,(DWORD_BIT-WORD_BIT)	; mm4=in1L=(10 11)
+	cvtpi2ps  xmm4,mm6			; xmm4=(12 13 ** **)
+	cvtpi2ps  xmm2,mm4			; xmm2=(10 11 ** **)
+	psrad     mm2,(DWORD_BIT-WORD_BIT)	; mm2=in3H=(32 33)
+	psrad     mm0,(DWORD_BIT-WORD_BIT)	; mm0=in3L=(30 31)
+	cvtpi2ps  xmm0,mm2			; xmm0=(32 33 ** **)
+	cvtpi2ps  xmm3,mm0			; xmm3=(30 31 ** **)
+
+	punpckhwd mm7,mm5			; mm7=(** 52 ** 53)
+	punpcklwd mm5,mm5			; mm5=(50 50 51 51)
+	punpckhwd mm3,mm1			; mm3=(** 72 ** 73)
+	punpcklwd mm1,mm1			; mm1=(70 70 71 71)
+
+	movlhps   xmm2,xmm4			; xmm2=in1=(10 11 12 13)
+	movlhps   xmm3,xmm0			; xmm3=in3=(30 31 32 33)
+
+	psrad     mm7,(DWORD_BIT-WORD_BIT)	; mm7=in5H=(52 53)
+	psrad     mm5,(DWORD_BIT-WORD_BIT)	; mm5=in5L=(50 51)
+	cvtpi2ps  xmm4,mm7			; xmm4=(52 53 ** **)
+	cvtpi2ps  xmm5,mm5			; xmm5=(50 51 ** **)
+	psrad     mm3,(DWORD_BIT-WORD_BIT)	; mm3=in7H=(72 73)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)	; mm1=in7L=(70 71)
+	cvtpi2ps  xmm0,mm3			; xmm0=(72 73 ** **)
+	cvtpi2ps  xmm1,mm1			; xmm1=(70 71 ** **)
+
+	mulps     xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movlhps   xmm5,xmm4			; xmm5=in5=(50 51 52 53)
+	movlhps   xmm1,xmm0			; xmm1=in7=(70 71 72 73)
+	mulps     xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[GOTOFF(ebx,PD_1_414)]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[GOTOFF(ebx,PD_1_847)]	; xmm0=z5
+	mulps	xmm3,[GOTOFF(ebx,PD_M2_613)]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[GOTOFF(ebx,PD_1_082)]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 01 02 03)
+	addps	xmm7,xmm3		; xmm7=data1=(10 11 12 13)
+	subps	xmm5,xmm1		; xmm5=data7=(70 71 72 73)
+	subps	xmm0,xmm3		; xmm0=data6=(60 61 62 63)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps    xmm1,xmm6		; transpose coefficients(phase 1)
+	unpcklps  xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps  xmm1,xmm7		; xmm1=(02 12 03 13)
+	movaps    xmm3,xmm0		; transpose coefficients(phase 1)
+	unpcklps  xmm0,xmm5		; xmm0=(60 70 61 71)
+	unpckhps  xmm3,xmm5		; xmm3=(62 72 63 73)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
+	movaps	xmm5, XMMWORD [wk(1)]	; xmm5=tmp3
+
+	movaps	XMMWORD [wk(0)], xmm0	; wk(0)=(60 70 61 71)
+	movaps	XMMWORD [wk(1)], xmm3	; wk(1)=(62 72 63 73)
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm0,xmm7
+	movaps	xmm3,xmm5
+	addps	xmm7,xmm2		; xmm7=data2=(20 21 22 23)
+	addps	xmm5,xmm4		; xmm5=data4=(40 41 42 43)
+	subps	xmm0,xmm2		; xmm0=data5=(50 51 52 53)
+	subps	xmm3,xmm4		; xmm3=data3=(30 31 32 33)
+
+	movaps    xmm2,xmm7		; transpose coefficients(phase 1)
+	unpcklps  xmm7,xmm3		; xmm7=(20 30 21 31)
+	unpckhps  xmm2,xmm3		; xmm2=(22 32 23 33)
+	movaps    xmm4,xmm5		; transpose coefficients(phase 1)
+	unpcklps  xmm5,xmm0		; xmm5=(40 50 41 51)
+	unpckhps  xmm4,xmm0		; xmm4=(42 52 43 53)
+
+	movaps    xmm3,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm7		; xmm6=(00 10 20 30)
+	unpckhps2 xmm3,xmm7		; xmm3=(01 11 21 31)
+	movaps    xmm0,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm2		; xmm1=(02 12 22 32)
+	unpckhps2 xmm0,xmm2		; xmm0=(03 13 23 33)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=(60 70 61 71)
+	movaps	xmm2, XMMWORD [wk(1)]	; xmm2=(62 72 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+
+	movaps    xmm6,xmm5		; transpose coefficients(phase 2)
+	unpcklps2 xmm5,xmm7		; xmm5=(40 50 60 70)
+	unpckhps2 xmm6,xmm7		; xmm6=(41 51 61 71)
+	movaps    xmm3,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(42 52 62 72)
+	unpckhps2 xmm3,xmm2		; xmm3=(43 53 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
+	movaps	XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
+
+.nextcolumn:
+	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 4*SIZEOF_FLOAT_MULT_TYPE	; quantptr
+	add	edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; FAST_FLOAT * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.rowloop:
+
+	; -- Even part
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[GOTOFF(ebx,PD_1_414)]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[GOTOFF(ebx,PD_1_414)]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[GOTOFF(ebx,PD_1_847)]	; xmm0=z5
+	mulps	xmm3,[GOTOFF(ebx,PD_M2_613)]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[GOTOFF(ebx,PD_1_082)]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 10 20 30)
+	addps	xmm7,xmm3		; xmm7=data1=(01 11 21 31)
+	subps	xmm5,xmm1		; xmm5=data7=(07 17 27 37)
+	subps	xmm0,xmm3		; xmm0=data6=(06 16 26 36)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps	xmm1,[GOTOFF(ebx,PD_0_125)]	; xmm1=[PD_0_125]
+
+	mulps	xmm6,xmm1		; descale(1/8)
+	mulps	xmm7,xmm1		; descale(1/8)
+	mulps	xmm5,xmm1		; descale(1/8)
+	mulps	xmm0,xmm1		; descale(1/8)
+
+	movhlps   xmm3,xmm6
+	movhlps   xmm1,xmm7
+	cvtps2pi  mm0,xmm6		; round to int32, mm0=data0L=(00 10)
+	cvtps2pi  mm1,xmm7		; round to int32, mm1=data1L=(01 11)
+	cvtps2pi  mm2,xmm3		; round to int32, mm2=data0H=(20 30)
+	cvtps2pi  mm3,xmm1		; round to int32, mm3=data1H=(21 31)
+	packssdw  mm0,mm2		; mm0=data0=(00 10 20 30)
+	packssdw  mm1,mm3		; mm1=data1=(01 11 21 31)
+
+	movhlps   xmm6,xmm5
+	movhlps   xmm7,xmm0
+	cvtps2pi  mm4,xmm5		; round to int32, mm4=data7L=(07 17)
+	cvtps2pi  mm5,xmm0		; round to int32, mm5=data6L=(06 16)
+	cvtps2pi  mm6,xmm6		; round to int32, mm6=data7H=(27 37)
+	cvtps2pi  mm7,xmm7		; round to int32, mm7=data6H=(26 36)
+	packssdw  mm4,mm6		; mm4=data7=(07 17 27 37)
+	packssdw  mm5,mm7		; mm5=data6=(06 16 26 36)
+
+	packsswb  mm0,mm5		; mm0=(00 10 20 30 06 16 26 36)
+	packsswb  mm1,mm4		; mm1=(01 11 21 31 07 17 27 37)
+
+	movaps	xmm3, XMMWORD [wk(0)]	; xmm3=tmp2
+	movaps	xmm1, XMMWORD [wk(1)]	; xmm1=tmp3
+
+	movaps	xmm6,[GOTOFF(ebx,PD_0_125)]	; xmm6=[PD_0_125]
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm5,xmm3
+	movaps	xmm0,xmm1
+	addps	xmm3,xmm2		; xmm3=data2=(02 12 22 32)
+	addps	xmm1,xmm4		; xmm1=data4=(04 14 24 34)
+	subps	xmm5,xmm2		; xmm5=data5=(05 15 25 35)
+	subps	xmm0,xmm4		; xmm0=data3=(03 13 23 33)
+
+	mulps	xmm3,xmm6		; descale(1/8)
+	mulps	xmm1,xmm6		; descale(1/8)
+	mulps	xmm5,xmm6		; descale(1/8)
+	mulps	xmm0,xmm6		; descale(1/8)
+
+	movhlps   xmm7,xmm3
+	movhlps   xmm2,xmm1
+	cvtps2pi  mm2,xmm3		; round to int32, mm2=data2L=(02 12)
+	cvtps2pi  mm3,xmm1		; round to int32, mm3=data4L=(04 14)
+	cvtps2pi  mm6,xmm7		; round to int32, mm6=data2H=(22 32)
+	cvtps2pi  mm7,xmm2		; round to int32, mm7=data4H=(24 34)
+	packssdw  mm2,mm6		; mm2=data2=(02 12 22 32)
+	packssdw  mm3,mm7		; mm3=data4=(04 14 24 34)
+
+	movhlps   xmm4,xmm5
+	movhlps   xmm6,xmm0
+	cvtps2pi  mm5,xmm5		; round to int32, mm5=data5L=(05 15)
+	cvtps2pi  mm4,xmm0		; round to int32, mm4=data3L=(03 13)
+	cvtps2pi  mm6,xmm4		; round to int32, mm6=data5H=(25 35)
+	cvtps2pi  mm7,xmm6		; round to int32, mm7=data3H=(23 33)
+	packssdw  mm5,mm6		; mm5=data5=(05 15 25 35)
+	packssdw  mm4,mm7		; mm4=data3=(03 13 23 33)
+
+	movq      mm6,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm6=[PB_CENTERJSAMP]
+
+	packsswb  mm2,mm3		; mm2=(02 12 22 32 04 14 24 34)
+	packsswb  mm4,mm5		; mm4=(03 13 23 33 05 15 25 35)
+
+	paddb     mm0,mm6
+	paddb     mm1,mm6
+	paddb     mm2,mm6
+	paddb     mm4,mm6
+
+	movq      mm7,mm0		; transpose coefficients(phase 1)
+	punpcklbw mm0,mm1		; mm0=(00 01 10 11 20 21 30 31)
+	punpckhbw mm7,mm1		; mm7=(06 07 16 17 26 27 36 37)
+	movq      mm3,mm2		; transpose coefficients(phase 1)
+	punpcklbw mm2,mm4		; mm2=(02 03 12 13 22 23 32 33)
+	punpckhbw mm3,mm4		; mm3=(04 05 14 15 24 25 34 35)
+
+	movq      mm5,mm0		; transpose coefficients(phase 2)
+	punpcklwd mm0,mm2		; mm0=(00 01 02 03 10 11 12 13)
+	punpckhwd mm5,mm2		; mm5=(20 21 22 23 30 31 32 33)
+	movq      mm6,mm3		; transpose coefficients(phase 2)
+	punpcklwd mm3,mm7		; mm3=(04 05 06 07 14 15 16 17)
+	punpckhwd mm6,mm7		; mm6=(24 25 26 27 34 35 36 37)
+
+	movq      mm1,mm0		; transpose coefficients(phase 3)
+	punpckldq mm0,mm3		; mm0=(00 01 02 03 04 05 06 07)
+	punpckhdq mm1,mm3		; mm1=(10 11 12 13 14 15 16 17)
+	movq      mm4,mm5		; transpose coefficients(phase 3)
+	punpckldq mm5,mm6		; mm5=(20 21 22 23 24 25 26 27)
+	punpckhdq mm4,mm6		; mm4=(30 31 32 33 34 35 36 37)
+
+	pushpic	ebx			; save GOT address
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
+	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
+
+	poppic	ebx			; restore GOT address
+
+	add	esi, byte 4*SIZEOF_FAST_FLOAT	; wsptr
+	add	edi, byte 4*SIZEOF_JSAMPROW
+	dec	ecx				; ctr
+	jnz	near .rowloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jsimd.h b/simd/jsimd.h
new file mode 100644
index 0000000..c21cf29
--- /dev/null
+++ b/simd/jsimd.h
@@ -0,0 +1,503 @@
+/*
+ * simd/jsimd.h
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * 
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ *
+ */
+
+/* Bitmask for supported acceleration methods */
+
+#define JSIMD_NONE    0x00
+#define JSIMD_MMX     0x01
+#define JSIMD_3DNOW   0x02
+#define JSIMD_SSE     0x04
+#define JSIMD_SSE2    0x08
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_simd_cpu_support                 jSiCpuSupport
+#define jsimd_rgb_ycc_convert_mmx             jSRGBYCCM
+#define jsimd_extrgb_ycc_convert_mmx          jSEXTRGBYCCM
+#define jsimd_extrgbx_ycc_convert_mmx         jSEXTRGBXYCCM
+#define jsimd_extbgr_ycc_convert_mmx          jSEXTBGRYCCM
+#define jsimd_extbgrx_ycc_convert_mmx         jSEXTBGRXYCCM
+#define jsimd_extxbgr_ycc_convert_mmx         jSEXTXBGRYCCM
+#define jsimd_extxrgb_ycc_convert_mmx         jSEXTXRGBYCCM
+#define jsimd_ycc_rgb_convert_mmx             jSYCCRGBM
+#define jsimd_ycc_extrgb_convert_mmx          jSYCCEXTRGBM
+#define jsimd_ycc_extrgbx_convert_mmx         jSYCCEXTRGBXM
+#define jsimd_ycc_extbgr_convert_mmx          jSYCCEXTBGRM
+#define jsimd_ycc_extbgrx_convert_mmx         jSYCCEXTBGRXM
+#define jsimd_ycc_extxbgr_convert_mmx         jSYCCEXTXBGRM
+#define jsimd_ycc_extxrgb_convert_mmx         jSYCCEXTXRGBM
+#define jconst_rgb_ycc_convert_sse2           jSCRGBYCCS2
+#define jsimd_rgb_ycc_convert_sse2            jSRGBYCCS2
+#define jsimd_extrgb_ycc_convert_sse2         jSEXTRGBYCCS2
+#define jsimd_extrgbx_ycc_convert_sse2        jSEXTRGBXYCCS2
+#define jsimd_extbgr_ycc_convert_sse2         jSEXTBGRYCCS2
+#define jsimd_extbgrx_ycc_convert_sse2        jSEXTBGRXYCCS2
+#define jsimd_extxbgr_ycc_convert_sse2        jSEXTXBGRYCCS2
+#define jsimd_extxrgb_ycc_convert_sse2        jSEXTXRGBYCCS2
+#define jconst_ycc_rgb_convert_sse2           jSCYCCRGBS2
+#define jsimd_ycc_rgb_convert_sse2            jSYCCRGBS2
+#define jsimd_ycc_extrgb_convert_sse2         jSYCCEXTRGBS2
+#define jsimd_ycc_extrgbx_convert_sse2        jSYCCEXTRGBXS2
+#define jsimd_ycc_extbgr_convert_sse2         jSYCCEXTBGRS2
+#define jsimd_ycc_extbgrx_convert_sse2        jSYCCEXTBGRXS2
+#define jsimd_ycc_extxbgr_convert_sse2        jSYCCEXTXBGRS2
+#define jsimd_ycc_extxrgb_convert_sse2        jSYCCEXTXRGBS2
+#define jsimd_h2v2_downsample_mmx             jSDnH2V2M
+#define jsimd_h2v1_downsample_mmx             jSDnH2V1M
+#define jsimd_h2v2_downsample_sse2            jSDnH2V2S2
+#define jsimd_h2v1_downsample_sse2            jSDnH2V1S2
+#define jsimd_h2v2_upsample_mmx               jSUpH2V2M
+#define jsimd_h2v1_upsample_mmx               jSUpH2V1M
+#define jsimd_h2v2_fancy_upsample_mmx         jSFUpH2V2M
+#define jsimd_h2v1_fancy_upsample_mmx         jSFUpH2V1M
+#define jsimd_h2v2_merged_upsample_mmx        jSMUpH2V2M
+#define jsimd_h2v2_extrgb_merged_upsample_mmx jSMUpH2V2EXTRGBM
+#define jsimd_h2v2_extrgbx_merged_upsample_mmx jSMUpH2V2EXTRGBXM
+#define jsimd_h2v2_extbgr_merged_upsample_mmx jSMUpH2V2EXTBGRM
+#define jsimd_h2v2_extbgrx_merged_upsample_mmx jSMUpH2V2EXTBGRXM
+#define jsimd_h2v2_extxbgr_merged_upsample_mmx jSMUpH2V2EXTXBGRM
+#define jsimd_h2v2_extxrgb_merged_upsample_mmx jSMUpH2V2EXTXRGBM
+#define jsimd_h2v1_merged_upsample_mmx        jSMUpH2V1M
+#define jsimd_h2v1_extrgb_merged_upsample_mmx jSMUpH2V1EXTRGBM
+#define jsimd_h2v1_extrgbx_merged_upsample_mmx jSMUpH2V1EXTRGBXM
+#define jsimd_h2v1_extbgr_merged_upsample_mmx jSMUpH2V1EXTBGRM
+#define jsimd_h2v1_extbgrx_merged_upsample_mmx jSMUpH2V1EXTBGRXM
+#define jsimd_h2v1_extxbgr_merged_upsample_mmx jSMUpH2V1EXTXBGRM
+#define jsimd_h2v1_extxrgb_merged_upsample_mmx jSMUpH2V1EXTXRGBM
+#define jsimd_h2v2_upsample_sse2              jSUpH2V2S2
+#define jsimd_h2v1_upsample_sse2              jSUpH2V1S2
+#define jconst_fancy_upsample_sse2            jSCFUpS2
+#define jsimd_h2v2_fancy_upsample_sse2        jSFUpH2V2S2
+#define jsimd_h2v1_fancy_upsample_sse2        jSFUpH2V1S2
+#define jconst_merged_upsample_sse2           jSCMUpS2
+#define jsimd_h2v2_merged_upsample_sse2       jSMUpH2V2S2
+#define jsimd_h2v2_extrgb_merged_upsample_sse2 jSMUpH2V2EXTRGBS2
+#define jsimd_h2v2_extrgbx_merged_upsample_sse2 jSMUpH2V2EXTRGBXS2
+#define jsimd_h2v2_extbgr_merged_upsample_sse2 jSMUpH2V2EXTBGRS2
+#define jsimd_h2v2_extbgrx_merged_upsample_sse2 jSMUpH2V2EXTBGRXS2
+#define jsimd_h2v2_extxbgr_merged_upsample_sse2 jSMUpH2V2EXTXBGRS2
+#define jsimd_h2v2_extxrgb_merged_upsample_sse2 jSMUpH2V2EXTXRGBS2
+#define jsimd_h2v1_merged_upsample_sse2       jSMUpH2V1S2
+#define jsimd_h2v1_extrgb_merged_upsample_sse2 jSMUpH2V1EXTRGBS2
+#define jsimd_h2v1_extrgbx_merged_upsample_sse2 jSMUpH2V1EXTRGBXS2
+#define jsimd_h2v1_extbgr_merged_upsample_sse2 jSMUpH2V1EXTBGRS2
+#define jsimd_h2v1_extbgrx_merged_upsample_sse2 jSMUpH2V1EXTBGRXS2
+#define jsimd_h2v1_extxbgr_merged_upsample_sse2 jSMUpH2V1EXTXBGRS2
+#define jsimd_h2v1_extxrgb_merged_upsample_sse2 jSMUpH2V1EXTXRGBS2
+#define jsimd_convsamp_mmx                    jSConvM
+#define jsimd_convsamp_sse2                   jSConvS2
+#define jsimd_convsamp_float_3dnow            jSConvF3D
+#define jsimd_convsamp_float_sse              jSConvFS
+#define jsimd_convsamp_float_sse2             jSConvFS2
+#define jsimd_fdct_islow_mmx                  jSFDMIS
+#define jsimd_fdct_ifast_mmx                  jSFDMIF
+#define jconst_fdct_islow_sse2                jSCFDS2IS
+#define jsimd_fdct_islow_sse2                 jSFDS2IS
+#define jconst_fdct_ifast_sse2                jSCFDS2IF
+#define jsimd_fdct_ifast_sse2                 jSFDS2IF
+#define jsimd_fdct_float_3dnow                jSFD3DF
+#define jconst_fdct_float_sse                 jSCFDSF
+#define jsimd_fdct_float_sse                  jSFDSF
+#define jsimd_quantize_mmx                    jSQuantM
+#define jsimd_quantize_sse2                   jSQuantS2
+#define jsimd_quantize_float_3dnow            jSQuantF3D
+#define jsimd_quantize_float_sse              jSQuantFS
+#define jsimd_quantize_float_sse2             jSQuantFS2
+#define jsimd_idct_2x2_mmx                    jSIDM22
+#define jsimd_idct_4x4_mmx                    jSIDM44
+#define jconst_idct_red_sse2                  jSCIDS2R
+#define jsimd_idct_2x2_sse2                   jSIDS222
+#define jsimd_idct_4x4_sse2                   jSIDS244
+#define jsimd_idct_islow_mmx                  jSIDMIS
+#define jsimd_idct_ifast_mmx                  jSIDMIF
+#define jconst_idct_islow_sse2                jSCIDS2IS
+#define jsimd_idct_islow_sse2                 jSIDS2IS
+#define jconst_idct_ifast_sse2                jSCIDS2IF
+#define jsimd_idct_ifast_sse2                 jSIDS2IF
+#define jsimd_idct_float_3dnow                jSID3DF
+#define jconst_fdct_float_sse                 jSCIDSF
+#define jsimd_idct_float_sse                  jSIDSF
+#define jconst_fdct_float_sse2                jSCIDS2F
+#define jsimd_idct_float_sse2                 jSIDS2F
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* SIMD Ext: retrieve SIMD/CPU information */
+EXTERN(unsigned int) jpeg_simd_cpu_support JPP((void));
+
+/* SIMD Color Space Conversion */
+EXTERN(void) jsimd_rgb_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extrgb_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extrgbx_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extbgr_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extbgrx_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extxbgr_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extxrgb_ycc_convert_mmx
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+
+EXTERN(void) jsimd_ycc_rgb_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extrgb_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extrgbx_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extbgr_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extbgrx_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extxbgr_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extxrgb_convert_mmx
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+
+extern const int jconst_rgb_ycc_convert_sse2[];
+EXTERN(void) jsimd_rgb_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extrgb_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extrgbx_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extbgr_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extbgrx_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extxbgr_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extxrgb_ycc_convert_sse2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+
+extern const int jconst_ycc_rgb_convert_sse2[];
+EXTERN(void) jsimd_ycc_rgb_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extrgb_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extrgbx_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extbgr_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extbgrx_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extxbgr_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+EXTERN(void) jsimd_ycc_extxrgb_convert_sse2
+        JPP((JDIMENSION out_width,
+             JSAMPIMAGE input_buf, JDIMENSION input_row,
+             JSAMPARRAY output_buf, int num_rows));
+
+/* SIMD Downsample */
+EXTERN(void) jsimd_h2v2_downsample_mmx
+        JPP((JDIMENSION image_width, int max_v_samp_factor,
+             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+             JSAMPARRAY input_data, JSAMPARRAY output_data));
+EXTERN(void) jsimd_h2v1_downsample_mmx
+        JPP((JDIMENSION image_width, int max_v_samp_factor,
+             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+             JSAMPARRAY input_data, JSAMPARRAY output_data));
+
+EXTERN(void) jsimd_h2v2_downsample_sse2
+        JPP((JDIMENSION image_width, int max_v_samp_factor,
+             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+             JSAMPARRAY input_data, JSAMPARRAY output_data));
+EXTERN(void) jsimd_h2v1_downsample_sse2
+        JPP((JDIMENSION image_width, int max_v_samp_factor,
+             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
+             JSAMPARRAY input_data, JSAMPARRAY output_data));
+
+/* SIMD Upsample */
+EXTERN(void) jsimd_h2v2_upsample_mmx
+        JPP((int max_v_samp_factor, JDIMENSION output_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+EXTERN(void) jsimd_h2v1_upsample_mmx
+        JPP((int max_v_samp_factor, JDIMENSION output_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+EXTERN(void) jsimd_h2v2_fancy_upsample_mmx
+        JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+EXTERN(void) jsimd_h2v1_fancy_upsample_mmx
+        JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+EXTERN(void) jsimd_h2v2_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extrgb_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extrgbx_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extbgr_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extbgrx_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extxbgr_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extxrgb_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extrgb_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extrgbx_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extbgr_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extbgrx_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extxbgr_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extxrgb_merged_upsample_mmx
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+
+EXTERN(void) jsimd_h2v2_upsample_sse2
+        JPP((int max_v_samp_factor, JDIMENSION output_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+EXTERN(void) jsimd_h2v1_upsample_sse2
+        JPP((int max_v_samp_factor, JDIMENSION output_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+extern const int jconst_fancy_upsample_sse2[];
+EXTERN(void) jsimd_h2v2_fancy_upsample_sse2
+        JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+EXTERN(void) jsimd_h2v1_fancy_upsample_sse2
+        JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
+             JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+extern const int jconst_merged_upsample_sse2[];
+EXTERN(void) jsimd_h2v2_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extrgb_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extrgbx_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extbgr_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extbgrx_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extxbgr_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v2_extxrgb_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extrgb_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extrgbx_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extbgr_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extbgrx_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extxbgr_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+EXTERN(void) jsimd_h2v1_extxrgb_merged_upsample_sse2
+        JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
+             JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
+
+/* SIMD Sample Conversion */
+EXTERN(void) jsimd_convsamp_mmx JPP((JSAMPARRAY sample_data,
+                                     JDIMENSION start_col,
+                                     DCTELEM * workspace));
+
+EXTERN(void) jsimd_convsamp_sse2 JPP((JSAMPARRAY sample_data,
+                                      JDIMENSION start_col,
+                                      DCTELEM * workspace));
+
+EXTERN(void) jsimd_convsamp_float_3dnow JPP((JSAMPARRAY sample_data,
+                                             JDIMENSION start_col,
+                                             FAST_FLOAT * workspace));
+
+EXTERN(void) jsimd_convsamp_float_sse JPP((JSAMPARRAY sample_data,
+                                           JDIMENSION start_col,
+                                           FAST_FLOAT * workspace));
+
+EXTERN(void) jsimd_convsamp_float_sse2 JPP((JSAMPARRAY sample_data,
+                                            JDIMENSION start_col,
+                                            FAST_FLOAT * workspace));
+
+/* SIMD Forward DCT */
+EXTERN(void) jsimd_fdct_islow_mmx JPP((DCTELEM * data));
+EXTERN(void) jsimd_fdct_ifast_mmx JPP((DCTELEM * data));
+
+extern const int jconst_fdct_ifast_sse2[];
+EXTERN(void) jsimd_fdct_islow_sse2 JPP((DCTELEM * data));
+extern const int jconst_fdct_islow_sse2[];
+EXTERN(void) jsimd_fdct_ifast_sse2 JPP((DCTELEM * data));
+
+EXTERN(void) jsimd_fdct_float_3dnow JPP((FAST_FLOAT * data));
+
+extern const int jconst_fdct_float_sse[];
+EXTERN(void) jsimd_fdct_float_sse JPP((FAST_FLOAT * data));
+
+/* SIMD Quantization */
+EXTERN(void) jsimd_quantize_mmx JPP((JCOEFPTR coef_block,
+                                     DCTELEM * divisors,
+                                     DCTELEM * workspace));
+
+EXTERN(void) jsimd_quantize_sse2 JPP((JCOEFPTR coef_block,
+                                      DCTELEM * divisors,
+                                      DCTELEM * workspace));
+
+EXTERN(void) jsimd_quantize_float_3dnow JPP((JCOEFPTR coef_block,
+                                             FAST_FLOAT * divisors,
+                                             FAST_FLOAT * workspace));
+
+EXTERN(void) jsimd_quantize_float_sse JPP((JCOEFPTR coef_block,
+                                           FAST_FLOAT * divisors,
+                                           FAST_FLOAT * workspace));
+
+EXTERN(void) jsimd_quantize_float_sse2 JPP((JCOEFPTR coef_block,
+                                            FAST_FLOAT * divisors,
+                                            FAST_FLOAT * workspace));
+
+/* SIMD Reduced Inverse DCT */
+EXTERN(void) jsimd_idct_2x2_mmx JPP((void * dct_table,
+                                     JCOEFPTR coef_block,
+                                     JSAMPARRAY output_buf,
+                                     JDIMENSION output_col));
+EXTERN(void) jsimd_idct_4x4_mmx JPP((void * dct_table,
+                                     JCOEFPTR coef_block,
+                                     JSAMPARRAY output_buf,
+                                     JDIMENSION output_col));
+
+extern const int jconst_idct_red_sse2[];
+EXTERN(void) jsimd_idct_2x2_sse2 JPP((void * dct_table,
+                                      JCOEFPTR coef_block,
+                                      JSAMPARRAY output_buf,
+                                      JDIMENSION output_col));
+EXTERN(void) jsimd_idct_4x4_sse2 JPP((void * dct_table,
+                                      JCOEFPTR coef_block,
+                                      JSAMPARRAY output_buf,
+                                      JDIMENSION output_col));
+
+/* SIMD Inverse DCT */
+EXTERN(void) jsimd_idct_islow_mmx JPP((void * dct_table,
+                                       JCOEFPTR coef_block,
+                                       JSAMPARRAY output_buf,
+                                       JDIMENSION output_col));
+EXTERN(void) jsimd_idct_ifast_mmx JPP((void * dct_table,
+                                       JCOEFPTR coef_block,
+                                       JSAMPARRAY output_buf,
+                                       JDIMENSION output_col));
+
+extern const int jconst_idct_islow_sse2[];
+EXTERN(void) jsimd_idct_islow_sse2 JPP((void * dct_table,
+                                        JCOEFPTR coef_block,
+                                        JSAMPARRAY output_buf,
+                                        JDIMENSION output_col));
+extern const int jconst_idct_ifast_sse2[];
+EXTERN(void) jsimd_idct_ifast_sse2 JPP((void * dct_table,
+                                        JCOEFPTR coef_block,
+                                        JSAMPARRAY output_buf,
+                                        JDIMENSION output_col));
+
+EXTERN(void) jsimd_idct_float_3dnow JPP((void * dct_table,
+                                         JCOEFPTR coef_block,
+                                         JSAMPARRAY output_buf,
+                                         JDIMENSION output_col));
+
+extern const int jconst_idct_float_sse[];
+EXTERN(void) jsimd_idct_float_sse JPP((void * dct_table,
+                                       JCOEFPTR coef_block,
+                                       JSAMPARRAY output_buf,
+                                       JDIMENSION output_col));
+
+extern const int jconst_idct_float_sse2[];
+EXTERN(void) jsimd_idct_float_sse2 JPP((void * dct_table,
+                                        JCOEFPTR coef_block,
+                                        JSAMPARRAY output_buf,
+                                        JDIMENSION output_col));
+
diff --git a/simd/jsimd_i386.c b/simd/jsimd_i386.c
new file mode 100644
index 0000000..ab52519
--- /dev/null
+++ b/simd/jsimd_i386.c
@@ -0,0 +1,956 @@
+/*
+ * jsimd_i386.c
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright 2009 D. R. Commander
+ * 
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ *
+ * This file contains the interface between the "normal" portions
+ * of the library and the SIMD implementations when running on a
+ * 32-bit x86 architecture.
+ */
+
+#define JPEG_INTERNALS
+#include "../jinclude.h"
+#include "../jpeglib.h"
+#include "../jsimd.h"
+#include "../jdct.h"
+#include "../jsimddct.h"
+#include "jsimd.h"
+
+/*
+ * In the PIC cases, we have no guarantee that constants will keep
+ * their alignment. This macro allows us to verify it at runtime.
+ */
+#define IS_ALIGNED(ptr, order) (((unsigned)ptr & ((1 << order) - 1)) == 0)
+
+#define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
+
+static unsigned int simd_support = ~0;
+
+/*
+ * Check what SIMD accelerations are supported.
+ *
+ * FIXME: This code is racy under a multi-threaded environment.
+ */
+LOCAL(void)
+init_simd (void)
+{
+  char *env = NULL;
+
+  if (simd_support != ~0)
+    return;
+
+  simd_support = jpeg_simd_cpu_support();
+
+  /* Force different settings through environment variables */
+  env = getenv("JSIMD_FORCEMMX");
+  if ((env != NULL) && (strcmp(env, "1") == 0))
+    simd_support &= JSIMD_MMX;
+  env = getenv("JSIMD_FORCE3DNOW");
+  if ((env != NULL) && (strcmp(env, "1") == 0))
+    simd_support &= JSIMD_3DNOW|JSIMD_MMX;
+  env = getenv("JSIMD_FORCESSE");
+  if ((env != NULL) && (strcmp(env, "1") == 0))
+    simd_support &= JSIMD_SSE|JSIMD_MMX;
+  env = getenv("JSIMD_FORCESSE2");
+  if ((env != NULL) && (strcmp(env, "1") == 0))
+    simd_support &= JSIMD_SSE2;
+}
+
+GLOBAL(int)
+jsimd_can_rgb_ycc (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
+                       JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+                       JDIMENSION output_row, int num_rows)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+  void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+
+  switch(cinfo->in_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_extrgb_ycc_convert_sse2;
+      mmxfct=jsimd_extrgb_ycc_convert_mmx;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_extrgbx_ycc_convert_sse2;
+      mmxfct=jsimd_extrgbx_ycc_convert_mmx;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_extbgr_ycc_convert_sse2;
+      mmxfct=jsimd_extbgr_ycc_convert_mmx;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_extbgrx_ycc_convert_sse2;
+      mmxfct=jsimd_extbgrx_ycc_convert_mmx;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_extxbgr_ycc_convert_sse2;
+      mmxfct=jsimd_extxbgr_ycc_convert_mmx;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_extxrgb_ycc_convert_sse2;
+      mmxfct=jsimd_extxrgb_ycc_convert_mmx;
+      break;
+    default:
+      sse2fct=jsimd_rgb_ycc_convert_sse2;
+      mmxfct=jsimd_rgb_ycc_convert_mmx;
+      break;
+  }
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
+    sse2fct(cinfo->image_width, input_buf,
+        output_buf, output_row, num_rows);
+  else if (simd_support & JSIMD_MMX)
+    mmxfct(cinfo->image_width, input_buf,
+        output_buf, output_row, num_rows);
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
+                       JSAMPIMAGE input_buf, JDIMENSION input_row,
+                       JSAMPARRAY output_buf, int num_rows)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
+  void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
+
+  switch(cinfo->out_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_ycc_extrgb_convert_sse2;
+      mmxfct=jsimd_ycc_extrgb_convert_mmx;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_ycc_extrgbx_convert_sse2;
+      mmxfct=jsimd_ycc_extrgbx_convert_mmx;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_ycc_extbgr_convert_sse2;
+      mmxfct=jsimd_ycc_extbgr_convert_mmx;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_ycc_extbgrx_convert_sse2;
+      mmxfct=jsimd_ycc_extbgrx_convert_mmx;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_ycc_extxbgr_convert_sse2;
+      mmxfct=jsimd_ycc_extxbgr_convert_mmx;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_ycc_extxrgb_convert_sse2;
+      mmxfct=jsimd_ycc_extxrgb_convert_mmx;
+      break;
+    default:
+      sse2fct=jsimd_ycc_rgb_convert_sse2;
+      mmxfct=jsimd_ycc_rgb_convert_mmx;
+      break;
+  }
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
+    sse2fct(cinfo->output_width, input_buf,
+        input_row, output_buf, num_rows);
+  else if (simd_support & JSIMD_MMX)
+    mmxfct(cinfo->output_width, input_buf,
+        input_row, output_buf, num_rows);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_downsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_downsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+                       JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
+        compptr->v_samp_factor, compptr->width_in_blocks,
+        input_data, output_data);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
+        compptr->v_samp_factor, compptr->width_in_blocks,
+        input_data, output_data);
+}
+
+GLOBAL(void)
+jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+                       JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
+        compptr->v_samp_factor, compptr->width_in_blocks,
+        input_data, output_data);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
+        compptr->v_samp_factor, compptr->width_in_blocks,
+        input_data, output_data);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_upsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_upsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_upsample (j_decompress_ptr cinfo,
+                     jpeg_component_info * compptr, 
+                     JSAMPARRAY input_data,
+                     JSAMPARRAY * output_data_ptr)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor,
+        cinfo->output_width, input_data, output_data_ptr);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor,
+        cinfo->output_width, input_data, output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_upsample (j_decompress_ptr cinfo,
+                     jpeg_component_info * compptr, 
+                     JSAMPARRAY input_data,
+                     JSAMPARRAY * output_data_ptr)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor,
+        cinfo->output_width, input_data, output_data_ptr);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor,
+        cinfo->output_width, input_data, output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_fancy_upsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_fancy_upsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
+                           jpeg_component_info * compptr, 
+                           JSAMPARRAY input_data,
+                           JSAMPARRAY * output_data_ptr)
+{
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
+        compptr->downsampled_width, input_data, output_data_ptr);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor,
+        compptr->downsampled_width, input_data, output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
+                           jpeg_component_info * compptr, 
+                           JSAMPARRAY input_data,
+                           JSAMPARRAY * output_data_ptr)
+{
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
+        compptr->downsampled_width, input_data, output_data_ptr);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor,
+        compptr->downsampled_width, input_data, output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_merged_upsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_merged_upsample (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
+                            JSAMPIMAGE input_buf,
+                            JDIMENSION in_row_group_ctr,
+                            JSAMPARRAY output_buf)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+  void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch(cinfo->out_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_extbgr_merged_upsample_mmx;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_extbgrx_merged_upsample_mmx;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_extxbgr_merged_upsample_mmx;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx;
+      break;
+    default:
+      sse2fct=jsimd_h2v2_merged_upsample_sse2;
+      mmxfct=jsimd_h2v2_merged_upsample_mmx;
+      break;
+  }
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    sse2fct(cinfo->output_width, input_buf,
+        in_row_group_ctr, output_buf);
+  else if (simd_support & JSIMD_MMX)
+    mmxfct(cinfo->output_width, input_buf,
+        in_row_group_ctr, output_buf);
+}
+
+GLOBAL(void)
+jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
+                            JSAMPIMAGE input_buf,
+                            JDIMENSION in_row_group_ctr,
+                            JSAMPARRAY output_buf)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+  void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch(cinfo->out_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_extbgr_merged_upsample_mmx;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_extbgrx_merged_upsample_mmx;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_extxbgr_merged_upsample_mmx;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx;
+      break;
+    default:
+      sse2fct=jsimd_h2v1_merged_upsample_sse2;
+      mmxfct=jsimd_h2v1_merged_upsample_mmx;
+      break;
+  }
+
+  if ((simd_support & JSIMD_SSE2) &&
+      IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    sse2fct(cinfo->output_width, input_buf,
+        in_row_group_ctr, output_buf);
+  else if (simd_support & JSIMD_MMX)
+    mmxfct(cinfo->output_width, input_buf,
+        in_row_group_ctr, output_buf);
+}
+
+GLOBAL(int)
+jsimd_can_convsamp (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_convsamp_float (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_SSE)
+    return 1;
+  if (simd_support & JSIMD_3DNOW)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
+                DCTELEM * workspace)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_convsamp_sse2(sample_data, start_col, workspace);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_convsamp_mmx(sample_data, start_col, workspace);
+}
+
+GLOBAL(void)
+jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
+                      FAST_FLOAT * workspace)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
+  else if (simd_support & JSIMD_SSE)
+    jsimd_convsamp_float_sse(sample_data, start_col, workspace);
+  else if (simd_support & JSIMD_3DNOW)
+    jsimd_convsamp_float_3dnow(sample_data, start_col, workspace);
+}
+
+GLOBAL(int)
+jsimd_can_fdct_islow (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_ifast (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_float (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
+    return 1;
+  if (simd_support & JSIMD_3DNOW)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_fdct_islow (DCTELEM * data)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
+    jsimd_fdct_islow_sse2(data);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_fdct_islow_mmx(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_ifast (DCTELEM * data)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
+    jsimd_fdct_ifast_sse2(data);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_fdct_ifast_mmx(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_float (FAST_FLOAT * data)
+{
+  if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
+    jsimd_fdct_float_sse(data);
+  else if (simd_support & JSIMD_3DNOW)
+    jsimd_fdct_float_3dnow(data);
+}
+
+GLOBAL(int)
+jsimd_can_quantize (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_quantize_float (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_SSE2)
+    return 1;
+  if (simd_support & JSIMD_SSE)
+    return 1;
+  if (simd_support & JSIMD_3DNOW)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
+                DCTELEM * workspace)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_quantize_sse2(coef_block, divisors, workspace);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_quantize_mmx(coef_block, divisors, workspace);
+}
+
+GLOBAL(void)
+jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+                      FAST_FLOAT * workspace)
+{
+  if (simd_support & JSIMD_SSE2)
+    jsimd_quantize_float_sse2(coef_block, divisors, workspace);
+  else if (simd_support & JSIMD_SSE)
+    jsimd_quantize_float_sse(coef_block, divisors, workspace);
+  else if (simd_support & JSIMD_3DNOW)
+    jsimd_quantize_float_3dnow(coef_block, divisors, workspace);
+}
+
+GLOBAL(int)
+jsimd_can_idct_2x2 (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_4x4 (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(int)
+jsimd_can_idct_islow (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_ifast (void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(IFAST_MULT_TYPE) != 2)
+    return 0;
+  if (IFAST_SCALE_BITS != 2)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
+    return 1;
+  if (simd_support & JSIMD_MMX)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_float (void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+  if (sizeof(FLOAT_MULT_TYPE) != 4)
+    return 0;
+
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
+    return 1;
+  if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
+    return 1;
+  if (simd_support & JSIMD_3DNOW)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
+    jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
+    jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+  else if (simd_support & JSIMD_MMX)
+    jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
+    jsimd_idct_float_sse2(compptr->dct_table, coef_block,
+        output_buf, output_col);
+  else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
+    jsimd_idct_float_sse(compptr->dct_table, coef_block,
+        output_buf, output_col);
+  else if (simd_support & JSIMD_3DNOW)
+    jsimd_idct_float_3dnow(compptr->dct_table, coef_block,
+        output_buf, output_col);
+}
+
diff --git a/simd/jsimd_x86_64.c b/simd/jsimd_x86_64.c
new file mode 100644
index 0000000..4bff6c2
--- /dev/null
+++ b/simd/jsimd_x86_64.c
@@ -0,0 +1,680 @@
+/*
+ * jsimd_x86_64.c
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright 2009 D. R. Commander
+ * 
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ *
+ * This file contains the interface between the "normal" portions
+ * of the library and the SIMD implementations when running on a
+ * x86_64 architecture.
+ */
+
+#define JPEG_INTERNALS
+#include "../jinclude.h"
+#include "../jpeglib.h"
+#include "../jsimd.h"
+#include "../jdct.h"
+#include "../jsimddct.h"
+#include "jsimd.h"
+
+/*
+ * In the PIC cases, we have no guarantee that constants will keep
+ * their alignment. This macro allows us to verify it at runtime.
+ */
+#define IS_ALIGNED(ptr, order) (((size_t)ptr & ((1 << order) - 1)) == 0)
+
+#define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
+
+GLOBAL(int)
+jsimd_can_rgb_ycc (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
+                       JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+                       JDIMENSION output_row, int num_rows)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+
+  switch(cinfo->in_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_extrgb_ycc_convert_sse2;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_extrgbx_ycc_convert_sse2;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_extbgr_ycc_convert_sse2;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_extbgrx_ycc_convert_sse2;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_extxbgr_ycc_convert_sse2;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_extxrgb_ycc_convert_sse2;
+      break;
+    default:
+      sse2fct=jsimd_rgb_ycc_convert_sse2;
+      break;
+  }
+
+  sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
+                       JSAMPIMAGE input_buf, JDIMENSION input_row,
+                       JSAMPARRAY output_buf, int num_rows)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
+
+  switch(cinfo->out_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_ycc_extrgb_convert_sse2;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_ycc_extrgbx_convert_sse2;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_ycc_extbgr_convert_sse2;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_ycc_extbgrx_convert_sse2;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_ycc_extxbgr_convert_sse2;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_ycc_extxrgb_convert_sse2;
+      break;
+    default:
+      sse2fct=jsimd_ycc_rgb_convert_sse2;
+      break;
+  }
+
+  sse2fct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_downsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_downsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+                       JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  jsimd_h2v2_downsample_sse2(cinfo->image_width,
+                             cinfo->max_v_samp_factor,
+                             compptr->v_samp_factor,
+                             compptr->width_in_blocks,
+                             input_data, output_data);
+}
+
+GLOBAL(void)
+jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+                       JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  jsimd_h2v1_downsample_sse2(cinfo->image_width,
+                             cinfo->max_v_samp_factor,
+                             compptr->v_samp_factor,
+                             compptr->width_in_blocks,
+                             input_data, output_data);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_upsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_upsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_h2v2_upsample (j_decompress_ptr cinfo,
+                     jpeg_component_info * compptr, 
+                     JSAMPARRAY input_data,
+                     JSAMPARRAY * output_data_ptr)
+{
+  jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor,
+                           cinfo->output_width,
+                           input_data, output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_upsample (j_decompress_ptr cinfo,
+                     jpeg_component_info * compptr, 
+                     JSAMPARRAY input_data,
+                     JSAMPARRAY * output_data_ptr)
+{
+  jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor,
+                           cinfo->output_width,
+                           input_data, output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_fancy_upsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_fancy_upsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
+                           jpeg_component_info * compptr, 
+                           JSAMPARRAY input_data,
+                           JSAMPARRAY * output_data_ptr)
+{
+  jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
+                                 compptr->downsampled_width,
+                                 input_data, output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
+                           jpeg_component_info * compptr, 
+                           JSAMPARRAY input_data,
+                           JSAMPARRAY * output_data_ptr)
+{
+  jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
+                                 compptr->downsampled_width,
+                                 input_data, output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_merged_upsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_merged_upsample (void)
+{
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
+                            JSAMPIMAGE input_buf,
+                            JDIMENSION in_row_group_ctr,
+                            JSAMPARRAY output_buf)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch(cinfo->out_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2;
+      break;
+    default:
+      sse2fct=jsimd_h2v2_merged_upsample_sse2;
+      break;
+  }
+
+  sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+}
+
+GLOBAL(void)
+jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
+                            JSAMPIMAGE input_buf,
+                            JDIMENSION in_row_group_ctr,
+                            JSAMPARRAY output_buf)
+{
+  void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch(cinfo->out_color_space)
+  {
+    case JCS_EXT_RGB:
+      sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
+      break;
+    case JCS_EXT_RGBX:
+      sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
+      break;
+    case JCS_EXT_BGR:
+      sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2;
+      break;
+    case JCS_EXT_BGRX:
+      sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2;
+      break;
+    case JCS_EXT_XBGR:
+      sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2;
+      break;
+    case JCS_EXT_XRGB:
+      sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2;
+      break;
+    default:
+      sse2fct=jsimd_h2v1_merged_upsample_sse2;
+      break;
+  }
+
+  sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+}
+
+GLOBAL(int)
+jsimd_can_convsamp (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_convsamp_float (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
+                DCTELEM * workspace)
+{
+  jsimd_convsamp_sse2(sample_data, start_col, workspace);
+}
+
+GLOBAL(void)
+jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
+                      FAST_FLOAT * workspace)
+{
+  jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
+}
+
+GLOBAL(int)
+jsimd_can_fdct_islow (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_ifast (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_float (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_fdct_float_sse))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_fdct_islow (DCTELEM * data)
+{
+  jsimd_fdct_islow_sse2(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_ifast (DCTELEM * data)
+{
+  jsimd_fdct_ifast_sse2(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_float (FAST_FLOAT * data)
+{
+  jsimd_fdct_float_sse(data);
+}
+
+GLOBAL(int)
+jsimd_can_quantize (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_quantize_float (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
+                DCTELEM * workspace)
+{
+  jsimd_quantize_sse2(coef_block, divisors, workspace);
+}
+
+GLOBAL(void)
+jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+                      FAST_FLOAT * workspace)
+{
+  jsimd_quantize_float_sse2(coef_block, divisors, workspace);
+}
+
+GLOBAL(int)
+jsimd_can_idct_2x2 (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_idct_4x4 (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_idct_red_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(int)
+jsimd_can_idct_islow (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_idct_islow_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_idct_ifast (void)
+{
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(IFAST_MULT_TYPE) != 2)
+    return 0;
+  if (IFAST_SCALE_BITS != 2)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(int)
+jsimd_can_idct_float (void)
+{
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(FAST_FLOAT) != 4)
+    return 0;
+  if (sizeof(FLOAT_MULT_TYPE) != 4)
+    return 0;
+
+  if (!IS_ALIGNED_SSE(jconst_idct_float_sse2))
+    return 0;
+
+  return 1;
+}
+
+GLOBAL(void)
+jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+                JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                JDIMENSION output_col)
+{
+  jsimd_idct_float_sse2(compptr->dct_table, coef_block,
+                        output_buf, output_col);
+}
+
diff --git a/simd/jsimdcfg.inc.h b/simd/jsimdcfg.inc.h
new file mode 100644
index 0000000..4876038
--- /dev/null
+++ b/simd/jsimdcfg.inc.h
@@ -0,0 +1,168 @@
+// This file generates the include file for the assembly
+// implementations by abusing the C preprocessor.
+//
+// Note: Some things are manually defined as they need to
+// be mapped to NASM types.
+
+;
+; Automatically generated include file from jsimdcfg.inc.h
+;
+
+#define JPEG_INTERNALS
+
+#include "../jpeglib.h"
+#include "../jconfig.h"
+#include "../jmorecfg.h"
+#include "jsimd.h"
+
+#define define(var) %define _cpp_protection_##var
+#define definev(var) %define _cpp_protection_##var var
+
+;
+; -- jpeglib.h
+;
+
+definev(DCTSIZE)
+definev(DCTSIZE2)
+
+;
+; -- jmorecfg.h
+;
+
+definev(RGB_RED)
+definev(RGB_GREEN)
+definev(RGB_BLUE)
+
+definev(RGB_PIXELSIZE)
+
+; Representation of a single sample (pixel element value).
+; On this SIMD implementation, this must be 'unsigned char'.
+;
+
+%define JSAMPLE                 byte          ; unsigned char
+%define SIZEOF_JSAMPLE          SIZEOF_BYTE   ; sizeof(JSAMPLE)
+
+definev(CENTERJSAMPLE)
+
+; Representation of a DCT frequency coefficient.
+; On this SIMD implementation, this must be 'short'.
+;
+%define JCOEF                   word          ; short
+%define SIZEOF_JCOEF            SIZEOF_WORD   ; sizeof(JCOEF)
+
+; Datatype used for image dimensions.
+; On this SIMD implementation, this must be 'unsigned int'.
+;
+%define JDIMENSION              dword         ; unsigned int
+%define SIZEOF_JDIMENSION       SIZEOF_DWORD  ; sizeof(JDIMENSION)
+
+%define JSAMPROW                POINTER       ; JSAMPLE FAR * (jpeglib.h)
+%define JSAMPARRAY              POINTER       ; JSAMPROW *    (jpeglib.h)
+%define JSAMPIMAGE              POINTER       ; JSAMPARRAY *  (jpeglib.h)
+%define JCOEFPTR                POINTER       ; JCOEF FAR *   (jpeglib.h)
+%define SIZEOF_JSAMPROW         SIZEOF_POINTER  ; sizeof(JSAMPROW)
+%define SIZEOF_JSAMPARRAY       SIZEOF_POINTER  ; sizeof(JSAMPARRAY)
+%define SIZEOF_JSAMPIMAGE       SIZEOF_POINTER  ; sizeof(JSAMPIMAGE)
+%define SIZEOF_JCOEFPTR         SIZEOF_POINTER  ; sizeof(JCOEFPTR)
+
+;
+; -- jdct.h
+;
+
+; A forward DCT routine is given a pointer to a work area of type DCTELEM[];
+; the DCT is to be performed in-place in that buffer.
+; To maximize parallelism, Type DCTELEM is changed to short (originally, int).
+;
+%define DCTELEM                 word          ; short
+%define SIZEOF_DCTELEM          SIZEOF_WORD   ; sizeof(DCTELEM)
+
+%define FAST_FLOAT              FP32            ; float
+%define SIZEOF_FAST_FLOAT       SIZEOF_FP32     ; sizeof(FAST_FLOAT)
+
+; To maximize parallelism, Type MULTIPLIER is changed to short.
+;
+%define ISLOW_MULT_TYPE         word          ; must be short
+%define SIZEOF_ISLOW_MULT_TYPE  SIZEOF_WORD   ; sizeof(ISLOW_MULT_TYPE)
+
+%define IFAST_MULT_TYPE         word          ; must be short
+%define SIZEOF_IFAST_MULT_TYPE  SIZEOF_WORD   ; sizeof(IFAST_MULT_TYPE)
+%define IFAST_SCALE_BITS        2             ; fractional bits in scale factors
+
+%define FLOAT_MULT_TYPE         FP32          ; must be float
+%define SIZEOF_FLOAT_MULT_TYPE  SIZEOF_FP32   ; sizeof(FLOAT_MULT_TYPE)
+
+;
+; -- jsimd.h
+;
+
+definev(JSIMD_NONE)
+definev(JSIMD_MMX)
+definev(JSIMD_3DNOW)
+definev(JSIMD_SSE)
+definev(JSIMD_SSE2)
+
+; Short forms of external names for systems with brain-damaged linkers.
+;
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+definev(jpeg_simd_cpu_support)
+definev(jsimd_rgb_ycc_convert_mmx)
+definev(jsimd_ycc_rgb_convert_mmx)
+definev(jconst_rgb_ycc_convert_sse2)
+definev(jsimd_rgb_ycc_convert_sse2)
+definev(jconst_ycc_rgb_convert_sse2)
+definev(jsimd_ycc_rgb_convert_sse2)
+definev(jsimd_h2v2_downsample_mmx)
+definev(jsimd_h2v1_downsample_mmx)
+definev(jsimd_h2v2_downsample_sse2)
+definev(jsimd_h2v1_downsample_sse2)
+definev(jsimd_h2v2_upsample_mmx)
+definev(jsimd_h2v1_upsample_mmx)
+definev(jsimd_h2v1_fancy_upsample_mmx)
+definev(jsimd_h2v2_fancy_upsample_mmx)
+definev(jsimd_h2v1_merged_upsample_mmx)
+definev(jsimd_h2v2_merged_upsample_mmx)
+definev(jsimd_h2v2_upsample_sse2)
+definev(jsimd_h2v1_upsample_sse2)
+definev(jconst_fancy_upsample_sse2)
+definev(jsimd_h2v1_fancy_upsample_sse2)
+definev(jsimd_h2v2_fancy_upsample_sse2)
+definev(jconst_merged_upsample_sse2)
+definev(jsimd_h2v1_merged_upsample_sse2)
+definev(jsimd_h2v2_merged_upsample_sse2)
+definev(jsimd_convsamp_mmx)
+definev(jsimd_convsamp_sse2)
+definev(jsimd_convsamp_float_3dnow)
+definev(jsimd_convsamp_float_sse)
+definev(jsimd_convsamp_float_sse2)
+definev(jsimd_fdct_islow_mmx)
+definev(jsimd_fdct_ifast_mmx)
+definev(jconst_fdct_islow_sse2)
+definev(jsimd_fdct_islow_sse2)
+definev(jconst_fdct_ifast_sse2)
+definev(jsimd_fdct_ifast_sse2)
+definev(jsimd_fdct_float_3dnow)
+definev(jconst_fdct_float_sse)
+definev(jsimd_fdct_float_sse)
+definev(jsimd_quantize_mmx)
+definev(jsimd_quantize_sse2)
+definev(jsimd_quantize_float_3dnow)
+definev(jsimd_quantize_float_sse)
+definev(jsimd_quantize_float_sse2)
+definev(jsimd_idct_2x2_mmx)
+definev(jsimd_idct_4x4_mmx)
+definev(jconst_idct_red_sse2)
+definev(jsimd_idct_2x2_sse2)
+definev(jsimd_idct_4x4_sse2)
+definev(jsimd_idct_islow_mmx)
+definev(jsimd_idct_ifast_mmx)
+definev(jconst_idct_islow_sse2)
+definev(jsimd_idct_islow_sse2)
+definev(jconst_idct_ifast_sse2)
+definev(jsimd_idct_ifast_sse2)
+definev(jsimd_idct_float_3dnow)
+definev(jconst_idct_float_sse)
+definev(jsimd_idct_float_sse)
+definev(jconst_idct_float_sse2)
+definev(jsimd_idct_float_sse2)
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
diff --git a/simd/jsimdcpu.asm b/simd/jsimdcpu.asm
new file mode 100644
index 0000000..bdbcc23
--- /dev/null
+++ b/simd/jsimdcpu.asm
@@ -0,0 +1,105 @@
+;
+; jsimdcpu.asm - SIMD instruction support check
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Check if the CPU supports SIMD instructions
+;
+; GLOBAL(unsigned int)
+; jpeg_simd_cpu_support (void)
+;
+
+	align	16
+	global	EXTN(jpeg_simd_cpu_support)
+
+EXTN(jpeg_simd_cpu_support):
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+	push	edi
+
+	xor	edi,edi			; simd support flag
+
+	pushfd
+	pop	eax
+	mov	edx,eax
+	xor	eax, 1<<21		; flip ID bit in EFLAGS
+	push	eax
+	popfd
+	pushfd
+	pop	eax
+	xor	eax,edx
+	jz	short .return		; CPUID is not supported
+
+	; Check for MMX instruction support
+	xor	eax,eax
+	cpuid
+	test	eax,eax
+	jz	short .return
+
+	xor	eax,eax
+	inc	eax
+	cpuid
+	mov	eax,edx			; eax = Standard feature flags
+
+	test	eax, 1<<23		; bit23:MMX
+	jz	short .no_mmx
+	or	edi, byte JSIMD_MMX
+.no_mmx:
+	test	eax, 1<<25		; bit25:SSE
+	jz	short .no_sse
+	or	edi, byte JSIMD_SSE
+.no_sse:
+	test	eax, 1<<26		; bit26:SSE2
+	jz	short .no_sse2
+	or	edi, byte JSIMD_SSE2
+.no_sse2:
+
+	; Check for 3DNow! instruction support
+	mov	eax, 0x80000000
+	cpuid
+	cmp	eax, 0x80000000
+	jbe	short .return
+
+	mov	eax, 0x80000001
+	cpuid
+	mov	eax,edx			; eax = Extended feature flags
+
+	test	eax, 1<<31		; bit31:3DNow!(vendor independent)
+	jz	short .no_3dnow
+	or	edi, byte JSIMD_3DNOW
+.no_3dnow:
+
+.return:
+	mov	eax,edi
+
+	pop	edi
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+	align	16
diff --git a/simd/jsimdext.inc b/simd/jsimdext.inc
new file mode 100644
index 0000000..4ea3d17
--- /dev/null
+++ b/simd/jsimdext.inc
@@ -0,0 +1,372 @@
+;
+; jsimdext.inc - common declarations
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright 2010 D. R. Commander
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library - version 1.02
+;
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+;
+; This software is provided 'as-is', without any express or implied
+; warranty.  In no event will the authors be held liable for any damages
+; arising from the use of this software.
+;
+; Permission is granted to anyone to use this software for any purpose,
+; including commercial applications, and to alter it and redistribute it
+; freely, subject to the following restrictions:
+;
+; 1. The origin of this software must not be misrepresented; you must not
+;    claim that you wrote the original software. If you use this software
+;    in a product, an acknowledgment in the product documentation would be
+;    appreciated but is not required.
+; 2. Altered source versions must be plainly marked as such, and must not be
+;    misrepresented as being the original software.
+; 3. This notice may not be removed or altered from any source distribution.
+;
+; [TAB8]
+
+; ==========================================================================
+;  System-dependent configurations
+
+%ifdef WIN32	; ----(nasm -fwin32 -DWIN32 ...)--------
+; * Microsoft Visual C++
+; * MinGW (Minimalist GNU for Windows)
+; * CygWin
+; * LCC-Win32
+
+; -- segment definition --
+;
+%define SEG_TEXT    .text  align=16 public use32 class=CODE
+%define SEG_CONST   .rdata align=16 public use32 class=CONST
+
+%elifdef WIN64	; ----(nasm -fwin64 -DWIN64 ...)--------
+; * Microsoft Visual C++
+
+; -- segment definition --
+;
+%define SEG_TEXT    .text  align=16 public use64 class=CODE
+%define SEG_CONST   .rdata align=16 public use64 class=CONST
+%ifdef MSVC
+%define EXTN(name)  name			; foo() -> foo
+%endif
+
+%elifdef OBJ32	; ----(nasm -fobj -DOBJ32 ...)----------
+; * Borland C++ (Win32)
+
+; -- segment definition --
+;
+%define SEG_TEXT    .text  align=16 public use32 class=CODE
+%define SEG_CONST   .data  align=16 public use32 class=DATA
+
+%elifdef ELF	; ----(nasm -felf[64] -DELF ...)------------
+; * Linux
+; * *BSD family Unix using elf format
+; * Unix System V, including Solaris x86, UnixWare and SCO Unix
+
+; mark stack as non-executable
+section .note.GNU-stack noalloc noexec nowrite progbits
+
+; -- segment definition --
+;
+%ifdef __x86_64__
+%define SEG_TEXT    .text   progbits align=16
+%define SEG_CONST   .rodata progbits align=16
+%else
+%define SEG_TEXT    .text   progbits alloc exec   nowrite align=16
+%define SEG_CONST   .rodata progbits alloc noexec nowrite align=16
+%endif
+
+; To make the code position-independent, append -DPIC to the commandline
+;
+%define GOT_SYMBOL  _GLOBAL_OFFSET_TABLE_	; ELF supports PIC
+%define EXTN(name)  name			; foo() -> foo
+
+%elifdef AOUT	; ----(nasm -faoutb/aout -DAOUT ...)----
+; * Older Linux using a.out format  (nasm -f aout -DAOUT ...)
+; * *BSD family Unix using a.out format  (nasm -f aoutb -DAOUT ...)
+
+; -- segment definition --
+;
+%define SEG_TEXT    .text
+%define SEG_CONST   .data
+
+; To make the code position-independent, append -DPIC to the commandline
+;
+%define GOT_SYMBOL  __GLOBAL_OFFSET_TABLE_	; BSD-style a.out supports PIC
+
+%elifdef MACHO	; ----(nasm -fmacho -DMACHO ...)--------
+; * NeXTstep/OpenStep/Rhapsody/Darwin/MacOS X (Mach-O format)
+
+; -- segment definition --
+;
+%define SEG_TEXT    .text  ;align=16	; nasm doesn't accept align=16. why?
+%define SEG_CONST   .rodata align=16
+
+; The generation of position-independent code (PIC) is the default on Darwin.
+;
+%define PIC
+%define GOT_SYMBOL  _MACHO_PIC_		; Mach-O style code-relative addressing
+
+%else		; ----(Other case)----------------------
+
+; -- segment definition --
+;
+%define SEG_TEXT    .text
+%define SEG_CONST   .data
+
+%endif	; ----------------------------------------------
+
+; ==========================================================================
+
+; --------------------------------------------------------------------------
+;  Common types
+;
+%ifdef __x86_64__
+%define POINTER                 qword           ; general pointer type
+%define SIZEOF_POINTER          SIZEOF_QWORD    ; sizeof(POINTER)
+%define POINTER_BIT             QWORD_BIT       ; sizeof(POINTER)*BYTE_BIT
+%else
+%define POINTER                 dword           ; general pointer type
+%define SIZEOF_POINTER          SIZEOF_DWORD    ; sizeof(POINTER)
+%define POINTER_BIT             DWORD_BIT       ; sizeof(POINTER)*BYTE_BIT
+%endif
+
+%define INT                     dword           ; signed integer type
+%define SIZEOF_INT              SIZEOF_DWORD    ; sizeof(INT)
+%define INT_BIT                 DWORD_BIT       ; sizeof(INT)*BYTE_BIT
+
+%define FP32                    dword           ; IEEE754 single
+%define SIZEOF_FP32             SIZEOF_DWORD    ; sizeof(FP32)
+%define FP32_BIT                DWORD_BIT       ; sizeof(FP32)*BYTE_BIT
+
+%define MMWORD                  qword           ; int64  (MMX register)
+%define SIZEOF_MMWORD           SIZEOF_QWORD    ; sizeof(MMWORD)
+%define MMWORD_BIT              QWORD_BIT       ; sizeof(MMWORD)*BYTE_BIT
+
+; NASM is buggy and doesn't properly handle operand sizes for SSE
+; instructions, so for now we have to define XMMWORD as blank.
+%define XMMWORD                                 ; int128 (SSE register)
+%define SIZEOF_XMMWORD          SIZEOF_OWORD    ; sizeof(XMMWORD)
+%define XMMWORD_BIT             OWORD_BIT       ; sizeof(XMMWORD)*BYTE_BIT
+
+; Similar hacks for when we load a dword or MMWORD into an xmm# register
+%define XMM_DWORD
+%define XMM_MMWORD
+
+%define SIZEOF_BYTE             1               ; sizeof(BYTE)
+%define SIZEOF_WORD             2               ; sizeof(WORD)
+%define SIZEOF_DWORD            4               ; sizeof(DWORD)
+%define SIZEOF_QWORD            8               ; sizeof(QWORD)
+%define SIZEOF_OWORD            16              ; sizeof(OWORD)
+
+%define BYTE_BIT                8               ; CHAR_BIT in C
+%define WORD_BIT                16              ; sizeof(WORD)*BYTE_BIT
+%define DWORD_BIT               32              ; sizeof(DWORD)*BYTE_BIT
+%define QWORD_BIT               64              ; sizeof(QWORD)*BYTE_BIT
+%define OWORD_BIT               128             ; sizeof(OWORD)*BYTE_BIT
+
+; --------------------------------------------------------------------------
+;  External Symbol Name
+;
+%ifndef EXTN
+%define EXTN(name)   _ %+ name		; foo() -> _foo
+%endif
+
+; --------------------------------------------------------------------------
+;  Macros for position-independent code (PIC) support
+;
+%ifndef GOT_SYMBOL
+%undef PIC
+%endif
+
+%ifdef PIC ; -------------------------------------------
+
+%ifidn GOT_SYMBOL,_MACHO_PIC_ ; --------------------
+
+; At present, nasm doesn't seem to support PIC generation for Mach-O.
+; The PIC support code below is a little tricky.
+
+	SECTION	SEG_CONST
+const_base:
+
+%define GOTOFF(got,sym) (got) + (sym) - const_base
+
+%imacro get_GOT	1
+	; NOTE: this macro destroys ecx resister.
+	call	%%geteip
+	add	ecx, byte (%%ref - $)
+	jmp	short %%adjust
+%%geteip:
+	mov	ecx, POINTER [esp]
+	ret
+%%adjust:
+	push	ebp
+	xor	ebp,ebp		; ebp = 0
+%ifidni %1,ebx	; (%1 == ebx)
+	; db 0x8D,0x9C + jmp near const_base =
+	;   lea ebx, [ecx+ebp*8+(const_base-%%ref)] ; 8D,9C,E9,(offset32)
+	db	0x8D,0x9C		; 8D,9C
+	jmp	near const_base		; E9,(const_base-%%ref)
+%%ref:
+%else  ; (%1 != ebx)
+	; db 0x8D,0x8C + jmp near const_base =
+	;   lea ecx, [ecx+ebp*8+(const_base-%%ref)] ; 8D,8C,E9,(offset32)
+	db	0x8D,0x8C		; 8D,8C
+	jmp	near const_base		; E9,(const_base-%%ref)
+%%ref:	mov	%1, ecx
+%endif ; (%1 == ebx)
+	pop	ebp
+%endmacro
+
+%else	; GOT_SYMBOL != _MACHO_PIC_ ----------------
+
+%define GOTOFF(got,sym) (got) + (sym) wrt ..gotoff
+
+%imacro get_GOT	1
+	extern	GOT_SYMBOL
+	call	%%geteip
+	add	%1, GOT_SYMBOL + $$ - $ wrt ..gotpc
+	jmp	short %%done
+%%geteip:
+	mov	%1, POINTER [esp]
+	ret
+%%done:
+%endmacro
+
+%endif	; GOT_SYMBOL == _MACHO_PIC_ ----------------
+
+%imacro pushpic	1.nolist
+	push	%1
+%endmacro
+%imacro poppic	1.nolist
+	pop	%1
+%endmacro
+%imacro movpic	2.nolist
+	mov	%1,%2
+%endmacro
+
+%else	; !PIC -----------------------------------------
+
+%define GOTOFF(got,sym) (sym)
+
+%imacro get_GOT	1.nolist
+%endmacro
+%imacro pushpic	1.nolist
+%endmacro
+%imacro poppic	1.nolist
+%endmacro
+%imacro movpic	2.nolist
+%endmacro
+
+%endif	;  PIC -----------------------------------------
+
+; --------------------------------------------------------------------------
+;  Align the next instruction on {2,4,8,16,..}-byte boundary.
+;  ".balign n,,m" in GNU as
+;
+%define MSKLE(x,y)  (~(((y) & 0xFFFF) - ((x) & 0xFFFF)) >> 16)
+%define FILLB(b,n)  (($$-(b)) & ((n)-1))
+
+%imacro alignx 1-2.nolist 0xFFFF
+%%bs:	times MSKLE(FILLB(%%bs,%1),%2) & MSKLE(16,FILLB($,%1)) & FILLB($,%1) \
+	       db 0x90                               ; nop
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/9 \
+	       db 0x8D,0x9C,0x23,0x00,0x00,0x00,0x00 ; lea ebx,[ebx+0x00000000]
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/7 \
+	       db 0x8D,0xAC,0x25,0x00,0x00,0x00,0x00 ; lea ebp,[ebp+0x00000000]
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/6 \
+	       db 0x8D,0xAD,0x00,0x00,0x00,0x00      ; lea ebp,[ebp+0x00000000]
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/4 \
+	       db 0x8D,0x6C,0x25,0x00                ; lea ebp,[ebp+0x00]
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/3 \
+	       db 0x8D,0x6D,0x00                     ; lea ebp,[ebp+0x00]
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/2 \
+	       db 0x8B,0xED                          ; mov ebp,ebp
+	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/1 \
+	       db 0x90                               ; nop
+%endmacro
+
+; Align the next data on {2,4,8,16,..}-byte boundary.
+;
+%imacro alignz 1.nolist
+	align %1, db 0		; filling zeros
+%endmacro
+
+%ifdef __x86_64__
+
+%ifdef WIN64
+
+%imacro collect_args 0
+	push r10
+	push r11
+	push r12
+	push r13
+	push r14
+	push r15
+	mov r10, rcx
+	mov r11, rdx
+	mov r12, r8
+	mov r13, r9
+	mov r14, [rax+48]
+	mov r15, [rax+56]
+	push rsi
+	push rdi
+	sub     rsp, SIZEOF_XMMWORD
+	movlpd  XMMWORD [rsp], xmm6
+	sub     rsp, SIZEOF_XMMWORD
+	movlpd  XMMWORD [rsp], xmm7
+%endmacro
+
+%imacro uncollect_args 0
+	movlpd  xmm7, XMMWORD [rsp]
+	add     rsp, SIZEOF_XMMWORD
+	movlpd  xmm6, XMMWORD [rsp]
+	add     rsp, SIZEOF_XMMWORD
+	pop rdi
+	pop rsi
+	pop r15
+	pop r14
+	pop r13
+	pop r12
+	pop r11
+	pop r10
+%endmacro
+
+%else
+
+%imacro collect_args 0
+	push r10
+	push r11
+	push r12
+	push r13
+	push r14
+	push r15
+	mov r10, rdi
+	mov r11, rsi
+	mov r12, rdx
+	mov r13, rcx
+	mov r14, r8
+	mov r15, r9
+%endmacro
+
+%imacro uncollect_args 0
+	pop r15
+	pop r14
+	pop r13
+	pop r12
+	pop r11
+	pop r10
+%endmacro
+
+%endif
+
+%endif
+
+; --------------------------------------------------------------------------
+;  Defines picked up from the C headers
+;
+%include "jsimdcfg.inc"
+
+; --------------------------------------------------------------------------
diff --git a/simd/nasm_lt.sh b/simd/nasm_lt.sh
new file mode 100755
index 0000000..6cd7329
--- /dev/null
+++ b/simd/nasm_lt.sh
@@ -0,0 +1,57 @@
+#! /bin/sh
+command=""
+infile=""
+o_opt=no
+pic=no
+while [ $# -gt 0 ]; do
+    case "$1" in
+        -DPIC|-fPIC|-fpic|-Kpic|-KPIC)
+            if [ "$pic" != "yes" ] ; then
+                command="$command -DPIC"
+                pic=yes
+            fi
+            ;;
+        -f|-fbin|-faout|-faoutb|-fcoff|-felf|-felf64|-fas86| \
+        -fobj|-fwin32|-fwin64|-frdf|-fieee|-fmacho|-fmacho64)
+            # it's a file format specifier for nasm.
+            command="$command $1"
+            ;;
+        -f*)
+            # maybe a code-generation flag for gcc.
+            ;;
+        -[Ii]*)
+            incdir=`echo "$1" | sed 's/^-[Ii]//'`
+            if [ "x$incdir" = x -a "x$2" != x ] ; then
+                case "$2" in
+                    -*) ;;
+                    *) incdir="$2"; shift;;
+                esac
+            fi
+            if [ "x$incdir" != x ] ; then
+                # In the case of NASM, the trailing slash is necessary.
+                incdir=`echo "$incdir" | sed 's%/*$%/%'`
+                command="$command -I$incdir"
+            fi
+            ;;
+        -o*)
+            o_opt=yes
+            command="$command $1"
+            ;;
+        *.asm)
+            infile=$1
+            command="$command $1"
+            ;;
+        *)
+            command="$command $1"
+            ;;
+    esac
+    shift
+done
+if [ "$o_opt" != yes ] ; then
+    # By default, NASM creates an output file
+    # in the same directory as the input file.
+    outfile="-o `echo $infile | sed -e 's%^.*/%%' -e 's%\.[^.]*$%%'`.o"
+    command="$command $outfile"
+fi
+echo $command
+exec $command
diff --git a/structure.txt b/structure.txt
new file mode 100644
index 0000000..09e1cd2
--- /dev/null
+++ b/structure.txt
@@ -0,0 +1,949 @@
+IJG JPEG LIBRARY:  SYSTEM ARCHITECTURE
+
+Copyright (C) 1991-2009, Thomas G. Lane, Guido Vollbeding.
+This file is part of the Independent JPEG Group's software.
+For conditions of distribution and use, see the accompanying README file.
+
+
+This file provides an overview of the architecture of the IJG JPEG software;
+that is, the functions of the various modules in the system and the interfaces
+between modules.  For more precise details about any data structure or calling
+convention, see the include files and comments in the source code.
+
+We assume that the reader is already somewhat familiar with the JPEG standard.
+The README file includes references for learning about JPEG.  The file
+libjpeg.txt describes the library from the viewpoint of an application
+programmer using the library; it's best to read that file before this one.
+Also, the file coderules.txt describes the coding style conventions we use.
+
+In this document, JPEG-specific terminology follows the JPEG standard:
+  A "component" means a color channel, e.g., Red or Luminance.
+  A "sample" is a single component value (i.e., one number in the image data).
+  A "coefficient" is a frequency coefficient (a DCT transform output number).
+  A "block" is an 8x8 group of samples or coefficients.
+  An "MCU" (minimum coded unit) is an interleaved set of blocks of size
+	determined by the sampling factors, or a single block in a
+	noninterleaved scan.
+We do not use the terms "pixel" and "sample" interchangeably.  When we say
+pixel, we mean an element of the full-size image, while a sample is an element
+of the downsampled image.  Thus the number of samples may vary across
+components while the number of pixels does not.  (This terminology is not used
+rigorously throughout the code, but it is used in places where confusion would
+otherwise result.)
+
+
+*** System features ***
+
+The IJG distribution contains two parts:
+  * A subroutine library for JPEG compression and decompression.
+  * cjpeg/djpeg, two sample applications that use the library to transform
+    JFIF JPEG files to and from several other image formats.
+cjpeg/djpeg are of no great intellectual complexity: they merely add a simple
+command-line user interface and I/O routines for several uncompressed image
+formats.  This document concentrates on the library itself.
+
+We desire the library to be capable of supporting all JPEG baseline, extended
+sequential, and progressive DCT processes.  Hierarchical processes are not
+supported.
+
+The library does not support the lossless (spatial) JPEG process.  Lossless
+JPEG shares little or no code with lossy JPEG, and would normally be used
+without the extensive pre- and post-processing provided by this library.
+We feel that lossless JPEG is better handled by a separate library.
+
+Within these limits, any set of compression parameters allowed by the JPEG
+spec should be readable for decompression.  (We can be more restrictive about
+what formats we can generate.)  Although the system design allows for all
+parameter values, some uncommon settings are not yet implemented and may
+never be; nonintegral sampling ratios are the prime example.  Furthermore,
+we treat 8-bit vs. 12-bit data precision as a compile-time switch, not a
+run-time option, because most machines can store 8-bit pixels much more
+compactly than 12-bit.
+
+For legal reasons, JPEG arithmetic coding is not currently supported, but
+extending the library to include it would be straightforward.
+
+By itself, the library handles only interchange JPEG datastreams --- in
+particular the widely used JFIF file format.  The library can be used by
+surrounding code to process interchange or abbreviated JPEG datastreams that
+are embedded in more complex file formats.  (For example, libtiff uses this
+library to implement JPEG compression within the TIFF file format.)
+
+The library includes a substantial amount of code that is not covered by the
+JPEG standard but is necessary for typical applications of JPEG.  These
+functions preprocess the image before JPEG compression or postprocess it after
+decompression.  They include colorspace conversion, downsampling/upsampling,
+and color quantization.  This code can be omitted if not needed.
+
+A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
+and even more so in decompression postprocessing.  The decompression library
+provides multiple implementations that cover most of the useful tradeoffs,
+ranging from very-high-quality down to fast-preview operation.  On the
+compression side we have generally not provided low-quality choices, since
+compression is normally less time-critical.  It should be understood that the
+low-quality modes may not meet the JPEG standard's accuracy requirements;
+nonetheless, they are useful for viewers.
+
+
+*** Portability issues ***
+
+Portability is an essential requirement for the library.  The key portability
+issues that show up at the level of system architecture are:
+
+1.  Memory usage.  We want the code to be able to run on PC-class machines
+with limited memory.  Images should therefore be processed sequentially (in
+strips), to avoid holding the whole image in memory at once.  Where a
+full-image buffer is necessary, we should be able to use either virtual memory
+or temporary files.
+
+2.  Near/far pointer distinction.  To run efficiently on 80x86 machines, the
+code should distinguish "small" objects (kept in near data space) from
+"large" ones (kept in far data space).  This is an annoying restriction, but
+fortunately it does not impact code quality for less brain-damaged machines,
+and the source code clutter turns out to be minimal with sufficient use of
+pointer typedefs.
+
+3. Data precision.  We assume that "char" is at least 8 bits, "short" and
+"int" at least 16, "long" at least 32.  The code will work fine with larger
+data sizes, although memory may be used inefficiently in some cases.  However,
+the JPEG compressed datastream must ultimately appear on external storage as a
+sequence of 8-bit bytes if it is to conform to the standard.  This may pose a
+problem on machines where char is wider than 8 bits.  The library represents
+compressed data as an array of values of typedef JOCTET.  If no data type
+exactly 8 bits wide is available, custom data source and data destination
+modules must be written to unpack and pack the chosen JOCTET datatype into
+8-bit external representation.
+
+
+*** System overview ***
+
+The compressor and decompressor are each divided into two main sections:
+the JPEG compressor or decompressor proper, and the preprocessing or
+postprocessing functions.  The interface between these two sections is the
+image data that the official JPEG spec regards as its input or output: this
+data is in the colorspace to be used for compression, and it is downsampled
+to the sampling factors to be used.  The preprocessing and postprocessing
+steps are responsible for converting a normal image representation to or from
+this form.  (Those few applications that want to deal with YCbCr downsampled
+data can skip the preprocessing or postprocessing step.)
+
+Looking more closely, the compressor library contains the following main
+elements:
+
+  Preprocessing:
+    * Color space conversion (e.g., RGB to YCbCr).
+    * Edge expansion and downsampling.  Optionally, this step can do simple
+      smoothing --- this is often helpful for low-quality source data.
+  JPEG proper:
+    * MCU assembly, DCT, quantization.
+    * Entropy coding (sequential or progressive, Huffman or arithmetic).
+
+In addition to these modules we need overall control, marker generation,
+and support code (memory management & error handling).  There is also a
+module responsible for physically writing the output data --- typically
+this is just an interface to fwrite(), but some applications may need to
+do something else with the data.
+
+The decompressor library contains the following main elements:
+
+  JPEG proper:
+    * Entropy decoding (sequential or progressive, Huffman or arithmetic).
+    * Dequantization, inverse DCT, MCU disassembly.
+  Postprocessing:
+    * Upsampling.  Optionally, this step may be able to do more general
+      rescaling of the image.
+    * Color space conversion (e.g., YCbCr to RGB).  This step may also
+      provide gamma adjustment [ currently it does not ].
+    * Optional color quantization (e.g., reduction to 256 colors).
+    * Optional color precision reduction (e.g., 24-bit to 15-bit color).
+      [This feature is not currently implemented.]
+
+We also need overall control, marker parsing, and a data source module.
+The support code (memory management & error handling) can be shared with
+the compression half of the library.
+
+There may be several implementations of each of these elements, particularly
+in the decompressor, where a wide range of speed/quality tradeoffs is very
+useful.  It must be understood that some of the best speedups involve
+merging adjacent steps in the pipeline.  For example, upsampling, color space
+conversion, and color quantization might all be done at once when using a
+low-quality ordered-dither technique.  The system architecture is designed to
+allow such merging where appropriate.
+
+
+Note: it is convenient to regard edge expansion (padding to block boundaries)
+as a preprocessing/postprocessing function, even though the JPEG spec includes
+it in compression/decompression.  We do this because downsampling/upsampling
+can be simplified a little if they work on padded data: it's not necessary to
+have special cases at the right and bottom edges.  Therefore the interface
+buffer is always an integral number of blocks wide and high, and we expect
+compression preprocessing to pad the source data properly.  Padding will occur
+only to the next block (8-sample) boundary.  In an interleaved-scan situation,
+additional dummy blocks may be used to fill out MCUs, but the MCU assembly and
+disassembly logic will create or discard these blocks internally.  (This is
+advantageous for speed reasons, since we avoid DCTing the dummy blocks.
+It also permits a small reduction in file size, because the compressor can
+choose dummy block contents so as to minimize their size in compressed form.
+Finally, it makes the interface buffer specification independent of whether
+the file is actually interleaved or not.)  Applications that wish to deal
+directly with the downsampled data must provide similar buffering and padding
+for odd-sized images.
+
+
+*** Poor man's object-oriented programming ***
+
+It should be clear by now that we have a lot of quasi-independent processing
+steps, many of which have several possible behaviors.  To avoid cluttering the
+code with lots of switch statements, we use a simple form of object-style
+programming to separate out the different possibilities.
+
+For example, two different color quantization algorithms could be implemented
+as two separate modules that present the same external interface; at runtime,
+the calling code will access the proper module indirectly through an "object".
+
+We can get the limited features we need while staying within portable C.
+The basic tool is a function pointer.  An "object" is just a struct
+containing one or more function pointer fields, each of which corresponds to
+a method name in real object-oriented languages.  During initialization we
+fill in the function pointers with references to whichever module we have
+determined we need to use in this run.  Then invocation of the module is done
+by indirecting through a function pointer; on most machines this is no more
+expensive than a switch statement, which would be the only other way of
+making the required run-time choice.  The really significant benefit, of
+course, is keeping the source code clean and well structured.
+
+We can also arrange to have private storage that varies between different
+implementations of the same kind of object.  We do this by making all the
+module-specific object structs be separately allocated entities, which will
+be accessed via pointers in the master compression or decompression struct.
+The "public" fields or methods for a given kind of object are specified by
+a commonly known struct.  But a module's initialization code can allocate
+a larger struct that contains the common struct as its first member, plus
+additional private fields.  With appropriate pointer casting, the module's
+internal functions can access these private fields.  (For a simple example,
+see jdatadst.c, which implements the external interface specified by struct
+jpeg_destination_mgr, but adds extra fields.)
+
+(Of course this would all be a lot easier if we were using C++, but we are
+not yet prepared to assume that everyone has a C++ compiler.)
+
+An important benefit of this scheme is that it is easy to provide multiple
+versions of any method, each tuned to a particular case.  While a lot of
+precalculation might be done to select an optimal implementation of a method,
+the cost per invocation is constant.  For example, the upsampling step might
+have a "generic" method, plus one or more "hardwired" methods for the most
+popular sampling factors; the hardwired methods would be faster because they'd
+use straight-line code instead of for-loops.  The cost to determine which
+method to use is paid only once, at startup, and the selection criteria are
+hidden from the callers of the method.
+
+This plan differs a little bit from usual object-oriented structures, in that
+only one instance of each object class will exist during execution.  The
+reason for having the class structure is that on different runs we may create
+different instances (choose to execute different modules).  You can think of
+the term "method" as denoting the common interface presented by a particular
+set of interchangeable functions, and "object" as denoting a group of related
+methods, or the total shared interface behavior of a group of modules.
+
+
+*** Overall control structure ***
+
+We previously mentioned the need for overall control logic in the compression
+and decompression libraries.  In IJG implementations prior to v5, overall
+control was mostly provided by "pipeline control" modules, which proved to be
+large, unwieldy, and hard to understand.  To improve the situation, the
+control logic has been subdivided into multiple modules.  The control modules
+consist of:
+
+1. Master control for module selection and initialization.  This has two
+responsibilities:
+
+   1A.  Startup initialization at the beginning of image processing.
+        The individual processing modules to be used in this run are selected
+        and given initialization calls.
+
+   1B.  Per-pass control.  This determines how many passes will be performed
+        and calls each active processing module to configure itself
+        appropriately at the beginning of each pass.  End-of-pass processing,
+	where necessary, is also invoked from the master control module.
+
+   Method selection is partially distributed, in that a particular processing
+   module may contain several possible implementations of a particular method,
+   which it will select among when given its initialization call.  The master
+   control code need only be concerned with decisions that affect more than
+   one module.
+ 
+2. Data buffering control.  A separate control module exists for each
+   inter-processing-step data buffer.  This module is responsible for
+   invoking the processing steps that write or read that data buffer.
+
+Each buffer controller sees the world as follows:
+
+input data => processing step A => buffer => processing step B => output data
+                      |              |               |
+              ------------------ controller ------------------
+
+The controller knows the dataflow requirements of steps A and B: how much data
+they want to accept in one chunk and how much they output in one chunk.  Its
+function is to manage its buffer and call A and B at the proper times.
+
+A data buffer control module may itself be viewed as a processing step by a
+higher-level control module; thus the control modules form a binary tree with
+elementary processing steps at the leaves of the tree.
+
+The control modules are objects.  A considerable amount of flexibility can
+be had by replacing implementations of a control module.  For example:
+* Merging of adjacent steps in the pipeline is done by replacing a control
+  module and its pair of processing-step modules with a single processing-
+  step module.  (Hence the possible merges are determined by the tree of
+  control modules.)
+* In some processing modes, a given interstep buffer need only be a "strip"
+  buffer large enough to accommodate the desired data chunk sizes.  In other
+  modes, a full-image buffer is needed and several passes are required.
+  The control module determines which kind of buffer is used and manipulates
+  virtual array buffers as needed.  One or both processing steps may be
+  unaware of the multi-pass behavior.
+
+In theory, we might be able to make all of the data buffer controllers
+interchangeable and provide just one set of implementations for all.  In
+practice, each one contains considerable special-case processing for its
+particular job.  The buffer controller concept should be regarded as an
+overall system structuring principle, not as a complete description of the
+task performed by any one controller.
+
+
+*** Compression object structure ***
+
+Here is a sketch of the logical structure of the JPEG compression library:
+
+                                                 |-- Colorspace conversion
+                  |-- Preprocessing controller --|
+                  |                              |-- Downsampling
+Main controller --|
+                  |                            |-- Forward DCT, quantize
+                  |-- Coefficient controller --|
+                                               |-- Entropy encoding
+
+This sketch also describes the flow of control (subroutine calls) during
+typical image data processing.  Each of the components shown in the diagram is
+an "object" which may have several different implementations available.  One
+or more source code files contain the actual implementation(s) of each object.
+
+The objects shown above are:
+
+* Main controller: buffer controller for the subsampled-data buffer, which
+  holds the preprocessed input data.  This controller invokes preprocessing to
+  fill the subsampled-data buffer, and JPEG compression to empty it.  There is
+  usually no need for a full-image buffer here; a strip buffer is adequate.
+
+* Preprocessing controller: buffer controller for the downsampling input data
+  buffer, which lies between colorspace conversion and downsampling.  Note
+  that a unified conversion/downsampling module would probably replace this
+  controller entirely.
+
+* Colorspace conversion: converts application image data into the desired
+  JPEG color space; also changes the data from pixel-interleaved layout to
+  separate component planes.  Processes one pixel row at a time.
+
+* Downsampling: performs reduction of chroma components as required.
+  Optionally may perform pixel-level smoothing as well.  Processes a "row
+  group" at a time, where a row group is defined as Vmax pixel rows of each
+  component before downsampling, and Vk sample rows afterwards (remember Vk
+  differs across components).  Some downsampling or smoothing algorithms may
+  require context rows above and below the current row group; the
+  preprocessing controller is responsible for supplying these rows via proper
+  buffering.  The downsampler is responsible for edge expansion at the right
+  edge (i.e., extending each sample row to a multiple of 8 samples); but the
+  preprocessing controller is responsible for vertical edge expansion (i.e.,
+  duplicating the bottom sample row as needed to make a multiple of 8 rows).
+
+* Coefficient controller: buffer controller for the DCT-coefficient data.
+  This controller handles MCU assembly, including insertion of dummy DCT
+  blocks when needed at the right or bottom edge.  When performing
+  Huffman-code optimization or emitting a multiscan JPEG file, this
+  controller is responsible for buffering the full image.  The equivalent of
+  one fully interleaved MCU row of subsampled data is processed per call,
+  even when the JPEG file is noninterleaved.
+
+* Forward DCT and quantization: Perform DCT, quantize, and emit coefficients.
+  Works on one or more DCT blocks at a time.  (Note: the coefficients are now
+  emitted in normal array order, which the entropy encoder is expected to
+  convert to zigzag order as necessary.  Prior versions of the IJG code did
+  the conversion to zigzag order within the quantization step.)
+
+* Entropy encoding: Perform Huffman or arithmetic entropy coding and emit the
+  coded data to the data destination module.  Works on one MCU per call.
+  For progressive JPEG, the same DCT blocks are fed to the entropy coder
+  during each pass, and the coder must emit the appropriate subset of
+  coefficients.
+
+In addition to the above objects, the compression library includes these
+objects:
+
+* Master control: determines the number of passes required, controls overall
+  and per-pass initialization of the other modules.
+
+* Marker writing: generates JPEG markers (except for RSTn, which is emitted
+  by the entropy encoder when needed).
+
+* Data destination manager: writes the output JPEG datastream to its final
+  destination (e.g., a file).  The destination manager supplied with the
+  library knows how to write to a stdio stream; for other behaviors, the
+  surrounding application may provide its own destination manager.
+
+* Memory manager: allocates and releases memory, controls virtual arrays
+  (with backing store management, where required).
+
+* Error handler: performs formatting and output of error and trace messages;
+  determines handling of nonfatal errors.  The surrounding application may
+  override some or all of this object's methods to change error handling.
+
+* Progress monitor: supports output of "percent-done" progress reports.
+  This object represents an optional callback to the surrounding application:
+  if wanted, it must be supplied by the application.
+
+The error handler, destination manager, and progress monitor objects are
+defined as separate objects in order to simplify application-specific
+customization of the JPEG library.  A surrounding application may override
+individual methods or supply its own all-new implementation of one of these
+objects.  The object interfaces for these objects are therefore treated as
+part of the application interface of the library, whereas the other objects
+are internal to the library.
+
+The error handler and memory manager are shared by JPEG compression and
+decompression; the progress monitor, if used, may be shared as well.
+
+
+*** Decompression object structure ***
+
+Here is a sketch of the logical structure of the JPEG decompression library:
+
+                                               |-- Entropy decoding
+                  |-- Coefficient controller --|
+                  |                            |-- Dequantize, Inverse DCT
+Main controller --|
+                  |                               |-- Upsampling
+                  |-- Postprocessing controller --|   |-- Colorspace conversion
+                                                  |-- Color quantization
+                                                  |-- Color precision reduction
+
+As before, this diagram also represents typical control flow.  The objects
+shown are:
+
+* Main controller: buffer controller for the subsampled-data buffer, which
+  holds the output of JPEG decompression proper.  This controller's primary
+  task is to feed the postprocessing procedure.  Some upsampling algorithms
+  may require context rows above and below the current row group; when this
+  is true, the main controller is responsible for managing its buffer so as
+  to make context rows available.  In the current design, the main buffer is
+  always a strip buffer; a full-image buffer is never required.
+
+* Coefficient controller: buffer controller for the DCT-coefficient data.
+  This controller handles MCU disassembly, including deletion of any dummy
+  DCT blocks at the right or bottom edge.  When reading a multiscan JPEG
+  file, this controller is responsible for buffering the full image.
+  (Buffering DCT coefficients, rather than samples, is necessary to support
+  progressive JPEG.)  The equivalent of one fully interleaved MCU row of
+  subsampled data is processed per call, even when the source JPEG file is
+  noninterleaved.
+
+* Entropy decoding: Read coded data from the data source module and perform
+  Huffman or arithmetic entropy decoding.  Works on one MCU per call.
+  For progressive JPEG decoding, the coefficient controller supplies the prior
+  coefficients of each MCU (initially all zeroes), which the entropy decoder
+  modifies in each scan.
+
+* Dequantization and inverse DCT: like it says.  Note that the coefficients
+  buffered by the coefficient controller have NOT been dequantized; we
+  merge dequantization and inverse DCT into a single step for speed reasons.
+  When scaled-down output is asked for, simplified DCT algorithms may be used
+  that emit only 1x1, 2x2, or 4x4 samples per DCT block, not the full 8x8.
+  Works on one DCT block at a time.
+
+* Postprocessing controller: buffer controller for the color quantization
+  input buffer, when quantization is in use.  (Without quantization, this
+  controller just calls the upsampler.)  For two-pass quantization, this
+  controller is responsible for buffering the full-image data.
+
+* Upsampling: restores chroma components to full size.  (May support more
+  general output rescaling, too.  Note that if undersized DCT outputs have
+  been emitted by the DCT module, this module must adjust so that properly
+  sized outputs are created.)  Works on one row group at a time.  This module
+  also calls the color conversion module, so its top level is effectively a
+  buffer controller for the upsampling->color conversion buffer.  However, in
+  all but the highest-quality operating modes, upsampling and color
+  conversion are likely to be merged into a single step.
+
+* Colorspace conversion: convert from JPEG color space to output color space,
+  and change data layout from separate component planes to pixel-interleaved.
+  Works on one pixel row at a time.
+
+* Color quantization: reduce the data to colormapped form, using either an
+  externally specified colormap or an internally generated one.  This module
+  is not used for full-color output.  Works on one pixel row at a time; may
+  require two passes to generate a color map.  Note that the output will
+  always be a single component representing colormap indexes.  In the current
+  design, the output values are JSAMPLEs, so an 8-bit compilation cannot
+  quantize to more than 256 colors.  This is unlikely to be a problem in
+  practice.
+
+* Color reduction: this module handles color precision reduction, e.g.,
+  generating 15-bit color (5 bits/primary) from JPEG's 24-bit output.
+  Not quite clear yet how this should be handled... should we merge it with
+  colorspace conversion???
+
+Note that some high-speed operating modes might condense the entire
+postprocessing sequence to a single module (upsample, color convert, and
+quantize in one step).
+
+In addition to the above objects, the decompression library includes these
+objects:
+
+* Master control: determines the number of passes required, controls overall
+  and per-pass initialization of the other modules.  This is subdivided into
+  input and output control: jdinput.c controls only input-side processing,
+  while jdmaster.c handles overall initialization and output-side control.
+
+* Marker reading: decodes JPEG markers (except for RSTn).
+
+* Data source manager: supplies the input JPEG datastream.  The source
+  manager supplied with the library knows how to read from a stdio stream;
+  for other behaviors, the surrounding application may provide its own source
+  manager.
+
+* Memory manager: same as for compression library.
+
+* Error handler: same as for compression library.
+
+* Progress monitor: same as for compression library.
+
+As with compression, the data source manager, error handler, and progress
+monitor are candidates for replacement by a surrounding application.
+
+
+*** Decompression input and output separation ***
+
+To support efficient incremental display of progressive JPEG files, the
+decompressor is divided into two sections that can run independently:
+
+1. Data input includes marker parsing, entropy decoding, and input into the
+   coefficient controller's DCT coefficient buffer.  Note that this
+   processing is relatively cheap and fast.
+
+2. Data output reads from the DCT coefficient buffer and performs the IDCT
+   and all postprocessing steps.
+
+For a progressive JPEG file, the data input processing is allowed to get
+arbitrarily far ahead of the data output processing.  (This occurs only
+if the application calls jpeg_consume_input(); otherwise input and output
+run in lockstep, since the input section is called only when the output
+section needs more data.)  In this way the application can avoid making
+extra display passes when data is arriving faster than the display pass
+can run.  Furthermore, it is possible to abort an output pass without
+losing anything, since the coefficient buffer is read-only as far as the
+output section is concerned.  See libjpeg.txt for more detail.
+
+A full-image coefficient array is only created if the JPEG file has multiple
+scans (or if the application specifies buffered-image mode anyway).  When
+reading a single-scan file, the coefficient controller normally creates only
+a one-MCU buffer, so input and output processing must run in lockstep in this
+case.  jpeg_consume_input() is effectively a no-op in this situation.
+
+The main impact of dividing the decompressor in this fashion is that we must
+be very careful with shared variables in the cinfo data structure.  Each
+variable that can change during the course of decompression must be
+classified as belonging to data input or data output, and each section must
+look only at its own variables.  For example, the data output section may not
+depend on any of the variables that describe the current scan in the JPEG
+file, because these may change as the data input section advances into a new
+scan.
+
+The progress monitor is (somewhat arbitrarily) defined to treat input of the
+file as one pass when buffered-image mode is not used, and to ignore data
+input work completely when buffered-image mode is used.  Note that the
+library has no reliable way to predict the number of passes when dealing
+with a progressive JPEG file, nor can it predict the number of output passes
+in buffered-image mode.  So the work estimate is inherently bogus anyway.
+
+No comparable division is currently made in the compression library, because
+there isn't any real need for it.
+
+
+*** Data formats ***
+
+Arrays of pixel sample values use the following data structure:
+
+    typedef something JSAMPLE;		a pixel component value, 0..MAXJSAMPLE
+    typedef JSAMPLE *JSAMPROW;		ptr to a row of samples
+    typedef JSAMPROW *JSAMPARRAY;	ptr to a list of rows
+    typedef JSAMPARRAY *JSAMPIMAGE;	ptr to a list of color-component arrays
+
+The basic element type JSAMPLE will typically be one of unsigned char,
+(signed) char, or short.  Short will be used if samples wider than 8 bits are
+to be supported (this is a compile-time option).  Otherwise, unsigned char is
+used if possible.  If the compiler only supports signed chars, then it is
+necessary to mask off the value when reading.  Thus, all reads of JSAMPLE
+values must be coded as "GETJSAMPLE(value)", where the macro will be defined
+as "((value) & 0xFF)" on signed-char machines and "((int) (value))" elsewhere.
+
+With these conventions, JSAMPLE values can be assumed to be >= 0.  This helps
+simplify correct rounding during downsampling, etc.  The JPEG standard's
+specification that sample values run from -128..127 is accommodated by
+subtracting 128 just as the sample value is copied into the source array for
+the DCT step (this will be an array of signed ints).  Similarly, during
+decompression the output of the IDCT step will be immediately shifted back to
+0..255.  (NB: different values are required when 12-bit samples are in use.
+The code is written in terms of MAXJSAMPLE and CENTERJSAMPLE, which will be
+defined as 255 and 128 respectively in an 8-bit implementation, and as 4095
+and 2048 in a 12-bit implementation.)
+
+We use a pointer per row, rather than a two-dimensional JSAMPLE array.  This
+choice costs only a small amount of memory and has several benefits:
+* Code using the data structure doesn't need to know the allocated width of
+  the rows.  This simplifies edge expansion/compression, since we can work
+  in an array that's wider than the logical picture width.
+* Indexing doesn't require multiplication; this is a performance win on many
+  machines.
+* Arrays with more than 64K total elements can be supported even on machines
+  where malloc() cannot allocate chunks larger than 64K.
+* The rows forming a component array may be allocated at different times
+  without extra copying.  This trick allows some speedups in smoothing steps
+  that need access to the previous and next rows.
+
+Note that each color component is stored in a separate array; we don't use the
+traditional layout in which the components of a pixel are stored together.
+This simplifies coding of modules that work on each component independently,
+because they don't need to know how many components there are.  Furthermore,
+we can read or write each component to a temporary file independently, which
+is helpful when dealing with noninterleaved JPEG files.
+
+In general, a specific sample value is accessed by code such as
+	GETJSAMPLE(image[colorcomponent][row][col])
+where col is measured from the image left edge, but row is measured from the
+first sample row currently in memory.  Either of the first two indexings can
+be precomputed by copying the relevant pointer.
+
+
+Since most image-processing applications prefer to work on images in which
+the components of a pixel are stored together, the data passed to or from the
+surrounding application uses the traditional convention: a single pixel is
+represented by N consecutive JSAMPLE values, and an image row is an array of
+(# of color components)*(image width) JSAMPLEs.  One or more rows of data can
+be represented by a pointer of type JSAMPARRAY in this scheme.  This scheme is
+converted to component-wise storage inside the JPEG library.  (Applications
+that want to skip JPEG preprocessing or postprocessing will have to contend
+with component-wise storage.)
+
+
+Arrays of DCT-coefficient values use the following data structure:
+
+    typedef short JCOEF;		a 16-bit signed integer
+    typedef JCOEF JBLOCK[DCTSIZE2];	an 8x8 block of coefficients
+    typedef JBLOCK *JBLOCKROW;		ptr to one horizontal row of 8x8 blocks
+    typedef JBLOCKROW *JBLOCKARRAY;	ptr to a list of such rows
+    typedef JBLOCKARRAY *JBLOCKIMAGE;	ptr to a list of color component arrays
+
+The underlying type is at least a 16-bit signed integer; while "short" is big
+enough on all machines of interest, on some machines it is preferable to use
+"int" for speed reasons, despite the storage cost.  Coefficients are grouped
+into 8x8 blocks (but we always use #defines DCTSIZE and DCTSIZE2 rather than
+"8" and "64").
+
+The contents of a coefficient block may be in either "natural" or zigzagged
+order, and may be true values or divided by the quantization coefficients,
+depending on where the block is in the processing pipeline.  In the current
+library, coefficient blocks are kept in natural order everywhere; the entropy
+codecs zigzag or dezigzag the data as it is written or read.  The blocks
+contain quantized coefficients everywhere outside the DCT/IDCT subsystems.
+(This latter decision may need to be revisited to support variable
+quantization a la JPEG Part 3.)
+
+Notice that the allocation unit is now a row of 8x8 blocks, corresponding to
+eight rows of samples.  Otherwise the structure is much the same as for
+samples, and for the same reasons.
+
+On machines where malloc() can't handle a request bigger than 64Kb, this data
+structure limits us to rows of less than 512 JBLOCKs, or a picture width of
+4000+ pixels.  This seems an acceptable restriction.
+
+
+On 80x86 machines, the bottom-level pointer types (JSAMPROW and JBLOCKROW)
+must be declared as "far" pointers, but the upper levels can be "near"
+(implying that the pointer lists are allocated in the DS segment).
+We use a #define symbol FAR, which expands to the "far" keyword when
+compiling on 80x86 machines and to nothing elsewhere.
+
+
+*** Suspendable processing ***
+
+In some applications it is desirable to use the JPEG library as an
+incremental, memory-to-memory filter.  In this situation the data source or
+destination may be a limited-size buffer, and we can't rely on being able to
+empty or refill the buffer at arbitrary times.  Instead the application would
+like to have control return from the library at buffer overflow/underrun, and
+then resume compression or decompression at a later time.
+
+This scenario is supported for simple cases.  (For anything more complex, we
+recommend that the application "bite the bullet" and develop real multitasking
+capability.)  The libjpeg.txt file goes into more detail about the usage and
+limitations of this capability; here we address the implications for library
+structure.
+
+The essence of the problem is that the entropy codec (coder or decoder) must
+be prepared to stop at arbitrary times.  In turn, the controllers that call
+the entropy codec must be able to stop before having produced or consumed all
+the data that they normally would handle in one call.  That part is reasonably
+straightforward: we make the controller call interfaces include "progress
+counters" which indicate the number of data chunks successfully processed, and
+we require callers to test the counter rather than just assume all of the data
+was processed.
+
+Rather than trying to restart at an arbitrary point, the current Huffman
+codecs are designed to restart at the beginning of the current MCU after a
+suspension due to buffer overflow/underrun.  At the start of each call, the
+codec's internal state is loaded from permanent storage (in the JPEG object
+structures) into local variables.  On successful completion of the MCU, the
+permanent state is updated.  (This copying is not very expensive, and may even
+lead to *improved* performance if the local variables can be registerized.)
+If a suspension occurs, the codec simply returns without updating the state,
+thus effectively reverting to the start of the MCU.  Note that this implies
+leaving some data unprocessed in the source/destination buffer (ie, the
+compressed partial MCU).  The data source/destination module interfaces are
+specified so as to make this possible.  This also implies that the data buffer
+must be large enough to hold a worst-case compressed MCU; a couple thousand
+bytes should be enough.
+
+In a successive-approximation AC refinement scan, the progressive Huffman
+decoder has to be able to undo assignments of newly nonzero coefficients if it
+suspends before the MCU is complete, since decoding requires distinguishing
+previously-zero and previously-nonzero coefficients.  This is a bit tedious
+but probably won't have much effect on performance.  Other variants of Huffman
+decoding need not worry about this, since they will just store the same values
+again if forced to repeat the MCU.
+
+This approach would probably not work for an arithmetic codec, since its
+modifiable state is quite large and couldn't be copied cheaply.  Instead it
+would have to suspend and resume exactly at the point of the buffer end.
+
+The JPEG marker reader is designed to cope with suspension at an arbitrary
+point.  It does so by backing up to the start of the marker parameter segment,
+so the data buffer must be big enough to hold the largest marker of interest.
+Again, a couple KB should be adequate.  (A special "skip" convention is used
+to bypass COM and APPn markers, so these can be larger than the buffer size
+without causing problems; otherwise a 64K buffer would be needed in the worst
+case.)
+
+The JPEG marker writer currently does *not* cope with suspension.
+We feel that this is not necessary; it is much easier simply to require
+the application to ensure there is enough buffer space before starting.  (An
+empty 2K buffer is more than sufficient for the header markers; and ensuring
+there are a dozen or two bytes available before calling jpeg_finish_compress()
+will suffice for the trailer.)  This would not work for writing multi-scan
+JPEG files, but we simply do not intend to support that capability with
+suspension.
+
+
+*** Memory manager services ***
+
+The JPEG library's memory manager controls allocation and deallocation of
+memory, and it manages large "virtual" data arrays on machines where the
+operating system does not provide virtual memory.  Note that the same
+memory manager serves both compression and decompression operations.
+
+In all cases, allocated objects are tied to a particular compression or
+decompression master record, and they will be released when that master
+record is destroyed.
+
+The memory manager does not provide explicit deallocation of objects.
+Instead, objects are created in "pools" of free storage, and a whole pool
+can be freed at once.  This approach helps prevent storage-leak bugs, and
+it speeds up operations whenever malloc/free are slow (as they often are).
+The pools can be regarded as lifetime identifiers for objects.  Two
+pools/lifetimes are defined:
+  * JPOOL_PERMANENT	lasts until master record is destroyed
+  * JPOOL_IMAGE		lasts until done with image (JPEG datastream)
+Permanent lifetime is used for parameters and tables that should be carried
+across from one datastream to another; this includes all application-visible
+parameters.  Image lifetime is used for everything else.  (A third lifetime,
+JPOOL_PASS = one processing pass, was originally planned.  However it was
+dropped as not being worthwhile.  The actual usage patterns are such that the
+peak memory usage would be about the same anyway; and having per-pass storage
+substantially complicates the virtual memory allocation rules --- see below.)
+
+The memory manager deals with three kinds of object:
+1. "Small" objects.  Typically these require no more than 10K-20K total.
+2. "Large" objects.  These may require tens to hundreds of K depending on
+   image size.  Semantically they behave the same as small objects, but we
+   distinguish them for two reasons:
+     * On MS-DOS machines, large objects are referenced by FAR pointers,
+       small objects by NEAR pointers.
+     * Pool allocation heuristics may differ for large and small objects.
+   Note that individual "large" objects cannot exceed the size allowed by
+   type size_t, which may be 64K or less on some machines.
+3. "Virtual" objects.  These are large 2-D arrays of JSAMPLEs or JBLOCKs
+   (typically large enough for the entire image being processed).  The
+   memory manager provides stripwise access to these arrays.  On machines
+   without virtual memory, the rest of the array may be swapped out to a
+   temporary file.
+
+(Note: JSAMPARRAY and JBLOCKARRAY data structures are a combination of large
+objects for the data proper and small objects for the row pointers.  For
+convenience and speed, the memory manager provides single routines to create
+these structures.  Similarly, virtual arrays include a small control block
+and a JSAMPARRAY or JBLOCKARRAY working buffer, all created with one call.)
+
+In the present implementation, virtual arrays are only permitted to have image
+lifespan.  (Permanent lifespan would not be reasonable, and pass lifespan is
+not very useful since a virtual array's raison d'etre is to store data for
+multiple passes through the image.)  We also expect that only "small" objects
+will be given permanent lifespan, though this restriction is not required by
+the memory manager.
+
+In a non-virtual-memory machine, some performance benefit can be gained by
+making the in-memory buffers for virtual arrays be as large as possible.
+(For small images, the buffers might fit entirely in memory, so blind
+swapping would be very wasteful.)  The memory manager will adjust the height
+of the buffers to fit within a prespecified maximum memory usage.  In order
+to do this in a reasonably optimal fashion, the manager needs to allocate all
+of the virtual arrays at once.  Therefore, there isn't a one-step allocation
+routine for virtual arrays; instead, there is a "request" routine that simply
+allocates the control block, and a "realize" routine (called just once) that
+determines space allocation and creates all of the actual buffers.  The
+realize routine must allow for space occupied by non-virtual large objects.
+(We don't bother to factor in the space needed for small objects, on the
+grounds that it isn't worth the trouble.)
+
+To support all this, we establish the following protocol for doing business
+with the memory manager:
+  1. Modules must request virtual arrays (which may have only image lifespan)
+     during the initial setup phase, i.e., in their jinit_xxx routines.
+  2. All "large" objects (including JSAMPARRAYs and JBLOCKARRAYs) must also be
+     allocated during initial setup.
+  3. realize_virt_arrays will be called at the completion of initial setup.
+     The above conventions ensure that sufficient information is available
+     for it to choose a good size for virtual array buffers.
+Small objects of any lifespan may be allocated at any time.  We expect that
+the total space used for small objects will be small enough to be negligible
+in the realize_virt_arrays computation.
+
+In a virtual-memory machine, we simply pretend that the available space is
+infinite, thus causing realize_virt_arrays to decide that it can allocate all
+the virtual arrays as full-size in-memory buffers.  The overhead of the
+virtual-array access protocol is very small when no swapping occurs.
+
+A virtual array can be specified to be "pre-zeroed"; when this flag is set,
+never-yet-written sections of the array are set to zero before being made
+available to the caller.  If this flag is not set, never-written sections
+of the array contain garbage.  (This feature exists primarily because the
+equivalent logic would otherwise be needed in jdcoefct.c for progressive
+JPEG mode; we may as well make it available for possible other uses.)
+
+The first write pass on a virtual array is required to occur in top-to-bottom
+order; read passes, as well as any write passes after the first one, may
+access the array in any order.  This restriction exists partly to simplify
+the virtual array control logic, and partly because some file systems may not
+support seeking beyond the current end-of-file in a temporary file.  The main
+implication of this restriction is that rearrangement of rows (such as
+converting top-to-bottom data order to bottom-to-top) must be handled while
+reading data out of the virtual array, not while putting it in.
+
+
+*** Memory manager internal structure ***
+
+To isolate system dependencies as much as possible, we have broken the
+memory manager into two parts.  There is a reasonably system-independent
+"front end" (jmemmgr.c) and a "back end" that contains only the code
+likely to change across systems.  All of the memory management methods
+outlined above are implemented by the front end.  The back end provides
+the following routines for use by the front end (none of these routines
+are known to the rest of the JPEG code):
+
+jpeg_mem_init, jpeg_mem_term	system-dependent initialization/shutdown
+
+jpeg_get_small, jpeg_free_small	interface to malloc and free library routines
+				(or their equivalents)
+
+jpeg_get_large, jpeg_free_large	interface to FAR malloc/free in MSDOS machines;
+				else usually the same as
+				jpeg_get_small/jpeg_free_small
+
+jpeg_mem_available		estimate available memory
+
+jpeg_open_backing_store		create a backing-store object
+
+read_backing_store,		manipulate a backing-store object
+write_backing_store,
+close_backing_store
+
+On some systems there will be more than one type of backing-store object
+(specifically, in MS-DOS a backing store file might be an area of extended
+memory as well as a disk file).  jpeg_open_backing_store is responsible for
+choosing how to implement a given object.  The read/write/close routines
+are method pointers in the structure that describes a given object; this
+lets them be different for different object types.
+
+It may be necessary to ensure that backing store objects are explicitly
+released upon abnormal program termination.  For example, MS-DOS won't free
+extended memory by itself.  To support this, we will expect the main program
+or surrounding application to arrange to call self_destruct (typically via
+jpeg_destroy) upon abnormal termination.  This may require a SIGINT signal
+handler or equivalent.  We don't want to have the back end module install its
+own signal handler, because that would pre-empt the surrounding application's
+ability to control signal handling.
+
+The IJG distribution includes several memory manager back end implementations.
+Usually the same back end should be suitable for all applications on a given
+system, but it is possible for an application to supply its own back end at
+need.
+
+
+*** Implications of DNL marker ***
+
+Some JPEG files may use a DNL marker to postpone definition of the image
+height (this would be useful for a fax-like scanner's output, for instance).
+In these files the SOF marker claims the image height is 0, and you only
+find out the true image height at the end of the first scan.
+
+We could read these files as follows:
+1. Upon seeing zero image height, replace it by 65535 (the maximum allowed).
+2. When the DNL is found, update the image height in the global image
+   descriptor.
+This implies that control modules must avoid making copies of the image
+height, and must re-test for termination after each MCU row.  This would
+be easy enough to do.
+
+In cases where image-size data structures are allocated, this approach will
+result in very inefficient use of virtual memory or much-larger-than-necessary
+temporary files.  This seems acceptable for something that probably won't be a
+mainstream usage.  People might have to forgo use of memory-hogging options
+(such as two-pass color quantization or noninterleaved JPEG files) if they
+want efficient conversion of such files.  (One could improve efficiency by
+demanding a user-supplied upper bound for the height, less than 65536; in most
+cases it could be much less.)
+
+The standard also permits the SOF marker to overestimate the image height,
+with a DNL to give the true, smaller height at the end of the first scan.
+This would solve the space problems if the overestimate wasn't too great.
+However, it implies that you don't even know whether DNL will be used.
+
+This leads to a couple of very serious objections:
+1. Testing for a DNL marker must occur in the inner loop of the decompressor's
+   Huffman decoder; this implies a speed penalty whether the feature is used
+   or not.
+2. There is no way to hide the last-minute change in image height from an
+   application using the decoder.  Thus *every* application using the IJG
+   library would suffer a complexity penalty whether it cared about DNL or
+   not.
+We currently do not support DNL because of these problems.
+
+A different approach is to insist that DNL-using files be preprocessed by a
+separate program that reads ahead to the DNL, then goes back and fixes the SOF
+marker.  This is a much simpler solution and is probably far more efficient.
+Even if one wants piped input, buffering the first scan of the JPEG file needs
+a lot smaller temp file than is implied by the maximum-height method.  For
+this approach we'd simply treat DNL as a no-op in the decompressor (at most,
+check that it matches the SOF image height).
+
+We will not worry about making the compressor capable of outputting DNL.
+Something similar to the first scheme above could be applied if anyone ever
+wants to make that work.
diff --git a/testimg.bmp b/testimg.bmp
new file mode 100644
index 0000000..8603d15
--- /dev/null
+++ b/testimg.bmp
diff --git a/testimgcrop.jpg b/testimgcrop.jpg
new file mode 100644
index 0000000..18fa0b1
--- /dev/null
+++ b/testimgcrop.jpg
diff --git a/testimgflt-nosimd.jpg b/testimgflt-nosimd.jpg
new file mode 100644
index 0000000..462f39b
--- /dev/null
+++ b/testimgflt-nosimd.jpg
diff --git a/testimgflt.jpg b/testimgflt.jpg
new file mode 100644
index 0000000..832f4cc
--- /dev/null
+++ b/testimgflt.jpg
diff --git a/testimgflt.ppm b/testimgflt.ppm
new file mode 100644
index 0000000..78945b6
--- /dev/null
+++ b/testimgflt.ppm
@@ -0,0 +1,4 @@
+P6
+227 149
+255
+0/-0/-10.21/40-51.62/62/83/83/:3-:3-:3-:3-:3-:3-:2/91.91.80-80-91.91.:2/80-80-80-80-80-80-80-80-6.+6.+6.+5-*5-*4,)4,)4,)4,)4,)4,)4,)4,)4,)4,)2-)/*$/,%/,%0-&1.'2/(30)30)63,63,74-85.85.96/:70:7.A:0B<0D>2F@4IA4JB5KC6KC6MD5MD5OC3NB2OC3OC3PD4RE5R?1Y?2b@4nB5}E6‹H8™G9£F7¯G:¸G9¾E:ÅG;ÇG>ÊG?ËH@ÐE@çFLíCLëDKëEIîCIïBDñ>Bô=Aø;A÷:@ô:?ð<?é?@â@>×?<ËA7»=/µ@.µ@.´?-´?-²?,²?-¯@-­@,ªA,¦A-¢B,Ÿ@*›A)˜@*–A,”>-’?/’?/‘>.‘>,�=+’<+’<+”?+”?+”=*”=*”=*•>+–?,–@/–?6•>5—=2Ÿ?1©B3³D3¼D4¿D4¹?0µA2¬F8žH;‡H9oA2T8*C3&=5295495473271160050-50-72/72/72/61.61-50,50,41,//-.0-//-//-0/-0/-2.-2.-4+,4+,4*+3)*7(+=.1E69P:<jBC|IH�MM•OOžVZªdnªoƒŸt�”{£‡®€†º~ˆ½sy­`a‘TKvQEiSJgOG^MH^TQbfdo|}‚�‘�™ž˜£©Ÿ¢¨šž “�’�{|lfgUXWEQNEUR[UQbUQb0/-0/-10.10.40-51.62/62/83/83/:3-:3-:3-:3-:3-:3-91.91.80-80-80-80-91.91.80-80-80-80-80-80-80-80-6.+6.+5-*5-*5-*4,)4,)4,)5-*5-*5-*5-*5-*5-*5-*3.*0+%0-&0-&1.'2/(30)41*41*63,63,74-74-85.96/:70:7.@9/A;/C=1E?3H@3IA4JB5JB5LC4LC4MA1MA1MA1NB2OC3QD4P>0U?1^A3jC4xD6„E4’E5œC3§C4¯A4µA4¼B7ÀD:ÄE<ÅF=ÍC@áEIçBIèCIêDHíDGïBDó@Cö?Cø;A÷:@ô:?ð<?é?@àA>Õ@<Ê@6¹>/µ@.´?-´?-´?-²?,°?-¯@-­@,©@+¦A-¡A+Ÿ@*›A)˜@*–A,”>-’?/‘>.‘>.‘>,�=+’<+’<+”?+“>*”=*”=*”=*•>+–?,–@/”@5•>5˜>3 >1«A3µD4½C4¿D5»A2·C6¬F8œI;…G:l@3S9*B4)>63:6595484382271161.61.72/72/72/61.61-50,50,41,//-.0-//-//-0/-0/-2.-2.-3--5,-5+,3)*5)+<-0C47N8:d>=vEA†JI�NLšTV¤aj¥l}�rŽ‘{¢†€®…¹{„»ou©[[�QHuOCiOFeOG_PH_RN_[Yfnotƒ‡ˆ”™•™ž—š ”™���‘ƒ~ojkY][LVSJXSZVRaXQa/.,/.,0/-10.40-40-51.51.72.72.72.72.92,92,92,92,91.80.7/-7/-7/-7/-80.91/80.80.80.80.80.80.80.80.6.,5-+5-+5-+4,*4,*4,*4,*5-+5-+5-+5-+5-+5-+5-+3.*2-'1.'2/(30)30)41*41*52+63,63,63,74-85.96/96/:7.?8.@:.B<0D>2G?4H@5H@3H@3I@1I@1K?1K?1K?/L@0MA1NB2MA1QA1YB2dC2qC3|C2‡B2’A0˜<- :+§;.¯=2µ@6ºD:¿F=ÅD>ÙCEá@FãBGçBFêDFðCEôADø?Dú;@ù:?õ;@ð=@è@@ÜA=Ñ@;Æ@5·=.³@-³@-²?,²?-°?-¯>,­@,ª?-§@-¥@,¡A+�A,š?*˜@*•@+”>-‘>.‘>.‘>.�=+�=+�=+�=+‘>,‘>,’<+’<+“=,“=,”?+•?.•A6–?5š>3£>2¯A4¹C5¿D5ÁC5ÀD8¸F;®I=™J=�G;h@4Q:,B5,?74=77<66;5594183072/72/62/62/62/51.52-41,41,21,.0-,1-.0-.0-//-//-0/-2.-5//4..5,-4*+4*+9-/>24I56[97l?9|E@†ID�OM˜[`›fv”mˆŒwžƒ}­}‚¹u~·fm¤TV‰MEvLAkMAeOFcQHcMH^NK\[[eqty…‰ˆ‡Œ†Š�ˆ��…†Š|xzlfiXZ[MVSLZU[ZT`[S`.-+/.,/.,0/-3/,40-40-40-61-61-61-61-81+81+81+81+7/-7/-6.,6.,6.,6.,7/-7/-80.80.80.80.80.80.80.80.5-+5-+5-+4,*4,*4,*3+)3+)6.,6.,6.,6.,6.,6.,6.,4/,30+30)30)41*41*52+52+52+52+52+63,74-85.85.96/96->7-?9-@:.B<0E=2E=2F>1F>1G=1G>/H<.I=/I=/J>.L@0JA0LE3NE4UD4^D3iD2sB1~A/†?-Œ9)”9'�9*£<-¬@3³E8¸H<ÁF>ÒDCÚACÞBCâDCçCDìBCó@C÷?Aú;@ù:?õ;@î>@åA@ÚB=Í@9Â@3µ=-°@,°@,°@,¯>,®?,®?,¬?+©@-¦?,£@+ @*œ@+˜@*–@)”?*‘>,‘>.‘>.�=-�=+�=+�<*�<*�=+�=+�<*�<*’<+‘>,”>-’?-•A6–?5œ>2¦@4²B6¼C8ÁC7ÂB7ÂF<¹I>¬L@—K>|F:b@4L:.A7-@85>96=77<74:5294183083062/62/62/32.52-21,21,12--2.-2./1./1.00.00.10.10.5106005//5,-4+,6,-:01D22T71c;3qA7{E;‚HD‰RU�_l‹i‚ƒs˜}y«x}µowµ`f¢QR‹LEyL@pL@hPEgQFfLC^GBVMLZ^^fjnquyxy~xz€vwzokoa`bTWYLTTL]WY]V]]V^------.-+/.,0/-10.3/,40-3/,3/,4/+4/+4/+4/+6/)6/)4/,4/,3.+3.+3.+3.+4/,4/,50-50-50-50-50-50-50-50-3.+3.+2-*2-*2-*1,)1,)1,)4/,4/,4/,4/,4/,4/,4/,4/,41,41,41,41,52-52-52-52-52-52-63.63.74/85096196/<5-=6,?8.@9/B:/C;0C;0C;.D:.D:.F:.G;-H<.I=/J>0I@1JG6MH5RG5YF5bE3jD1uB/|?,‚;)‹:)“:*š=,£B2¬F8²J=»I?ÌGBÔDCØDBÜDAâBBéAAð=@ô<>ù:?ø:<ô<>í?>áB>ÓC:ÅA5¹?0²?-®?,®?,®?,­>-¬>-¬>-ª?-¨>.¤?- ?,ž?+š?,—?+•>*“>*‘>,�?.�>-�>-�>-Ž=,Ž=,Ž=,Ž=,Ž=,Ž=,Ž=,�<,�>-‘>.�?.”B4—A4�@1¨@3¶A7¿C9ÅB8ÄA7¾C;·H?¦LC�J@tE;Z>2E9-<5+@93@85?75>63=52<4194083/62/43/43/23.32.12-12-02--2.,2.-2.-2./1./1.00.10.3205105104..2,,4+,7./=/.N5.Y9.e=1n@3sB;yKK€Zeƒg€p—zxªu{·ks´_d¦TT”OGƒLBwNAmNBhMAeJA`GBYGFXKKWMPU]cc`fbbia`f\Z`TW[MUXMXXP^YV`WX`WZ,,,,,,.-+.-+/.,0/-3/,3/,2.+2.+3.*3.*3.*3.*5.(5-*3.+3.+2-*1,)1,)2-*3.+3.+3.+3.+3.+3.+3.+3.+3.+3.+2-*2-*2-*2-*1,)1,)1,)0+(3.+3.+3.+3.+3.+3.+3.+3.+41,41,41,41,41,41,41,41,41,52-52-63.74/85085085.;4,<5+=6,>7-@7.A9.A9.A9.C9/C9-E9-F:.G;/H<.J>0HA1JG6IH6NG5VF6\E3dC2lA0t?-{<+‚;)Œ;*”=,œ@1£F5ªJ:´J=ÄH@ÌEAÑFAÖE@ÞCAä@>ì>?ñ:<÷;<ô:;ð<=é@=ÜC=ÍC8¾@2±>,®?,«@,ª?+ª?+ª?-©>,©>,¨?,¥>-¢?, ?,›>,—?+•>*“>)‘?*�>+�>-�>-�>-Ž=,Ž=,Ž=,�<+Ž=,‹<+�<+‹<+‹<-Œ=,�>/Ž?0”B4—A2 @2¬B5¹C9ÂC:ÅB:ÂA;ºA9±I@£NG�NEoG=S?4A;/96-@93A75?74>63<4194083/74/43/43/34/23.23.02-02-.3--3/-3/.3/.3/02/02/11/11/21/32032040/2.-1-,4..8.,G4-O4)X8+`<0e?6mGFyYd‚k…€uŸ||²w|¼nu»dh¯[[¡RM�LB~OArL@hI=cH>`HB^ECX@BO;@FBGJDMJJQJJQIIQFKQEOUIVWO^YS`YS`XU++++++,,,---/.,/.,0/-0/-1-*1-*1-*1-*2-)2-)2-)2-)2-*2,,1++1++1++1++2,,2,,1++1++1++1++1++1++1++1++2,,2,,2,,1++1++1++0**0**3--3--3--3--3--3--3--3.+41,41,41,30+30+30+30+30+41,41,52-63.74/74/85085.:3+;4,<5-=6.?6/?6-?6-?7,B8.B8.E8/E9-G;/H<0J>2H@3HE6GE6KE5QD4XC2_B2f?.n=,v=,|:*…9)Œ;*“=.›B2¢F7¬F8¼G=ÂF>ÉF>ÐE>ÙD@âC?é??ð>>ò::ñ;:ì<<äA<ÖC;ÆD6µ@/ª=)ª?-©@-©@-¨?,¨>.¨>.§=-¥>-£=. ?.ž?-š?-–?,“>*‘?*�>)�>+�>-�>-Œ=,Œ=.Œ=.‹<-‹<-‹<-Š=-Š;,‰<,Š</Š=-‹=0Œ?/’C4˜B3¡A3®B6¼C:ÂC<ÄC=ÀC=¹FA²QJ¥XR�XQsRIWI>CC7<?6>93@72>71=60:5/94.83/63.43.43.34/13.13..3-.3-.3--3/-3/-3/-3/.3/.3/02/02/00.11/22021/0/-/.,2.-4/,?0+D0)K3)T8-Z<4dFFu]jƒs�‰€«…„¾~ƒÇtzÆmp½ce¯VSšLC‚K?qI<gG;cE>_FB]DBW?AN;?H:BE>HGDMHGQIGQHJRGNVKUXM^ZOaYNaXO++++++,,,---.,-/.,0/-0/-1-*1-*1-*1-*2-)2-)2-)2-*2,,1++1++0**0**1++1++2,,0**0**0**0**0**0**0**0**2,,2,,2,,1++1++0**0**0**2,,2,,2,,2,,2,,2,,2,,2,,3/,30+30+30+30+30+30+2/*41,41,52-63.63.74/85085092,:3+;4,<5->5.>5.>5.>5,B8/B8.E8/E8/G:1I=1J>2I?3FB6FB6JB5OA4TA2\@2b>0j<-q<.w9*}8)…8(Œ:,•=/›B4¤B5²F:ºE;ÁF>ÊG?ÔG@ÞFAçCAîB@í;;ë;;ç>;ßB;ÑD:¿D4¯@-¤>(¦A-¦A-¦A-¥@.¥@.¤?-¤?-¤>/¢>.Ÿ@.œ?.˜?-•>+‘?*�>)�>+�>-�>-�>-Œ=,Œ=.‹<-‹<-Š=-Š=-ˆ<.ˆ<.ˆ<.ˆ</‡>/ˆ>1‹?1‘D4–C3¢B4­C6ºC;ÁD>ÁD>»EA¹PL²[T¥e\‘f]u_T[UIGMACI?<92?82>71;6094.74-63.43.43.34.24/13./4..3-.3-.3--3/-3/-3/-3/.3/.3/02/02///-00.22022010./.,0/-3/,8,,;,)C0*K70S<6^IHtbn‡z”�Š¶Œ�Æ„ˆÏz€ÌrxÆik¶WWŸID„E=nG<dD<aC>^CAY@CV@DP>EKGQRKWUQ^WU`XS_UR^TT^SY_S^[LaZJaZJ,-/,-/,-/,.-------.-+.-+/.,/.,1-*0,)0,)0,)/+(/+(/+*/+*/+*/+*/+*/+*/+*0,+/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*0,+0,+1-,1-,2.-1-,1-,1-,1-,1-,1-,1-,1-,0,)1-*2.+3/,3/,3/,3/,3/,3/,3/,3/,3/,40-51.62/74/80-92,:3-;4.=4/>5.>5.>5.?4.?5,B5-C6.D7/F90G:1F<2G>5H@5J@6P?5T>3X<1^90b6+m9.t8-|8+ƒ9,�;/“=0˜?1ž>0§A3­A4µC8¾E:ÊG=ÔG>ÞE?åC@è@?êBAæDAÚE>ÈD8·B1ªA.¢B,¢A.¡@-¢?,¢>.¡=-¡=-¢>. ?/œ<,š=,˜<-”=,“=,�=+Ž=*Œ>*‹<+‹<+Š=-‰<,‰<,ˆ;+ˆ:-‡;-‰=/†</†</‡=0ˆ>3ˆ@2‰A5‹A4‘E5—D4£E9±I>ºG@»D>»EA¸MG´ZR¯f_£qf‘sh~rdjj^V^SIQFLLBJF=B>5<8/95,74+63,33+43.34.14-14-02-/1,,1+,1+-2.-2.-2.-2./1./1./1./1.02/02/11/11/11/11/11/40/4+0;/3A32C4/J;6]OOymy‹…Ÿ“�»”–Ï�•Ùƒ‰ÓtzÆjn·_b§Z[”LItHBdA>]>>X?BUAIVLU\U`bbqnn}xv†|rulyoguh_k_W_P]\Hb\Fc]G,-/,-/,-/,-/------.,-.-+/.,.-+0,)0,)/+(/+(/+(/+(.*).*).*).*)/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*0,+0,+0,+1-,1-,1-,1-,1-,1-,1-,1-,1-,1-,0,+0,)1-*2.+3/,3/,2.+2.+2.+2.+2.+2.+3/,40-51.62/80.91.:2/;4.=4/>50>50=4-?4.?4.A4.B5-C6.E80G:2H;3H>5H>5L=6O>6Q=4V;2Z90_7/h70p7.w7-�9-‰9.�<1–<1›=1¢@3§A3­C6´D8¾E:ÉF<ÔE=ÛD=ßD@àE@ÞFAÔG>ÄF8²C2¥B-žB-ŸC.žB-Ÿ@.ž?-ž?-ž?-ž>.ž>.š=,™>,–=-”=,�=+�>+Œ>*Œ=,‹<+Š=+‰<,‰<,‰<,‡;+‡;-…<-†</…=/…=1ƒ=1„>2†@6‡A7ˆB6“G9˜F8£G<¯J@¸IB¸GA·LF³TN±d\ªqfŸ~o�o|mmseZfZNXMLNAKI=EC7@=4=:188.44,11)23-23-03,/2+/1,.0+.0+.0+/1.-2./1./1./1./1./1./1.02/02/11/11/11/11/11/2015+49-7<23?53H?:^VTxr|Šˆž““¹”—Ê�–Ô„‹Ïy�Âqy¸kt­hnž\_€XZqSUjRWjT^hZgmfvvr‚tˆ~’‡ƒ•‰~Ž�w‡zr€qfteZeT[ZE`Z@b\D,-/,-/,-/,-/,-/,.-------.-+.-+.-+-,*/+(.*'.*'.*',+),*+,*+,*++)*+)*+)*+)*-+,-+,-+,-+,-+,-+,-+,-+,,*+,*+-+,-+,-+,.,-.,-.,-.,-.,-.,-/-./-./-./-./.,0,+0,+1-,2.-2.-2.-2.-1-,1-,1-,1-,1-,2.-3/.40/51.80.91.:2/;30=31=4/=4/=4/?40?4.A4.A4.C60D71F93F93H;5J;6K<7N=6P;6S:5V72[6.c60k6.t5,}7/‡9/�;0”<0–<1�?3 @4¢@3¨@3±C6ºD8ÅE:ÍD:ÕF@ÖG?ÔIBÌI?¾E:®C3¡B.œA.�B/œA.œ?.›>-›>-›>-›>/›?0˜<-–=-”<.“=.�>-Œ=,‹<+Š=+‰<*‰<*‰<,‡;+‡;-…<-„:-ƒ;-„<0‚<0‚<2‚>3ƒ?4…A8‡C:ˆD9”J=—H; H>«JC²KF³LG²SM®`V­sg¦�q�Œz‘Ž{‚‰ws€ocqbXcUNRDMN@HI;DD8@@49;.46+/1&01)01)/0*/0*./*./*//-//-//-.0-//-//-//-//-//-//-00.00.00.00.00.00.00.1/26+97+98/4;63HE>_^Yzz|‹�š�”±’—¿�•Ç„ŽÃ}‰»{‰¶|‹²}Œ«}ˆšwƒ�q~‡n}‚n~~pƒyŽ…ƒ™�‹¡”‘¨˜�¥”ˆ�Šƒ–ƒ|�{j{i\hTXX@]Y<_[@-.0-.0-.0-.0-.0-.0.......,-.-+-,*-,*.*'.*'.*'.*),*++)*+)*+)*+)**()*()*(),*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+-+,-+,-+,-+,.,-.,-.,-.,-/-./-./-./+*0,+1-,1-,1-,1-,0,+0,+0,+0,+0,+0,+1-,2.-3/.40/91/:20;31<41=31=31=31=4/>3/>3/@2/@3-A4.C60D71E82F93H94I:5J;6L:6N94Q83T50^72e60o6/x8/‚90Š:/�<1”>1™?4›?4›?2Ÿ?1¥A2®B5¸D7¿E:ÇG>ËH@ÊJAÃI>¶F:ªB5žA0™@.™A-™A-˜?-—>,—>,™>,™=.—>.•<,”=,“=.�=-Ž=,Š=+Š=-‰<,ˆ;)ˆ;)‡;+‡;+„;,„;,ƒ;-€;,�;/€<1€<1�>5ƒ@7ƒC:‡D<ˆE<”L@–H<žG>§JC®LI®QL­]Vªj^§€o¡�yšš‚�›ƒ†•€z‹xm{jbm]SZJQUFKO@EI:@D6;=057,13(01)/0(./).-(.-).-)/.,0/-/.,/.,/.,/.,/.,..,/.,..,0/-//-0/-//-//-//-//-2-17,:6*83-1961HJ?bfX{€z‹�“�“£�–°Œ•¶ƒ�²��¯…™²�¤¶’¨³“¨«�¤£Œ ž‡ž˜‚™�}˜‰�œ‹ˆ£�”°š˜±›”­—‹£�…›„~‘{k|iXfOSV;ZV9^Z=+/2+/2+/2+/2-.0-.0-.0-/.------.-+-,*-,*,+),+),+),*+,*+,*++)*+)**()*()*(),*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*++)*+)*+)*,*+-+,-+,.,-.,-/-./-./-./+*/+*0,+0,+0,+0,+/+*.*)/+*/+*/+*/+*0,+1-,3/.40/901:20;31<42=32=31<20<20=20=2.?1.?1.@2/A30B5/C60D63C84D95E:6G96H94K84N50X72_60i70r7/}:1†<1Œ>2�>2—@6—?5—?5—?3œ@3£C5«C6³E8ºE;½G=¾H>¹G=°D8¥@4›?0—@/—@-—@-–?.•>-”=,•<,–=/–=/•<.“;-’</Ž<.‹<-‰<,‰<,ˆ<,‡;+†:*†:*†:*ƒ:+ƒ:+‚:,€:.€:0€<1€=4?6‚B9ƒD=‡F@ˆH?‘KA’H=šG?£LE¨PL¨XQ©f]¦vh¢Œw›™€•¤‡Œ£‡†�ƒ~‘}t‚qjue\eRV_LNUCEL:?F6<@27:/58-01)/0(.-(/,'/+(/+*1-,2.-1-,1-,1-,1-,1-,/.,1-,/.,2.-0/-2.-0/-0/-0/-0/-2-18,:5)51++66,GL8`hQzƒpˆ’‡�—–�š �™¥†–£‡›¢�ª©›¹± À³˜·¨š¶§›·¨˜·¥‘±œ‰«’‰«��±”’³–•´•�­�ˆ¢…„›�{’xgzdTbIQT9VU9XW;,03,03,03,03./1./1./1./1....../.,.-+.-+.-+-,*-,*-+,-+,-+,,*++)**()*()*()+)*+)*+)*+)*+)*+)*+)*+)*,*+,*++)*+)*+)**()*()*(),*+,*+,*+-+,.,-.,-/-./-./+*/+*0,+0,+/+*/+*.*)-)(0,+0,+/+*0,+1-,2.-3/.40/:12:12;31<42=32<21<20;1/=20<1/>0->0-?1.@2/A4.?4.@51@72@93A:4B94C84F74H5/Q51X5/a6/l8-u9.€</†>0‹=1“?5•>5“?5“?4•B4šB4 C4¥D4¬B5°D8´E:±E;ªB7¡?4š>1–>0•?.”?+“=,“=,“<+“<+”<.”<.“;-’<-�<.�;-Š;,‰<,‡;+‡;+†:*†:*…9)ƒ:)ƒ:+�9*9-9-€<1<3~>5€A8‚C<…F?ˆIBŠICŽJA�G>—HCŸNJ§VR§`Z©pe¤ƒrœ”}–¡ƒ�¨Šˆ§ˆƒ „~–~yˆup{jdp\]iSR^JJS@BK:>E5:@29<134,22*1.)/,'/))0**1++2,,1++1++1++1++1++0,+0,+0,+1-,1-,1-,1-,/.,/.,/.,2,.8*75(13+(56&EK1\gGu‚d†“yŽ›Š‘ž”�Ÿš‹ž˜�¥——´ž¢Å¥©Î­¥É­¦Ê°§Ë±¤È¬—¾Ÿ‹³‘ˆ°��²�Ž²Ž�²��¬Š„¡‚€™{wŽrdx]Q_ENR7QQ5SR6,03,03,03,03./1./1./1./1/////////////.,/.,.-+.-+/-..,-.,--+,,*++)*+)**()+)*+)*+)*+)*+)*+)*+)*+)*+)*+)*+)**()*()*())'()'(+)*+)*,*+-+,.,-.,-/-./-./+*/+*/+*/+*/+*.*)-)(,('0,+0,+0,+0,+1-,2.-40/40/:12;23;23<34=32<21<21;10<1/<1/>0/=/.>0-?1.?1.>3/=60;60;83<94=:5>93@72C60G4.O4+Y4+d5+n8,x:-;.…;.�<4‘<5�>3�@3�A2“B1—A2™@0 >1¦@4ªB7ªB9¥@6Ÿ>5—=2•?2’?-’?-‘>,�=+‘;*‘;*’<-’<-‘;.‘;.Ž;-�;-Š;.ˆ:-†:,„;,…9)…9)ƒ:)‚9(‚9*�9*~8,~8,€<1€=4€@7�B9„E>…IAˆKFŒMFŒIAŽH@˜JF¡SO¨]X©hb©{n¤Žy™œ�’§†‰¬‹‚ªˆ€¡„™€~‹wx�njxadr[ZhQQ]IITCCL;>D6<?467/44,30+0+'1()1()2)*4+,2)*2)*2)*2)*2)*0**0**0**1++1++0,+0,+0,+0,+0,+1+-6)25)-4-%46!AH&Wc;q€W‚’m�œ}� †�¢Œ‹¡Š�¨‰•¸Ž È”¤Ð�©Ó«¨Ò®§Ñ­¡Ë¥“½—…¯‡‚¬„‡°†ˆ®…Œ±ˆ‹«†�Ÿ{{”tqˆk_sWM]BIO3JM0KK/./1,01./1./1.0/.0/.0////00.00.00.0/-0/-0/+/.,1-,2.-1-,1-,0,+/+,,*++)*+)*+)*+)*+)*+)*+)*+)*+),+)*+),+)*-(,*(),())'()'((&'*()+)*,*+,+)-+,.-+/.,/.,-,*-,*/+*/+*.*)-)(,(',('1-,2-*2-*2-*3.+4/,50-61.;31;31<42<42=31<20;1/;1/<1/<1/=/.=/.=/,>0-?1.=2.=4-=6.;819919:29:2;81?6/C2+J1,T2)^3*g7+o9-u=.z</‡=4‹=3‹?2Œ@2Ž@3�@1’?1•=/˜</œ>2¢B6¡C7žB7™?4–>2”>1”>/“=.�=-�<,Ž=,Œ=,Œ=.�>/Ž<.Ž<.‹</Š;.‰;.ˆ:-†:,„;,‚:+‚:+€;+~;*~;+|:*}9,|:.€>2�>5€@6‚C:ƒG=…JB‡NE‹OGŒICŽID“PJ˜\T�h`�uiž‡už˜€”¡…�ª‰‰¬‹ƒ¨‡€¡„~›z�ytˆoj~cby]\pUUgOO\HGRAAH8=A388.85.7/,3+)2()2()3)*4*+0*,/+*0*,0**0*,0*,2),2),3*-1+-1+-1+-0,-0,-0,-1+/4)/4*+4-%46!?F%T`8m|Q��g‰™t�žz�¡~ˆ¡zŠ§{‘¶‚›ÄˆŸË�ŸÉ™žÇ��Æš˜Á•Œ·Š‚­€…®‚Œµ‰„«‰®…‰©‚~œxv�olƒfZnSJZ?GM3FJ1DF./0+.0+/0+01,01,01,12-21-32.43/43/62/51.41,3/,4/,50-50-4/,3.+2,,1-,0,+0,+.,-.,---/--/,,.++-*).))+.)/.)-/(/.)-/)-.)-.*+.*+/+*/+*-,*.-)--+-.)./*./*------.,-/-./.,0/-2.+2.+2-*4,)5-*6.+8/*:1,:1*;2+=4-=4->50>50>50=4/<3.;2-=2.<1-<1/;0.=/.>0/>0/@1.A0)@2);4*77-39-19.47.83-?-)F*'L)'S*&Z/(`5,d<0k@0yA2@/†A2‹B3�@3“=0“;/’8-“;/“=.”A1•B2”C2’A0‘>.‘;,—;0—:2’;1Œ=0ˆ?0„?/‚?.ƒ>.ˆ@1ˆ?0Š>1Š<0‰;/ˆ:.†:-ƒ;-{9+~@3w<.q7)w>-w>-v;-}?2{;/‚@4‡E9‰I?ˆLA†MB„PC‡NE‘KI—SP”bY‘paŠ}j‡‰q‰–zŠŸ€�¤‡�¦‰Ž§Š�§Œ‡¤ˆ~Ÿ‚uš{o—ui�ib„_[zXZsUZnSWeNPWEJK=C?6@93;0.6**3')3'+3'+0'*++-)+***,*(),'+.(,1(-2'-3(.3(.3(02(00)00)0/*1/*0/)-1++0-(//#<?*V\@pzX‹g�œu‹�sŒ¡vŒ¦wŠ¨v‰ªuŽ´{™¿†ŸÁ��¾�–ºŠ�³ƒˆ®}…«|ƒª{…ª~ƒ¨}ƒ¦~€ž|y”ur‰mh|cYkUMZFIQ<CH4?B//0*01+01+01+12,12,21,32-43.43.74/74/63.52-50,50,7/,7/,6.,6.,3.+2-*0,+0,+.,-/-./-0--/-,1+*/)*.)(-.)//(//(//(//)-/)-.*+.*+/+*/+*.-).-)./*./*./)./*.....0....../.,/.,1-*3.+5-*5.(7.)8/(:/);0*<2)=3*>5,>5,>5.>5.>5.=4-<3.<3.=2.<1-;0.;0.=/.=/.>0/@1.C0*C0)A2+>4+:6-95,:3-<1-?-+D*)I*(N+'T/)Z5-_:1e=1pA1x@/€A0ˆ@2�?4‘=3“:2“;1�9.‹</‹>.Œ@0�A1�A1ŽA/‘?1•;2•;2�<2‹=0†>0ƒ@/�?/‚?.…@1†>/‰=0ˆ<.‡;.„;,‚:,€;,x8,x<1s9-n9+s>.s>.r8*u:,}=1ƒA5‰E:‹I=‰K@†NA„PC‡NE‘JH—SR’f]Œwfƒ‡p“w€Ÿ„¦…ˆ§ˆ�§Š�§‹�§�Š¤‰�Ÿƒwœ}o˜ve’i`‹`Y‚ZY{X\wXZnSSaJNUCFH;C@7<737/-3*+2)*1(+.(*+++)*,+),+),-(,/)-2(02(03(02(02(10)1/*1-*1-*3-*12-12.-0,)--%8:,SXBox]€‹i‡–o†™l†žnˆ£p†¤p„¥pŠ­w“¶€™¹‡–¶„“³��¯|†«x‚¨w‚¨yƒ¨|ƒ¦|‚¤�{x’uqˆnh{eYjWMZHEL<@D5;=/12,12,12,12,23-23-43.43.54/54/85085085074/72.61-80-80-7/,6.+4/,3.+1-,1-,1-./-.0.1/-0--/,,.+*/+)./(/1'//)-/)-/)-/)-.*+.*+0,+0,+/.*/.*/.)0/*/0*/0+//-///0./0.//.,/.,1-*2-*6.+70*90+:1*<1+=3*>4+?5,?6-?6-@7.@7.?6/>5.=4-=4-=2.<1-;0.;0.;0.;0.<1/?1.C2+E0+H/+K--L,/K+.I*/E+.A-,@.*A.(E.(M.)X1*b3-g5.j:,o;-w;0�=4‡<6Œ<5�:4�:4Š<2‡=2†>2…?3†A2‰A3ŒB5�C5Ž?2�=/Œ>1Š>.‡>/…>,†=.ˆ<,‰;.‰<,ˆ:-†:*ƒ;,<)~=+{<+}=1z<1v:/u;/x@1x@1v<.v;-?5‚B8‰F=‹H?‰JA‡KA‡NCŠNF‘JH˜TQ–f\�ve‡…n�‘t�{�¥�…¨‡ˆª‰�ª‹�©ŒŽ¥‰ˆŸƒ�š}y˜xi–ma’e\‹a\†`]‚`[|]UrVPhPFYEBP?:D93:2.3,*/)*,)**(.((1&*1&*1&*0'*1(+2).1+//*.-+.,+0+,0)-0(-0(-1)-01/23/./+(**"57*PUAmu]|Ši€�k~“j™jƒŸn€ n~ m€¥q‡¬x�¯|‹­zˆªw„©u�§t§u¦w¦y¤y€¢}}›yw‘tp‡mh|cYkUMZHDH9=?299-23-23-34.34.34.45/54/54/650761961:72:72961940940:2/91.91.80-50-4/,3/.3/.3/03/01/01/20.1..0--/-+.0)02).0*.0*.0*.0*,/+,/+*1-,1-,0/+0/+10+10+01+01,11/1111/010.10.2.+3.+3.*92,92,;2+<3*>4+@6-@6-A7.B8/A8/B90A8/A81?6/>5.>5.=2.=2.<1/;0.;0.<1/=20=2.B3.E2.L..R+0V'/U&.P'/I).C/.<1+;2)?2)H1)R/)_.*d/)i9/k;/u<3}<6…;8‰;9Š;7ˆ:6‰>8†@8†B9…B9…C7†B7‡A7‰A5ˆ@1ˆA/‡@.‡@.‡>-ˆ<,ˆ;+‰:+Š8*ˆ9*‡:*ƒ:)<){=(x>(x>*‚>1{7.z7.z<1v;-w=/|A3{@2€B7ƒE:†H=ˆJ?‡KA‡MBˆODŠQHŽSK�\QŽh[‹tb…g€‹m~–vž|ƒ§ƒ„ª…‡ª‰‹ªŠ‹¦‡‡ ‚ƒš~}˜yq•oi”if�gfŒeg‹gd†e_~_Zw[PhPK^JBP?8D62:/.4*,/(+*%2&&5%(4%(2&(1'(/)+/+,.,-+++*+-*+-(,-(,/',/',/*+-.*+1++0+'.+"88,PUAiqYvƒe{Œhy�h|–iœl}Ÿmyžkz m¥r‚§t€¥q}¤o}£p}¥s~¥v~¥x}¤x|¡x}Ÿzz™wu�qn‡je|bXlSN[GDF9?=1:8,45/45/45/560560560761761872872;83<94<94<94<73<73<41<41;30;3083072/61.61.5106216213123121/00./1-.2,04+.4+.4+.2,.2,,2,,2,,3/.3/,3/,3/,21,21,32-32.22022032032051051.61.61-;4.<5/=4-?6-A7.B8.E8/C9/D:1D:1E;2D:1C90B8/@5/@5/>3/>3/=2.=2.=20=20>31>31@51E31M02T,4X)3W(2R)1K,1B30:6-77+:6*B4)M2)X/)^/)f:1j;3s<7z=:‚<:†<;‡;;†::�:6�>8�A8�B9ƒC:ƒC:…@9„@7‚C2‚C1ƒB0„?/†=.ˆ<.‰:-‹9-‰7+‡8+…9+‚:+~=+x>*v?*x>*�9-|/'‚8/„>4w4+s5*}A6}C7E:€G<‚I>ƒJ?„KB†MD‡QG…WJ}`N{iQ€pY�u]‚|bƒ†i†“uˆ�|ƒ£~�§€‚ª…‚¬†�©„¥€{ž}z™wz’pz�lx�lwŒku�mr‰lm„gkd`rX[iRR[HHL=@@4;7,70&5*$6('5''3''1'&.)&+*(++)+-*',('+*)+*)+***,+),+),-)*,#$2*'50*86*BC5UZFfpWn}^tˆeqŠbu�ex˜ivšjs™htšiwŸkz¢nx lwŸkx ny£q{¥u{¥w|£wyžuzœww–tsŽom†hd{_WkPN[GCC7>:195,560560671671782782872872983:94=:5>;6>;6>;6>95>95?74?74>63=52;63:5294194184184395484373243151240/6-06-.6-06-.4..4..4..4/,40-40-40-51.32-32-43.43.43/331542540841850940:5/=60>7/@7.A9.C9/D:0G:1H;2F<3F<3F<3F<3E;2C90B71A60@51@51?40>3/>31?42@53?53@72C52I35P16T/6S.5P05J22C52=90<:-=:+C7)I6(Q3)W2)]2+d3,l50v64}77‚87ƒ77‚66~75}:4}<6}>5€?9ƒ@8†?9…A8€B3€C1�B1ƒ@0…=/‡;.ˆ:.‡9-…9,ƒ9,‚:,<,|=,y>,x?,|=,‡5*‹2,¡LE¨XQ‹A8|90‚F;€K=yH:zJ<{M>|N?}OB�QE‡UJ‚_LrkOosR|uX†v\�yb˜iŸ‹s ”z––z��|‹£ˆ¨�…§‚€¤~{Ÿy~™vƒ‘p…ŽoƒŽp€�p�p{ŽpwŒms‡kj~bfv\_hSV[GOM>GA3@6*=0';,'9+(6+'3+(/,',-'+.').().(+-(-,*/+*3)*4(*7'*7'(3($<3,E>4IG:QR@]bKgqVjyZn‚]k„\l‰]p‘bq•eo•do—eršfuŸmrœjq›itžnx¢rz¤vy¢vyŸvvštw™vu”ro�kj…fc|^UlON\ECC7@91;4,671782782782893893983:94:94;:5>;6?<7?<7?<7@;7@;7B:7B:7A96@85=84=84<73<73<73<74<74<74;639529338308/09/.8/080.80.80.61.61-61-61-52-52-63.63.74/54/540540841952:63<94=84@93@70A8/C90D:0G:1H<0I=1I=1J=4J=4J=4I<3F<3D:1B8/A7.A81@70@72?61?61@72@72A83?74@85B86D97G96H96H96H94E80E8/E9-E9+G9,I9*K9+Q7*Z/&d/'n3-z63ƒ98‰;9‹;:‹=;‹A>‡@:‚=6<3:3‚<4ˆ=7‰@7ƒA5�B3‚@2ƒ?2ƒ=1„<0�;/€</~<.|=.{=.|>/|>/}>/=/†9/�1+£;8ÎkfÛ~y­ZTŒC<ˆLAN@tJ<qM=pQ?qR@tS@zWD‚[J~eOmsOo{U„y[˜u_©oc¶mf¾qk½wo»‚w±Œz§—~žŸ€–¡��¡~…�y…™vˆ‘r‹�rˆ‘r„“r€–r|–sx“pt’pm‹ii„edx\]kQV^GMP;ED0B;+@3+?2,;0*70*30)00(./'./)/0*2/*6,+:*+>(+C&+E$+C&(F5-LC4VP@[XE`bLgmQiwVj{Wl‚[g‚Wf†WlŽ\o”an”an–bršfs�ko›jo™irœnvŸsy¢xxžwu›vs—su—vs’rn‹li„cb{[TmMM]CGH:E<5@707827828938938939:4:94:94;:5<;6?<7@=8@=8@=8A<8A<8C;8C;8C;8B:7?:6>95>95=84>95>95>95>95=85<73:51;30:0.:0.91.91.91/91.91.72.61-61-63.63.63.74/74/74/540651952;83<94?:4B;5B;3A8/B:/C9/E;/H<0I=1J>2J>2K>5K>5J=4J=4F<3E;2C90B8/B92B92A83@72@72A83B94A:4?82@93B;5D=7F=6G<6K<5N;4M6.N6,Q6+Q6+Q7*P9+P9)V6'f6*r6,~;3‰@9•D@›HDŸJGŸLF QJ™LDŽD;…;0„7-…7-Œ91�=4ˆ>5‡>7†=6…<5…<5�=4}=3z>3x@3vA3x@3z>3<3ƒ:3ˆ73’42˜(&³=;í{zþ•’Åhc–G@‰K@xH:nM<jQ=fT>hV>lX@t[E`L€hPysQ„wUžt\´l^É__ÓV\ÙQ[×T\äouÙzx̆|À�~µ•€«•}£’xŸ�vœŠrœŠt™�u•�u‘‘u‹’s†‘s‚‘r|‹ny†ju{amoWgbN_TBUE5R</O4-M1-I0,D/*>/(9/&7.'6/'81):/)=.+A,+F)+H(+K'+H)'TB4YQ<d^HgeLilOnuVm{Xl~Xk„Zg„Vg‡Vm�]p•bq–br™du�it›lo™ko–jr™mw�vyŸxwšyt–us”uu”us�qoŠki„ea|[TmMM]BMN@KB9F=69:49:49:49:49:4:;5=<7=<7=<7=<7@=8@=8A>9B?:D?;D?;E@<E@<E@:D?9D?9C>8E>8D=7B;5B;5B;5B;5B:7B;5A:4A:4A83A83A83@72@64>71>71>71<71<71;60:5/96/85.74/74/961961961:70<71=82A:2B;1C:1D<1F<2J>2K?3L@2N@3N@5N@7N@7L?6K>5I<3H;2E;2E;2B90A8/@91?80?80?80@93@93<5/MD=M@:K:3T@9R62O0+a<6\3-`5.`4+^/%^/%f6,m=1q=0ŠJ>�G;—I?ŸKAŸF>›>7›>7žD;¦ND¤PE¨VJ«YK¥OB™A5—;0›>6“98‘98Œ65†52†84„?8|@6r>1rB4oA2q=0v:0‚72Œ43’-1š',ÈHIèbaÅEDÍWUáyv«VO†F<„VFmR=i[AibFjeHj`En^D|dL�jT’bL²m]ÑnhÛX]à;Lç-Eò+Hõ0Lö@YìI\ä[eånrâ{|Û€}Ûƒ؉‚Ãv½�wº„x¶†x²ˆx­Šw¨‹y¥Œxš€o™~m˜xi“oa�bY„UM{IBxA>u:<i.2d04a03P&'G%#L0-M51G4.E2,D1-F1.F.,F*)K--Q6/YH4`W:f_BgdEkoNu|Zu„]o�YlƒWl‡Zp�]r’as•bu—dw™fv›hq•is˜os—qq•os•tw™xz˜|y–zx•ys�rm‡jk„fi‚dazZTmMO_DNO?SK@PG>9:49:49:49:4:;5;<6>=8?>9>=8>=8A>9A>9B?:C@;D?;D?;FA=E@<E@:E@:D?9D?9E>8E>8E>8E>8E>8D=5C<6C<4D;4C:1B92B92B92B92B94A83@93@93>71<71;60;6096/85.74/74/96196/96/:70<71>:1A:0B<0C;0E;/I=1J>0L@2L@2N@3N@3N@5N@7M?6K>5I<3H;2E;2E;2C:1B90@91@91@91@91A:4B92D93J71L/+V.,j76u99{;;…ECƒD?„H@…I?†H=ŽJA™NH¡PL¢OG“D7”B4šB6ŸC8 >3œ7-œ7-Ÿ=0£C5›?0™@0 E3¤G6§F6®H:¶NE´LM¯HL¥@D™9;‘98�>:ƒ@8x>3o>0zI;…LA„?8„,+’).³9DÒLWÜKNßMMÊ>=À@?Ð`\³YQ„@5‚TDyaIe[@_^@ihIslOvgJ~dK–cN¼dXÖ_[æUZëANò,Eþ%Dÿ&Hÿ'Iÿ.Nù3Nò<RðIZëP^åQ]ãS]àXbÛbiÕflÔjnÒmqÏqrÍutÉxwÇywÊ||ÉyzÈvxÇqtÆkrÅfnÅakÃ_kÅ`n³S_¥KWšGQ�DK†EIm69HT1-M0*H1+J6/K81K:2O>6UE8[O5cZ9gaAifEnrOz‚]z‰bt†^r‰_p‰_p‹`qŽ`u’dw”dw”dw”fs’ix–rz—xy–x~˜}ƒ�„…�‡ƒ›…�™ƒz’zr‰om„hh�d`y[TmMNaERSAVPBUOC8938938939:4;<6<=7?>9@?:@?:@?:C@;C@;C@;C@;D?;D?;FA;FA;FA;E@:E@:E@:G@8F?7JA:I@9H?8G>5F=6E<3F<3E;1D:1D:1D:1D:1D;4D;4D;4C<4?80?80<71;60:5/:5/94.94.96/96/:5/;7.>7/@9/B90C;0E;1F<0I=1K?1M?2M@0P@1O?0P?5M?4L>3K=2I<3H;2E;1D:0C:1C:1A:0A:0?;2@<3@<3D;4UD<T3.j23ŽAG­LWÃUbË]fÆ_c›B>—I?–LAœNB¬RJ¾URÌOSÄLK¥@4�B0 @0¥?1¥;.£7+¤8+§=/«E6¡>+˜7$�:'¥>-¬A/²C2¶D:§02¯:B¹DL½LRºMR­KL–A>‚71{8/v6,x2*„1-œ37¹=GÕGWèM[âFIÏ53Ð:9½31ÈNKÆc]”J?Œ[J}cLj_CgdEtoOpQ�hJ�hN°hYäbbõQZôDQö7Iý1Jÿ2Nÿ/Mÿ+Iÿ2Oü0K÷1Jõ5Lò4Mð2Kó3Ló7Pò@XïC[îF]íIaëNcêQeèUhèVkåSjäRiãOiäNiçNlëPoòSsóVuþh…ña|å[uÕUlÍYlÉdr¨R]r-2^&'Y0,W:4Q?5B:-:8)>B1LN9VO3aV6e_?heDquT†e~ŒiwˆdxŒiu‹er‹dr‹cvŒey�hz�iz�j€•v„›~Šž…� Š’£�—¨–˜©™–¦™‘¢’‡˜†z�ypƒmg~b^wYTmMPcGSXDXUDYVG7827828939:4;<6=>8@?:A@;BA<BA<DA<DA<DA<DA<E@<E@<GB<GB<GB<FA;FA;FB9HA9HA9LC<KB9JA8I@7H>5G=3F<2E;1D:0D:0E;1E;2D;2E<3E<5E<5@91?80=82<71;60:5/:5/:5/96/96-;7.;7,?8.@:.C;0D</F<0F=.J>0K?/M@0M@0O?/O?0O?2O>4L>3K=2J<3G:1E;1D:0D;2D;2B;1@<1@<3A=4B>5G=4S81[*&‰:?¾T`ÙTeæTgãUcËJOŸ30–:/‘>0™A3®F=ÅIGÕBHÍ>@±B7§F5©B3¬@3¯?3°>3³B4´D6²G5©B/£<)¤;(©<(«:(®6&­3(¶97º9=½7>¼7<¿<BÆJLÄTS¾XS™?7Œ4*‰,%�61ÃHKÜOWâAQÛ3@Þ:9Í.*Ó84È74ÆE@È]U©WK�WFwW@veI~pSpQ„fJ�dK©o[ÕuiîSWüDP÷;Jø9Iþ=Nÿ@Qÿ:Nù4Hÿ?Sþ9Mû4Ký3Ký1Jÿ/Iÿ0Nÿ6Sÿ3Tý4Vý6Wú7Wø9Zø=\÷>^õ@aøCfö@fõ?eõ?gø?hüCmÿErÿIsÿOuÿQuþUvòSqçTnåaxÖfv·Wb}15j0/Y2+P8,G>/>A.;D/?G/PK._T8g`CokN||`ˆŽr‡‘v}‹qz�q|�s|�sz�ozŠm}Œoƒ’uˆ—z•£Šœ©•£®�¦±¡¬´§±¹®°·°«µ­¡«¢•¡•ƒ‘‚s„qh|c]tXSlNPdHRZCWYDYZH671671782893:;5<=7?>9@?:BA<BA<DA<DA<EB=EB=FA=FA=FC<FC<GB<HC=HD;HD;JC;JC9LC:KB9KA8J@6J=4I=1H<0G;/E;/E;/E;/E;1E;1E;1D;2D;2@9/@9/@91?80<71;60;60:5/;7.;7.;7.<8-?8.A;/C;0D</G=1G>/K?1M@0M@0NA0O?/O?0O?2N>1N=3J<1I;0G;/D:0D:0D<1D<1B;1@<1A=2A?3B?6J<3\5.w32­LSÛ]iæM_å?S×6E·&+¨1+�7(’:&—<)¬@3ÃD=Ô>?Í;;±?5¨B4ª@2­?2²@5¸C9¼H;»J<¬>/ª?/¨=+¦;)ª9)°:,·;/¼<1ÊFAÎDAÐ>?Ð79Ö7<ÞBEâLNÝROÄC>ÆKCÍTLÔSMÙKJÛ?Bà5>á27Û4.Ú7.Ð1+Ï:4¿84³@9µ[P–UC{N9‚bI†kP„`F‘ZE­fTËrdçmhêEKò:Dê9Cç<DçBHçFKèGMêFMôJSóAMó8Iö5Hù2Gý/Fÿ2Lÿ8Rÿ9Vÿ9Xÿ:Xý:Zþ;]ÿ=_ÿ@bÿAeÿCiþDkþDmüCmüCoüCoüBqþCpÿGpÿ@hÿGmÿMpòIhéOkå[rÙcs½YcŽ?D`&$Q+"V@3VO=IL7@D-RJ3eYCujV€{gŽ�y˜œ‹”�Œ‡”ƒ‚Ž€‰—ˆ�›�Œ™ˆ‡’‚Š“‚—žŽ¤¨™²¶§¸¼®¿Á¶ÀÁ¹ÂýÅÆÁÃÃþ¾¾°µ±¢©¢�—Œy‡vi{e\sYTmORfJQ[BUZCW\F560560671782893:;5=<7>=8@?:@?:C@;DA<EB=EB=GB>GB>GD=GD=HC=ID>IE<IE<KD:LE;LC:KC8LB8KA5L@4K?3J>2I=1G=1G=1F<0F<0E;1E;1D;2C:1A:0A:0@91@91=82<71;60;60;7.;7.;7,<8-?9-A;/D</E=0H?0J>0L?/M@0NA0NA0O?/O?/O?0N>1M<2L;1I;0G;/F90C9/C;0B<0B;1@<1@>2A?3B?6N;4m84—FEÄY_ÛWbßCQÞ8FÔ3;½++­1'¡8%˜:!—9 ¤;&¶>.Ã:2¾71§7,¢:-¥9-§7,¯:1¸B8ºE;·E:®>2­?0§<,¤6'ª6)¶>0ÃD;ÊG=¾8-Î@6áD?ìBBô=Aó9>ë27Þ.0Û75âGBèSLåNGÜ=9Ù10ã/2ê67Ù1(Þ</Í,"Ð71¾3,°7/Ég\¸l\ŽWC‡ZCƒV?‰R=©\LÎlaágbäTTðEKñ=Fä;@Ù=>ÒA>ÒGDÝOMëWWíNSïDMò9Gö5Fø1Fø.Dû1Gþ7Nþ8Sý8Tý8Vþ9Wÿ;\ÿ>aÿAeÿCkþ;eý<hý>jý?mú>nø>oô=mö:jÿHrÿ;aÿ?dÿKnÿKn÷NmïUoãZnèr‚Ä`j—HKt:8\6-P9+PC2UM:^QAreUˆ|n•�‚Ÿ�‘¦§Ÿ §Ÿ”�˜•ž›ž§¤¦¯¬¦¬¨¢§£¦¨£µ´°ÂÁ¼ÌÉÀÑÎÅÕÐÌÕÐÍÔÎÎÕÏÑÑËÏËÆÊ¿½À°²¯˜Ÿ˜€Œ~m~k`t[WnRVjNS`FV`EX`H560560560560671893:94;:5=<7>=8A>9C@;DA<FC>HC?HC?GD=HE>ID>ID>JF=JF=MF<MF<ME:LD9MC9MC7NB6MA3N@3MA3JA2JA2I@1H?0G=1F<0D<1C;0C:1C:1A:2@91?82?82<71<71<8/<8-<8-<8-@:.A;/D</E=0H?0K?1M@0NA0P@0P@0O@-O?/O=/O=1M=0L</H:/F:.E8/C9/B:/A;/A:0?;0?=1@>2@@4P91r1/¨JKÀRU»>D½06Â03Á//º1)©1!£9#œ=!™;›< ¡=#¥;%¤6%ž6) 8-¢6,¥4,­81·@:¸C<²@6±B7­A4¨</¦8+¯:0¼C8ÃD=Å@7Ã>-Ï>-Ù5+á*&í"%ú%+ÿ*1ÿ.4í)*ç.+á3,ß4,à3,ã2,é0-ç2+Ú2%Ö6&Ð2&Ï7,Å6.½>7È`UÒq¦eS�VBŠM:£WIÈf[ßd_çRTèCIõFMí@Dß??Ô@<É@8Æ?9ÑFAãMLïJPô@Kù9Hý7Hü5Hø3E÷6Gú;Mú8Pü7Rû6Rý6Uÿ6Xÿ9]ÿ;aÿ<gÿ>kÿ?mÿ@qÿBtÿBuþBuüBtùBpÿFmúAaþEeÿKjþKkÿVtÿ_zõ]tòj~æp~Ónv¯]aƒCAg6/hB7sUJye\Œ}v£–�­£¡²®«¶¶¶²¶¹¨°³°·½¶½Å½ÁÊÀÃÊÄÃËÉÆÍÕÎÖÝ×ÛâÚØçÝÛéÝÝæÚÜåØßä×ààÔÞÙÏØÊÅ˺ºº¢§£ˆ’‡t‚qexb\sW[oS[jM\iK]gL561561561560561671872983<;6=<7@=8B?:DA<FC<GD=HE>HE>HE>HE<JF=JF;KG;KG;NH:MG;MG;ME:MD;MD;MD=NC?NC=MC7MD5KB3JA2H>2G=1E=2D;2C:1C:3A:2@93@93?74<73<71>7/>7/<8/=90>:1A:2C<2D=3H@5I?3L@2MA1NA0P@0O@-O@-L?/L>1M=0L;1J91I81E80C90@9/>:/;;/<<0=>0>?1A?0S8-‡<7Ä\[ÊZY®86«.*±0*®.%­2#¬9&¬B*ªF,¦D'£A&£A&¢<#Ÿ9# 9*£;0¦:0¦7.­<4¶C<·E;°@5«=0¨</©;.¬<0»E9ÆLAÅE<¼8,¿6$Í:(Ý;.æ3,ñ+*ù(+ý&+û%'ø**õ.+ï2,ë4,è3*å3)ä2(à3%à:*Ñ2Ü?0Î6)Ç7,Â?5³A6Ñl`¿gY¥RB¨OAËcZçhbæRRêAFóBJë>Dà:<Ö<<Í@9Á>4¶8,¹6.Ä94âHHé@Eð:Fö:Hú:Iø8G÷;Iû?NþAUÿ?Vþ<Tþ9Tÿ8Uÿ9Zÿ:^ÿ;cÿ>iÿ>lÿ>oÿ>pþ?rú@rø@t÷Aq÷CjÿMmÿSs÷JhìAaýTsÿg…ÿe‚÷]wï_xèg{Üp}Æqx§gg�^X{ZQ—~wª™’¾°­Ä¼ºÆÂÃÈÇÌÇÇÏÃÃÏËÊØÌÊØÑÊÚ×ÎßßÔäæÛéêàëíãëïåæòèæôèèðäæîáèíàéêÝçäØâÑÊÑÁ¿Â««©‘–�z†xl{hbu_`rXbpVboScmT21/320431651875984984984;:5<;6==5??7AA9CC9EE;EE;HH@HH>HH<JH;KJ8LK7ML8NK8NK:JF:GD;JFCNIMNIPGDOD?ENE>LC4J@4KA5MC9LB8F=4B90E<5C:5@93?74@85?75<74<42C:5B94A96>95>96>:7?;8@<9B?:D@7G@6J@4L@0O@-O@+L?,G@.F?/I;0K81M53L76J88F;9B?:6904</9D4:A/69$?=(bC1¼j^¹JA¦7,¥5)§7+¥5'¥7&¨=+£8$¥<&ª@*­@)¬=)¬9&®6%«6%¤6'¡7)¢8*¦<.ª@2¬B4¬B2¬A1©>,«=,­<,³=/½C4ÄF8À>1º2&Ç9+Ì8*Ð8+Ö8-Û7-â5.é3/î1-ñ.,ô.-ó0,ñ1,ê5*ã7)Ý:)Û:(Ü9&Ô3Ð3 Ï8'Æ7'¾6(ÂB5ÏSIÈMEÏSKá\Wî]ZðPRí>Cò9?÷?GêAFßCDÕAAÈ?9»<3±;/°</²<0ÊKBÔKCÜFEá>?ë<A÷?Gý?Iù;GþDQüCQþDRÿDUý@Vü<Uÿ<ZÿBdÿFkÿAjú;gõ:gö<mùCsýIyÿKzÿMyÿMtøPsðOpçMiêMjüYxÿgˆÿ_ƒõGl÷Vwè_yÃ^n½{‡a^�xp§‘„»§œÑÄ»ÝÖÐÝÜÚÜÛàÞÜçâÛëæ×ìëØîñÙñöÝó÷âó÷çôõëôôîòùôñùôðùñïöíîóèìîãééÞäæÝâÖÐÔÍÉÊ»º¸£¥ ‹�‰yƒxr}op{k}†sx�lv}k0./1/0320542653762873872:94;:5==5??7AA7CC9DD8EE;HH>HH>HH<JH9KJ6LK6MM5NM8KJ8KI=KJEPPRWU`YXjVTjRO`OIKMD?H?:F=8G>9H?:F=8B;5D:8A96?74?74@85@85=85;62=4/=4/=52<74=96>:9=<:>=9B?:D?9G@6J@4M@0O@-P?+L@*B?,B?.F<0H:1J65I56F35@65@<9:=6;>5@@4E<-J9)^B4€L>®QB«;- 2#¡6&£:'ž7$ž9%£>*¡;%§<(®?+³@-¶=,·9*¹7)µ7)ª9)¥:*¥:*¥;+¨>.«A1«B/«@.®?,«:(±<+¿E6ÆH9Á?1¼8+À6)Ë7+Ï7,Ð9.Ñ;-Ô</Ø:.à8/ç4-î1-ô.+ô.+ñ0+ê4)á7(×:'Õ:$Þ<'Ù6#Õ8%Ô=*Ì;*Á5&Ä:/ÏF<ÑF?ÚIFéOMóPQôGIð;@õ:AøCHåBEÚDCÒBAÅ>8·;1®:-ª<-«=.¶E5ÆL?ÔNEÜGCã?>ï?Bú?Fý?IøCJ÷CLúDPÿFTýBSù>Sü>XÿBaÿGiÿCiü?iù@jùBpüFvþJzÿK{ÿJzÿHtõKpñQsõ\zù_{úWvüNqýEkúDlïEiÚKgÈ_p±nukLI�~t°–‰Ì²¥äÓÉìãÜëçæêéîêçòíâóöãùûáúÿâýÿçÿÿìÿÿñýÿöþÿùûþú÷ýúõü÷ôúõòøïðôëîñèíïæëæÝàÜÖØÊÆÅ´³¯ž¡šŽ“Œ‰�…‡�‚�•‡‰�€†Œ~/.,0/-10.21/43/540762761:94::2<<4>>6@@6BB8CC9DD:IF=IG;JH;LI8MJ7NL7NL7OL9KI:NKBTRS_^fihxmm…lkŠkhƒ`YiYQ\OGRH@KH>GG=EE<A@:<C:;?:7<74<73=82>93>:1=9083-94.;60<92=:3>;4?<5@<3E>4G?4I?3L@0O?0P?/P?-L?,<;&:=(?<+A;-B71A62>42;30=84B;5H94N2.Y,)l/.‡;=¡FE£;0¢2$Ÿ1 ¤9'¥<)ž7$ž8"¢<&¤;&ª=)±A-¶A/¹>.»<-¾:-¾</²<.¬<.©9+©9+«<+­>-­>-­<,²?-°8(¹>.ËL=ËG:¼6*º2&É9.Ð6,Ô6+Ô8,Ö:.Ø:/Ü8.ã6/è3,ð0-ô.+ô.+ð1)ç2'Ý6&Ô8"Ò7!Ý8$Ü5#Ú9'Ú>/Ó=.È6)Å9,ËA7É<5Í>8ÙEAåKIêJJéCCì?AëCCÞCAÕD?ÍB=Â=4¶:.®8*«:*¬;+¬8)ÀD8ÔNEÝIEâ?@ê=A÷BIÿHPôAGô@IùCOýGTþEUú@Uû?XþA_ÿEgÿCiÿCkÿFpÿIwÿK{ÿL|ÿJzÿFxÿHwÿKtûOuÿZ|ÿa€ýUvñCdûIkÿTvÜ?\ÍI`Ø{…¸||aC;�|n®Ž�Õ³§óÛÑûìåúòðù÷ú÷ôýôìûúëÿÿêÿÿìÿÿîÿÿóÿÿ÷ÿÿûÿÿýüÿþûÿýùÿúøÿøöÿõöýóôûðôùðóðçêåßáÓÏξ½¹­®¨¢¥ž�£™�£—£«žž¦—›£–/.,/.,0/-10,21-32.54/650880991;;3==5??5AA7BB8CC9HE<JF;KH9MJ9PK8OM8QL8OL;LJ>QPL^]bmlzzz’�‚¡€‚¨€€¦{u—rkŠe_{YSmTLdMEZG@PB<HD?F@;?<87;63;60<8/=9.;8/:70991;;3>;2?=1@<1?;/A;-F?/H?.K?/M@/O?0O?/P>0L?/@=*?>,@=.?;/?;2>93=:5:94<94D95M51V-+j)-†18ž9A¬>A¦7.¥7(¥:(ª?-¨?, 9&¡8#¤;&¥:&«<)³>,¹@/½>/Á=0Å=1Ä>2½?3¶=2²9.®8*¯9+°:,²:,³9*¶;,¹;-ÄB4ËE9È>3¾1'À2(Í9/Ö5+Ø4*Ú6-Û7-Þ7.á6,æ3,ì1*ó0,ô.+ô/)ï0(ä2&Û4$Ò6 Ð5×2Ú0Ú4$Ü<.Ö=/Ð9.Í=2ÒD:Ä5-Å60Ê;5ØD@âMIåKIÞC?Ö=8Ó@9ÎA8Ç>4¾</µ:+°8*¯9+±;-¬4&½?3ÑH@ØGBßA@æ@DôGMýNUð?Eð?EöBMþHUþHWüBWú@XüA^ÿAcÿCfÿFmÿIrÿMyÿO|ÿL{ÿIyÿDvÿN~ÿR}þKrøMoÿVuÿUu÷MjðFcïKfÒ>VádtþŸ§Ó�Žd92lJ>¥ynÐ¥œõÔËÿéãÿôòÿýÿÿüÿùóÿüïÿÿïÿÿðÿÿòÿÿôÿÿ÷ÿÿûýÿüûÿýúÿüøÿúøÿù÷ÿö÷ÿõöÿôøþó÷öëïìãäÚÕÒÉÆÁº¹´³´¬±´©°¶ª¶¾³°¸­¬´©10,10,0/+0/+10,21,43.54/77/880991;;3==3??5AA7DA8IE:LF:NH:PJ:RK9RM:SL:QK=OJDVTUfdqwx�‡‡©��¹�“Ã��ÆŽ‹À‡ƒ¶{w©pmšid�_[~UPnNJaKGXEBM?=B;7895296/85,85,671783891;;1></?<+B=*C<)G?,J?+K@,LA/M@0M?2L@4K>5L?6K>5F<3B92=82:946;46=57<5>:1H4-Y2-w78”?D¦>E©8:ª;0ª<+«@.¬A/©@-¥<)¦;'«>*§8$­:'µ<+¼=.Á<-Å;.Ê<0Ê=3Å?6¾>3º:/·7,·7,¸8+º8+¹7*»7+ÇA5ÍC8Æ9/Á1&Æ3)Ë7-Ï5+Û4+ß3)á4-â6,ã5,ç5+ì1*ð/*õ/,ö/*ò/)ì0'â2#Ù4!Ð5Ï4Ö1Ù/Ø2"Ø6)×9-Ô:0Ö?6ØE=ÖF>Í@7É<5Ë@9ÖKDÙNGÒE>Å;1Å=1Â<0À</º;,·9*´:+¶<-¹?0µ9-¿<2É@:ÒC?ÛCBãDHíJOöOVì?Eí>CòBLûIUÿJYýFXúCYûC]û@_ûBbÿEjÿJsÿNzÿO|ÿLyþIvÿJwÿT�ÿT}ýJqùImÿStÿVtøNiëD^äF]êXkÿ–¢ÿµ¼ó�žŽHFk.)¦kcÍ–�óÈÁÿãÞÿñðÿüýÿýÿü÷þþöÿÿöÿÿøÿÿøÿÿúÿÿüÿÿýýÿýúÿüøÿûøÿùøÿø÷ÿöøÿõ÷ÿôøÿôøýñóóéêäÜÚÖÑÍÍÊÃÉÉ¿ÊÊÀÊÍÂÉÐȾȿ¸Â¹65143.32.10+10+21,32-43.66.77/880::2<<2>>4@@6B@4JD8ME8OH8RK;TK:TM;SL<RKAQLIYV]li|~�Œ�¸–™Ì™žØšžÞ™˜Ú”‘ÔŒŠÉ†…¿�~µxv§nk–fe‡]ZwVTiMKYDAJ><?;:8;74762555457664872<:.?<+C>*F@*G?*H@+IA,IB0IA4HB6HA9G@:K=<I;;C9:=77875384/83.916=574+C1']80~C=–GCž>?�40§;/ª<+ª?-ª?-©>,©>*¬?+°?-«8%²9(¹:+¿9-Ã9,È:.Í:0Ï<2Ë>5Æ;4Â91Á8.Á8.Â9/Â8.Ã6,Å7-ÑC9Ð@7Ã0&Ã,#Ð7/Ö<4Ó2*à3,ä1*ç2+è3*ê3+í2+ð/*ó-*ö/*õ.)ò/)ì1(á4$Ú5"Ñ6 Ð5Ø7#Ù3#Ö3$Õ3&Õ5)Ô8,Ô<1Ö?6ìYQáRJÓHAÌC;ÊE<ÌI?ÈE;À>1¼:*¼;(¼:*¹:)·:(¸:+¼>0¿A3ÁA6Á<3Æ=7ÏA=×EEÝGHäIMëLPèBFê@CðCIùKTþNYüJZüF\üF]øA]ùBaýEiÿJpÿNwÿOzÿNxÿLvÿQ{ÿRyþOvýOrÿVwÿ]{ÿSoîD^úTlîQd÷dtÿ’žÿ‹•ôƒ‰Ç`d§MMµjgÎ�Šï¼¸ÿÞÙÿïîÿúùÿüþýûÿýúÿþûÿÿûÿÿûÿÿüÿÿûýÿûûÿûøÿûøÿüùÿûúÿûúÿøúÿöøÿó÷ýñóùíïòææçÝÛßØÒÛ×ÎÝÛÏàÞÒßáÖÜãÛÏØÓÅÎÉ<94;8185052+41*41*52+63,74-85.96/;81=:1?<3A>5C?4JB7MC7PG8SJ;WK;UL=UK?SJCSJK]Wcnl‚‚‚¦‘•Åš ÚŸ¦ê¢©ñ¡¥ïž ë™›å•˜Ý”•Ö�‘ʉ‰½‚ƒ±zy¡rq‘fc~XVkPN\IGRFCLCBJ??K==G;:@;9:<94?;/C=-E@,F@*FA+EB/EC4CC7BC;AC>@ACB<FD>LCANEEOCHNAIL>HI=HDCHAA=2L:.gF7ƒOA�I<Ž?2”8)¥>/©>,«>*¬?+­@,¯@-°?-²=+±8'¶8)½9*Ã9,È8-Ì8.Ò:/Ô;3Ï:4Í:3Ë81Ë81Ì92Í:2Ì70Ë6/Ó<5Õ>5Ò91Ì2(Î4*Ø;2Ú<3×3*ä1*é1)ì1*î2)ð1)ò/)ô-(ö,(÷-)õ.)ñ0)ê2(â5'Ù6%Ó7!Ï7 Ô9%Õ8%Ó6%Ñ4%Ò4(Ó7+Ò8.Ð7/ãNGèXPçZSÚQIÌG>ÄD9¿@7¹>/¹<*¹<&¹<(¹:'¸9(º;*À>0ÃA4ÉD;Æ@7É@:ÑFCÕIHÖHGØGJßIKæFHèBDíDIõLSüQZûO]ûL]ûK`öD^÷DaùFfýIlÿNuÿOxÿOxÿPy÷UzóUvôTvüYxÿ]{ÿZwûQkòI`ýWköUgêM^ö^mâKZæTaåTa×XaÁefȃ~嫧ÿÔÑÿëèÿóòÿøùÿþÿûüÿûüÿûüÿüûÿüúýýùúý÷÷ü÷ôÿú÷ÿûøÿüûÿûúÿøøÿóóýîñúëîðáäéÝÝäÙÕâÙÒæßÕëçÛñíáññåìóìÛæâÎÙÕ@=6=:3:7074-52+52+52+63,74-74-96/;81=:1?<3@=4B>3JB7MC7RF8VJ:WK;XL<WK?TICSJM^Xfpmˆ‚„«‘—Ëœ¤ã£­õ§°ýª±ÿ¦¬ú¡§ó ¤îŸ£ê�¡á™›Ö•–Ì’’ĉˆ´yyŸji‹`]|XUpRPhOMeNJcKG^FBS@=H?:>?:6@<1A>/C@/CB0BC3BD7?D=>D@<ED;BJ>CVCIaLRhU\n\br`dobbjd`afXWaJDlG>ƒSE•XF“J7‘@+™>+¨A.®@/®A-¯@-°A.´A/²=+°7&µ7(¼8+Â8+È8-Í6-Ñ7/Õ8/Ø;4Ò72Ð72Ð72Ñ82Ô94Ô94Ò72Ó6/ÞA:Õ8/Ñ3*Ö8/Ü>5Þ=5Ú91Ü5-ç2+ë0)ï0(ñ0)ô/)÷-)÷,(÷,(÷-)ô/)ð1)ê4)á5'Ú7&Ô7$Ð8#Î7"Ï8%Î7$Ï6&Ð7)Ò9+Ñ7+Î4*Ë4+ãNGód\ë`YÕOFÄA7º;2µ9-¸>)¸>'¹<&¹;%º9&½9*À</Ä>2ÊD9ÊA9ÎE?ÕKHÓLIÎGDÍEEÖHGãIIæCDèEHòMSøSZùR\÷O^úOaôH^õG`øGdûIiÿNrÿQxÿRyÿTzðUuïZxù^}ý^|ùUpõMgùOiÿYnÿ[mÿctðM^øUfôO`üUgõL_äR_À^_»vqל˜úÉÅÿåãÿíëÿôôÿÿýüÿÿûÿÿûÿÿûÿÿýýýýüúþùöýøõÿøõÿùöÿúùÿúùÿööüððøéìôææëÝÝèÚÙäÙÓèÞÕðèÝùóåÿúìþþòôúöâìëÓÝÜB>5A=4@<3>:1<8/:6-84+73*62)62)73*84+;60>93A<6E>6I@7MC7RF8UI9WJ:XK;ZLA[NHTIMXR`gd�|~§�’Ê— ãžªö¥°ÿª³ÿ«²ÿ¬´ÿ¬²üª±ùª¯ó¨¬ì©«èŸ¡Ú™šÒ��Á�±tr¡he’]Z‡XR~XQ{YQxVOpMH_C?M?:@?;:B?8>>4@B5BE:?D=<B>:CB>FH?LUEWoQfƒarŽisŽtv��x�„o€€^lŠ[c’WYžWU¦VM¦M?Ÿ@.œ9$ :$¨=)¬=*®?,±@.³@-´?-¶=,¹;,º6'¾6(Ä6*Ë7-Ñ7-Ö8/Ú91Ú83×84Ø95Ù:6Ø93Ø61Ø61Ú83Ý:3æC<à=6Ú70ã@7þ[Rÿlcÿ]Tç?6ä2(í2)ñ2*ô/)ô*&õ(%ø(&ø+(÷-+ó0,í2+ã1'Ú0#Ô1"Ô4$Ó:(Ì;&Ê<(Í<)Ï<*Ò:,Ò9+Ò8,Ñ7-Ú@8Ï81ÜGAôc^åXQËB<Å@;·4*·:&¶<$¹<&º<&½:(¿:+Ä<0Æ>2Ç>4ÌC;ÙPJßXRÔOJÃ?:À?:ÏHDÞHGâGEæJKïPTñRWïNVíKXñL\ùRføOf÷Ke÷JfüMlÿQsÿRvýRvôWvîXuõXuû[wÿ\xÿZsÿWnÿVkÿbuýVgÿ\mÿ[lñFXôDXÿQeîVeÆfgªieÊ�‹ç¶²øÒÏÿëçýïîüø÷ùýüùÿÿùÿÿúÿþüþýÿþüÿûøþöôþùõÿû÷ÿùöÿöôÿööÿõõöèèæØØãÕÔÞÐÍáÖÐñçÞüôçþøèÿüéÿÿóúÿùèñðØáàC?6B>3A=4?;2=90;7.:6-95,73*73*73*84+:5/=82@;5D=5H?6LB6OF7TH8VI8WJ9YK>ZMEUJNWQ_eby{¤Š�É”�⛦õ¢¯þ«´ÿ¬µÿ®¶ÿ¯·ÿ°·ÿ¯¶ü¯´ø¯²õ©ªë¢£ã˜˜ØŒŠÉ€»uq®jf¡d_™`WŽ`VŠ]TVPtMIbFCTBAIAAC<<:AB=FEACD?A?@DBEOMRWVdch†quš}¢‡}Ÿ’}œ |˜¥pŠ¤`u¬Yi¶VaºQU¸HF°@4¬;)¬<&­='¨8$¨7%«8&®9(±9(²9(µ7(¸6(»3%Â4(Ê7-Ó:2Ù<3Þ=5ß<5ß<5á>9ß<7ß<7á>9åB=èC=çB<æA;å@:æB9ä@7Ü8/Ô1(Ø6+ëI>ÿ[Oñ@6î6,ê+#î)#ø-)ÿ/-þ,+ö((ú0.õ4/ï61ç6.Þ3)Ö2&Ô2%Ï6&Ë:'È;'Ê<(Í:(Ï9*Ð7)Ñ5)Ð4(Ó9/Ï5-Ô=6åPJåTQÛLHÍB?·1(¼;(»=&½<'¾;'¾9(À8(Â:,Ä:/È>4ËB8ÕOFÜWPÒRIÃD=¿C;ÌHCÛLHßJFåMLëSRíTWëPVéNVìOZ÷VføUhùSiúRkþQmÿSqÿStÿRuòMmøUtÿ\zÿ_{ÿ\vÿUoûRiûReüUf÷RbÿZhÿZhóN^øScþYißS^¶`_—^W¶�}Ù©¥îÈÅýáÞùééýøõúüùùÿÿøÿÿ÷ÿýúþýÿÿýÿûûÿ÷õýøôþùõþùõþöóÿ÷öÿõõöêêêÜÛäÖÓÝÐÊßÕÌïåÛüõåÿùçÿýéÿÿïúÿøèñîÛáßEA6EA5C?4A=2?;0=9.<8-;7,84+84+84+84+:5/=82?:4A=4F?5JB7ND8RF6TG6WG7WJ:XKBWKKXP]b^ysuœ…ŠÂ�™Ü˜¤ðŸ¬úª³þ¬¶þ¯·ÿ±¹ÿ³ºÿ´»ÿµ¹ÿµ¹ÿ¯²ù¬¬ô££ë™—àŽ�Óƒ€Çxu¼sm³k`¤i]›cZ‘]W…VRwNKhEDV@@L85<A<@KABL@BQ>BZEJpW]‚cs—s—¤{§®«²{£¸vœÀp“ÂcƒÀTnÂG\ÐK\ÑHPÄ<<º5.¸9*µ<'°<%°;'°;)²:)´;*·<,»<-À<-Ä<.Ä6*Ë8.Ó<1Ù?5ßA8á>7à<3ß:4Ü71Ù4.Ø3-Þ93æA;ìE?éB<ä=5ã<4Û4,Ø4+Û7-Ô2'Ì, Ö6*éG:õK>ò@6ï4-ð-)ö,*ú,,ü,,ø*,ð*)ì/+è2.ã5.Ü4+Ø1(Õ1'Ñ3'Í7(Ë:)Í:(Ð:)Ò;*Ó:*Ô8+Ô8,Ñ4+Ö<4Õ:5Õ<7åONíYWÙEEÂ3/À;,¿<(À;*¿:)Á9)À8(À8*À8,Æ=3ÇA6ÏJAØUKÒRGÆG>ÀD:ÇG>ÖKDØICÝLIäSPåSTâPSâMSåNWòXdöZhý[pþZrýUnûQlýPlÿQpýGmÿPvÿZ}ÿ[{ÿVrýTkþWkÿ\mûYhøZhûamö`kí]gïfnìfmÉ\_ ]W{OF˜mfÁ˜”ݺ¶ðÔÑôàßÿõôùù÷ùÿýùÿÿ÷ÿýúþýÿÿÿÿýÿÿøùüùôüùòüùôýøôÿùöÿøöùîìïäâåÚÖÛÐÊÜÒÈîäØýöäÿûèÿýçÿÿíúýôêðìÞãßIE9HD8FB6D@4B>3@<1?;0>:/:6-95,95,95,:5/<71>93@;5C?6HA7LD7OF5RE4UF3UH7VJ>XMKXNW^Zqpp–€†ºŒ–Ô•¡é�©õ§±ù©±ù¬´ý°·ÿ±·ÿ²¹ÿ´¸ÿ´¸ÿ¯±ü«­ø¥§òŸžê—–⌋ׂ�Í}yÅum¶pgªf`ž_Z�YX„QQuGGcA>Q=5DH8BT=C`@EnCJ€OU–]dªgx½k�ÉnšËo˜Éi�È`ƒÉWxÆIgÄ;UÆ2HÙAPàEMÑ;<Æ71Ä?0½@,±:$­6"­5$®5$³5&¶7(¼8+Ã;-Ç;.Ó@6×@7Ú@6Ü?6Þ=5Þ93Ý6.Û4,Ó.(Ò-'Ó.(Ø3-á:4ã<4á81Ü5-Ü5-Ð,"Í)Ù7,âB6Ü?0Ð4%Ë.éF7óI<ýH?ü<7ô-*ñ#%õ')û/2ñ-.ë/.ä2.Þ3,Ú1*×0*Ö/)Ó0'Ó7*Ò9+Ö:+Ø<-Ú>/Û?0Ü@3Ü@3Ò6*æLBåJEÎ50ßGFóZ\ßIKÕA?È</Ä;+Ã:*Ã:*Â:*Â:*Á9+¾9*Â<0Â>2ÈF9ÏOBÎPDÆH<¿C7ÁC7ÎH?ÑF?ÕJEÛPKÝROÙMLÚKMÝLQëX`ó[güaqÿauûXmõOgõMgÿNkÿOuÿRyÿUwÿTqüTmý[pÿcuÿjxÿguúboõamídlãflÙhjÎghµfa�^TeG<^U©ˆ�Í«©æÈÆïÚÙÿóóùøöùýüøÿÿ÷ÿÿûÿÿÿþÿÿüÿÿùúþûöûûóûúõþûöÿýùÿûøþôòøíéêßÙÞÔËÝÓÉíæÖÿøåÿþèÿÿæÿÿëøúïêïèãæßNH:MG9LF8JD6HB6F@4D=3C<2?80>7/=6.=6.=60>71@93?:4C>8EA8KC8NE6QD3RE2VG4VH;ZMGVMR[Whlk‹~‚±Š“Î’Ÿãš¦î¦®÷©¯ù«±ý¯³ý°³ÿ°´þ°³ÿ¯³ý«®û¨«ø¤§ô ¢ïšœé“•âŠŒÙ†…уÉ{u»mk¬dež_a’Y[„RQsNHbWFY_CQmBL�EO˜LV¬U]¼[dÊZhÙRpâRußTuÙPlÒJbÍBWÇ8JÅ.?Ô8EãCKåFJØ>>Í>6ÌF:ÃH6¶?+·>-¸=-º<-½>/Ã?0ÊB4ÐD7ÕE:ÙE9ÚB7Û>5Ù;2Ù6-Ù5,Ú3+Ù2*Õ0*×4-Û7.Ý90ß80Þ7/Ý6.Ü5,Ö/&Ù5+×7+Ï2#Î2#Ò9)Õ>-Ô;)Ø<-çC7øJAýD?ù64õ*-ö*-ö.1÷67í55â30Ú1,×0*×/,Ù0-Ù0+ã81ä91â:1â;2à<0ß=0Ü?0Ú>1Ñ7+ïWLøaZÓ;6Ñ;:æPQÜFHæPQÏ<4Ê:/Æ8,Å9,Ä:-Ä<.Ã;-À;,¾:-¾<.ÃA3ÇH9ÇI;ÂD6¾@2¼>0ÉF<ÊE<ÎIBÕPI×RMÕNKÓKK×LOéZ`ð^hüfrÿhxü]qóPeóMeýNkÿVzÿTxÿRrøTmö]qûhxûjw÷erüetó_mì`kêkrßruÃjfªd\žla}fVXH9kXJ”|r½ œÞÂÁîÖÖþîïûõõûûûûÿÿùÿÿýþÿÿþÿÿûÿÿúýÿþùûþõûüöÿþùÿÿúÿþúÿûõÿ÷òñèáäÜÑáÙÌðé×ÿúäÿÿæÿþåÿÿê÷ùëíðçèéáQK=PJ<OI;MG9KE9IC7G@6G@6B;3A:2@91?80?82?82A:4@;5B?8EB9KE9MF6PE3RE2UF3UH7YLCUKLZScii…|�«‰“È“žÞš¤ë¦¬ø©¬û¬­ý­°ÿ®±ÿ®±þ­°ý­°ý©¬ù§ª÷¤¨ò¢¦ðŸ£í›Ÿé•™ã‘•Þ’’Ú‰‰Ï|€Àsx²ou©mp�igŽj_}y_x‚WjŽP_¢N[»S^ÍT]ÖPWÛHRå@Qè?RäCRÜDPÖCKÏ@DÊ9<É46äJLçIJáAAÕ74Î95ÎA8ÈD8¾>1ÃE6ÃE6ÅF7ÉE8ÎF8ÑG:×G<ÚG=Ö>3Õ;1Ö8/Õ4,Ö3*Ø4+Û4,Ú6-Û81Þ;4à=4à<3Ý90Ü5-Ü5,Û7-Ò.$Ý;0Ú=.Í1"Æ-Î7&Ò=)Í:&Î8'Ô8)â:/ó=9ÿ@@ÿ<>þ37ó,/ê,.á+*Ø*)Ô+(×/,Þ44ä88é99ï75ð74î73è71ã7-Ü6*×5(Ñ5&Í5(çPEÿmcÛHAÈ42×CCÓ>BêVVÙA<Ò>4Ë8.Æ6+Æ:-Æ</Ã>/Á=.»9+»<-¾?0ÀA2¿A2½@.¼=.»<-ÄD7ÄD9ÈH?ÐPGÓRLÐOIÒNLÕONç_cîagûitÿnzÿduõVjõQiÿUoÿWvÿVtúXpñ^pônyöw€íksÞXañ`mñ`mìboïs}숊Ì�|ªth™yjskXON:_VGƒqg¯–’ÜÀ¿ïÕØüéëþôõýûüýþÿúþÿþýÿÿýÿÿûÿÿùþÿÿûùÿõúýöÿÿúÿÿúÿÿøÿý÷ÿþöùðçíåÚéáÔôïÜÿûåÿþåÿÿãÿÿèùúêòôçîðåSM=RL<QK=OI;MG9KE7JD8IC7E>4D=3B;3A:2@93A:4A96@;7A@;CC;IE9MG7OG2RF0UF1UH5WK?RHFXR^jg‚}�§‰“Ä’žÚ˜¢ç¤§ö¦§ú«©ü«¬þ¬­ÿ­®þ¬®û¬®û¨¬ö¦ªô¤¨ñ£§ð¡¨î §í�¤è›¢æ—žâ’šÛ‹“Ò†�ʃ�„ˆ·ƒ�¨ˆyš�mˆšcz©Xi»Q_ÏMYàIRèAIê=Aé=;ã>8ÝC9ÖH:ÎI8ÅF3ÂC0Å@1äTIàG?Ú;7Ø64Ø88Ø:;Ñ98È74À:/¾<.¿;.À:.Â8+Ã7*Æ4'Æ2&Î4*Ð3*Ñ3(Ô3)Ö3*Ù5,Ü5-Ü8/Þ=3Ý<2Û:0Û8/Û7-Ú6,Ø4*×3)Õ3&Ó3%Ð4%Ï6&Ò=)Ñ>*Ê9$À2È:&Ê4%Õ1'ç51ú<<ÿ=@ÿ7=÷37ê-1á-.Ù--×/.Ü43ã9:é;=ð9=ö26ø03ô01í1/ä1*Þ2(Õ1%Ï3$É1$Ñ=1új_äTLÈ95Ð@?Ë:=äRSäJHÜC=Ñ:3É6.Æ8.Å;0Â<0¿=/º;*¼?-½@.¼?-º;*º;(½;+¾<,¼>0¼>2ÀD8ÈLBÌPHÊNFÌLIÐNLç`dìaføhrÿo{ÿhx÷Zk÷UjþZrÿYsû]tîaræjtë}€ñ‡‰èwyÚ_dñgtþm|óeuëlwö‘•ì¢ŸÄ”Š¢�|hjUHQ<WUFth\¤ŒˆÛ¿¾òÕÙùãæÿóöÿúüþþÿüýÿþûÿÿüÿÿúÿÿøýÿÿûùÿõ÷ýóýÿ÷ÿÿøÿýöÿý÷ÿÿöÿùïõïáòìÜúõáÿýçÿýäÿþâÿÿçýþìùúì÷÷ëVO?TN>SK>PJ<OG:LF8LD7JD8H@5F@4E<3B;1B92A:2C:3A<6C@9DD:JF:MG7OG2QF0TH2UH5UI=QGEWQ]jh€~‚¨Œ”Å“�Ø™¢çŸ ò£ ÷¦£ú§§ý¨ªÿ©«þ©¬ý§­û¥«õ£ªò£§î¢§ë¤©í¦©î¥¨í£§î˜ è”¡å”Ÿß”Ÿ×•›Íš•¾ Œ¯ªƒ ¥e´]pÂQcÐIWÜBNå>Fë<Aì<<ê?8ãA4ÜE4ÔI4ÊI3ÂF.¿B,Â?-âSEÞD<Ý97æ>>êDFå@DÙ8=Î65ÍB;ÈD8ÉC8ÇA5Æ>2Ã9,Â6)Â2'Æ2(Ë3(Î4*Ó5,Ø5.Ù4.Û4.Û4,äB7Ý=1Ù7,Ú6,ß9-ß9-Û5)Õ1%Ø8(Í2 Ê1!Ï9(Î8'Ç4"Ê7%ÔC0È7&Î8*×6,Þ5.æ3/î53ö:9û?>ÿLKúHFòBBì>=ê<=ç7:å26æ,1ô+1ø)/ô,.ï/.è1+à4*Ù5)Ó7(Æ0!À.ôdYë]SÏ@:ÔDCË;;ßMMêRQàHEÓ;6É6/Æ8.Æ9/Ã;-¾:+½;+Á@-ÀA.»<)·8%¸9&½<)¾?.µ7)³7+¸>1ÁG:ÅK@ÅIAÆJBËJEåa_æ^`ñdjþnwþkuó]iðXeø]mú^sõbtèdoãlrì~ù�‹ö†…éqsøryÿy†õaqàXföˆ‘ÿ´´Ð®¢ ›‡]jPDR9QT?jbUœ‰‚ÚÀ¿òØÙöàãÿô÷ÿúýÿþÿýüÿþüÿÿýÿÿûÿþùýûÿúôÿòòþòùÿôüÿöûýòÿýôÿÿôÿÿóûõç÷ñáþùåÿýçþüãþýáÿÿêÿÿñüþóüüòXO@WP@WN?UN>TK<RK;RI:PH;MC7KC6KA5H@3H>2G?2H>2F@4FB9GE9JG8NI6RJ5UJ4UJ4VK9XNDUMKZTbkhƒ{~§‰�љטŸç£¤ö¥¢ù¢¢ü¡¢ü¡¥ÿ¢¨ÿ£«ÿ¤­ü¡ªõ¡©ñ¡¨ì¥¨íª©ï­©ó±©ö¬©ø ©ø—§ò’¡â–ŸÖ¦£Î¶ž¾¹…œ¸gz¿L]ÑERß>Mç;Gç<Dæ=Bà@@à@@â>?ß<=ÝEBÉ:4ÊD9ÊF:º1)ÛJEçIHéCEê=Aè;?æ<?á=>Ú<=Ô<;ÏB;Ç>6ÊD;ÑMAÍK>¿=0¸6)¼8+¿7+Ã6,È5-Ë2*Ñ/*Ô/+Û2/Þ5.çA5ß9+Ù1$Ý3&å9+ç;-ß8&Õ2Ï2Ì4Ê5!Ç4"Ç4"Ê4%Í5'Ï5)Ê0&Í3)Ñ5)Ò6*Ò4(Õ3&×3'Ü4)à2)æ3.ë52ë33é/2ç-2è-4í.5ö-3ö+/ï+,ì/-ê5.ã9,Ù7(Ð4%Ã-Ä3"çXHéYNË;3ßNIÅ41ÚHHáOPÜJJÔC@Ì;6È80Æ8,Ç9+È<+Å<*Â;(¾9&¼;(»<)º=)º=+·<,´8,µ;0¹=1»?3»?3¾@4À@5ÄA7ïjaõpiêc_ômjûqqìadôgmÿpyÿlzûhxðdoébiëdjîlnîqoîqoûy{ÿ}…ùZlücwßbpû¬¯¶ªš}‘u]rQJY:LR8snZ¥•ˆË·°íÕÓÿîðÿö÷ÿ÷ûýøüüüþÿÿÿÿþÿÿþÿûÿþðÿôåþéæûêïÿîøÿöüÿôþÿóÿþñÿýñÿûîÿúêÿùæÿùãÿùáÿùáüúåÿÿõþÿûÿÿûYPAXO@XO@VM>UL=TK<SJ;RI:NE6MD5LC4KB3JA2JA2JA2IA4GC8HD9LF8OH6RJ5SK4VK7TK:XNDULMYUckhƒ{}¦†ŒÀ�•Ó–šáŸžì žòžžôžŸù�¢ýŸ¦ÿ ªÿ¡¬û ¬ô «íŸ§è¢¥è¥¤ê©£í« î¤Ÿï §÷š¦ðœ¡á£œÐ±˜À¿�ªÂq„ÄTbÑCOâ<Fì8Aï7?ê:=å<?Ü@AÛACà<CÚ9?ÚADË;;ÌB?Ê@=Á31åOPæCFè?Dê;Bç8?â8;Û89Õ:8Ï;7ÔE?Æ=5Å<4ÌG>ÌLAÁC5¸:,·9+¹7*¿7+Æ6-Ë4-Ñ2.Ø3/á53ä84ä>2á;-Þ6)ß5&á5'á5'Ú5"Ó4Ñ9$Ë9"Ç9%Ç:&É<+Ë<,Ì8,Î4*Ó4.Ö5-Ó7+Ï7)Ë5&Ç4"Ë4!Ï4"Ø2$Þ2&ã0+æ.,æ,-æ,/æ-2è.3ì+.í,-ì0.è2.à4*Û5'Õ8'Ò9'Õ?.»*ÖD5Ð>1ÙF>ëVPÚB?ØBAçUVâRRÜKHÔC>Í=5É9.È9+Ç8(È;)Ä;(¿:'½<)½>+º?-¹@/¶>.²:,´;0·=0¸>1º?0½?1ÁB3ÅC5ëi\òmdçb[ðkfõolèbañklütxÿq~ÿo|öirí]fêY`ñbf÷qpþzxûvwÿx~ùRdþ[pä]nö¦©©¦“jŒkZtON_;RY:us\©��ÖĸöáÜÿõñÿúúÿúûÿûüüüüüþýûÿþúþÿõÿúíÿóáÿéáýæéÿëóÿñøÿòýÿòÿÿñÿýïÿýíÿúéÿöåûóàúòÝüôßü÷äûûóüüúÿÿý\PB[OA[OAYM?XL>WK=VJ<VJ<SG9RF8QE7OC5OC5OC5OC5MC7JC9JF;MG9PI7SK6VK5VK7TK:WMCVNLZWbki�y{¡…‰¹‹‘Ë‘–Ö˜˜Þ™™ã˜™éš�î› ô�£÷Ÿ§ø ©ô ªï�¨èœ¤ãœ¢àžŸàž�á�™ã›™â–›ßššÚ¤•Ð±ŒÀ¿‚¬Çr“ÉYqÇBS×ALß>Dã=?å==å<?ã=?á>Aá>Cà<C×8=Õ?AÎ@?ÊC?Ã<8Ä96éWWàBCä>Bè;?ã9<Ü68Ö66Ñ96Ì;6ÕHAÄ;1¾5+ÅA5ÍK>ÆH:»=/·8)º6)¿7)Ç7,Ï6.×50Þ71æ95é=9×3*×5*Ú6*Ü6*Ý5*Ý7)Ú:*×>,Å4!¿4¹4!¹6"¼8)½8)½3(À-%ã@;å>8Þ=3Õ<.Î8)È7$È7"Ì7#Ò7%Õ3$×/&Ù.'Û.*Ý/.Û//Ü./Û,)ß0+à5-Ü5,Õ3&Ð4%Ð9&Ñ>*ÔA/Â1 çSEÒ;0ÚA;ÞC?èJIðTUíUTçSQãOMÝJCÖC;Ï<2Ê8+Å6&Ä7&À7%¼7&¹:'¹<*·>-µ=,²=,°:,±;/´<.µ=/·=.¹?0¾C4ÃE7Ü\Qêg]äaYðlgöpoçabìfgójqàR^ï^kõhqòemñ`gõfjùqqþxwútuÿnuóJ]øRhä[mñœ¡£œŠa€`UoHOd=W`Aww]­¥’áÑÂÿîæÿøñÿüøÿþúþÿúûÿüûÿýûÿýùÿûóÿöèÿîÜþãÚøÞáùáéüæïüèöýëýþìýúéÿúêÿøçúñàôèØòçÕöëÙøñáþúñýüøÿÿû\PB\PB[OAZN@YM?YM?XL>XL>UI;TH:SG9RF8RF8RF8RF8PF:LE;KG<OI;RK9TL7WL6WL8WK;VLBUMK\V`jh~yy›‚…²Š�Â�”Γ•Ò•”Ö••Û–˜ã˜›è˜žìš¢í›£ëœ¥è˜¢ß–ŸÚ•›Õ”˜Õ“—Ö’“Ö‘�Ò…„½‘ƒ¶¥}±·s¢Èf�ÓW{ÕGaÔ<KÞAJàBCÞCAàB?áAAä?Cç>Eç=Fã?FÖ:>Ñ?@ÐDCÆA<º61Å>:êZYÝABã@Cæ=@ã:=Û89Ó97Ð<8ÎA8ÏE;Â:.»2(À</ÇE7ÄE6½>/¸9(½8)Â9)Ë8.Ò:/Ù80à91è;7ë>8Ø1+Ù2*Ü3,Ý5,à5-ß7.Û9.Ö=/É7(Æ:)Á<+¾<,¿;.Á;0Å<6Î95ëC@í@<ä@7Ú<1Ï9*É8%È:&É;%Ï<(Ð7'Ï1%Ñ1%Ô1(Ö3,Ô1,Ñ/*Ó2*Ó5,Ó7+Ð7)Í5'Ê7%Ë:'Ì>*Ì;(Ï<,ô^PàF<Ò3/Ò.,æ@@ûWXÿusÿroÿjgö^YçPGÕA5Ç5(½.Ä8'À8(»8&¸;)·<,µ=-²=,°<-¯;.­<.°</°<-±;-¶>.¼B3ÀF7ËMAß`Wâa[ôpløstæ_cä]aæ]dÖHTçYeógrôgoöelükpþsvþvvûpsûenñDXôKbåXkë’– •ƒ_zYQlCSh?\gEy|_°ª”çÜÊÿõèÿùðÿûñþþôýÿ÷ûÿúùÿúöÿøòüóêûëÚøÞÎóÒÊëÌÏìÎ×îÒàðÖéóÛò÷áöõáûöãüõãõìÛíáÑéÝÍíáÑñèÙüõíü÷ñþùó\PB\PB[OA[OAZN@YM?YM?YM?WK=VJ<UI;UI;TH:UI;UI;SI=OG<OI=QI<SL:UM8XM7YL9XL<UKAUKI\U]hexut“|~¥…‡·Œ�ÂŽ�Ã��È�’Ï�”Ô’•Ü“—à”˜ß•šÞ•�Ü’›ÖŒ•Î‡�Ç…‹Å�‡Ã}ƒÁ~¸}tŸŽp”¨g�Á\„ÖNvâAcè:Uê9KæAGàECÜGAÛGCßEEæAGî<Jí=JæBI×>@Ð@?ÐIEÀ?9³2,ÈC<àUPÜDCáACã?@á>?Ú?=Ô@<ÑD;ÏF<Ç=2À8*½5)½9*¿=-¿=-½;+½<)Â:*Ç;*Ï;/Õ;/Û8/à8/ç83é:5æ95ä52ä20æ21ç32æ40Ý2+Ô0'Ë1'Ë7-È:0Ä7.Á4-Ã40Ì:;Ý?@ë8;í76ä71Ø5,Ì4&Æ5"Ä7#Å:%Ë=)Ë8&Ì4&Î5'Ô8,Ô:0Ó9/Ð8-Ð>1Ê;-Å8'Ä7%Æ9'Ç<)Ç<)Ç:(Í>-Í:*Ø>2áC:Ô/-ä::ß13ß56×53Ö;6ÞC>åKCêQIêSHèTHäUGË?0Ä?.¾<,¸=-·>-´?.°?/®>0¬<.¬<.¬<.¬=,­<,±=.·A3ºD6ÆLAÝcXàc]ïolôrrå`cå`eç`g÷kvûoz÷kví`hð_fýlqÿy|ÿy|úmsø^jóAWôC]èUhæ‡�£“ƒg]YtI^uIhwP�ˆf³²–èâÌÿúéÿýíÿþïýÿòûÿôøÿôôÿñíûêäòáØî×Åèǻ伺޺¿Þ¼ÇàÀÍâÃØæÌãëÓìîØóðÝ÷òßóêÙìàÒèÚÍêÜÏïáÖòèßñèáòéâ[OA[OA[OAZN@ZN@YM?YM?YM?XL>XL>WK=VJ<VJ<WK=WK=XL>QI>QI>SK>UL;XM9XM7YL9YL;WK?ULGYSWe^nnjƒut–~}¥„…±†‰´‡‰¹Š‹Á‹�È�ŽÏŽŽÔ��Õ��Ó�”Ó‹�ʃŠÁ{‚¸u|²ov­io«ljœ€lˆ“cy®YvËPoãBdð7Vö4L÷8Hï@EåFBÝJCÛJEáGGèCJò=Nð>LäCI×CCÌA>ÎIDº;4°4,ÉI@ÕLFØBAÚ?=Û;;Ú<;Ø@=ÔE?ÍG<ËG:À8*Á8(À8*¾9(¼9'»8&½:(À;(Ä;)Ê=,Ò<.×;.Ü8.á6.ä6/è50é32è./é,0î02õ47õ77î45æ21Ú.*×4/Õ62Ñ32Ï/1Ò/4Û6=ê:Dí06î02æ3/Ü3,Ñ5)Ê7'È9(É<(Æ8$É6$Î6(Ó:,Ú>2ÝA5ÝA5ØA6ÓG8ÉA1Â;(À9&Ã<)Æ=*Å<*Å8&Ì:+Í7)Í0'æC<Û2/ï?Aã/2Ú*,Õ1/Ò5.Õ81Ö90Ò8.Ì5*Æ2&À1#ÎE5ÆA2¼=.µ:*±9)¯9+«:,ª:,«;/«=0¬>/¬<.«<+­<,²>/µA2ÉSGàg^ßd_èkiînoæchðjqõoxúq{üs}ömuîbködnÿrzÿx~ÿquúipøWfùAYô@[êQfá~ƒ¬—†y�kmˆ]o‰ZyŠ`�—r¶·˜ßÞÂùôÞüúåÿÿïûÿïõÿïïÿëçúäÜïÙÐãÍÅÞÁ°Ö­¨Ó¦ªÐ§°Ó«¶Ô®¼Ö±ÇÚºÒßÁàæÌèêÔòíÚóêÙîâÔìÛÑêÙÏìÛÓéÚÓæÙÑä×Ï[N>[N>[N>[N>ZM=ZM=ZM=ZM=YL<YL<XK;XK;XK;YL<YL<YM?SI?TJ>UL=XL<YN:ZM:ZM:ZM<YM?VKEXOR_Wbd_sjf�so�yy�|~¤~€©�‚²ƒ„¼††Æˆ‡Ë‰ˆÌ‰ˆÊˆˆÈƒƒ¿|}µvw­pq§kl¤fg l`Ž‚`q•V^³N^ÒJ^ì@X÷6Kþ3Fþ9Có?BéD@áHBÞICâFIéBJò=Nï>NÝAEÔFDÈA=ÉHB´:/±7,ÎOFÉB<Î=8Ð64Ï10Ð51Ó>8ÑD;ÈF9ÃD5À8(Ã:(Â:*¿:'½8%¾9&¿:'Á:&Ä9&Ë:)Ó:,Ù9+Ý7+á5+ä3+ê3-ë*+ñ*-ø-3ü/4ÿ17ÿ37ÿ38ý58ù8;õ8<ò9>ò9Aô9D÷7Fû6Hþ5Eó*4ô-2í12ã4/Ù6-Ð8*Î;+Î=,Ë5&Ñ8*Ù;/ß=2â>4ä=4å>5ßA6ÙJ<ÏG7ÉA1Ä<,Â9)Â9'Æ8*È9)É2'Ó9/á@8ëD>Þ0/à..ë46è66Ø3-Õ7.×90×:1Õ;1Ò;0Í:0É;/ÖL?ÌF:¿@1´9*°6)¬6(«7*©8*¬;-­=/®>0­=/«<+«<+­<.±=0ÀKAÚdZÞc^èkiðosêinõrzüvîfp÷oyùryöjsþlvÿwÿu|ùhoùcn÷RbþB[÷<YêKaÝt{µš‰�žz}˜k{—f‚•h‘�u¬²ŽËÌ­ààÄéìÑîöÞåóÙÙíÑÏæÉÅÞÀ¼Õ·´Ë®ªÈ¤šÁ”–Â�™Ã“¡È™§Ë�¬Ë¡µÎ§ÁÒ°ÎÙ»ÙÝÄåãÎìåÕïáÖëÚÒçÒÍãÐÊàÏÈÚÌÃÕǾ[N>ZM=ZM=ZM=ZM=ZM=YL<YL<ZM=YL<YL<XK;XK;YL<ZM=ZL?VJ>VJ>XL>YN<ZM<ZM:ZM:ZM<[OAWMDWML\RZ^Xfb]qjfsp�wv–xxœ|{§�~µ„¿…‚ň‚ȉ„Ç‚}¿y·zu­xq§un¤sj¡pg za‹„QZ™HE³EHÓHOîBNø:Fþ6Aþ9@õ==îA=æD?ãEBäCHèAKî=Më?MØ?BÑGDÅA<ÃG?²9.³:/ÑUKÃ>5É83Ç/,Å*&È/*Í:3ÍC9ÃC6¼>/À;*Ä;)Ã<)Á:'Â;(Ã<)Â;'Ã8%Ä5$Ê7%Ô8)Ù7*Ý5*á3*å3)ê2*õ33ÿ58ÿ8<ÿ4:ÿ-2ÿ(-ÿ(-þ+1ü-3ö)0ò&1÷)6ÿ,?ÿ*Aü 9ó/ö"0ó(.í,1æ0/Ú1,Ô3+Ñ5)Ð6*Ô6+Ý90æ=6ë>7ë:4è71è50â92ßH=×K<ÓE7Ë=/Ä6(Â3%È6)Ï8-Î4*Ô6-æC<à93à21Ò  ì89å63Õ1(Ò4)Ò4)Ï5)Î6+Ê6*È5+Å7+ãYNØRFÊH;¾@2·;/µ;.µ<1µ<1¯9-±;/±=0±=.¯;,­9*­9,®:-±<2ÐYQÜb]ìppôvyîmrõq|ör}ÿzƒÿ‰ÿy€ôhq÷epÿq|ÿuÿktø^jöOaÿA^÷8WéG^Ûntº�‹™©„‚žn}™f€–g�œq¡©‚´º–ÈÍ­ÕܽÌÙ»ÁÔ´²Ì©§ÂŸž¼˜š¶�–²Œ�±†‰³�†µ~�¹„•À‹œÂ�ŸÄ‘¨Å™²È¡¾ÎªÉÒµÚÙÄæßÍêÞÒéÖÏáÌÉÛÆÃÔÁ»Î½µÈ·°ZM=ZM=ZM=ZM=ZM=ZM=ZM=ZM=XK;XK;XK;YL<YL<ZM=ZM=ZM=ZL?ZL?ZM=[N=]M=^O<\O<\O>YM=YLCWLHXNOZPX]Ub`Yia\rjf�okŽsp›vq§{u³ƒ~ÁŠƒÉŠƒÇˆÀŠ~¼ˆy²ƒr¨†r§�t«‡l£†\„�HL¤B9¶A:ÍC@Þ@?è:;ô<>ÿDE÷><ó=<ì>=ê@AëBIêCMëANãALÐ>>Å@9¾?6·>3°</²<0ÀD8ÒMDÇ61Ô;6Í2.Ì3.ÖC;È>3µ6'»>,»6%Á:'Ã<)Â;(À9%Â9&Æ;(Ç:(Ë:'Ð9(Ö6(Ø2$Ú."à/%ì7.õ=5õ82÷40÷0-ø**û&(û%'û%'û%'û&*÷!)ø".ÿ'7ÿ(>ÿ!;ÿ6ý1ÿ'8ï#,æ#+ã+-Ý--Ô+(Ò-)Ü41â62è64î66ò65ò12ï./î,,å/,Ü92Ñ:1Ò91ìSK¼#Ä+#ßF>È.&Í.(Þ<7èE@â:7Ø/*Ú.*â51å<7ß>4Õ<.Ï5)Ë3&Ì5*Í;.Ë;0Å8.ìbXàZOÈC:¹6,º:/¸8-±3'´6*°0%°2&®2&¯4%°6'²:*´<.³=1°:0ºC;Ö\Wìppñsvöx|ûyƒõq|ÿ|„ýw€út}üs{ÿr~ÿr}ÿmxÿgsü`nôI\ÿ>\ÿ>[äAVÔek´•€’¡zƒ�m}™f~”c‡—j”Ÿwž¦� ¨ƒœ¨„’£Š£|€Ÿvzšqyšo|�r� vƒ¤u‚­wƒ³y‹¸‘¼„“¾†—¾‡œ¾Œ¢½�§¼“¯½š¾Â§Ï̹ßÓÇãÒÊÜÇÄÔ¿¼Ìº¶Á²«¹ª¥ZM=ZM=ZM=ZM=ZM=ZM=ZM=ZM=XK;XK;XK;YL<YL<ZM=ZM=ZM=ZL?ZM=\L=]M=]N;^O<\O<\O>]P@[OC[NFYNJZPQ\SX_U^_Wfd\tibƒng‘qj�xp¬€y¼‡€ÄŠ�Ä��¿�~º�t«Œn¢—q¤¡u¨šjž™W{�=Aª5+±5+Â<3Ó@8Ý>8è?:óA=õ=;õ;<ô;@ñ>Bî@IèAIâ@KÚAFÇ<9¼<3¶<1±;/«:,®:+½A5ÐJAÚG@ÜA=Ô63Ð51Õ@9ÍC8»9+²5#½8'Â;(Ä=*Ä=*Ä;(Æ;(Ç:(Ê9(Ï9(Ï6&Ô2#Ú2%ã5*ê8.ï80ð91ç1&ç/%ê.%ï,&ö+'û+)þ,+ÿ-.ü*-ú%+ú$.ÿ'7ÿ$;ÿ7ÿ3ü0õ0ö.9õ3<å*1Û(,Ü.0Þ02Û+.ñ:>ô7=÷4:ö26ö/4ø03ù25ô87Û2-Ú;5×82äE?äE?Î/+Ñ2.Á"Ô51á?:èE@à=8Ø3-Ù2,Ý60Ý:1Ô:.Ï9*É5'Ç5(È9+Ê<0Ê<2Æ9/ícYáXNÉ@8º4+Â91Ã:0¿6,Â9/¿5+¾4*½4*»5)¸6)µ5(²7(°6)´;0»B9ÓYTèljïqtöx|þ|„øv€þ|„üy�ùv~þuÿsÿp}ÿkwüdqý_nöI]ÿ>\ÿ>\éCYÓeh®Žw‰˜oz•bu’\xŽ]�cŠ•k�˜p‹•p…”mp‡]l‰]h‰\g�^l�`q•ey›izŸk�¬t„±vŠ·|�»}�¼“¼‚–½†›»‰›µ†¡³‹­¶—¾½¨ÏÆ·ØÇ¿ÖÁ¼Ï¼¸Æ·´»­ª²¤£\L<\L<\L<\L<\L<\L<\L<\L<ZJ:ZJ:ZJ:[K;[K;\L<\L<\L<\L=\L=]K=]M=]N;^O<^O<\O<_RA]QA\PD[NFZOK[PN]QS]RZ_SgcXvh]…kb‘sk¤vµ†}¾‰}»—…¿™€·›v©žpž®u¢ºv¥µi˜°Rt¸<D¾5-º4+À<0ËC7ÓE9ÛC8ä?9ñ=<÷:>ù:Aö=Eï@GæAHÙ@EÏ@BÀ;6¸90±9+­9*©8(­9*»?3ÏF<ÙD>×96Ù74Ú;7ÜC=ÝOEÎH<µ3#¾9(Á:'Ä;)Æ=+É=,È<+É:)Ì9)Ï8'Ï2#Ó/#ß4*î=3ô?6ñ91é4+à4&Þ4%â2%å/$ë,$î+%ñ+(ò+(ó+-ð(+ò&/ø(5þ%8ü4ü3ø3ô%7ò0;ó5?ò9Að;Dð<Eé5>ä,6ð2<ó/;ö.9ö-7õ.3õ.1ó/1í42Û/+â=9Ú64ß;9ÿmjåCAÑ/-Î/,Ù:7ß@<àA;Û<6Õ7.Ö5-Õ4*Ò6*È6'Ã6%Â6'Ä8)Å;.Æ<1Æ<1Å;1ë`YáTMÉ<3¿0(Ë;3Ñ>6Î;3Ò=6Ø=8×<7Ô=6Ï<2É;1Â:.»7*¶6)¹;/¼@6ÏTMãgeíorøz~ÿ�ˆû|ƒû|ƒûy�ûx€ÿw�ÿs�ÿn}ÿgvû`pü[mõG^ÿ<Zÿ>\ìFZÎ`až~ev…Zi„Qh…OmƒRt…X{ˆ]}ˆ^x„\q‚X`zM_Pa…Ue�Yn•`uœg} j~£mƒ¬r„²tˆ¶x‹¹y‹¹y�¹z‘º~”º�“±•¬€›©†©­’ºµ¢Ç¹®Ë¸²È¶´¸¬¬ª¡¢ —š[K;[K;[K;[K;[K;[K;[K;[K;ZJ:ZJ:ZJ:[K;[K;\L<\L<\L<]K=]K=]L<^M=^M;^O<^O<^O<^Q@^Q@]OB[OC\NE\OG\OI]OO^P_bTmdY{i^‰pgœ|t­…{·ˆ{³™…¸¥…´¯€ª¸z¡ÆxŸÑuœÊdŒÆMlÜFRßA@ÕA=ÏE;ÌH;ÍG;ÓE9ÞA:î@Aõ<Aú=Dø?GðAHâAFÒ@AÆ?<¿?6¶<1¯9+«:*ª9)®8*¼>0ÎD:Õ<7Õ31à;9åA?áD?èSLßUJ¿:+Á9)¾7$À7%Ä;)É=,Ê=,Ë9*Î8)Î5'Ò2$Ù3'å9/ñ>7õ>6ð5.ä/&à6)Þ6)à4*ã1'ç.)ê-)ì*(ì**ê),è',ë'1ò*7ö';÷!9ü>û%Aû6Hä(6ç/;ÿP[ÿ`lÿP\ò<Ió9Gè)8í*8ñ-9ô0:ñ27ê01á+*×(%Ø0-Ú72Ô2/åC@ÿspõVSÝ>;åGDÛ=:Ú<9Õ:5Ò80Ñ7-Ð6*Í4&Æ3#¾5#¹6$»8&¿;,À</¿9-À:/Å<2åXQÜMGÊ70Ç0)×<7ÞA:Ú;5ß<5ç>9é=9æ?9á@8Ù?5Ï<2È:.Á9-»9,»=1ËNHàc_ìnoú|ÿ„ˆý~…øy€øy€üy�ÿx‚ÿt‚ÿn}ýetù^n÷VhóE\ý8Vÿ<ZîH\ÇYZŠlRcrG\tB\yCdzIl}Pt�VvƒXs‚Yo�Wm‡ZlŽ\o•bu�h{¤l�¨o„ªq…«p…­q…±tˆ´u‰·v‰·v‰·v‹¸w�¹z‘´|�«{’¤|›¤…­«–¾²¤È·¯Ç¸µ´«®¡ž¥—’™[K;[K;[K;[K;[K;[K;[K;[K;ZJ:ZJ:ZJ:[K;[K;\L<\L<\L<]L<^K<^K<^M=^M;_N<^O<^O<^O<^N>^N>\O?^NA]OB^PE^OJbNYbQdcTqdZ}le‘xr¢w¨ƒv¤”}§¬†«Â‰©Î‚žÚw”ál‰ÛXwÚD_óBTúDPïJPáIHÐE>ËE<ÐF<ÙD>èBBð@Cõ@GõBHíDGßCDÏA=Â?7¿@7µ=/®:+«:*ª9)®8*»;.Ì?6áFAâ=;îBBëAAÞ;6ãJBæXLÕK>È@2¿7'¼3#Á8(È<-É:*Ë7)Ð8+Î0%×3)à8/é;4í:5í60ë0+å,'ã0,ã1-ç10ì31ó25ö37÷48÷6;ì-5ê,6í.=ñ0Có-Fõ(Fý+Nÿ3Rô9Lå3?ð@Mÿ_lÿkyÿWfõCSõ>Pî3Dí3Aí3>é6<ã99Û83Î4*Ç/$Ë2*Ç.(Ð72òYTö\ZòXVÛA?âHFÜB@Ô<7Î70Í6-Ï8-Ï9+Ê7'Â5#¹8#³9$µ:(¸=-¹;,¶6)¼7.Æ=5ÛNGÙEAÎ50Ð1-á>9ç@:ã81ç51ë20î21ë52ç83á:2Ø:/Ð9.È:.¾6*º8+ÈIBÜ_Yënlû}€ÿ…‰ýƒõv}öw~üy�ÿx‚ÿsƒÿk}ýdvù^pôSeôF]ý8Vþ=ZïL]ÀTRz_BWh<Yq?[vCfyKn~Qt�Vx…Zyˆ_wŒay•exšhz m|§q�ªr„«r…«p†©o‡­rˆ°t‰³tˆµt‡µt…·r‰¸t‹¸w�¶{Ž¬xŽ£x™¤‚­­•À¹§ÎÀµÒÄÁ¼¹À¨ª¶œœ¨ZJ:ZJ:ZJ:ZJ:ZJ:ZJ:ZJ:ZJ:ZJ:ZJ:ZJ:[K;[K;\L<\L<]L<^K<^K<^K<_L=^M;_N<^O<^O<]N;]N;^O<_O?_O?`P@aQAbPFbNPcN]_Pe`Usga…to—{t�}q—�v–®ƒŸÌ‹¡Û‚–ær…ëduéPdé;Rù3Lÿ<QûERéDKÕ?@Î@>ÑC?ÖEBãEFèBDîAEíCFçDEÚEAËB:¿?4»?3±;-©8(¨9(ª9)°8*¼:-Ê;3èIEé??ë=>å78Ú2/ÞA:îZPôh[ÚPCÉA3¾5%Â6'Ç9+È9+Ë7+Ò8,Ñ0&Ú6-æ;4é;4ê40è/,ê-+ë--ó49õ3;ø3=ü3=ÿ1>þ0=ý1=ù1>õ0Aò1Bô4Kõ5Nõ0Nö-Oÿ1Xÿ>_æ3HùO\ÿanÿ_nÿWfûP`ôDXé8JóAQê:GÝ2:Ò/2Ê2-Â8-½;+¼:*Ä:/À2(ÙJBúkcÝLGáPKÏ;9Ï;9àLJÔ@<Ê70Ê7-Î</Í<+Ë:)Ä;(´;&­:%¯<)³>-±9)¯3'º:/ÊD;ÖGAØC=Ô72×2.ç;7ì:6è2.í2-ò,-ô,,ò./ï31è71â:1Ú<1Ò>2Â6)º4)ÅB:×XRçjhú|}ÿ†‰þ€„õv}÷xýz‚ÿx‚ÿq�ÿi{ûbtø]oòQc÷I`ÿ:Xý?[ðO_ºQNqX:Sf8]uEa|In�Sv†Y{ˆ]~‹`��g‚—l€œl}Ÿm|¤o}¨p€©o‚ªnƒ©l…©lˆ«qˆ®qˆ°r†³r„³oƒµp…·r‰¸t‹µv‰¬tŽ¥wš¨„°´™ÈïØ̾ÛÒÍÂÂÌ«±Á›¡±ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8[L9[L9\M:\M:]L:^K<`J<^K<_L=^M;_N<^O<^O<^O:_P;`O;aP>aP>aP>aP<bOAeMMdMW`O_aUmgb€uq’|v˜€s‘�uŽ­€”Ά”ß~‡îqyùclúP[ú;Mÿ*Fÿ2Qÿ<Sñ<Mà<E×@EÕCDÔBCÜCEàBCäADâBDÞDBÔC>É@8¾>3¸>1¯9+¨7'©:)­<*´<,Á=1Ï?7ß=:å99ã03â/2à42Û94æMEülaòh[ÙQCÆ</Æ:-È:.É7*Í6+Õ8/Õ4,Þ7/æ93è62é1/è/,î1/õ15ÿ2Cÿ0Eÿ-Fÿ*Bÿ%@ÿ#=ü :ö :÷'Aõ+Eõ.Mõ/Pó+Pô)Tÿ2`ÿAhóCXÿ`mÿhuøVeíKZñL\ëBUÚ2Cë@PÝ7CÉ-1º((±.$«8&¥?&¨@'¸?.¼:,ÝYMá]QÌF=ÍD>Á63É;7äUQÕF@É91Ç7,É:,È9(È9(Ä=)±?'¦<&¨;&¬=*«7(¬3(¼=4ÓLFÖGAÝE@Û96Ü30é73ì51ë0+ô1-ý,/ÿ+-ú./ô1/í4/æ81ß;1Ø>2É9.½3(À;2ÏNHáa^÷yzÿˆŠÿƒ‡ùz�û|ƒÿ{†ÿx„ÿn~þew÷^põZlëL`õIaü9Wû=YîM]²JGhQ1Qd6\tDf~Nuˆ[Žc„�h†’jˆ—n‰žsˆ¤t„¦t§r}¨p¨n�©mƒ©l…©l…¨n‡«n…­o…¯o�°l€²k‚¶n…·r‡³t‡­t�§wœ¬…²¹šÈƯ×νÙÒÌ»¿Ê¤¬¿’š­[J8ZK8[J8ZK8[J8ZK8[J8ZK8[J8ZK8[J8[L9\K9\M:]L:]L:^K<^K<^K<^M=^M=^O<^O<\O<`Q>aR=aR?bQ?bQ?aP>aO;bL>hOKgNRbP\cWkjfuv”}|œ�x—‘z–«�•Æ„’Ú}…ðuzýklÿY[ýDLþ/Iÿ5Rÿ<Uó<Né?LãCKÝBHÖ=@ÛACÜ@AÞ@AÛA?ØC?ÐA9Ç>4½=0µ?1¬;+©8(­<,´?.»@1ÊA7ØC=Ù74á85Þ..â30æ:6Ó0'Ñ5)ëUGÿwjë[PÑC7Ê<0Ë;0É7*Í6+Ô:0Ú70ß82å63ç32ê01í12ô36ý4<þ%6ÿ#9ÿ"9ÿ"9ÿ!8ÿ6ý5ø6ï2î9ð%Bï(Gî&Jñ'Mÿ3\ÿEhÿbuÿ`l÷WcñQ]òP]ïMZâ@MÔ2=Ú=DÌ79»/.®/(¦5'¢;(›?&œ>%ª9'²8+ÑUI»;2Ä?8¾41½2/ÎC@åWSÕHAÈ91Ä6*Å6(Ä5%Æ5$Ã:(±<(§:&¨7%­9*¬4&¯1%Â?5ÛRJÙIAáHBÞ;6Þ2.ê41í1/ï-+ü22þ,/ý+.ù+-ó++í-*æ1*ß4,Ö8,Ñ=1¿2(¼7.ÈIBÙ\Xówwÿ‰‹ÿ‡Šû€…þ†ÿ}ˆÿv„ÿj|þ_s÷XlöTiêD\õD^õ6Uñ:VàLZ¥HC^K-Mc5Uo?b}JtŒ\€–g‡˜lˆ›nŠ r‹¥vŒ¬z†«w�ªr~§m|¥i}¥g�¥g‚¦i‚¥k‚¨mƒ«m‚¬l®h~°i�³j„¶o‰¶s‰¯rŽ¬x™°„­»˜¿Å©ËʵËͶ½Åžªº‹—§]K7\K7]K7\K7]K7\K7]K7\K7^L8]L8^L8]L8^L8]L8^L8]L:]J<]K=^L>\L<\L<[N=]P?^SA`SB_R?`P@bQAcRBdQBdNAdLBjMIhKMdPYf[lnk†w{ž€…­‰ˆ²�ƒ©¥‡©¾ˆ Ð‚’ây~ðsqújbÿa^ùJWñ>Që:Lî?NïCQê@Kæ<Gå>Fá<Bß<?Ú<=Ö<:Ï>9Ê=4Á;0¸:,§6&¯A0­<,¬6(¿A5ÌF=Ì=5Ð72Ú85Ü71Þ5.Þ6-Þ8*Ý<*Ù:'Ó6%Ù;0ðSLòZOÙE9Ç8*Í>.ÔB3Ò:-Ô3+ðGBá//î5:ó6=î,7ÿBMñ'3ú&2ÿ)2ÿ(2ù'ý"*ÿ'/þ'/ð&ð)ù+7ê .ì&7ñ-Aç#;ÿXrç-Dÿ]lÿ_iþZcùU\õPWëHMÝ?@Î95Ë>7ÄA7·?1¦7&™2!™6#š;'Ÿ<)§8-®7/ÂE?¿;7½31Ä64Å54ÔDCÛLHÚKEÑD;È:0Ä4)Ç5(Ë4)Å5*¼</³9,­/#«(³-$Ã9/ÒC;ØH?äPFßE=Ú70Ý0*ç0,ò21ù13ý14÷/1ö01ö01ó0.ò0.í2-æ5-Ý9/Õ=0¾2%À>1¹=3Ö_Yésqÿ��û„†û‡Šÿ„Œÿxˆÿh}ÿ_wÿ[uÿUoøLføEcñ:Yô;[æ?YÛ]i‹@;VF-L^6YuBg†Mw–]}œc�g€¡jƒ¦n‚¨o€§n¦m~¥l}¥i|¤h}£f£e£f~¤k¦m�«l¬k~­g®f�±gƒ³i‡µm‹´r�°vŽ®|™´‰§½™±Á¤²Á®¦²°�˜žz…‹]K7]K7]K7]K7]K7]K7]K7]K7^L8^L8^L8^L8^L8^L8^L8^K:\K;^L>_M?^N>[N=[N=\Q?^RB_SC]QA]OB_OB`NDbPFcOHeNHiMJgNQdR^g_tnn�y~¨€‡»‡ŠÁž•Ê£‹»«£¹x�Óyƒê|{ôvjöi`ø]aóOZìFRòHSøIVõEPï?Jë;Eä9?à:<Ú::Õ=8Î?7Ç>4À</·<,«<+ª?-©8(²:,ÁA6Å<4Ç61Ô<7Ò50Ö5-×5*Ø4(Ú7&Û:&Û;%Ú9'Ý6.æA;ëMDáI<Ð>/Ç8(Ê8)Ó;.Þ;4öJFä01ð6;ö6Añ.<ÿ@Nï'4ó$,÷%(û),ý+,ü*+ø((ö((÷+,ï$'í%(ë&-ý;DÙ'ð4CüARØ&6ÿjuÿbjþZaüW]ôOSåBC×<8Ñ>6¾8,¿D5¹H6©>,™/�)�(’(ž-%°:6ÑTPÎHGÃ54Æ45Ñ==ëWU×FAÖG?ÐA9Ç9/Æ3)Ê6,Ð7/Î70Ç=3Â91Â5.Ã2-É4.Ñ83Ø?9ÞE=äJ@ÞA8Ú6-Ý0*ç0,ò21û03û03ô02ó12ô01ó/0ó0.î1-ç4-ß9-×=1¿3$¾</·=2Ó^Wévsÿ��ù…ˆø…Šÿ‚Šÿt…ÿf|ÿ\vÿXtÿQoþIhü@aó:Zñ?_áI`ÁS\{?7RF.M_9ZvCe‡Ks•Yxœ_w�b{¡f}¥i}¦j|¥i|¥i|¤h{£e{¡d{¡d}¡c~¢e|£j}¦l}©l}ªi|«e}­c¯c‚±c…°h‡²kˆ¯pˆ«s‹«|�¬„�ª‡Œ¢‹x‡€dppR^^]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8^K:_N>aP@`P@]P?\O>[O?[RC[QE[QG\OG\OI_PMcQOfTTgUUgUUgV\h]nli†sv£|ƒº‚ŠËŠ�Ô˜‘× ŠÊª‚´¹}ŸÏ€“䄆îvósjüfh÷S\íGQðFOöHR÷EQòAKî=Eä7=à7:Ù99Ô<7Í@7ÅA4¾?0¶>-­>-ª<+«:,¸B6ÇH?À;2»0)Ê:2Í6/Ð6.Õ7,Ø8,Ù7(Ù8&Ü9&à:*Ý4-Û4.âA9éOCÞH:È5%È2#Ù@2â>5õHDè13ò5;ù7@ø4@ÿESø2?ð)0ë#&ì$'ù13õ-/î&&ì&%õ12ë)*ö8:è+1â)1ë3=ÿR^ï=Kÿ\iþ^fñV\îOTðMRêEIÝ:;Õ74Ô?9À6,À>1¼B5³=1¬7-¥3)¡,#œ' š%©/*ÇECÂ::¼,,È35×ABô^]Ñ@;ÒC;ÐA9Ë=3Ì9/Ô=4Ú@8Û@;Û@<Ø=;Ý>;ãA?ä>>à::ã?=ëIDàB9Ý<4Ù5,Ý2+ç1-ð31õ12ø02ô02ô02ô01ó/0ó0.î1-ç4-ß9-Ø>2À4%º8+µ;0Ð[Tízwÿ“’û‡Šúƒ‰ý~‰ÿr‚ÿdyÿ[tÿTqÿNmÿFgû=_ô=\îEbßTg¢BFp<1QI2Pb<]xEf‡Nr“Zu˜^tš_xžc{£g|¤hy¢fy¢fy¡cx byŸby a| bz cz¢f{¤h{§h{¨e{©a{©`}«`®`‡²jŠµn‹²s‰¬t†¦w‚žuy“no†lSeYAPM2A>^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^M9^L8`N:aP<`Q>]P?ZN>YOCXPEXOH[QO_UTbXYeX_iYcl\fk^glaildqok‚rtšy~µ†Ê…ŒÚŒŽáŽ‡×Ÿ‹Ó­ŠÄ³€©½yŽÌ{�Þ�zò‚wÿsrÿcgùU\ôMTõGPôCMò>Gï<Bå6;à88Û97Ô<7Í@6ÅA4¼A1·?.«:*­>-±=0¼F:ÌRGÇH?º5.º1)È91Í81Ó;0Ù=1Û;-Ù6'Ü6&à8+à5-Û2+à=4ìNBãK=Ò:,Ñ8*àB6ß7.ð>:é/0ð16÷4<ü8BÿNZÿLUý>Eò38â$&í/1ê,,ð22è**ë/.ò::Ñæ37Ý,2è9@ðENÿXcòNWØ=CÚADáBFá>AÞ9=Ü89Ø88Ó97É83Â91½:2ÀA:ÍNHÖVSÐNNÃC@²72¶95ÏKIÙMLãQRïYZêRQñYVÉ51Î>6ÒB9Ð@7Ó?5ÙB9àC<â@=é?@ç8=ì9=ò?Cï:?è58ì>=öMJÞ93Û60Ú3+Þ3,å4.ì30ò21ô01ô02ö/2ö01õ/0õ/.ð0-é4-à8-Ù?3Â6'·5(³9.ËVOï|yÿ“’üˆ‹ý�‰þz‡ýoücwÿXqÿOlÿFhÿ@cù:\ô?^éKdØZh…43f<.VP8WgCa{Kj‰Ps’Yt•\u˜^w�by¡ez¢fw dv awŸavž`xŸ`xŸ`{Ÿaz cz¢fz£gz§fz§bz¨`z¨]|«]~­_†²g‹´nŒ³tŠ­s†¦u€�qtŽii€dQcUDSL8G@^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M9]K5_M5`P9_P;\O>ZN@XOFXQKYPQ`W\g`hnfsshysg{sg}qh}ol}pq†rw—v}«z‚Á‡Ò„Šà‹Œæ“�ã�ŒÚ „¿žuŸ l‚¬nq¼vlÒviòsmÿllÿgiþ^`ûRWõFMï<Bê7=æ5;á78Ý98Ö=7ÎA7ÆB5½B0¸@/­7)´@3¯:0°;1ÈRHÕ\QÉMC»;0¾90Á7-Ì9/Ö>3Û=1Ú6*Ü4'â6*à3,â70ä=5åE9âF9ÞB3ÞA2â>2Û0&ê72ì0/ï.1ñ.4ú7?ÿPXÿ]dÿZ_ÿOSá.1ã03à,-÷EEè66Ü,,Û-.ë>@Ü25Ù37ÿbhà<CË*2Ê-4Á&*Õ;=á@EÛ6:Ù37ß9=Ü7;Í/0Ì43Í;;ÕGFÞRSì`cõilôekê^aÄ@>ÆE@å^[øhgÿopÿjjåKK×=;É40Ñ>7×D<ÖC;Õ>5Õ<4Ø:1Ü41î5;ñ-9ð,6ò.8ï,4ë,3ó9<þJKß3/Ü3.Ú3+Ü3,ã5.ê40ï4/ò21ô02ö/2÷/1ö.0õ/.ð0-ê3-â7-Ù?3Æ8*´2%²8-ÃNGï|yÿ’‘ÿˆŒÿ‰ÿw…úl|øatûTnýHgÿ>bÿ:_ú6\óAaàNcÂW_p/+_?0[W>^lIe}Mk‰St’\u–_v™_xžcy¡cx buŸ`uŸ`u�_vž_wž_xŸ`{Ÿaz¡bz¢d{¥e|¦d|§`y§^y¨Z{ª\}¬^€¬a…¯g‡®m‡ªp…¥sƒ r{•nrŠjexe[k^RbU_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N:^L4_N4`N8^O:[N>YOEZSMZTT_Zahbprm�zuŒ{u‘{q“zp“vp’ps’rxšt}¨w�¶zƒÄ~†Ï„‰Ú‹ŒÞ–�Ü™ŠÍ›…¶¢ƒ£ª„‘«~yŸl[šQ>¼MBÔQIâWRëWUñSRñKMí>Cå4:é9<æ9;à;9Ø=8Ó@8ÊB6ÀA0º?/³9,ºA6«5+£.$¾I?Ùg\ÙdZËUI¸<2¸6)À3)Î:0Ù;0Ú6,Þ3)ä6-ß1*ä71ã;2Ý9/Þ>0ãF5äB3Þ8*ß1(é4-ð31ï-.í*.ô17ÿEJÿV[ÿ]bÿ\aâ9<ã9<Õ+,ôJKâ89×//Û57Ì()Ì'+øXZÝ>BÖ:>Ñ6:º#Ä(+Ü<>æAEÝ49Ù,2Þ17Ù05Í(.Ò37äKNöaeûhnùcl÷`iú`løbkÖLLÕNJê\ZêVTçMMãHFÏ42Î3/ÙA<àKDäOHßJCÙ@8×:3×6.Ü.-õ1;û-:ø*7õ'4õ)4ø0:ÿ>CÿJJà1.Þ3,Ü3,Ý5,á4-ç4/ì3.ð3/ô02÷/2ù.1ø-0ö..ñ/-ë2-â7-Ù=1É;-³1$±7,ºC=ìyvÿ��ÿˆ�ÿ|ˆÿq�öhxö_rùRlùDcÿ9^ÿ6]ù7\ðFcÑN^¢IKe5+VD0[Y@^lIe|NkˆRt‘[w•_x™`{ždz cvž`s›]rœ\s›\tœ]v�^wž_{Ÿaz¡by¡bz¤b|§b{¦_x¦[x§Yz©Y{ªZ~ª_ƒ­e…¬k‡©m‰§s‰¥u„�s|”ru‰pnlfwe_P=_P=_P=_P=_P=_P=_P=_P=^O<^O<^O<^O<^O<^O<^O<_N:aP6aP6aO9^O<[OA[RK\VV_Zakgvso†}z—‚~¡�|¥}w£{u¥xv§sy©s}°x�¸{„¿~†Ä„‰É�ŽÎ”“Íš“Ç •½«›µ¼¦²É­©Å¦”ªˆmœfJœE2¯A4¶@4ÂB9×IEèNNîJKì?Cë=?é;=ä<;ß=:Ø?7Ð@5È>1Á=0½:0¼@6¯4,§1'»I?ÑcVÖh[ÕeWÃOBº>2º4)Ç7,Ö90Ù5,ß2+ç60â0,ã4/â7/Þ7.à>1åC4ä>0Þ4%æ4*é0+ð3/ð..ï-.ò14ó78öBEöMRüY^éFKöQWØ37ëFJÝ8<æAEÚ7:Î.0Ñ05äHKÒ7;Ð7:¹#%Õ;=Û8;å;>ë<Aç6<â-6Ý)2á1;ç<DðLSý]eÿksÿjsú`jöZgùXh÷[fôddèZVèWTØ@=Ï10Ô44Ó33áC@ãJEéRKêSLáJAÙ?7Ø93Û81ã33ó.7ÿ0>ÿ0>ú*7ü.:ÿ;EÿCHûABã1/ß4-Ý4-Ý5,â4-æ3,í4/ñ40ô02÷/2ù.1ø-0ö..ñ/-ë2-ã7-Ø:/Í>0µ1%±7,²;5ésqÿŠŠÿ‡�ÿv†þi}óató[p÷PjúCbÿ8`ÿ8aù;_ìLf¿O[~86]=0NF1VX@ZeEcxMj…RrŽ[v”^yša{ždxžas›]p˜Yp™WqšXqšXt›\uœ]y�_xŸ`x¡_y¢^z¥^y¤\x¤Yx¤Wy¥V{§X©_‚¬d„«h†¨l‡¥o†¢q€šmx‘jq…ikdbv]_P=_P=_P=_P=_P=_P=_P=_P=^O<^O<^O<^O<^O<^O<^O<_N:cQ9cQ9aP<]P@ZPG\TR_Zaa_mtrŠyx˜�€¨ƒ�°}}¯yy¯yu°wwµw|¾x€Á}…Ä‚ˆÄ…ŠÁ��¿—–¾ �¼¬§»¶¬µ¾³±È¸«Ð¾¦ÖÁ¢Ñ»–Ьˆ¯oT®\G¤G5¨>0ÂG?àWQïWVñNOé@Cê=?æ<=â=;ß=:×=5Ñ:1É90Á80À=5¹:3·>5ÁOEÇXMÇ]OÑeXÓcUÂL>¼:-Å8.Ò91×4-ß2,ç60æ40á2-à5.ã<3çC7æC4â:-ß3%ê5,æ.&ð0-ï/,ð31ñ54ç-.â24çBF÷X]öW\ÿouåDJëHMà<Cþ[`ÿouÞ=Bÿ}‚ÿmqêQTÔ>@Æ02Ö:=òHKî9>ð7?ø=Fò7Bè-8ô>Jÿ]gÿmwÿjrÿcn÷]gú^kÿaoÿ]nùYeò\]ãPIâJGÕ74Ò/0Ù54Ò0.Ø95ßD?ãJDâIAÙ@8Ó6/Ö4/Ü71é77ê'/ü0<þ2=÷+6ù0:ÿ>Eý?Aì44â3.à5.Þ6-ß4,á3*è3,ï4/ô41ö01÷/2ù.1ø-0ø..ò/-ë2-ã7-×9.Ñ?2¶0%²6,¯50çonÿ‡ˆÿ†Žÿp‚ücxñ\pôXn÷OiûBbÿ9aÿ<döAbäRg°QWd/'WD3GG/RT<U^?btLi�Qq‹[v’_z˜b|�dxœ_r™Zn—Un—Un—Uo˜Vr™Zt›\w›]wž]wž[x¡[y£[w£Xv¢Wv¢Uw£Ty¥Xz¤Z~¦_�¦c¡d~œf{—frŒ]i‚X`vRZoPRgH\L<^N>`P@`P@_O?_O?_O?aQAbRB_O?]M=]M=aQAcSCbRB`Q>dS?aP>]P@^QI[QPZSZeapsq‰yxšz|¥~€±€�¹}ºxz·rs´lo¶twÈv{Ì€ƒÊ‹�Æ–—䤾²°»¼¹´ËƳÓÌ°ÙϬÛϧÜÏ£ÝУÙ˜٘Ѩˆ¸€g ^HËzgÖtg½H?×QN÷ccéKLå?Aâ89ä::â:9Ü75Ù61Ö;6Ë81À5.ÇB;¹=3¿I?ÆXK¹OAÎdVÑcTÌXIÅG9Ä:/Í6-Ø7/á6/ã4/ß0-ß2,à70ä=4à<0Ý7)ß5(å7,ä/$ê/&ì/)î1-î20ë40è51â64×79æJNõY]íNS×8=×6<ùX`ÿ~„ÿ~†ÿouÿgoÿipÿflõ\añX[ûZ_ÿUZÿDKü8Bû7A÷3?î+9ì0>ñ=HÿS_ÿ`iÿgrÿepþbmÿcnÿ`oý]gíSSßF@Ö85Ö41×11×/.Ø31Û94áD=ÜB:Ù?7Ö<4×90Û81à93æ66ñ7<ó4;õ3;ö4<÷6;ò59ë33ã1/Þ3,Þ5.ß7.à5+ä3+é2,ï2,õ20÷12ø03ú/2ù.1ù//ó0.ì3.å7.à?5Ì8,¿6,¯/&¸;7ÒXWÿ˜šþ{ƒÿm€ÿe|ø\rðPhôGeüCeÿ;bù9^õNlÓRd‰=?Q-!C;&CG.KM5Z`DZjEcxMqˆZx’bx–`t•\r–Yq˜Yl•Sm–Rn—So˜Tr™Xt›Zuœ]uœ[t›Xt�Wv Xw¡Wv¢Wv¢Uu¡Tt S{¥[{£]zž^u—[m‹Ub~MXrCSj>AW0?U/<R,[K;^N>`P@`P@`P@_O?`P@aQAcSC`P@_O?`P@bRBcSCbRB`P@fVG_QF\OG^TSaZad`omkƒvw–wy¢{}®}¸x|¹quµkp´lo¶nq¾uvÐ|~Õ‰‰Ó˜˜Î¨§Ç¸¶ÁÉźÓαÛתâÛ¥çÞ¥æÝ¢äÛ¤âØ£ÛÑ ÔǛӼš»œ€©|eÅ‹wÊ}m´VJÒa[örpìZ[çKLã?@ç=>ë>@ç;;â66Û64Ö=8É83Ê?8¹6.¸>3½K@¶H;È]MÝo`Ô`QÆH:½5)Å1'Ô6-Ü5-Û0)â51å95ä=7à<3Ü8.Û5)ß4*ã5*ì7.í5+í2+ë0+ç0,ã2,á4.Ý52Ö87Ø>@äJLóW[ú^bü\dþ]eÿaiôS[ïNVñQYÿ_gÿjqÿflõZ`óNTí4:ñ,5ô*6÷-9ñ'5å!-æ'6í7CþPZÿZdÿblÿdmÿenÿcmý[hõQZåEEÞ<7Û64Þ63ß55Ü30Ú2/Û62à>9Û>7Ø;4×:3Û81ß82å95ç:6ê7:ë7:í6:ï79ï79ì65å31ß2,Ü3,Ý6-ß7,â6,å2+ê2*ò1,õ20÷12÷03ú/2ù.1÷//ó0.ì3.å7.ß<3Ô=2Â8.¶1*·74ØZ[ÿ�“ûs}ÿfzÿ^vùWo÷OiùFfû?bü<aôAaåOh¹KVx;8L2#<:#<B(EG/OS8VdA`rJnƒXu�_v“]t“Xq“Vp”Tk’Ok”Pl•Qm–Rp—Vr™Xsš[sšYuœYu�WuŸWuŸWtžTržQq�RpœQn˜Pl”NhŒLb�FXu?Ni6F]/@W+@W-@W-AX.YL<[N>^QA^QA^QA^QA_RB`SCaTD`SC_RB`SCaTDaTD`SC^PCbUM\QM[QRcYbkdtqm„vu•z|¥y{¬y~¶x|»nsµei°`g¯jn¸ruÆ}}ׇ„Ý—”Û©§Ø»¹ÏËÊÈÜØ¿åá´éäªíè¦ðé¥íå¤çá§âÛ§×Ï¡ËÃœ¸¯�©›�›ƒk¨ƒp§o`œRGº]UÝmi÷uuï__äJLå@DëADì>@è8:â66ß=:Ô;6ÔA:Â70·7,·>3®=/¶F8Ûj\ßi[Ù[MÊB6Å2(Ñ4+Ù6-Ù2,Þ52éA>ìGCà=6Õ2)Õ1'Ú2'Þ0%æ4*è1)ç.)æ/)ã2,â7/á=4àB9Æ-(Ê43ÜDCðVXú^b÷ZaóS[ñQ[èFQêHSïOYü\fÿmvÿntôY_ã>Dé.5ô-4û2<ü2>õ+9í'4ö6EÿKW÷ISõPWøW_þ^fÿckÿ`i÷RYéBIÞ97Þ71ã75é<8ë;;æ95á51Ý60Ý:3Ù;2×90×90Ü71á83ç:6é<8ã99ã99æ87è88è88æ74á40Û2+Ú6-Û7-Þ8,â6,ç2)ë0)ó0,÷1.õ12õ13ø02÷/1÷//ò0.î3.ç6.Ý90ÜD9Ç80»4.µ/.ä_bÿƒˆølwþbxþWsúPmþLlÿCfö:]ó;_ìIfÔSg�DJl;4O>,=<':?(ED/FH0Q\<YkEg|Qr‰[t�Zq�Um�Rk�Oi�Mi‘Ki’Lk”Pm–Tn—Up—Xq˜WuœYu�WtœUrœTpšRm—Mk•Mj”Le�GaˆE[AUt;Li3E`-AX*>U'B\/E_2Ic6WJ:ZM=\O?^QA^QA^QA_RBaTD`SC`SC`SCaTDaTD`SC_RB\PDZOMZQVbXcjcsqm†vu•xz£z|­w|´sy·mr´bi¯]d¬ag±lr¾x{ʉ‰ß•‘夡䶵áÇÄÙÕÑÎáÝÄçä¹ëæ¯îê­îè¬éâ«ãÛªÚÓ©Ë¡»¶™¥¥�›™„“ˆv˜�q™rc›bW³i`Ïrmû‹‰õutê[]æLNèCGí@Dì<?é;<ß:8Û<8ÞE?ÓC;Ã:0º:/°6)§1#¿I;ÛaRêk\ÛSEÉ:,Ì4'Ô6+×4-Ù42ñMKþZXíKFÚ91Ö3*Û4+Þ2(ß.&á,%á,%Þ-%Ü1)Ú6,Ù;0×?4Â/'Ä4,Î:6ÙCBßFHàEIäGNêMVñQ]öVbøXbù\eÿgpÿmuôYaÞ;@ê17ñ.4ò-6ò*4ë%2ì(4û=IÿS_í?IèAIêGLôQVÿY`ÿX]ôGMä68á53ä73ê;8ò@>ô@?ï=;æ95à72Ý82Ú91Ù80Ù80Ü71ã73è96é:7â:7à;9ã:7æ:8ç:6å84à5.Ú3+Ú6,Û7-Þ8,â6,ç2)ë0)ò/)ö0-ô22ô23÷12ö01ö0/ò0.î3.ç6.Ü8/àF<É91¾3.¶,,îfjÿy€øgtÿbyýTqþKkÿGjÿ>cò6Yë=^ãQhÄXeˆBBgC7[N;GF1AC-JH3EE-LU6Sc>btLlƒUp‹Xn�Rk�PiŽKg�Ih�Jh‘Kj“Mk”Pm–To–Wp—Vt›Xs›Uq™So—Pj“Mg�JeŽHd�G`‡D]�AWy=Rq8Mi6Lf6Ne7Of8Lf7Oi9Sm>UH8WJ:[N>\O?\O?]P@_RBaTD_RB`SCaTDaTD`SC_RB_RB^QIVMR_Xhnf{tp‰xt—xw¡xz«y{´sy·jr³`g­[c«_g°jr»v|Æ€„Γ•àžžä­®ç»½æËÉßÔÓØÝÛÎáßÆåá¾æâ¼äß¹à׶ÚеÎƱ¾µ¦­§›˜šŒŒŽ€‰…yŠ|q‘vk›si¨meºoi扄î�~ósrñedñWYðMNîDEì@@ã:7à;7äB=ÞE?Ì<3Æ=3ÄB5¯1#®0"ÇI;ÜXKÖN@Ì=/Î8*Ñ7+Ñ3*Ö42õUUÿkhú]XãD>Ü;3ß;2ã80â4-ã2,â1+Ý0)Ô0'Í/$Æ/$Â0#Å9,À4'Á3)Ë:5×CAÛEGÛBGÚ@HßCNëOZòVaòVaú^iÿgoóYaáBGê:<í57ï28ï28ò2=÷9CÿHSÿV_ä6?Ý4;ß6;ëADùJOüIMô;Aé13è51ë95ð=9ò<9ó;9ð:7é73â70ß;2Ü;1Û:0Ü90ß82ä73è64é75æ:8ä;8æ:8è:9è;7æ95á6/Ü3,Û7-Ü8.ß9-á7*å3)é1)ñ0)ô1-ó32ñ33ô22ô01ô1/ò0.î3.ç6.Þ7/àC:Ê70À1-À13økqÿqzúftû]túPmÿCfÿ=dÿ9cõ8^æEdÙZm«V[wE>dJ=aYFTO;KI4PI6GE.FO2L\8ZlFf|Nm…SmŠRlŒMh�JfŽHg�Hg‘Ii’Lj“Ol•Sn•Vo–Ut™Vs˜Un•Rj’Lf�KcŒHaŠF_ˆFZ�BX|?St;Pn8Ok:Rl=Wn@ZqCUo?Vq>WqARE5UH8XK;ZM=[N>\O?^QA`SC`SCaTDaTDaTD`SC`SCbUEaVP^Vcje|yu�}yšzy¡xy©vy°sw´jr³cj°[c«]e®hp¹v~Ç�‡ÑˆŒÓ™�Ý£¥à¯³ã»¿åÆÇãÏÎÞÔÓØ×ÖÑÙ×ËÙÖÇ×ÑÅÒÉÀÌþŹ¹´¨¬£š�Š‹†x{tyvqulele�ogŽa[�`[ÀsmØyuð{ü|yÿrpýccöSTðHGïC?ê>:á<6àA;Ñ:1ÓC:ÝTJÂ>2¸6)½9,Â=.È?/ÏB1×C5Ö@1Ñ7-Ð2/êLMú_]òWSßD?Ù:4Û81à70æ93ç92ç92â:1Ú9/Ñ9,Ê8+Å9*°(¸3"È@2ÓI>ÒD@Ë;:Ì7;Ï:@Ñ7AâHRíV_ðYbõ^g÷`iêU[ÜAEå@>ì?;ó@CûGJÿLTÿQYÿT]ÿV^á4:Ú/5Ü/3ç7:ô=Aø;?÷48ó12í42ï95ð86í42ë20ê20ç40á4.ã<4à=4ß<3Þ;2á83ä73è43è43è88ç98è88ë99ë97é75â5/Þ3,Û7-Ü8,Ý:+ß7*ã2(ç0(ï0(ò1,ñ42ð43ó32ó11ó1/ñ1.ì3.ç6.à91Ù<3Ì71Á0-Ñ@CÿnuýkuùbqðRiõHfÿ;aÿ4^ÿ6bû>däNiÊ\i‹GFjF:]L<]VCXQ>RJ7OG4ID.DJ.GU2Sc>^sHgMj‡OjŠKgŒGgŒFfŽGf�Hg�Ji’Nk”Rm”Um”Uq•Up•Rj‘NgŽKcŒH`‰E_ˆF^†G\‚E[~DWxAVr?Uo@YpB]rG^uG[sCYtAXs@NB2QE5UI9WK;XL<ZN>\P@^RBaUEbVFbVFaUE_SCaUEfZJh]YngwupŽ|x™{y wx¦vw­pu¯jo¯ah¬^f®_g°em¶pxÀ}…͉�Ø�—Ø™ Ö¡¨Öª±Û´¹ßº¿ßÀÁÝÃÃÛÇÆØÆÃÔÅÁÐÁ»É¼³Ä»°Áµ¨¼¦˜¯–‹œˆ…Œqqqtpopfewhe†nj„_Y•d_›ZT¹f`Ùsnðzvÿ{wÿwsÿhfüZWþRNõHBæ=6çD=Ø;4ÛG=ê\RÅ;0Ä<0¼6*¾6(Ç>.ÑD3ÔB3ÔA1×A3Ò86Ø>>ÝCCÜB@Ö=7Õ81Ø7/Þ7/å:3æ93å:3â;3Û>5ÕA5ÑE8ÌG6®-¼;%ÍJ8ÑL=Æ=5º/*À00Ê9>×AJáKTëU`ðZeð]eìYaßLRÔ>@Ï4/Ø7/à;9ê@@òCHõDJñBIî?Fâ5;ß26ã36ï8<õ79ø14û/2ü14ï20ð95ð74ì0.è,*ç0,æ3.à3,ä=5à=4à=4à<3ã:5æ95ê65ì65ì57ì57í57î68î66ë54ä2.Þ1+Þ7.Ý9-Ý:+ß7*á4&æ1&í1(ñ1,ð42î53ñ42ò21ó1/ñ1.ì3.ç6.â;3Ò4+Ò:5Æ20èVYÿpxùcnòZiçI`óFdÿ8bÿ/]ÿ5bÿBhÜLe±PYj5/^G7XK:UN;[P>VI8M@0JC0@F,AO.K[6Xj@bxGf�JhˆIf‹FeŠDc‹Bc�EeŽHg�Lh‘Oi‘Rk’Sn’Rl�Ph�Nd‹JaŠH_ˆF_‡H_‡H_…H_‚H]~G]yF]wH`wKcxOczN]uC[vAZu@JC3MF6QH9TK<WK=YM?\O?^QAdWFeXGdXHbVF_UIaWMg^Ulcfwr‰yušxvžtsŸrr¤tu­lp¯bg©Za©_e¯fl¸ms¿v|È�‡Ó�“Ý—œÞ›¡Õ §Ó©­Ú®³Ý±´Ý²´Û´´Ú¶´Ûµ±Ö³­Ñ®§É«¡ÄªžÂ¥™½™‹¯ˆ~™|x†dbgiefcYZj[X{c_yVP’c]ŠNF¢UM½`XÖkcñwrÿzÿzwÿnjþa\÷TOéGBîOIÝB=ÞG@êWOº,"¾1'¾4'Ç>.ÑE6Ë>-Á2"Ê7'ÛE7ÞE?Ó97Ì3.Í50Ô;5×=5Ý<4á=4ä;4ä92Þ5.×4+Ï5)Ê8+Å<,Â?-ÔS>ÈG2½<)¾:-Ã=2Ä;5Â74Â43ÜKNÞMRáPUåTYèY]åVXÜMOÓD@É5+Ñ7+Ù80Ý52â24å26æ18æ.6ì4>ì3;ó5?û9Bý4>ú+5ÿ(3ÿ/9ø-3ù57ù59ð./ë*+é0.è51â5/ã<4à=4à=4á=4ä;6ç:6ë76î66ì46ë35ì25í34í55ê41ã1-Û0(Þ8,Ý:+Ý:+Þ8(à4&ã3$é3&ì4*î50î52ð42ð3/ò2/ï2,ì4,ç6.ä=4Í/&×>9Ì35ùbgÿoy÷[iðNcêC]ùFfÿ:eÿ0]ÿ6aöGhÉJ]“CFR,#VI8SL:OH6[O?UI9F?-HF1>B'@J(IU1Sd:_rEe}Ig„JeˆHcˆCa‰BbŠCc�EeŽHf�Mh�Pi�Ql�Pk�Mf�JcŠIaˆG`‡F`‡H`†I^�G^F_}G`|Ia{Kd{Mf|Nf~N]xC]z@^{CCA5FD8JF;PG>RH>WI>]M>_O?`Q>aR?`SB_VGbYRgb_ojnso~wq“zt |y¨yw©pp¦gg£`a¢\_¤`b­fhµopÀyzÌ„‚Ö�‹ß–’眙栞ݡ¡×££Ù¤¤Ú¦£Ø¥¢×¥¡Ô¤ Ó¥žÑ£�Í¢šÉ¢™ÈŸ–Ù�»Ž‚°ƒ|žrn}igld^`fZZkYUrWP}WNŒWO•RI¨ZPµXPÁZSØkdìyrù‚|ÿ‹†ùuqÿusí_]ßPLÙEAïZTÌ71Å2*Ä6*Ã7(À7'Ã7&Æ9(Ê;+Ð<.Õ?1Ò8.Ö<2×?4Ó;0Í5(Í3'Ø:/â@5à90á90Þ7.×7+Ï7)É:*Æ?,ÄA/Â=.Æ@4ËE:ÍG<ÌC;Å<4»2,¶+&Ç<7ÐEBÙOLàVSãYVáXRØOIÐC:Ì;*Ó:(Ü8,â5.è./ì*2ð(3ô'6ú->ú+?û*?ý)?ÿ(?ÿ&?ÿ$<ÿ$<ÿ$8ÿ'9ÿ+<þ,9ó+5í.5è45ç;9Ý84Ý:3Þ;4à;5ã:5æ95é73ë54ë35ì46ë54é54ç53â5/Ý5,Û5)Ü8,Û9*Ü9(Þ9&ß8&â7%ã6%æ4&ì5-í4/ð50ñ4.ñ5,í5+ê5*å7,à90Ø7/Õ:6Û?BÿmwÿbsûSjõIcÿKjö7Yÿ6_ÿ<gø>cçNj¯IV]#!D, ?:'HE4PM<PI9EB1>C/?G/BH&EK%KS.S]8[kDavKd~NdƒJb†F`ˆBa‰@c‹Be�Fh�Jl�Pm‘Ql‘LiŽHgŒGf‹HfŠJe‰Ic…HaƒGdƒJd�IbGc~GdJdHe~Ga}C[{<\=_‚B=?4?A6FC<JE?PE?TG>[K>]M=`O;`Q<^SA^VIc[Xhcinlyso†tp•vpžtp¢pm¢gfŸaaŸ]^¡\\¤ed°kjºtrÅ}{φ‚Ù�‰à”Žæ˜“å›–Ùœ˜Õ�™Öžš×ž˜Ö�—Õœ–Ô›•Ñž–Ñœ”Íœ“Ê›’Çš�Ä“‰»ˆ®‚x�tk|mdif]^f[YiZSoWM{UJŠVKšVK°\RºYRÁXRÓfaàqjãvoí~wý‡…ÿ‡…÷usîgdãUSëZUÇ2,Â/%Å7)Ã7&À7%Á8&Ä8'È;*Ï;-Ó=/Õ9,Õ9,Ó:,Ò:,Ñ9+Ñ9+Ô8+Ø8,ß;1à90Ü8.Ö8,Ï9*È9(Á:&¾9&¾6*Á80Æ=5É@8ÊA9È?7Å<4Â91¿4-Ç>6ÑH@ØOGÚQI×NFÌC;Ã7*Ë8&Ó8$Ý7)å4,í//ô+3ù)7ý(:ÿ+@ÿ*Aÿ)Aÿ&Aÿ#@ÿ!<ÿ=ÿ:ÿ:ÿ 9ÿ%;ÿ);÷-9ð19ê7:å;;Ü94Ü:5Þ;4à;5ã:5æ95é73ë54ë35ì46ë54è64ä71ß6/Ü5,Ù5)Û9*Û9*Ü9(Ü9&Þ9&ß8%á8%ã6%ç5+ê3+ë4,ë5*ë5*ê6(ä7)á7*ß;1Õ4,Ó54óTYÿesÿ[põGbÿIhÿAdÿ;`ý3[ù;aòMmÍLb‰8?S*$A5'69(;<,B@1DA0@?-=B.>G,=CKP'\b<eoJhvRezQd~Od‚LaƒFa†Aa‰@cŒ@g�DiŽHk�Ml‘Nl‘Kj�Gh�Hf‹FgŠJf‰If…JdƒHfƒKd�IdHd€Fe�Ge�GfEb~A]€>_„?b‡B69.:<1??7DA:KB;OE<VH=YL<^O<\O<\SB^WMc]]helom{sq‰so”qk™jf˜c`•\[”ZZ˜\] \_¦hjµno¿wxÊ€Ò†„ØŠˆÜ�‹à‘ŽÝ“�Ó“‘ÑÑД�Ñ“ŽÐ“ŒÏ’‹Î•ŒÍ“‹Ç”‹Ä”‹À“‹¾Œ„³ƒz¥€r•|ewu^fi[[d[Vd[RfXKrVH„TH¢YP½`YÉ\YÏYWÛcbákißokãvqûŠ„ÿŠ†ÿ…�ÿ}y÷nhîaXÄ4+À2&Ä8)Á:'À9&À9%Â9&Å:'Ì:+Ð:,Ö:-Ó5)Î5'Ð8*Ô>/Ô>/Î6(Ë/"Ú:.Ú8-Ù7,Ö8,Ð:+É:)Â9&¾7$½5)À6,Â8.Ä:0Ç=3Ê@6ËA7ÌB8Ç:1Ê@6ÏE;ÐF<ÒH>ÑG=Ê@6Å7+Ì6%Ö6&ß7,ç4/ñ03ö-5û+9ÿ*<ü';ü'=û%=ù#;û!:ú7ü7þ6ÿ8ÿ6ÿ$8ý*;÷2<ñ6=ë8;ä::Ü94Ü:5Þ;4à;5å95ç85é75ë54í36î47í55ê65å61á6/Ü5,Ø4(Ú8)Ù9)Ú9'Û8%Ý8%Þ7$à7$á6%á5'â4)ã5*ã6(ã6(á7(Þ6)Û7+Ý<2Ð1+Ø88ÿkrÿ]nþPiñ<[ÿHlÿ7]ÿ>eù5[î>bæVqª?Qa#(L/);7+07'37&9<)@=*A<(@>)BB&?FXa4s~T€Œdw‡`hSa}La€G`ƒC`…?a‰@cŒ@hŽEj�Jk�Ml‘Nn“Mk‘Hh�HgŒGh‹Kh‹Kh‡Lg†Ke„KdƒJc€Fd�Gf‚Hf‚Ge�Fb€Bc†DeŠEh�H25*58-;<4A>7G@:KB;RE<VJ<XL<YM=YQD^WOc^bigron€roŒrm•jg”`^�WW‹TUŽWX˜\_¢ac­ln»rsÃ{|΂ƒÕˆ†ÙŠˆÛŒˆÜ�Š×�ŽÒ��Î�ŒÏŽ‹Î�‰ÏŽˆÐŽ†Ï�…Î�…ÊŽ…Æ�…Á�‡¼�‡¶‰‚¬€x�oŒ€_p|Y`lXYd[T^_Q_\IjWF}SEžUL¾\YÐY[ÙX\äcgçkkáplåxsì}vízsþƒ~ÿ†�ÿ�{õkaÉ<2Ä8)Â;(¿:'¿;&¾:%Á:&Å:'È9)Î8*Ø:.Ô6*Î5'Î8)Ñ>.Ñ>.Ë8(Ç/!Ò6)Ó5)Ò4(Ò6)Ñ9+Í<+È=*Ä;)Ä:-Ä:/Æ90Æ9/Ç:1È;1É<3Ë=3ÓD<ÓE;ÐA9Ë=3Ë<4ÏA7Ï@8Î;1Î5'Ù5)â5.ë31ò/5÷+6ù)7ü):ù&9÷&;÷&;ø'<û&<ý%:ÿ$:ÿ#:ÿ"7þ#7ú%7ù,;÷5>ò9?è8:à87Ü94Ü:5Þ;4à;5å95ç85é75ë54î47î47î66ê65æ72á6/Ü5,×5*Ø8*Ø8(Ø8(Ù8&Ú7&Ü7$Ý6$Ý6$Û5%Ü6(Ý7'Ü9(Ü9(Û:(Ù9)×:+×:1Ð3.ëHKÿoyÿVjøD_ø<_ÿAgÿ3\ÿ8bø>cèMlÅOe€0=J C1-11'.6'4=,=B.B@+B<$E>$HE$U\0p~M�Ÿr—¨|ƒ˜mj„W_{J]~E_ƒCa†@bŠAf�Ck�Jl‘Lm’Om’Oo”Nl’IiŽIh�HiŒLiŒLj‰NiˆMf…LdƒJd�Ge‚HgƒHh„If‚Gd‚DgŠHiŽIl‘L-3'36+891<;6B=9G@:MD=QG=SI=SK>UNDZUQa^eigumk�lkŠjh�b`�XVˆPP†QQ�XYš`cªfj´orÁuxÉ|Ђ…Ö†‡Ù‡ˆØŠ‰Ù‹ŠÖ‹ŒÏ‹ŠÌŠ‰Í‰‡Ð‰…ш„ÑŠƒÑ‰‚ÐŒ„Ï‹„È�…Á�†º�†°‡�¥~w–m…„Yj€SZnVTc\RZaOZ_HeZDxUB’LD¸TRÐSWÛRYå^dèejãklæuqäunâoh÷~vÿ‚{ÿ‡~ôl`É?2Á8(À;(¿='¿='À<'Â;'Å:'Ç8(Í7)Ø:/Ø7-Ò8,Ï9+Ì:+É:*È9)É7(Ï9+Ñ7+Ò8,Ñ7+Ï7*Ì8*Ë<,Ë>-Ê<0É=0Ê<2É;/È:0Ç9-Å7-Å5*Ð@7Ð@5Í=4É9.Ë80Î;1Í:2Ì5,Ð/%Ù0)ä20ì25ô/8ù-9û+;ü,<ü-?û.?ü/@ü/@þ/Aÿ/?ÿ-?ÿ.@ÿ)9ü)8÷+7õ0:ñ6=í8=ã77Ü43Ü94Ü:5Þ;6â:7å95ç85ê65ì44ñ48ñ48ï56í76æ72á6/Ú6,×5*Ø8*Ö9(Ö9(Ø9&Ø9&Ù8&Ú7&Ù8&Õ8%Õ:&Ö;'Õ=(Ô=(Ô=*Ô=,Ô<.Ñ7/Ú;8ÿ^eÿaqÿPi÷<[ÿAiÿ4_ÿ4`û1YôKlÛYq•>NZ%-C),:2/-0)08+7@/<A+<<"A<UH(`[5{…S�£lª¾‹¨¾�Ž¥wo‹[aK^F`„DcˆBe�Dj“Go”Np•Po“So“So”Nl’IiŽIgŒGiŒLiŒLi‹OhŠNg†Ke„IdƒHe„Ig…Ih†Hg…Ge…DiŽIi‘Kl”N+1%.4*470893>:7B=:HA;KD<NH<NG=OJDWSR_^ffeugggf†ba‰[YŠQQ…NO‡RS“[^£ei³lo¾uwÊz|Ñ�Öƒ…؆‡Ù†‡×ˆ‡Õ‰‰Ó‡ˆË†‡È†…Ë…ƒÎ…€Ð…€Ò‡Ò‡€Ð�…ÒŒ…Ë�…ÁŒ…¸‹…«„�{v�|jz„Vc�PVnUQd^P[eMXbIb^EsXC�OE´WRÐUZÚSZç\cèagågjèpoìyræqh÷|tósjþxmæ\QÄ;+¾5#Â;'Á=(Á=(Ã<(Ä;(Ç:(Ë9*Ï7*Ö5+Ø7-Ô:.Î:,Ç:)Ä8'Æ:)Ê=,Ï=.Ö@2ÙA4Ö>1Î8*È4&Ê8)Ë<,É:,È:,É;/É;-Ê:/È9+È5+Ç5(Ê6,Ï;/Ò>4Ó?3Ö?6Õ>3Î7.É,#Ò+%Ü-*ç02ð39ö1;ú0<ý0?ý0?ù0@÷1@÷1@ö0=÷/<÷-9ù+8ú*7ü0<ù/;ô0:ñ4;í6;æ69ß55Ù42Ü94Ý:5ß:6â:7å95ç85ê65ì44ò59ò59ð67í76æ74ß6/Ú6-Ö6*Õ7+Ô8)Ô8)Ô9'Õ8'Õ8%×7'Ô9%Ð9$Î<%Ï=&Ï?'Î@(Î@*Ð?.Ò;0Ñ61ëHIÿfrÿOdÿHfû7[ÿAmý/]ÿ7aò4XãPj¼Ufm19E(*E697325406906=-3;#38CCja:…�TŸ®w©Á‡µÎ—¯Ç“’®}v”be†Q`‚F`…BcˆBg�Fm–Jr˜Or—Rq•Up”To”Nl‘Kh�Hf‹Fh‹KiŒLi‹Ni‹Og†Kf…Je„If…IhˆIi‰Hi‡Gf†Cj�Jk“Mn–P(0#+1'/4-350764<85@=8C>8IE<HE<KHCRPQ[[eabtbb|``‚_^ˆZXŠTTŠTT�\] gi³psÄvxÍ{Ø~‚Ûƒ„Þ„†Ý„…؃„Ö…„Ô†…у†Ë‚…ȃƒË‚�̓€Ñ…€Ôˆ€Õˆ€Ó�…Ò‹„È‹ƒ¾‹ƒ²‰‚¤�|“wp€veo~U[|PQnUNf_O]eMZdIaaGq]E‰TF­]TÈZ[ÔVZâ[aæ]déaeîllõzsírj÷xoÞYPå[PÒD8Ã5'Ç:)Å:'Ä<&Å='È=(Ê;*Í:*Ð8*Ô6*Ó0'Ô3)Ò8,Í;,Ç;*Â;(À;(Ã<)È;*ÕC4áK=ÞH:Ó=/Ê4&É5'Í;,È9)È9)È9+È9)Ê8+Ê8)Ë7+Ë7)Í6+Ò<.×?4Ø@3Ú@6ÛA5×:1Ò1)Ü0.å14î49õ6>ù4>ù1>ú0@ø/?ò.<ï/<í/9í07ï-5ï,2ò+2ñ*1õ2:ó4;ò5;ì59å57á55Þ65Ü75Ü86Ý97ß:6ã:7æ87é77ê67ì46ô5:ô5:ñ7:î87æ74à70Ú6-Ô6*Ó7*Ñ8*Ð8*Ð9(Ð9(Ð9&Ñ8(Ð9&Ë9"È:"È>$Ç?%È@(È@*Ì=-Ï;1×96üU\ÿ^qÿGcÿ?aÿ:aÿ7gþ3`û9^íFbÅNa‰@IS12A64@:<:46?56>7179+4="9CSZ.‚‚P¢§q±ÅŠ¯ËŽ°Ì’¦Ã�’®}{™egˆQ_�E^ƒ@c‰@i�Do•Js™Ps˜Sq•Uo“Sn“Nk�JgŒGeŠEgŠJh‹Ki‹Ni‹NfˆKd†Id†IeˆHgŠJh‹Ii‰Hf‰Ej’Ll–No™Q&.!(0%-2+13.331764:95=<7GD=DD<GFBOOQZZd_ap_az\\~`_‰\ZŒYX‘\\šgg­qsÀz|Ñ€Ü�„ႅ䅆ㄅჃÛ�‚Õƒ‚Òƒ‚΄‡Ìƒ†É„„̃‚Î…‚Ó‡‚ÖŠ‚׌ƒÔ�…ÐŒƒÄ‹‚·Š‚«†€œxˆsnuqbevWUtSLjVKf_MbfO_eIcbFp^FzQ?�[M·ZSÅSR×VZâW\êY`òce÷rmðme÷qhÑD;ÔA7Å1%È4&ÔC2Ç9%Ç:&É;'Í<)Ð:+Ó:,Ø8,Ù7,Ï,#Î0%Î6)Ë<,Ç@-Á>*¾;'¼7$À4#ÑB2ãOAäN@ØB4Ë7)Ë7)Ð>/Ê;+É:*Ç8(Æ7'È6'Ê6(Î8*Ð8+Ó;.Õ;/Õ9-Ò6*Õ7,Ú<1Ý<2Ü71è88ñ8=÷:Aû9Bú4Aõ/<ò,;ï,:í/;ê19ê38ë48í57ñ56ô36ó57ð37ï6;í9<é69â45Ý33ß76à;9Ý86Þ97á96ã:7æ87é77ê67ì46ô5:ô5:ñ7:ì89æ95ß82Ø7/Ó7+Ð8+Í:*Í9+Ì;*Í:*Í:(Í:*Ì;(Å9"Â:"Ã=$Â>%Ã?(Æ?+Ê<.Ð92ß<?ÿ\fÿQhÿIfÿ9]ÿ=hù.]ý<hé;\å\p¦LV^..G83BC><89?48E26C41=:)>E&FU*bs?�›a­½±ËŒ«É‹¦Ã‹�º„Ž¬z{™ef‡PZ|@^�?a‡>hŽCo•Jr˜Or—Rp”Tn’Rn“Nj�If‹Fd‰Df‰Ih‹Ki‹Ni‹NfˆKe‡Jd‡GeˆFgŠHh‹Gh‹GeŠDk“Lm—OpšR(0#'/$(-&).(.0-3317839:4==5?@8DE@JLKORYVXeZ\s]^}YX‚[Y‹^]–ee£nn¶wxÈ�Ø„…⌌ðŠŠì†‡ä‚‚Ü€€Ö‚€Óƒ‚΀ƒÈ�„ǃƒË„„ΆƒÐˆƒÓ‹„ÒŒ„ÏŠ€ÅŠ€¼‰€¯‡}Ÿ€x�wozkegf]Xm[Qm[Mi\Lf^Kd_IeaHhaGo_FoP;…UA¥]O¿aYÑYXÝQTîS[ø\`ùheæYPÔD<Î70Ñ7/×:1Ö=/Ô;+Ï9(Í:&Ï9(Ñ:)Ô8)Ø8*Ü5,Ü5,Ð/%Í3)Ì8,Ê>/Å@/Á@-¾=*¼9'Â9)¾0"éWJßK?Ã/#Ñ=1Î</Ç8(Ê;+É:)È9(Ç8'È7&É6&Ë5&Ì4&Ð7)×;.Ü>2Ü<0Ù7,Ú8-â>4ìC<õBEò9?ð2<ö4?ö2@ñ-;í-:í2=ì7>æ5;ä68å78ë99ð:9ñ78ï77ò;=î<<ê::æ87á85ß74Ü75Ü75Þ97á98ã99æ::é9;ì9<í9<ï8<ô7;ó6:ð8:í9:ç;9à;5×90Ï7,Ê8+È9)Ç9+Æ:)Ç:)Ç:(È;*Ç<)Á9#Ä='¾:#¶4¼9%ÉD3Ï?6Ê1,ÿ^eÿVgÿIbÿ<\ø1X÷0Yû6cõBiçYo¬FQo33N3,B?6>C<@78@-1L/3K2.A9&DJ(Zn=~š_ž¸w¨Å‚«É‹Ÿ¾‚˜¶€”²~‚ nf„RXvBY{?]€>c‡Ai�Dn”Im•Ll”Nk’Ol‘NiŽIfŠDc†Bc†Bf‰Ih‹KfŠLdˆJg‹MhŒNg‹Kf‹HeŠEf‹Ej�Ik“Jq›QržSt U*0$)/#).').(-/,130561782;<4>?7AD=HJGOPTSUaXYkZ\u\Z^]‰db”kj¤utº~·‡ß‹Œé��ð‰‰ë…„ ؀~Ò€Ï��Ë…†É†‡È‡ˆË‰ˆÎ‹ˆÏŒ†Î�†ÌŽ…ÈŠ€¼‰±†|¡‚x‘{s€tknlb`e]Re`Me^Kf^Kf^Kh]Ij^Hk_Gn^Ew_G[C’YF­ZLÇVPÞRSõQXÿX_ðUSãLEÙ>9Õ60Ù6/Ü90Û9,Ö9(Ó8&Ñ8&Ñ8(Ô8)×7)Ü6*ß4,Ý5,Î0'É5)È9+Æ=-ÄA/Á@-¿<*¾9(Ä8)Ë<.äREÜH<È4(Ë9,È:,Â4&Ê=,Ë<+Ê;*Ê;*Ë:)Ì9)Î8)Ï6(Ó7*Ö8,Ù7,×3)Õ1'Ø1(Þ7.ç<5øDGñ8>í/9ï/:ð0=î.;ë1<ë6?ä5:Ü36Ø22ß:8îEBøIFøDCó?>é:7ç98å97â96á85à85á98á98à87â88å99ç9:é9;ì9<í9<î7;ó6:ò59î68ê88ä;8Ü:5Ò8.Ë7+Æ8*Ã:*Â:*À;*À;*Á<)Á<+Á<)½9$Á?)»:%³5¹:'Á;/Ê:2Ø<=ÿ^gÿPdÿA[ù8Wö6[õ;`ô>dãIe«?Lƒ?>_:2G;/:>08>2F<:R>=N0.S8/VK5ciEx�[�°p�Â}¡Æ�š¹}š¹€š¶ƒ�©xrŽ^YuETp=\{B_@b†@i�Cm“Hl”Kk“Li�Mi�MgŒGe‰Cc†Bc†DgŠJiŒLhŒNg‹Mi�Oi�MiŽKh�HgŒFfŽEl’Im•LpœQo�Rt U-1#-1#,/&,.)//-11/34/45-9;0<>1@B7EHAKMLQRWVVbVWi\Yt^]fdŒpn zx·…ƒÎ�‹ß��ê�ŽïŽŒíŠ‰ç‡„ß�Ó|Ë}È€ņ…LjˆÆŠŠÊŒŠÉŒŠÉŒˆÅŒ„¿‹‚·‰€­†|¡€v�{p€ujpndbi_Vc]M_aL_aKc^Jg\Jl[In\Ho]Gq]EwaIxX?‡S=£VDÇXOãUSøOTþMSâ><Þ<7Ý84Þ71á6/á6.Ý7+Ú7(Ô7&Ó8$Ó8&Õ8'Ø6)Ý5*à3,Ý5,Ë3(Ä6(Â9)Á<+Â?-Â?-Á<+À8(Ã5'ÛI<ÚF:ÕA5Î</Ä6(Ä;+¾6&Æ:)Ç:)Ç:)É:*Ê8)Ë7)Ì6(Í3'Õ9-Õ7,Ö3*×3*Ú3+ß6/å:3é<8ð<=í49ë07ì18î3<í4<ë7@ê=CÝ7;áAAìNKøYUýZUúQLïB>å84â62à72à72à72á85â96ä;8ä;8á77ä88æ8:è8;ë8<ë8<ë8<ì8;ï58î47ì57ç77â:7Ù:4Ï8/Ç7,Â8+¿:+½;+»<+»<+»<)»<+»<)¸9&½@*³9$±6$¶;+µ0'Ã40ìPTÿ[iÿJ`õ:Uð7Vñ@]ïHfçHdÍNa‚89e?6R@2DA09=,57)>6+L:0E-#R=,g]B�ˆ_’«t—»{–Áz–¿{”³xœ¸…š´…€šm\uKHb5Nh8[xB_}?cƒ@hŒDl’Gj’Ih’Jg�Lf�Kd‰Dd‡Cc†DeˆFh‹KkŽNk�OjŽNi�Oi�Oi�Mg�Ig�Hh’Hl”Ko™Ms¡Vs£Wv¤Y24&04&01)01+12-23.34.46+9;-;>-?A3DF9IJDMONRQWSS]XVd\Zoda~nl”zw¬„‚ÃŒ‰Ö‘�㔑ê“�뎋≅قÌ|Ã|¿€~¿„‚Á†…Á‰ˆÂ�ŠÃ‹ˆ¿‰…¸‡®…}¤…{�€vŽyn~rgmmaak^Vi[Pc\J]`K]aJc^Jh[JmZKqZJtZIv[FsWAxR;ŠQ=ªYFÍ[PãTPðFIñ>AÛ2/Ü5/á51ã60ã5.á5+à6)Û8)Ö7$Ó8$Ñ8&Ô9'Ø6)Ü6*à3,Ú6-È6)¿7'¼7&½:(¿>+Â?-Ã;+Ä8)Æ4'éUIÓ<1Ì8,ÑB4¾5%À=+À=+À8(Ä8'Å9(Ç:)È9)Ë7)Ì6(Í3'Ò6*Ô3)Õ2)Û4,â92é>7ì?9ì=:è45ì59ï6;î5;ê2:å28ä7=â<@óUVübbÿmjÿidóVQäA<ß63Ý4/ß61Þ71Þ71ß82á85â96å97æ:8å78æ89è8;é9<ë8<ë8<ë8<ì7<î5:î68ê67å97ß<7Ö<4Ì9/Å9,¾9*º;*¶;)µ<)µ<+´=)³;*³<(³:'·@,­8&­8'³;-«+"Ã54ý`gÿTdúG\ð<Uë=XçG_áOdÒO_´SZvGAXH9KD2FE1BE0=?*:8#;3?3PG*nkHŠ“f•¬t‘´tŽ·s‘ºxž»…¡¸Š�§}l‚[K_:AW0Lc7Wr?_{@dƒ@iŠCk‘Fi‘Hg‘Gd�HeŽJb‡Bc†Bd‡EgŠHj�Ml�Om‘Qn’Rj‘Pj‘Nj’Lh’Jg‘Gi“Gm—KpœOv¥Wv¦Zx¨\78(68*66,56.56056067/68+;>-=@-@C.DG4IK>MNFQPNRRRVUSYX]a_lki�us›€}²ˆ†ÇŒŠÓ•’ᔑâ��Ü‹‡Ó„�Ä�}º|µ|³�~·„�º‡„¹‰…·Š„²†¨�z›€x�s‡znzrfjm_^j[Ti[PjZKfZJb_Lb_Le^Li\Kn[Lr[Kw\K{ZG~UA†R=›RA¶VFÎRHÚG@â88å33Ý1-ß3/â5/ã5.á3,ß3)Ý5(Ü9(Ó8$Ð9$Ï:&Ñ:'Õ8)Ù7*Ý5,Ø7-Æ:+»9)µ6%·8'½;+Â=.Ç;.Ç7,Ð90õ\TÖ=5É5+ÏA5¸3$·:(¼?+¿:)Ã:*Æ:+È:,Ê;-Ì:-Ï8-Ñ7-Ð3*Ò1)Ö1+Þ71å<7ê>:è;7æ74é75ï;<ð<?ê69ä15ä5:èBFíMOÿxwÿqnñ`[ÝJCÑ83Ñ2.Ü73ç?<à93à93à72à72â62ã73å76å76æ68ç79ê7;ë8<ë8>ë8>è7=ê7;î7;î79ê7:ä::ß=:Ö=7Ë;2Ã;/º8*µ:*³;*¯<)¯<*¬<(¬;)¬;)ª;(¬=*¦8'§9*ª8-¬/)Ë@CÿhrôM^óH[íDYåH[ÙO^ÉT]·TWœXUdM?OL9LG4JE1FD-?B'=D#>H#KV.\g<x�T‹™f‘¤m�©n�²r—¹}Ÿ¶ˆ’§€u†dRcC?P0BS1Oa9Vl=b{Be‚BiŠCj�Ei’Fe‘Fb�HcŽGb‡Bf†CgŠHiŒJkŽNm�Pn’Ro“Sj‘Nk’Oj“Mi“Ki“Ij–InšKržOv¥Wv§Xw¨Y?=.=>.==1==3==5;<4;=2;=/?B/@D-CG.FJ3KL:NOAQQIRRHYWJZZN_^\gerqn‰{y¡ƒ€·‡…Ä�ŒÑ�ŒÓ�ŠÏŠ…ǃº~y¯zw¦yv¥{v¬|y®€|®ƒ}©ƒ}£x—|t‹zr�vksrffm_\j\Sk[NlZLl[Kk[Ki]Mg^Mi]Mj]Mn]Ms^My\L�ZI“`O›VG§N@¸J=Å@7Ï6.Ø1+Þ1+á51á6/â5.â4-à2)Þ4'Ü6&Ú;(Ñ9$Í;$Ì;&Î;'Ñ:)Ö9*Ù7,Ó9-Â=,µ:(°5#²5#»9)Ä<.É;/Ë7-Ù?7ø[TãIAÎ:0Ì@3¸6&¬3 µ<)¼9'¿7'À7'Ã7(Å7)É7*Î7,Ñ7-Ù<3Û:2Þ93å<7ê>:ê=9ç85ã41è96ë<9ì::è8:ì>@øPPÿbcÿppùheâTPÊ=6À1+É40Ö;7ß=:à;7â;5á:4â94á83ã73ä73å55æ66é69é69ê7;ë8<ë8>ê7=è7=é6:ï8<ì89è8:ã;:Ý>:Ó>7É<2À</¶8)²:)®;)ª;(©<(¨;'¥:&¥:(¥<) 9( ;) 9*¢6*µ>8ÛTXûepîK\ïI]éJ\ÛM[ÉQZ´VV ZRŠ`RPI7HM9NI5H@+;567>O%Lf6lŠTw”\„›e‰šd�›f�£k”­t˜°|ƒ•mn}\Q`C<J09F,AP1O_;Wj=czDe‚Bj‹Dk‘Fi’Fe‘Fa�FcŽGcˆCi‰Fj�Kl�Ml�Ol�Om‘Qn’Rm”Qm•Om–Pl–Lk—Jm™JrŸNs¢Q{¬]{¬]z«\DA0DB3DB6CC9CC;CC;BB6BC3DF0EH-GJ-JM2ON9RQ?TREUSD\ZC\\D\]Oaaaihxrq�{z¦�~µ†‚¿ˆƒÃˆ„Á„�º{®ys¡to—om•plžrn¡vpžwršwr’to…skzrinmb`l_Wj\Qk[Lm\Lo\Mo\Kq[Mq[No\Nm]Nl_Om`Os`Oz_N†ZMŸ_S¤PE®B8º;2Æ6-Î1(Ö2)Ý5,â70á6/á4-à3,à4*Ý5(Ø7%Ô9%Ï:"Ê="É=$Ê>'Í<)Ò;*Õ9,Ï;-¿@-°;'¬5!®3!º8(Ä<.Ê:/Î5-Ø93ðNIñTMÙB9ÌB5¾?.¦1³<(¾<,Á9+Á7*Â6)Å7+É9.Ï;1Ô;3áD=ãA<ä?;ç>;é=;é;:è88ç85ç85æ95æ74é;:ôJJÿ\[ÿdcøc_Ì=9Á82½4.Ã81ÕA=ÞE@ß=;Ù42â;5â;3â94â94ä84å84ç77é77ê69ë7:ë6;ë8<ë8>è7=è7=é6:ë7:é69æ89à;9Û>9Ñ>6Æ<1¾<.³8)®;)ª;(¦;'¤;(£:'¡:'Ÿ:&¡>+•6$œ=+š9)š2'ÃPKíkmî\fìP^éJ\ßHWÍKS¹RS¤[R’cQ€jUJM8DL7JB/H9$B7FH#Up=g“V~³o‚³qƒ¥h}’Yƒ�YŽ–c‰—d|Œ_XfCGT89E-6B*:D,=J0JX7Wh>e|Fh‚Cl‹El�Fh’Fc‘Fa‘Ga�Gf‹FlŒIm�Nn‘Ol�OkŽNk�Om‘Qp—Tq™Sp™SpšPo›Nq�Nt£Rw¦U|­^z­^y¬]IC3JD6IE9IE:IG;IG;HF9HG5IH3JJ2LL2NN4RP;TQ>XRDYTA]Y>[Z>\YH^[Tdahnk~xs“}x �{§ƒ}«…­ƒ|¦}w�un�oi…kg‚ieˆjf‹mhˆmi‚mh|lfrkbgj`^g]Th\Nh[Kk\In]Kq_Kq_Kp]Lp]Np]Op]Ns]Ow]N~]N‡\L’XL¡UH¤F<«9/º7-Ê7/Ó6-Ú6-Þ6-ß6/ß4-ß4,ß4,ß7,Ý7)Ø6'Ñ6$Ï:$Ì<$Ê<&Ë='Î;)Ñ;*Ó:,Î<-¾A/¯<'¨3!¬3 ¶9'Á=.É;/Ì8.Ì2*àA;ø[TàI@ÐD7ÊF7«2¹>,ÊF7ÌB5Ê@3Ë?2Í?3ÔA7ÜE<ãF?â@;â=9ã:7â64ã54ä65å97å<9ä?9à=6Ü:5ß@:êMHêQLÓ@9¸)#¸-&¾5-É>7Ñ@;Ö>9Ú;8ã;;ç==â96â96â96ä86å95ç:6é99ê::ê88ê88ë8;ë8;é9<è8;ç7:ç7:é6:æ68â88Þ:8Ö=8Ì=5Â<1¹;-®9(ª;*¦;)¢;( ;'ž;(ž;(š;'˜?-Œ5"™=.˜8*—/&Í\X÷y|ßX_æXdÚQ[ÊKR¹LO©VP�_T�gU€nVPW8?J*<;C=UQ+lvDƒ¢f”Á~‡¼v†¸s|žan‚Mu~Q‚‡^u}V[b@;D)4<%5=(=D2<E29C+CP2Ue>g{HiƒFl‹Gm‘Gj“Gf’Gb�GcŽGh�Jl�Mo’Pn“Pl�MiŽKj�Jj’Ls›Ts�Us�SržQržOs¢Rw¦Uz«Zv¦Zu¥[s£YRK;RK;RJ=RJ=RJ=RJ=RJ=RK;UN<VO=WP>XQ?YR@[TD\UE\UEb[Hb[IaYLbYRe][kaiqftuj{zn„|pˆs‹s‰|p„uj{mbre]jc`k`_g_\c_Z^^ZY^YUaZRbZOe[Og[Ki\Kk\Ik\Gm]Fm]Fj^HicMqfTqZJuOB‹WJœ\P£YL¬ZL¥N=¦E4§<,±7(Â8-Ñ80Ü41Ý1-ß6/Þ7.Þ7.Þ7.Þ7.Û7+Ø6)×7)Õ8)Ô8)Ò9)Ò9)Ò9+Ò9+Ô8+Í;,½;+¶>-®9(©1 «2!·9*À</Ã9.È5+Î5-Ô:0ÛC8ÝK>ÔH9Â=,·2#ÝPFØJ@ÓE;ÔD;ÚG?ÞG@àC>à<:à74æ87ì::ì::ç77â64à93Ü=7ÙF>ÏB9Ä:/¼4(·1&¹5)¿;/Ã?3Ã:0Ê=4Ò?7Ú?:â=;ç;9ì89í9:ç7:å8:å8:å99å97å97å95å95à40á51ä65å76æ89ç9:ç9:æ8:ë;>æ9;ß99Ü=:ÖA=ÍB;ÁA6·?1©:)£<+ =*›<(™:&–9'•8&�9&�A0‹<-‡/#‘/&¼LJämoçloÍVZ½LNµNOªQM¢UO›XO—]R•aTŠjSWZ/:M7FTa)~ŽP›°oŸ¼yœ½x’µs~�ah€L\mCYdBT]BHP9?D0:=,@C2<@1:>0<C3:C0AN2Rd>]sBgƒHp�Ko“Kl’Gj’Ii“Kg�Jl�Pm‘So“Sl“Pm’Mk“Jm–Jm˜IqœLqžMrŸNr¡Pt£Rv§UyªYzª^y¦cm™ZbŽOWN?WN?WN?WN?WN?WN?WN?WN?XO@YPAYPA[RC\SD]TE^UF_VGd[JcZIcYMcYOf[Uj_]nbdpdhreltgpvirvirsfmm`gg[_bXY^[VZZRYVOWTKXTIXTH\VH^WGcZKcZIf[Ig\Hi]Gj^Fk_Gh`IcbMngTv\M�SFšVM«VO°RH·QE¯H9¬C0¬=,±9)¾8,Ë80Õ62Ù40Ý6.Ü8.Ü8.Ü8.Û7-Ú8-Ù7,×7+×7+Õ7+Õ7+Õ7+Ô8+Ô8+Ô8+Î:,¾<.¶>.®9(¦1 §2 ²:)¹>.¾<.Æ:-Ç4*Ë4)Ò;0ÜE:áOBßQCÝOCÛF?Ø@;Ó<5Ò;4Ö=7Ú=8Ü86Ý33ç77ë78î79ë78ç77å97â=9ÛB:ÊA7¿?4¸:,±6'¯5&²8)¸=-¿@1ÅA4Í@6×@7ß=8ã:7é77ï56ð69é6:ç6<ç7:æ89æ87å95å95ã:5â92ã:3ã:5ä;6ä;8ä;8å99å99ä:;ã;;ß<=Ù?=ÐA;Ä?6·=2¬;-¥;+ =*œ=+—<)–;(’:&‘8&Œ9'€9'€8)‰5*�<5¿QP×eeÑ^a·MM¬NL¤TM�YP˜\Q—^S—^S˜_VŒfQ`_/Sg*^r3z�O™¯n¥¾|™²pˆ bj‚HauBTe;JX7FP7?H59?158-:</AB4=?4:=2=A3:C0@M3Rd>^tEgƒHq�Mp”Ln’Hk”Hk“Lh‘Kk’Qn’Tm”Sm”Qk“Lk”Hl—Hm˜Hp�Jp�Jp Ls¢Qu¦Tx©Xx©Xx¨^m—Xb‹QY‚H]TE]TE]TE]TE]TE]TE]TE]TE[RC\SD]TE^UF_VG`WHaXIaXGe]Je]Hd[Jd[Le[Qf[Uh]Yh][i]]j^^k__k__i^\f[YbWS_VQZVMWTKUQHRNCQMBSOCWQCXRB^WG^WEaYFc[Fg\Hi^Hk`JhbLbaMngUy_PˆXN¢\T±VQ²IE¸E@¹@7¸>1·;/¸8+¼8+Ã9,É;/Ñ:/Ù8.Û7-Û7-Û9.Ú8-Ù7,Ù7,×7+×7+Õ7+Õ7+Ô8+Ô8+Ô8+Ò9+Í;,À>0¶>.¬9'£2 £3ª:&³>,¹>.Ä?0Ä8+Å2(È4*Ð<0ØF9ÜM?âNDÙ:6Ø43Ô20Ö42Û97á=<æ<<é;<í9<î5:ë27ê37è58å99á<:Ù@:¾8-µ:+±6'®4%­5%°8(µ<+»=.ÈD5ÏC6ÖB8Ü=7â94è64í55ï79ê69é69é69ç77ç77æ95å95å:3ã:3ä;4ä;6ã<6ã;8â:9á99Þ88Ý9:Ü<<Ú@>ÓB=È?9¹:1­7+¤6'¡;,�<+™<+”;)“:(�9&�8%Š9&z6#y6%ˆ9, F=¹SOÀWT¸PO«IF¢MHœSL™YP—^S—aW—aW˜_V‹ePsrFz�V“¦n¤¹€«À‡ž³{|�[[o<?R$AS+AP/?L2:C06<.69058/9;.@A3=?49<1=A39B/@M3Qc=`vGi…Js’Or–Np”Jm–Jm•Nk”Nn•To–Uo–Sn–Pl”Kl•Il—Hn™Ip�Lp�LpŸNu¤Sx©Xyª[u¦Wq X_‰JVEOx>aXIaXIaXIaXIaXIaXIaXIaXI_VG`WH`WHaXIbYJcZKd[Ld[Je]Hf^Gf^Ig_Lg^Of\Pf\Rf\Sh]Wh]Wh]Wh^Ug]Te[QdZPc[N]YMZVJVRFRNBPL@PM>TN@UO?WQAYTA[VC]XDaZGd]Jf_LfaMdcQleUv\OŠ[Q©b\¸[V¶FD¹<:Ã<8Æ;6Ä:0Â8-½8)¼:*¾?.Ç?/Ó9-Ù7,Ú8-Ù9-Ú8-Ø8,Ø8,Õ7+Õ7+Ô8+Ô8+Ô8+Ô8+Ô8+Ò9+Í;,Á?1·?/­:(¢3 �3¢8"¬=)µ@.¼?-¿:+Ã7*Å5*Ë7+Î:.Î</Ó:2Þ65à24Û12Ý34à88å;<ê=?ì<>î7<ì38é06æ25å58á77Ú65Î61·5(¯7'®6&¬7&¬7&®9(²:)µ:*ÃA1ÊB4ÔA7Ü?8â;5æ95ë76î87í68ë78ë78ê86ê86è94ç:4æ;4â92á:2ß:4à;7Þ:8Ü:8Û97Ù99Ö<:Õ@<ÒC=ÊA;»;2¯6+¤6'ž7(œ;*˜;)”;)’;(�:&�8$Œ7#‡9%€=*v5#�5'™E:ªPH¬MG©LG¦QL�PJ˜UL”ZO”^R•aV•aV—aW�gT‰…_™©x­½Ž¦¸ˆ�¡ts†YN`8/A->3B%:F.=F3;A3:=2:=4<=599-?@2<>39<1<@29B/?L2Pb<awHj†Kt“Ps—Or–Lo˜Lo—Pm—Op—Tq˜Up˜Ro—Pm–Jm–Hn™Jp›KqžMrŸNt£Sy¨X{¬]w§[m�Se’M[…FS|BNw=cZKcZKcZKcZKcZKcZKcZKcZKcZKcZKd[Ld[Le\Mf]Nf]Nf]Lf^Gg_Hh`IiaLiaNi`Oh_Pg^Oj`Ti`Qi`Qh_Nh_Nh`Mh`MhaNd^N`ZL[UGVPBPM>NK:NK:NK:PM<RO>TQ@VS@XWC[ZF]\H^]Kb`Qf^QmWJ†[R¬ic¾c`ÁMMÈBCÐ;=Ô89Ó84Ì70Á9+º=)µC+¼B+Î;+×7)Ø8*×:+Ù9+Ö9*Ö9*Ô8)Ô8)Ô8)Ô8)Ò9)Ò9)Ò9)Ò9)Í;,Á?1¶>0«:(¡6"š4›7 £=&¯B-³;*º;,Â:,Ç9-Ë8.Î:0Ð<0Ù:4é9<ì4<æ39ã28â38á48â38â25é6<é49ç4:ç7:ã9<Û76Ð21À0(´6(­:(­:(«:(«:(«:(¬9&¯7&¸:+Á;/Î>3Ø?7à=8ä;8è96ì87î66î66î66í74ê84è:3ç:3ä<3á:2ß;2ß<5Ü=7Û=:Ù><×=;Ô>=Î@<ÉB<ÅB:¹=5­7-£5(ž7(š:*–;)“<)‘;*�:(Š9&‰8%ˆ7$ƒ8%ƒ@-u4"{3%“G:¤QI¡NFžNG¡WN˜UL’XL�ZN�\N�]R”`U—aW�hW�‰f˜¥z�©�€Žj]kHDS42@&$2-:&0;*5=.9?3<?4=?4:;3::099->?1;=28;0;?18A.>K1Pb<bxIk‡Lu”Qt˜Ps—MqšNq™Ro™QršSršSršQp™Mm—Km˜Ip›LsžOt Qv¢Sx§Y{©^z¨_qŸVc‘IY†C]†JVGQzBe\Me\Me\Me\Me\Me\Me\Me\Me\Mf]Nf]Nf]Ng^Og^Oh_Ph_Ng_Jh`IiaLjbMjbOjbOjaPjaPjaPiaNh`Mh`Kh`Kh`IiaJjbKf_Mb]J^XHXRBSM=MJ9KH7IH6KJ8LK9LM;NO=PQ?QTASVESVE[YL^VIcQEzXN¡ha¶eaÄVUÖQRÙ@Cß:>Ü87Ô70Ç;.¼?+±C*µA(Ê;*Ô7(Õ8)Õ9*Ö9*Õ9*Õ9*Ô8)Ô8)Ò9)Ò9)Ò9)Ò9)Ò9)Ò9)Í;,Á?1µ=/«<)¢9$™6•5›;"§A*²B.¹@/¿;,Ã7*Å2(Ë4+Ò;0ß=8ê5:î3<ë6=é8>ç:>ä;>ã:=â9<ß58á7:â9>á=>Û=>Ñ96Ã2/¶0'¯9+«>*«>*©>*ª=)©<(©:'«8%¯6%¹7)Å;0Ð=3Ù<5ß<7ä;8é:7î87ï75ï75î85ë:4é;4ç<4ä=4â>5à?7ÞA:ÚA;ÙA>ÕA?Ñ@=ÍB?Á@:»@8³=3ª8-¡5(›5'˜8(•<*‘<(�<)Š<(ˆ:&†9'„7%ƒ6$6%}:)t3!~9*—OA£YNœRG•OE˜VJ•YN�[MŽZMŒ[MŽ\Q“_T™`W‘gW~xXyƒ^r|ZU`B;E,/;%0;*0<.3=24<16<2<?6@A9?A6<<077+89+>?1:<17:/:>07@-=J0Oa;bxIk‡Lu”Qu™Qt˜NqšNršSpšRs›RsœPr›Op™Kn™Jp›LsžOv¡Ry¥X{§Zz¨]x¦]qŸWg”O]ŠEVƒB_ˆNX�IR{Cg^Og^Og^Og^Og^Og^Og^Og^Og^Og^Og^Oh_Ph_Ph_Ph_Pi`OjaPjbOjbOjbOi`Oi`OiaNiaNiaNh`Kh`Kh`Ih`IhaGhaGhaGf_Le^Kb[I\WDWR?PM:LI8IH6IH6GH6GJ9GK:HL;IM<HO?JN?PQCXRF[OCiRDƒYM–VL°SLÔZYÛIJâ?Bá99Ú72Ï;1Â?-´B*³<$É:)Ó7(Ô8)Ó:*Õ9*Ó:*Ó:*Ó:*Ò9)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Í;,Á=0µ=/¬=,£<)˜8 �4•9 ¡A)­C-¶A/¿=/À6+Â/%Ç0'Ð6.Þ:8è59î6>ì;Aê@CæAEàBCÜ@AÙ??Í12Ð66Ó:<Ñ=;É;9¾71µ4.­4)ª<-¥>+¥>+¥@,¥>+¤>(¦;'©:'ª7%²7'»7*Å8.Ì70Ô:2Þ<7æ?9ë97ì95ì95ë:4é;4è;4å=4â>4àB7ÝC9ÚC<ÕB;ÏA=Ê?<Ç=;¿>9±<3©:/¢8+œ6(˜7'”8)‘:)�:(‹=)‰<*‡<)ƒ:'‚9(€7&6%}6$x3#x5%…B2˜UE�WK”PC�PD•ZL“YM�ZN�\Q�\Q’[T•\U˜[VŽbUkbEV`=JS4=G,6?*2=,4>35?66=68?8=B;CF=EG<CD6@>/;9*78*=>0:<17:/:>06?,=J0N`:awHk‡Lu”Qu™Qt˜Nr›OršSpšPq›Or�NqœMp›LošKr�Nu¡Ty¥X«`~©az¥]qœUf‘L_‰G]‡E^ˆIb‰PY€ISzCh_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Pi`Qi`QmdUlcTjaRi`Qh_Pg^Og^Mh`MiaNiaLiaLiaLiaJh`IhaGh`IgaKf_Le^KaZG\WDUR?PM<ML:EF4DG4DH7DH7BI9BK:BK:CJ:IK>SQDSPAXL<eK:tE5•H>ÂYSÙROàDEà::Ü75Õ;3Ê>/º=)µ8"É:)Ï8'Ð9(Ñ:)Ò;*Ò;*Ò;*Ò;*Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Í;,Á;0¶:.­<,§>+™9#�3“7Ÿ?'¢9#¯<)À>0Æ<1Ì8.Ð7/Ù<5ã?=ì=Bî?DéCEâDCØ@?Ê;7À50»1.¹/,¼51¾:6»<6³:2ª8.¤8,¡:+¡=-Ÿ>+ ?,¡@-¡@-¡A+£>*¦=*©;*­:(²7(¹5)¿5+É90Ö?8àC<å<7è;5è;5ç<4å<5ä=4á>5Þ@5ÜD9ØD:ÒC;ÊA9Ä?:¼;6¸85°93¢7-™8(•5'“6'‘:)�<*‹:)…8&ˆ=*…<+ƒ<*€<)~:'|8%{6&{6&y1#€8*ŒG8–SC’PB‹L=�RE™_S‘WL’YN”[R–]T–]V—\V–YTŒ^QaX;EM(6@8A&>G2<E43=2-7.1817>6BE<GJ?HI;BC1?>*;:&78(=>09;069.:>06?,<I/N`:awHj†Ku”Qt˜Pt˜Nr›OršSpšPqœMqœLqœLp›Kp›LsžOw£V{§\‚­e€«dx£\j•P]‡EYƒA^ˆHe�PcŠQZ�JSzCi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi_Sj`Wj`Wj`Wj`Wj`Vj`Vj`TjaRjaRjaPjaPjbOjbOjbOjbMjbMg_Jg_Jf^Ie_Ib]IZXCSP=ML:HI7EH7AE6@D5?E7?G8?G8>F7GK=IK=AE4KJ6\Q;aE/ƒJ9Àj]Üg`ãUSäFEåA@àA=Õ>5É=0Æ=-Æ4%Ë5&Ì6'Í7(Í7(Î8)Ñ;,Ó=.Ò<-Ò<-Ò<-Ò<-Ñ;,Ñ;,Ñ;,Î<-»2(¶7.±;/¨:)š7"”5•7œ9"¬A-±<+º6)Á4*Í6/×<7âC?éGDèDEèEHÝCCÉ:6»61´80®8.©4+«9/©9.¤8. 8-›9,–:+“<+’<+–=+˜=*™>+š?,›@-�A, A-£@-¢;*¥:(«7(°6)¶6+¾8-Ä;1Ì<3Ú=6à=6â?8âA7âA9ÞB6Ú@6×@5ØH=ÒE;ÇA8»;2²5/ª3-¦1*Ÿ2+™9-’;*‘9+�:*‹:)ˆ;)‡;+…<+€9'~9)|9(|9(|9({8'y6&x3$w- ‹A4—OC“MAŽL@‘QE“UJŽRG˜\R˜\R‘UM�TL—ZU•XS”WT“gZ]T5IQ*AK(BK,>H/7@+2=-4?14<1HPCZ^PW[JMO:DF.AA'==%:;)>?1;=2:=2;?14=*;H.Pb<cyJlˆMu”Qs—Or–Lp™MršSq›QqžMqžKrŸNrŸNsŸPv¢U{§\ªb‚­fx£^j“O`‰G^‡Ea‰JfŽPh�Td‰S[€LTyEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi_Sj_Yj_[j_[j_Yj_Yj`Wj`Vj`Tj`TjaRjaRjaRjaRjaRjaPjaPh`Mh`Kg_Jg_Jd]J^YEUR?ON:IJ:EH7BD6@D5?E7>F7>F7=E6@F8>E5>F1HL5MI0U?'Q:¼sbËcXà_ZëUTçGGâ@>àA>Ø?9Î:0Í9+Í<+Ï=.Ï>-Ï;-Î;+Í9+Í:*Î:,Ð:+Ð:,Ð:+Ð:,Ð:+Ð:,Í;.É?5ÅB:¼B7®=/ 7$–1–1›2¦5#²7(Ã;/ÑA8ÞE?æIDéJFêKHÞCAÛEDÐB>¼<3¯9/©=0¥?1Ÿ=0Ÿ?1œ>2—>0”>/�>-Š>.‡@.ˆ?,Ž=*‘;*’<+“=,”?+–?,—>,™>,Ÿ@.¡=-¤:,¨8,¬8+±8-¶:0¼9/Ê<2Ð<2Ó?5ÕA7×C9ÖC9ÔD9ÑE8ÊB6ÃA4¹=3°:0©6/¢5.Ÿ4,™7,‘;.Ž=,Š;,ˆ;+†:*„;*ƒ<*�<,~;*|;){9)y:)x9(w8'u6'w4$€4'�D7˜NC‘K?ŒJ>�QD’TI�RG“SJ˜XO–UO—VPœ]X”WRŽSOŠcT]W5Xc9_jBZfBIT66C)4@*<H4OXGYbQ_fTW\FJN7BE*??#<<"9:(=>0:<19<1<@25>+;H.Pb<dzKlˆMu”Qt˜Pr–LqšNršSq›Qs OrŸLp�LqžMt Qx¤W|§_ªc{¦aošUcŒJ]†D`ˆIe�Ne�Nc‹Od‰S[€LTyEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi_Uj_[j^^j^^j_]j_Yj`Wj`Vj`Tj`Tj`Tj`Tj`Tj`Tj`Tj`Vj`TlaOk`LiaNiaNf_La\HYVCSR>MK<HI9DF8AE6AE6?F6>D6<D5=F57@-@H1IK3F?%W?'‡ZC·s`¶VJÙbZðebìTSæFHëGHæCDØ:7Ë7-É:*É:,É:*Ë9,Ê8)È6)È6'Ê8+Ë7)Ë7+Ë7)Ë7+Ì8*Ì8,Ë8.Â3-¾5/¶6-­3(¡0"ž/ 1 §3$¼A2ÇC7ÖG?áLFçLGéJFåFBßD@×CAÏDAÁ@:±;1¤</¡A1œE4˜E3“D3‘D4�B2ŠA0…A.‚A/~A.@.†:*‹9+‹:)Š;*‹<+�<+Ž=*�<*–?.–=-›;+�9* 8+¥9-¨:-®9/¸:.¾90Á;0Ä>3Å?3ÅA4ÃA3ÁA4¶;,²:,ª8-£7+ž6+™7,˜7.“;/‹</ˆ<,‡;-„;,‚:+€;+€;+<,};+z;*z;,w9*v8)u7(t6)v4&ˆ;1”F<–NBŽJ=ˆH<�OB‘SH’RI’OI™TO™SQ˜WSš]Z“YU�[T�l[�}X~‰^€‹alxRLX46D#=J.LZ@eqYeoW]eMPU>EI0?C(=@#:<$89';<.8:/9<1=A36?,<I/Pb<dzKm‰Nv•Rt˜Ps—MqšNršSpšPu¢Qp Ln�LpŸOu¤Vy§\}¨`|§`p™Uf�M\„E\„EeŒMi�Qf�N`†IcˆRZLSxEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QjaRjaRjaRjaRjaRjaRjaRj`Vk`\k__k`^k`\k`ZkaXkaWkaUkaUkaUkaUkaUkaWkaWkaXkaWnbRmbPjbOjbOhaNe^K_ZGYVERO>MK<GH:DF8CE7BD6@D5<C3<H49B-FJ3LF0O:%qL9šcOª^N¡A5ÆSLçc_ð^^íSUëJOéDJáAAÌ91Ç9-Å7+Å7+È8-Ê:/Ë;0Í:0Ñ>4Ñ>4Ò>4Ò>4Ó?5Ó?5Ó?5Ó>7Ó?;ÑB>ËD>ÆF=ÂF<ÃI>ÇM@ÎPDÙSJÝPGâMGäIEäGBâC?àA=ÛB=ÕFBÉD?º?7«=0Ÿ>.šC0•G3”I6�F4ŒE3ˆC3…B1€A/}@.zA.|?-‚:,†8,…9+†:,†:,‡;-‡;+ˆ;+‰:+‹:)Ž8)�8*•9,™;/�=/¢<0ª:/­8.¯9/°:.°:.°:,¯9+®8*¨7)¥7(ž6)›6*•7+’8-�9/Ž<0ˆ<.…<-„;,�9*:*~9*}:*};+|:,z;,y;,w9*s8*q6(r7)u5)ŒD8’H=‘K?ŠH:‡I<‹OD“TK—TL–QL›TP—QO”SO–\X�_X�h^—~j””p…’fu‚WYh?CR+?O*P`<crQetU\iKNX=CL1?E+>B)=A(;='78(:;-79.9<1=A37@-<I/Oa;bxIlˆMv•Ru™Qt˜Nr›OršSo›Pq OožMožMs¢Rz¨]|ªay¤]tŸZcŒJ^‡E[ƒD_‡HgŽOj‘Pf�L`‡Hb‡QY~KRwDi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QjaRjaRjaRjaRjaRjaRjaRj`Tk`Zk`\k`ZkaXkaWkaUkbSkbSkbSkbSkbSkaUkaWkaWkaXkaWocUnbRkbQkbQkbQhaOd]M`ZJXRBRO>NK<IG8GE6DE5BC3>B1=I1?G/LG1S>+g=-•ZL­_Sœ>4ž3+¶?9ØXUóigödeëRUæJNèNPÚIDÓF<Í@6Ë>4Î@6ÒD:ÕG=×G>ÚJAÚJAÝJBÝJBÞKCÞKCßJCàKEëSRèTTãSRÜROØSLØSLÛTNàUNàLHáIFãEBâC@äB@âC@ãDAàFDÙJFÊE@¹>7¬=2¡?2˜B1’E3�G4ŒE3ˆC3‡B3ƒ@0�?/~?0z?/|>/€:.ƒ9.ƒ9.ƒ9.ƒ;-ƒ;-ƒ;-ƒ;-�9*ƒ:+…9+‡:*‰:+�;-�<.“;/š:.ž9-Ÿ9- :,¡9,¢8+¢8*¢8*Ÿ8)œ8)˜8(”8)�9*Œ:,‰:-‡;-„:-‚:,�9+€8*}8)}7+|8+{9+z8,y9-x:-v8+q5*o5)q7+t9+ŒH;�I<‰I=„I;…K?ŒRF“WM˜XO˜QMœUQ—SP“VQ’`YˆbWƒg[†zdnrQZi@JY0AR(IZ0Wi?dvNj|VXiEL[:?L.:D)<D,>C-<A-:=,89+:;-68-8;0>B47@-;H.L^8_uFi…Jt“Pt˜Pt˜Nr›Os›TpœQn�MmžLp¡Pw¨Y|ªaz¨`r�Xi“QZƒA\„EaˆIeŒMgŽMgŽKf�Jd‹Lc†PY{IRtBi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QkbSkbSkbSkbSkbSkbSkbSkaUlbYlbYlbYlbVlbVlcRlcRldQldQldQlcRlcTlbVlbXlbXlbXocUnbRkbQlcRlcRkbQg`Pd]M^WGYSCSM?MJ;KH9GE6EB3?B/;H.?H-OC-a=-ƒE:«XR´NJž,+²:9¹=;ÓSRòpnþvvñefëX^ð^_ãUQÚPFÔG@ÐC:ÑD=ÔG>ÖGAÖG?ÙHCÚJBÜHDÝJCÞJFÞKDàKGàKGâHHáGIÜHHÙGGÖHD×HDØGDÛEDàBAäABæBAèBBçCBäB@ßA>ÛA?ÛJGÊC?¹<6®>3¢@3˜A0�A0�D3ŠA2ˆ@1‡?1…=/‚<0€</=1=1�;1�;1ƒ:3�;1�;1€</€<1=/€</€=-�<-€;,‚:,�9*‚8+ƒ7)Œ:.�8.‘9-”:/—;.š=.ž</Ÿ>.™9)–9(“:*�:+Š;,‡;+„;,‚:+‚:,�9+7)~6({5){5){5)z6)x6*x8,w9,s8*p4)o5)r:-v<0‹M@‰K<„J<‚M=…QC‹VH’XM—WN“NI™TO—VR–]V�e\{^PfVF`^GEO->N)?O(J\2_rEk~QexKXj@M_7@Q-6E&7D*=F1=D4:>07;-8:,:;-68-8;0>B47@-9F,I[5[qBe�Fr‘Ns—Ot˜Nr›Os›TpœQmœLo Os¤Ux¨\z¨`s [i“Q`ŠHY�B^†GeŒMgŽMf�Je�Ge�Gf�La„NXzHQsAi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QkbSkbSkbSkbSkbSkbSkbSkbSlbXlbVlbVlcRlcRldOldOldMldMldMldOldQlcRlcTlbVlbVocUnbTkbSlcTmdUlcTi`Qf_Ob[K^WGXQARL<OI9JG6IC3AC.:G+@F*U@+xG9£PL´HH¶8<º39ÈAEÄ?@ÊJIßb`ñqpðnnçbeâZZÛQNÕLDÏD?ÍB;ÏD?ÐE>ÐB>Í@9ÒC?ÒC=ÔC@ÕD?×CAØD@ÙECÚDCÝADÝADÞDFÝGHßIHàJIâHHåGHçACêADìBEìBEçAAâ@>Û=:Õ=:ÚFFÉ=<¹83°;2¦@4˜?1�>0ŽA1Š</Š<0‰;1ˆ:0‡81†91…:4„;4�;3€<3�:4€<3<3}=1}=3|<0|<0|=.|=.{<-|:,{9)z8*{7*‚:.ƒ9.‡9-‰9.�9.‘;.•<.–=-’9)‘:)�:*Š;*‡;+ƒ<*€;+~;*€8*€8,~6*}5)|4(z4(z4*y5*u2)v6,v8-r6+o3)o5*s:/x@3‹QC†N?�N=‚Q@…VF‰XIŽUJ‘RI’OI—TN“TO‘\VŠfZn[JPK7EK1AN0KZ9WgC`rJgyOdzLVl>G]/AV-7H$1@!6B*=F5<D98=67:39;.;<.68-9<1?C57@-7D*FX2Vl=a}BoŽKq•Ms—Mr›OtœUq�RožPq¢Qu¦Wv¦Zs¡Yj—RbŒJ[…E^†H`ˆIcŠKeŒKf�JfŽHf‹Fe‰I`ƒMWyGPr@i`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QkbSkbSkbSkbSkbSkbSkbSkbSlbVlcTlcTlcRldQldOldOldMldMldOldOldQlcRlcTlcTlcTnbTmaSmaSmaSmdUlcTjaRg`Pd]M_YIZSCTN>PK8MH5ID1DC.@F*DD([B,ˆTG¯\X«>A¬/3ÎINÆ?EÀ;>»;<ÄFGØZ[ãefÛY[ÌGHÑJGÐEBÍB?ÐB>ÔFBÔFBÐB>É>7ÏD=ÏD=ÒD@ÔEAÖEBÙECÚDEÝDFÞ?Cß@DÞBEßEEáGGàFFàBCá>?èBDé@Cè>Aç=>ä>>ãA?áC@ßECÖBBÅ98¶50°;4¦?6š>3�=/�?2ˆ</‡:0‡:0‡81‡81‡81ˆ94ˆ;5‚92€:2€:2;2~<0~<0|<0|<0x8,y9-z:.z:.{<-{<-z;,z;,};/9/€7.�7,„6*…5*†7*†7(Š8*ˆ9*ˆ;+…<-„<-€=,~<,};+}9,}7-|6,y5*z4*x4)x4)x5,r2(u5+v8-r6+n4)n5*t;0xB6ŠUG„PB€O@„SDˆWIŒVJŒRG�NG•TN”UN‹RK„XMdSeZDGK2=J,IX9ZkIhyUcxQYnESh=Nc8J_4:L&0?,95>)=C5<B8:<79<5;=2;=079.9<1?C58?-8B)EU1Uh;azCnŠMr“Nr–NršQs�Sq�Rr Uu£Xv¤[t¢ZošUf�N^ˆH[ƒDdŒNa‰K`‡Hb‰Hf�JgŽKeŠEa…E`ƒMWyGPr@i`Qi`Qi`QjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbQkbQkbQkbQlcRlcRlcRlcRlcTlcTlcTnbTk]Pp`QtdWrdWnbVj`Th`ShbTgaSebQc]M[VCVQ=SN8NH2GA)KG,I?$bI3‘gW¶xm»le¸YW¹OO¶BE½BEÂDGÈFHÍHKÐJKÒLMÓKKÏEEÐDEÓEDÖFEÕFBÒC=ÏD=ÏF<ÈD8ÉE9ÊE<ÌE?ÒD@×CCÚADÛ?CáBFâBDáCDàDEàDEáFDãEDäDDãCCäB@ä@?ã?>ã?>â@=â@=ÞC?ÓB?Å>:¶93ª70Ÿ:0—=2Ž>3†>2ƒ?2�?3�=2�;3„93‡83ˆ81‰92…;0„<.„<.ƒ;-�;/€:.€:.€:.~:-~:-~:-~:-};/};/};/};/}90}90~80~8.€7.€7.€7.~8,€8,~8,}9,|8-{9-z8,x8,x8,x8.x8.w7-u7,v6,t6+s5*q5+m1'r6,n4)m3(o6+m4)q8-I=…SH…SJ†RGˆRH‹PH�QJ”QK•RL“TM�TK‰[N‚cQo_HYU:MV7O_;j~YbxQYoHUjCRg@Mb;IY5CR17D&4>#2:#6;'<>0@B5@@6<>358-39-4:.7:/9=/;?.=B,<E(S_;arFn†Rs‘Ur•QršSs�Sr�UužZ~§e~§ep˜YbŠK^†G^†H^„G`†IbˆKdŠMf�NeŒMcŠK`‡H^„G_‚LTvDKm;i`Qi`Qi`QjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTnbTrbSqaRrbUrdWrfZogZmeXjdVgcWkhYmjYjgThcOc^H[U=RL4PJ0H>%Q?)kP=XG„NBˆE=“D?±WV¸TT»QQÀPOÀNMÁMMÂLJÃKJÐTTÈHGÂ=>Ä=:ÐB@ÓEAÏB;Ç>4ÊD9ÉE9ËF=ÍG>ÔFBØDDÜAEÝAEáBFáCDáCDâDEâDCáFDáFDáFDàEAßD@àB?ßA>àA=àA=àA=ÞC?ÔE?ÈC<º>6¬:0 90”:/Š<0‚<0€A2~@3~>4<3ƒ:3†91ˆ81‰90‡;.…<-„;,„;,ƒ;-‚:,‚:,‚:,9-9-9-9-~:/~:/~:/~:/}90}90}90}90}90}90|90|90|90z:0z:0y9/y9/y9/w9.w9.w8/w8/t8.s7-r6,r6,p6+p6+k1&q7,m4)l3(o6+l3(p7,~H>„RI…SLˆQJ‹PJ�QL”PM•QN”SOŠOG…SH€[K{ePsiPhkLfqQgxT^tMXnGPf?La:K_:K\:GV7CP4;E,8@)6;'7:)<=/?@2>>2;=04:03;05;16<0:</<?.>A,>D*Q[9^mDk�RpŒRp“Sq™Rs�SsžVxŸ^|¢exžak‘T`†I]ƒF]ƒF\‚E`†Ia‡Jc‰Le‹NdŠMbˆK`†I^„I]LRtBIk9i`Qi`Qi`QjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTnbTqdTo_Pm]Pm_RqeYog\ldYhbVfbWqma{xi~{j€zjzubpiVe`JTT<FF.C>(MB.S@/U8(]7*l=3n70u91~;5‡@:’IB›PJ¥XP­ZR»^YºSN·GE¼B?ÈDBÑFCÓD@Ï@:ÐA;ÐC<ÒE>ÔG@ÖGCÚFFÛEFÞDFßCDßCDßCDßCDàDEßEEßECßECÞE@ÞE@ÜD?ÛC>ÜC=ÜC=ÜC=ÛC>ÖE@ÎE?ÃB<³>5¢:1”8-ˆ9,ƒ;-‚@2@1~>2=1ƒ:1†91ˆ81‰90‡;.„;,„;,„;,‚:,‚:,‚:,‚:,9-9-~:/~:/~:/~:/~:/~:/}90}90}90}90|90{8/{8/{8/{;1{;1z:0y;0x:/x:/x:/x:/t8.t8.s7-s7-r6,q5+o5*o5*j1&o6+m4)k2'l6*i3'm7+}G=‡PI‡PI‹PJŽQL‘RM’SN”UP‘VPŒWO‡[P~^OtbNoiQorUlwYfwUQeBK_:EY6DU3EV6GV9GT8ER8?I0<D-9>*7;*:<.<>0;=/9;.6<25=26<26<0:</<>0=@-=C)MW5Zi@f|MlˆNn‘Qq™RuŸUtŸWz¡`w�`n”Wc‰L\‚E\‚E\‚EZ€C_…Ha‡JbˆKdŠMc‰La‡J_…H]ƒH[}JPr@Gi7i`Qi`Qi`QjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTnbTrdWoaTm_Rk_SkaWlcZle[jf]jf]xtiˆ„x‘��•��’Œ~‰‚r}zg`cNOT>DG2FE1HC0F=,L<,VC4P9+R6*V4*Y5)\6+a9/e=1m=1ŠKBœRI®UO¶NK¼FDÆABÒDCØFFÔ@>ÕA?ÖB>ÖE@ÙECÚFDÛEDÝEDÞDDßCDßCDÞDDÞDDÝEDÝEBÝEBÛFBÚE?ÙD>ØE>×D<×D<ÙD=ØE>ÔD<ÒE>ÉF>ºA8§;1˜8,Œ9+…<-ƒ@0@1~?0=/ƒ:1„:/‡9/ˆ:.„:-„;,„;,ƒ:+‚:,‚:,�9+:+9-9-~:/~:/~:/~:/~:/~:/}90|90}90{8/{8/{8/{8/x8.{;1y;0y;0y;0x:/x:/x:/v:/s7-s7-s7-r6,q5+o5*o5*m4)j1&o6+l3(h2&k5)h2&l6*|F<ŠOGŒOJ�PK�RM�SN‘VP�WPŒZQ‰]R…aUy_PiZGd_IdhO\fKN]@EU8AQ4=M0<K.?M3CQ7FR:GS;BL4>G2:A/9=,9=.9=.8<-7;-7=36=56;46<2:<1;=/<?,<A*JS4Ve>bxIj†Ln‘Qs›Tw¡Wx¢Z{¢ar˜[e‹N\‚EZ€C\‚E\‚EZ€C_…H`†Ia‡JbˆKa‡J`†I^„G]ƒHY{HNp>Eg5haQhaQhaQibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlcTrfXth\sg[mcYjaXle]snhzwp�~w��† �”¨¥œ«§œ¨¤™£�‘™–‡z{kcfSOR?GJ7EH5BE2BE2EH5KL<JH9GD5D>0A9,A7+@6*F4(S5*qB8ŒPHŸPK¯IG¿GIÌEIÓBEÖ@BÙ@BÚBAÛCBÚEAÚEAÜDAÜDAÞDDÞDFÞDFÞDFÞDDÝEDÝEDÛFBÚFB×FAÖF>ÔE=ÔE=ÓF=ÔE=ÕF>ÔA:ÓC;ÎE=ÁB9®>3ž9-‘;.Š</„?0�?/€>.€=-�;/„:-†:-‡;.„:-ƒ;-‚:,‚:,‚:,�9+9-9-9/~:/~:/~:/}:1}:1}:1}:1|90z:0|90y9/y9/x8.x8.v8-y;0y;0x:/v:/v:/u9.u9.t:.r8-q6.q6.p5-o4,n3+n3+m4+j1(m7-j4*h2(j4*f2'j6+}D;�NG�NH�QJ�TL�UOŽWPŠYRƒ[Q{YMv\OkZJ]UBYXDY^HOYA?K3:H/7E.4B+5A+8D.=I3CL7EN9BK6?H5<E2:A/8?/7>.5<,4:,5<45<56;56;49;0:</;>-;@)GP1Ra:`vGi…Ko’Ru�Vy£Yy£[xŸ`n‘W`ƒIZ}C[~D^�G^�G]€F`ƒIa„Jb…Kb…Ka„J`ƒI_‚H^�IW|ILp@Bf6haQhaQhaQibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlcTkaUndZpf\lcZkd\rmg„�|‘�Œ§£ µ±®ÄÁ¼ÉÆÁÉľÅÀºÀ¹±¹²¨¤œ�ˆ‚rfcTPQ?FI8>E3;D1:F28E38E38C2:B3<B4@C8CE:GE9>4(P:/e?6€EA¡PO¿X[ÊQVÊCIÕFJ×EFÚFFÛEDÝEBÞDBÝD?ÞCAÞDDÞDFÞDFÞDFÜDCÜDCÜDCÚDCÙEA×FAÕF@ÓF=ÒE<ÐF<ÑG=ÔE=Ô?8Õ@9ÐC:ÆC9¶A7¥=0–=/Œ=.†>/‚?/�>.€=-�<-ƒ;-„:-…;.‚:,‚:,‚:,‚:,�9+:+9-~8,~:/~:/~:/~:/}:1}:1}:1}:1z:0z:0y9/y9/x8.v8-v8-t8-w9.u9.u9.t:.t8-s9-r8,r8-q6.q6.p5-o4,n3+m4+l3*j4*h2(l6,i3)f2'h4)e1&i5*{B9ŽMG’MH‘PJ�TL�VO‡XN�YO|\Qs\Nk[L`WFVSBPSBJQ?@I64@,3<+2;*09(09(2;*5>-9B1<E4?H7?H5>G4:F2:C25@/3>-1<,3:33954954928919;.;>-;@*CL/O^7]sDi„Mp“SwŸYy£Yw¡Yq˜Yf‰OZ}CX{A\E]€F]€F^�G`ƒI`ƒI`ƒI`ƒI`ƒI_‚H^�G]€JV{HJn@Ae7haQhaQhaQibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlbVj`VjaXkdZkg^upjˆ…€Ÿžš°°®ÃÂÀÓÒÐâáßçæäçãàâßÚÜ×ÓØÏÈμ²¸¥——‡zym]_YIKL:@E1;D/>J6=H7=F5;C4;>5:;3983880==3C<2N71a84ˆHH¯]_Å`dÊWZÍQSÓNO×MKÛKJÞHGáFDãDAâBBßCDÞDFÞDFÝCEÝCEÜBDÜBBÛCBÚDCØD@ÕD?ÓF?ÒE<ÐF<ÐG=ÔE=Ù@:Ù@:ÑB:ÈC:¼C:¬@4™;/Œ9+‰=/ƒ>.�>-€=,�<,ƒ;,ƒ;,„<-‚:,‚:,‚:,�9+9-9-~8.}9.~:/~:/}:1}:1}:1}:1}:1}:1z:1y:1x90x90w8/t8.s7-r8-t8.s9.r8-q8-r8-p7,p7,p7,o6-o6-n5,n5,m4+l3*k2)i3)g1'k7,h4)e1&h4)b0%f4)zA8�MG’MH‘PJŽULˆXN�YOxZOr\Nk^N^WGRPAKN=CJ:9B12:+08)17+17-06,/5+/5)08+2:+4<-9D4;F5=H7<I7;F56C13@,1>-.800621622717829;0:<.:?)?H+KY5[qChƒLq“Vwž[w¡YsœVg�P^�IUx@WzB[~F[~F[~F]€H_‚J_‚J_‚J_‚J^�I^�I]€H]€JUvGJk@@a6haQhaQhaQibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlbVpg^ng_lgatqj‡„¡ œ¼¼ºÍÏÎÏÏÏàààòòòúøùû÷ö÷óðñìéîãßèÐÆÝÁµÁ«�¢‘�}ub^\GKM7CH2AI2AI4@G5BD7AA7B=9C97?:69<5>?7B71J.+d66�QR³giÆnmÀ\\ÇZWÐVS×QNÜLKâHHæCDåBCßCFÝDFÝDFÝCEÝCEÜBDÜBBÜBBÜDCÙCB×C?ÓD>ÒE>ÐF<ÐF<ÔE=ÜC=ÛA9ÒB:ÊD;¿E:°A6œ:-‹5&Š</…=.‚=-€=,�<,�<,ƒ<*„<-‚:,‚:,‚:,:+9-~8,}9.}9.~:/~:/}:1}:1}:1}:1}:1{;1y:1w;1x90u9/t8.s9.r8-q7,r8-q8-p7,n8,p7,m7+o6+o6+o6-o6-n5,m4+m4+j4*i3)h4)f2'k7,h4)c1&e3(b0%e3(xB8ŽMG’OIŽRJŠUM„XOzZOq[Mi]MZTDIJ:?C4>D69A208+/7*6<247058157247025.06,08-19,6A3:E5=J9>K9<I76E22A,1>-,6..400511606718:/:<.:?+<E*IW4YoAg‚Mp’Vv�\užXp™U`†IX{CSv>WzB[~FY|DY|D\G_‚J_‚J^�I^�I^�I]€H]€H]LStGHh??]7icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSjcSkbSkbSkbSlcTlcTmdUmdUmdUmdUmdUmdUmdUmdUmdUmdUldWibZgd]vsn“’Ž¯¯­ÄÆÅÚÜÛêîïóôöö÷ùüüþÿÿÿÿÿýÿûúü÷ôýòì÷áÔòØÇàʵƶŸª ‡ˆ„ibbFEH-DH/CG0FE1GD5F<3C60F42H:9>=9>?:B:7I75cGF’jjÅ“’è­©Ò‹‡ÇtnÀ]XÆSPÕONßJLçFKéFKßEGßIKÚDFÔ;>Ø>@áGIàDGÖ:=Ú@@Ú@@ÙA@ÖB>ÔC>ÑB<ÐA;ÑA9ÞC>ÛA9Ó@9É@8¾B8°@5Ÿ:.�8*Œ=0‡>/ƒ>.�>-€=,<+�=*�<,:+9-9-~8,~8.|8-|8-|8-{8/|90|90|90z:1z:1z:1z:1y=5w<4w;3u:2t91q8/p7.p7.o6-m7-m7-l8-m7-l8-m7-m7-i2+m6/p92o81k4-g2*g2*h3+e0(i7.d2)]-#`0&_/%b2(s>6ŒOJ�PKˆQJ�RJ}YMx^QhXITN>DE5>E5:C25@04</3;04:04:068378366446135005./6..6+.9+2=-6C2:G6:I68G44C.1?.*4+,2.-2..3-45/79.:</:?+6>&DR1WlCh‚Rs”]wž_r›Wk”R[~DXyDTu@Tu@WxCZ{F[|GZ{F^JbƒNbƒN^J\}H_€K`�L^~MTsJFd@<W4icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSkbSkbSkbSlcTlcTmdUmdUmdUmdUmdUmdUmdUmdUmdUmdUldYgd]onj†…ƒ¤¤¢ÁÃÂÖÚÛêîï÷ûü÷ûþøüÿüýÿþþþÿÿýÿþúÿüùÿúóÿðàûëÔíßÅ×ͲÀ»�¤£„‚ƒcgjKLQ3GJ/EC.GB/H>2I:3L95K<9?:7BA?MHE]RPyjg£Ž‹Î²®èÅ¿ÿ×Ñ벩͂|Àc^ÊVVØRSßHMÚADÖDEÕEEØHHÜJKÛEFÕ<>Ö<>ÛACÙ??Ù??ÙA@×A@ÖB>ÓB=ÒA<Ó@9ÜA<Û@;Ó@9ËB:¿C;±A6 ;1’8-Œ=0‡>/ƒ>.�>-€=,<+�<,�<,:+9-9-~8,}9.|8-|8-|8-|90|90|90|90z:1{;2{;2z;2w<4u<3u:2s:1r90n8.o6-m7-m7-l8-l8-l8-l8-j8-l8-l8-k6.n70o81n91l7/i4,g2*e3*c1(d4*_/%^.$b2(^.$`0&p>5�JE‡RL‡XR~XOrVKdRDRI:>?/:A16A05@03>04<14<15;17:379477577557246116//6./7,.9+1<.4A08E49H58G44E24B1.8/.5..3--2+23+46+8:-9=,7?(DQ3WlEh‚Ut•`w�bp˜Yg�PY|DWxCTu@Tu@VwBYzEYzEYzE[|G_€K`�L\}H[|G^J^J^{MSnKF_A9R4icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSkbSkbSkbSlcTlcTmdUmdUmdUmdUmdUmdUmdUmdUneVneVlfZed_y{x—™˜²¶·ÍÑÒäéìôùüúÿÿûÿÿüÿÿýþÿþþþÿþüÿÿûÿÿúÿÿôþüçúùÝïðÑáåÄÒØ´½Å  ¨ƒ‡‘l`gEQX7EH-DC.JD4LD7KA8I>:C;9SJKia_�wvœ’�¾³¯ÙÎÈêÜÓþèÝÿóéÿäÛð©£ÃecµBEÇDIÚRVÛUTÑMKÊDAÍCAÔFE×EFØDDÙCDÖ=?×>@Ø@?Ö@?ÖB@ÓB?ÓB=ÔA:Ù@:Ù@:ÒB:ËE<ÀD<²B7¡<2“9.Ž<0‰=0ƒ>/�>.=-~<,€=-€=-9-9-~:/}9.}9.|8-{8/{8/}:1}:1{;2{;2{;2{;2z;4x<4t;2q;1r90o9/n8.l8-l6,k7,l8-j8-j8-j8-j8-i9-j8-j8-n91m80m80m80m80i7.f4+c1(d2)a1']-#_0&c4*\-#_0&sD<}PJ�VP~YQpTI^J?OC7?<-46(2;*0;+1<.2=/4<15;17:17:379479668357257227007/08-.9+0;-2?.5B17E48G48G49F54>33:2/4-,2(/0(13(57*6:)6>'ER6YmJk„Zw—fwœfl“Z_‡KWxCWuCTr@Tr@VtBXvDYwEYwEZxF]{I_}K]{I]{I_}K^|JZwKPgJAW@3I2icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSjcSkbSkbSlcTlcTmdUmdUmdUmdUmdUmdUneVneVneVneVmg[jkfƒ‡ˆ£§¨¹¾ÁÐÕØåíïóûýõýÿûÿÿûÿÿüÿÿüþûýþùþýøÿþ÷ÿÿóøýæôýÞî÷ØæòÎÝéÃÎÚ´¶Á™¡«†€‡efmLMR4FH0HJ5KI:HD9D@7OGDj`_‹�€¤š™·¯¬ËÆÂÞÛÔêæÝúñèÿûñÿ÷íÿÝÕû¶±å�ŽÊcd°@?ÃPKÉTMÉPHÃF@ÈGBÔMJÖKHÐ@?Ò>>Ó??Õ?>ÕA?ÕA?ÓB?ÓB=ÓB=Ö>9Õ@:ÒC=ÌE?ÀE>²A9¢=5•;2Ž<1‰=0ƒ>/�>.=/~<,€=-€=-9-9-~:/}9.}9.|8-{8/{8/~;2~;2|<3|<3{;2{;2z;4x<4r90o9/o9/m9.l8-k7,j6+h6+i7,i7,i7,h8,h8,h8,h8,h8,m;2l7/j5-i7.j8/j8/f4+a1'd4*a2(].$a2(b3)Y* a2(yNE�\VvWRiNGWC:I<3@:.:;-8<.08)/:*3;.4</6<25;169069079468368349238139/19.19..9+/:,0=,2?.5B17E4;H6<I89C87?428.-3'.0%/1$24&48'4<'ER8]pPr‹d{šnwœif�XU|ESsATr@Sq?Sq?Tr@UsAVtBVtBWuC[yG]{I\zH^|J_}K\zHWrIJ^E<M;.?-gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSkbSkbSlcTlcTmdUmdUmdUmdUmdUmdUneVneVofWofWng]qqo‡Œ�¤©¬·¿ÂÍÕØãíïòüþôþÿøÿÿûÿÿûÿüýÿúüýõúüñøøìöøêöÿèóÿäðýáîùÙéôÒÞèÅÊÒ­¶¾™�£�}‚bY]BGL5FK7FJ9CG9AC8c^X…}z«¡ŸÁ·µÌÂÁ×ÏÍæáÝñîéú÷ðüõëÿñèÿòéÿôìÿåß騤·hc¦LAµPDÁYNÇ[QÆQH¿D=ÂA<ËDAË@=Î@>Ð@?ÑA@ÒC?ÑC?ÑC?ÒC=Ó?;ÒA<ÐE@ÉE@½C>¯@9 >5•=3Ž<1‰<2ƒ=1�=0=1~<.=/=/~:/~:/~:/}9.|90{8/{8/{8/|<3|<3|<3{;2z;4y:3y:3w;3n70m80l7/l7/k6.h6-h6-h6-i7.h8.h8.g8.g8.g8.g8.h8.k92h6/f4-g5.i70h70e4-b1*c2+b3+_0(]1(]1(V*!b6-{ULtYRaNHN?8A7.=6,;9-9;.9<139-3;.5;/6<06<07:/68-47.6904923812:/2:/2:/19.19,.9+.9+/:,0;-3>.6C2:E5<G7=E8:B55;/06(02%/1$13%15$2:%GS;buWw�k|›rq•g\‚QJp=Oo>Qn>Qn>Qn>Qn>Qn>Sp@Sp@WtD[xH]zJ]zJ_|L_|LZwGRlECU?6D7(6)gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSkbSkbSlcTlcTmdUmdUmdUmdUmdUneVneVofWpgXpgXng]lnmz‚…“š «µ·ÄÎÐÜæèí÷ùôþÿõþýöÿüúÿúûÿùûþóøúíõöèñôãôýêôýèôüåóùßòöÛéìÍØ×¹ÅÄ¥±°’�ŽregOMP;GK:FM=GOBKQG{{sŸš”ž¸ØÎÌßÓÓçÛÛóëéü÷ôðïêüüôÿÿöÿýôÿùïÿóêÿçÞÿ×ËØ“ƒ¸eS£J:´TF½WIµF;·@8ÈKEÇC>ÊC?ÌB?ÎC@ÎC>ÎC>ÍD>ÍB=ÐA;ÒC?ÎDAÅD?¸A=ª=8ž;5•<4Ž;3‰<2ƒ=3�=2=1~<0=1=/~:/~:/~:/}9.|90{8/{8/{8/|<3{;2{;2z:1y:3x92w81u91n70l7/l7/j8/i7.i7.h6-g7-g7-g7-f7-f7-f7-f7-f7-f7-h70h6/g5.g5.g6/h70g6/f5.a2*e6.^2)\0'`4+a5,i=4uSJXG@E>6<5/95,;8/8:/57,36+69.5;/7:/69.69.58-57,36+28,19,19.19,19.19,19,19,.9+.9+.9+/:,0;+3>.6A17B2<D5:C28?/5;-57*24'13%04%18&HT>ex\u�mt’lf‰_RwKBh9Kk<Nk;Pm=Pm=Nk;Nk;Ol<Qn>VsCZwG]zJ^{K_|L_|LVsCNg@<J9/:2$.&gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSkbSkbSlcTlcTmdUmdUmdUmdUmdUneVneVofWpgXqhYoh^hikmtz…Œ”£¬±¿ÈÍÓÞâåðòðüüòüûõþùøÿ÷úÿöüÿòúýìøùçö÷çööê÷÷ëú÷èüöæýõâ÷ìÖçØÃÕƯÀ±šž“}un[ZWFPQCSVK[bZeld•˜�²²¨ÕÎÆèÝÙóããúêëÿñôÿ÷ùÿþûþýùùúôøùñüüðÿÿóÿÿñÿûêÿæÑñª”¼o[©VD®RCµPD»LA¿JAÂG?ÆE?ÉE@ÊE@ÊE>ÉD=ÈC<ÉB<ÑC?ÐE@ÌEAÂC=³=9¦;5œ;5•<6Œ<3‡=4…<5�=4<3~<0=1=1~:/~:/~:/}9.|90{8/{8/{8/{;2{;2y:3x92w81v70v70s7/n70j8/j8/j8/j8/h8.h8.h8.g7-f7-f7-f7-d8-d8-d8-f7-g6/h70i81i81h70i81j;3l=5g80k<4b6-a5,oC:xLCtI@nMDC9056.45/8918;247.14+25,58-58-57,46+46+46+46)37)/7(-8(.9+.9).9+.9)08)08)19*08)08+08+08+19,3;.4<-9B1:C0;B0:A1:<.68*24&/3$29'IU?cu[m„gf„bWyTInE?d8Hg;Li;Nk=Nk=Li;Li;Mj<Ol>UrDYvH[xJ]zL^{M\yKQn@G_;6B4+4/#)%gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSjcSjbUlbVlbVmcWmdUmdUmdUleUmfVngWogZnh\oh^mjekormtzƒŠ�¥®³ÃÌÑÕßáæîñôüþôýú÷ýùùþ÷úÿôüÿòüÿñüÿîýþðûøïþúñÿùïÿúíÿùéÿòßïàËÝηÁ²›¢—�|ubb_NXZL]cWmwn|†}¦¬¢¾ÀµÚÕÏîäâüîîÿôôÿô÷þôõþøøûúøûýøüÿúøÿõôüñ÷ýñÿÿñÿùèÿôãÿÌ»½p›QD©SFµSH²C:ÀJ@ÄH@ÇG>ÉF>ÈE=ÇD<ÆC;ÆC;ÍG>ÌG>ÇG>¼C;­>5Ÿ:2—:3‘>6Š=5†=4ƒ=5�=4€<3;0�;1�;1~:/~:/~:1}90|90{8/{8/{8/z:1z:1x92w81v70u6/u6/q6.m80k90j8/j8/i9/i9/i9/h8.g7-f7-g7-f7-f7-f7-f7-f7-f5.j81m;4k:3j92j:0m>4oC8l@5oD;d<2c=2zVJ†dZyWMbLA>:13814927<54:0/5+25,9<336+25*24)03(13(25*46+47,.6),7)/7*-8*/7*/7*/7*/7*19,08+08-/7,/7,/7,08-08+7?09B1<E4=D4;A38<.26'-4$2;*JVBcr[h|a]vXOmKFfAAa:Hd;Kh<Nk?Nk?Kh<Jg9Li=Nk=TqEWtHZwK[xL]zNZvMNjAC[;2>4*00!''heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjdVjdVjdXldYlcZmeZmeXmfVg`NjeRnkXol]he\feasqr�‚†Š�“–ž¡«³¶ÃÈËØÝàêîïö÷ùùûúüþýþþüÿÿúýþöúýòùüñúýòûþóùúòúúòùùíüúëþüçù÷ÞéåÊÕÔ¶º¹›˜˜|psX_dMbkXr}l„�‚‘›�´¹²ËÌÇãâÞòîë÷óòû÷öýùøüø÷ÿþüþþüþþüýÿúýÿúýÿúýÿúýÿúúü÷ÿÿúÿûöÿêåا £`X›D=µNE¿MCÂF<ÄD9ÊE<ÎI@ÎKAËH@ÇG<ÆG8ÃH8½J8³H8¦E5šB4‘B5‹C5†B5„B6‚@4‚>3‚<2„:1†91†9191~:1|91{80z7/z7/w7.v6-w7.w7.w8/w8/v7.u6-q5+n3+l7/i81i81h70h70g6/g6/g6/f5.g6/i70i81h70g6/e4-e3,h3-j5/m80k90j:0i;.i=0kA3gB2jI:dH:v_O~k\ŠxjŒoRJ=79.4:04:039/39/39/28.28.17-17-36-06,25,/5+14+/5+/4-.5-/4-.5-/4-/4-/4-/4-05./4-/4-.3,.3,/4-/4-/6.1;23=26@58B79D67B45@03>-4?.DQ=WeN[kQPbHEX<@T8BW6Ic>Ke>Kf=Kf=Je<Je:Lg>Oj?UpGYtK\vO]wP]wRWqNHb?;O6&1+&+.$),heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjcYkdZlc\md[meZmfVidQjhSmjWjk[gg_lll~„�“š¨¯µ¸ÂÄÑÖÙâçêòóõûüþÿþÿÿþÿþýûÿþúÿÿúþÿùûüôøûòöüò÷ýó÷úñøúïøúí÷úéùûåòõÚßâÃÌЯ´¸—‘–vkqUYaI_kUt�oŒ˜Š�§œ¿Á¼ÓÓÑççåóóñ÷÷õûûùüüúûûùþþüþþüþþüþþüþþüþþüþþüûÿþõÿÿõÿÿþÿÿÿûúÿíêð¿ºÂ~uŸKA¯M@¹K>ÂL@ÅK>ÃE9¼=4¾?6ÃE9ÄF7ÃJ7½L:³J7¦F6™C4‘C6�E7…C7‚B8�A7�>5‚<4ƒ:3†91„93~92}:2z:1z:1y90x8/v7.v7.u6-v7.t8.t8.t8.r6,o5*m4+l7/i81j81h70i70h70g6/g6/e3,e3,g5.h6/j81j81j81j81k60l71k90k;1k;/j>1j@0hC1gH6kP=gRAufS�ue‹†s†ƒrKL<69.39/39/39/28.28.28.17-17-17-17-06,06,/5+/5+/5+/4./4./4./4./4./4./4./4.05//4..3-.3-.3-.3-/4./6/.80/:21=34@66B66B66B45B16C1CP>P]IR`IIW@AO6BP7FV;Jb@Jd?Ke@Ke>Ic<Hc:Id;Je<SnEXrK\vQ]wT[tTTmOC\>6I5&1-',0$)-heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjcYkdZlc\md[meZmfVmhUkiTmjWjk[kkcwww‘’—§¬²¿ÆÌÒÜÞìñô÷üÿþÿÿþÿÿÿþÿÿþÿþýûÿþúÿÿúýþøúûóõøïñ÷ëñ÷ëô÷ìõùëöøêõøåõ÷áîîÔÛÛ¿ÉÊ«°±’“–yuy`kpZt{iˆ��Ÿ§œ®µ­ÍÏÊÞÞÜïïíøøöûûùþþüþþüýýûþþüþþüþþüþþüþþüþþüþþüýÿþûÿÿûÿÿþþþÿûùÿúôÿ÷íýÑÆØž�”L=¥RB©L;§@1·G;ÏYMÏUJ¼@4ÃC6ÆH:ÃK=¹I;ªB7�?5—@7“D=‰@9„A9�A8?6~>5}=4<4}=4z;4x<4w;3w;3u:2t91t91t91r90r90r90r90o9/n8.l6,k5+l7/j81l71j81k60i70i70h6/g5.g5.f5.g6/h70i81k:3k:3l;4j;3i:0h<1k?2jC4iD2fE2gJ8lVAjYGujV†m‘Ž{€mDG658-28.28.28.17-17-17-06,17-17-17-06,06,/5+/5+/5+/4./4./4./4./4./4./4./4./4./4..3--2,-2,.3-/4..5.+5-+6.-9//;12>24@46B47D38E3?L:ER>CQ:;I28F->L3DT9J_@Jb@LdBKc?Ia=G`9F_8G`9RkDXpL]uS^vVZqTPgK=T8.A.&1-',0%*.heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjcYkdZlc\md[meZmfVniVljUlkWlm]pqi€‚�ž¡¦¹¾ÄÄËÑÛåçõúýûÿÿüýÿüýÿþüýüûùÿþüÿÿûÿþùûüôõöîîñæéïãçíßíñâðôåñôãñôßòñÜêêÐØؼÇÇ«²±•¡¡‰‘“}�“€œž‘ª­¢¶»´ÀÅ¿ÜÜÚééçööôûûùýýûÿÿýÿÿýýýûþþüþþüþþüþþüþþüþþüþþüþþþûûýÿþÿÿþÿ÷òïüóìÿüóÿûíÿôäìñ¹�p•P@¢O?¸XJ·M?µE9ÀG<ÁA6ÈE;ÅI?½G=¯@9¡;6š=8–A>�?;‡@<�@:}@;z?9y@9x?8x?8v=4v=4v=4u<3s=3r<2r<2q;1o;0o;0o;0n:/k9.j8-j8-i7.m82n72m61l71l71k60i70i70l:3j92h70f7/f7/f7/h91f:1j?6f>4e=1f>2hC3iE5fG3cG2cJ6hV@i\IskV…‚o��zor_9=,28,28.17-17-17-06,06,06,17-17-17-06,06,/5+/5+/5+.3-.3-.3-.3-.3-.3-.3-.3-.3-.3--2,-2,-2,-2,.3--4-*4,)4,*6,+7-.:.1=14@26B46C2:G5<I78E12?+2@)8F/>N4DX<E]=K`AJb@K`?G_;H^:F^:SiEXpN`uV^uXZnSLcI9M4):(%.+%*.$)-heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjcYkdZlc\md[meZmfVlgTkiTnmYop`svmƒ‡†¤§¬¿ÄÊÎÕÛæðòûÿÿûÿÿûüþýþÿÿþÿþýûÿÿýÿÿûþýø÷øðïðèçêßàæÚÞåÕçëÚêïÛíðÛìðÙîîÖèèÐÚØ¿Ìʱ½§·µ ±®�¶´§Á¾µÉÈÃÒÓÎÙÙ×èèæòòðúúøüüúýýûÿÿýÿÿýüüúþþüþþüþþüþþüþþüþþüþþüÿþüÿüÿÿüÿÿûúüû÷ýÿùþÿøþÿôÿÿíÿþìøßËÅœŠ�bP–J:¥L<´P@¼L>¿@7ÆC;ÅF?¿D=±>;¦:7ž:8™>=‘=;Š?<„@=}@=xA<tB;rC;rC;q?6s>6s>6r=5r=5r=5p>5o=4n>4m=3l<2k;1j;1i:0i:0j:0n72p62p62m61m61l71i70i70l;4k:3h91e90e90e90e:1d<2fB6cA5cC4dD5dG5dI6bI3_I2^L6eV?jbMrmW„ƒo„‡r[`L3:(28,28.17-17-17-06,06,06,17-17-17-06,06,/5+/5+/5+.3-.3-.3-.3-.3-.3-.3-.3--2,-2,,1+,1+,1+,1+-2,,3,,6.+5-*4+*4++5,.8-0;-2=/2=-5@/6A05A-1=)1=)5A-8F/=O5@T8FY=H\@J]?I^=K^>K`?TgG[pQcvZ`tYYkSJ]G6G4&4%$-*$),"'*heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjcYkdZlc\md[meZlfVjgTkjUlnYorasvm€†„Ÿ¤¨¸¿ÇÖßäí÷ùûÿÿûÿÿûüÿþÿÿÿþÿþýûÿÿûÿÿúüüôóõêëíâãçÙÝãÕÜãÑãèÒçìÕéíÖêìÔëëÓèæÏßÚÄÓθÓκÌƶËøÐÉÁ×ÒÎàÛØêæåðïíóòðùù÷ýýûýýûýýûÿÿýþþüüüúþþüþþüþþüþþüþþüþþüþþüÿýüÿûüÿøùÿþýþÿýùÿýôÿøïÿõ÷ÿöð÷çÿÿíÿîÛͧ”¤kX¢[GªWE­L;¹D:ÀD<ÀE>¼E?´A>ª=:¢:9š;9•=<�?=…?=}@;wB<rC;oD;oD;p?8q?6q?6q?6q?6p>5o?5o?5o@6n?5m>4k<2h<1h<1h<1l<2o83q62p62p62m61l71j81i81h70g80e90e:1d<2e?4f@5dB6^B4aG8cL:dM;cM8`K6]K3]M4]Q9bX?mhRss[€ƒnv{eFO:4=*39-28.28.28.17-17-17-06,17-17-17-06,06,/5+/5+/5+-2,-2,-2,-2,-2,-2,-2,-2,-2,,1++0*+0*+0*+0*,1+,3,.5.,6.*4+*4+*4++5,-7,.9+-8*0;+4?/5@/4?.3?+3>-3@,7F/9K3@O8BT:GW<H[?M]BM`DUeJ\oSdtZ`rZXgRFWD4B1$2%$-*$),"'*heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjdVjdVjcYkdZlc\md[lfZlfVliVlkVkmXjn]lrhy}•� ®·¾ÒÛàêóøûÿÿùþÿùúþþÿÿÿþÿúù÷ÿÿûþýøùùññóèéëÞãçØßæÖàèÓäéÒçíÓèìÓçéÑèèÐçåÎàÛÇ×оÚÓÃÙÑÆÞÔËæÝØîäãóéêúñôÿùûúù÷þþüÿÿýýýûýýûÿÿýÿÿýüüúþþüþþüþþüþþüþþüþþüþþüÿýüÿùøÿúùÿþýüÿý÷ÿþîÿùñÿÿòÿýõÿúóùëÿýíÿúèôλ«wb�Q:©[G²N>¸J=¹I>¸I@´G@¯D>¥=:œ:7–=9�>:ˆ?9€A:xA:tD:pF:pE<p?8q?8q?8q?8q?8q?8p?8p?8qB:pA9n?7l=5i=4j>5j>5n=6o83r73p62p62n72l71j81i81d8/d90c;1d>3d@4cA5cA5_C5ZE4_N<gVBhWCcS<^O8[O5\R7]V<]Y>nlUtv^|�kcmU2>(6B.3;.39/39/39/28.28.28.17-17-17-17-06,06,/5+/5+/5+-2,-2,-2,-2,-2,-2,-2,-2,,1+,1++0**/)*/)+0*,1++2+-4-+5-*4,*4,+5,+5,,6+-7,,6+/:,2=/6A17B27B15@04?.2?-4C.:G3=L5AO8DT:JX?L\BTbI\kTcqZ^mXUbPDRA2>0#/%&/,&,,$**heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjdVjdVjcYkdZlc\md[lfZlfVolYmlWjlWgkZflbr{xŽ˜š¨±¸ÊÓØãìñøýÿ÷üÿúûÿþÿÿÿþÿúùõÿÿúýýõøøîðòåéëÝåéØãêØåíÖæìÒéíÒêíÒèèÎèæÏèãÏáÚÈØÐÃØÎÄÝÒÌéÞÚøíëÿôöÿõùÿõûÿùüÿýþÿÿýÿÿýþþüþþüÿÿýÿÿýýýûþþüþþüþþüþþüþþüþþüþþüÿþúÿýûÿýûýüúùýüøÿÿõÿÿóÿÿíýúóÿûûÿúÿÿôÿöæÿóßÿãÍ×­•©oY¨WD®Q@¬O>¯OA±OB­K@¤C<™<5•>7�?6ˆ?8�A7{B7uD6qE8qE8p?8q?8s>8q?8q?8q?8q?8p?8sB;qB:p?8m>6l=5j>5m>6o>7o83q73o83m82m82i81h91f:1f;2d>3d@4bB5`C5]A3\@2WB1TG4[S>f^Gi^HbX?ZS9ZS7\W:ZV;XW;kmUsw^u}fUaI&28D03;.4:04:039/39/39/28.28.17-17-17-06,06,/5+/5+/5+-2,-2,-2,-2,-2,-2,-2,-2,,1++0*+0**/)*/)+0*+0*,1++2+)3+*4,+5-+5,,6--7.-7,,6+.8-2<16A39D69D47B46A10;+2?-5B09F2<I5@N7FT=JX?R`I[hTanZ\iWQ^MBN@/;/",#+1-(.,&,,heVheVheVheVheVheVheVheVifWifWifWifWifWifWifWifWkeWkeWkdZle[md]ne\mg[mgWqn[kmWknYkr`jrgpyv‡‘“Ÿª°»ÄÉÐÙÞêïóöûþüýÿþÿÿÿþÿþýùÿÿúþþöúúðöøëôöèðôãêñßçïØðöÚíòÔëîÑììÒèæÑáÜÉÝÕÈÞÔÊáÔÎèÚÙòääúëîþòôÿ÷ûÿùÿÿüÿÿýþþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþÿúÿþúÿþúþþüýÿþûÿÿúÿÿúÿÿûÿÿþÿÿÿÿýÿûõÿûñÿüìÿûåÿóÛÿêѺ‚i¦bK—O9ŸT?¨ZF£RAžK; M?Ž<0‹=0†>2=/x=/t?/sB3tD6r=5r=7t=8s>8u@:u@:t?9q?8sA:q@9r@9q@9p?8n?7o>7o>7k92m82l:3l;4l;4h<3g<3c=2c?3^>1dG9cG9T=-N9(M8'@1WQ;PN7KI0OK2]Y>ieJeaDVU7XX<]`CosZz€fdlU>J2-9#2>*4</5;14:04:04:039/39/39/39/39/39/28.17-17-06,06,16016005/05//4./4./4./4.,1+,1+,1+,1+,1+,1++0*+0*.5.-4--4-,3,-4--4,.5-.5-+2*-4,07/5=29A6=E8?G<?J<7B25@/3>-1=)2>*9E/BN8IU?O[GVbN]hWYdTLWI<F;/9.)3*)/+(.*'-+heVheVheVheVheVheVheVheVifWifWifWifWifWifWifWifWkeWkeWkdZle[md]ne\mg[khWonZjlVknYkr`jtiq|x‰”–£®´¾ÇÌÒÛàëðô÷üÿýþÿþÿÿÿþüþýùÿÿúþþôûûï÷ùëô÷æðõáêòÝèíÖêïÑæëËäçÊææÌãáÌßÙÉÞÔÊßÔÎçÙØíßßöçêûïóÿôøÿ÷ûÿúþÿüÿÿþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿüûÿúûÿúýÿüýþÿþýÿÿüÿÿüÿÿûÿÿüÿÿüÿÿûùÿüõÿþñÿÿíÿûåÿòÙÿëÑà±—§u\‘[C–ZBšYCšVC™R@ŽG5‹F6ˆE4ƒD3|C2wB0s@/o>/u@8t?9u>9u@:t?9o=6p>7sA:q?8p?8p?8o>7o>7m>6n=6m>6k<4l=5m>6k?6j?6gA6eA5bB5dG9[A2^G7^I8N=+F7$G8%?7"LL4GK2DG,BE*MM1[[?baC`aBYY=dgJsw^sy_X`I:D,/;%7C/6>16<26<26<25;15;15;15;15;15;15;14:04:039/39/39/27127127116016016005/05/.3-.3--2,-2,,1++0*+0*+0**1**1*)0))0))0)*1*+2*,3+.5--4,.5-07/4;39A6?F>BJ?;F8:E57B14?.3?+6B.<H2@L6LXDS_KYdTWbRKUJ=G<1;2+5,+2+*0,)/+heVheVheVheVheVheVheVheVifWifWifWifWifWifWifWifWkeWkeWlcZmd[md]ne\mg[khWmlXimVjoYktaiuiq}yŠ˜™¥²¸¿ÊÐÔÝâíòö÷üÿüýÿýÿþÿþüþýøÿÿ÷ýýóúûí÷úéô÷äïôÞçïØåëÑâçÉÞáÂÝÝÁÞÞÆÞÛÈÝ×ËàÕÏäÙ×îââóçëúîòÿôøÿ÷ûÿùüÿúýÿýýÿþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿüøÿüøÿüûÿþþþþÿüÿÿúÿÿùÿÿøÿÿùÿÿúÿÿúúÿýøÿÿöÿÿóýÿïÿúæÿóÛÿøàÿìÓ㿧¬‚j‡YB…Q;’[F—]I�S?F2u>)r>)s@+tD0vE4q?4n<3q<4tB9sA8o?5qA7xH>n>4m>4m>4m>4l=3j>3l=3j>3g?5gA6gA6gC7eC7cC6`C5^D5bK;UB1ZI7`S@RG3C;&E=(FA+AE,BH.AE*<@%?B%LO2_`AijK``DnqTvzagmSJO94<%4=(=F38>28>48>48>47=37=37=36<28>48>47=37=37=36<26<26<25:449349349338238238227105/05//4..3--2,,1++0*+0*).().().().().(*/)+0*,1+160/4.,1+,1+/4.6;4=B<AH@@K=>K:<I78E34A-2?+3A*5C,DQ=KXDQ^MQ^MHTH<H<2>4.8/-4--4-,3,gdUgdUheVheVheVheVifWifWifWifWifWifWifWifWifWifWkeWkeWlcZmd[md]mf\mg[jiWkmXimVjoYjs`hthm|wˆ˜˜¦³¹¿ÊÐÓÜãëðööûÿüýÿýÿþÿÿýÿþùþþöýýñúûí÷úçô÷âîóÜåíÕâèÎÛàÀ×Ú»ÖÖ¼Ù×ÂÝ×ÇÞ×ÍåÛÙìààøìðûðöÿõûÿùþÿúýÿûûÿûûÿýüÿþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿüøÿþøÿþûÿþþþþÿüþÿúÿÿùÿÿøþÿúýÿùüÿùùÿýúÿÿøÿÿøûÿõüþðÿÿíÿüéÿñßÿðÝÿïÚñÒ½´�zyS>sI3uH3vI4zM8~O;yL7nA.e7'{M@sD:qB8sD:oC8g;0d8-g=1i?3i?3i?3g?3f>2f>2f>2e?2b@4bB5`C5_C5_C5[D4YD3WD3XI6OB/\T?oiSc^HMK4IG0IK3>D*?H-@F*<B&<A#EJ,XY:deFmmQxx\tv^[_F@C.6;%8@+<C19?39?59?58>48>47=37=37=38>48>48>48>48>48>48>48>47<67<67<66;56;56;55:45:438238216005//4.-2,,1+,1++0*+0**/)*/)+0*,1+-2,-2,05/.3-+0*+0*-2.2718=9;B:@K=@M;@M;=J68E13@,1?(0>'<I5BO;HUCIVEDPB;G;2>4.:0,6..5.-4-gdUgdUgdUheVheVifWifWifWifWifWifWifWifWifWifWifWkeWldWlcZmd[md]mf\mg[jiWkmXinWiqZgs_drejyt…••£²·½ÈÎÒÛâêïõôùýûüÿýÿþÿÿýÿÿúÿÿøÿÿóþÿñûþë÷úåðõÞæîÖâèÌÙÜ¿ÕÕ¹ÒйÔѾÚÔÈàÙÑêàßôéíþóùÿ÷üÿûÿÿýÿÿþÿÿþüÿþúÿþúþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿþúÿÿúÿÿûÿÿþþþÿýþÿüüÿûüÿûúÿúøÿù÷ÿú÷ÿýøÿÿûÿÿûýÿúýþùùôîÿþöÿýôÿüóÿûîÿ÷çûæÕéÒÀÏ´¡¶š…”uawVClI6mH6rM;uP>mF7iB3gB2jE5nI9nJ:oK;oK;eA1cB1cB1cB1cB1cB1bC1`C3]F6\G6\G6ZG6ZG6WH5TG4RG3PH3NH2fdM~~frrZVX@JL4GM3<E*<F+?H+AG+AF(CH*MN/TU6xw[{z^nnVUU=CB.?A,>A.:>-;>39?59?58>48>47=37=37=38>48>48>48>48>48>48>48>49>89>89>88=78=78=77<67<66;55:449338227105//4./4.-2.-2.,1-+0,+0,,1--2.-2.,1-,1-,1-,1-.210513763:3;G9=L9@O<@O:<K67F/4C,2A*6E.:I4@O<BP??M>8F71?2-9-,6--4,,3+fcTfcTgdUheVheVifWjgXjgXifWifWifWifWifWifWifWifWldWldWlcZmd[md]mf\mg[jiWkmXinWiqZgs_bpcgxr‚””£²·ÀËÑÔÝäëðöõúþûüÿýÿþÿÿýÿÿúÿÿøÿÿôÿÿóÿÿïúýèò÷àçðÕâèÌ×Ú½ÑѵÍË´Î˺ÖÏÅÞ×Ñìâã÷îóÿöþÿúÿÿýÿÿþÿÿÿýþÿúþÿùþÿúþÿúþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþþýþÿýþÿýþÿýÿþýÿüþÿúþÿùÿÿ÷ýýóýýõþýøþýùÿþüþþþÿýÿÿþÿþùÿÿüÿÿûüûöóÿøòÿýôÿþñÿþíÿúéÿóà÷äÓçÒ¿Ôº©ºŸŽŸ„s�rb|]K}^LtUCaD2W:(Y<*[>,X=*`E2`E2`E2_F2`G3`G3`G3]H3\K7\M:\M:YL9XM9UM8RL6QL6NL5QQ9ikS|€gmqXSY?FL2>G,<F+<F+?H+DJ.EJ,DG*HI*LM.xw[tsWfdMTR;LI6KJ6FE3<=-<?49?59?59?58>48>48>48>47=37=38>48>48>48>49?59?5:?9:?9:?9:?99>89>89>88=78=78=77<66;55:4493382382/40.3/-2.,1-+0,+0,+0,+0,+0,,1--10.21/32/32/32-4-3?17F3=L7@O:?N7<K49H18G05D-8G0<K6>M:<J97E61?2-9-+5,-4,,3+fcTfcTgdUheVheVifWjgXjgXifWifWifWifWifWifWifWifWldWldWlcZmd[md]mf\mg[jiWjlWhmVhrZfs_cqdhys…——¥·»ÇÒØÚãêðõûøýÿüýÿýÿþÿÿýÿþùÿÿõÿÿòÿÿðüÿìøûæïôÝâëÐÝãÇÕÕ¹Î̳ÉÄ°ÊÄ´ÐÉ¿ÚÒÏéàãöíòÿ÷ÿÿúÿÿýÿÿþÿÿÿýýÿúüÿøüÿøýÿúþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþþÿüÿÿüÿþýÿýÿþúÿúøÿùøÿõøÿóøÿóøÿõüÿøýÿúÿýþÿüÿÿüÿÿûÿý÷ÿÿúÿÿüÿÿýÿÿþÿþýûûûóúøìÿÿñÿÿïÿÿíÿþíÿýíÿöçÿîàýêÛúãÑÿæÐãÊ´¥�xt^GeO8bL5\H0^J2\J2\J2]K3]K3^L4^L4\M6YN8WO:XP;VP:UO9SN8PN7NN6MO7PT;`fLfoTU^CEO4AK0<F+?I.?I.CL/HN2HM/FI,OP1YX:qmRjfKa\F[V@YTATQ>MJ9BC3=@5;A7;A7;A7:@6:@6:@69?58>48>48>49?59?5:@6:@6:@6;@:;@:;@::?9:?99>89>89>8:?9:?99>89>88=78=77<67<6495273162/40-2.,1-+/.+/.,0/,0/-10.21.23.23-12+1-+9*/@-6H2;M7=O7<N4:L2:L25G/7I1:L6<M:;L:6G70@3-;.,6--4,,3+cdRcdRdeSefTgeVhfWigXigXhfWifWifWifWifWifWifWkeWkeWkeWldYle[md]mf\kg[jiWikVhmWhrZgt`dqgj{u‰›�ª¼ÀÍØÞàçíôùÿúÿÿûÿÿýÿüþÿúþþöúúîúûíúûéøúåóöáêî×ßãÊÙÜÁÓѸÌDZž¬Æ¾±ÌüÖÎËæÝàôêòþ÷ÿÿúÿÿþÿÿÿÿþþþýÿúüÿúüÿúýÿüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüÿþüÿþüÿþüÿýþÿüÿÿüÿÿýÿýÿþûÿüøÿù÷ÿ÷÷ÿõøÿöùÿöüÿøÿÿûÿþýÿüÿÿûÿÿúÿÿúÿÿûÿÿüÿÿþÿþýÿûýüüþûþÿúùúòúüñýûïþúîÿùìÿûìÿýïÿýìÿðÛÿûâìÜÞŽu`P7RB)XH/XH.[M3[M3ZN4ZN4[O5[O5\P6[Q8UO7TO9TO9QO8PN7NN6KM5IM4JN5LR8S\APY<@I,;F(AL.?J,DM0EN1JQ2MR4JM0LL0YY=kjNhdIb^E^ZA_ZD`[GZUAQO:KJ8AB4@B5@B5>A6>A6>A6=@7=@79?59?59?5:@6:@6:B7:B7:B7;@9;@9;@::?8:?99>79>89>7;@:;@9;@:;@9:?9:?8:?9:?98?87=94;4382160/4..3/.3/.21.21,2.,2.+2++2*,4))6%$6)=!2F+8L1;O6;M5;M5;M58J29K3;M5<N6<N87H51B0.<--7,.5--4,]eN^fO_gP`hQdiUfiVgjWijXfgUhgUheVheVjdVkeWkeYlfXlfXkhYmg[kg\mf^jf[ieYgeVghVfkUenYbo]dqhsƒ€–¨¬¸ÇÎÒÛââéïò÷û÷üÿøüûüþùüÿöüþñ÷øêùúêø÷åòñÝîíÙëêÖàßÊ×Ò¾×ϼËÀ®½´¥½³©Â·±ÌÂÁÛÒ×êàèúóûüöÿþûÿÿþÿþýÿüýÿüþýüþýþþþþþþþþþþþþþþþþþþÿþüÿþüÿþüÿþüþýùþýùÿüùÿýúÿþûÿþûÿþÿÿýþÿüýÿýþÿÿÿþÿýýÿüûýøÿÿûüý÷ÿýøÿþúÿýøÿúöÿ÷ôÿùùÿúÿÿüÿÿýÿÿÿÿÿÿÿýÿüýÿþüþýþÿÿÿÿÿÿþüÿþùÿþóÿþîÿýêÿýæúôÜÿúàçàÆ�ˆlUN2WN1TK.SJ+WN1WN1UN1UN1TO2TO2TN4SO4QM2RN5QO6PP6MO7KO6HM6GL5JP6FO4AJ-=F'>H&DK*FM+EL*PV4PT3KN/HK,TT8bbFccIZZ@[Y@\ZA][B][B[Y@XV=US:SQ8HF1GD1DA0A?0@>1>>2??5>@5;>59?59?59A69A48C58C58C5:B79A69@88@58?79A69@8:B7:A9:B79@8:B7;B:<D9=D<>E=8B:8B:7A88@56>14</39/271.40/51-7/-9-.;)0@&6G'7M';V+:W+;U0;T4;R8=P:>O<>P::L6@R8EW=FY=DX=@T;5H2+<),6+,3++2*ZgM[hN\hP]iQbjSckTfkWhkXghVghVifWifWkeWlfXmeZmgYmgYkhYkg[kg\jf]jf[hfZefVhiWglVfo\erajwnzŠ‰œ«²ºÈÑÐÙàáéìò÷ûøþþùþúúÿøûþóùýïøùë÷øæõôâíìØéæÓåâÏÜ×ÄÒ˹ʿ­Á´¤¹­¡¼¯¦Á¶²ËÁÀÛÐÖæßæûóþý÷ÿþûÿÿþÿþýÿüýÿüýÿüýÿþþþþþþþþþþþþþþþþþþÿþüÿþüÿþüÿþüÿýúÿüùÿüùÿýúÿýùÿþúÿþýÿýüÿüýÿýþÿþÿÿþÿÿýþýûüÿþýÿþûÿüúÿøóøíçòåÝöèßûðêÿúýÿüÿÿýþÿÿýÿÿûýÿüýÿþüýÿýþÿýþÿÿýÿÿþüÿþõÿÿîÿþèÿÿäúöÝýûâäàÇ‹‡lRM0SN0QK+QK+UO/UO/TO1TO1SO2SO2QP4QP4QM2PO3PN5NQ6MO7JP6HM6EM5HQ6EN1BK.?I'BI(FM+JP,KQ-NR/PT1VZ7^bA_bCY\?TW<TT<VV>WU>XV=YW>YW>XW;VU9XT9QK3OI3LG4ID1DA2B@3A?3>@5<=59?59?59A67B48C57D37D3:B79A69A68@58@59A69A6:B7:B7:B7:B7:B7;C8<D9=E:<F;=G?<H><F;;F6<D5:B39?38=67>74>63=40>-3C)9M*BY/Ga1Op;Ln;Li=Fa>BY?<R=<M;:K8=O9DV<K^BM`BL`DH\@:Q71D.+5*+2*)0(ZgM[hN\hP^jRbjSdlUglXilYijXijXkhYkhYlfXmgYnf[nhZmgYkhYkg[jf[jf]ieZhfZfgWhiWejTdmZgtco|s�Ž�¬³¶ÄÍÍÖÝÞæéñöúøþþúþýùþøøûòôøêóôæðñßêé×áàÌÜÙÆ×ÔÁÐ˸ÇÀ®¿³£½° ¼°¤Ã¶®ËÀ¼ÕËÊãØÞëäëüôÿý÷ÿþûÿÿþÿþýÿüýÿüýÿýÿþþþþþþþþþþþþþþþþþþþÿþüÿþüÿþüÿþüÿýúÿýúÿüùÿüùÿüøÿüøÿýüÿüûÿûüÿüýÿþÿÿþÿÿýþþüýÿþýÿþûÿýûüóîêßÙâÕÍëÝÔ÷ìæÿúýÿüÿÿýþÿÿýÿÿûýÿüýÿþüýÿýþÿýþÿÿþÿÿÿýÿþõÿþíÿýçÿþãû÷ÞþüããßƉ…jPK.QL.QK+QK+TN.UO/TO1UP2TP3SO2QP4PO3QM2ON2PN5MP5MO7JP6IN7GO7HQ6FO2CL/CM+FM,JQ/OU1SY5QY2W_8jpJy[qvVY^@KO4MP5QQ9QQ9RP9SQ8TR9UT8YU:ZV;XR:VP:RM:NI6HE6DB5B@4?A6<=59?5:@69A67B47B47D37D3:B7:B79A69A69A69A6:B7:B7:B7:B7:B7:B7;C8<D9>F;<F;?IA>J@?I>>I9?G8>F7>D8=B;>E>9C;6@74B19I/BV3Pg=XrB^J[}JXuIMhECZ@9O:6G54E2<N8EW=NaESfHSgKNbF@W=6I3/9./6.-4,ZgM[hN\hP^jRblTemVhmYinZklZklZlj[liZnhZnhZog\oi[liZkhYjfZjf[ie\ieZhfZfgWfgUbgQajWerao|s~Ž�–¥¬«¹ÂÇÐ×Úáçîó÷öüüøüûõúôóöíîòäéêÜãäÒÜÛÉÓÒ¾Î˸ÉƳþ«½¶¤Ã·§Ç¹¬ÍÁµÖÉÁßÔÐèÞÝóèîøñøýõÿþøÿÿüÿþýÿýüÿüýÿüþýýÿþþþþþþþþþþþþþþþþþþþÿþüÿþüþýûÿþüÿýúÿýúÿýúÿüùÿû÷ÿû÷ÿýüÿüûþúûÿûüÿýþÿþÿÿýþþüýÿüûÿþûÿýûýôïíâÜæÙÑðâÙüñëÿúýÿüÿÿýþÿÿýÿÿûýÿúýÿþüýÿýþÿýþÿÿþÿÿÿýÿþõÿþíÿüæÿüáþùãÿþçäßÉŠ†mPJ0RM0SL/SM-TN.UO/UP2UP2TP3SO2PO3ON2PL1NM1OM4LO4LN6JP6JO8HP8JS8GP3GP3IS1MT3OV4V\8\d=grHt�U‡’h�›s…�lnuTZaBRV;LN6NN6MM5NL5OM6RP7WS:YU<ZT>XR<TO<QL9KH9GE8EC7@B7=>6:@6:@69A67B47B46C26C2;C8:B79A69A69A69A6:B7;C8;C8:B7:B7;C8;C8=E:>F;=G<=G?<H>>H=>I9?G8>F7>D8=B;=D=9C;6@74B19I/DX5Ri?[uEZ{FWyFTqEIdA=T:2H3/@.->+6H2@R8L_CReGSgKOcG@W=6I31;007/-4,YgMZhN\hP^jRblTemVinZjo[mn\lm[mk\mj[nhZnhZoi]oi[liZkhYjfZieZie\ieZig[ghXghVchRajWerao|s{‹Š�Ÿ¦¢°¹ÁËÔÔÝäéñôòúüôúøñöðíðçèìÞâãÕÚÛÉÒÑ¿ÌɶÉıþ«¿¸¦¼³¢Ê¾®ÓŸÜÐÄæÙÑíâàöìíÿôúÿùÿÿ÷ÿÿúÿÿüÿþýÿýüÿûüþüþýþÿÿþþþþþþþþþþþþþþþþþþÿþüÿþüþýûþýûÿýúÿýúÿýúÿüùÿúöþùõÿýüÿûúýùúþúûþüýÿýþÿþÿÿýþÿüûÿýúÿüúÿøó÷ìæòåÝøêáÿôîÿúýÿüÿÿýþÿÿýÿÿûýÿúýÿþüýÿüýÿýþÿÿþÿÿÿýÿþõÿýìÿûãþúßÿüæÿÿéçâÌŒˆoRL2TO2TM0UN1TN.UO/VQ3VQ3UQ4SO2ON2NM1PL1NM1NL3KN3LN6KQ7KP9JR:LU:IR5JS6OY7SZ9T[9]c?fnG‚’c’¦sŸ±�›ª’¡zˆ”pr{\\bFLP7KM5LK6LJ5LJ5PK5TN8VP:WQ;VP:SN;QL9LI:IG:GE9CE:=>6:@6:@69A67B47B46C26C2;C8;C8:B79A69A6:B7;C8;C8;C8;C8;C8;C8<D9=E:>F;=G<=G?<H>=G<=H8>F7=E6=C7=B;:A:7A95?64B19I/BV3Ne;Uo?Tu@RtAPmAE`=9P6-C.+<*+<)1C-<N4H[?PcERfJNbF@W=5H2,6++2*'.&
\ No newline at end of file
diff --git a/testimgfst.jpg b/testimgfst.jpg
new file mode 100644
index 0000000..a586047
--- /dev/null
+++ b/testimgfst.jpg
diff --git a/testimgfst.ppm b/testimgfst.ppm
new file mode 100644
index 0000000..e158bd1
--- /dev/null
+++ b/testimgfst.ppm
@@ -0,0 +1,4 @@
+P6
+227 149
+255
+0/+0/+0/+10,21-32.52-63.83/83/83-83-:3-:3-:3+:3+72.72.61-61-61-61-72.72.50,50,50,50,50,50,50,50,4/+3.*3.*3.*2-)2-)2-)2-)1,(1,(1,(1,(1,(1,(1,(1,(.+$.+$/,%0-&1.'2/(2/(30)52+63,63,74-85.96/96/:7.@:.A;-C=/F@2GA3IC5JC3JC3LE3LE3LC2LC2LD1ME2PE3QD3P?/V?/`@1mC3}E6ŠG7˜G6¢E4¯H9¶F8¾F8ÂF:ÆH<ÈH=ÊG?ÐE@åFKëDLëDKëEIíBHîACñ>Bó<@õ;>ô:=ó9>î;>è>?ß@=Ö>;Ê@6»=.´?+´?+´?+²?,²?,°@,¯@,­@,©@+¥@,¡A+�A*™A)˜@(•A)’@+�?,‘>,‘>,�=+�=+�>)�>)“>*“>*“>*“>*”=)•>*–?+•?.”@5“?5—=2ž>0¨A0³D1¹D2½D1·?/³B2ªF7œI;†H9mB2T9(C3&<5/95484373262/51.40-3/,63.52-52-52-41,41,41,10+-/*+0*-/*-/*./*./*1-*1-*2-*1,)3)(2('5)):..D66O:9iAB|IH‹NM”POžVY©ck¨o€œtŽ’|¢…«†º|ˆ¼rz«^aŽTKvOEhRIfMH^KH[SQ_ddn{|��‘�˜�—¢¨ž¡§™�¡’Ž’�z{iefTXWEQNETQZTPaTPa/.*/.*0/+10,21-21-52-52-72.72.72,72,92,92,92*92*72.61-61-50,50,61-61-72.61-61-61-61-61-61-61-61-3.*3.*3.*3.*2-)2-)2-)1,(2-)2-)2-)2-)2-)2-)2-)2-)/,%/,%0-&1.'2/(2/(30)30)52+63,63,74-85.96/96/96-@:.A;-C=/E?1F@2HB4IB2IB2JC3JC1KB1JA0KC0LD1OD2PC2O?/U@/]@0iB1wD3„E4‘D4›B2¦B2¬A1³B4ºB4¾D7ÁE9ÅF=ÊC=ßEGçBHèCIêDHìCFïBDó@Cõ>Bö<?÷:>ó:?ï<?ç??Þ@=Ô?;É@6¸=-³@+´?+²?*²?,°@,¯?+®?+¬?+©@+¥@,¡A+�A*˜@(—A(•A)‘?*�?,‘>,‘>,�=+�=+�>)�=(“>*“>*’=)’=)“>)•>*•>*•?.”@5“?5—=2Ÿ?1«A1³D1¼D3½D3º@1´C5«G8šH:‚G9j@0S9*A3&=60:6595484373051.51.40-63.52-52-52-41,41,41,10+.0+,1+.0+.0+/0+/0+0/+2.+3.+2-*3+)3)(5))9--A55M87d>=vEA†KGŽMK™SS£`g¢l|šrŒ�{ …­~…¹y„ºmt¨Z\ŒPIuNDiNEdMH_NI_PM^ZZfmnsƒ‡ˆ“˜’˜�–™Ÿ“˜œ��‘€|nijX]\JURIURYUQ`UQ`..,..,0/+0/+10,21-41,52-52-52-61+61+61+61+81)81)61-61-50,4/+4/+50,61-61-61-61-61-61-61-61-61-61-3.*3.*3.*2-)2-)2-)1,(1,(3.*3.*3.*3.*3.*3.*3.*3.*1.'1.'1.'2/(30)30)41*41*52+52+63,74-74-85.96/96-?8.@:.B<0D>2E?1F@2G?2HA1HA1HA1H?.H?.I@/JA0MA1MA1L@0QA1YB2bC1nC2{C2‡B2‘@/—<*Ÿ;+¦<,®>0³B4¸D7¼F:ÄD;×CCß@DâAFæCFéCEïBDôADø@Bø;?÷:>ô<>ï==æA?ÜA<Ð@8ÅA5¶>-±A-²?,°@,°@,¯?+¯>,­@,ª?+¦@*£@+Ÿ@*›A)˜@(–@)•@+‘?*�?,�?,�>+�>+�>+�<*�<*�>)�>)�>)�>)�>)“>*”?+•?.”A3“?4˜?1¡@0­B2·D2¾C3ÀB3¾D5¸G9«I<™J=€F:fA1P:,B6*?82<74;63:5294183062/51.63.52-32-32-21,21,21,01+.0+,1+,1+,1+.0+/0+0/+0/+3/,4/,5-+3+)4*)7-,=32G53[:5k@9{F>…JB�OM—[]™gs’o‡‰xœ�~«{‚¸s~¶fm£TV‰MEvLAkKAcOFaOJaLI\LKYZZdqux„‰……Œ…‰‘†Š�‚…‰zwykfiXZ[KSSIYTZYS_YS_--+--+/.*0/+0/+10,41,41,41,41,50*50*50*50*70(70*50,4/+4/+3.*3.*4/+4/+50,50,50,50,50,50,50,50,50,3.*3.*2-)2-)2-)1,(1,(1,(4/+4/+4/+4/+4/+4/+4/+4/+2/(2/(30)30)41*41*52+52+52+52+63,63,74-85.85.96-=6,>8,@:.B<0C=/D>0E=0E=0E>.E>.F=.F=,G>-H?.K?/JA0KD2MD3UD2]D0gD1rC/|A/†?-‹:'‘8&š9(¢=+ª@0°E5¶H9¿F;ÒDBÚACÝCCáCBçCBìBBò@@ö>@ø;?÷:>ô<>í?>åA?ÙB;ËA7¿?2³>,¯@,¯?+®?+¯?+­>*­>+¬?+¨?*¥?)£@+ž?)š@(–@'”@(’@*�?*�?,�>+�>+�>+Ž=*�<*�<*�>)�=(�=(�=(�<*�>)‘>,’?-“A3–@3›?0¤@1°B3»C5ÁC5ÁC5ÀF9¹I=¬L@—K>{G:`@3L:,?7*@93=85<74;63:5294173062/63.32-32-32-21,12,21,01+,1+,1+,1+,1+.0+.0+/0+0/+21-40-4/,4,*3+)5-+91/A2/R7.b<1pB5yE:‚IBˆRR‹`j‰i€‚r—|y¨w~µmw²_eŸQRŠLEyL@pL@fODdOFcID[FCVLLX[^cimnsyuw~vyutzlkoa_aSVXJSSI[WX[V\\U\,,*,,*--+..,0/+0/+10,10,30+30+30+30+4/)4/)4/)4/)4/+3.+3.+2-*2-*3.+3.+4/,4/,4/,4/,4/,4/,4/,4/,4/,2-*2-*2-*2-*1,)1,)0+(0+(4/,4/,4/,4/,4/,4/,4/,4/+30)30)30)41*41*41*52+52+41*52+52+63,74-85.85.85,<5+=6,>8,@:.A;/A;/C;.C;.C;.C<,D;,D;,E<-G>/J>0I@1JG6JG4PH5WF2`E2jD/tA.|?,‚;'ˆ:&‘:'™>+¡B0ªF6°J;ºJ>ÊFAÓCBØDBÜDAâC@è@?ï==ó;;ø:<÷;<ó=<ë?=àC>ÓC:Ä@3¹@/¯?+­@+®?,¬?+­>+¬?+«=,©>,¥>+¤?+ ?,œ@+˜@*•?(”?*‘?)�>)�>+�>+�>+Ž=*Ž=*�<)�<)Ž=*Ž=*�<)�<)Ž=,Ž=*�>-�?.”C2–@1�@1¨A2´C5½C6ÂB7ÂB7¼C:´H>¥KB�J@tE;Y?2E9+:6*>93?74?74=52:5194074/74/43.32-32-23-12,03,12,/2+-2,,3,,3,,3,-2,/1,/1,01,12-32.40-3.+2-*2-*50-;0,M7,W:,c=0l@3rC9xKHZb�h~~o–yw©tz¶js´\d£SU’MGƒKAvNAmNBfKAcGA]EBWFEUILUMPU]ca`g`ah``f\Y_SVZLUYKVXM\YT]WW_VW++),,*,,*--+/.*0/+0/+0/+2/*2/*2/*2/*3.(3.(3.(3.*3.+2-*1,)1,)1,)1,)2-*3.+2-*2-*2-*2-*2-*2-*2-*2-*2-*2-*2-*1,)1,)0+(0+(0+(3.+3.+3.+3.+3.+3.+3.+3.+30+30)41*41*41*41*41*41*41*41*52+63,63,74-85.85.;4,;4*=6,>8,?9-?9-@8-@8+B:-B:-C9-C:+D;,F=.I=/G@0IH6IH6MH5SF3[D2dC0lA.s?*{<*‚;'‰;'’=)šA/¡F3ªK9²K<ÃH@ÌE?ÐE@ÕE=ÝD?äA<ë=<ñ;:ö:;ô:;ð<;ç@:ÚC<ËC7»@0±>+¬?*ª?+ª?+ª?+©>*©>*¨=+§>+£>*¢?*ž?+›?*—?)”?*‘?)�>)�>+�>+�>+Ž=*Ž=*‹=)�<)‹=)�<)‹<+Š;*Š;*‹<+Œ=,�>/�>-’C2—A0 @0«A3¸B6¿C7ÃC:ÁB9¸A9¯H?¡ODŒNCoG=S@2@:.77+>93?74>63=52:5194074/74/43.23-23-14-03,03,03,-2+,3,,3,,3,,3,-2,-2,/1,/1,01,12-32.10,1-*1-*2.+8/*F5+M5)V9+`<.e?6jGCwYc�k‚tœ{{¯w|¼luºbi¯[\ŸRM�LC|MAoK?eH>aF?^FB[EDV?AN;@FBGJDMHIPIJQIIQFJRENVIUXM^ZQ`YQ`YS*,+*,+,,*,,*--+..,0/+0/+/.*/.*1.)1.)1.)1.)2-)2-)1-*0,)0,)/+(/+(0,)0,)1-*/+(/+(/+(/+(/+(/+(/+(/+(1-*1-*0,)0,)/+(/+(/+(/+(1-*1-*1-*1-*1-*1-*1-*1-*30+30+30+30+30+30+30+30+41,41,41,52-63.74/74/85.92*:3+;4,=6,=6,>7->5,>6+A9.A9,B8,C9-D:.F<0J>2H@3FD5GF4KF3PE3VC2]B/e@.l=+u=,|;)‚9(‰:)“=,™B/¡F4ªF7¹G<ÁE;ÇG>ÏF<ØE>àC<é@=ï=;ò::ð:9ë<9ã@9ÖC;ÅC5´A.¨>(©@+§@-¨?,¨?,¨?,¥>+¥>-¤?-¢?,Ÿ>+œ?-™>+•>*“>*�>)�>)�>+�>-Œ=,Œ=,‹<+Š=+‹<+Š=+‹<+‰<,‰<,‰<,‰<,Š=-‹=0Œ?/’C2—A0¡A1¬B5¹C9ÀD:ÃB<¿C;¸G?¯QI£YPŽZOpRGVH;BB6<?4<90>71>71<5/94.83-63,63,43.23-23-14-03,.3,.3,,3+,3,,3,,3,,3,,3,,3,-2,-2,.0+12-23.12-0/+/.*0/+4/+>1+C1'K3'S9,Y<4cGDs]iƒs�‡€ªƒ„¼~ƒÇs{Äkq»ce¯VSšIC�K?oI=eF<aE>]EAZDCU>BN8?E:BD=HDDMHFPGGQFHRGLVKUXM^ZNaYLaYN)+**,+++),,*--+..,/.,0/+.-).-)0-(0-(0-(0-(1,(0-(0,)0,)/+(/+(/+(/+(0,)0,).*'.*'.*'.*'.*'.*'.*'.*'1-*0,)0,)0,)/+(/+(/+(.*'0,)0,)0,)0,)0,)0,)0,)0,)3/,30+30+2/*2/*2/*2/*2/*30+41,41,52-63.74/74/85092,:3+;4,<5-<5-=6,=4+=4+A8/A9.B8.C9-D:.F<0J>2I?3EB3FC4JC3NB2SB2Z@1`?0g<,p<.u:*|9)„8(Œ;*”>-šC2£C3±G:·E:¿F;ÈH=ÒH>ÝF?çD?íA=ë<9ê;8ç>9ÞA:ÏE:¿D4®A,¢>'¥@,¤A,¥@,¥@,¤?+¤?+¤?-¢?, ?,ž?-š?,—?+”?+‘?*�>)Œ>*�>-Œ=,Œ=,‹>,Š=+Š=+Š=+ˆ=*‰<,ˆ<,ˆ<,…<-†=.‡>/ˆ>1Š>0‘D2–C1¡A3­C6ºC;¿D=ÀC=ºE>·PI°\R¤f[�fZu_R[UGGM?BJ=;8/>7/=6.:5/94.74-63,33+34.23-14-14-.3,.3,-4,,3+-4--4--4--4--4--4-.3-.3--/*/1,23.12-/0+./*/.*2.+7-+:,)B1)J6-Q=6]IHscn…{”�ŠµŠŒÅ„‰ÍyËqwÃhl¶WWŸHD�E=lE=bD=_B?\B@V@CT>DP=FKGQRKWSQ^US`VR`SR^RS_SY_Q]ZI`YG`YG+-,+-,+-,+-,+-,+-*,,*,,*/.*.-).-).-)/,'/,'/,'/,'.*',+),+)-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*,+),+)-,*-,*.-+.-+/.,/.,/.,/.,/.,/.,.-+.-+.-+.-)/+(0,)1-*2.+3/,3/,2.+2.+2.+2.+2.+3/,3/,51.62/74/70*81+92,;4,<5-<5-=4-=4+>5,>5,?5,@6,A7-C9/G:1F<2G?4G?2I?3M?2R>3W=0[9-a7)k:,r8*{9+ƒ9,Š;,‘>.–>0›>/¦B3¬B4³C5¾F8ÉG:ÓF<ÝD>åC>è@=éA>åC>ÙD=ÈD7¶C1§A+¡A+¡A+¡@- ?,Ÿ>+¡>+Ÿ>+Ÿ>-ž?-œ=+™<+–=+”=*�=+Ž=*Ž=*Œ>*‹<+‰<*‰<*‰<*ˆ;+‡;+‡;+…<+ˆ<.†=.†=.…=/†>0ˆ@2‰A5ŠB4�D4•D1£F7±I>¹F?ºC=ºE>¸ME´ZQ­g]¡qc�tf}qcik]T_QGRDLLBIE:B>3;8/85,74+63*33)34,34,14-03,-2,,1+,1+,1+-2,-2,-2,-2,-2,-2,-2,-2,02-02-02-02-12-12-12-21/4+.:/3A32B5/H;5ZNNwmxŠ…œ’‘»“–Í�•×ƒ‰ÓtzÄgn´^c¥Z[”LJrFCb@=Z=?V?BUAITJU[Sabapmn}vu…{p€uiylguh^l]W`O]\Ha[Eb\F,.-,.-,.-,.-,.-,.---+--+.-+.-).-)-,(/,'/,'.+&.*',+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+)-,*-,*-,*.-+.-+.-+/.,.-+.-+.-+.-+.-+.-+.-+/+(0,)1-*2.+2.+2.+2.+2.+2.+1-*1-*2.+2.+40-51.62/7/,81+:3-;4.<5/<5-=4-=4->5.>5,?5,@6-A7.C9/F90G:1G=3H>4K>5L>3Q=4T<2Y8/]7,g8.m7+w7-€8,ˆ:-�<.•=1™=0¡A3¥A2ªC4²D5½E7ÈF9ÒE<ÙD=ÞC>ßD?ÞG@ÓF=ÁF7±C2£C-�B-žB-�A,œ@+›?*�>*�>,ž?-œ?-™<+˜=+”=*“>*�>+Œ>*Œ>*Š=)‰<*‰<*‰<*‡<)‡;+…<+„;*„;,…<-„<.„<.ƒ=1„>2„@3†B7ˆC4‘H9•F5¡H:¯J@¸IB·HA¶KE²TL¯eZ¨sež}l�o}{llteZfXMYKKM@JH<EC7>>2=:177-44*11'23+12*03,/2+.0+,1+.0+,1+.0+,1+.0+,1+.0+,1+.0+.0+/1,/1,/1,/1,01,01,01,10.3,48.6<23>42G@:[VSvszŠˆ��“¶“—Ç�–Ò„ŒÍx‚Àpz·jsªgo�\_~XZqQViQXhT^hYfleuuq�~s‡~}‘…�–‡|Ž€u‡wp�qdueZfRZYD_Y?a[C*./*./*./*./,.-,.-,.-,.+,,*,,*-,(-,(,+',+',+'+*&+*(+*(+*(+*(+*(+*(+*(+*(,+),+),+),+),+),+),+),+),+),+),+)-,*-,*-,*-,*.-+.-+.-+.-+.-+.-+.-+.-+.-+/+(0,)1-*1-*2.+2.+1-*1-*0,)0,)0,)0,)1-*3/,40-51.80-91.:2/;30<41<5/=4/=4-=4-=4->3-?5,@6-B8/E80F91H;3H;3I<4M<4N;4R94U82[6.b70h6-r6,{7,…9,�;/’</–=/œ@3�@1¡A3§A3°B5¹C7ÄD9ÌC9ÔE=ÖG?ÓI?ËH>¼F8­C3ŸC.™A-œA.›@-›@-š?,š=+š=,›>-š?-˜<-•<,“=,�=+Ž=,Œ=,Š=+‰>+ˆ=*ˆ=*‡<)‡<)‡;+„;*„;,ƒ;,ƒ;-�;/‚<0�=2ƒ?4„B6…B9‡C6“K=–I9 H<«JA²KD±KF±TM­_U¬rd¦�o�Œx�Žz�Šwr€ocqbWcUKRBKM?GI;CE7??39;.35*/1&01)/0(/0*./)./*-/*./*.0+/0+.0+/0+.0+/0+.0+/0+/0+01,01,01,01,01,01,01,1/04,76,78/2;62HE>]^Vxy{‰�˜Ž”®�—¼�•Ä„ŽÁ}‰¹{Šµ|Œ°|‹¨|‰šw„�q~†n}€l~~nƒ~x�„‚š�Š¢’�¨•�¦’‡œ‰€•‚z�zj{i\hRWW?\X;_[@*./*./*./*./,.-,.-,.-,.-,,*,,*-,(,+',+',+'+*&+*&+*(+*(+*(+*(*)'*)'*)'*)',+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+)-,*-,*-,*.-+.-+.-+.-+.-+/+*/+(0,)1-*1-*0,)0,)/+(/+(/+(/+(/+(0,)1-*3/,3/,80.91.:2/;30<41<41<3.<3.=4/<3,=2,>3-?4.@6-C6.D7/F93F93G:4I:3L:6M83P72T50]6/c7.l5.v8-9/‰;/�=1“=0˜@4™@2˜?1�@1¤@1­A4·C6¾D7ÇG<ÊG=ÉI>ÂH=¶F8¨B3œA/—@-˜A-—@,–?+–>*–=+–=+™>,—>,•<,”=,�=+Ž=*Œ=,Š=+ˆ=*ˆ=*‡<)‡<)†;(„;(„;*‚;)‚:+€;,€:.;.€<1€>2‚?6‚B8„D:…E;‘K?•I<œH>¥KC«LF¬RJ«]S©k^§�n Žxšš‚Žš‚„–€x‹wlzian]R[HNUCJN?DH9?C5:>057,13(/0(./'-.(,-',-(.-)/.*0/+/.*./*/.*./*/.*./*/.*./*/0+/0+/0+/0+/0+/0+/0+0./5-85+63-/961HJ=aeWy€xˆ‘�Œ”¡Œ–¯Š•³ƒ�°€‘­…™±Œ£³�©°“¨©�£¡‹ ›‡ž–€š�}˜‰€�‹‡¤Ž“±™–²™“­”‹£‹„œ„|’{i|fXfOSV;XW9]Y<+/0+/0+/0+/0+/0+/0-/.-/.--+,,*,,*,,*,+',+',+',+'++)*********))))))((((((************************************************++++++,,,,,,---------/.,/+*/+*0,+0,+0,+/+*/+*.*)/+*/+*/+*/+*0,+1-,2.-3/,91/91/:20;31;30;30<20;2-<3.<3.=2.=2,=2,?4.B5/C60B73B73C84D93G96G83J71M60V80]7.g6/q8-z:0„<0‹=0�>2–?5–?5•?2—?1œ@1¢B2«D5²D5¹E8¼F:¾H<¹G<¯E8£B2™@0•?.–A-•@,”?+”=*”=,”=,•<,•>-“;-’<-�<,Ž=,‹<-‰<,‡;+…<+†;(†;(„;*ƒ:)ƒ:+‚:+�9+:+;.~<0<3?5�A8ƒD;…F=†G>�K@‘I=˜H? KD§PI¨XQ§g[¤vf¡�u™™}”£†‹¤‡„œ‚|’}s‚ojudZfRU^ILUBDM:>E59@07;-47,01)/0(.-(,+&,+'/+(0,+1-,0,).-)0,).-)0,).-)0,).-)/.*/.*/.*/.*/.*/.*/.*1-.7-85)31,)66*GL8_iQwƒm†’„‹—“�š ‹™¤†—¡‡œ¡�©¨™¹® À±˜·¨—¶¦˜·¥˜¸£�²š‰«’‰«�Ž±“’³”’³”�¬Žˆ¢…ƒ›�{’xg{bTbIPT9TU6XW9+/0+/0+/0+/0+/0+/0-/.-/...,--+--+--+-,(-,(-,(-,(,,,,,,+++******)))((((((*********************************))))))))))))(((******+++,,,,,,---.......*)/+*/+*/+*/+*.*)-)(-)(/+*/+*/+*/+*0,+1-,3/.3/.:20:20;31;31;30;30;1/;1/;2-;2-<1-<1-<1+=2,@3-?4.?61>71?82@93B94B94D71G6/O6/V5,^6,i7,t:.~<.…=/‹=0“?5’>4’>3“@2”A1™B1ŸD2¤C2ªC4¯C6³E8¯E8©C7 @2˜?1•?0“@.’@+“>*’=)“<+“<+“<+“=,‘;,�<,Ž=,‹<+‰<,‡;+…<+„;*…:'…:'ƒ:)ƒ:)�9*�9*~9*}9,~<0}=1~>4A6‚C:ƒG=…IA‰JC�JA�G=—HCŸOH¦UQ¦`X¦qc¢ƒo›•{”¡ƒŽ©ˆ…¦‡�Ÿƒ~–~x‡tp|hcoY[iRP]IGS?@K:<E49?19=/24)01)/.).+&-)&/*'1++2,,0+(/+(0+(/+(/+(/+(/+(/+(0,).-).-).-).-).-).-)1+-6*44)11,&56&CL/ZhEs‚a„’x�œ‰�Ÿ”�ž—Šž•Š¥–•µž¡Æ¥§Ï¬¢É¬¥É¯¤Ë®¡È©—¿�‹³�‡¯Š�³ŽŒ²‹Ž²ŒŒ«‰ƒ¡~™zv�rcw\P`ELR6QR3QR3+/2+/2+/0+/0+/0+/0-/.-/..0/.0/..,..,..,--+.-).-)------,,,+++******))))))******************************)))))))))((((((((()))******+++,,,---.......*).*)/+*/+*.*)-)(,(',('0,+0,+0,+0,+1-,2.-3/.40/:12:12;31;31;31;31;1/;1/;1/;1/;0,;0,;0,<1-?2,>3-<5/;60;83<94<94=82?82@5/G4-N3*W5+b6+m8*x:+<,…<-�=2�<2Ž>3�@3�A2‘B1•B0™@.ž>.¥?1©A4©C7£A4œ>2–>2”>/’?-�?*�=+�<*‘;*‘;*‘;,’<-‘;,�<,�;-Š;,ˆ;+‡;+„;*„;*ƒ:'ƒ:'‚9(‚9(�9*€8)~9*}9,=1~>2?5€A8ƒD=…IA‡JE‰MEŠJAŽH@•JE¡TN§\W©i`§|l£Žy—œ~�§…‡­ˆ€©‡~¡ƒ~˜}}Švw€miw`brXXgPO]FHT@CL;>D6<@257,34,0/*/,'.)&1)'2)*3*+2*(0+(2*(0+(0+(0+(0+(/+(0,)0,)0,)0,)0,)0,)0,)1++6)25)-1-"46 ?I&Ud9o€T€’j‹œ|Ž †�¡‹Š¡‡‹©‡”·�žÉ”¢Ñ›¦Óª¥Ñ­¤Ð¬žË¤�½”‚¯„�¬�…°ƒ…®‚Š±…ˆ«ƒŸzy”sp‰i_sWM]BIO3JM0IL/,01,0/,0/,0/,0/.0/.0-.0-/1.00./0+/0+0/+/.*/.*/.*/.,/.,.-+-,*,+)+)*+)*)))))))))))))))))))))))+)))+),+)**(+*())'(((((&''''))))))***++),,*--+..,..,,+),+)-,*/+*.*)-)(,('+'&0,)0,)0,)1,)2-*4/,50-61.:20;31;30;30;30;30;1/:0.;1/:0.;0,;0,;0,;0,>0-=2,=4-<5-:70991891891991>7/@1*I0)R1(\4*f6*o:,t<-z<-†>2‰=0‹?2Œ@2ŽA1�@1‘@/“=,–=-œ?0ŸB3 C4�A4™@2•?0’?/“=,“=,�=-Ž=,�<+Œ=,Œ=,Œ=,Œ=.‹<-Š=-‰<,‡;+†:*„;*ƒ:)�:(:*:*~;*}:)|:*|:,y:+~>2~>2€@6�B9‚F<…JB‡NEŠOGŠIC�HC’OI˜\Tœh]œuf�ˆu�™€“¢…Œªˆˆ­‹�©‡~¡ƒ}š~y�xs‡ng~baz\ZqUTfLM\EFR>@I8=A379.55-4/+0+(0(&1'(3)*3*+/)).*'.*).*)/))/))/))/))0**0**0**0**/+*/+*/+*0*,4)/4*+2.#35<F#Q`7l}Q�f‡™sŠžy‹¡}ˆ¡z‡§x�µšÃ‡�Ê�œÈ—œÇœ›Æ™–Á”‹¶ˆ�¬~ƒ®€Šµ‡‚¬~‡®‚†©�|œwu�ol„dZnSIY>DM2EI0CG.-/*.1*.1*.1*/2+01+01+01+23-32-43.32-52-41,30+2/*50,4/,4/,3.+1-*0,+0,+-,*.,-.,---/--/+,.*+-()-))+.)/.)-.)/.)-.)-,*-.*+,*+,+)-,*,,*,-(-.)-.)-.)-.),,*,,,---.,-/.,/.,2.+2.+1,(2-)3.(6/)70(81):1*:1*;4,<5-<5-<5-=4/<3.<3.;2-;2-;2-;0,;0,;0.;0.<1/>0-A0)?1(;4*57*28,19.36-83-?-)E)&K($R)%Y.'`5,d<0k@0x@/@.…@0ŠA2�@1‘>0‘;.‘9+’:,“=.“@0”C2“B/’A.�>+�<,•;0–:/’</Œ=.‡>-ƒ?,‚?,�>+…@0‡?0‡>/ˆ<.‡;-†:,„;,‚:+z8*~@1w<,q8'w>-w>-u<+{@0{<-�?3‡E9‰I?‡K@†MBƒOB‡NE�LI”SO”bY‘pa‰}g†‰n‡–y‡Ÿ‹¤†‹¦‡�§ŠŠ§‰…£‡|Ÿ�t™zn–tg�h`„^Y{XYtSZnRVfLNWDIK=A?3=829/-4*)3')3'+2&*0'****',((*))))*()-)*.(*1(+2).2).2(02(00)0/*0/*0/*./)-1++0-(/0";>)S\?nyW|‹d‹›t‰žs‹¢t‹§v‰§s‡ªrŽ´{—¾…œÁ�š¿Œ•¹‰Ž´ƒˆ®}ƒ«y‚©zƒª}‚§{‚¥}Ÿzv“tp‰lf}aYkUMZFHQ<CH4>C/.1(.1(.1(/2)01)12*12*12*44,44,63,63,63,52+41*30)7/,6.+6.+5-*2-*2-*0,+0,+.,-.,---/--/+,0*+/)*.)(--(..'..'..(,-(,-)*-)*-)(-,*-,*,-(-.)-.(-.(./)./*,.+,.----..,/.,/.*1-*1.)2-'3.(6/'70(90':1(<2);2)=4+<5+=6,=6,>5.=4-<3,;2+;2-;2-;0,;0,;0.;0.<1/>0-B/(C1'?2)<3*95,95,83-:1,?-)C+'G)'N+'T/)Y4,_:1d=.o@.w@,@/ˆ@2Ž?2‘=2’;1‘:0Œ:,Š;,Š>.‰@/ŠA.‹B/ŽA/‘?1“;/”:1�<1Š>0„?/�@.@.€?-ƒ@/„?/†=.…<-„;,ƒ;,‚:+<,w7+x<1r:+n9)s>,r=+q9(t;*|=.‚@2ˆF:‰I=‡K@…M@ƒOB‡NE�IG”SQ’f]Œwf‚†m~“tž~‚¥„ˆ§‡Š§‰�¨‹Ž§Šˆ¥‡ ƒt›|m™vd“i^‹`X‚ZX|VYwUXoRRaJMVCDH:@@6:724/,1++2)*/)+-)****)+*))))))+),-(,/)-0*.2(00)00)0/*1/*1-*1,+1,+02-12.-/+(,,$8:,SXBnwZ}‹g…•nƒ™k†žl†¢o…£o‚¥m‡¬v�µ˜¹„–·‚‘³€Š¯{…ªw‚¨u‚¨y‚§{‚¥{�¤|~�{w’soˆkh|cYkUMZHDM<=D4:>//2)/2)03*03*12*23+23+23+55-55-74-85.74-63,63,52+81+70*50,4/+3.*2-)1-*0,)/.,/.,......--/,,.)*,**,.)-/)-/)-/)-.)-.*+.*+.*).-+.-).-)/.*./)./)/0*/0+//-///..,..,/.,/.*1-*1.)4/)4/)70(81);2)<3*>4*=5*?6->7-?6-?6-?6->5,=4-<3,;2-;2-;0,:/+:/-;0.<1/>0-@1*D1*G/+J.+K+,J*-G*,C++A.*=.'@/(C/&M/'W0)b3-e5+i9+n:,v<.=1‡=4‹;4�:2Š:1‰<2‡=2„?0ƒ@0†A2‡B3‹A4ŽB4�>1�>/‹>.‡>-†?-ƒ?,„=+…<+ˆ<,‡;+‡;+†:*‚;)<){=(z=*|>/z</u:,u;-w?0w?0u=.v;-~@3‚B6ˆE<‰I?‰K@‡K@†MBˆME‘JH•TP–f\�ve‡…l�’r�{€¤€…©…†©ˆ‹ªŠŽ©ŠŒ¥‡†Ÿ‚€™|x—wg–l_“eZŠ`Z‡^[ƒ^Y|[SqUPhPFYE@Q?8D60:/,3+)0)(-))+(.((0'(/&'/&'.(*/)+0*,0,--+.,,.+,.)-0)-0(-0&.0(-01/23/..*'**"57)PU?luZ{‰f‘k}”h™j�ŸmŸm} j€¥q‡¬vŠ¯y‰®x†«w‚©t€¦s~¦r~¥v¦y¤y¢z|›yv‘roˆkg{bXjRMZFBI9<@18:,14+14+14+25,34,34,45-45-66.66.96/96/96/96/85.74-92,92,61-61-50,4/+2.+2.+0/-10.1/0//////---,,.+++/*.0*,0*.0*,/+,/+*/+*/+*/.*/.*/.*0/+/0*01+01+01+00.00.00.01,0/+0/*2/*2/(61+73*:3+;4*=4+>5,@6,@6,A8/@9/A8/A8/@7.?6->5.=4-<3.;2-;0,;0,;0.;0.<1/=2.A4.E2,L.,R+.U(/U&.P'-H*,A/+;2):3)<2&G1&R0'^/)b0)g8.j:.t;0|=4ƒ<6‰;7‰:5ˆ;5ˆ?8…A8„B6‚B6…C7†B7‡A7‰A5ˆ@1‡@.‡@.†?-…>,†=,‡;+ˆ;+ˆ9*ˆ9*†9)ƒ:)~;({=(v>'w=)�?1z6+x8,y;.u:,v<.z@2z@2~C5ƒE:†H=‡I>†J?†L@‡NC‰PEŒSJ�[PŽiYŠt_ƒfŠl~–tŸz‚¦‚ƒ©„‡«‡ˆª‡‰¦‡†¡‚�š||—xp–oh“hdŽfd�ee‹fc‡c]~_Zw[OgOH^I?P>6D30;-,4)).'*+%1'&3%%1%'0&'/&'.((.*),,,)+*'+*'+,&,,&,,$-,&+.'+,.*+1,)/*&.+"79+PUAhrWtƒb{�gy�fz•hœl|žlx�iy k~¥p€§r~¥n}¤o{£o|¤p}¥s}¤u}¤w{ u|Ÿwy˜vr�nm†he|`WkPM[DBF7<=/78*25,25,36-36-56.56.56.56.77/880;81;81;81;81:70:70<5/;4.94083/72.61-51.51.51.32.32032021/10.0.//.,1+-3*+1+-1++1++1++0,+0,)2.+2.+10+10+10+21,12*12,22022022012-21-21,41,50*;4.;4,=4->5,@6-A7-B8.C9/C;0C;0C;0C;0B90@7.?6->5,=4-<3,<3.;2-;2-<3.<3.=4/@51E3/L0/S,1X)1W(2R)/I,.A2-:6+66*97(B4'L1&V0'[.(d90h<3q<4z=8‚<:…;:†;8„:7�:6€=7€@7�B9ƒC:‚B9…A8„@5�B0�C.ƒB.ƒ@-…>,‡;+ˆ;+ˆ9*‰7)‡8)…9);({=(x>(t@(v=)€8*{.$€8,ƒ?2v4(s5(|@5}C7}E8F;‚I>ƒJ?‚L@„NB‡QE…WH|aLyiP~pV€v[€|aƒ†i…“r‡�y‚¢}�§€€ª‚€­„€ª„|¤yŸzz™wz’py�kw�iu�itŒjrŠjm„gkd_qWYhQO[GFM;?A3:8+40%4+$5'&5'&2&&0(&.)&+*&+,'*,)',(',('+*)+*)+****+),-)**%"0+'50*86*@C2RZCcpTl}[tˆcqŠ`s�bv—hvšjs™fs›gwŸjy¡lu jtŸjv¡ly£q{¥uz¤t{¢ux�ryœtw–tpŽll…gd{^UlOL[DBD6=;/86*47.47.47.58/67/67/780780991991<92=:3=:3=:3=:3<92=82>71<73;62:5194084173084184165165154243/21/40-3--5-+3--3.+3--3.+2.+2.+3/,30+21,21,21,32-23+23-34/34/34/45/54/85.94.95,=6.=6,?6-@8-C9/D:0E;1E;1F<2E=2E=2D<1C:1B90@7.?6-?6/>5.=4/=4/=4/=4/>50?61?61C52I43O13S.5S/3O/2J20C60<8-;9,<9*A8)H6(P2'V1([3)b3+k4/t72}75�76‚66�74~73}:4|;5}>5?6‚?7…>8„@5€C1~D0€A/�@.ƒ>.…<-†:,‡9,…9+ƒ:+€;+~=+z=*x>*v?*{<*„4)‰2)ŸLD¦XN‰@7{9-‚F;J<yH:xJ;zL=yN>}O@€RC…UG�^JrkNosP{vX…w\�y`–€hžŠr ”z•–wŽœy‹£ˆ¨�„©€¥|z y|šv��o„�n‚�mŽm~�oz�nv‹lqˆki}aeuZ\hPS[DLM;EB1?7*:0&:-'9+(5*&2+%.+$+,&*-&(-&).(+-(-,(/+(2)*3)*6'*6((1(#:3)D>2HF7OR=[cKdqShyWmƒ\i„YjŠ[m‘ap”dn”an–bošetŸjr�hpœgržkw£ry¤vy£uxŸsušqv˜st“qnŒjj…fb{]UlOL\BAC5=9.:3)58/58/58/690780891891891::2;;3>;4>;4?<5?<5?<5>;4?:4?:4>93>93=82<71<73;62:72;83;83:72:6385073063.50-7/,7/-7/,50-50,50,50,41,41,52-52+33+44,44,43.43/540651761:72;81=82=90?80?8.A8/B:/D:0F<0H<0G=1G=3G=3G=3G=3E;1D:0A8/@7.@70@70?6/?6/>71>71?82?82?82@93B94D95G96H96H94G:2C90C9/C9-E9+F9)G:)J9)Q7(Z0$d/'m2,w60‚85†;8Š<:ˆ=8‹B<†A:�>6~;2;2�;3‡<6ˆ?6ƒA3€A2€A2�?1�=0�;/€</€</~<.{=.{=.y>.{=.{=.|=.ƒ9,�1)¡;6ÌkeÙwªZQŒC:ˆLA~M>tJ:qN;oP<nR=sT@zWD�\I}eMmsMnzR„y[–v_§o`³mc»pj»wn»ƒv±Œy¦–}�ž”¢‹¡{„žy„švˆ’p‰�o†‘o„”p€–r{•ru“or‘okŠhhƒbax[[kPT^ELP9DC/@:*>4+=2,:1*51(2/&./'-0',/&/0*2/*6,*9+*=)+A&+D%+A'(C5,KD4VQ>[YD^bIenQhxTg|Ui‚Xe‚Tf†Wj�\n“`m”_n–aošesžio›hm™fpœkv rx¡uxžutšss—qt–ss’pm‹ii„ca|[TmML\AGH8D=3@7.6906906906908918919:29:2::2;;3>;4?<5@=6@=6@=6?<5A<6A<6@;5?:4>93>93=84=84<94<92<94<92<73:70:5194.91.92,91.92,91.72,72.72.61-52-52+52+44,44,44,55-54/54/761880<92=:1?;2@<1@9/@:.B:/C;.E;/F=.I=/J>0H>2H>4H>4G=3F<2D:0A8/@7.B92A81@70?80?82?82@93A:4=82>93A:4C<4E<5G=4J;4L;3K7.M7,N6*O8*O8(O8(O:'T7'f6(q5*}:1‰@9“D?™IB�JDœLEŸPI–LAŒD8„<0ƒ6,…7+Œ:/�=4ˆ>5†=4†=6ƒ=3‚<4€<1}=3y>0w?2u@0u@2y?1}=3�;1‡83�4/•(#°=8ì|zý–‘Ãia•H>ˆJ=wH8mL9gQ:eU<fV=lX@t\D|`JgMxrNƒvS�s[´l^Ç_\ÒUY×QZÖS[ãpsØyuË…{À�~´”}ª•z¢’xž�u›‰q›‹r˜Œr•�sŽ‘t‰’s„’q��q{Œlw†gs|amoWgbL^S?UF3O<.L3,M1-H1+B/(=.'9/&6/%5.$81)90)=.)@-)D*+G)+J'+G)'RB3YQ:d^FfeIglLkuSk{Vi~Uk…Xg„Tf†UlŽ[o”`p—br™dtœhrœjo™im—ipšlw�tyŸvw›wr–rr”su”tr‘qlŠhi„ca|[TnKK^BMN>KB9F=48;27:17:17:19:2:;3;<4<=5<<4<<4?<5@=6A>7B?8B?8C@9E@:E@:D?9D?9C>8C>8B=7B=7@;5@<3@;5?;2A:4?;2A:4A:2@93@70@72?80>71>7/=60;60<71;60:7096/85.74-74-63,66.66.85.96/;60=90@91A:0B90C;0F<2H>2K?3L@2L@2L@2MA5L@4K?3J>2G=3F<2E;1D:0B90A8/?8.?8.>7/=90=90>:1;4,MD=K@:J92S?8R70N/*^;5Z3,^6.^3*[/$\0#e7*k=0q=0ˆJ=ŒG8•I<žLAžG>˜>5š=5�C:¤MC¤QC¨VHªXJ¥OB˜@2—;.›?4’95�:7Œ65…40…73ƒ?6|@6r>0rB4n@0o>0v:/�82‰41�-0˜')ÇGFçc_ÂECËXUàzu«WM…E9„VFmR=g[AhcFhdGhaDm_DzdL‹kR�bJ²n[ÑnhÚY]à;Kæ,Cð+Gò0Kö@WêIYä[eãorâ|zÚ�{؃~Ö‰�Á~u¼‚v¹„v¶‡w¯ˆw«Šw§‹v¤‹w—�l˜k•xh“oaŒbVƒTJzH?w@=t9;g/2d02_11N&$G&!J1,K70G4.D3,D1+E2,F.,D*)J.-O6/XI4`W:e`BeeCioKs}Xsƒ\m‚Wj„WkˆXnŽ]q“`t–cs˜bušdv›hq–jr™ms˜oq–mr”qw™vz™zx—xw”vqŽpj‡ii„eg‚a`{ZSnKM`DNO?QK?OI=7:17:17:17:19:2;<4<=5=>6>>6>>6@=6A>7A>7B?8B?8C@9E@:E@:E@:D?9D?9C>8C>8C?6C?6B>5D=5D=5C<4B;1B;3B;1B92A8/A81A81@91@91@91@91<71;60:7096/85.85.74-74-66.66.96/96-;7.=9.A:0B<0C;0C;.F<0H?0K?1L@0L@0L@0L@2L@4K?3J>2I<3F<2E;1D:0B90B90@9/?8.=90>:1>:1A:2C:1I81J.*U-+i65t88{;;„DB�D?ƒH@„J?„H=�JA˜OFŸPKŸOF‘C6‘B3˜B3�D6Ÿ?3™7,œ7+ž</ C4›@.™@.ŸD2£F5§F6®H9µMB²LJ­IK¤@B˜89‘97Œ=8‚?6v>1n=.yH9„L?„@7‚-*‘*-²:CÒLUÚLKÞOKÈ>;À@=Ï`Y³YP„@3‚UBx`Hc\?]^?hiIrmOugJ}eK“cMºdWÕ_[äUYè@Mò,Cü&Bÿ%Gÿ&Hÿ-Kø2Kð=RîIYèP\ãQ[âR\ßWaÛbiÕflÓimÑmoÍqrÊtsÈwtÆxtÇ|yÈzxÆvwÆqtÅjoÅflÄ`jÃ_iÄ_m³S_¥KU™HO�EI„FIk87FT1-K0)H1)I7-J91K:2M?6SF6[O5cZ9fb?ggCnrMx‚]z‰br‡^o‰\n‰\oŒ^qŽ^s“bu•cu•cu•fr’iv–qz˜vy–w~˜{ƒ�‚„ž…ƒ›ƒ�™�y‘yq‰ok„gg‚c_zYSnKNaCPS@TQBRNB6906907:18;2:;3<=5>?7?@8@@8@@8B?8B?8C@9C@9C@9C@9EB;DA:DA:DA:EA8D@7D@7D@7HA9HA7H?6G>5F=4E=2D;2D<1C:1C;0C:1C:1C:1C:1B;1B;1<8/<8/<8/;7.96/85.85.74-96-96-96-:7.<8/=9.A:0B<0C;0D</G>/J>.L?/M@/M@/M@0M?2M?2L>3J>2I=1G;/E;1D:0C;0C;0A:0?;0?;0?;0?<3B;1UD<R3.i33�@F¬KTÂT_Ë^dÅ^ašB>–H>”L>šNAªQI½TPÊORÄLK£A4›B0žA0£?0¥;-¢6)¢8*¤=.«F4 =*˜7$œ9$¤=*«@.²C2¶D9¥11®9?¹DJ¼KO¸NP­KJ–A<€7.y9-u7*w3(„1+œ45¸<DÓGTåMYáGGÏ61Î:6½30ÇMHÆd[”J=‹[G}cJi`AedEspO~qQ€iJŒhN°hYâbaóRZóCPõ6Hý2Hÿ1Kÿ/Lÿ+Hÿ1Lû/Hõ1Iô4Kò5Kï2Hò2Kó7Oò@XïC[íE\ìI^ëNcéReçTgçUhãThâSgâPgâNhçNjëQmòSsóVuýg„ña|ã[sÔVkË[kÉdr¥RZr-0\&&W0+V;2O?2A;-89)=A0KO8VO2`X4e_=ffBptQ|†c~Œiv‡cw�gsŒer‹dqŒct�ev�ew�fw�h}•sƒš}ˆŸƒŠ¡‡�£�–©•—ª—•¦–�¡�†™†yŒxn…kfb^wWSnKOdESXBWVDXVG58/58/6907:1:;3<=5>?7@A9AA9AA9C@9C@9DA:DA:DA:DA:EB;EB;EB;EB;FB9FB9FB9FB7JC9IB8I@7H@5G?4E=2D<1D</C;0C;0C;0C;0D<1D;2C<2C<2?80=90<8/<8/:7096/85.85.96-96-96-:8,<8->:.A;/B<.D</F=.I=-K?/M@/M@/M@/M@/M@0L>1K=2J<1H<0G;/D:0D:0D<1B<0@<1@<1@<1@>2@=4F<3Q9/Z+%‡;=½U^ÙUcæUfâT`ÊINŸ30–:/�?.—A0­F=ÄJEÔAGË?@±B7¥F4¨B3«A3®>0¯?1±A3³E6±H5¨A.£='¤;&¨;'ª9'«6%­3(¶97º9=¼7<»6;¾=AÄIKÄTR¾XS™?6Š3)†,#�61ÂGIÜOWàBPÙ3=Ý98Í.(Ñ82Ç82ÅD>Ç\R¦WH�XDvV=veG{pPpQ‚gJŽdK©o[ÒthìSVûCOö<Iø9Hþ=Nÿ@Qþ;L÷4Eÿ?Sþ9Mú5Iü2Hü0Iþ.Hÿ/Kÿ5Rÿ2Qý4Tû6Tú7Wø9X÷<[õ>]ô?^öCcõAdó?dó?fø?hûBlÿEpÿHqÿOuÿQrþUvòSqæSmãbvÖftµXb|15g0-W2*N8+E>,=@+8D.>F.PK.]T5f_BnjMy|_‡�q„‘u|ŠpyŒpzŽr{�syŽox‹m{Œl�’r‡–y•£Š›©’¢¯�¥²¡ªµ§®¸­®¸¯ª´«Ÿ«Ÿ” ’�’€q„pe|b\uWRmLPeFPZBUXCWZE47.58/58/7:19:2<=5>?7?@8AA9AA9C@9DA:DA:DA:DA:EB;FC<FC<FC:FC:GC:GC:IB8IB8KB9KC8JB7H@3H>2G=1F<0E<-D:.D:.D:.D:.C;.D<1D<1C=1@9/@9/=90<8/<8/;7.:5/:5/96-97+;7,;7,>7-@:.C;0D</F<0G>/J>.K@.M@/M@/M@-M@/N>/N>1K=0J</H<0G;/D:.C9-C;0B<0@<0@<0?=1@>2A?3I<3\5.u40¬LPÚ^håM\ã@QÖ6B¶%(¨1)›7'’:&–;&ª@2ÁE;Ô>?Ë;:°>3§C3ª@2¬>1±?4¸C9»G:¹I;ª?-¨?,¦=*¥:(©8&°;*·<-º<0ÊG?ÎDAÏ==Ï68Ô8;ÝADàLLÝROÃD=ÆKCÌSJÓTMÙKIÛ?Bà5=à16Ú3-Ù8.Ï2)Î92¾71³A7´\N•TBxN6€bF…jM„aE�[C­fRÊqcämgèEJï:Cè9@æ=BæCHæEJçFKêGLòKSñALò9Gö5Fø1Fý/Fÿ2Jÿ8Rÿ9Uþ9Uþ9Uý:Xý<[ý>_ý?aýAdÿCiþDküClüCmüCmüCoüCoýBmÿGpÿAfÿGmÿMpñJhèNhå\pØbr»Zc‹?A^'$O*!TA2TO;IL7?C,QK3cXBskVzf‹�x—œˆ“œ‹†“‚��~ˆ–…�›Š‹˜‡†‘�‰’�–��¢©—±µ¦¸¼­½Àµ¿Â¹ÂýÄÅÀÁÃÀ¼¾»¯´®¡¨ ‹—‹x†uh{e\sWSlLPhHN[AUZCW\E47.47.47.58/891:;3;<4<=5??7@@8B?8C@9DA:EB;FC<FC<FC<GD=GD;GD;HD;IE<KD:KD:KC8KC8JB5IA4I?3H?0G>/G>/F=.F=.F<0E;/D</C;.C;0C;0A:0@9/>:1=90<8/<8/;60;7.:7.:8,;7,<8-?9-A;/C;.D=-G>/H?.K@.LA/M@/M@/M@-M@/N>/N>1K=0J</H:/F:.C9-B:-C;0A;/?;/?;/?=1@>2@@4L;1k92•FBÃY]ÚW_ÝCOÜ8CÒ38¼+(¬0$ 7"•9 •9 ¢<&´?.Ã:0¾8/§7+¡;-¤8,§7+¯:1¶A7ºE;·F8¬>/ª?/¦;+£5$©5&µ=/ÃE9ÊH;¼8+Î@6ßD?ëAAñ=@ñ8=ê37Ü./Û75áFAçRKåNGÜ=9Ø0/â/2ê65Ø0'Ý;.Ê,!Ï8/½3)¯6-Çg[¶m\�VA‡[B‚V=‰S;©\JÍk^ÞgaãUTïFKî=Câ<>Ø>>ÒA>ÑFAÝOKêVVíNRíDKð:Gõ4E÷2Dø/Bû2Eþ7Lþ8Qý8Tû9Tü:Uÿ:[ÿ=`ÿ@dÿChþ;eý<hý>jü>lú?l÷=ló<lõ:gÿGqþ:`þ>cÿLnþLlõNjïUoâYmçqÁ`i•IKq:5[5*O8*OB1UM:\P@qeU…|m“��œœ�¤§žŸ¦ž“�•“ž˜œ§¡¤¯©¤­¨¡¦¢¥§¢´µ°ÂÁ¼ËÈ¿ÐÍÄÔÑÊÓÐËÔÎÎÔÎÐÐËÏÈÆɽ½½¯±®—ž–‹}l}j^uYVoQTlLRaDU_DV`E47.47.47.47.67/7809:2:;3==5>>6A>7B?8DA:EB;FC<GD=GD=GD=GD;HE<IE<JF=LE;LE;LD9LD7MC7LB6NB4MA3MA3L@0JA0I@/I@1G>/E=0D</C;.C;0A:0A:0@91@91=90<8/<71;7.:7.:8,;7+<8,?9-A;/C;.D=-H?0H?.K@.LA/M@-M@-O@-N>.N>/M=0L</I;.H:-F:,C9-B:-A9.@:.>:.?;/?=1@>2??3N:1p1,¦KJ¿RUº=A¼/5Á/2¿0,¸2'©1 £9!œ=!–;™;Ÿ="¥;#£6"�6'Ÿ7,¢6,¤4)«81¶A8·B9±A5°B5­A4§;.¥7(®:-»B7ÃD;ÃA4Â=*Ì>*Ù5)à)#ì""ù$(ÿ*/ÿ.1ì)'æ-(à2)ß4*à4*ã2*ç0*æ1*Ú3#Õ5%Ð2&Î6+Ä7.»?5Æ`TÐ~p¦eS�VBˆN:¡XGÇeZÞc^åSTçDIôGKëABÝ?<Ñ@;È?7Ä?6ÑF?âLKîIMó?Hø8Gü6Gû5Fø3Dö6Eú;Kú9Nù7Qû6Rû6Rþ7Xÿ9]ÿ;aÿ=eÿ>jÿ?mÿAoÿBqÿBuþBtüBtùBpÿElúAaýFeÿJiýLiÿVqÿ^yò]sñj{åo}Ònv­]^ƒD?f5.gA6qUIwe[Š}u •�«¤ž¯®ª´¶µ±µ¶§¯±°¸»µ¼Â½ÂÈ¿ÂÇÃÂÈÈÆËÓÎÔÜÖØáÙÖæÝØçÝÜåÛÜä×ÞãØàÞÔÜÖÏÖÉÄʹ¹¹¡¦ ˆ’‡tƒpey`\sWYqQYiL[hJ[hJ47036/36/36/56067/780891;;3<<4>>6??7C@9CC9GD=EE=GD=GD;HE<HF:IG;JF:KG;MG9LF8LF8ME:MD;MD;LC<LC<LC:MD5LC2KB3I@1G?2E=0D</D<1B;1A:0A:2@91=90<71<71<71;7.;7,;7,<8-=9.?;0B;1C=1F@2H@3JA2JA0LA/M@-O@-N?,L?/J>0L</K:0J9/I80E8/A8/>:/=;/;;/:<.<=/=>0A?0S9,…<5Â\XÇZW­73ª-)°1*­/#¬3"ª:$«A'ªG*¤E'¡B$¢@#¡>!ž8!Ÿ9*£;0¤:-¦7,¬<1µE:µE9¯A4¨</¨</©;,¬<.ºD8ÆLAÃD;»7+½6"Ì;&Ü:+æ4*ñ+(ø((ü&(ù$&ø**õ.)î2)ê4)æ4(ä2&ã1%à3%à;(Ð1Û>-Î6)Æ8,À@5±A5Ïm`¾hY¥RB¦N@ÊdXåhbæRRé@EóBJê=Aß;:Ô<9Ë@9À=3¶8,¹6,Ã81âHHè?Dï:Cö;Fù;G÷9Eö;Fú>LþBSÿ?Tü<Sü:Tÿ8Uÿ9Xÿ:^ÿ:bÿ>hÿ>lÿ>lÿ?ný>qú@rø@r÷AqöBiþLlÿTsôJgêA^ûTpÿhƒÿeö\vî^uægxÛq}Åpu§geŒ]W{ZQ–w§˜‘¼±«Ã»¸ÃÁÂÆÇËÄÇÎÀÃÌÊÊÖÌÊØÑËÙÕÏÝÝÕâãÛæèáéìãèîäãñçåòèçîäåíàçìßæèÝåâØàÐÉÐÀ¾Á©«¨�–Žy…wkzgbv]`rXbpVboS`mQ01,12-34/5616727828938939:2:;3<>3>@5AA7BD7DD:DF;HH>GG=GH:GH8HI7KJ6LK6NL7NK8JG8GE9IFAMIJNIOGDMD?CNE<LC4JA2KB3LD7KC6F>3B:/D=3B;3@93?82>93=84;62:51B94B94@93>93=:5>:7=<8=<7@@8B?6D@5H@3JA0L?,N?*L@*G@-E?/I;0J70K61K63J88D:8@?:4:01<.9D39B/58#?=(bC/¼j\¶J>¥7*¤6)¦8)£5$¤6%¨=)¡8"¤;%¨?)¬?(¬>'¬9$«6"ª7$¢7'Ÿ8) 9*£<-©?/¬B2¬B2ª@0¨=+«=,­<*³>-½C4ÄF8À>1·1%Æ8*Ë7)Ð8*Õ9,Û7-á6.è3,í0,ñ.,ó-*ò/+ï2,ê5*ã7)Ü9(Ú;(Û:&Ó4Ï2Í7&Æ7'»6'ÀB4ÎTGÇNCÎSKß[Ví^ZîPQë>Bñ8>ö?Dé@CÞCAÓB?Æ?9º;2±;/¯;,±;-ÉK?ÔKCÛFBà=>é<@÷?Gü>Hø:DýCNûCOüCQÿDUý@Vû;Tÿ<ZÿCbÿFkÿAhù<fô9dô=køBrüHxÿKzÿNyþNtöQsïPnåLhçMgûXuÿf…ÿ_ƒòGiöWwæ`yÂ]k¼z~‡a^�yn§‘„»§œÐúÛÖÐÛÛÙÛÜàÝÜäâÜêåØìê×íñÙïôÝñ÷âñõèòôêòóíïùôñøóïöñîôîîñèéìãæèßäãÝáÕÐÔËÉʺº¸£¥ Š‘‰xƒuq}oo|j}†sw€kt}h//-00.11/3315616727827829:2:;3;=2=?4?A4AC6BD7CE8FH=GG;GH:HI7HJ5LK6LL4MM5KJ6KI<KKCPPPWU`XWgUSiRO`NHJLC<G>7F=6F?7G@8E>6B;3C<6?:4?82<71=82>95<73;60<3.=4/=60<73<94<;7=<8>=9@?:B?8D@5H@3L@0L?,O>*L@*A?*@?+D</H:1J73G53D45>64>=97<59?3?A3C<,I9)\B3L;­QBª;*Ÿ1 Ÿ6#¡:'�8$�9"£?(Ÿ;$¦=(­@+²?,´<+´9)¶7&³8(¨9(£9)£9)£<+¨?,ªA.ªA.ª?+¬?+ª9'±<*¿F5ÅG8Á?1»7*¿5(Ë7+Î6)Í9+Ï;-Ò<.×;.Þ7.æ3,í0,ñ.*ô.+ð1)è3(ß8(Õ:&Ó;$Ü<&Ø6!Ô7$Ó<)Ë:)À4%Â:,ÌF:ÐE>ØICçOLñQQòHIï<@ô:?÷BGåBCÚEAÐB>Ã>7·;1®:-©;*ª<+µD4ÅK>ÔNCÚG@â@>î>Aú?Fý@GøCJöCIúENüFRüCSù>Sû=VÿB^ÿFhÿCiú@gö@iøCnûGtüKyÿKzÿKzþItóKoñRrô[w÷^zúWvûMnûEkùEjíEhÙJfÆ_p°nrjKF�r°–‰Ì²¥ãÓÆìãÚéèäèéëéæïêâñõâøûáúÿãûÿçýÿíýÿòüÿ÷üýùøýùöüùôû÷ôøôñöððòììïéëíçéäÞàÚÖÕÈÇŲ³®� ™Ž”Šˆ�…ˆ�ƒ�–…‰�~†�}--+..,/0+12-23.4505606718919:2:<1<>3>@3@B5BD7BD7GG;IG;JH;JI7KJ6MK6NL7NK8KI:NKBTSQ_^dihxll„klŠiiƒ^ZiYQ\OHPHAIE>EE?CC=?@::A;;>:7<73;60=82>:1>:1=9.84+95,;7.<90=:1>;2>;2><0C?4E?3G?2K?/M@0N>.O>,L?,:<$9>'=<*A;-A8/?6/<41940<94@;5G:4L3.W+(k/.†<= GC£;0 2!�2 ¢9&¢<&�9"œ8!¡=&¤;%¨>(¯@,µ@.¹>,»<+¾:+»<-±>,«<+©:)¨:)ª;*¬=,¬=*¬=*±>,­8&¸?.ÉK<ÊH8»6'º2$È:.Î6)Ò6)Ò9+Ó:,Ö:-Û9,â6,è3,ð0-ô.+ò/+í1(å3'Û6#Ó8"Ð8!Û9$Ú5!Ú9'Ú?-Ò<-È6)Å9,ËA6Æ;4Ì?8ØD@äJHéJGéCCì@@ëCBÜC>ÔD<ÌA:Â=4µ9-­7)©:)ª;*«7(ÀD8ÔNEÛJEà@@è>AöAHÿGOôAGòAG÷CLýGSþEUù@Rú>VþA]ÿEeÿDgþDküFoþItÿKxþM{ÿJzÿFwÿGvÿJsúOsÿZzÿa€ýVtïEbùJkýTsÚ@ZÊI]Ø{…·{{_B:�|n¯�‚Õ³§óÛÏúìã÷òîøöù÷ôýóëúùêýÿêÿÿëÿÿîÿÿôÿÿøÿÿüþÿýüÿýúÿýøÿû÷þùöýôõûòóúñôùðóïæçäÞÞÓÏξ½¹¬¯¨ ¦œ�£™œ¤—£«œ�¦•š¢“--+-.)./*/0+01,23-34.45/7808919;0;=2=?2?A4@B5BB6HF:JF:KH9LI8MJ7NL7OM8OL;KI=QPK]\amlzyy‘€� �§}¥yu–qj‰d^zYTkQLbKGXD@N@=HA>E=;>986952:70<8/<8-;9-88.991::0>;2><0></>;,>;*E?/F?-H?.K@.M@0N>.N>.K@.?=(<>)>=+></=;/<90::29:4::2C:3J50T,*g)*ƒ15�9Aª=@¤8,¢7%¥:(¨?,§>)Ÿ9#ž8"¤;&¤:$«<(²?,¸?.½>-À<-Ä<.Â>1»@1³=/¯9+®8*®8*¯:)¯:)±9)´;*¹;,ÁB3ÊF9Æ>0¼2%À2&Í9-Ó5*Ø4*Ø6+Ú8+Ü8,ß7,æ3,ê2*ñ0+ô/)ñ.(ì0'ã1#Ù4!Ð5Í5Õ3×0Ø5$Û;+Ö=/Ð:,Í=2ÒD:Ã6-Ã6/Ê;5ØD@âMGäKFÞC>Õ<6Ð@7ËA6Å?4¾</´9*°8*®:+°<-«3%½?3ÎH?ØGBÜA?ä@AóFLüMRî?Dï>DôCKüHSþHUúCUú@WüA\þCbýDfþFlÿJrÿMxÿN{ÿLyÿIwÿEtÿO|ÿR}ýJq÷LnÿVuÿVs÷MhðFaîLdÐ>SádrþŸ¥Ñ�Žd92kI=¤znÐ¥œõÔËÿéãþóïÿüþÿûÿøòþúïÿýïÿÿïÿÿòÿÿõÿÿøþÿúúÿûøÿþùÿýøÿû÷þùõÿö÷þõöþó÷þôõöìíëãáÚÕÒÈÅÀ¹º²²µª±´©°¶ª¶¾±¯·ª«³¦10,0/*0/*0/*0/*21,32-44,77/88099/;;1==3??5@@4AA5HD8KE9MG9OI9PK8QL8QL9OL=NKBVTUeeowx�„‡¨Œ�¸Ž’ÂŽ‘ÄŒŒ¾†‚´zw¦ol—fd‹\[{SPmLJ`IFWDBM=<A97876285.86*55+56.581891::.=;.?<+A<(B<&G?*G?*J?+K@.L@0K?1K?1K?3L@4H>4E;2@91<928915<45<47=3<:.F3,W2,u87’@B¥=D¦89§;/¨=+ª?+«B-¨?*¤;&¤;%¨>(§8$­:'µ<)º=+¿;,Ã;-È</È>3Ä@4½=0¹9,·7*·7*¸8+¸8+·8)¹7)Æ@4ËC5Å9,¿1%Å2(Ê8+Ï5)Ú3*ß3)ß4*à6)ã5*æ4*ê2*ï0(ò/)ô/)ñ.(ë1%â2#Ù4!Ð5Ì4Ô1Ö/Õ2!Ø6'Ö9*Ô:.Ó?5ÕE<ÔE=Í@7È;2ËA7ÕKAÙOEÐF<Ä:/Ä<0Â=.¾<.º;,´9)³9*µ=-¸>/µ9-¼<1È?9ÑB>ØBAàDEìINõOSê@Cí>CòCJúISýKWûFWùBVúBZùA]úCbýEiÿKqÿNwÿOzÿMwþItÿJwÿS~ÿT}üIpùImÿStÿVq÷OhéE]ãG\çYiÿ–¡ÿµºòžœ�HCi.&¦kcÌ—�ñÉÁÿâÝÿòîÿüûÿýÿùöýý÷ÿÿ÷ÿÿ÷ÿÿøÿÿúÿÿûþÿüüÿýúÿüøÿûõÿú÷ÿùöÿ÷øÿö÷ÿôöþôõüòóñéçäÜÚÕÐÌÍÊÃÈȾÉ˾ÉÍ¿ÈÐŽǾ·Á¸54/43.21,10+10+11)33+44,66.66.88.::0<<2>>4??3B@4IC7LD7OH8QJ8TL9SL:SL<QK?PKGYW\kjz}~œ‹Ž·“™Ë—�טžÜ—˜Ø’’ÒŒŠÉ…„½~~²uv¤lk“ee‡[[wUShKJXCBJ==?:9787376246526556177/;9,?<+C>*E?'G?(G?(H@+HA/HB2GA5G@6G@8H=;H::C98=85664173.91.:05=275)A1$Z8.~C=”FB�><�40¦<.¨=+¨?*©@+¨?*©>*¬?*¯@,ª7$±8'¸;)¾:+Ã9,Ç9-Ì9/Ì<1É?4Å<2Â9/À7-Á8.Á9-À8,À6+Ä7-ÑC7Ð@5Ã0&Ã,#Ï8-Ô<1Ñ3(ß3)ä2(å3)è3*é2*ì1*ï0(ñ.(õ.)ô/)ð/&ë1%à3#×4!Ð5Í5×6"×4!Õ4"Ô4$Ó6'Ô8,Ô<1Ó?5ëXPáRJÓI?ËB8ÊF:ÌJ=ÅE8½>/¼:*º;(º;(·:(¶9'¶;+º?/¿A3ÁA6Á<3Å<6Î@<ÖDDÜHHãJLëLPèBFê@CðCIøJSýMXûKXùFYúG]õA\÷C`úEfþJmÿNuÿOxÿNvüKuÿQ{ÿRyþPuüNqÿVwÿ]{þTníE\÷TiîQb÷erÿ’œÿ‹“ò„‡Ç`c¦LK´keÌ�ˆî½¶ÿÞØÿðìÿùøÿûýüúÿüûÿýúÿÿûÿÿûÿÿûýÿûûÿûøÿü÷ÿüöÿüöÿüùÿûøÿùøÿö÷þòôýññøììðæäæÜÚÞ×ÑÚÖÍÜÚÎÝÞÐÝáÓÛâÚÎØÐÄÍÈ::288055-33+22*22*33+33)55+66,77-99/;;1==3>>4A?3IC7MC7PG8SJ9VK9UL;TK<SJAQKK[Xalm‚‚¥�”Ä™ŸÙž¦ç ¨ð ¤îœ é˜›â”—Ú“”ÔŽ�ɇ‰¼�„¯xz pq�cc}XVkNM[IGREDLBAI>>H:=D;:?:97<92>:.C>+D?)E@*E@*DA.DC1CC7AB:@C<@BA?<EA>ICALBELBGM?IK=IG=HBCI?@<0K9+fF7‚N@ŒJ<Ž?2“7(£>,¨=)©>*©>*ª@*®A,®?+°=*°7$¶9'½9*Â8+Æ8*Ê8+Ð9.Ò;2Î;3Ê:1É90É90Ì91Ì91Ë80É6,Ñ=3Ô=2Ñ9.Ë1'Î4*Ø<0Ù;0×3)ã2(ç1&é1'ì1(ï0(ñ.(ó.(ó,'÷.'ô/'ð1'é3&à5$Ø5"Ò7!Î8 Ô9%Ô7$Ñ6$Ð5#Ñ5&Ó7+Ñ7-Ï8/ãNGçWNæYPÚQGÌG>ÃC8½A5¸?.¶<'¶<%¶<%·:&·:&¹:)½>/ÃA4ÈC:Å?6É@:ÑFAÔIFÕGFÖHGÞJJåEGçACìCHõLSûPYúNZúK\úJ^öD\öC`øEeûIkÿNrÿPwÿPwýQwöVxñVvôUuûXwÿ\xÿ[uúRiñJ^ýWkõTdéM[õ]jßKYæTaåTaÖYaÀfeÇ„~嫧ÿÓÎÿìèÿóñÿøöÿýþúûýøüÿûüÿúûýûûûûúøû÷ôûøóÿûõÿüöÿüùÿüùÿø÷ÿóóüîîùëëïááèÝÛâÙÔáÚÒåÞÔëçÛïíÞïñãëòëÚåßÍØÔ==5;;188044*22*22(33)33)44*55+66,88.::0<<2>>4A?3JB7MD5RF8UI9WL:WK;VJ<SJAQKM[Wemm‡�ƒª�–Ê›£â¢¬ô¦¯ü©°þ¤«ù ¦ò�¤ìž£çœ ß–›Õ”—Ì�’‡‰²yyŸji‰^^zWToRPfOMcMKaIG\DAR@=F<:;=:3?;/@=,BA-@B-AB2@D5?E;=D=;DA:CH=DTBH^LSfU\n\bp_cnabgd`afXUaJBiG=ƒSE”WE’K7�?*˜=(§B.«@,«@,­@,°A-²B.°=*¯7&µ8&»7(Â8+Æ8,Ë7+Ï7,Ó9/Õ;1Ñ80Ï81Ð71Ñ82Ó:4Ó:2Ò80Ð6,ÜB8Ô7.Ï2)Õ7,Û=2Ý<2Ú9/Ü5,ä2(ë0'í1(ð/(ó.(ô-(ö,(õ+'÷.'ô/'î2'è4&à6%Ø7#Ó8$Ï8#Í8"Í8"Î7$Í6#Ð7'Ò9+Ñ7+Í5*Ë4+âMFód\ëaWÔNEÃ@6¹;/´9*¶=(·=%·=&¸;%¸9&»:'À<-Ä>2ÉC8ÈB9ÌE?ÓLFÓLHÍFCÌEBÕGEâHHæCDèEHñLPøSZøQ[÷O\ùN`ôH^ôF_õFcùJiþOpÿQuÿSwÿUyîUsíZv÷^zü]yùUpóMeøPgÿZlÿ[mÿcsïM\÷UdóN^üUgôK^äR_À^_»vqל–ùÊÄÿåàÿíêþôòÿÿýûÿþúÿÿøÿÿúÿþûýúûûùýùöüùôýøòÿúôÿú÷ÿú÷ÿöõûïï÷ééóåäêÜÛçÙÖãØÒæÝÔïçÜøòäýúéýÿñóúóàëçÒÝÙ@=4@>2>;2<:.:7.85,74+63*52)52)52)74+96-=:1@=4C?4H@5LC4RF8TH8VI8VJ:YM?YOFRILUQ_ee�z}¦Œ“Ê–Ÿâ�©õ£°ÿ¨´ÿª³þ«³üª²úª±÷§¯ð§¬ì¦ªçŸ¡Ú˜™ÏŒŽ¿€°sržhe�]Z…VT|XQzXRvVOnMH_C?M=;@=<8@@8=?2?C5BF8?E;;B;:C@>FH>KTDVlQf�`q‹htŒsuŒxŠ�o_j‰Zb’WYœWR¥UL¤L>Ÿ@.œ9"Ÿ9"§<(ª=)¬?*¯@,±A-´?-¶=*¸;)º6'¾6(Ä6(É7*Ï7,Õ8/×9.Ú91Ö71×:3Ø93×82Ö71Õ7.Ú91Ü;3åD<Þ=3Ø7-á@6ü[Qÿmbÿ]Rç?4ä2(ì1(ñ2*ò/)ô*&ó($ö)&÷,(õ.)ò/)ì1(ã1%Ø1!Ò1Ò5$Ñ:'Ë:%Ê<&Ì;&Î;'Ð:)Ñ9+Ñ7+Ñ7-ÙA6Î7.ÜG@ôd\äWPÊA9Å@9µ5(·:&¶<$¹=%º<&¼;(¾;)Á<-Å=1Æ=3ÊD;ÖOIÝXQÓOJÀ?9À?:ÏHDÞHGáFDæJKîPQðQVîMUíKVñLZøQcöOcõKeôJeúMkÿQrÿStüStòWvîXuóYuû[uÿ\uÿYpÿWlýVhÿbsýWeÿ]kÿ[jñFVñDWÿQeîVcÆfgªicÉ�‰å¶°öÓÏÿêæúïëúöóøüûùÿÿøÿÿ÷ÿýùþúþþüÿûøûöòýúóÿúôÿøõþöóÿ÷õÿöôöèçæØ×ãÕÔÞÐÍàÖÍïçÜüõåýøåÿüéÿÿñúÿøçðíØáÞB@4A?3?=1><0<:.:8,85,74+52)52)52)74+96-<90?<3B>5G?4KA5OF7SG7UH7VI8XL>WMDTKLWQ]eb}wz£‰�Ç’œáš§õ¡®ý©µÿª¶þ­·ÿ®¸ÿ®·ü­¶û­´ø®³õ¦ªê¡¢â–˜Õ‹‹Ç~¸sr¬gfŸc^–_V‹`Vˆ]TUOqMIbEBSBAI@BA<=8AB<EF@CD>??=CABNLQWWcbg„pt—~ †}œ‘|›ž|—¤p‡£_t«Xh¶VaºRS¶ID°@4¬;)«;#¬>%¦7#§8%©9%¬9&®9'±8'´7%·5%»3%Á5(É7*Ñ:/Ù<3Û=4ß<3Þ;2à=6Þ<7Ý;6à=6äA:çD=çB<åA8ä@7åA8â?6Û9.Ó1&×5*éI=ÿ[Oñ@6í5+ê+#ì)#ø-)ý0-ü,*ö)&÷0-ó3.í6.ä6+Ü4'Ô2#Ó3%Ï6$È:$È<%É;%Ë:%Î8'Î7&Î5'Î4(Ó9/Ï5+Ô=6åPJåTOÛLFÌA<¶0%¼;(º<%»=&¼;&½:(¾9(Á9+Â:,Æ>2ÉC8ÓNEÛXPÒRIÀD<¾B:ËGBÛLFÞIEäLKëSRìSUéPSéNTëNWöUd÷TeùSiúRküRmÿSqÿSqþSsòMm÷Tsÿ[yÿ_{ÿ[týUlúSgúSeûVf÷R`ý[hý[hñO\öTcýXhßS^¶a^–]Tµ€zÖ©£ìÉÅúáÝøéæûöóùûøøÿþøÿÿöÿýùþúþþüÿüûýøõüùôýúóýøôþöóÿøõÿöôõêèéÛØäÖÓÛÐÊÞÔÊîæÙüõãÿúæÿýèÿÿïùþ÷çðëÙâßEA5D@4C?3A=1?;/=9-;7,;7,73(73(73*84+:6-<8/?:4@<3E>4IA6NE6RF6SF5UH7WJ:XKBVLKVPZa_wsuœ„‰À�˜Û–¤ïž«ù¨´ü«µü®¸ÿ°ºÿ²»ÿ²»ÿ³ºÿ´¸ÿ®²ù©¬ñ £è˜˜Þ��Ó‚�Çwv¼rm°i`¡g]™cZ‘]W…TRwLLfCDV??I76;A=>JB@L@@N>?ZFHoV\€ds–r–£z¦­~¨±z¡·u›½p’Áb€ÀTnÁG\ÎKYÐIPÄ=:¹6,¶9'µ=%°<#®;&®;(°;)³;*µ<)º=+¾<,Á<-Ã7(È9+Ò;0Ù?5Þ@5á>5à<3Þ:1Û81Ø5.×4-Þ93æA;êF=éB:ã<4â;3Ù5+×5*Ù7*Ò2$Ë+Ó6'éG:õK>ò@4ì4,î-(ö,*ú,,û++ö*)ï)(ë.*è3,á5+Û5)Ö2&Ô2'Ð4%Ë8&È:&Ë:'Î;)Ð:)Ñ:)Ò9+Ô8+Ð4(Ö<2Ó:4Ó;6äOKìXV×EEÀ3,À;,¿<(¾;'¿:'¿:)À8(À8*¿7)Ä>2Æ@5ÏK?ÕUJÑSGÃG=¿C9ÆF;ÖKDØIAÝLGãROåSSáOPâMQãNTòXdöZhú[mýZoüVnùQkúPkÿPoüHkÿQtÿ[{ÿ[xÿWqüTküVjþ[lúYh÷Ygú`ló`jë^fîgnêgmÉ]] ]U{OD˜md¿˜‘ÛºµïÓÐòàÞýõóùù÷ùÿû÷ÿþöÿýúþýÿÿÿÿüþþøøúùôúúòûøñýøòÿùöÿ÷õøïêîãßåÚÔÛÑÈÜÒÆìåÕýöãÿüæÿþçÿÿì÷ýóéðéÝãßHD8GC7FB6D@4B>2@<0?;0>:/:6+95*84+95,:6-<8/>93?:4B>5GA5KC6NE4QF4SF3UH7VJ>WMKVOW^[pmp“…¹‹•Ó” è›©ô¦°ø¨²÷¬´ü®·ü°¸ÿ°¸ÿ±¸ÿ±¸ÿ®²ûª®÷¤¦ð�Ÿé”–ጋ×�€Ì|zÅtm³mf©e_›^[�VWƒPPrFF`A>Q<6BF9@S<B_?DlCGNR•^d©fw½k�Èm˜Ëo˜Èh�È`ƒÈVwÆJfÄ<TÆ3FØ@OÞDLÑ;<Å8/Ã@.»A*°9!«6"«6$®5$°5%µ8&»9)À;,Å;.ÐA3Ô@4Ø@5Ü?6Þ=3Ý90Ü5-Ø4+Ò/(Ð-&Ó.(Ø4+à91â;3ß80Û4+Ü5,Ï+!Ì*×7)ßB3Ú>/Ð4%É.çE6óI<úH>ù<6ó,)ð$%ô&(ù//ï-+é0-ã2,Ý2+Ø1)Ô0'Ó/&Ò0%Ó7*Ò9)Ô;+Ö=-Ú>/Û?0Û?2Û?2Ñ5)æLBãJBÍ50ßGDðZYÞJJÓB?Ç;.Ã<)Â:*Á9)Â:*Á9)À8*½8)Á;/Á=0ÈF8ÎNAÍOAÃH9¾C4ÀB4ÎH=ÏF<ÔIBÚQIÜQNÙNK×KLÛLPêW_ò\güaoÿ`rúXmôNfôLeüMjÿOrÿRvÿVuÿToûUkü[mÿesÿiwÿfsúbnôblìeláfi×ijËge´g_�_ReH:~^S©ˆÊ«¦äÈÅíÙØýòðø÷õøýù÷ÿýöÿüúþýÿþÿÿýÿÿúúûúõúûóúúòýúóÿüøÿüöýöð÷îçêà×ÝÕÊÜÔÇëæÓýøäÿÿæÿþåþÿê÷ûíéðèàåÞKH9JG8IF7GD5EB3C@1B>2A=1<8-<8-;7.:6-;60<71>93?:4B>5EA6JD6MF4PE1RF0TG4UH8XMGVMRYVgkl‹}�°ˆ’Í‘žâ™¥í¦®÷§¯÷ª°ú¬³û®´þ®´þ¯³ý¯³ýª®ø¨¬ö£§ñŸ£í™�璔ߊŒ×††Ð�€Æyv¹llªbeœ]a‘WZƒRQqNI`WFY_CQmBLFM–MT¬U]»ZaÉYgÙRpâRußTsÙPjÒJ`ËBTÆ8HÂ.<Ô8EãCKãGJÖ>;Ë>5ËG8ÂH3¶?)¶=,¶=,º=+½>-Â@0ÇB3ÎE5ÓE7×E8ÙB7Ø>2Ù;0Ù6-Ø4+Ù2)×3)Ô1(Õ4*Ù6-Û8/Ü8/Û7-Ü5,Ü5,Ó/%Ø6)×7)Ï2#Í2 Ñ:'Ô=*Ò;(Õ<*æD7öJ@üE?ù64õ*-õ),ö.0ö65í53â3.Ú1,×0*Õ0*Ø/*Ù0+ã80á90â:/â<0à=.ß=.Û>/Ú>1Ð6*ïWLøaXÐ;5Ð;7äPNÛGGäPNÍ=4È:,Æ8,Ä8+Ä:-Ä<.Á<-¾:+¼:*»<+ÀA0ÅG8ÆH9ÂD5½?0¼>/ÉG:ÈF9ÍH?ÓPFÖQLÓNIÒLKÕMOçZ`ï]güfrÿgwü]oñOdòLbýOjÿVwÿSuÿSp÷Skõ\nùhuûkuöfpùeqñ_jë_héjqÜrtÁie©e\œl^~gUYI9kXJ”|r¼Ÿ›ÝÁÀëÖÕýííûõõûûùùÿýøþþüýÿÿþÿÿüÿÿùüþÿùúýôùüóýþöÿÿøÿþøÿúôÿøððçÞãÛÎàÙÉîéÕÿúäÿÿæÿÿãýÿçõùêëñåæéÞNK:NK:LI:JG8HE6FC4EA5D@4@<1?;0=90<8/<71=82>93?:4B?8DB6HE6MF4NF1PE/UF1UH7YLCTKLXTbhi…|�«ˆ’Ç’�Ý—¤ê¥­ö§­ù«®û«±ý®±þ­±û­±û¬°ú¨¬ö¦ªô£§ð¡¥îž¢ëšžç•™â‘•Ü�’׆‰Ìz€¾rx²mt¨koœigŒi^|x_u€WiŽP_¡PYºR[ÍT]ÕOVÛHRå@Qç@QáCQÜENÔCHÎ?AÉ99È44ãIIçIJßA@Ó84Ì92ÎA7ÇC6½?0ÂD5ÂD5ÄE4ÇE5ËF7ÐF9ÕG9×H:Õ>3Ô<1Ô8,Õ4*Õ2)×3*Û4+Ú6,Ú7.Ü;1ß<3ß<3Ü8.Ù5+Ù5+Ú6,Ñ/"Û;-Ú=.Í1"Å.Í8$Ð=)Ì9%Ì9'Ò9)à:.ñ>7ÿ@?ÿ<>ü46ñ-/é++ß+*×*&Ó+(×/,Ü41ä88è88ï75ð74í81è7/â6*Ü6(Ö4'Ñ5&Í5'æODÿlbÚG@Æ50ÔBBÑ?@êVV×B<Ñ=3Ê7-Æ6+Æ:-Å=/Â=.¿=-¹:)º=+½@.¿B0¿B0½@.»>,º=+ÂD5ÂD6ÇI=ÎPDÑRIÏPIÏNIÓOMç_aíbgúhrÿnzÿeuôUgôQfþUlÿXuÿWsúXoñ^nómvõv}ëmqÜY_ïbkïamëblîrz뇉Ì�|¨tg–ygsmWON:_VGƒqg®•‘Û¿¾íÕÕûèêþôõüúûüýÿùýþýýÿÿýÿÿûÿÿùþÿÿûùÿõøþôýÿ÷ÿÿøÿÿøÿþõÿþôøñçìæØèâÒòðÛþüåÿÿãþÿàþÿæøûèðôåîðãQN=PM<NK:MJ9KH9IF7GC7FB6C?4B>3@<3>:1>93>93?:4?<5@@8CC9IE9LF6NF1PE/UF/UH5WK?QIFUR[hh€}�§ˆ’Ñ�Ù—¡æ¢¨ô¥¨ù¨©û©¬ý¬­ý«®û«®ûª®ø§«õ¥©ò¡¨î §í §ëŸ¦ê�¤è›¢æ–žß‘™ÚŠ”Ñ…�È‚Œ¿�ˆ´‚€§‡z˜�m…˜cw¦XhºR]ÏMWßHQç@Hé<@è<:ã?6ÜD7ÕH7ÌI5ÅG1ÂD.Ä?.âTHßH?Ø;6Õ63×85Ö;9Ñ96Æ71¾:-»<+½;+¾:+À8*Â6)Ã4&Ä2%Í5(Î4(Ñ3(Ó2(Ö4)Ø4*Ú6,Ú8-Þ=3Û=2Û;/Ú:.Ú8-Ø6+×5(Õ3&Õ3&Ó3%Ï4"Ï6$Ñ<(Ð?*Ç9!¿1Ç9%Ê4%Ó1&ä6/ù;;ÿ>@ÿ8;õ26ê./ß--Ù-+×/.Ü43â88è:<ï8:ô23÷12ò0.í2-ä2(Ý1%Õ2#Î2#È2#Ï=.új_ãTLÇ82Ð@?Ë;;ãQQâJEÛB<Ð92È5-Æ8.Å;0Á=.¾<,¹<*º?-»A,»>*¹<(º;(»<)¼=*º?/º?0¿C7ÇK?ËOEÊNFÌMGÎNKæabéae÷gpÿnzÿix÷ZköUgýZoþZrù]rëapäkré}}æxw×_aðhrünzñerêlwö‘•ì¢ŸÃ•ˆ¡ŒyhlUHR:WVDth\¢�ˆÙ¿¾ðÖ×÷ãåÿôöÿúüþþÿûüÿýüÿÿýÿÿúÿÿøýþÿû÷ÿô÷ýóüÿöÿÿøýýõÿýôÿÿôþúïóðáïìÙøößÿýäýýáýþßÿÿåýÿê÷úéö÷éTN>SM=RL<PJ:NH:LF8JD6JD8GA5E?3C<2B;1A:2A:2B;3@<3AA9CC9HF9LF6NF1QF0SG/UH5SI=PHETQZgh}}�§‹“Ä’œ×˜¡æž¡ð¡¢õ¤¤ú¦¦ü§©ü¨ªý§¬ý¦¬ø¤ªô¢©ï §ë¢§é£¨ê¤©ë¥¨í¡¨î– å“ ä’ Ý“žÖ”œË˜–¾ž�¯¨„ £e|³\mÂRbÏITÛAKã>Dê;@ë;;ê?7â@1ÚE1ÓH1ÉI0ÂF,¾B*Â?+áSEÜC;Ü:7ã==éCEåABÙ9;Ì65ËB8ÇE7ÇC6Æ@4Æ>2Ã9,Á5(À2$Å3&É2'Í3)Ó5*×4+Ù5,Ù5,Ù5,äB7Ý=1Ø6)Ù5)Ü8,Ý9-Ù6'Ó1"Ø8(Í2 Ê1Î9%Ì9%Å4É8#ÑC/Å6%Ë8(Õ7+Ü5,å4.í53ô;9ù@>ÿLKúHFðBAë?=é;<ç7:ä15å+0ó+.ö*-ó-,ï/,è1)à4(Ù6'Ó8&Æ0½.òdVê]SÎA:ÓEAÉ;9ßMMèSOßGBÑ<6É6/Å7-Æ9/Â:,¾;)½<)¿@-¾A-º=)·8%·8%»<)¼?+³8(²8)¸>1ÁG:ÅK>ÅI?ÅI?ÊKDä`\å_^ðehþouýjtó]iðXeø]k÷^póbqçeoàloë}÷ŽŠó†ƒèrr÷qxÿ{„ôaqÞWeô‰‘þ´³Ï¯¢�›„]lOCS8QT?jbU›ˆ�ØÀ¾ðØØõßáÿô÷ÿúýÿýÿüûÿýüÿÿþÿÿúÿýøüùÿúòÿññýñöÿóúÿöùüñýýóÿÿôÿÿóùöç÷òßüúåÿþåýüàþýßÿÿçÿÿñûýðüüðWP@VO?UN>TM=SL<QJ:PI9PH;KC6KC6IA4H@3G?2F>1F>1F@4EB9DD8HF7MH5PJ4RJ3UJ4SK8VNCRMJXU`hh€z}¦ˆŽÂ�˜Ö—žä£¤ô¤¡ø¢¢ú¡¢ü ¤ÿ¡§ÿ¢ªÿ¡¬û¡«ó ©î ¨é¤§êª©í­ªñ¯¨õª©÷ž©ø•§ñ’¡â•ŸÔ¥£Ë³ž½¶…›¶hx¼LZÐDOÞ>Jå;Dæ;Aã=?ß??Þ@=á=>ß=;ÜD?É:2ÊD9ÉE9º1'ÚKCåJFèDCè>Aè;?å;>à<=Ú<;Ò=9ÎA:Å?4ÈD8ÑMAÌJ=¾<.¸6(¼8+¼6*À6)Ç4*Ë2*Î/)Ó0)Ù2,Ü5-åA5Ý:+Ù2"Ý3$ã9(æ<+Ý8%Ó2Î3Ê4Ç5Æ5 Ç4"È5%Ì6'Î6(É1$Ì2&Ï5)Ð7)Ñ5(Ô4&×3'Û3&à2)å4.ë52ë33è.1æ,1ç,3ì-4õ.3ó+-î,*ë0+è6,â8+Ø7%Ð5#Ã.Á3æYHèZLÊ<2ÝNFÃ4.ØIEàPOÚKGÒC?Ë<6Æ7/Å7+Ç9+È<+Å<*À;(½:&»:%º;(º=)·=(µ<)²8)´<.¸>/¹?0»@1½?1À@3Ã@6îkaôohéb^ômjúppë`côhkÿqwÿmxùhuîemèchêehímlîqoîqmúxxÿ}ƒ÷[iûdwÝcpø¬®³«˜z‘t[sQIZ8LR8pnY£•ˆÉ¶¯êÕÒÿîíÿö÷ÿ÷øüøùûûýýþÿþÿÿþýÿùÿýíÿóåþéäûéíÿíøÿöûÿôýÿòÿýðÿýîÿûëÿúèÿøåþøâýùàýùàûúåþþôþÿúÿÿúXO@XO@WN?VM>TK<SJ;RI:RI:NE6MD5LC4JA2I@1I@1I@1H@3FD8GE9JG8NI6QK5SK4SK4SK8VNCSNKXU`hh‚z|£…‹¿�•Ó“šÞŸžêŸ�ð��ó›Ÿø�¢ýž¥ýŸ©þ¡¬ûŸ«óŸªìŸ§æ¡¥å¥¤è¨¢ê¨ í£ ïž§ö™§î™¡ß¢›Î±˜¿½Œ©Âq„ÂTaÑDMá=Dê9?í8=é;=ã=?ÜA?ÛAAÝ<AÚ9?ÚACË;:ÌB?Ç@<À2.äPNäDFè?Dé:?æ7<à7:Û89Ó97Î:6ÒE>Æ=5Ã=4ËI<ÌL?ÁC5¶;,¶8)¹7)¼6*Ä6,É5+Ð3,Ö3.Þ52â92â>2Þ;,Ý6&ß5&á6%à6%Ú5"Ó4Ï9!Ê: Å9"Å:%É<*Ê;+Ë7+Ì4)Ò4+Ó5,Ñ7+Í7(È5#Ç4 Ê3 Ï4 Ø2$Ý1%ã0)æ/+æ,-å-/æ-2ç.3é+-ë--é0-å2+à4(Û5'Ô7$Ð9$Ò?+¸*ÓD3Î?1ØE;êWO×B>ÖB@åUTàRPÚKGÒC=Í=4É9.Ç8*Æ9(Ç<)Ã:'¿;&½<'¼=*º?-·@,´?-¯9+±;-¶>0¸>/¹>/¼>/ÀA2ÂC4ëi\ñndåbZðkfõolça`ðjiütxÿq{ÿoyöirë^déZ^îbcöpmûzuøvtÿw{øSdüZoâ]nõ¦©§¦’iŒkZtON`:PY:ttZ§ž�ÔÄ·óàÚÿóïÿúùÿùøþúùûûùúþýùÿýùýüóÿ÷êÿòàÿèàüåèÿêòÿð÷ÿñüÿñÿÿðÿýìÿüìÿúèþõäùòßøòÜúôÞú÷äúúòûü÷þþüYPAYPAXO@WN?VM>UL=TK<TK<PG8OF7NE6MD5LC4LC4LC4KC6HD8HF9KH7OJ6QK5SK4VK5TL9VNASNKZWbii�y{¡„ˆ¸Š�Ê�•Õ–™Ü˜˜à˜šç—œí™ ñ›£ôž§ö ©ôŸ©î�¨æ›£á›¡Ýœ ßž�á�šáš˜á“šÞ™šÚ¢•Í®Œ¾¾�ªÆq�ÆYpÆCQÕBLß>Cã=?å==å<?â>?á>Aß>Cà=BÖ7;Õ?@Î@>ÉB>Á<7Ã83éXUßABä>Bå;>â8;Û78Ô65Î93Ë<4ÔG>Ä;1¼6+ÄB4ËL=ÅG8»=.µ8&¸6&¼7(Å7+Í6+Ô6-Û60å95æ=6Ö4)×5(Ù5)Ü6*Ü6(Ú7(Ú;(Õ>)Ã5½5¸6 ¸7"¼9'¼8)¼4(¾.#á?:â=7Ü>3Ô;-Ì9'È7"È7"Ì7!Ò7%Ô2#Ö0$Ù.&Û.*Û/-Û/-Ú.,Ú-)Þ1+Þ6-Ú6*Ô4$Ï4"Ï8%Ð=)ÓB/À2æUDÏ;/Ù@8ÝB=èJIïUSëUTåTQáPKÚJBÖC;Ï<2È9+Ä7&Ã7&¿8%»8$¸9&¹<(·>+³>,°=+®:+°<-²<.²=,µ=-¸>/½B3ÂD6Û[Pçg\âbYðlgôplåa_êefòkoàS\î^hõhqòekñ`eôegùroýyuøtrÿnuòI\øRhãZl¡œˆ`]UoHPe>Va?ux[«¤‘ßÒÂþíåÿ÷ðÿüøÿýùýþøûÿúûÿýùÿýùÿûñÿöæÿíÛýâÙ÷Ýßùàèûåîûçõüêüÿìüûéÿûèÿøæùðßòê×ïçÔóëØöðàüùðýü÷ÿþúZQBYPAYPAXO@WN?VM>VM>UL=SJ;RI:QH9PG8OF7OF7PG8OH8JF:KG;NH8QJ7SK6TL5VK5TL9VL@UMJZW`jh}xxš�„±‡�Á�“Í�•Ï“”Ô“–Ù”˜á•›ç—Ÿê™¡ê›£ë›¤å˜¢ß•ž×’›Ô“™Õ’–Õ’“Ö��Ð…„½�„¶£}®¶t¢Çf�ÐWxÓGaÔ<KÜAGßABÝB@ßA>à@@ã>Bæ=Dæ=Dâ?DÔ:<Ð>>ÏEBÆA<º61Ä=7éZVÜBBâ?Bæ=@á;=Ù99Ó97Í=5ËA7ÎD:Á9-¹3'¿;,ÅF7ÃF4¼=,·8%»7(Á9)É:,Ð:,Ù80à91æ:6é>7Ø1)×3*Ú3+Ý5,Þ6-Ý7+Ú:,Õ<,Ç8'Ã:'¿<*¾<,½;-¿;/Â<3Í:3ëC@í@:ã?6Ø</Í:(Ç9%Æ:#É;#Ï<(Î7$Î1"Ð0$Ô1(Õ2+Ò1)Î0'Ñ3(Ò4)Ð7)Î6(Ë5$È7"É;%Ì>(Ê<(Î=,ó]OßG<Ñ2,Ñ/,å??ùVWÿtrÿrnÿkeô_XåQGÔB5Ä5%¼/Ä8'¿8%»8&·:(µ<)µ=,±>,­<,¬;-¬<.®=-¯;,±<+´?.»A2¿E6ÉMAÞ_Váb[ñpkõsqä_`â]`ä]cÕIRæYbógpôgmöelúkoûssüvuùqsúdmðCWóJaåXkë’–�•‚\zXQlCRi?ZhDw|\®ª‘æÛÉÿôçÿøïÿûñýýóüÿöùÿøùÿúöÿøñýóéûëØùÞÍòÑÊëÌÎëÍ×îÒÝïÕèòÚñöàóõàùöãûôâôëÚêáÐèÜÌíáÑñèÙüõíü÷ñþùó\P@[O?[O?ZN>YM=YM=XL<XL<VJ:VJ:UI9TH8TH8TH8UI9SJ;MG;LI:PJ:RK8UM8WL6WL6UM:UK?TMG[V\hevtt�|~¥ƒ‡¶‰�Á��ÃŽ‘Æ�’Í�”Ô’•Ú“—Þ”˜ß“šÞ•�Ü‘šÕ‹”ˇ‘ƃŒÃ�‡Á}ƒÁ~·}tŸŽp”¦h�¿]„ÕNváAcæ:Tè9HæAGàECÜGAÛHAßECæAEî<Hí=HãBGÔ>?Î@>ÎID¿?6²3*ÇB;àUPÚEAàBAâ?@à=>Ù>:ÔA:ÎD:ÌF;Å=1¿7)»6'¼9'¿=-½>+½<)¼;&À;*Ç;*Î<-Ó;-Û8/à8/å82ç:4æ95ã62á2/ä2.ç40ã5.Û3*Ò/&Ë1'É7*Ç:0Â8.À3,Â40Ë;:Û@>é9;ì63ã60×5*Ë5$Ä6"Ã8!Å:#É=&É8#Ê4#Í6%Ò9+Ô:.Ó9-Ï9+Ï=.È;*Å8'Ã8%Å:'Å='Ç<'Ç:&Í>-Ì9)×=1àB9Ô/+ã99ß12ß56Ô52Ö;6ÝD>äKCèQHèTHçUFãVEÉ@0Ã>-¾<,·<*¶>-³@.°?-®?.ª<-«=.¬=,«<+¬;+¯>.¶@2ºD6ÅLAÜbWàd\ïpjóqqâ`bå`eæ_föjuûoxöjsë`eð_fýlqÿyyÿzzúnq÷]iò@VôC]èUhäˆ�£”�f€]YtI\vGgxN�ˆf²±•çáËÿúçÿüìÿþïüÿñûÿô÷ÿôôÿñëüêãñàØî×Åèǻ伹߸½ßºÅà¿ËãÃØæÌáëÒëïØòñÜ÷òßóêÙëßÏçÙÌêÜÏîâÖñçÞñèáòéâ[O?[O?ZN>ZN>YM=YM=YM=XL<XL<WK;VJ:VJ:VJ:VJ:WK;UL=OI;PJ<RK;UM:XM9XM7XM7XM9UL=TMEXTUc_mmk�st“{}¤ƒ†±†‰´†Š·‰‹¾‹�ÆŒ�ÎŒ�ÒŒ�Ô��ÓŽ”Ò‰�É‚‰¿{‚¶u|²ov­io©lj›€lˆ“cy­ZtÉPoâCcï6Uö4L÷8Gí@BäEAÜIAÚJBßGDçBHñ<Kï=KâCGÔBBËA>ÍID¸<2®4)ÉI@ÔKEÖB>Ø><Ú<;Ú<;Ø@;ÓD<ÌH<ÊH:¿7)À8(¾9(½8'»8&º9$¼;&À;(Ä;)É<*Ð=-×;,Û7-Þ6+ã5.ç60è21ç//é-.í//ó55ó76í53ä2.×.)Ö3.Õ62Ñ30Î.0Ò/4Û6=é<Bí06î02å2.Ú3*Ð4'È7$Æ9%Ç<%Ä7#È7$Í7&Ò;*Ù=0Ü@3ÝA5ØB4ÒF5ÉB/À;(¾9&À;(Ä=)Ä;(Å8&Ì;*Í7(Ì0$æC<Ú1.í?>â/2Ù)+Ô1,Ò5.Ô70Ó9/Ñ9.Ì5*Ä2#¾1 ÍD4ÆA2»<+´9)®9(­:(«:*©:)ª<-«=.«=,«=,«<+¬=,°?/µA2ÉSGßf]ßd]èkgînoåbgíjpônuùpzüs}ömuîcjöelÿsxÿx|ÿquúipõWeø@Xó?ZèPeá~ƒ«–ƒxŽhmˆ[nŠYx‹^Œ˜p³·–ÞÝÁøóÝûùäÿÿíûÿïõÿïîÿëåûäÚñ×ÐãÍÅÞÁ°Ö­¨Ó¦¨Ñ¥®Ó¨µÕ®¼Ö±ÇÚºÑàÁàæÌçéÓïíØñêØîâÔéÛÐêÙÏëÚÒèÛÒäÚÐä×ÏZN>ZN>YM=YM=YM=XL<XL<XL<XL<WK;WK;VJ:VJ:WK;XL<VM>QI<RJ=SL<VN;XM9YN8YM7XM9WN?ULCVPP^Y`c`qigqp�yy�|¢}�§�‚²ƒ„º††Äˆ‡É‰ˆÌˆ‰Ê†ˆÅ‚„½{~³ux«pq§kl¢efžlaŒ�`q•V^°N[ÑJ[ì@V÷6Ký2Eþ9Cò>?èC=ßHAÜIBàFFçBHñ<Kî>KÝADÔFBÇB;ÇH?³:/°7,ÎPDÈC:Î=8Î63Í2.Ï61Ñ>7ÏE;ÆG8ÂC4¾9(Á:'À;(¾9&¼9%¼:$½;%Á:&Ä9&Ë:'Ò;*×:)Ü6*ß5(ä3+è3,è**î*,õ-0û03ÿ26ÿ47ÿ47ü58ø7:õ8<ñ8=ñ8>ó8C÷7Fú5Gû5Dó*4ó,1ë20á4.Ø5,Ð8*Î;+Î=*È5%Ï7)Ø:.Þ>2á=3â>4â>4ßA6ØK:ÍH7ÉA1Ã<)Á:'Â9'Å9(Ç8(Æ2$Ò8,à?7êC=Ý1/Ý--é56æ66Ø3-Õ7.Ö8/×;/Õ;/Ñ;-Í;.È:,ÔL>ËG8¾?0´9*®6&«7(©8*¨8*«;-¬<.­>-­>-«<+ª;*¬=,®=-ÀL?ÙcYÝd\çlgîpqéhlôqwüvígp÷oyøqxójqþmtÿwÿu|ùhoùclõSbþB[ô<XéLaÜuy´›‡Œžx}™i{—d€–g�žs¬²ŽÊË«ßßÃçêÏíöÛäóÖÙíÑÍæÈÄßÀ»Öµ³Ì®ªÈ¤˜Â”“ÂŽ˜Ä‘¡É—§Ë�¬ËŸµÎ¦¾Ò­ÎÙ¹×ÝÁäãÎìåÓíáÓëÚÐæÓÌãÐÉÝÏÆØÌÀÕǾYM=YM=YM=XL<XL<XL<XL<XL<XL<WK;WK;WK;WK;WK;XL<YM=SJ;TK<UL;XM;YN:YN8ZN8YL9[O?WMCVNKZTX\Yd`]pig}ppŠvw•xxœ|{§~~²‚�½…ƒÄ†ƒÈ†ƒÆ}¼~zµyv«vr¥un¢sjŸpgžza‰„QZ˜IE²FFÒGNíCNø:Fþ6@þ:>õ==íA=æD?áFBäDFçBIî>Kë?KÕ?@ÐGAÂB9ÂH=¯9-°:.ÐTHÃ>5Ç82Æ.)Ã*%Æ.)Ë;3ËB8ÁC4¼?-¿<*Ã<)À;(¿:'¿;&Á=(À<'À8"Ä6"É6$Ñ8&Ø8(Ý5(à4(å3)é2*õ31þ68ÿ9:ÿ58ÿ-1ý(,ý(,ý*0û,2õ*0ò&/ö(4ÿ,=ÿ*?ü 9ó/õ!/ð'-ë-/ã/.Ú1,Ó2(Ð4(Ï5)Ó5*Ü:/æ>5ê=6é;4è71ç60á90ÝI=ÖJ;ÒF7Ë=/Ã5'Á4#Å6(Í9+Í5*Ó6-æC<ß82Þ20Ïë99ã62Ô1(Ñ3(Ð4(Î4(Í5(Ê6(Æ7)Ã7(áYKÖREÊH;½?1¶</´<.²<0³=1¯9-¯;,±=.°<-¬;+«:*«:*«:*°</ÐZPÛbZêpmówwëmqóqyõrzýz‚ÿˆÿy€òiq÷eoÿpyÿt|ÿltø^höO`ÿB\ö9UéG\Ùosºž‰˜¨ƒ‚žm|šd—e‹œo ¨�³¹•Ç̬ÑÛ¹ËÙ¸ÀÔ±±Ë¨¤Âœ�»•˜¶�–²Œ�±†‰´†µ~‹ºƒ•ÀŠ›Ã�žÃ�¨Å—°ÉŸ¾Î©ÉÒ³ØÚÂäßÌèßÐçÖÎàÍÇÚÇÁÓº˽´Å·®ZM<ZM<ZM<ZM<ZM<ZM<ZM<ZM<WJ9XK:XK:XK:YL;YL;YL;YM=XL>YM?YM=ZM<[N;[O9[O9\O<YM=WMAVMFWNOXQX[Ua^Zi_\ohe€mlŽroštr¤zv³‚}¿ˆ‚ÈŠƒÇ‡~½‰»‡z°ƒr¦†r§Œt¨‡l£†\‚�HL¢B6´A:ÊC?Ý?>è:;ô<<ÿDCö=;ñ=<ì>=ê@AèBFéBJê@KãBJÏ?>Ã@8»?5´>2®=-°<-¾D7ÑLCÆ7/Ó;6Ì1-Ë2,ÓC:Ç?3³5&»>,¹6"À9%Ã<(Á:&À9%À9%Ã;%Æ;&Ë:'Ð9&Ô7&Õ2#Ø.Þ1#ë6+ô>3ô71õ4/÷0+÷,(ù'&ú&%û%'û%'û&*õ"(ø",ÿ'5ÿ(;ÿ"9ÿ4ý/ÿ'8ï#,ä#(ã+-Ü,,Ó*'Ò-)Ü5/á51ç51î64ñ54ò21ï/.í-,ä.*Û:2Ð9.Ð90êSJº#Â+ ÞE=È.$Ë.'Ü>5çD=à;5Ö/)Ù-)â51ä;4Ü>3Õ<.Í5(É3%Ê6*Í;.É;/Å9,êbVßYMÇC7¹7*º:/·9-±3'³5)¯1%¯1#®3$®4%¯7'¯:)±<+²>/°:.¹C9Ö]Uêpmðtvõw{úx€óq{þ|„ûx€÷tzût{ÿs}ÿr}ÿmxþhsü`mòJ[ÿ>Yü>Zã@UÓdj´•€’¡z��j{™c}•c…–i“ už¦ž¨ƒ™§ƒ’£‰¢{€Ÿuxœpw›oy�o s€¥r‚­uƒ³w‰¹}�¼�’½…—¾‡œ¾‹ ½�¦»’®¼™½Ã§Î͸ÝÔÅàÒÉÜÇÂÓ¾»Ê»´¿²ª¸«£YL;YL;YL;YL;YL;YL;YL;YL;WJ9XK:XK:XK:YL;YL;YL;ZM<YL<ZM=ZM<ZM<[N;[O9[O9\O<[P>[OAXNEXOJYPQ\SX]V]^Xda\sgb‚lg�njœuo«zº†É€Ã��¿�}·�u©‹mŸ—q¤ t§šjœ˜Vz�=A§5*°6)Á=1Ð@7Ü?6ç>7ò@<ô<:ó;;ò;?ð=Aí@FèAHá@HØAFÅ;8¼<3´;0¯;.ª9)®:+¼B5ÎJ>×G?ÛB=Ô72Ð50ÔA9ÌB7¹:+±4"»8$Á:&Ä=)Ã<(Ã<(Ä;(Æ;&È:&Î8'Î5#Ó3#Ù3%á5'é9,î9.ï9.å0%æ0%ê.#î,#ô+$ø+&ý+*ÿ-.ú),÷$*ù#-ÿ&4ÿ%9ÿ5ÿ0ü.õ.ö.9ô2;ä+1Ú'+Ü.0Þ01Û+-ð9;ó6:õ47õ13ö/2÷/1ø23ó76Ú1,Ú<3Ö71ãD>âE>Í0)Ï2-Á"Ó4.Þ?9æD?à=6Ö3,Ö2)Ý6.Ü90Ò:,Ì9)È6'Ç5&È9+Ê<0È</Ä:/ìbWàXLÈ?5¹3(À:/Á;0¼6+Á8.¾4*¾4)½5)º4(·5'´6'±6&¯7'´<.»B7ÓZRçmjíqsõx|û|ƒ÷uþ|„úx€ùv|üu|ÿs~ÿp{ÿjuüdpû_lôI\ÿ=Yÿ>[éCYÒdg­�vˆ—nx”at’\vŽ\}‘`ˆ•iŒ™nŠ–nƒ“lp‡]kˆ\h‰\fŒ[j�_p–cv›hy i�­r‚²t‡·y�»}�½‘½€•¼ƒš½‡š´…¡³‹¬·—¼¾¦ÍÆ´ÕǼÔÁ»Î»·Ä¶³¸­©°¥£YL;YL;YL;YL;YL;YL;YL;YL;WJ9XK:XK:XK:YL;YL;YL;ZM<YL<\L=\L<\L<]N;]N;[O9\O<_RA]QA\PDYOEYNJYOM[QP[QY\SfbYvf^…jb‘pk£}w³†}¾ˆ~»—…¿˜€´šv¦žpž­t¡ºw£´h–°Rt¸<D»5,¹5)¿;.ÈC4ÒD8ÚB7ä?9ð><õ;>÷:@õ<Bï@Gå@FÙ@CÎ@?¿:3¶:.¯9+«:*¨9(¬8)»@1ÍG<×D=Õ:5×84Ù<7ÛD;ÝOCÍI:³4#¾9&À9&Â;(Å>+Ç>+È=*É:)Ê9&Ï8'Ï2#Ó0!ß5(ë=2ò@6ð91è3*Þ4%Þ4%á1"å/"é-$í,%ð*'ñ+(ò,+î'*ò&/÷'4ü%7û4û2÷1ó$6ñ/:ñ6?ð8@ï;Dí<Dé6<â-4ï29ò09ó.7õ.5ô-2õ.1ó/0ì31Û/+à=8Ù53ß;9ÿliäB?Ñ/-Î/+Ù:6Þ?9àA;Ú<3Õ7.Ó5*Ó5*Ï6(Å6%Â7$Â6%Â9)Ä:-Ä<.Å;0Å;1ê`VÞTJÈ;2¿0(É:2Ï?6Î;3Ñ=3Ö=5×=5Ô=4Ï<2É;/Á9+»7(µ7(·;/»A6ÏTMãhcëooöy}ÿ€‡ú{‚û|ƒûy�ûx€þw~ÿtÿn{þfsú_mù[jôG[ÿ;Wÿ>YìFZÍ_`�ev…Zh„Qf„Nl„Rs‡Vy‰\{ˆ]u„[o�W`zM_P`„Te�Yl”_t›dzŸi|£j�­pƒ±q‡µuŠ¸x‹¹yŒºzŽº}’¹€’°|”«}š¨„¨¬‘¹´ Å¹«Ê¹±È¶´¶¬«§¡¡�—™YL;YL;YL;YL;YL;YL;YL;YL;WJ9XK:XK:XK:YL;YL;YL;\L<[K<\L<\L<\L<]N;]N;[O9\P:^Q>]P?[O?ZN@ZMD[NF\OIZNN^P_`TldY{f^‡oh›zu­ƒ{´…z²˜„·¤„³¯€ª·y Æy�Ðt™Êd‹ÆNjÙEQÝB@ÓC;ÍD:ËG:ÌF:ÓE9ÜB:ì@>ô=Aù<C÷>FïBFâBDÒ@@Æ?;¾>5µ;.®:+«:*©:)­9*¼>0ÌC9Ô=6Ó40ß;9åA?ßE=çTJÞVH¾:+¿:)¾7$¾7$Ä;)É=,É<*Ë:)Î8)Í4&Ñ1#Ø2&ä8,ð?5ó>5í5-ã.%ß5(Ü6(ß3'ã1'æ.&é,(ë+(ë+*ê)*è',ê&0ò*7õ';ö"9ú ;ú$>û6Gâ(5å/;ÿP[ÿ`lÿO[ñ=Hñ9Eè*6í+6ð.7ó19ð16é/0á+(Ö)#Õ0*Ø61Ó1.äB?ÿspõVSÜ=:åGDÚ<9Ù<7Ô:2Ò8.Ð8-Ï7*Ì4&Ä3"¼5!¹6"»8&¾:+¿;,¾:-À:/Ä;1ãYOÚMDÊ70Å0)Õ<6ÜB:Ù<5Þ;4æ=8ç>7ä@7à?5Ø>2Ï=0Ç;,¿:+¹9,º<0ËOGßd_ëooø|~ÿ„ˆü}„÷x÷xúx€ÿx‚ÿtÿmzýerù^löUeòEYý8Tþ<WíG[ÅYYŠlRbsGZuB[xBbzHkNq�Ts‚Wq‚Vl�Vk†Yk�[n”atœg{¤l¨n‚©pƒ«o„®o„°qˆµtˆ¶u†·u†·uŠ¸wŽ¸y�³{�«z‘£{š£„««“¼³¢Å·®Å·´±«­ ž£•’™ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8YJ7ZK8ZK8ZK8[L9[L9[L9\M:\K;]L<]L<]L:]N;]N;]N9^O:\O<[N;[N=[N=\O?]OB]OB^OJ`OYbQdaTncY{jeŽur¡~x¨�u£”}§«…ªÂ‰©Î‚žÙv’Þl†ÚYuÙE]òCTùEPïJNÞHGÏF>ÊD;ÏE;ÙD>çCBïACô?DõBHëEGÞDBÍB=Â?7¼@4µ=/¬;+ª;*©:'­8'º:-Ë>5àE@à<:ëB?éA>Ü:5ãJBæXLÓK=È@0½8'º2"À7%Ç;*É:*Ë7)Ï7)Í1$Õ3(ß7.é;2í:3í60ê/*ä+&á0*à1,ç1.ë20ò12õ26ö37ö5:ë,4ê,6ì-<ð/@ó-Dô)Fü+Kÿ4Qó8Iä2>ïAKÿ`jÿkwÿWdôDQô?Ní4Bì4@ì4>ç6<â88Ú72Í3)Æ/$Ê3*Å.%Ï81ñZSö]XñWUÛA?àHEÛC@Ó;6Í6/Í6+Í9-Í9+É6&Â5#¶8"±8#³:%¸=+¶;+µ7)»9,Ä>5ÙOE×G?Í6/Î1,ß>6åA8á81å61ê1/ë2/ë51æ81à90Ø:/Î:,Ç;,¼6*º8+ÈI@Ü_Yênlù}}ÿ†ˆû~‚óv|öw~úx€ÿx‚ÿs€ÿl{üdsù^nñRdóE\ý8Tþ=XîK\¾URz_BVg:Wr?[vCdzIl~Nr‚Uu„Yx‰]w�_w”dušgx k{¦n«p‚«oƒ«o„ªm…­o‡¯q‡´s‡µt„¶q…·r†¸s‹¸wŽ¶zŒ­x�¤v–¤€ª­’¾¹¥ÍÁµÏÄÀ»¹¾¨«´™›§ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8YJ7ZK8ZK8ZK8[L9[L9[L9\M:]J;^K<]L:]L:^M;]N;]N9^O:ZM:[N;[N;\O<_P=`Q>`P@`OEbNPbNZ_Pc_Upga…qo–ytœ|r—�u•­‚�Ê‹ Ú�“ãsƒêctçPcè=Pø4Jÿ;PûERèCJÔ@@Ë@=ÏD?ÖE@âDCèBDîAEíCFåEEÙD@ÊA9¿?4¹?2°:,©8(¨9(¨9&­8'»9,É:2çJEè@?ê<;ã75Ø3/ÞA:íZPôh[ÙQAÆA0½5%À7'Ç9+Ç8(Ê6(Ð8+Ï1&Ù5+å:2é;4é3/ç.)è-(ë.,ò46õ49ø3<û2<ý1<þ0=ü0<ù1>ô/@ò1Bó3Hô4Mõ0Nô-Nþ2Wÿ=^å2EøNYÿakÿ_kþVcùQ`òFVç8Gñ@Pè9FÛ29Ð02È3-Â8-½;+º;(Â:,¿3&ØKAùjbÛLFáPKÌ;6Í<7ßKGÔA:Ê7/Ê8+Í;,Í<+Ë:'Ã:'´<$«;%¬<&±>+¯:)®4%¹9.ÉD;ÔG@×D=Ñ61Ö3.å<7ê<5æ3.ì1,ñ+*ó++ñ/-ì3.è7/à90Ø</Ñ=/Â6)º4)ÃC:×XQæieø|zÿ††ü€‚ôw{öyûy�ÿx‚ÿq~ÿixúbqø]mïPbõI_ý;Vý?YðO_¹RMoX8Re7\tBa|Il‚Qt…Xz‡\|‹`€‘e€—i~›k{ l|¤o|¨m~ªm�«l‚ªlƒ©l…«n†®p‡±q…²o„³o‚´m„¶o†¸sŠ´u‡­t‹¥u˜¨�¯¶—ÅîÖ;ÚÑÌÁÁÉ«±¿š ®ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8YJ7ZK8ZK8ZK8[L9[L9[L9\M:]J9^K:^K:]L:^M;^M;]N9^O:^O:^O:_P;`Q<aP<aP<aP<`N@cNKcNU`P]^Ujgc~rq‘zw–}sŽŽvŒ­�’̆‘ß~‡îqyøbk÷PZ÷;Jþ)Eÿ2Nÿ;Rñ<MÞ=EÕ@DÓCCÓAAÜCEàBCáAAáCBÜD?ÒC;Æ@5¾>1·=0­9*§8%©:'«<)³;+À</Î>5ß=:ä88ã11ß//Þ52Û94åNCûm_ñgZ×RCÅ=/Ã:*È:,É7*Ì5*Ó9-Õ4*Û7.å:3è71è1-ç.+ì0.õ13ÿ3Aÿ0Bÿ-Cÿ*Aÿ&>ÿ":ü 9ö 8÷'Aô*Dõ.Kõ/Pó+Pô)Tÿ2`ÿBhñDWÿ_jÿhr÷WcìJYðK[êCTØ1Bè@OÛ7BÇ,0¸('¯/$«8%¥?&¨@'·@,¹:)ÜZLà\PÌF=ÌC;À71Æ;6âUNÔE=È:0Ç8*É:*È9(Ç9%Ä='°>&¦<$¦<&ª=)©8&«3%»<3ÑLCÔG@ÚE?Ø95Ú3-ç83ë51ê/(ó0,ü,,ý+,ú.-ô1-í5-æ8/Þ<1×?2É9.¼4(¿<2ÎOFßb\öywÿ‡‡ÿƒ…ø{ù|‚ý{ƒÿw�ÿn~ýdvö]oôYkëL^ôH^û9Tú<VëM[±KFfR1Oe4\tBf~LuˆZ}Žaƒ�e…’g‡˜l‡žp‡¤t‚§s§r|¨m}©l©j�©kƒªk…©l„ªm…­n‚¯l�°l€²k�µm„¸p‡´s†®rŒ¨wš¬„²¹šÇÅ®ÖͼØÑË»¿Ê¤¬¿’š­YJ7YJ7YJ7YJ7YJ7YJ7YJ7YJ7YJ7ZK8ZK8ZK8[L9[L9[L9]L:\K9^K:]L<]L:^M;]N;]N;\P:^R<`Q<aR=aR=aP<`O;_N:`M>gNIgNRbQ[`Whiguv”{|›y•�{”©‚•Æ„�Ú}…ïtwýklýYZûCKü1Gÿ4Qÿ<Tó<Nç?LãCKÝBFÖ=?ÛACÛ?@Û@>ÛA?ÖC<ÏB9Å?4¼>0´>0«:*¨9&«<)³@.»@1ÉA5×B;Ø63à74Þ/,á2/ä;4Ó0'Î5'éUGÿwié[OÐB6Ê<.Ê;-É7*Ê6*Ô:.Ù8.Þ7/ã62æ21é/0ì01ò46û4;ý&6ÿ"6ÿ"9ÿ#7ÿ 5ÿ5ý3ø4ï2í8ï$Aï(Gí%Ið(Mý3[ÿFhÿbsý`iôW`ðPZðP\îLYá?LÒ2<Ù>DË68º0-®/&¥6%¡=&›?&œ>$¨9&²8)ÑUIº:/Ã>7»4.»2,ÍB=ãXQÕH?Æ9/Ä6*Ä7&Ã4#Å4!Â;'±<(¦9$§8%ª9'¬4$­1%Â?5ÛRJØIAßHAÝ;6Û2-ç40ì1,î-(ù2/ü,,ü*-ø*,ó,)ë.(ã1'Ü4)Õ7+Ð>1½3&»9,ÇH?Ø]Vñwtÿ‰ˆÿ†ˆû€…ý€†ÿ}…ÿv�ÿj|þ_s÷XlõShêDZôC]õ7Sñ;TßMZ¢H@\L+Lb3To<a|GtŒZ€–e†™k‡šlˆ p‰¥t‹«y…¬w€©q{§l|¦g|¦f¦g¦g€¦i�©k�«l€­l~­g~°g€´jƒµlˆ·qˆ°q‹¬w™°„¬º—¿Å©ËʵÊÌÁ´½Äžªº‹—§[J6[J6[J6[J6[J6[J6[J6[J6\K7\K7\K7\K7\K7\K7\K7\K9[J:\K;\L=ZM<YL;YN:\Q=^S?^S?_R?^Q>aR?cRBbQAbO@bL?iMIfLMdPYf[lkk…v{›€…«‡‰°Ž„§¢‡¦¼ˆŸÎ‚�ây~ðsoùkaÿ`\öKTð=Pé:Ië?MíCPç@Jä=Eã>Dá<BÜ<>Ú<;Õ=:Ï?7È>3À</¶;+§6$­B.«<)«8&½B3ÊF:Ì>4Ï60Ù72Û60Ý6.Ü6*Ü9*Ü;)×;%Ó6%Ù;0ïRIñZO×E8Æ7)Ì=-ÓA2Ò:-Ô3)ðGBà..î5:ò5;î,5ÿALï'2ù&/ÿ(0ÿ'1ù&û"*ÿ&.þ'.ð$î)÷+6é!.ì&7ñ-Aä":ÿWqç-Dÿ^lþ^hü[cùU\óPUéHMÚ>?Í84É>7ÄB5·?/¥8$™2—7!š;'œ;(§9,®7/ÂE?¿;7¼2/Ä64Ä51ÓD@ÛLFØKBÑD:Ç9-Ã3(Ç5(É5'Ä6(º<-±9)­/!©)²,!Á9-ÐC9×G>ãPFÝD<Ù80Û0)å/+ñ10ù11û11÷/1õ/0õ/0ó0.ð0+ë3+ä6+Û9,Ô>0¾2#¾?0·=0Õ^VèrnÿŽŒø„„û‡Šÿ„Œÿx…ÿhzÿ_wÿ[uþTnøLf÷Dað9Wó:Yå?WÚ\gŠ?9UG,L^4Xu?g†Kw–]}œc|�f€¡j‚¥m�§n¦m~§m|¥i{¥f{£e|£d|£d}¤e}¥i~§m~ªk¬k~­g~®d~±d‚²f†´k‰´oŠ°uŽ®|˜³ˆ§½™±Á¤²Á®¤³°‹˜žw„Š\K7\K7\K7\K7\K7\K7\K7\K7]L8]L8]L8]L8]L8]L8]L8]L:\K;]L<^N>\O>ZO=ZO=ZR?[S@\TA]QA\P@\NA_ND`OEbPFcOHgNJfMPcS]f^snoŽv}§}‡¸…‹¿œ•É¡‹¹ª~¡¸w�Óy‚ê}zôvjõh^ø]aðNYëEQïHRöJVôFPí?Hê=Cã:?Þ:;Ù;:Ô<7Î@6Ç?3¿=/¶=,©<(©>*¨9&¯:)¿A3Ã=2Æ7/Ô=6Ò5.Õ4*×5*×5&Ø7%Ú9%Û;%Ú9%Ü5,æA;éLCàJ<Ð>/Æ7&Ê9(Ó;-Ü;1õIEä01ð6;õ6>ð.9ÿ?Mî&1ó$*ö%(ú(+ü,,û++õ''ö((÷+*î#&ì%(ê%,ý;DÙ'î4Bû@Q×%5ÿjuÿcjü[aûX]óNRåBC×<8Ñ>6½7+¾C3¹I5©>*—0�*Ž(’(œ-$°;4ÐSOÌHDÁ54Æ44Ð<:êVT×FAÕF>ÏA7Æ8.Å2(É5+Ï7,Ì8.Æ<2Â91Á4+Ã3+È3,Ï81×@7ÞE=âJ?Ý@7Ù6-Ü1*æ0,ñ10ø00ú00ô02ò01ó/0ó0.ñ/-í2+ç5+Þ8,Ö>0¾2#¼=.µ=/Ó^Uévqÿ‘�÷……÷„‡ý�‰ÿuƒÿfyÿ\vÿXtÿQnýIfû?`ó:Zñ@]áI^ÀR[{?7RF.L^6Yv@e‡Jr–Yu›^w�by¡e|¥i}¦j|¥i{¤hz¤ey£d{¢c{¢c}¡a{¢cz£gz¦k|¨i}ªg|«c|¬`~¯`€±b‚°e†±j‡¯p†¬s‰ª{Ž­„�©†Œ£‰v‡aooO]]\K7\K7\K7\K7\K7\K7\K7\K7]L8]L8]L8]L8]L8]L8]L8]L8]L:_N<`Q>^Q@[P>ZO=YP?ZQB[RC[QEZPG\OI]OL`QNeSSgUUgUUfW\f^mkhƒrw¡y‚¹�ŠËˆ�Ñ–‘ÔŸ‹È©‚±¹}ŸÏ�‘䄆í~uñriúegõT\íGQïENöHRöFPòAKí>Eâ7=Þ8:Ø:9Ò=7Ì?5Ä@3½@.¶>-­>+¨=+©:)·C4ÅI=¾</º1'È:0Ë7-Ð6,Ó7+Õ8)Ø6'Ø7#Ú:$Ý:)Ü5,Ù4.ßA6çOBÛG9Ç4$È2#Ø?1á?4õHBç10ò59ø6?÷3?ÿDPø3=ï(-ê$%ë#%÷10ô.-í'&ë%$ô01ê()õ79å+0à(0é4=ÿR^î<Jÿ[fý]eñVZîOTîMRèEHÜ9:Õ74Ó>8À6,À>1¼B3²>/©8*¥3(Ÿ-#›&š%§0*ÅEBÁ;8»-,Å33ÖB@ó_]Ð?:ÑB:ÐB8Ê<2Ì9/Ô=4Ú@6ÛA9Û@;Ø=9Ý>;ãA?â><Þ:8ã@;ëIDàB9Ü;3Ø5,Ú1*ä1,î20õ20ö0/ó12ò01ó/0ó0.ñ/-í2+ç5+Þ8,×?1À4%¸9*´<.Î\Rëztÿ“�ù‡‡ù‚†û~†þrÿewÿ[tÿTqÿNlÿFeú<^ó<[îEbßTg¡AEp<1QJ0Pb:]xCfˆLp”Wr˜[tš_wŸcy¢fz£gy£dx¢cw¡bw¡bxŸ`y a{Ÿay az¢fz£gz§f{¨cy©_zª^|­^~¯`„²g‰´mŠ²s†¬s„¦t€Ÿux’mo†lRdX@OL2A>]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8\L5^N7`O;`Q>\O>ZN>XO@XPEXOFZSM^TSaWXeY]gZak[ek^glaikeonl‚rtšw~´~‡Ê„‹Ù‹ŽßŒ‡×œ‹Ñ«ŠÁ³�§¼x�Ëz€Ü‚yð�vþrqÿcföU[óLSóHPóDKò>Gí<Bä5:ß77Ø96Ò=6ËA6ÄB4¼A/µ@.«:(­>-¯>.»G8ËSEÅI=¹6,·1&Æ9/Ì9/Ò<.Ù=0Ú:*Ø7%Ù6#Þ8(Ý5,Ù2*ß>4êNAãK=Ñ;*Ð7'ßB3Ý6-î@9è0.ï13õ3;û7AÿNXÿKTý>Cð45á%$ë/.é-,î21æ*)è/-ñ99Ñå26Ü-2æ9?ïFMÿX`ñMT×<@ÚACÞBEÞ>@Þ:;Ü89Ø96Ò94É91Á80¼9/ÀA8ËOGÕVPÐNLÃC@²72´92ÍLGØNKâRQîZXèRQïZVÈ5.Í=4ÑA8Ï?4Ò>4ØA8ÝC9â@;è>>ç9;ì9=ñ>Bî;>è58ë=<öMHÝ82Ù6-Ø4+Û3*â4-ë40ð3/ò0.ó12ó/1õ/0õ/.ò/-î1+é4+ß7,Ù@2Â6'´5&²:,ÈVLî}wÿ“�û‡ˆü�†üz„ûo|ûbtþXpÿOlÿFfÿ@bø9[ó>]éKd×Zh„32e;-UO5VhBa|IiˆMq“Ws—Zr˜]vžby¡ex¢cv av auŸ`vž`wž_xŸ`{Ÿay ay¡cz¤ez§dz§bx¨^yª[z«Z|­\„³e‰´l‹´rŠ­s…¥sœps�fhbPbTCRK7F?]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9^N7\L3]M3_O6^O:[N=WN?WODWPHVQN]WYg`hnfsphwpgxrfzqh{ol}oq†rw—u|ªy�À‡Ò„ŠÞŠŒã“ŽâšŒ× „¿›užŸl¬nq¼vlÑuhñrlÿljÿgiþ^`úQVõFKï<Bé6:æ69á77Û96Ô=6Í@6ÅC3¼A/·?.¬8)´@3¯;.°</ÆREÓ]QÇMB¹;/¼:-À8,É:,Õ?1Û>/Ø6'Ú4&à6)ß2+á6/â>5âD8áE8ÜC3Ý@1á?0Û0&é61ê1.ï./ð-1ù6<ÿPXÿ\bÿZ^ÿORá.1â/2Þ,,öDDç55Û++Û-,ê>>Û33×34ÿcfÞ=BÊ+0Ê.2À')Ô::àBCÚ7:Ø24ß9;Ü89Í/.Ë30Í<9ÒGDÝSQëaaõilôeké^aÂA<ÅF?ã^Y÷igÿppÿkjãKJÖ>;Ç4-Î>5ÕE<ÓC8Ó?5Õ=2×:1Û30î5:ð,6ð,6ñ-7ï,4ë,3ò:<ýKIÞ2.Ú3+Ø4+Û4+á4-ç4-ì3.ñ1.ô02õ.1ö.0ö..ô.-ð0+é4+ß7,Ø?1Æ8*´2$±9+ÀNDí|vÿ’Žþ‡‰ÿ‰þv„øk{øatûUmýIfÿ=aÿ:_ø6YñB_ÝNbÁV^p/)\?-YW>]mHe}MkŠQq’Ys–\s™^wŸcy¡cx btž_tž^u�^u�^wž_xŸ`{Ÿaz¡by¡bz¤dz§bz¨`y§\y¨Z{ªZ|«[«^ƒ®f†¯k†ªm…¦qƒ rz•lq‰idwcZj]QaT^O:^O:^O:^O:^O:^O:^O:^O:]N9]N9]N9]N9]N9]N9]N9]N7^M3_N2_O6]N9ZN>XNBYRLYSS^Y_ebmrnyt‹{u‘yr‘wp‘tp‘ps’pyšt}¨v€µy‚Ã}…΃ˆÙŠ‹Û“�Ù—‹Ë™†´ ƒ¡©ƒŽ«~yžkZšQ>ºNAÓPHáVQëWUðRQðJLí?Aå58è8:å99Þ:8Ø=8ÑA6ÈC4ÀA0º?/³9,¹@5«5)¢.!¼J?×g[ÖeWÉUH·=0µ5(¾4'Í;.Ø</Ø6)Ü4)â6,Þ0)ä70â;2Ü:-Ü?0âE4ãC3Þ8*ß1&é4-ï2.î.-ì*+ó25ÿEJÿUZÿ^`ÿ[^á8;á8;Ô,,òJJâ89×//Ú44Ì('Ê'(öXYÚ>AÔ:<Ð7:¸!Ä()Û;;æAEÝ47Ö,/Û14Ù03Ì'+Ñ34ãJLôbcùhköciõ_hù_k÷ajÔMJÓNGç\YçVSæNMáGEÎ31Í4/ÖA:ÞKCãPFÝJ@ØA6Õ;1Ô6-Ú.,ô19ú,9ø*7õ'3ô(3÷07ü>@ÿJIß0+Û2+Û4+Û4+á5+æ5-ì3.ð3/ô01ö.1ø-0ø..õ--ñ0+ê3+á6,Ù=0É;-³1#±7*¹D;êysÿ�‹ÿˆŠÿ{‡þp€õgwõ^qøRjùEbÿ9^ÿ7]ù7ZïGaÏO^¡HJd4(VD.ZZ@^lIe|NkˆRs‘[t•\wš`xžcx bu�^rœ\rœ\s›\s›\uœ]wž_zž`y ay¡bz¤b{¦_y§^x¦[x§Wy¨W{ªY}©\‚¬b…¬i†ªl‡¨qˆ¥uƒžs|”pu‰pm~kevd^O<^O<^O<^O<^O<^O<^O<^O<]N;]N;]N;]N;]N;]N;]N;]N7`O5aP4`P7\O<YPAZQJ\VV_Zaiguqo„|y–~ ~|£zw¢xu¢uv¤rx¨r~°w�¶zƒ¼}…ÈȊŽÍ’“Ì—“Å –»«œ³»¦¯É­©Ä¥“ªˆl™fIœE1¬A1¶@4ÂB9ÖIBèNLíIHê@Aê<=è<<ã;:ß=:×@7ÏA5Æ>0À</º:/»?5®5,¦1'ºJ>ÎbUÓgZÓeVÃO@¸>/¸4'Æ8,Ó9-×5*Þ3+ä6/à2+â5.ß7.Ü8,à>1åC4â?0Ý3$å5(è0(ï2.ï--í-,ñ01ò67öBEöMPûX[éFIôQT×47éFIÜ89äABÚ78Î.0Ï12âHJÑ8:Í78·"$Ô;=Û89ã;;é<@ç6<à-3Ý*0á29ç<BïLQú^bÿlqÿipù_iôZf÷Ygö\fócbè[TçVS×?<Ì1/Ò41Ñ30ÞC>âKDæRHèTJßKAØ@5×:1Ù80á32ó.7ÿ0=ÿ/<ú*7û-9ÿ:BÿBEúBBà1,Ü3,Ü5,Ý5,à4*å4,ì4,ñ40ô01ö.1ø-0ø..õ--ñ0+ê3+á6,×;.Í>0µ1$±7*²;3çtoýŠ‡ÿ†Šÿvƒüi{òarò[n÷PjùBaÿ8^ÿ8_ù;_êLe¾NZ~86]=.LF.UX=ZeCcyKh„QrŽ[v”^wš`yŸdwŸarš[o™Yo™Wp™WqšXr›Yt�[wž_xŸ`w ^x£^y¤]y¥Zv¥Wu¤Tw¦Ux§V}©\‚¬b„«h„¨j„¥n…£qšmx‘ip„hj~cbv]_P=_P=_P=_P=_P=_P=_P=_P=^O<^O<^O<^O<^O<^O<^O<^O:bR9aQ7`Q<[O?ZPG\TR_Za`^lsq‡yx—€€¦�‚®}}¯xx®wv¯vx³t|»w€Á{…ˆÁ…ŠÁŒ�¿•—¾��¹«§¸´­´¾³¯Æ¹©Ð¾¦ÕÁ Ñ»”Ьˆ¯oS®\F¤G5¨>0ÁG<ßVPïWTðNLè@@ç==æ<<â=;Ý>:Ö=5Ï;1Ç9/Á80À=5¶:2·>5ÀPDÅYLÅ^OÏeWÒdUÂL>º:-Ã9,Ð9.Ö3*Ý2+ä6/ä6/à3,Þ6-ã=1æB6åB3á:*Þ2$é4)å-#í0*ï/,ï20ð42æ.,â22æBCôX[óWZÿosãDHèGLÞ=BûZ_ÿptÝ>Bÿ}€ÿnpêQSÓ=>Å02Õ;=òHKí:>ï8=÷>Dò7@ç.6ô?Hÿ]gÿltÿjrþck÷]eù]hþ`nþ]löYbò\]âOHáIFÒ73Ð1.×52Ò0-Ø93ÞD<áJAáJA×@7Ò5,Õ4,Ü71æ74è&.ù0:þ2=÷+4ù08ÿ>Bý?Aë52à3-Þ6-Ý6-Ý5*á3*å2+ì4,ô41ö01ö.1ø-0ø..õ--ò/+ë3+â6,×9-Î?1µ1$±5)¯6.çplý‡…ÿ‡‹ÿoücwñ\oòYm÷OiûBbÿ9_ÿ<aõ@aäRg°QWd/'VE1FF,QU:T_?btJg�Qq‹[v’_x™bzžav�^p™Wm–Tl–Tn—Uo˜Vp™Wr›Yuœ]v�\vŸ[w Zw¢Zw£Xt£Us¢Ru¤Tv¥Ux¤W}§]~¥b~ c}›ez–crŒ]h�W`vRZoPRgHYL;[N=^Q@^Q@]P?\O>]P?^Q@_RA\O>ZM<[N=^Q@`SB_RA`Q>cT?_P;]P@\RHZPNZSZd`orp†wx—y{¢~€°€�·}ºwy¶qr²lo´rxÆv|Ê„Ȉ�Å•˜Á¤¤¼±°¸¼¹²ÊŲÑÌ®×ϪØϤÛТÜРÙËšØÁ•Ñ¨ˆ¶€f ^FÊyfÕtd¼G>ÖQLöb`èJIã?>á99â:9à;9Ú64Ø61Õ:5É91¿6.ÅB:¸<2¾I?ÄXK¶O@ÌeVÏdTÉXHÃH9Ã;/Ê6,×6,ß6/á4.Þ1+Ý2+ß7.ã=1à<0Ú7(Ü5%ã7)ä/$ç/%ì/)î1+î3.ë40ç4/â62Ö66åKMóX\ëOS×8=Õ6;øW]ÿ~„ÿ~ƒÿpuÿhlÿimÿgjó]_ïY[ú[_ÿVZÿCJú8@ú6@ö2>ì,7ë0;ð<GýS\ÿ`hÿhpÿemýalÿbmÿ`lü\dìRRÞE?Õ74Õ30Ö1/Ô/+×2.Û94àC<ÛC8×?4Ô<1×90Ú70ß82ä73ï6;ò5;ó49ö5:ô5:ñ56ê22ã1-Þ3,Ü5,Ý6-à5+ã2*è1)ï2,ô20÷12÷/2ù.1ù//ö..ó0,ì4,ã7-Þ@4Ë9,¼6*®0$·<5ÒXUÿ˜—þ{�ÿlÿd{ö]rðPhôGcûBbÿ;bù9^ôMkÒQc‰==P,B:%BH.IM4W`AZjCaxJo‰Yw“`w•_s•Yp—Xo˜Vl•Sl•Qm–Ro˜TqšXr›Yt›\uœ[r›Us�UtŸWu¡Vv¢Uu¡Tt SsŸRz¤Xy£[wž[t–Yl‹Rb~KWqBQk>?X0=V/:S,YL;[N=]P?^Q@]P?]P?]P?_RA`SB^Q@\O>]P?`SBaTC`SB^Q@dWF_QD\OF^TR`Z^c_mmk�uv•vy¢y}­}€·x|¹quµkp²lo´mq»uwÎ{}Ò‡ŠÑ–˜Ë§¦Å¸·¿ÇŸÒÏ°ÚÖ©àÛ¤äÞ¤ãÝ¡äÛ¢âÙ¢ÛÑžÔÇšÒ»™»œ€¦|cÄ‹wÉ}m³UIÑ`XõqmìZZæJKâ>?æ>>é?@å;;à66Û64Õ<7È72Ê?8·7.¶=2¼L>´H;Æ\LÚo_ÒaQÄI:º4(Ã0&Ó5,Ü5-Ú/(à40â94ä=5à<2Û7-Ù5)Ý5(â5'ë6+ì4*ì1*é1)æ1*ã2*à3,Ú5/Ó86×>@ãJLòW[ù]aû\aü\dÿ`hòRZíMUïRYü_fÿjpÿgjóZ]ñNSì38ð+2ó+5õ-7ï'4ä ,å&5ì6BûPYþZcÿcjÿemÿemÿblüZeõQXäEBÜ=7Ù53Ý52ß63Û2/Ú2/Ù6/Ý?6Ø>4Ö<2Ô:0Ù80Ý82â94ç:6è88é69ë78í68î87ê65å3/Þ1+Ü4+Ü5,Ý7+à5+ã2(é1)ð0+õ20õ12ö/2÷/1÷//ö..ó0,ì4,ä6-Þ>2Ò>2Â8-µ2(¶71Ø[Yÿ��úszþeyÿ]uøVnöOiùFdù@`ú<`óAaäNg¹KVx;6K1 ;9":C(EG/OS8Vd@^sHlƒUt�\v“]q“Wo“Um”Qi’Nj“Ok”Pm–Rn—Up™Wr™ZsšYsœVs�UtžVsŸTsŸTržQpœOo›Pm—Mk“MhŒL`‚EWv=Lh5D^.>X)=V,>W-?X.XK:[N=]P?^Q@]P?]P?^Q@`SB`SB_RA_RA`SBaTCaTC_RA]P@bUL[PL[QP`Y`jdrqm„vu•y{¤w{«x}´v|¸ns³di­`g­im¶ruÄ|~Õ……Û–•Ù§¨Ø»¹ÏËÊÆÙ×¾äá´èã©íè¦îé¥ëæ¤çá¥áÚ¤×Ï ËÃœ¶°�§›�™ƒk§ƒm§o^›QDº]UÜlh÷usî`^äJJäABéAAë??è8:â66ß=:Ó;6ÒB:Á80·7,·?1­<.¶G6ÙjYÞj[Ù[MÇA5Å2(Ï5+Ù6-Ø1+Ý52çB>êG@ß<3Õ3(Õ1%Ú2'Ü0$æ4*è1)æ/)ä/(ã2,â7/á=4ßA8Å-(É32ÙCBîUWø]a÷[_óS[ðPZçGQèHRîNXú]dÿmsÿntóX^â=Cè/4ò-4ù1;ü3=ó+8ì(4ô5DÿJVôIRôPW÷X]û_cÿdiÿ`föQXèBFÝ86Ü71á85è<8ê=9æ95à40Ü5-Û:0×;/Ö:.Õ9-Ú7.à91æ;4é<8ã:7â96ä86æ87ç:6å84à5.Ú1*Ù5+Ú6,Þ8,à6)ä2(ë0)ñ0)ö0-ô22ó12ö01ö0/õ/.ò1,ì4,ä6-Ü:/ÚC8Æ9/º4+´0,á__ÿ„†÷ktþbwýVpúPmýKküCeô;]ñ<_éIeÓRdœCGk<4O>,==%9?%EE-FI.Q\:YkCg}Oq‰Yr�Yp�Tl�Oi�Mg�Jh‘Ki’Lj“Ml•Qn—Uo—Xo˜Vt›Xs�UrœTq›QnšOl˜Mj–Ki”Ld�G`‡D[ARt8Ki3D`-?Y)>U'B\-D^/Ic4WJ9YL;\O>]P?]P?]P?_RA`SB_RA_RA`SBaTCaTC`SB^Q@\PDYNLZQV_X`gcrqm„ut”xz¡z|­v{²syµjr³bi­\c«`g¯lr¼v|ȈŠÝ”“㣣㵴àÆÅ×ÓÒÎßÝÄçä¹éæ¯íé¬ìè«çâªâܪÙÒ¨Ëߺµ—¤¤Œ™˜ƒ“ˆt—€n˜qb›cV²h_ÏrkúŠ†ôtqê\[åKKçDEê@Aì<?è:;ß:8Ú;5ÝF?ÒB:À:/¸:.¯5(§1#¾H:ÚbRéj[ÛSEÉ:,Ë3&Ô6+×4-×31ñMKüZUëMD×9.Ö4)Û4+Ý2(Þ-%ß-#Þ+$Þ-%Û0(Ù5+Ø:/Ö>3Á.&Ã3+Î:6ØDBÜFGÞEHãFMéLUîQZóV_õXaø[dÿhpÿnsòY^Ý:?é26ð/2ñ.4ð+4ê&0ë)4ø=HÿS^ë@HæAGçFKòQVÿZ^ÿW\óFLã68á53ã81ê;8ñ?;ò@>î<8æ93à70Û8/Ø:/×9.×9.Û8/á81æ93è;5â:7ß:6á96ã:5æ:6ä92Þ5.Ù2)Ú6,Ú8+Ü9*à6)ä2(ë0)ò/)ô1-ó32ò12ô01ö0/õ/.ò1,ì4,ä6-Ú8-ßG<È8/¾3,´-*ìfgÿy}öirÿcxüTnýJjÿHjþ>cò6Yè=]áQhÂXb‡A?eC7YN:GG/AC+JH1DD*KU3Sc>btJkƒSoŠUn�Rj�Mh�Hg�Hf�Hg�Ji’Lk”Pl•Qm•Vn—Ut›Xs›UpšRm—Oj”Lf‘IcŽFcŒF_†CZ�@Vx<Qp7Mi6Lf6Ne7Ne7Ke5Ni6Sm=SG7VJ:YM=[O?[O?\P@]QA_SC^RB^RB_SC_SC^RB^RB^RB\RHULQ^Xfkgxtp‰xt•xwŸxzªv{²syµiq²`g«Zbª^f®iqºu}Å}„Ì’–àžžä­®æ»½äÉÊßÓÒ×ÚÛÍàÞÅäá¾æâ¼äß¹ÞضØѵÎƯ¾µ¤­§™—™‹ŒŽ€‰†w‰{n�ujšrh¨me¹nh䉄í€}ñtpðdcðXWïLMîDEë??â96à;5áB<ÝF=Ë=3Ã=2ÃA4¯1#«0!ÇI:ÚXJÕM?Ë=/Ì8*Ð6*Ï2)Ô51õVSÿjgú]XâE<Ü;1Þ:0á9.á3*â1)ß1*Û0(Ó/&Í/$Å.#¿0"Å9,½3&¿2(Ê;3ÕDAÙEEÛBGÚ@HßDLëOZñU`ñW_ù_gþgnñZ_àAEè:;ì65í36í38ð3:õ:CÿHRÿV^â7=Û59Ü6:é@CøKMüIMó<@ç02ç51é;4ï<8ò<8ñ;8í:6ç83â70ß;2Ü;1Ù;0Ú9/Ü8/â70å61æ72ä;8â:7å95ç:6è;7æ93ß6/Ú3*Ú6,Ú8+Ü9*Þ6)ã3&è0&ð/(ò1,ñ42ð22ó11ô1/ó0.ñ1,ì4,ä6-Ü8.ÝC9É6.¾1*½10öknÿrxùgrû^sùOlþBeÿ>dÿ9`ô8[æFbØ[kªUZvD=bK;_YCQO:KI4PI6FD-FO0L\8ZlDe{Mm…Ql‰Ok‹JgŒGfŽGf�Fg‘Ih‘Kj“Ol•Qm•Vm–Tr™Vp˜Rm•Oi’LeŽHb‹E`‰C_ˆDZ�@V}>St;Pn8Ok8Qk;Wn@YpBUp=Ur<Vr?PD4SG7WK;YM=YM=ZN>]QA^RB^RB_SC`TD_SC^RB^RBaUEaVP^Vaje{xt�}yšzy¡wx¦ux­rx´iq°bi­[c«\d¬go¸u}Å€ˆÐ†�Ñ—œÜ ¥Ý¯³ã»¿åÄÈáÌÎÛÒÓ×Ö×ÑÙ×ËÙÖÇÖÐÄÐÉ¿Ìüù·³©ª£š›‰Š…wzsvvnulcle�ogŽbY›`X¿rlÖyrï{û{xÿqoýccõRSïGFïC?ç>7ß<5ßB9Ï;1ÑC9ÝUIÁ=0·5'¼:*Â=.Æ>.ÏB1ÖE4Õ?0Ï7*Î3/èMKú_]òWSßE=Ø:1Û8/ß8/ä91å9/å81à90Ú9/Ñ9,È9+Â9)°)·2!È@2ÒH=ÒE>Ë;:É78Ï:>Ï8AßHQìU^ðYbô]döagéTZÚACã?=ë?;ñAAúHHÿMRÿQVÿU[ÿW^Þ39Ù/2Ú03æ8:ô=Aø;?õ47ñ01ì51ì94ð95ì51ê1.é3/å3/á4.ã<4ß<3Ý<2Ü;1à91ã81ç51ç51ç98æ87è96ê86ë95ç83â5.Û3*Û7+Û9,Ý:+Þ6)ã3&ç1&ï0(ñ1,ð42í42ò21ó1/ò0.ñ1,ì4,æ5-à90Ø;2Ì70À/*ÐBAþosükröbnïQhõHfý;`ÿ5^ÿ5_û?bãMfÊ\g‹GFiE9\M:\WAWR>QJ7OG4ID.DJ.FV2Sc<]sEgKi†Li‰Hf‹Ee�FdŽDe�Gg‘Ii’Nj“Ok“Tl•So–Sm•Oj’Lg�Ib‹E`‰E^‡C^‡E[‚CZ~AVw@Tr>Tp?VpA\sG^uGYtAWt>Vs=NB2QE5TH8WK;XL<YM=[O?]QAaUEaUEbVF`TD_SCaUEfZJg^YmfvtqŒzy™yyŸvw¥uwªpu¯jo¯`g«^f®_g¯em¶pxÀ}…͈�Ø�—Õ—¡ÔŸ§Õ©²Û²ºß¹¿ß½ÁÜÂÄÛÄÅ×ÅÂÓÄÀÏÁ»É»µÃ¸¯À´©º¥™­”Œ›ˆ†‹qqqrqmnfcwhc†oi‚`W”c\›[R¶f]Ösmízuÿ{wÿvrÿheûYVýQMôG@å>6æE=Ö<2ÚF<ê\PÄ:-Á<-¼7(¾6(Ç>.ÏC2ÑB1ÓB1ÕA3Ð83Ö>=ÜDAÚB=Õ<6Ô70×6.Ü8/ã;2å:2ã:3ß;2Û>5ÕA5ÐD5ÌG4®-»:$ÍJ6ÐK<Æ=3º/*¿10Ê9<ÕBJßLTëU^íZbï\dëZ_ÞMRÓ>@Î3.×6.à;7è@=ðCEóEGñBGì?Câ59ß26ã36î7;õ79õ13ú.1ú24î3.ï82ð72é0+ç+)ç0,å2-à3,ã<3à>3Þ=3ß<3â92å82é62ë52ê67ë46ì44í55î64ë52â4-Ü1)Ý7+Ü8,Ý:+ß7*á4&æ1&ë1%î2)î52ì42ð31ò2/ò0.ñ1,ì4,æ5-â;2Ñ3*Ò;4Æ2.çUVÿqvöckðZfåI_òEcÿ9`ÿ0[ÿ4aÿBhÛMc°OVj5/]F6WL8TO9YQ>TI7LA/ID.@F*@O.K[6Uk=`xDe‚Hg‡FeŠDbŠCc‹Bc�Ee�Gf�Kh‘Mi‘Rj“Ql“Pj‘NgŽKd‹HaŠF_ˆF^‡E^†G_…H^�G\}F\zF\xG^xIaxLczLZuBYv@Yv@IB2LE5QH9SJ;WK=XL>[O?]R@dWFcXFcXF_VE^UFaWMf]Ticewrˆyu˜wuœsržrr¤ru¬ko®af¨Za§^d®fl¸ms¿v|È€†ÐŒ’Ü–›Ýš¡ÕŸ¨Ó¦­×­²Ü¯´Ü°´Ú±³Ù´´Ú²°Õ²®Ñ­§É¨¡Ã¨žÀ£™»–‹­‡—{xƒcadhdebZXi[Xzc]wVM�cZ‰MC¡TJº`WÖkcïxrÿ€zÿzwÿniý`[õSNæGAìOHÜB:ÝF=éVL¹+¾2%¾5%Æ>.ÏF4Ê?,Á2!È7$ÙE7ÝF?Ò:5Ê2-Í6/Ó:2Ö<2Ü;3á=4ä<3á90Ý6-Õ4*Ï5)È9)Å<*Â?+ÔS>ÇF1½<)¾:+Â<1Ã:4Á63À43ÛLNÜMOàOTåTWçX\åVXÙMNÓD@É5+Ð6*Ø7-Ü5/à21ä14å05æ.6ì4<ì3;ñ6?ú8Aü3=ú+5ÿ)3ÿ07÷,0ø46ø46ð./ë*+é0.ç40â5/â;3ß<3ß>4à=4ä;4ç:4ê84í74ë33ë33ì23í42ì42ç4/à2+Ù1&Þ8,Ü9*Ý:+Þ8(à4&ã3$è2%ë3)í60ì51ï31ð3/ñ1,ï3*ë4,ä6+ã<3Ì.%×>8Ì43ùceÿoxö\hîOaèD\øEeÿ;cÿ1[ÿ6aöGhÈK[“CFR,!VI8RM9NI5YQ>SK8F?-GG/=D%?K'GU1Rd:]sDc~Ie„HeˆFb‡Ba‰BbŠAc�Ed�Gf�Kg�Nh�Nj‘Nh�Je�GcŠGaˆE`‡F`‡H`‡H^�G]€F_}G_{H`{Ha{Ke}Mf~L[x@[z?\{BAB4FD7JF;NG=RH>VJ>ZM=]P=`Q<^R<^S?]VDaZPgb_nimrp~vq‘zuž{x¥yw¨pp¦gg£`a¡[^£_a«dg´opÀyzÌ„‚Õ�‹ß•“盘域ݟ Ö¢¢Ø££Ù££×¥¢×£¡Ó¤ Ó¤�ТœÌ šÈ ˜Ç�–—Ž¹Œƒ®ƒ|�pn{hfic_^e[YkYUqVO|VM‰WL’RH¨ZP³YP¿ZRØkdêyq÷‚{ÿŠ…÷vqývsë`]ßPLØD@ïZTË80Ã3(Ä6(Â6%À7%À7$Ä9&Ê;*Ï>-Ô>/Ð8+Õ;/×?2Ó;.Í5'Í4&×9-â@5à:.ß9-Û7+Õ7+Ï7)È;)Å>*ÄA-À<-Å?3ÊD9ÍG<ÌC;Ä;3»2,µ,&Å<6ÏF@ÙOLßVPãYVáXR×NHÍC9Ì;(Ó;&Ú8)á5+ç/-ë*/ð(2ò(6ù,=ù+?ú+?û*?ý(>ÿ&>ÿ$<ÿ$:ÿ$8ÿ(9ÿ+:û+8ó+5ë.4ç55ä;8Ý82Û:2Ü;3Þ;4ã:3å82è62é62ë33ë33ë54é62ç51â5.Ý5,Ø4(Û9*Û:(Û:(Ü9&ß8%â7%ã6%æ4&ê5,ì5/ï4-ð5,ð4+í5)è6*ã7)Þ:.Ö8/Õ:5Ú>?ÿlvÿbqúThôH`ÿKhõ6Xÿ5\ÿ=dö>båOh®JT]#!C-=:'GF2ON:OJ7DC/=B.>G,BH$EL#HR-R^8ZkAawIdLdƒJa…E`ˆAa‰@cŒ@e�Dg�Hj‘Nk’Oi‘Jg�Fe�Fc‹EcŠGdˆHb†H`„FbƒJb�Ha€GbGdHdHc~Ga}C[{<\=_‚B;?1>B4DD:IE<NE>QG=XJ=[N=_P;]Q9\T?^WG`[Uecfmkxqo…so”sp�tq omŸgfŸ``œ\]žZ]¢ce°ik¸trÅ}{Ï…�ÖŒˆß“�嗒䘕ؙ˜Òœ˜Ó�™Ö�™Öœ˜Õœ–Ò›•Ñ�•Ð›“Ìš“É™’Ƙ�Á‘‰¸†«xšskzlcfe][d[VhYRmWLzVJ‰UHšVK¯]RºYRÀYRÑf^Ýpiávnë~wú‡„ÿ‡„övsífcâTRëZUÆ3+¿0"Ä6(Ã7&À7$À7$Ä9&Ç:(Í<+Ò<-Ò9)Ô8)Ó:*Ñ:)Ñ:)Ð9(Ô8)×7)Ý;.Ý9-Û9,Ö9*Ï9(Ç:&Á:$½:&¾6*Á8.Å<4È?5É@8Ç>4Ä;3À:/½4,Ä>3ÑH@ÕODÚQIÕODËB:Á7*É8#Ñ9"Û8'å4*í0.ó*0ù)6ý(:ÿ+?ÿ+Aÿ(@ÿ'?ÿ#=ÿ!<ÿ;ÿ9ÿ9ÿ!9ÿ%:þ);õ-8ï29ç79å;;Ü94Û94Ü;3Þ;4ã:3å82æ72é62ë33ë33ë52è62ä71ß7.Ú6*×5(Ù9)Ù:'Ú9%Ú9%Ý8$Þ8"à7"á6$å4*é4+ê4)ë5*ë5(ç5'ã6&ß8(Þ</Ò4+Ó52óTXÿesÿZmõG`ÿIgÿ@bÿ:^ý3Yù;_ñNkËL_ˆ9>R+$@4&69&9<)@A/DA.@?+:B+>G*=DJQ%Zb;dpJfvOezQd~Od‚LaƒFa†A`‰=b‹=e�Dg�Hk�Kl‘Lj’Ih�Ge�FdŒEf‹HeŠGd†Ib„GdƒHc‚IbEbEd€Fe�FcDaA]€>_„?b‡B4:,9=/??5DA8IB:ND;TH<WK;\O<[P<[RA]WKb^[gfknm{ppˆqo”nk˜jg–ca“\[”ZZ–[\�\_¤hj´mp½wxÈ~ц„׊ˆÜŽŠß��Ü‘�Ò‘‘Í“’Γ‘Ð’�Ñ‘�БŒÎ‘ŒÌ’ŒÊ‘‹Ç’ŠÃ’‹¿’Š»Œ…±‚{¥~s“zevs^cf[Yc\Tb\PeYIrWFƒSE¢YP»aYÉ\YÏYWÚd`ÞkfÜohâwoú‹„ÿ‰…ÿ…€ÿ|wömgîaXÃ5+¿3$Â9)À9&¾:%½9$Á8%Å:'É:)Í:*Ô;-Ò6)Î5'Ð8*Ó=.Ó=.Î6(Ê.!Ù9-Ø8,×7+Ô8+Ð:+É:)Á8%¾7$½5'¿5*Á7-Ä:/Ç=3É?4ËA7ËC7Å;1ÉA5ÎD:ÏG;ÑG=ÐH<Ê@6Ä6(Ì6%Ô7$Þ6)ç4-ð/0õ,4ú*7ý*;û(;ú':ú%;ù$:ù!8ù7û6ý3ÿ6ÿ 6þ$5ú*8÷2<ð7=é9;ã;:Ü94Û94Ý:3ß:4ã:3å82è62ë52ì44ì44ë54è62ä71ß7.Ù5)×5(Ù9)Ù:'Ù:%Ú9%Û9$Ý8$Þ7$à6%ß5&á5)â5'ã6(â7&á7&Ý7'Ù7(Û=1Î1(Ö87ÿkqÿ]kýQgð=ZÿIjÿ7[ÿ=d÷5Zí?`äVp©@O_#%J/(:7(07%19$9<'@>)?=&>>&?B#?FW`1s~T}Œcvˆ^hSa}La€G`ƒC`…?`‰=cŒ>e�Dg�Hk�Mk�Mk“Li‘Hh�GgŒFgŒIf‹HfˆKd†Ie„Ic‚Gc€Fc€Fd‚FeƒGd€Eb€Bb‡DbŠDe�G06(39-:;3>>6D?9KB;PF<SJ;VM<WN=XQA[WLa_`hgolmoo‹om”ih”`^�VVŠSTŒVX•[^¡`d«ko¹qtÁz{Ë�‚Ò‡…؉‡ÚŒ‰Ú�Š×Ž�ÏŽŽÌ��ÍŽ‹Î�ŠÏ‹ˆÏŒ†ÎŒ†ÌŒ…É‹…Å�…ÀŽ‡»Ž†µˆ�ªy�~pŠ€_pyY^jXVc\R]^N^[HjXD}SCžUL½\VÐZZÙYZâdeåkhßpiäwpé|uì{sý„|ÿ‡€ÿ�{õkaÇ=2Â9)Â;(¿;&½;%¾:%À9%Ä9&Æ9'Ë8(×;.Ô6*Î5'Î8)Ð=-Ï>-Ê7'Æ0!Ñ5(Ñ5(Ò4(Ð7)Ï9*Ì;*Ç<)Ä;)Ã9,Ã9,Ã9.Ã9,Æ9/È</É<2Ê>1ÒE;ÒF9ÎA7É=0Ê=3ÎB5Í@6Ì=/Î5%×5&â6,ë40ñ.2õ,4ø*7ú*:÷&9÷&9÷&9÷&9ù&9û&:þ%:ÿ$8ÿ"7ý"6ú&5ù-9ö4=ð9>ç9:ß76Ü94Û94Ý:3ß:4ã:3å82è62ë52î45î45ì65é75ä71ß6/Ù5+×5(Ø8*Ö9(Ø9&Ø9$Ù8$Ú8#Ü7#Ü7$Ù6%Ù6%Ú7&Û8'Û8%Ù:'×:'Õ9*Ö:.Ð3,éIKÿoyÿVh÷E_ö=]ÿAfÿ3[ÿ8_÷?cçMiÅPc/:G@1*02%,7&4=*=B,@@(@<!C>!HE"T\-o}J�Ÿp–©|‚˜jiƒT_{H]~E^ƒ@a†@b‹?f�Ah�Gj’Kl‘Nl‘Nl”Mj’Ih�GgŒFgŒIgŒIg‰LfˆKe„IdƒHd�Gd‚Ff„Hf„Ff‚Ec�Af‹FfŽHj’L,4'06*690;<6A>9E@:KD:OG<RJ=QK=SOCYVO`_dggsjk€klŠih�b`�VVˆOP†OQŒUY˜_b§ei²or¿rxÆ|Ђ…Ö†‡×‡ˆØ‰ˆÖŠŠÔŠ‹ÌŠ‹Ë‰ŠË‰ˆÎ‡…Ї„чƒÐ‰‚ÏŒ„ÍŠƒÆŒ„À�†º�†°‡�£}x–mƒƒYg€SXmUQc]QZbMW_GbZCxUB’LB·TOÏTVÚSWå^bæfgákiåvoãtkàpeö}uÿ‚yÿ‡~ôl`È@2À8(À;(¿='¾<&¾<$Á:&Ä9&Ç8'Ì6'×9-Ø8,Ñ8*Ï9*É:)È;)Ç:(Ç8'Î8)Ð8*Ñ8*Ð7)Í7(Ë8(É<*Ê>-È<-È<-È</È<-Ç9-Æ8*Å7+Ä6(Î@4ÏA3Í=2È9+Ê7-Î</Ì9/Ë4)Ð/%Ù1(ä2.ì23ó.5ö-7ú,9ú+;û.?û.?ù/?ú0>ý.>ÿ/=ÿ.=ÿ.=ý)8û(7÷+7õ0:ñ7<ë8;á77Ù42Ü94Û94Ý:3á:4å95æ74è64ë54î45ñ56î66ê65ä71ß6/Ù5+Ö6(Õ8)Ô9'Õ8'Õ8%×8%Ø7#Ø7#Ø7#Õ8%Ô9%Ó;&Ó=%Ô=(Ó>(Ò=)Ò<-Ð6,Ù;8ÿ^dÿanÿOhô;ZÿBgþ5_ÿ5^ú0VóLjÚZq•?LY&+B**83/+1'.9+5A-:B*9<!?<RG'_Z2{…Q�£l©½Š¦¾Œ‹¥uo‹ZaI^€D_„A`ˆ?eŽBi’Dl”Km•No”Qn“Po”Nl’Ih�GgŒFgŒIh�Ji‹NhŠMd†Jc…IdƒGdƒGf…Ig‡Hg…Gd„Ch�Hi‘Jl”M*2%-5(2707:3;:5@=8EA8HD9NH<LH<NKBTSO]^cceteg~ef…a`ˆYZŠPR…LO†RS“[^£ei²jp¼sxÉw|Î�Ôƒ…Ø…†Ø…†Öˆ‡Õ‰‰Ó†‡Ê…†Ç„„ʃƒËƒ€Ïƒ€Ñ…€Ò…€Ð�…Ò‹„ÈŒ„ÀŒ…¸‹…«„�yv‹{kx�V`~PSnUPd^PZdLVcI_]DrWB�OB´WPÎVXÚSYä\`èaeäfgèqmëyoãqfö{sòriþxmå]QÄ;+½6#À<'À<'¿=%Á=&Ä;(Ç:(Ê9(Î6(Ö6*Ø8,Ô;-Í:*Æ9'Â9&Ä;(É>+Ì=,Ó@0ÙA3Õ=/Î8)È5%Ç8'Ê=,Ç:)Ç;*É;-É<+È:,È9)Æ7)Ç5&È6)Í;,Ò>2Ò>0Õ>3Ô>0Í6+Ç-!Ð+%Ú-)ç10ï26ö1:ù/;ü/>û1?ù0@÷1>õ1=ô0:ö.9÷.8ù+7ù+7û/:ø.:ô0:ð3:ì59å78Ý54Ø42Ü94Ü94Þ93á:4å95æ74è64ë54ñ56ò67ï77ë76å82ß6/Ù5,Ö6*Ó7*Ñ8(Ñ8(Ñ8&Ó8&Ô7$Ö7$Ó8$Ï8#Í;$Í=#Ì>&Í?'Í?)Í?+Ð<.Ð50êGHÿgpÿPbÿHeù7ZÿAjú/Zÿ7_ð4UáQjºUem19E(*C7754034/39/4=*2:"18ACh`9„‚RŸ®w¨À†µÎ—¬Ç’’®{v”`d…N_ƒE]…?a‰@g�Dl•Go—Np˜Qq–So”Qo”Nk‘HgŒFf‹Ef‹Hh�Ji‹Ni‹Ne‡Kc…HdƒGe…Fg‡HhˆGg‡Fd‡Cj�Ii“Km•N%0"*2%/4-36/672:95?<7B?8IE<HE<IHCQQQ[[c`asac|]`�\^‡WW‰RS‰SU�\] fh²psÂsxÊz~Õ|‚Ú€„Û�†Ú‚„ׂ…Öƒ…Òƒ…σ†É‚…È�„Ë‚‚̃€ÑƒÓ…€Ô†�Ó�…Ò‹„È‹ƒ¾‰ƒ±‡�£|‘uqufm}VYzQOlVKd_L]eMZdI``Dp\C‡TC¬\QÇ[YÓUXáZ^æ^bç`díkkózrërgöwnÝXOäZOÐD7Ã6%Æ;(Ã:'Â;%Ã<&Å='Ê<(Í:(Ð8*Ó7*Ñ0&Ô3)Ñ9,Í;,Å<*Â;(À;(Ã<)Ç;*ÓD4ÞJ<ÝG9Ò<.È4&È6'Ë<,Æ9(Æ9'È9)È9(È9)Ê9(Ê8)Ë8(Í7)Ñ;,Ö>1×?1Ú@4ÚA3Ö:.Ñ0&Û/+ä22í57ô5<ø3=ù1>ø0=ö0?ð.9î.9í/9ì/6ì-4ï,2ñ*1ñ*1õ2:ó4;ò5;ì59ä46à44Û62Û83Ü94Ü94Þ95â96æ87æ66è66ë46ò59ò59ð67ë76å84ß6/Ø5,Ó5)Ñ8*Ð9(Ï9(Ï9(Ð9(Ð9&Ñ8&Ð9$É9!Ç;"Ç=#Ç?%Ç?'È@*Ë>-Ï;/×96ûT[ÿ^nÿH`ÿ?aÿ:aÿ7dý4`û9^êFaÃN`ˆ?FQ12@64@::956>64;607:)3<!7CQZ-€‚P §q°Ä‡®Ê�¯Ì’¤ÂŒ‘¯{x™bgˆO]�C\„>`ˆ?g�Dm–Hp˜Op˜Qn•Rm”Qn“Mj�Gf‹Ed‰CeŠGgŒIhŒNhŒNe‡Jd†Ic…Hd‡Gf‰GgŠFf‰Ed‰Ci‘Jj•Mn˜P#. (0#,2(.3,241553984<;6FC<DD<FGANPOWZa^`o^`wZ]~]_ˆZZŒWX�[]šff¬qsÀz|Ï|€Ù€„ß�†âƒ‡â‚†ß�ƒØ‚Ó€�Ñ�ƒÍ„‡Êƒ†É‚…̃ƒÍ„�ІƒÔˆƒÕ‹„ÔŒ„ÏŒƒÄ‰‚¶ˆ�ª„€™|x‡pnsnbbuXTrTJiWIe`LaeL^dHcbFn^DzQ=�[K´ZRÄTR×WXâWZéX]ïcdõqlîncöpgÐC:ÑA6Ä2%Ç4$ÔC0Ç9%Æ9%È;'Ê<(Î;)Ò9)Õ7+×7+Î-#Í1%Í7)Ë<,Æ?,Á>*½:&»6#½4"ÐC2âN@äN@×A3Ë7)Ê8)Í>.È;)Ç:(Ç8'Å6%È7&Ê7'Ë8(Ï9*Ò:,Ô;-Ò9+Ñ5(Õ7+Ú<0Ý=1Û7.ç85ð9;õ;@ø9Aø4@õ/<ñ-;ì,9ê/8é19é27ê37ì46î45ò46ò46î47ï6;ì8;é69â45Ý33Ü73Þ;6Ü94Ý84à85â96æ87æ66é56ë46ò59ó6:ð67ë76å84Þ71×6,Ð6*Î8*Ì9)Ë:)Ë:'Ë:)Ë:'Ì9'Ë:%Ä8!Â: Â<!Â>%Â@(Å>*È<-Í81Þ;>ÿ]fÿRfÿIeþ8\ÿ>hø-Zý=fç=Zã\m£LT]--F93BC><87>56C34A4.<;)<E&EV*`s<�›`­½°Ê‹ªÈŠ¤Ãˆ›¹ƒ‹¬wx™df‡NY}?\�<_‡>f�Cl•Gp˜Op˜Qn•Rl“Pm’Lj�Gf‹Ed‰DeŠGgŒIhŒLhŒNe‡Jd‡Gc†Fd‡EgŠFgŒFg‹EeŠDj”Lk–Nn™Q(0!'/"(.$).'.0-2206729:4==5?@8DE?IKHORWUWdY[p\]{YX€YZŠ]^”ce¢mmµvwÇ~€×�„ቌ뉌퉉酆₂Ú€Ó‚�уƒÍ€ƒÆ€„ăƒÉƒƒË…‚φƒÐˆ„ÑŠ„·€Ãˆ€»‡€¬„}ž€y‹upwjfee^Vl\Ok[Kh]Ke_Ib`IcaHf`Fm_DlP8„V?¥]O¿aWÐYUÛQQíRX÷\`øieæYPÔD;Î7.Ñ7-×;/Ö=/Ó<+Î8'Í:&Í:&Ï:&Ò9)×7)Ù5+Ú6,Í/$Ë3&Ë9*É=,Å@-Á@+¾=(¼9%Á8&¾1 éWHßK=Â. Ï=.Ì=-Å8'È;)È;)Ç8'Æ7&Ç6%È5%Ê4%Ì4&Ð7)×;.Û=1Ü<0Ù7,Ù7,â>4ìC<õCCñ8=ð3:ó4<õ3>ð.9ë-9ì2=ê7=ä5:ã57ä67é99î:9ð86î64ï;<ë;;é99æ87á85ß74Ü73Ü73Ý84Þ95â88æ::è:;é9;ì8;î7;ó6:ó6:ï77ì87ç;7ß:4Õ8/Ï7*É7(Ç8(Æ9(Æ;(Æ:)Æ;(Ç:(Æ;(À8 Â>%½;!µ5»:$ÉD1Í?3É0+ÿ^eÿVfÿIbþ;[ø2Võ1Yú7aóCgåZoªFNm32K2+@@6>C<>96>./J03J2.@9&BI'Xn=}›_›¸v§Ä�ªÈˆ�¼€—¶}“±{�Ÿmd„RVw@Xz=]€<a‡>hŽCm“Fl”Kk“Li’Nj‘Nh�HeŠDa†Aa†Ad‰Ff‹HfŠJdˆHg‹Kg‹Kg‹KeŠGeŠEdŒEj�Ik“Jo›Po�Rr U)/!)/!(.$).'-/*02/450782;=2=?4AD;GIDLPSSV_WXjY[t[Z|]\†bb”jk¤ttº}̆†ÞŠ‹èŒŒì‹‹íˆˆè…„ကÖ}ЀÍ��É„…ƃ‡Æ†‡È‰ˆÌ‰ˆÎŠ‡Î‹…Ë‹…ň€»‡°„|¡€x�{s~tkljc]d^Rd_Ld_Kd^He]Hh]Gh^Ei_Fm_Dt_D~Z@‘XD«YKÇVNÜRPõRWþX\îUPãLEØ>6Õ7.Ù6-Û9.Ú:,Ö9(Ð7%Ð9&Ð9&Ñ8&Õ8)Ù5)Ü4+Ü5,Í1%È4&Æ9(Å>+Ã@,Á@+¾=(½8%Á8&Ê;+ãQBÜH:Ç5&É:*Æ:)Á5$Ê=+É<*Ê;*É:)Ë:)Ë8(Í7(Ï6(Ò6)Õ7+Ø8,Ö4)Ô0&Ø1(Þ7.æ;4÷CDñ8=í07í07ï/:í/9é1;é5>ã6:Û25Ø22ß:8îEB÷JF÷EAñ?;ç:6å97ä86á85à85à85à85à85ß74à85å99ç9:é9;é9;ë8;î7;ð69ï58í76è96ã:5Ü;3Ò8.Ê6*Æ8*Ã:(Â:*À;(À;*¿<(À;*À;(¼:$À@'¸:#²6¸;'¿;,È91×==ÿ]fÿPcþ@Zù8Wõ7Yô;]ó?dãIeª?Iƒ?>];2E;/8>07?2E<7Q>:N0.Q9/VK5bjCvŽZŒ°pœÁ{ Å—ºz™¸˜¶€ŒªxrŽ^WtDSq=\{@^~=b†>hŽCl’Ek“Ji“Kh‘Mh‘Mf‹FcˆBa†Ab‡BeŠGgŒIg‹KfŠJhŒLiŽKh�Je�GdŒEfŽEi‘Hl–Lo›Po�RqŸT-1",0",/&,/(-/*/1,25.36-8<.:>0?C5EH?KMJQRVUUaUVh\Yt^]}edŒnn yy·ƒƒË‹ŒÞ��ê�ŽìŽ�늉愄܀~Ñ~{Ê}ƀņƆˆÅŠŠÈ‹‹É‹‰È‹‡Ä‰„¾Šƒ·ˆ¬…{ uŽzotimne`h^Tb\L]`K]aJc^Hg\HkZFm[En\Fo^DwbGxY=…S:¢UCÆWLàUP÷NQýOQâ><Ü=7Û81Ý6.ß7.ß7,Ý7+Ø6'Ó6%Ñ9$Ñ8&Ó8&×7)Û5)Þ3+Ü5,Ê4&Ä6(À8(À=+Á@-Á@-¿<*À8(Â4&ÛI<ÚF:ÓA4Ì=/Ã7(Â:*¾7$Å:'Å:'Ç:(È9(Ê9(Ê7'Ë5&Í4&Õ9,Ô6*Ö4)Ö2(Ú3*Þ6-ã;2è;5ï;<í49é06ì18í4<ì4<è7?ç<BÛ78à@@ìNK÷XTýZUøQKîC<ä71à72à72à72ß82à93á:4â:7ä;8á77ä88å99ç9:é9<é9<ë8<ì8;î68í57é56æ87á96Ø93Ï8/Æ6+Â8+¾;)½;+º=)¹<*¹<(¹<*¹<(·:$»A*²9$¯6!¶;+³1$Ã4.ëORÿZfÿJ]õ;Sï7Sñ@\ïHdæIdÍN_�78d>3Q?1DA06=+48'>6+K9/E-!R=*e^Aˆ]‘ªs•¼{•Ày•Ày‘³w›·„˜´ƒšm\uKGb5Mi8[x@_}?c„?hŒBk‘Dj’Ih’Je�If�KcˆCc‡Ac†Bd‡CgŒIiŽKjŽNjŽNh�Li�Mh�Lf�Ie�Gg‘Gk•Im™Lr¡Ss¤Uu¦W13%04%/2'/2)/1,02-14+25*8<-:>->B1CG8HKBMOLQRVSS]XVd[Ynbb~mm“ww«‚‚‹‹Õ�Žá”‘ê��莋⇅؂Ì|Ã|¿½‚‚¾…†¿ˆ‰ÂŠ‰Â‹ˆ½ˆ„¶…­ƒ}£„zœu�xm}qflmb`k^Ui[Nb[I]aJ]aJc^Hg\HmZIqZHr\Gu]ErX?wR8‰S;ªYDÍ\NâSMïEFñ>AÚ1.Ù4.Þ5.á6.á5+à6)Ý5(Ù8&Ó6#Ñ9$Ñ8&Ñ8&×7)Ù5)Þ3+Ù7,È6'½8'»8&»:'½>+À?,Á<+Ä8)Æ4'èTHÑ=1Ë9,ÏA3½4$¿<*¿<*À9&Ä9&Å9(Ç:(È9)É8'Ë5'Í4&Ò6*Ó3'Ô1(Û4+â92ç?6ê?8ê=7ç55é58í6:í6;è38ã28á6<á<@ñVTûc`ÿmhÿhcòUPãA<Þ71Ý4/Þ71Ý60Þ71Þ71ß82á83ã:7æ:8ã77å78æ89ç9:é9<é9<ê7;ë6;î5:ì57é77å97Þ;6Ö<4Ì9/Â8+¼8)¹:)¶;)´=)³<(±<(³<(³<(±:&´?+¬9&«8%²:,«+"Â42û`fþTaøG[ï<Rë=VçH^áObÑN^³RYsG>WG7JC0ED/AD/;?(97";3>2OF)kkEˆ”f•¬t‘´tŒ·r�¹wœº„ž¸‰�§}k�ZJ`:@V/Lc7Wr=_{@bƒ>g‹Aj�Ch’Fe‘Fb�HcŽGa†Ac‡Ad‡Cf‰Eh�Jk�Mm‘Qk’Qi�Mi�Mh‘Kg‘Ig‘Gi“Gm˜Io›Lu¤Vu¦Ww¨Y78(58'57*56.56056057,68*;>+<?*?C,DG4HJ<LMEOOMQQQVUSYY[``lhh€ttš~~°‡…ÆŒŠÓ”‘à’‘á��Ú‰‡Ðƒ€Ã€|¹|µ|³µ��·„„¸ˆ†·‡„¯…�¦�z›w�}t‡yozrfjm_\j\ShZMjZJf[Ia_Ja_Jd]Jh[Hl[Iq\Iu\Hz[G|V?…S<šS?´WEÍSFÚG?â96â30Ü0,Ý4-â5/â5.á3*ß3'Ü6(Û:(Ò7#Ð:"Ï:$Ð9$Ô9'×7'Û5)Ö8,Ã:(¸9&´7#µ8$»<)Á>,Å<,Æ7)Ï8-ô\QÖ>3È4(ÍA2¸3"¶9%¼?+¿:)Ã:(Ä;+Æ:)É:,Ë9*Î7,Ï7*Ï2)Ò1'Ô1*Ý6.å<7ê?8è;7ä73ç85í;;ï;<é69â25ã68çCDëMLÿyuÿqlñ`[ÝJCÐ71Ð3,Û83ç@:Þ93Ý90ß80Þ7/à72â62ã75å76å78æ68ç7:è8;è8;è8;è7=é6:î7;ë78è88ä;8Þ<7Õ<4Ê<2À:.¸9*´;*°;)¯<)¬<(«<(«;'ª;'ª;'«>*¤9'¦8'ª8-ª/(Ê@@ÿgoôN\óHZíDWäGZØN]ÈS[¶SV™XRcN=NM8JH3HF/FD+?B%=D#>H#KV.\g<v‚RŠ˜e�¤l�©n�²r—¹}�·‡‘¦}s‡bPdA>O/AR0Mb9Vl;_{Ac‚?i‹Aj�Cg‘Ee‘Fa�Gb�Fa†Ad‡Cf‰Eh‹GiŽKk�Mk’Qm”Sj‘Nj“Oj“Mi“Kh”Gj–GnšKožNu¦Ws¦Wt§X<=+<=-<<0;=2;<4;<4;=0;>->A,@D+BF-FJ1IL9MOAPPHRRHWXJZZN^^\ffpnn†yx ƒ€µ†„Ã�ŒÑŽ�ÓŒ‹Ïˆ†Åƒ€¹|z¬zw¦yv¥zxªzz®}®�~©�} ~z•|t‰yq€tjrpfem`Zj\QjZKl[Kl[Ik\Ih]Kf^Kh]Kj]Lm]Mq^Mx]L€[I‘aM™VE¦O>µJ:ÅA5Î7.×0(Ý0)à5.á6.â5.á3*à2'Ý3$Ü6&Ù:'Ð8!Í;"Ì<$Í;$Ñ:'Ô9'Ø6)Ò:,À=)³:%®5 °6!¹:'Â=,È:,Ê6*Ø>4÷ZQãI?Í9-Ì@1·6#«4´='»:'¿7'À7'À7'Ä6(É7*Ë7+Ï7,Ù<3Û:2Ý:3ã<6ç>9é=9å84á40ç:6ê;8ê::æ87ê>>õOOÿccÿpmùhcâUNÇ<5¿0*É4.Ö;6ß=:à;7à;5ß;2à91à70â62ã62ä65å55ç79ç79è8;è8;è8;è8;ç6<ç7:ì8;ê88ç98ã;8Ü=7Ó?5É<2¿;.³8(¯:(¬<(ª;'¨;&¦<&¥;%¤;&£='Ÿ:&Ÿ:(Ÿ9*¡5)µ>8ÚTUúdmíKZîIZèI[ÛMYÇRX´VTŸYQˆaPOH5FN7KI2F@(;648=N"Kg6j‹Tv•\ƒšbˆ›cŠ›d�¢j“¬r—°yƒ•kn}\P_@<K.8G*AP1O_:Wj<b{Be‚BiŠCj�Cg‘Ec‘F`�F`ŽFa‰Cf‰Ej�IkŽJj�Lk�Mj‘Pl“Pk”Pl•Ol–Nj–Kj–Im™JqžMr¡Pz«Zx«\x«\CB0AB2BC5BB8BB:BB8AC5@C0CG0DH-GJ-JM2MO7OQ<SQDUSD[[CY[C\]Oaa_hguqq�zy£~~²„ƒ¿…ƒÂ…„À„�¸}{¬wtŸsn–om”omœqo tqžusšur�spƒqkwoimlb`j`Wj\QjZKm\Jo\Ko]Iq\KpZLn[Mm]Nk^MlaOq`Nz_N…[Kž`Q¤PE­B8·;1Ä6,Ì2(Ô2'Ü4)à8/á6.á4-à4*à4(Ý6&Ø7%Ó8"Î9!Ê="É=$É=$Í<'Ð;'Ô8)Ï<,¿@-°;'©4 ¬3 ·8'Ã;-Ê:/Î5-×82íNHñTMÖB8ËA4½>-¦1±<(¼=,¿:+¿7)À6)Ä6*È8-Ï;1Ô;3áD=ãA<ã@;å=:æ=:è<:ç98å84æ95æ93å84è<:ôJJÿ\[ÿdböb^Ë=9À71¼3+Ã81ÒB:ÝF?Ü=9×4/à;5ß;2á:2á81ã73å84æ66é77é69é69ê7;ê7;è7=è7=æ7<ç7:ë7:é77å76à;7Ú=8Ð=5Æ<1½;-±8'­:'©:'¦;'¤;&£:% :$Ÿ;$ž=*“6$™<*™9)™3'ÁPJìjlí[eëO\èLZÝIWÍKS·SQ¤\P�cP~jRIM6CK4HB,E9!@7DI!Tp=g“V}²n�³n‚¥e|‘V�ŽX�—bˆ–a{‹\WeAGT89F,6C)8E+=J.JX5Wh<d{EfƒCj‹Dk‘Fh’Fc‘F`�F`ŽEc‹EiŒHm�Lm�Lk�Mj�Li�Mj‘No˜Tp™SošRnšOo›NožNs¢Qw¦U|­\y¬]x«\GD3GD5GE8HF:HF:HF:FG9EF4FH2GI1KK1NN4PP8SR>VSDWTA[Z>[Z>ZYE^[Rdbgmj{vs�|x�~{¦€}ª„~ª‚}¦|všto�nj…jf�hd…gfˆkh‡lj‚khyifoiceha[f]Tg[MgZIk\In]Ip_Kq_Kp^Jo\Mo\No\Mp]Nu^N|]K†[K‘WIŸVG¢F9ª:/·7,È8/Ñ7-Ø6+Ý6-Ý6-Ý5,Þ3)ß4*Þ6)Ý7)×8%Ñ6"Î9#Ì<$Ê<&Ê<&Í<'Ð:)Ò9)Î=,¾A-¯<'¨3ª3¶9'À<-È</Ë7-Ì2*ßB;õ[SÝI?ÐD7ÈF6«2·>+ÉE6ÌB5É?2Ê>1Í?3ÔA7ÛD;àF>á?:â=9á96à74á53â64å97å<9â?8Þ=5Û94Þ?9èMHèQJÐ@8¶)"¶-%¾5-È>4Ï@8Õ=8Ù<7á<:æ>;á96á:4á83ä84ä84æ95è96é99é77ê88è8:è8:è8;ç9;æ8:å79è58å55á77Ý97Ö=7Ì>4Á;0·<-­:(¨;'¥<'¢;( ;'ž;&�:%š;'—@,Š5!—>.–9*–/&Í\Xöx{ßX_äXaÙQ[ÈKQ¶LN¨UO›`RŽhS€oUOV5?J(<<B>TR)jvD‚¡e“Â~‡¼t…·p{�`k‚JsO�ˆ\t|SXb=;D'2<#4=(<E0<E09C*CQ0Uf<e|HgƒFkŒGl’Gh’Fe‘Db�Ga�FeŒIj�Lo’Nn“Nk�Kg�Ig�Ih’JrœTq�RržQožNpŸOr¡Pw¦UyªYu¥Yr¥Zr¢XQK;QK;QK=QK=QK=QK=QK=QK;TO<TO<UP=VQ>XS@YSCZTD[UEa\H`[G_YKaZPe][icgogtsj{wnƒyp…|sˆ}t‡yp�rjyjboe]hb_h`_e^]b][\]YX^YU^ZQa[Od]Me]Jg\Hh\Fi]Gk\El\Ej^FicMqfTpYIsO?ŠVIœ\P£YL©ZK¥N=¤E3¦;+°8(À8,Ï8/Ù40Û2-Ý6.Ý6-Þ7.Þ7.Û7+Ù7*×7)Ô7(Ô8)Ò9)Ñ:)Ñ:)Ñ:)Ñ9+Ò9+Í;,º;*µ=,®9'¦1ª2!µ:*À<-Â8+È5+Ì5,Ó;0ÚC8ÜJ=ÔH9Â=,¶1 ÜPCØJ@ÓE;ÔD;ÙF>ÞG@ÝC;ß=8ß63æ87ê;8é:7å84à72à93Ü=7×G>ÍC8Ä:/»3'·1%¹5(¾<.Ã?2À:.Ç=2Ò?7Ú@8à=8æ:8ê88ì89å78ä79ä88ä88ä86ä84ä84ä84à40á51â62ä84å97å97å99æ89é;<æ::ß99Û<9Õ@:ÌB8À@5µ?1¨:)¢;(Ÿ<'š;'—;&•9$”8#�9%�B.‰<*…/"�/&ºMHânnälnÌUY¼KM´MNªQM¡TN›XO–^Q•aS‰iPVY,9L6EQ`%{�O™¯nž»xœ½x‘´p~ž_h€LZm@WeAT]@HP8>C/9>*>C/<@18?/:C28D.?N1Rd>\t@f‚Eo�Km“Hi’Fh’Fg’Jg�Ji�Ok’Qn“Pl”Nk“Lj”Jk•Il˜Ip�Ln�Lp Lq£Ns¥Pu¦Tx©Wzª^w¦blšZbŽOVO?VO?VO?VO?VO?VO?VO?VO?WP@WP@XQAYRB[TD\UE]VF]VFc\Jb[IaZJbZOd[Th^\lbcodhpekrgmtiotioqfllaef[_`XV]ZSYYOWWMWUIWSHXTH[UG\VFaZHc[Hc[Ff[Eh]Gi_Fj^FhaGbbJngTv\K~SC™WKªXM®RG¶QE­G8¬C0«<)±9(½9,É90Ô72×4/Û7.Û7-Û7-Ú8-Ú8-Ø8,×7+Ô6*Õ7+Õ7+Ô8+Ô8+Ò9+Ò9+Ò9+Í;,½;+µ=,­8&¥2¦3 ¯:(¹>,½;+Å9,Å5*É5)Ñ:/ÚF:ÞOAÝQ@ÜN@ÛG=×@9Ò;4Ò;4Ô=6Ø=8Û94Ü41ç77ë78í68ê65æ74â94á<6ÛB:ÉA5¾@2·9+°5&®5$±8'·>-¾@1Ä@1Ì@3Ö?6Ü=7ã:5é75î66ï79é6:æ8:æ8:å99å97ã:5ã:3ã:3á81â92ã:3ã:3ä;6ã:7ã:7â88ã9:à::Þ<:Ø?:ÐA;Ä?6·=0¬;+¢;*�<)™=(—<'“;'‘:&‘9%‹:%8$~:'‰5*›=5½QNÕedÏ^`·MM«MK£SLœXO—[P–]R•_S—^S‹eN_^.Rf']r1y�N—°l£¾{—²o† _i�G_v@Sf9IY5CP6?H58>059+9;-?B1=?28>0;A37C/?M3Od=]uCf„Fp‘Ln”Il’Gj“Gi“Kh‘Kk’Ql“Tm”Ql”Nj”Jj”Hj–Gl™Hm�Im�Gn Iq£Nu§Rw¨VvªXx¨^j–W_‹PV‚G[TD[TD[TD[TD[TD[TD[TD[TDZSCZSC[TD\UE]VF_XH_XH`YGd]Jc]Gc\Ic\Ld\Oe\Sf]Xg][h^\i_^j`_j`^h^\d[V`WR]VNYULVTHSQEQOBQMARO@TQBXRB\WD^WD_XEbZEe]Hg_Hj_IhbJbaLmfSx^O†XK \S¯VP±JC¶E=·A7¸>1¶;,¸9*¼8+Â:,É;/Î:.Ö8,Ù7*Ú8+Ú8+Ú8+Ø8*×7)Ô7(Õ8)Ô8)Ô8)Ô8)Ò9)Ò9)Ò9)Í;,½>-´?.¬9&¡2¡2©9%³>*¹>,Ã>/Ä8+Â3%Ç3'Î</ÖG9ÛN=àNAØ93Ø42Ò3/Ó40Ù:6à<:å=<è:;ì8;í57ë35é24ç55å97á<8Ø?7¾8,³9*°6'¬4$­5$°8'µ<+»=.ÈD5ÍC6ÕA5Ü>5á:4å63ì65ï79ê69é69é69ç77æ95å:3å:3ã;2ã;2â;3ã<4á<6à;7ß:6Þ97Ý98Ü87Û=<Ú@>ÒA<È?7¹;/­7+¢7' ;)›<(—<)“<(‘:&�:%Ž7$ˆ:&z6!x5"‡9,žF<¶SN¾VS¶PN©JF¢MHœSJ˜ZO—_R•aT•aT•_U‰eOrqEz�U“¦n£¸ª¿†œ´zy�XXo9?R$@R(AQ-=L/8D.4=,39-58-9;-?B1<>17=/:@27C/>L2Od=^vDh†Hr“Np–Kn”Il•Il–Nj“Mm”Qn•Ro—Qm•Nk•Kj”Hk—HmšIm�Im�Ip Lt£Rx©Wx©Xt¥VpŸW]‰JU~DNw=`YI`YI`YI`YI`YI`YI`YI`YI^WG^WG_XH`YIaZJb[Kb[Kc\Jd^Hd^Fe_Ie^Ke^Le]Pd\Qd[Rg^Ug^Uf]Tf]Te]Rd\Qc[Na[M\XLXVITREPN?OL=OL=QN=RO>UR?XS@ZUA]XD`[Gb]Gf`JfaKcbNkdRt]O‡[P§b[µZU¶FD¹<8Á<7Å:3Ä:0Â8-½8)¼;(½>+Ä?,Ñ9+Ø8*Ù7*Ù9+Ø8*Ö9*Õ8)Ô8)Ô8)Ô8)Ô8)Ò9)Ò9)Ò9)Ò9)Í<+¾?.´?.ª:&¡4œ3¢8"¬>'´?+»>,¿:+Â6'Å6(É7*Ì:-Ì=-Ò:/Ý54Þ22Ú00Û32Þ97ä<;é==ë;=ë7:ê37é05æ23å57á85Ù53Î70¶4'®6&¬7&«6%«8%¬9&±9(³:)ÃA1ÉA3ÓA4Ú@6à;5ä84ê65í76ì57ê67ê67é75ç83æ93å:2ã;2à90Þ:1Þ;4Þ;6Ý;6Ú;7Ù:7×98Õ;9Ô@<ÒC=ÇA8º<0¬6*£5&�6%š;'—<)“<(�;&�;%‹9$Š8#…8$€=*v5!6'—E9©OGªMF¦KF¥PI›QH—UI“[N’^P’aS•aT–`V‹fSˆ†_˜¨w­½Ž¦¸ˆ�¡tr…WN`6.@,=2A"8F,:F09B17=/9<1;=29:,>A0;=06<.:@26B.=K1Nc<`xFi‡Is”Or˜Mp–Ko˜Ln˜Pm—Op˜Rp˜Rp˜Qn–Mk•Ik–Gl˜In›Jn�LožMs¢Qx§Wz«Zv§XlœPd“KZ„ER{AMv<b[Kb[Kb[Kb[Kb[Kb[Kb[Kb[KaZJb[Kb[Kc\Lc\Ld]Me^Ne^Ld^Fe_Gf`HgaKhaNg`Mf_Mf_OibRhaQg`Ng`Nf_Lf_LgaKfaMa^M^[LXUFSP?OL;NK:LK9ML:ON:PO;RQ=UT@XWCZYD]\G^]IbaOf_OlVH…[Oªja½d^¿MLÇABÏ;;Ó97Ñ83Ë6/Á9+¹<(´B*»C*Ì;(Õ8'Ø8(Ö9(Ö9(Ö9(Õ8'Ô9'Ô9'Ò9'Ò9'Ò9'Ò9'Ñ:'Ñ:'Ë<+Á?/´?.ª;( 7"™5™7¢<%¬B,±<(¸;)Á9+Ç9-Ë9,Í9-Ð<0Ø;2è8:ë49æ37â46á46ß56à35á35è59è38ç48æ68ã9:Û75Ï1.À0(´6(­:(«:(«:(©:&©:&¬9&­8&·:(À<-Í?3×@7ß>6ã<6è96ë76í76í76í76ë84ê84è:3ç:3ä<3Þ:1Þ;2Ý<4Û>7Ù>9×>9×=;Ó=<Î@<ÉB<ÅB:¹=3¬8+£5(œ6'š;)–;(‘<(Ž<'Œ;&ˆ:$†8$‡6#ƒ8#‚A-r3!x3#“G9¢RG NCœND VK–VJ’XJŽZL�\N�]R“_T—aWŽgV�‰f–¥z›ª��h\lHDS22A$$2,:#.;'2=-7?0:@4;>3:<199-89+=@/;=05;-9?15A-=K1Mb;`xFjˆJt•Ps™Nr˜Mp™Mo™Qn˜PpšRq›QpšPn˜Lm—Km˜Io›Lq�Nq Pt£Sx§W{ª\y§\pžUb’HX‡A]‡HU~DPy?d]Md]Md]Md]Md]Md]Md]Md]Md]Md]Md]Me^Ne^Nf_Of_Of_Mf`JgaIhbJhbLicMicMhaNhaNhaNhbLgaKf`Hf`HgaGhbHhbJe`L`]J[XEVS@QN;MJ9IH6HG5KJ6JL7KM8LO<OR?QTARUBSVCXYI]WIbPBzXL g^µe^ÃUTÖQRÙ@BÝ:;Û76Ô70Ç;.»>*°D(³A&É;'Ó8&Ô9'Õ:(Õ:(Õ:(Ô9'Ò9'Ò9'Ò9'Ò9'Ñ:'Ñ:'Ñ:'Ñ:'Ë<+À>.³>-ª;( :$—7”6™;!¦B*±C,¸?,¿;,Â6)Å3&Ë4)Ò;0ß=8è59í4:ë6;è:<å;<ã;;á;;à::ß58à67â9<à<=Û=<Ð85Ã2-µ/&¯:)ª=)ª=)¨=)¨=)¦<&§:%¨8$­5$·8'Ä<.Ï<2×=5Ý<4ã<6è;7ë84í74í74ë83ê93ç:3æ;3ã<3á>5Þ@7ÛA9ÙB;ÖA;Ô@<Ñ@=ÌA<Á@:º@5³=3©9- 6(š6&–9'•<*�;'�<'Š<(‡:&„9$ƒ8#ƒ6$6#{:&q2 |9(”O?£YLœRE•OC˜VH”ZL�[MŒ[MŒ[M�[P’^S˜_VŽgV}wWx„^p|XSa@9F*.<#/<(0=,1=/2<15=29?5>A8?A6;;/78*78(<?.:</5;-8>04@,<J0La:`xFjˆJt•Ps™Nr˜MqšNpšRo™Qq›Qq›Oq›OošKn™JošKržOu¡Rw¦Vx§Yz©[w¥ZqŸVf”L[ŠBVƒ@_ˆLW€FR{Ae^Ne^Ne^Ne^Ne^Ne^Ne^Ne^Ne^Ne^Nf_Of_Of_Og`Pg`Pg`NibOicMhbLhaNhaNhaNhaNhaNgaKgaKgaKf`Hf`FgbEgbEgaGf`Jc^J`[G\WCUS>PM:KJ6IH4FG5FI6FI6GK:HL;IM<GN<JN?NPBUQEXO@hRD�ZK–VJ¯RKÓYVÛIIà@@Þ88Ù72Î;1Á@-±B'±>#È;'Ð7%Ñ8&Ò9'Ó:(Ó:(Ò9'Ñ:'Ñ:'Ñ:'Ñ:'Ñ:'Ñ:'Ð;'Ð;'Ë<+¿=/´<.«<)£='—9 �5“9 @'­C-¶A/¼=,À6)Â/%Æ/&Ð6.Ý:5ç48í5=ì;Aé@CäABßA@Û@>Ù?=Ì22Ï53Ò:9Ð<:Ç<7½81´5,«5)§<*¤=*¥>+¤@)¤@)¤>(¥<'§:%ª7$°7&º8*Â8+Ë8.Ó:2Û<6ã>8è94ê93ê93è:3ç:3æ;1ã<3à>3àB7ÜB8×C9ÒB9ÎA:È?9Ä=9¿>8°;2§;.¡7*œ6(—7'”9'�;'Œ;(‰<(ˆ=(„;(ƒ:'€9%8$6%}6$v3"w6$…B2—TDœXK”PC�PD”YK’ZM�[M�[N�\O’\R•\S—\V�cUkbCV`;IS1<G)4@(2>*4?15?44>56@8<C;AG=DG<BD6=>.;:(78(<?.9;.4:,7=/4@,;I/La:`xFi‡It•Ps™Nr˜MqšNq›SpšPpœOpœMpœMn›Jn›JpœMt Qx¤W~ª]~ª_x¦]o�Ud‘L\‰D[ˆC\‰H`‰MY€IRyBg`Pg`Pg`Pg`Pg`Pg`Pg`Pg`Pf_Of_Of_Of_Og`Pg`Pg`Pg`PkdTjcSibRg`Nf_Mf_Mf_Mf_Lg`MgaKgaKgaIgaGgaGgaGf`FgaIf`Jd^H`[E\WCUS>ON:JL7EF4DG4CG6BI7BI9AJ9BK:CJ:HL>PQCRPAXM;dK7sD4•I<ÁXR×QNÞDBÞ:9Û75Ô;3Ê>/¸>)´7!È:&Ï8%Ð9(Ð;'Ò;*Ñ<(Ñ:)Ð;'Ñ:)Ð;'Ñ:)Ð;'Ð:)Ð;'Ð:)Ë<+¿;.µ;.­<*¤>(˜:!Ž4’8ž@&¡8"¯<)¾?.Æ</Ê7-Ï6.Ø;4â><ë>@î?DéCEâDCÕ@<É:4¿4/¸1+·0*º50½94º;4°91¨8-£9, :+ž=*ž?+ ?,Ÿ@* @* @*¢?*¤>(¨;'­:(²7(·5'¿5*È:0Ó>7ßB;ä=5ç<4ç<4å=4ä=4â>4à@4ÝA5ÛC8ÖC9ÑC9ÊA7Ã>7¼;5¶93¯:1¢7-˜7'’5&’7%‘:)Ž;)‰;'„9&†=*„=)�=(;(}:'{8%{7$z5%w2"~9)ŒG8•RB’PBŠK<�QD™_Q�XK�ZL’\P•\Q–]T–[S–YTŠ^QaX9EM&5?7B$=G/:F23>0-7,/905?6@F<EK?GI;AD1>=);:&67%;>-9;.4:,7=/4@,;I/K`9_wEi‡Is”Os™Nr˜MqšNq›SpšPpœMp�LoœKn›Jn›Jq�Nv¢Uz¦[‚®cªbv¤\h•PZ‡DVƒ@\‰Fe�OcŠQY€IRyBhaQhaQhaQhaQhaQhaQhaQhaQg`Pg`Pg`Pg`Pg`Pg`Pg`Pg_Rh`Uh_Vh_Vh`Uh`Uh`Sh`ShaQhaOhaNhaNhaNhbLhbLhbLhbLg_He_Ge_Ic^Ha\HZXCRQ=KM8HI7DG4@D3=D4>E5>G6>F7=E6EL<HL;AF2JI4[Q8aE/�J6½j\Üg^àUPâGEãA>ß@:Õ>5É=0Æ=-Æ5$Ê4#Ì6'Ê7%Í7(Ì9'Ð:+Ð=+Ñ;,Ï<*Ñ;,Ñ;*Ñ;,Ñ;*Ñ;,Ì=-º2&¶8,°<-¥:(š7"‘5“7œ9"¬B,±<*¸6(À3)Ê6,Õ<4áD=çHDçCDæFHÛCBÉ:6º5.´8.®8.¦4)©9-§8-£9,ž8*š8+–:+‘;*‘;*–=+–>*™>+˜@*š?*›@+�A, @* ;'¤9'¨7'®6&µ7)½9-Ä;1Ë;2Ú=4Þ=3à?5âA7àB7ÛA5Ø@3Ô@4×H:ÐF9Æ@5»;0°5-©2*¤1*ž3)™9-’;*�:+Œ9'ˆ9(‡:(„;(ƒ<*~:'|9&z9%z9'z9'z9'w6$w4$w- Š@3—OA“N?ŽL>‘RC“UH�RD—[P—[Q�TL�TL—ZU”WR“VQ“hX\S2HP'AK&@K)>H-5A+1<+3>.1<.EPBX^PU\JLO:CG.@@&<<$9:(<?.:</7=/9?12>*9G-Nc<bzHjˆJs”Oq—Lp–Kp™MpšRpšPp�LoŸKoŸKožMpŸNs¢Rz¦Y~ª_‚­ew¢[i”O_‰G\†D`ŠJdŽNg‘RcŠSZ�LSzEhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQh`Si`Yi`[i`[i`Yi`Yi`WiaViaTibRibPibRibPibPibPibPibOh`Kg_Hf`Je_Ic^J^YEUR?ON:IJ8EH5AE4>E5>E5>G6>F7=F5?H7=F3=G/FL2MI.S?&}P9»raÉcUÞ_XéUQåGDß@<ß@<Ø?9Í9/Í9+Í<+Î<-Î=,Í;,Ì;*Í9+Í:*Î:,Î;+Î:,Í:*Ï9+Ï9*Ï9+Ì:-É?5ÄA7¼B5®=-�6#•2•2›2¥4"±6&Á;/ÐB8ÝF=äJBèKDèKFÝB@ÚDCÍB=»;2¯9-©=0£?/�=-�@/š>/–?.“=.�>-Š?,„@-†?+�<)‘;*‘<(’=)“>*”?+–?+—?+�@.Ÿ>-¢;*¦8)©8*±8-´:-¼:-É;1Ï;1Ò>4ÕA5ÕC6ÔE7ÒD8ÎD7ÇA5Â@2¹=1¯9-§7,¡5+�5,™7,‘;,‹<+‰<*ˆ;)†;(ƒ:)�=*€=*|;'y:(y:(x9'x9(v9't7%v4$}3&�D7—OA‘L=ŒJ<ŽPA‘SF�QF’RI—WN•TN–WP›\W“VQ�TM‰bQ]W3Xc8^i?ZfBGU46C'4@(<H4LXDWcO]fQT\EHN4AE*=@#;;!8:%;>-9;.7=/9?12>*:H.Nc<bzHk‰Kt•Pr˜Mq—Lp™MpšRo›NpŸNoŸInžJn�Lq Pu¤V|¨]~©a{¦_ošUa‹I\†D_‰Ic�MdŽNbŒMb‰RY€KRyDhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQh`Ui`[i_]i_]i`[i`Yi`WiaViaTiaTibRiaTibRibRibRiaTibRjbOiaLh`KgaKe`La\HYVCSR>KL:GJ7CF5@D3?F6?F6>E5<E4;G35A+>H0GK2D>$U@%…[Cµr_³VGØcYîe_ëSPåGFéGEæDBØ;6Ë7+È9)É:,É:*È9+Ç8(È6)Ç5&Ê8+Ê8)Ê8+Ê8)Ë7+Ë7)Ë7+Ê7-Á2*¾5/¶6-¬3(¡0 �/Ÿ1 ¦3!»@1ÇC7ÔG=àKDæMGçJEåFBÞC?ÖB>ÎC>Á@:±;1£<-žA0šE1–E2“E1�C1ŽC0ˆA/„A.A,|B.~@+„;*ˆ9(‰:)‰:)Š<(‹=)�<)�<*•@,–=+˜;*›:*ž8)£9+¦:-«:,¸:.¼:-À<0Â>2Ä@3Ä@3ÃA3ÁB3µ:+¯9+ª9+£7*�7+™7,–8,’:.‹<-ˆ<,…<+ƒ:)‚;):*~;*~=+|;)z;*y:)w:(u7(r7'q6&t5&†<1“F<•MA�I<ˆH<ŒNA‘SH‘QG‘NF˜SL—TN—VP™\W‘ZSŒZS�lY€~X}Š\~‹_ixOJX45E!;J+L[>cqWboU[eLMV;CI-?D&=@!:="79$:=,8:-7=/:@23?+:H.Mb;bzHk‰Ku–Qs™Nr˜Mp™MpšRo›Nr¡Pn IlžInŸMt£Sy¨Zz¨_z¨`ošUdŽL[„B[„BcŒJh‘Od�K^†Hb‡QY~JRwChaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQh`Ui`[i_]i_]i`[i`Yi`WiaViaTiaTibRibRiaTiaTiaTiaViaTkcPjbMjbMjbOhaNd_K_ZGYVCPO=KL:GH8CF5AE4AE6?C4<C1<H27C+DI2KG.M;#oL6˜dO§^M @2ÅSIçc^ï`\ëSRéKLèEHßA@Ë80Å9*Å7+Å7)Æ8,É:,Ê:/Ê;-Ð=3Ð>1Ñ>4Ñ?2Ò>4Ò>2Ó?5Ò?7Ó?;ÑB>ËD>ÅE<ÀF9ÁI;ÇM@ÎPDØRGÝPGáLEãJDãFAáD?àA=ÚA<ÔE?ÈC<º?7ª</Ÿ>.˜C/•G3‘H5�F2ŠF3ˆD1ƒB0~A.{@.xA-z@,:*„8*…9+…9+†:*†:*†:*‡;+ˆ;)‹:)Œ9)�8*“:*—;,š=.¡=.¨9.«9.®9/¯;.°:,¯9+®8*¬8)§6&¤6'ž7(š6'”7(�8*�:-�;-‡;+ƒ<*‚:+:*~9)}:)|;)|:*z;*z;*x:+t9)r7'p7&p7&t6'‰C7‘I=�K>ˆI:†H;‹PB“UJ—TK•PI›TN—RM”SM•\U�_U�g]•j’”mƒ“dr‚UXi=BS)>N'O_:csOdsRZiJLY;AL.<E(=C'<@%;=%78&9<+79,6<.;A34@,:H.La:ayGjˆJt•Ps™Nr˜MqšNq›So›Pq OmŸJmžLr£Ry¨Z{©^w¥]rŸZa‹I\†D[„B^‡EeŽLh‘Od�K_‡Ha†PX}IQvBhaQhaQhaQhaQhaQhaQhaQhaQibRibRibRibRibRibRibRiaTjaZja\jaZjaXjbWjbUjcSjcQjcQjcQjcQjcSjbUjbUjbWjbUlcRkcPkcPkbQjcQgbOb]J]ZGURAQP>LK9GH6EF4BE4@C2=B.<J1=G,KG.S>)d>+“ZI«_Rš>3ž3+¶?7×XRñjfõedéSTäIMçNPØIAÐF9Ì?5Ê>1Í@6ÒD8ÔF<ÕG;ÚJAÚJ?ÚJAÛK@ÝJBÞKAÞKCßJCéTPçSQãTPÜSM×RK×RKÚTKßTMÝMEáIDãEBâC@áB>âC?âE@ßGBØIEÉD?¹>7«<1Ÿ?1–C1’E3�H4‰E2ˆC3†A1ƒ@0~?.|>/y>.{=.9-‚8+‚:,‚:,€;,�<-�<,�<,�9*‚9*…9)‡:*‰:+Š;,Ž=,’<-˜:.›9,ž9-Ÿ9+ 9*¢8*¢8(¡7'ž7&›7'—7'“7(�9*‹:)‰:+‡;-ƒ;,€;+:+}8)}8){8(z8(z8*x9*y:+x:-t9+q6(o5'q7)t9+‹I;ŒJ<ˆJ;„I9…K=ŒRD’VK—WN–QLšUP—SP’UP‘`YˆbW‚hY†zbmsOZi>IX-AR&HY-Wi?dvLi{SXjDL\8=L+8E':D+<D,;@,9<)78(8;*68+6<.<B45A-9G-J_8^vDh†Hs”Os™Nr˜MqšNq›So›Pn�MmžLp¡Ov§V|ª_y§^p�Xg”QYƒA[…E`‰GcŒJgŽKgŽKeŒId‹L`…OW|IPuBhaQhaQhaQhaQhaQhaQhaQhaQibRibRibRibRibRibRibRiaTjaXjaXjbWjbUjcSjcQjcPjcPjdNjdNjcPjcQjcSjbUjbUjbUlcRkbQkbQkbQkdRibPf_Mc^K\VFVSBQN=LK9IH6GF4DC1?B/:H.?H+NB*`=*ƒE:«XP´NJœ,(²;7¹>9ÒROñpküvuñefêY\î^^âUNÙODÑG=ÍC8ÑD;ÔG=ÖG?ÖH>ÙIAÙI@ÚJBÚJAÝJCÞKCÞKDàKEáGGàFFÜHFØGDÖHDÕGC×FAÚEAÝB>ãA?æBAèBBæB@âC?ßB=ÙA<ÚIFÉB>¹<6¬=2¢@3—A0�B0�D3‰A2†A1†>0…=/€</=/}=1=1€:0€:0€:0;.€<1=/=/=-=/=-<,€;+:+:*€8)�8)‰:-�9.�:-’:,—;,™<+�=-�>,˜8(–9(’9)�:(‰:)†:*ƒ:+‚:+:+~9*}8)|7({6'z7'z6)y7)u6'v8)w9,s8*o5'm5&q9,v<.ŠL?†K;„K:�L:„Q@‹VH‘WL–VM“NI™TO–UQ“\U�d[z^PdWD`^EEO,>N'>O%J\2^qDiQcyKVl>M_5@R,6E$5D':F.<E28>06:,79+9<+68+6<.<B45A-7E+F[4Yq?d‚Dp‘Lq—Lr˜Mr›OrœTpœQmœLnŸNs¤Sx©Zz¨_s¡Yg”Q]ŠGX‚B\†FcŒJf�Mf�Je�Ge�Gf�L_„NV{HOtAhaQhaQhaQhaQhaQhaQhaQhaQibRibRibRibRibRibRibRibRjbUjbUjcSjcQjcPjdNjdLjdLjdLjdLjdNjdNjcQjcSjbUjbUlcTkbSjaRkbSlcTjcShaQf_OaZH\WDVQ>OL9LI6IF3FC0AC.:G+@F*U@+wF7¢OI²HH´7;¸36ÈBCÃ>?ÊJIßb^ïrpînmæabâZZÚQKÔKAÎE=ÌC9ÏD=ÐF<ÐC<Í@7ÑB<ÒC;ÓB=ÔD<×C?ØE>ØD@ÚEAÜ@AÝADÝCCÜFEßIHßIHâHFäFEæBCé@CëABëABæ@@á?=Ù>9Ô<7ÚFDÈ=:¹83®<2¦@4˜?/�?.�@0‰=/ˆ<.ˆ:.†9/†91„:1ƒ:3„;4€:2;2;2~<0~;2|<0|<0|=.{=0{=.{=.{<+z;,y:)z8(y7'9-‚:.…9,‡9,Œ:,‘;,“<+•<*‘8(�9(�:(Š;*‡;+ƒ<*€;+~;*~9*7)~6({6'z4(y5(y5*x6*s3't6)v8-r6+n4(m5(s;.x@1‰QBƒN<€M<€Q?…VFˆWH�TI‘RI’OI–SM‘TOŽ\S‡eYl[INL5DJ.@N-JZ6WgB_qIezObyKUm=E\.AV+5J#/@ 4B(;G3;C67<57:39;.:=,68+6<.<B44@,6D*CX1Um;`~@mŽIp–Kq—Lr›OrœTpœQnŸPo£Qr¥Vt§Zq¡Wi˜P_ŒIY†E]‡G_‰Ib‹Id�Kf�Je�GdŒEcŠG^„KUzGNs@haQhaQhaQhaQhaQhaQhaQhaQjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdRkdQkeOkeOkeMkeMkeMkeMkeOkdQkdRkdTkdTkdTnbTmaSl`RkbSlcTlcTibRg`Pd]M^YFXS@SN;OJ6LG3ID0CC+>G*DD([B,ˆTF­]V¨>>ª/2ËHMÄ?B½;=º:9ÃFDØ[YâecÛYYËGEÐIEÎE?ÌA<ÎC<ÑF?ÒG@ÐC<É?5ÎC<ÏD=ÑD=ÓD>ÕD?ØDBÚDCÜDCÞ?CÞ@AÞBCßEEàFFßECßA@Þ>>çACè@@ç=>å==ã==â@>áD?ßFAÖB@Å:7¶5/¯:1¦@4—>0�=/‹?1…<-…;.†9/†91†91‡81ˆ94‡:4‚92;2;0;0~<0};/{;/{<-w9,w9*x:+y;,y;,y;,y;,z;,};/}9.~8,7)‚6(„6)†7(†7&ˆ9*ˆ9*‡:*…<+�<,€=,~<,};+|8+{7*{5)y5(x4'w5)w5)u5)p2%s5(s7,q7+m5(n6)r=/xC5ˆUD‚Q@P@‚SCˆWI‹UIŒRG�NE”SM“TMˆRHƒYM}cRc[DGK0<I+HW6YjFezScxOYnCQh<Mb7I^3:L&/@+:3?'<E4<B6:=69<5:</9=.59+6<.<B46?,5B(DT0Si:azAm‹Kp”Nq—NršQrœRq�RqŸTs£Wu¥[r¡YlšRc�K]‡EYƒCbŒL`ˆI_ˆFaŠHf�Jg�Ic‹E^…D^„KTyFMr?h_Pi`Qi`Qi`QjaRjaRjaRkbSjaRjaRjaRjaRjaRjaRjaRjaRjcSjcSjcSjcQjcQjcQjcQjcQkdRkdRkdRkdRkdTkdTkdTlcTk^Np`QqdTpdTlcTi`Qh`ShbRgaQebOa^KYWBVQ;TO9OI1GC(JF)I?$bJ2‘hV´xmºme·YW¹OOµCCºBCÁCDÆFEÌGHÐJIÑKJÓKKÎDBÏCBÓECÔFBÓF?ÐC<ÎD:ÎF:ÇC7ÇE8ÊE<ÌF=ÒD@ÖBBÙ@CÛ?CàAEàBCáCDßDBàECàECáFDãEBâC@äB@ãA?ã?=â@=â@=àA=ÜC>ÒC=Ä?8µ91¨8-Ÿ:.•=/Œ>1„?0�?1?3€>2€<1ƒ:1†91ˆ81‡81…;0ƒ;-ƒ;-�<-€;,€;,<,<,~:-~:-};-};-};-};-};-};/{9-|8-|8-}7+~8,~8,~8,~9*}7+}7+|8+{9+z8*x9*x9*v8)x8,v8+w7+u7*t6)s5*s5*q5*k1%p6*n4(k3&o7*j5'n9+}I;ƒSEƒSG„RG‡QG‹PH�PI‘PJ’QK’SLŠTJ†ZMbPm^GVU9KV4M^:jXaxNYpFSj@Qf=Mb;GX4BQ06C%2?#0:"3;&;?.?A3??3<>339+2:+39+5;-7>.9A,;C,;F(Q`9_rDn‡Pr�Rq”Pq™RrœRržSužX}¦b}¦dp™Wb‹I]†D]…F\„E^†GaˆId‹LeŒMeŒKb‰H_†E]„E^�KSuBKm;h_Pi`Qi`Qi`QjaRjaRjaRkbSjaRjaRjaRjaRjaRjaRjaRjaRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTlcTqaRqaQobRpdTpgXohXlfXgdUgdUkhWmjWjhShcMc_F[U=RL2PK.F?#P@'kP;XGƒOB†F<’E?°WS¶TQºQN½PM¿NJÀMJÂLHÃLHÏSQÇGFÁ=;Ä=:ÎC@ÑF?ÌB8Æ=3ÈD7ÈF9ÊF:ÍG>ÓEAÖDDÛBEÝAEÞBEàBCáCBßDBàECßECàEAáFBßD@ßD@àC>ßB=ÞA<ÞA<àA=ÜC=ÒE<ÇB9º>4«;/Ÿ:.“;-‰=/�>.@1|A3}?2=1�;1†91ˆ81ˆ8/†:-„;,ƒ;,ƒ;,‚:+‚:+€;,:+:+:+~:-~:-~:-~:-};/};/|8-|8-|8-|8-{9-{9-|:.z:.z:.z:.y9-x:-x:-w9,w9,w9,t8-t8-s7,s7,r6+p6*p6*o5)j2%o7*m5(k3&l7)j5'n9+}I<ƒSGƒSI†QI‹PJ�PK“OL”QK’SL‰NFƒSG[KzdMqjPhkLcqNfxR]tJUnDOe>Ka:J^9J[7FU6BQ4;E,6@(3;&6;';>-?@0>?1;=03;.1<.4:.5;-9=.;@,<A+;D)Q[9^mDk�PpŒRo’Pp˜Qs�SržSwž[{¢cxŸ`k’S`‡H]„E]„E\ƒD_†GaˆIcŠKd‹Ld‹Lb‰J_†G]ƒF\IQs@Ik9g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTlcTpcSo_Pk]Pl`RndXogZkeYeaUfbVom^zwh~{jzgzuaojVcaJRU:DG,@>%LA+SA-S8']7*l>1n8.t8.~;3‡A9‘H?›QH¤WM¬YQ»^WºSN·GCºC?ÅCAÏEBÑC?Î?9ÏB9ÐC:ÏE;ÓF?ÕF@×FCÚDCÜCEÞBCÞBCßCDÞDDÞDBÞFCßECÞFCÝE@ÜD?ÛD=ÚC<ÚC<ÚC<ÜC=ÙD=ÔE=ÌF=ÁB9²=3¡9.‘8*‡9,�<-@1~@1~?0=/�;/„:/‡9/‡9/„:-ƒ;,ƒ;,‚:+‚:+‚:+:+:+~:-~:-~:-~:-};/};/};/};/{8/{8/{8/{8/{8/{8/y9/y9/z:0y;0y;0x:/x:/v:/u9.u9.t8.s7-r8-q7,p6+p6+o5*n5*j1&o6+j4(i3'l6*i3'm7+}G;†RG†RGŠQHŒQI‘RK’SL“TM‘VN‰WN†\P~^OsaKoiQnsUkvVevTPd?I_9DX3BV1BV3FW7ET7DR8?I0:D,6>)7<(9=,;>-;=/8<.5=22<16<26<09=.:>-<A+:C(MW5Yh?f|KlˆNm�Np˜QtžRt Uz¡^wž_n•Vb‰J\ƒD[‚C\ƒDZ�B_†G`‡Hb‰JcŠKcŠKaˆI^…F]ƒFX}GMr?Di6g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTlcTqdToaTl`Ri_SiaTjdXieZhfZig[wuiˆ„x��~•��’Œ|ˆƒp{zf^cLLT<CG0DF0FD/E=*L<,UB3O8(R6(T4'V4(\6)`8,d<0m=1ŠLAœRI­VOµOK»ECÃAAÐDC×GFÑ@;Ô@<ÔD<ÕE=ÙEAÙEAÛFBÝEDÝCCÝCCÞDDÞDDÝEBÝEBÝEBÛFBÛF@ÙF?ØE=×D<×D<×D<×D<ÕE<ÔD<ÐF<ÈE;ºA6¦</•8)‹9+„<-�?/@1~?0=/�;/„:-‡9/‡9-„;,ƒ;,‚:+‚:+‚:+�9*:+:+~:-~:-~:-~:-};/};/};/};/{8/{8/{8/{8/{8/x8.x8.w9.y;0y;0y;0v:/v:/u9.u9.u9.s7-s7-q7,p6+p6+o5*m4)m4)j1&m7+j4(h2&k5)h2&k5){E9ˆOFŠOGŒQIŽRJ�TL�UM�WN‹YNˆ^RƒbSv_MiZGc^HdhO\fKN]@EU8@Q1<L/;K.>N1BP6ES9GS;@L4>G29B-7?*7>,7>,8<-4:,5=23=46;46<29=/:>-;@*9B'IS1Ud;bxGi…Km�NršSv Tv¢Wz¡^q˜Yd‹L[‚CY€A[‚C\ƒDZ�B^…F_†GaˆIaˆIaˆI_†G]„E\‚GW|HLq>Ch5g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTlcTrfXsgYqg[ldYhaWieZrnezwn€}v��…Ÿœ“§¥™ª¦›¨¤˜¢œŽ˜•†x{hbgQMR<EJ4CH4@E1AD1DG4KL:HI7ED2A>-@:,?7*@6*F4&R5'oC8‹PHžOJ­IG½FHÊEHÑBDÕAAÖ@?ØC?ÙD@ÚE?ÚE?ÜD?ÜDAÞDDÞDDÝCCÜDCÜDCÜDCÚEAÙEAÙF?ÖF>ÕF>ÔE=ÒE;ÒE;ÓF<ÔF<Ó@8ÓC;ËE:¿C7­?2œ:-‘;,‰=-ƒ>.�?/~?.=-€=-ƒ;,†:-†:,ƒ:+‚:+‚:+‚:+:+:+~9*~9*~:-~:-};/};/};/};/};/};/z:0y9/y9/y9/x8.w9.v8-v8-x:/v:/v:/u9.u9.t:.s9-s9-q7,q7,p6+p6+n5*m4)l3(l3(g1%l6*i3'f2%i5(f2%i5(|C8ŠND�NE�PI�SKŽULŒXM‰YOƒ[OyYLt]MiZG\UBYXCY^GNX@?K39G.7E,4B+4B+7E.<H2@L6CO9@L6=I59E19B/7@-5>+4=*2:+3=42<45:35;18;09=.:?+9B'GQ/Ra8_uDh„Jo’Pu�Vx¢Vw£XxŸ^k‘T]ƒFW}@YB\‚E\‚E[�D^„G^„G_…H_…H_…H^„G]ƒF\‚GV{GKo?Bf6g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTkbSndXnf[jcYhd[pmfƒ€y��Š¤£Ÿ²±¬ÁÀ»ÈžÇÄ»ÄÀ·¿¸®¸²¦¢œŽ…‚qedROP>DI5<E08D.8E18E17D07D28C3;C4>D8BE:GE9<4'M:,d?6~E>¡PO¿XYÊQVÉDGÒFGÖFFÙECÚFBÝEBÝE@ÝD?ÜC>ÞDDÞDDÜDCÜDCÜDCÛCBÙD@ØD@ÙEAÖF>ÔE=ÒE<ÐF;ÐF;ÏG;ÓF<Ó>7Õ@9ÎD:ÄD9³A6¤>0•<,‹<+…>,‚?.€>.>,€=,�<,„;,„;,‚:+‚:+‚:+�9*:+~9*~9*}:*~:-~:-};/};/};/};/};/{;/z:0x:/y9/w9.w9.v8-u7,s7,u9.t:.t8-s9-s9-q9,r8,r8,p6+p6+p6+n5*m4)l3(l3(i3'g1%k7*h4'f2%h4'e1$h4'{B7�MD‘NF‘PJŽSK‹WL‡XN€ZMy\Ns\Lk[K_XFUT@PS@JR=@I64@,3<+1:)09(.9(0;*3>-7B19D3=H7<H4<H4:F28D04A-2>*0;+0:10943824927:18<.9=,8@(CL-O^7]sDhƒLp“SvžWx¢Vw¡Wp—Xc‰LW}BV|?YD[�D[�F[�D]ƒH]ƒF^„I^„G]ƒH\‚E\‚G[�HUzGJn>@d6f`PgaQgaQgaQhbRhbRhbRicShbRhbRhbRhbRhbRhbRhbRhbRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTiaTicWjcYkg^spi…„ž�™¯°«ÁÁ¿ÑÑÏààÞæåáæãÞáÞÙÛÖÐ×ÎÇÌ»±¶¤––†wvm\_ZGKM8>F/;D/=J6<I5;F5;C4;>3:;3983880<>1A=1L8/_81†IH¯^]Å`dÈWYÍQSÑOO×MKÙKIÞIEßFAáC@áB?ßCDÞDFÝCEÝCEÜBBÛCBÛCBØC?ÙEAÖE@ÔE=ÒE<ÐF<ÐF<ÏF<ÒE;×@9×@7ÑC9ÈD8ºD8«A3˜<-‹:)†=,ƒ?,€?->,€=,�=*ƒ<*ƒ;,‚:+€;,:+:+~9*~9*}9,}9,~:/};/};/};/{;1{;1{;1{;1y:1w;1x90u9/t8.s7-s7-q7,r8-q8-r8-p7,p7,n8,o6+o6+o6-o6-n5,k5+j4*j4*i3)h4)f2'k7,h4)c1&e3(b0%e3(zA8�ME’MF‘PJ�TK‡WMYLwZLo\Mk^N]XERQ?KN;BI79B/2;*08)06*06,06*.6).6'/7(2:+4<-9D4:E4<I7<I7:G35D/2?+0=,-7/-6116027058/7:/9=.8@)?H+KY5[qBg‚Kp“SvžXv Vs�Ug�P]€FUx@Wz@[~F[~D[~F]€F_‚J_‚H_‚J^�G^�I]€F]€H]€JSwGGk=>b6f`PgaQgaQgaQhbRhbRhbRicShbRhbRhbRhbRhbRhbRhbRhbRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkcVoi]kg\lh_spi„ƒ~Ÿ ›¼¼ºÍÏÎÍÏÎÞàßñññøøöø÷õöòïðëèíâÜæÐÅÛÁ´Á¬› ‘~}u`^\EJN7AI1@J2@I4@G5BD7AA7B=7B:7>;67=3=@5B71H/*c74ŽRQ³ggÆnm¿[[ÆYVÍVRÕQMÛMKàHEäEBåBCßCDÝDFÜCEÝCEÜBBÜBBÛAAÚBAÙD@ØD@ÕD?ÓD<ÑD;ÐF<ÏF<ÒE;ÛB<Ù@8ÒB9ÊD9¿E8­A4š:*‰6$ˆ<,ƒ?,�>->,€=*€=*�=*�=*€;,:+:+:+~9*}:*}9,|8+};/};/};/};/{;1{;1{;1z<1w;1v<1v:0t:/t8.r8-q7,p7,p7,p7,p7,m7+m7+m7+l6*l6*m7-l6,l6,k5+j4*h4)h4)g3(f2'i7,e3(c1&e3(a/$e3(wA7ŽMG‘NHŽRJŠUMƒYMzZMq[Mh\LWTCHI9?C2=D49B108)/7(6<047047058117-06,/7*08+19,6A38E4<I7=L9:I46E02A,0>-*5-,50/4.16/4707:/9=.8@+<E*IW4YoAf�Lp’UuœYt�Wo˜T`†IX{CRu=VyAZ}EY|DY|D[~F^�I^�I^�I^�I]€H]€H]€H\~KStGEh>>^5icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUkdTkdTkdTjdVhdYee]ttl‘’Œ®¯ªÃÅÂÙÛÚéíîñõöôø÷ûýüÿÿýÿÿýýüøúöóûòë÷áÔòØÇÞɴŵž©Ÿ„ˆ„gbbFEH-DH/BF/FE1GD5E<3C60D52F;9>?:>A:B:7G83bFC�jgÄ“�符ъ„Çtn¿^WÅTNÔPNÞLLçGIêGJÝDFÞIKÚDEÔ;=Ø>@àFHßEEÕ;;Ù??ÙA>×B>ÖB>ÓB=ÑB<ÏB9Ð@7ÜC=Ú@8Ò?7É@6½A5®@3�<,Ž8'Š=-†?-‚=-€=,~=)<)<)€<):+~9*}9,}9,|8+|8+|8-z8,{9-{9-y9/y9/y9/z:0z:1y:1w=2v=2v<1s:/s9.q8-p7,m7+m7+m7+m7+l8+l8+l8+l8+l8+g3(k7,o;0n:/j6+g3(f2'e3(c1&h6+c1&]-!_/#^."b2&q?4‹OG�PK‡PI�RJzXLu^PeXHSM=BE2=D29B14?/2=-3;.2:-39/68357257227016/06,.6+.6)-8*0=,6C19H5:I45G11C-0>-(4*)2--2,.3,25.58-8<-8@+6?$DR/WlAg‚Or“Zv�\r›Wj“QX~CVyCSv@Sv@VyCY|FY|FX{E]€J`ƒM`ƒM\I[~H]€J_‚L]MRrIDd?9W3icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUleUleUleUjdVee]noi…†�£¥ À¿ÕÙØéíîõûûöúû÷ûüùýüýÿüþþüÿþúÿüùÿûóÿðàúêÑìÞÄ×αÀ»�¤£„€„cflJLQ3FI.CC+GB/H>2G:2J84J<9>:7CB>NIF\RPyje£Ž‰Í²«çżÿ×Î벧̓zÀc\ÉXT×RSÞIKÚADÔDDÓEDØHGÜJJÚFFÓ=>Õ=<ÛAAØ>>Ø@?ÖA=ÖA=ÕA=ÒA<ÐA9ÑA9ÛB<Ù@8ÑA8ÊA7¾B8°@4ž</�8*‹<-†?-‚=-€=,~=+~=)<)<+:+~9*}9,}9,|8+|8+{9-z8,|:.|:.z:0z:0z:0z:0z:1y:1v<1t;0t;0p:.q8-n8,m7+k7*l6*k7*k7*i8*k7*i8*i8*i8*j6+l8-n:/n:/k7,f4)e3(d2'b0%c3'^."]-!b2&]/"_1$o?5JB†QK†WQ}WNqUJdRDPI9>?/9@.5@/3@/2?.2=/3;.4:069268368557246116/16//7,/7,-8*/;-4A07F39H56H23E/3A0,8.,6.-2,-2+03*36+6:+6>)6>&DR1VkBh‚Ss”]vœ_p˜YfŽOV|CUxBRu?Ru?UxBWzDX{EWzDZ}G]€J^�K[~HZ}G\I]€J[{LPnJD_@7R3gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUleUleUleUieYde_y{x—™–²¶µËÑÑãééóøû÷ÿÿûÿÿûÿÿûÿþüþûþÿúÿÿúÿÿúÿÿôýûäùøÚïðÑàäÃÑ׳½Åžž©�‡’j^hEPW6CG,DC.HE4LD7J@7H?8C;8QLIia_wtœ’�½²¬ØÎÅéÛÐüéÛÿôçÿäÙï«¢Áfa´DCÆEIÙSTÚVTÎMHÉC@ËDAÑFCÕEDÕCCØBCÖ>=Ö>=Õ?>Ö@?ÕA=ÓB=ÑB<ÒB:ØA:×@9ÑB:ÊD;¿D<°A6¡<0’:,‹<-‡>/‚=.€=,~=+~=+~=+~=+~;+}9,}9,}9,{9-{9-{9-z8,{;1{;1{;1{;1z;2z;2y:1w;1s:1q;1p:0n:/n8.l8-k7,h6+k7,i7,i7,h8,i7,h8,h8,h8,k90k90j8/j8/j8/i7.e3*b2(c3)`1'[,"_0&c4*Y-"\0%pD9}PJWO}XPoSH^J?OC7?<-46(1:)0;+0=,1=/2<14<15;169268368368338227027008-08--8*.:,2?.4C07F38G46G48F52>20:1/4-,2(-0'03(37(4;)6>'DQ3XlGk„Zv–dw�dl”X_‡KUxBTtBRr@Rr@TtBVvDVvDVvDWwE[{I\|JZzHZzH]}K[{IXwKNgG?W=1I1gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUleUmfVmfVjfZikf‚ˆˆ£§¨¹¿¿ÍÕ×åíïòúüôüþùÿÿùÿÿúÿüúÿùüÿøýþöÿÿõÿÿó÷ýãóüÝîøÖåñÍÝéÃÎÚ²µÁ™ ¬„}‡bfmKLS4DH/HJ5KI:GE9D@7OGDia^Š‚€¢š—·¯¬ËÆÀÝÚÑéåÚùóçÿüïÿ÷ìÿÝÕú·±å��Çcc®@?ÁPJÈSJÇPHÂG@ÇF@ÒMHÕJGÐ@?Ï>;Ò><Ó?=Ô@>ÓB=ÓB=ÒC=ÒA<Ô?9ÔA:ÑD;ÊE<¾E<°A8 >3”:/‹</‡=0‚=.€=-~<,~=+~<,~<,~:-}9,}9,}9,{9-{9-{9-z8,|<2|<2{;1{;1z;2y:1y:1w;1q8/n:/m9.l8-k7,h6+h6+g5*i7,h8,h8,h8,h8,g9,h8,h8,k;1i7.g5,h6-j8/h8.d4*`0&c4*`1'].$`1'`4)V*^2'vND~\SuXPgNGWC:H<0@:.:;-7;,/7(.9)0;-2=/4<15;169058/69249249238138119.19.19,-8*.9+/<+2?.5B07F39H5:H78D64?117+,2&,/$-1#15&18&4<%ER8]pPqŠc{›lv›heŒWT{DQs@Qq?Pp>Pp>Qq?SsASsATtBUuCYyG[{IZzH[{I]}KYyGVsGH_C:M9,?,gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSibRjcSjcSjcSkdTkdTkdTleUkdTleUleUleUmfVmfVngWkg\pro…��¤©¬·¿ÁÌÔ×ãíïñûüóýþ÷ÿÿøÿþúÿúúÿøûþóùûîøùëöùèõÿçóÿäðýßíøØèóÑÞèÅÊÒ­¶¾™›£~{‚aX^BGL5FK7DK9@F8@B7b]W…~x«¡ŸÀ¸µÊ¿ÕÐÌäáÚîîæùöíûôêþðåÿóéÿôìÿåß騤¶ic¥K@²PC¿YKÇ[OÃQG¾E=Á@:ÊE@Ê@=ÌA>ÏA=ÐB>ÐB>ÑC?ÐC<ÑB<Ò?8ÒB:ÏD=ÇG>»D<­A7Ÿ=4”<2‹;0‡=2‚<0€</~<.~<,~<.~<.~:-}9,|:.|:.{9-{9-y9/x8.|<2|<2z;2z;2y:1y:1x92v:2n8.k9.j8-i7,i7,h6+g5*f6*g7+f8+f8+f8+f8+d8+f8+f8+j:0g7-e5+e5+h8.h8.d4*`1'a2(b3)\0%\0%\0%V*a7+zVJqXQ_NFL?6@6,<6*:8,9;.9=/19,2:-4</4</5;/69.58-47.39/2:/2:/2:/19.19,19,/:,-8*-8(.9)/<+2?.5B18E4:G6;F67B24:,/5'.2$.2$04%/6$19$FR:buWvŽj{šqq•g\‚QJp=Mo=Nn=Oo>Oo>Oo>Oo>Pp?Qq@TtCXxGZzI[{J]}L]}LWwFQmDAT>3C6&6)gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdShbRjcSjcSjcSkdTkdTkdTleUkdTleUleUmfVmfVngWohXlh]jnmz‚…“›ž«µ·ÃÍÏÚåçë÷÷òþüóþúõÿúùÿùùÿöúþð÷ûìõöæñôáóüçóüçóûäóùßñõÚéìÍ×Ö·Ää°¯‘��qdfNMP;DK9EN=EPBIQF{{qžš‘Ľ·×ÎÉÝÓÑåÛÙðëçú÷òïîéûûóÿÿöÿþóÿøîÿóéÿæÝÿ×Ë×’‚·dR¢K8±TC½WI´E:¶@6ÈLDÄC=ÈC>ÊC=ÍD>ÍD>ÍD>ÌC;ÍB;ÐA;ÑB<ÌE?ÄE>¶A:©>6�<5•<4‹;2‡=2‚<0€</~<.~<.~<.~<.~:-}9,|:.|:.{9-{9-y9/x8.{;1{;1y:1y:1x90w8/w81t91l8-j8-j8-i7,i7,g7+g7+g7+f8+f8+f8+f8+d8+d8+d8+d8+h8.f6,e5+f6,g7-g7-f7-d5+a2(d5+^2'[/$`4)`4)g=1tTIUG>C?695,86*99/79,48*37)4:.4:.4:.4:.58-47,46+36+28,19,19,08+08+08+.9+.9+.9)-8(-8(.9)/<+2?.4A06C2:E48C28?/4;+37)15'04%-4$07%HT>dw[tŒls‘keˆ^RwKBh9Ij;Ll;Mm<Mm<Ll;Ll;Mm<Nn=TtCXxGZzI[{J]}L\|KSsBLf?:K8.;1#.&feSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSgdShbRjcSjcSjcSkdTkdTkdTleUkdTkdTleUmfVngWohXohXmi^fjiluz„�’¡¬°½ÈÌÒÝßäððïûùòýùóþöøÿ÷ùÿôûÿðùýìøùçö÷åõöèö÷éùöçûöãüôáöëÕæ×ÀÔƬ½±™›“|tmZXWENPBQWKYcZcmb•˜�±±¥ÓÏÆçÞÙñãâøèèÿññÿöõÿýúüû÷÷úóöùîúüïÿÿóÿÿñÿûêÿåÐºoZ§VC¬SAµQB¸L@¿JAÂG?ÅF?ÈE=ÉF>ÉD=ÉD=ÈC<ÈB9ÐC<ÏD=ÊE>¿C;³>7¥:2›:3”=4‹;2‡=4‚<2€<1~<0~<0~<0~<0};/|:.|:.|:.{9-{9-y9/x8.z:1z:1x90x90w81v70s7/r7/l8-j8-i9-h8,h8,h8,h8,f8+f8+f8+d8+d8+d8+d8+d8+d8+f6,g6/h70h70g80g80i:2k<4d8/i=4b6-`4+nC:vKBsH?kMCB9047.25,7:/8;247,14)25*58-47,37)37)35(35(46)37)/7(-8(-8(-8(-8(-8(-8(-8(.9).9)-8(-8(.9)/:*0;+1<+7C/7C/:A/9@.9=.59*15&,3#18&IU?cu[m„geƒ_WzRHmD>c7Gg8Li9Nk;Nk;Li9Kh8Mj:Ol<UrBXuE[xH\yI^{K\yIQn>F`95C2)4, )$feSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSgdShbRicSicUjbUkcVkcVkdTleUkdTkdRkeUlfVmgYnh\mi^jidjpplvx�Š�¤®°ÂÌÎÔÞàãíîñûúóü÷öýö÷þöùÿóüÿòûÿðüÿîüýíúøìýùîÿùíÿúêÿøèÿòßîßÈÝη¾²šŸ—€zua`_KVZK\dWkwk|†{¦¬ ½¿´ÙÖÍìåßøíëÿòòÿóóüòóý÷÷ùúõøý÷úÿùõÿôóûîõýîÿÿñÿùèÿôãü˺ºošRC§TF´TH°D8¾I?ÃG=ÇG>ÉF<ÈE=ÇD<ÆC;ÅB8ÌF;ÌH<ÆH<¹C9ª>2Ÿ:0•;2�>3‰<2„>4�=4€<1;0;0€:0;0};/|:.|:.|:.{9-{9-y9/x8.z:1y90x90w8/v7.u6-r6.q6.l8-i9-i9-i9-i9-h8,g9,g9,e7*e7*e7*e7*e7*e7*e7*e6,f5.j81l:3k:3i9/i:0l=3nB7kA5mE9c=2b>2xVJ…cWwWLbLA<902:/2:/5=24:0.4*06*6<025*25*04&04&04&15'37)28*.6),7),7),7),7)-8*-8*-8*.9+.9+-8*-8*,7)-8*-8*-8(4?.6A0;D3<E4;B26=-07',3#2;(JVBbqZg{`]vVOmIEe@@`9Ed8Kh:Mj<Mj<Kh:Jg9Kh:Nk=TqCWtFYvH[xJ]zNZwKMj>BZ81?2'0-(%heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheVhdXjdXkdZkeYldWmfVe`MjeQnkXmk\hfZed_qqo€�ƒˆ�’•Ÿ ª´¶ÁÉË×Üßèîîó÷øøúùüþûýÿúýÿùüÿöúýòøûðùýïúþðøúïøúïøøìûùêþüçøöÝèäÇÔÓµ»ºœ˜™zptY]eMbkVq~lƒ‘‚�œ�´¹²ËÌÆãâÝïîé÷óðû÷ôüøõû÷ôÿþúþÿúýÿúýÿúýÿùýÿùýÿùýÿùùûöÿÿúÿüöÿêäר  `V™E;³ND¿MBÂF<ÄD9ÊH;ÎKAÎKAÉI>ÆH:ÅH6ÂI6ºJ6±H5¥F4šC2�C3‹C5„B4�B3€@4€>2�;1ƒ9.…8.ƒ90~:/}:1|90{8/x8.w7-v6,v6,u7,v8-v8-w9.t8-r6+p4*m4)j8-g8.g8.g8.f7-f7-g7-f6,f6,f6,h8.h8.h8.g7-e5+e3*g2,j5/l7/k90i9-i;.i=0hA2eA1iJ8dI8u^N|kYˆxh‹~nQK=79,3;.2:-2:-19,19,19,19,17+06*06*06*/5)/5)/5)-5(.6+.6+.6+.6+.6+.6+.6+.6+.6+.6+-5*-5*-5*-5*.6+,6+0:/2>26@58C58C57B24?/3>-3?+DQ=WeN[kQPcGDW9@U6@X6Ic>Ie=Jf=Ie<He9If:Kh<Mj>SpDWtHZvM\xP\xRVrLFa>9P6$1($,.!)+heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgeVhdXhdYkd\kdZldYlfVidPjhQlkWjk[fh]jli{‚Ž“™¨°³¸ÂÃÎÖØßçéïóôøüýÿÿÿÿÿýýüøþýùþÿùýþöúýôøûòöüðöüðöùîöùî÷ùë÷úçøúäñôÙßâÃÌЯ´¸—‘–virUXbG_kUt�m‹™ˆ›§›¼ÁºÒÓÎçèãòóî÷øóúûöûü÷úûöþÿúþÿúþÿúþÿúþÿúþÿúþÿúúÿþôÿÿóÿÿûÿþÿüúÿîèïÀ¸Á~užL@®L?¹K>ÂL@ÅK>ÁE9»?3¾@4ÁF7ÂG5ÁJ4»L8±J7£F4˜B1�D4‹F7ƒD5�C6€@6€=4€<1ƒ:1ƒ90‚90}:1{;1z:0y9/w9.v8-v8-u7,s7,s7,t8-t8-r8,q7+o5)m4)j8-h8.h8.g8.g7-g7-g7-g7-c3)d4*e5+g7-h8.i9/j8/j8/i70i70k90j:.j</i>.hA0gD1fG3jQ;fQ>ufQ~vcŠ…q„ƒqIL;39+2:-2:-19,19,08+08+08+08+/7*/7*/7*.6).6).6)-5(-4,-4,-4,-4,-4,-4,-4,-4,.5-.5--4,,3+,3+-4,.5-,6--7..:01=34@65A56B45B14A05B0BO;P^GQ_HHV=AO5BP6EU8Ia?Ie?Ke@Ie=Hb;Fb9Hc:Hd;RmDVrJ[uN[wQ[tTRmNB[=4J5$1*$,/"*-heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgeVhdXhdYkd\kdZkeYlfVmhTkiRkjVijZjlauwtŽ’•§¬²¾ÆÉÒÜÞêòôõýÿüÿÿüÿÿÿÿÿþþüýüøÿþúÿÿúýþöøûòõøíñ÷ëñ÷ëóöëôøêõøçô÷âôöÞìïÔÛܽÆɪ®±’’—yuy`kpYt{iˆ‘€�¨š¬¶«ÌÏÈÞßÚïðë÷øóûü÷ýþùþÿúüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿüùÿÿùÿÿýÿþÿüùÿúôÿ÷íýÑÄן�’M=£RA¨K:¥?0·G9ÎZKÏUH¼A2ÃD5ÅG8ÃK=¸H:©C5�?3•A6‘D:‡A9„A8€@7~?6~>5}=4}=4}=4z;2w;1v<1u;0u;0t:/r9.r9.q8-q8-o9-o9-n8,m7+l6*k5)l8-j8/j8/h8.i7.h6-h6-h6-e5+e5+e5+f6,g7-h9/j:0j;1k<4i:2i:0h</i?1iB1iD2eF2eJ5jV>iXDujT…~k‘�z~mCF528*19,19,08+08+/7*/7*/7*08+/7*/7*/7*.6).6).6)-5(-4,-4,-4,-4,-4,-4,-4,-4,.5--4,,3+,3+,3+,3+-4,,6-*4+*6,,8./;12>24@26C27D37D2?L8ES<BP9;I08F-=K2CS8I^?Hb=Kc?Jd=Ha:E_8F_8E_8QjCVpI\tP\vSZrRNgI=T8.A-%0*$,."*,heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgeVhdXhdYkd\kdZkeYlfVkiTkkSlkWkn]nqf~ƒž¢¥¹¾ÄÄÌÏÛåçóûýùÿÿúþÿúþýüüüúúøþýùÿþùþÿ÷úüñô÷ìîñæèîâçíßíñâïóâñôáðóÞðòÚèëÐØؼÇÇ«±±•ž¡†�’|�“€›Ÿ�©­Ÿ¶¼²¿Ä½ÛÝØéêåö÷òûü÷üýøþÿúþÿúüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúþþüúúúÿÿÿÿþÿõñîûôìÿýñÿûíÿôäë°¶‚m“P= O<µXGµN?´D6ÀH:À@3ÇE8ÅI?¼F<®?6¡<4˜>6•B<Š?:…@;�@:|@8z?9x?6x?8x?6v=4u<1s=3s=1r<2r<0p<1p</o;0o;.l:/l;-k9.j9+i7,i7,l7/l7/l7/l7/k6.i7.h6-h6-k;1j:0h8.f7-e6,d8-e9.f:/j?6f>4c=0e@0hC3iF3eF2aF1aK4hV>i]GrjS…ƒn��yor_9=,17)08+08+08+/7*/7*/7*.6)08+/7*/7*/7*.6).6).6)-5(-4,-4,-4,-4,-4,-4,-4,-4,-4,,3+,3++2*+2*,3+,3++5,*4+)5+)5++7--9-0<03?15B15B19F4;H47E.2@)1?&7E,=M3DY:E];J_>Ia=J`<F^:H^:F^:RhDWoK`uT^vVZnRKbF9M4(9'$/)#+-!)+heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheVhdXhdYkd\kdZkeYlfVjhSjjRmlXnq`svkƒˆ„¢§ª¼ÃÉÍ×Ùåïñùÿÿùÿÿùýþúþýÿÿÿýýûÿÿûÿÿúüýõ÷ùîîñææêÜàæØÞåÕæêÙéîÚìïÚìðÙìîÖèèÎÚØ¿ËÉ°¿½¦¶´Ÿ±®›µ³¤¾¾²ÈÈÀÑÒÍÙÚÕèéäñòíúûöüýøüýøþÿúþÿúüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿþúÿýýÿûþþúùúûöüÿøüÿöþÿóÿÿíÿþêõßÈÄ›‡›bN“J7¤M:´P@»K=¾@4ÅB:ÅF?¾C<±>9¥96�97˜=:‘=;Š?<ƒ?<|?:w@;tB9qB:qB8p>5r>3r=5r>3p>5o=2o=4o=2m=3m=1k<2j</i:0h:-h9/i9/n70p70o6/m6/m6/k6.i7.i7.k;1i:0h9/d8-c9-d:.e;/d<0cA5bB5bB3bE3cF4cH3aH2]I1]M4bV<iaJrnU‚„nƒˆqYaL29'08)08+08+08+/7*/7*/7*.6)08+/7*/7*/7*.6).6).6)-5(,3+,3+,3+,3+,3+,3+,3+,3+,3++2*+2**1)*1)+2*+2**4++5,*6,)5+)5+*6*,8,/;-1>-0=,3@.5B03@,0=)/=&3A*6F,=O5?S7EX<H]>I\>I^=J]=J_>SfFZoPbuY`tYYkSI\F5G1$5%"-'!*) )(heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheVhdXhdYkd\kdZkeYlfVigRjjRlnYnq`pvj€†‚œ¤¦·¾ÄÕßáí÷ùùÿÿùÿÿøüýüÿÿÿÿÿýþùÿÿúÿÿ÷úüñóõèéíßâæ×ÜãÓÛâÐãèÒæëÔéíÖéëÓëëÓèæÍÜÚÃÒÍ·Ò͹ÌƶÉõÍɾÖÑËÞÛÖéåäïîìòóîøùôýþùüýøüýøþÿúþÿúûü÷þÿúþÿúþÿúþÿúþÿúþÿúþÿúÿýúÿûúÿ÷øÿýüþÿýøÿûòÿ÷îÿòôÿõïöäÿÿìÿíØ˧‘¢kW¡ZFªXC­L;¸D7¿C9ÀE=»D<³@;ª=:¢:7š;9”<;Œ><…?=}@;wB:rC9oD;oE9o?5p>5p>5p>5o?5o?5o?5o?5n?5m>4j>3i=2h<1h<1h<1k;1n70q60p70o6/m6/l7/i7.h8.f7-d8-d:.b:.d<0d>1e?2cB3^B4^G7bK9dN9aL7`L4]K3\L3]Q7`Y=lhOssY€„mt|dEN92>*19*19,19,08+08+/7*/7*/7*08+/7*/7*/7*.6).6).6)-5(,3+,3+,3+,3+,3+,3+,3+,3++2*+2**1)*1)*1)*1)+2*+2*,6-*6,)5+(4*)5)*6*+7+-9+,8*/<+2?.3@.3@.2?+2?+3A*6E.8J0>P6BU9EX<GZ<J]?L_AReG[nRbtZ_qYVhRFWD2C0#1"#.("+* )(heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheVjdXjcYkd\kdZkeYlfVljUkkSjmXjn]kqey{”œž­¶»ÑÛÝéóõùÿÿ÷ÿÿöúûüÿÿÿÿÿøùôÿÿúþþöøúïðòåèìÞãçØßæÖßçÒäêÐæìÐèìÑçéÑèèÐæäÍàÛÇ×оÙÒÂØÐÃÜÓÊåÜ×íãáòéêùðñüø÷úùõýþùÿÿûýþùüýøÿÿûÿÿûüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿýúÿø÷ÿùøÿþýüÿýõÿýëþøðÿýñÿûõÿøñøèÿüêÿûèôιªwbŽP9©[E±M=¸J;·I<µI=´G@¬C=¤>9™:4•<8Ž=9ˆ?9@9xB8rD7oE9oE9o?5p>7p>7p>7o>7o>7p?8p?8pA9o@8k?6j>5i=4i=4j>5m<5n70q60p70p70m6/l7/i9/g8.c9-a9-b</b>0d@2bA2aA2^C2WD3]N;eVAfW@bS<[O5[O5\R7\W:ZY;mmSrv[y�icmT1?(5B.0;+2:-2:-19,19,08+08+08+08+/7*/7*/7*.6).6).6)-5(,3+,3+,3+,3+,3+,3+,3+,3++2**1))0()0()0()0(*1)+2*+5-*4,*4+*4+*4++5*,6+,7),7).9)2=-5@07B26A05@/3@,0?*4C,9H1=M3@P6DT9HX=K[@RbG[kQapY^mXSbODRA1?.!/"%0*$-*"+(heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheVjdXjcYkd\kdZkeYjgVmlWkmUilWelZem`qzu�—˜§°µÊÔÖâìîõýÿôüþ÷ûüüÿÿÿÿýøùôÿþùýýó÷ùìïñãèìÝåéØãêØåíÖæìÐèïÐèíÏæéÎææÎåãÎáÚÈØÑÁ×ÍÁÜÒÉéÞÚøíëÿóõÿõöÿôøþøøþýûÿÿûÿÿûþÿúýþùÿÿûÿÿûýþùþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿþúÿþúÿþúüû÷øüûøÿÿôÿÿðÿÿëþúïþ÷ûÿøÿÿóÿõåÿòÞÿãÌ×­•¨pW¦XB¬Q>«N<¬O>¯OA¬L@¡C9™<4”=4Ž>5ˆ?6€@6zA6sC5pE5oE7o?5p>7p>7p>7q?8p?8p?8p?8rA:pA9n?7k?6l=5j>5k?6n=6o81p70n70m80j8/i9/h9/e;/d<0c?1bA2aA2_B2]B1[@/VA.SF3ZT>f^Gg`F`Y?YR6XS6[W:XW9WX9imRqw[u}eUbH%36C/1<,3;.2:-2:-19,19,19,19,08+/7*/7*/7*.6).6).6)-5(+2*+2*+2*+2*+2*+2*+2*+2**1)*1))0()0()0()0(*1)*1)(2*)3+)3**4++5,,6-,6+,6+,6+.9+2=/6A18C59D47B25@/.;)0?*4C.7F/;J3?O5DT:HX>QaGYhQ_nY[jWP^MAO@.<-!-#(1,&/,#,)heVheVheVheVgdUgdUgdUgdUifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZkhWpoZjnUinXjq_hsenys…‘‘ž©­»ÄÉÏÙÛçïòóûýúþÿûÿþþþüýþùÿÿúýýóøúíõ÷éòöåïôàêòÝçï×ïö×ëòÑêïÑéìÑææÎÞÛÈÜÕÅÜÔÉßÕÌçÙÖñãâøììþòôÿöúÿùýÿüýÿþüþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿþùþÿùþÿúýÿþûÿÿúÿÿøÿÿúÿÿüÿÿþÿúÿüóÿûïÿýìÿûäÿóÛÿéι�f¥bH•O6ŸT=¨ZF£RAžK; M?�=1‹?1†>0~?.w>-s@-sD2tF6p>5r=7s>8s>8u@:sA:r@9q?8r@9q@9q@9o@8p?8n?7n?7n=6j8/m80k90k;1j;1g=1d<0b>0`?0^>/bG6aG6T=+N9&L7$@1WQ;OO7KI0ML0[Z>gfHbaCUT6WX9]`AosXw€ccmT=J0,:#1>*1<,3;.3;.3;.2:-2:-2:-19,2:-19,19,08+08+/7*/7*/7*/6./6./6..5-.5-.5--4,-4,+2*+2*+2*+2**1)*1)*1)*1)+5-+5-*4,*4,*4++5,+5,,6-(2)*4).8-2=/7A6:E7=H:>I;5B04A/1>,/<(1?(7E.AO8HV?N\EUcL[hTXeSKXG:F8.:.(4*&/*%.)%.)heVheVheVheVgdUgdUgdUgdUifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZjiWnmXimTjoYjs`itfo|u†”” ®±½ÈÌÒÛàéñôôüþúþÿûÿþþþüüý÷ÿÿøýýóùûîöùèóøäðõßéñÚæïÔéðÏåêÊäçÊææÌãáÌÞØÈÜÔÉÞÕÌåÚÖíßÞõççûïñÿôöÿ÷ûÿùýÿýýÿþúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿúûÿùûÿùýÿüýÿþþþÿÿýÿÿýÿÿüÿÿýÿÿýýÿûøÿýóÿÿñÿÿìÿúâÿòØÿìÐß²•§uZ‘[A–ZB›ZDšVAšSAŽG5‹G4‡F4‚E2}D1wC.qA-p?.r@5q?8t?9sA:t?9o=6p>7rA:p?8p?8o>7n?7m>6m>6m>6l=5k<2k<2j>3j@4h@4eA3cB3aA2cH7[A0\G4\J6N=)E6!F7">6KK3GK2CG,BE(JM0Z[<abC_`AWZ;cfGsw\rx\WaH7D*/;%7C/5=.5=05=04</4</4</3;.3;.4</4</3;.3;.2:-2:-2:-19,18007/07/07//6./6./6./6.,3+,3+,3++2*+2**1)*1))0((2*'1)'1)'1)'1((2))3*)3*+5,+5,,6-.8-2<37A6<F=?I>9F58E35B02?+1>*4B+:H1?M6JXAQ_HWdRUbPJVH;G90<2*6,(2*(1,'0+gdUgdUgdUgdUheVheVheVheVifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZjiWmlWhlShpYiuaiugq~w‰——¥³¶¿ÊÎÔÝâêòõõýÿúþÿûÿþþþüüý÷ÿÿ÷üüðùûíöùèò÷ãîóÝçïØãìÑàçÆÜáÁÛÞÁÞÞÆÞÛÈÝ×ÉßÖÍãÚÕîââóçéúîðþôõÿ÷øÿøùÿúûÿüüÿþúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿúøÿúøÿüûÿüýÿþÿüÿÿûÿÿùÿÿùÿÿúÿÿûýÿúøÿþøÿÿôÿÿñüÿìüúãÿóÛÿøàÿíÓâÀ§«ƒj‡YA…S<’[F–]IŒS?F2u>)r>)s@+tD.vE4p>3l<2o=4sC9sA8n>4qA7wH>m>4m>4l=3j>3j>3i=2i=2h>2g?3f@3fB4dC4cC4aD4_D3[D2aL9TB.YJ5_S=RG1C;$D<%EA(AE,@I.?E):A"=B$LO0\_@fiJ^aBmpQuy^flPHP83=$4=(=F37?07?27?26>16>15=05=05=06>16>16>16>15=05=05=05=03:23:23:2291291180180180/6..5-.5-,3++2**1)*1))0((/('.''.''.'(/()0)*1**1*/6/-4-+2++2*.5.4;3<C<>H?>J<=K::I66E02A,0?(2A*3B+CR;IXAP_LO^KGUF:H91>4,8.+5-*4,)3+gdUgdUgdUgdUheVheVheVheVifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZjiWkmWimThpYht`gsem|u‡—–¤³¶¾ÉÍÓÜáéñôóûýùýÿúþýþþüýþøþþöüüðùûí÷úçò÷áíòÛåíÕâèÌÛàÀÖÙºÖÖ¼×ÖÁÜÖÆÞ×ÍäÛÖëáß÷íîûðôÿõùÿùüÿúúÿúúÿüùÿþúÿþúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿúøÿüøÿþúÿþýÿþÿýþÿûþÿúþÿùþÿùüÿùùÿúùÿýùÿÿøþÿöúÿóúþïÿÿíÿüéÿñÝÿðÛÿðÚñÒ½³�zxR;rI3rH0uH1yL5|O:xK6nA,d7$xM=pD7nB7qE8nB7f:-c9-g=/h>2f?0f>2f?0f>2e@0e?2d@2aA2_B2_D3^D3^D3ZE2WE1UD0WH3NB,\T=nhPb^ELJ1IG.IL1;D)=G,>G*9B#;@"DI)UX9adEjmNuxYtw\Z^C?D-4<$8@+<C18@17?27?27?26>16>16>16>17?27?27?27?27?27?27?27?26=56=55<45<44;34;34;34;329118007//6.-4,,3++2**1)*1*)0))0))0))0)*1*+2+,3,.5.,3,*1*)0),3,1817>79C:?K=?N;?N;<K67F/2A*0?(/>';J3AP9GVAHWDBP?9G81?2,9/+6.+5-*4,fcTfcTgdUgdUheVheVifWifWifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZjiWkmWhnThrZfs_cqbixq‚•“¢±´½ÈÌÑÚßèðóòúüøüÿúþýÿÿýþÿùÿÿöÿÿòüÿîúýêöûåïôÝåîÓâèÌØݽÒÕ¶ÑÑ·ÓÒ½ÚÔÆßØÐéáÞòéêýô÷ÿ÷üÿûÿÿþÿÿþüþýùýþøþÿùþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿüúÿþúÿÿûÿþýÿþÿþüÿýúÿüúÿüùÿù÷ÿú÷ÿû÷ÿýøÿþúþÿûûÿúüÿø÷ôíÿþöÿýôÿýñÿûëÿ÷çùçÓéӾʹžµ™ƒ“t_vVAiH5kH5pM:sP=kF4gB0fA1jE3lH8nK8nJ:mL9bA0bA.bA0aB.aB0bC/bC1`C1\E3[F3ZH4YH4XG3TH2RG1PH1NH0MI0ecJ~~dpsXTX=IM2FL0;E*<F+>G*>G(@E'BG'JM.ST5vwXyz[nnTRU:@B,>B+>A.9=,8>07?27?27?26>16>16>15=06>16>16>17?27?27?27?27?28?78?77>67>66=56=56=56=54;34;33:229107//6..5--4,+2++2+*1**1**1**1*+1-,2.+1-+1-*0,+1--3//511731;2:H9;L9>O<=O9:L45G/1C)0B(3E-8J2=O9?P=<M;5F6.>1,:-*6,+5,*4+fcTfcTfcTgdUheVifWifWifWifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZjiWkmWhnThrZer^ao`fwo‚•“ ²´ÀËÏÓÜáéñôòúüøüÿûÿþÿÿýÿÿúÿÿöÿÿóÿÿñþÿìùþçòøÞæïÒâèÌÕÚ¼ÏÒµÌÌ´Î˸ÕÏÃÞ×Ñëãáöíðÿ÷üÿúþÿýÿÿþÿÿÿýüÿøýÿ÷ýÿ÷þÿùþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþþüýþÿûÿÿýþÿýÿþýÿúþÿùþÿ÷þÿõüþóüýõýýõüý÷ýýûþþþþþÿþþÿûùþÿûÿþúûúõòýøòÿþôÿþñÿþíÿúèÿòÝôãÏæѾѺ¨¸ž��ƒr�rax\Gz^IqU@aE0W;&W<'Z?*X=(_D/^E/_F0_F0_F0^H1^H1]I1\L5YM7YM7XM7UM6SM5QM4NL3MK2OR7ilQ|€elrVQZ=DM0=H*9F*9F*?H+BK,DI+CF'GH)LM.xwYsrTecJSS9KJ5HJ4DE3:=,9?18@38@37?27?26>16>16>16>16>16>16>17?27?27?27?29@89@89@88?78?77>67>67>67>66=56=55<44;33:2291180-4--4-+2+*1*)0))0))/+)/+)/+*0,,2.-3/.40.40-3/+5-2@/5G1;M7=O7=O79K17I/5G-2D,5G/9K5;M7:K94E3.?/,:-*6,*4+)3*ebSfcTfcTgdUheVifWifWjgXifWifWifWifWifWifWifWifWjdTkeWkeYkdZle]le[jfZjiWjlVhnThrZfs_`qahyq„—•¥·¹ÆÑ×Úãèíõøõýÿùýÿûÿþÿÿûþÿ÷ÿÿõÿÿòýÿïüÿê÷üåîôÚâëÎÝãÇÓÖ¹Ì̲ÇÅ°ÇijÏȾÚÒÏéàáöíòþ÷þÿùÿÿýÿÿþÿýÿúúÿøúÿõüÿöýÿùþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþþüÿýÿþýÿýþÿûÿþúÿúøÿ÷÷ÿó÷ÿóöÿòøÿóúÿ÷ûÿùýýûÿýÿÿüÿÿûÿûõÿýùÿÿýÿÿýÿÿýÿüýøúúò÷øêÿÿñÿÿïÿÿíÿþíÿüìÿõæÿíßüéØøãÎüæÎàʲ¤Žvt^FdP7aM4\H/]I0[J0\K1\K1]L2\L2\L2\N4YO6VP8VP8VP8TP7PN5NN4KN3LO4NT8`fJenQS^@DO1@K-9F(=J.<I-BK.FO0GL.FI*NO0YX:qmPjfI`\CXV=VT?SR=KJ8BC3;A3:B5:B59A49A48@38@38@36>16>17?27?28@38@38@39A49@89@89@89@88?78?78?77>68?78?78?77>67>66=56=56=538227105/.3-,1-+0,+0,+0,+0,,1--10-10.21-10,0/)3+*8'-@*5H2:N5;O6:N39M28L04H-5I08K5:M79L95H5.@0+<,*6*+5**4)bcQbcQcdRdeSefThgUigXigXhfWifWifWifWifWifWifWkeWjdTkeWkeYkdZle]jf[jfZjiWhjTglUhrZgt`aqdh|sˆšš©»¿ÍØÞßèíòúýøÿÿúÿÿûÿüýÿùýÿôùùíùúìùúè÷úåòößêîÕÞäÊ×ÛÀÒÒ¸ÌDZĿ¬Ä¾°ÊûÕÍÊæÝàóêïþ÷ÿÿúÿÿýÿÿÿÿýÿüúÿùúÿ÷ûÿøýÿúþÿúþþüþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿþúþÿúÿþúÿþüÿüÿÿýÿþþÿýÿþúÿúøÿ÷÷ÿõöÿõøÿöùÿöûÿøþÿúÿþúÿüþÿûÿÿúÿÿüÿÿüÿþüÿþýÿüýÿùûúùþøüÿú÷úñùûðúúîýùíÿùëÿúëÿýíÿþêüðØÿûáêÜÂ��t^P5PB'WI.WI.ZL1[M2YN2ZO3ZO3ZQ4ZQ4XQ5SO4SO6SO6QO6OO5KN3JM2HL1IO3IR5R[>OZ<>I);F&@K+?J*CN0DM0HQ2KR3IN0JM.YZ;jiKedH_^B]Y@][B][DXV?PN9KJ6@C2>B3>B3<B4;A3;A3;A5;A58>27?28@38@39A49A4:B5:B5:B79A69@89A68?78@58?77?49@89A69@89A69@89A69@88?77>66=64;318005./4..3--2,.3/-2.,2.+2++2++2*)3((5$#7'>!/F)6M18O3:N39M29M26J/7K09M4;O6:M76I50C0,=++7+,6++5*\fN\fN^fO`hQaiRejTgjWgjWfgUgfTheTgdSjdVjdVkeWlfXlfVjgXjfZjf[jf]ieZhdXefTehSdlUenYan\aqfq„€•§«¶ÇÎÑÜàáéìï÷ùôýü÷ýùùþøüÿöûÿñõ÷éøùçö÷åñðÜîíÙëêÕàßÊÔÒ½ÖϼÉÁ®½´¥¼²¨Â·±ËÁÀÛÒÕêáæùòúü÷þþûÿþýÿýþÿùýüùþúúÿûýÿüýÿüýÿþýÿüþþüþþüþþüþþüÿþúþýùþýùþýùÿýøþýøÿþùÿþûÿýüÿüûþüýþþüþÿýþÿýúÿùøý÷ýÿúûüöýýõÿþøÿþøÿúõÿøòÿúøÿúÿþûÿþýÿþþþýÿþýÿüúÿûúÿûýÿþþþþÿþüÿþùÿýòÿýíÿýêÿþæøôÛþúßåßÅŽ‰lTM0WN1TK,SJ+VM.VM.TN.TM0TO1SO2SO2QP4ON2ON2PO3PP6MP5JN5GM3DM2HQ6EN1AJ-=F'=G%AK(EL*EM(PV4OU3JO/HK,RU8bbFbbFZZ>YY?ZZ@\ZA[[AZZ>WW;UT8SQ8FE0ED0CB.?@.>?/=?1=?2=A3:=28>28@38@37B47B48C38C57B47B46@56A36@56A37A67B47A67B47A67B48B79D6;E:;E:8B98B97A66A36>14</2:/07/-4-.5.-7.,8*-:(0@&3F&6L&9V*8W+:V.9T3:S6:P9<O9=P:9K3@R8DX<EY=BY<=T82I/)<(*6(*4))3(YgMYgM[hN]jP`jRblTfkUglVghVfgUhgUheTkeWkeWmeXlfXjgVjgVjfZjfZjf]ieZgeXefTgjWemVfo\erahxnwŠˆš¬°·ÈÏÐÙÞàèëï÷ùõþûøþúøÿ÷ûþóøüíöùèöùæôóßìë×éæÓåâÏÚ×ÄÐ˸ɾ¬Á´¤¹­Ÿ»®¥Á¶²ËÁÀÚÏÓåßãúóûú÷þþûÿþýÿüýÿûüþùýüúþýýÿþýÿþýÿþýÿþþþüþþüþþüþþüÿþúþýùÿüùÿüùÿýøÿýøÿþùÿþùÿýúÿüùýüúþýûÿÿÿÿÿÿýýýûûûÿþýÿþûÿüùÿøð÷ìæïåÛõçÞûñèÿûýþüÿþþþþþüýÿúýÿúúþýúþÿýþÿýþÿÿýÿÿþüÿþõÿþíÿþèÿÿãöôÛûûááàĉˆjQL/RM/RL,QK+TN,TN,TO/SP/SO2QP2PP4OO3ON2ON2OO3MP3MO7IO5HN4EN3FP5DO1AJ->H&AH'EM(IO+JP*MS-NT0U[7\b@_bCY\?TW:TT:UU;VV<WW=WW=WW;VW8VU7UT6PL3NI3IG2FC0CB0?@0??3>@3:=28>29?38@37B47B47D37D37B47B46A36A36A36A37B47B47B47B47B47B48C5:E7;F8<G9;G=;G;;G9:G6;D39B18@37?46=54>52<30>-3C)7M)AX.Ea0No:Ln;Ih<Db>@Y<;S;:M:9L8<N6DW;I]ALaBJbBE\?9R5.D-)5'(2''1&YgMYgM[hN]jPakSblTglVglVijXhiWjiWjgVlfXlfXmgYmgYjgVjgVjfZjfZie\ieZhfYefTgjWckTdmZfsbm}s|��š¬°³ÄËÌÕÜÞæéï÷ùöÿþøþú÷þö÷úïôøéñôãîñÞéèÔàßËÜÙÆ×ÔÁÏÊ·ÇÀ®¿´¢½° ¼°¢Ã¶­ËÀ¼ÕËÊâ×ÛêäèûôüûøÿþûÿþýÿüýÿûüþúþýúþýýÿþýÿþýÿþýÿþþþüþþüþþüþþüþýùþýùÿüùÿüùÿýøÿýøÿýøÿýøÿýúÿüùýüúýüúþþþÿÿÿþþþüüüÿþýÿþûÿýúúóëêßÙàÖÌëÝÔöìãÿûýþüÿþþþþþüýÿúýÿúúþýúþÿýþÿýþÿÿýÿÿþüÿþõÿþíÿþåÿþâøöÝûûáàßdžhOJ-QL.RL,RL,TN.TN.TO/SP/SO2QP2OO3OO3NM1ON2NN2MP3LN6JP6IO5FO4FP5DO1CL/BL*FM,IQ,OU1RX2QY0W_6iqJx€[pvTX]=KP2MP5NQ6NP8QQ9QQ7TR9UT8VU9WV:VR9TO9OM8KH5GF4BC3@@4>@3;>39?39?38@37B47B46C26C28C57B47B46A36A37B47B48C58C57B47B48C58C5:E7;F8<G9=I?=I==I;=J9?H7>G6<D7;C8=D<9C:6@73A08H.AW3Of<VrA^JZ|IUtHKiEB[>8P85H52E1<N6EX<MaEQfGPhHLcF?X;3I2.:,,6+*4)XfLYgM\iO]jPakScmUfnWhmWjmZjkYkjXkjXkhYnhZnhZnhZkhWkhWjfZieYie\ieZhfYfgUehU`hQajWeram}s{ŽŒ”¦ª¨¹ÀÆÑ×Ùàæìô÷ôýüöüøôûóòõêíñâçêÙáäÑÛÚÆÒѽÍÊ·ÈŲ½ª½¶¤Ã¸¦ÇºªÌÀ´ÖÉÀÞÓÏèÞÝóèìøòöýöþüùÿþûÿþýÿüýÿûüþúþýûÿþýÿþýÿþýÿþýÿþþþüþþüþþüþþüþýùþýùÿüùÿüùÿýøÿü÷ÿü÷þûöÿýúÿûøüûùüûùýýýþþþþþþýýýÿûúÿþûÿýúüõííâÜãÙÏðâÙüòéÿûûþüÿþþþþÿúýÿúýÿúúÿûúþÿüýÿýþÿÿýÿÿþüÿþõÿýìÿýäþýáúøßþþäáßƇ†jPK.RM/SL/SM-TN.UO/UP0TQ0TP3QP2NN2MM1NM1NM1MM1LO2LN6JP6JP6HQ6HR7FQ3FO2HR0LS2OW2V\8\d;grGs€T†“h�›s„ŽikuRY`?RW9KO4KM5LL4LL4OM6QO6TR9VT;XS=VQ;RP;NK8JI7DE5BB6@B5;>39?39?38@37B47B46C26C28C58C57B47B47B47B48C58C58C58C58C58C59D6:E7;F8<G9<H><H<<H:<I8>G6=F5<D7<D9=D<8B96@73A09I/CY5Ri?ZvEZ{FWyFRqEGeA<U81I1.A.,?+6H0@S7J^BQfGQiIMdG?X;3I2/;-.8-+5*XfLYgM[iO^kQblTdnVfnWinXkn[klZlkYkjXliZliZnhZoi[liXkhWjfZieYhd[hdYhfYghVfiV`hQajWeral|rx‹‰�Ÿ£Ÿ°·ÁÌÒÔÝâçñóñùûóù÷ð÷ðíðççëÝàâÔØÛÈÑмËʶÇŰþª½¸¤º³ Ê¾®ÓƶÜÐÄæÙÑíâÞõëêþó÷ÿúþþ÷ÿýúÿÿüÿþýÿüýÿûüþúþýûÿüýÿþýÿþýÿþýÿþþþüþþüþþüþþüýüøþýùÿýúÿýúÿýøÿü÷þûöýúõÿüùþú÷ûúøûúøüüüþþþþþþþþþÿûúÿüùÿûøÿøð÷ìæðæÜøêáÿõìÿûûþüýþþüþÿúýÿúýÿúúÿûúþÿüýÿüýÿÿýÿÿÿûÿþóÿýêÿüãüûßýûäÿÿèäâÉŠ‰mRM0TO1UN1UN1TN.UO/UP0UR1TP3QP2NO0LM.ML0ML0MM1KN1LN6JP6KQ7IR7KU:HS5JS6NX6RY8T\7\b>emD�‘b’¦sž²�š«~’¡x†•nq|ZZcDKO4KM5KJ5JI4LJ5PK5SN8UP:UP:TO9QO:OL9KJ8FG7DD8BD7<?4:@4:@49A46A36A35B15B19D68C57B47B47B47B48C59D68C58C58C59D69D6;F8<G9=H:;G=;G;<H:<I8>G6=F5<D7;C8:A96@75?64B19I/AW3Ne;Tp?St?RtANmACa=7P3,D,*=*)<(1C+;N2G[?OdEPhHLcF>W:3I2*6((2'%/$
\ No newline at end of file
diff --git a/testimgint.jpg b/testimgint.jpg
new file mode 100644
index 0000000..2501c61
--- /dev/null
+++ b/testimgint.jpg
diff --git a/testimgint.ppm b/testimgint.ppm
new file mode 100644
index 0000000..59c50a8
--- /dev/null
+++ b/testimgint.ppm
@@ -0,0 +1,4 @@
+P6
+227 149
+255
+0/+0/+0/+10,21-32.52-63.83/83/83-83-:3-:3-:3+:3+72.72.61-61-61-61-72.72.50,50,50,50,50,50,50,50,4/+3.*3.*3.*2-)2-)2-)2-)1,(1,(1,(1,(1,(1,(1,(1,(.+$.+$/,%0-&1.'2/(2/(30)52+63,63,74-85.96/96/:70@:,A;-C=/F@2GA3IC5JC3JC3LE3LE3LC2LC2LD1ME2PE3PE3Q>/S@1c?/iE5�C4†H9œE4žG6´F7µG8¿E8ÁG:ÇG<ÈH=ÊG?ËH@êCKëDLëDKëDKíCFìBEò>Añ=@÷:>ö9=ò9>ò9>ä@?â>=Ó?;Ð<8µ@,´?+´?+´?+²?,²?,°@,°@,«@,ª?+¤A,£@+œB*›A)—A(–@'‘@-�?,‘>,‘>,�=+�=+�>)�>)“>*“>*“>*“>*”=)•>*–?+–?+“@8’?7™=0›?2«@.°E3¼C0½D1·?/¸@0¥H9£F7K=s?1N;*E2!95495484373262/51.40-3/,63.52-52-52-41,41,41,30++0*+0*-/*-/*./*./*1-*1-*2-*1,)3)(2('5)):..D66J<<o??zJJŽMK’QO UY­bf¤p†¥q‡Š©‰~¨{ˆ½{ˆ½sy©[a‘WKsPDlPKbMH_KHYSPaddl{{ƒ�’Ž˜�™¢¨œ¡§›�¡�Ž’�z{iefTXWEQP>TPaTPaTPa/.*/.*0/+10,21-21-52-52-72.72.72,72,92,92,92*92*72.61-61-50,50,61-61-72.61-61-61-61-61-61-61-61-3.*3.*3.*3.*2-)2-)2-)1,(2-)2-)2-)2-)2-)2-)2-)2-)/,%/,%0-&1.'2/(2/(30)30)52+63,63,74-85.96/96/96/@:,A;-C=/E?1F@2HB4IB2IB2JC1JC1KB1JA0KC0LD1OD2OD2Q>/S@1b>.gC3~@1ƒE6˜A0šC2­?0®@1¹?2»A4ÂB7ÄD9ÈE=ÈE=ç@HèAIéBIêCJíCFíCFô@Có?Bø;?÷:>ó:?ó:?ä@?â>=Ó?;Ð<8´?+´?+´?+³>*²?,²?,¯?+¯?+ª?+ª?+¤A,£@+œB*š@(—A(–@'�?,�?,‘>,‘>,�=+�=+�>)�=(“>*“>*’=)’=)”=)•>*•>*–?+“@8’?7™=0œ@3¬A/°E3½D1½D1¹A1¹A1¦I:¢E6}I;q=/N;*E2!:65:6595484373051.51.40-63.52-52-52-41,41,41,30+,1+,1+.0+.0+/0+/0+2.+2.+3.+2-*4*)3)(5))9--C55H::k;;uEEŠIGŽMKœQU©^b l‚£o…ˆ}§‰~¨z‡¼x…ºnt¤W]�THpOCkLG^MH_MJ[PM^ZZbmmuƒˆ„“˜”˜ž’™Ÿ“˜œ‹�‘€}~lijX]\JUTBUQbUQbUQb..,..,0/+0/+10,21-41,52-52-52-61+61+61+61+81)81)61-61-50,4/+4/+50,61-61-61-61-61-61-61-61-61-61-3.*3.*3.*2-)2-)2-)1,(1,(3.*3.*3.*3.*3.*3.*3.*3.*1.'1.'1.'2/(30)30)41*41*52+52+63,74-74-85.96/96/?9-@:.B<0D>2E?1F@2G?2H@3HA1HA1H?.H?.I@/JA0MA1MA1JA0LC2YB0\E3pC0tG4‰B0ŠC1—<*™>,¦<,ª@0²B4µE7¹H:ºI;Ü@CÝADáCDãEFëCCìDD÷@B÷@Bù:?ø9>ó<>ò;=âC?ß@<ËB8È?5±A-±A-°@,°@,°@,¯?+®?,®?,©@+¨?*¡A+ @*™A)˜@(•A)•A)�?,�?,�?,�>+�>+�>+�<*�<*�>)�>)�>)�>)�>)‘?*”?+•@,“A5’@4›>/žA2±@0µD4ÁB3ÁB3¾D7¾D7¥K@ F;vJ=j>1I=-B6&=85<74;63:5294183062/51.63.52-32-32-21,21,21,10+,1+,1+,1+,1+.0+.0+0/+0/+3/,3/,5-+3+)4*)7-,=32A76^90gB9|F:‚L@�OMšZX“hy˜m~‚y¤…|§wƒ½r~¸gl£QV�PDtL@pLAaPEeOJaMH_LKYZYgquv„ˆ‰…Œ„‰�ˆŠ�‚ƒ‰{wykfhZZ[KSTDYS]YS]YS]--+--+/.*0/+0/+10,41,41,41,41,50*50*50*50*70(70(50,4/+4/+3.*3.*4/+4/+50,50,50,50,50,50,50,50,50,3.*3.*2-)2-)2-)1,(1,(1,(4/+4/+4/+4/+4/+4/+4/+4/+2/(2/(30)30)41*41*52+52+52+52+63,63,74-85.85.96/=7+>8,@:.B<0C=/D>0E=0E=0E>.E>.F=,F=,G>-H?.K?/L@0LC2MD3YB0[D2nA.qD1…>,†?-’7%“8& 6&¥;+®>0³C5¸G9ºI;Û?BÜ@CàBCáCDêBBëCCõ>@õ>@ù:?ø9>ó<>ó<>âC?Þ?;ÊA7Å<2°@,°@,¯?+¯?+¯?+®>*­>+­>+¨?*§>)¡A+Ÿ?)˜@(—?'”@(”@(�?,�?,�>+�>+�>+Ž=*�<*�<*�>)�=(�=(�=(�=(�>)“>*”?+“A5“A5œ?0ŸB3²A1¶E5ÂC4ÂC4ÀF9ÀF9§MB F;uI<h</H<,B6&>96=85<74;63:5294173062/63.52-32-32-21,21,21,10+,1+,1+,1+,1+.0+.0+0/+0/+40-40-6.,4,*4*)6,+:0/=32Y4+a<3u?3zD8†FD�OMŠ_p‘fw|sž€w¢s¹lx²`eœNSŠPDtL@pLAaODdMH_ID[EDRLKY[_`imnszrw~vyqtzlln`_aSWXHSTD[U_[U_[U_,,*,,*--+..,0/+0/+10,10,30+30+30+30+4/)4/)4/)4/)4/,3.+3.+2-*2-*3.+3.+4/,4/,4/,4/,4/,4/,4/,4/,4/,2-*2-*2-*2-*1,)1,)0+(0+(4/,4/,4/,4/,4/,4/,4/,4/,30)30)30)41*41*41*52+52+41*52+52+63,74-85.85.85.<5+=6,>8,@:.A;/A;/C;.C;.C;.C;.D;,D;,E<-G>/J>0K?1FI6FI6PH5QI6aE0cG2tA,uB-‚;'ƒ<(�;'•@,ŸB0¥H6­L;°O>ÎDAÏEBÖE@×FAäB?äB?ò<;ò<;ú9<ú9<ð><ï=;ÙF>ÕB:½D3¹@/­@+­@+­@,¬?+¬?+¬?+©>,©>,¤?+¤?+Ÿ@,ž?+—A*•?(’@*‘?)�>+�>+�>+�>+Ž=*Ž=*�<)�<)Ž=*Ž=*�<+�<+Ž=,Ž=,�>-Ž?.”C2“B1¢>/¦B3º@3½C6Å@7Å@7¹D;ºE<�OE–H>jJ?]=2<=/78*@85?74?74=52:5194074/74/43.32-23-23-03,03,03,/2+,3,,3,,3,,3,-2,-2,/1,/1,12-23.40-2.+1-*1-*50-72/N7)S<.c>.hC3pD;xLCz[j†gvwqŸ|v¤p{½it¶^c¥PU—OGƒIA}PAjPAjKAcI?aEBWGDYILUMPY]ca`fdah`_f^Y_STZNSYKSYK_WT_WT_WT++),,*,,*--+/.*0/+0/+0/+2/*2/*2/*2/*3.(3.(3.(3.(3.+2-*1,)1,)1,)1,)2-*3.+2-*2-*2-*2-*2-*2-*2-*2-*2-*2-*2-*1,)1,)0+(0+(0+(3.+3.+3.+3.+3.+3.+3.+3.+30)30)41*41*41*41*41*41*41*41*52+63,63,74-85.85.;4*;4*=7+>8,?9-?9-@8+@8+B:-B:-C:+C:+D;,F=.I=/J>0FI6FI6OG4PH5_C.aE0r?*r?*�:&‚;'Ž9%’=)�@.¢E3«J9®M<ÍC@ÎDAÔC>ÕD?ãA>ãA>ñ;:ñ;:ù8;ù8;ï=;î<:×D<Ò?7ºA0µ<+¬?*¬?*¬?+¬?+«>*«>*¨=+¨=+£>*£>*ž?+�>*–@)•?(‘?)�>(�>+�>+�>+Ž=*Ž=*�<)�<)�<)�<)�<)Œ;*Œ;*�<+Ž=,�>-�>-”C2”C2£?0§C4»A4¾D7ÆA8Å@7·B9¹D;žPF—I?jJ?\<1;<.78*@85?74>63=52:5194074/74/43.32-23-23-03,03,03,/2+,3,,3,,3,,3,-2,-2,/1,/1,01,12-51.3/,1-*1-*3.+50-K4&M6(\7'a<,i=4qE<wXgˆix{u£€z¨s~Àkv¸di«X]ŸUM‰KCPAjM>gH>`H>`FCXFCX?BK<?HBHFFLJIPHJQIJPDKQEOUGSYK`XU`XU`XU*,+*,+,,*,,*--+..,0/+0/+/.*/.*1.)1.)1.)1.)2-)2-)1-*0,)0,)/+(/+(0,)0,)1-*/+(/+(/+(/+(/+(/+(/+(/+(1-*1-*0,)0,)/+(/+(/+(/+(1-*1-*1-*1-*1-*1-*1-*1-*30+30+30+30+30+30+30+30+41,41,41,52-63.74/74/85092*:3+;4,=6.=6,>7->5,>5,A9.A9.B8,C9-D:.F<0J>2K?3DE5EF6LE3MF4XC2YD3f?.f?.v=,v=,‚9(…<+“=,˜B1ŸF4¢I7»F<¼G=ÆH<ÈJ>ÙD=ÚE>ì?;ì?;ò::ñ99ç>9ç>9ÍG;ÇA5­D.§>(¦A-¦A-¦?,¦?,¦?,¥>+¤?-¤?- ?,Ÿ>+›@-™>+”?+“>*�>)�>)�>-�>-Œ=,Œ=,Š=+Š=+Š=+Š=+Š=-‰<,ˆ<,ˆ<,ˆ<.‰=/Š>0‹?1’D0’D0¥?1©C5½A9ÀD<ÃB<ÂA;±JCµNG˜^S’XMgWJXH;;F87B4?80>7/>71<5/94.83-63,63,34.23-14-14-.3,.3,-4,,3+,3,,3,,3,,3,,3,,3,-2,-2,.0+02-23.12-/0+./*0/+10,>1+?2,J4'P:-V=6aHAo^p…t†��µ„„¸{ƒÎs{Æmp¿be´VSšFCŠL?jI<gD=_E>`DBXDBX<CM8?I9CB=GFCNFEPHFRFGSGLVKPZOaZJaZJaZJ)+**,+++),,*--+..,/.*0/+.-).-)0-(0-(0-(0-(1,(1,(0,)0,)/+(/+(/+(/+(0,)0,).*'.*'.*'.*'.*'.*'.*'.*'1-*0,)0,)0,)/+(/+(/+(.*'0,)0,)0,)0,)0,)0,)0,)0,)30+30+30+2/*2/*2/*2/*2/*30+41,41,52-63.74/74/85092*:3+;4,<5-<5+=6,=4+=4+A9.A9.B8,C9-D:.F<0J>2K?3BC3CD4JC1KD2VA0WB1d=,d=,t;*t;*€7&‚9(�9(”>-›B0žE3¸C9¹D:ÃE9ÇI=ÙD=ÛF?îA=îA=ñ99ð88ç>9æ=8ÌF:Æ@4«B,¥<&¥@,¥@,¦?,¦?,¥>+¥>+¤?-£>, ?,Ÿ>+š?,™>+”?+“>*�>)Ž=(�>-Œ=,Œ=,Œ=,Š=+Š=+Š=+‰<*‰<,‰<,ˆ<,‡;+ˆ<.‰=/Š>0Š>0’D0“E1¥?1ªD6¾B:ÀD<ÃB<ÂA;·PI½VO¡g\›aVqaTbREDOA@K=>7/>7/=60<5/94.83-63,52+34.23-14-14-.3,.3,-4,,3+-4--4--4--4--4--4-.3-.3--/*/1,23.12-/0+./*/.*0/+:-':-'F0#L6)T;4aHAsbt‹zŒŠŠ¾ŒŒÀ�‰Ôx€ËsvÅhkºYV�FCŠI<gH;fD=_D=_B@VCAW>EO>EOGQPLVUR]UT_WS_SR^RT^SV`U`YI`YI`YI+-,+-,+-,+-,+-,+-,,,*,,*/.*.-).-).-)/,'/,'/,'/,',+),+),+)-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*-,*,+),+)-,*-,*.-+.-+/.,/.,/.,/.,/.,/.,.-+.-+.-+.-+/+(0,)1-*2.+3/,3/,2.+2.+2.+2.+2.+3/,3/,51.62/73070*81+92,;4.<5-<5-=4-=4->5,>5,?5,@6-A7-C9/G:1H;2G?4G?4K?3L@4T>3T>3]9-]9-m9,n:-}7+€:.Œ:,�>0—>0˜?1¦B3¨D5²D5·I:ÉG:ÌJ=ÝD>ßF@æA=çB>ßF@ÜC=ÀH8¹A1 D-žB+ A- A-Ÿ@,ž?+Ÿ>+Ÿ>+ž?-ž?-š=,™<+•>+”=*�>+Ž=*Œ>*Œ>*Š=+‰<*‰<*‰<*‡;+‡;+…<+…<+†=.†=.…=/…=/„>2†@4‡A5ˆB6�E2‘F3¦D7­K>¼E?ºC=¶G@»LE©_T°f[”wg‘tdsugik]P`SCSFOK@IE:A?3;9-85,74+44*33)34,34,14-03,-2,,1+,1+,1+-2,-2,-2,-2,-2,-2,-2,-2,02-02-02-02-12-12-12-12-4*39/8?42@53E<7YPKrn}Š†•�‘”–ÇŽ•ÝƒŠÒs{Ãfn¶^d¢V\šNKjFCb=?V=?V<DQAIVHVWSab`qim~vt†zo�uiyleuh^l]SaRa[Ca[Cb\D,.-,.-,.-,.-,.-,.---+--+.-).-).-)-,(/,'/,'.+&.+&,+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+)-,*-,*-,*.-+.-+.-+/.,.-+.-+.-+.-+.-+.-+.-+/+(0,)1-*2.+2.+2.+2.+2.+2.+1-*1-*2.+2.+40-51.62/70*81+:3-;4.<5-<5-=4-=4->5,>5,?5,@6-A7-C9/F90G:1F>3G?4K?3K?3S=2S=2\8,\8,k7*l8+{5)~8,‹9+Ž<.–=/—>0¤@1¥A2¯A2²D5ÄB5ÈF9ÚA;ÜC=å@<æA=ßF@ÛB<¿G7¸@0 D-žB+ A-Ÿ@,ž?+�>*ž=*ž=*ž?-ž?-™<+™<+”=*”=*�>+Ž=*Œ>*‹=)‰<*‰<*‰<*ˆ;)‡;+‡;+„;*„;*…<-…<-„<.…=/„>2…?3‡A5ˆB6“H5“H5§E8­K>½F@¼E?¹JC¾OH¯eZ¶la›~n˜{kz|nprdWgZJZMOK@KG<EC7@>2=:196-44*11'23+12*03,/2+,1+,1+,1+,1+,1+,1+,1+,1+,1+,1+,1+,1+/1,/1,/1,/1,01,01,01,01,5+48.7=20?42H?:^UPvr�Œˆ—�‘”–ÇŽ•Ý„‹Óx€ÈpxÀlr°gm«a^}[XwSUlTVmU]j\dqftur€�u†~�ˆƒ•‰|Ž‚v†yp€sfteXfW^X@_YAa[C*./*./*./*./,.-,.-,.-,.-,,*,,*-,(-,(,+',+',+'+*&+*(+*(+*(+*(+*(+*(+*(+*(,+),+),+),+),+),+),+),+),+),+),+)-,*-,*-,*-,*.-+.-+.-+.-+.-+.-+.-+.-+.-+/+(0,)1-*1-*2.+2.+1-*1-*0,)0,)0,)0,)1-*3/,40-51.80-91.:2/;30<41<41=4/=4/=4-=4->3-?4.@6-B8/E80F91H;5H;5J;4K<5O:5O:5V72V72c60d71t5,w8/†8,Š<0“;/•=1šA3šA3ŸB3¡D5®B5²F9ÃE9ÅG;ÑG=ÓI?ÏJAÌG>µJ:®C3›D0˜A-šB.™A-™A-˜@,™>,™>,š?-š?-–=-•<,‘>,�=+Œ=,Œ=,‰>+‰>+ˆ=*ˆ=*‡<)‡<)…<+„;*ƒ;,ƒ;,�;/�;/�=2�=2‚?6„A8ƒC9„D:“K<’J;¡G>¦LC±KF±KF¨XO®^Užyf¦�n�’|��y{Œyp�ncqbVdUMQ@JN=GI;CE7>@39;.35*/1&01)/0(/0*./)./*./*./*/0+/0+/0+/0+/0+/0+/0+/0+/0+01,01,01,01,01,01,01,01,6+<6+<600;55EG:]_Ru{{‰���”±‘˜µ‹–Ä�¼|‹²{Š±y�§xŽ¦|‹�v…Šn€€l~~j�ym„|u��€›Œˆ¤Ž�©“�§‘†žˆ€–�xŽyj{hXiVZV9\X;_[>*./*./*./*./,.-,.-,.-,.-,,*,,*-,(,+',+',+'+*&+*&+*(+*(+*(+*(*)'*)'*)'*)',+),+),+),+),+),+),+),+),+),+),+),+),+),+),+),+)-,*-,*-,*.-+.-+.-+.-+.-+/+(/+(0,)1-*1-*0,)0,)/+(/+(/+(/+(/+(0,)1-*3/,3/,80-91.:2/;30<41<41<3.<3.=4-<3,=2,>3-?5,@6-C6.D7/F93F93H92I:3N94M83T50T50a4.c60r3*v7.…7+‰;/“;/”<0™@2™@2›>/�@1©=0­A4¾@4ÁC7ÎD:ÏE;ÌG>ÉD;³H8¬A1™B.—@,™A-˜@,—?+–>*˜=+˜=+™>,™>,•<,•<,�=+�<*Œ=,‹<+ˆ=*ˆ=*‡<)‡<)†;(†;(„;*ƒ:)‚:+‚:+€:.€:.€<1�=2‚?6„A8„D:…E;“K<’J; F=¥KB°JE´NI¬\Sµe\¦�n­ˆu˜š„•—�„•‚yŠwlzk`n_UYHPTCKM?EG9@B5;=057,13(/0(./'-.(,-',-(-.)./*/0+./*./*./*./*./*./*./*./*/0+/0+/0+/0+/0+/0+/0+/0+7,=5*;3--:44HJ=bdWyŠ��‹’¯Ž•²ˆ“ÀƒŽ»€�¶‡–½Œ¢º�¦¾—¦«’¡¦Œžž‰››‚™‘–Ž�œ�ˆ£””°š–²œ”¬–‹£�…›†|’}j{hUfSXT7ZV9]Y<+/0+/0+/0+/0+/0+/0-/.-/.--+,,*,,*,,*,+',+',+',+'+++*********))))))((((((************************************************++++++,,,,,,---------.../+*/+*0,+0,+0,+/+*/+*.*)/+*/+*/+*/+*0,+1-,2.-3/.91/91/:20;31;30;30<20;1/<3.<3.=2.=2.=2,?4.B5/C60A83A83B94C:5E:6D95I81I81V80W91g7-k;1|:.€>2‹=0�?2–?6–?6“@2”A3™B1œE4©E5«G7¶F:¸H<ºJ>¸H<«F:¥@4—A2•?0”B-“A,”?+“>*”=,”=,”=,•>-’<-’<-Ž=,Ž=,Š=-‰<,…<+…<+†;(†;(„;*ƒ:)‚:+‚:+:+:+~<0~<0}=3?5€A8ƒD;ƒG?„H@�K@ŽJ?šGAŸLF¤QKªWQ�l]¦uf““w™™}‰¨ˆ‡¦†��„x”{u€ojudZfPS_IJVBBN:=F58A07;-48*01)/0(.-(,+&.*'/+(1++2,,0,)0,)0,)0,)0,)0,)0,)0,)/.*/.*/.*/.*/.*/.*/.*/.*9+:6(70-$85,DO1_jLv„j†”z‹˜‘�œ•‹š�ˆ—š„Ÿ–�ª¡–½¢�Ä©—¸£–·¢–º –º �´•ˆ­Žˆ­‹�²�‘µ‘‘µ‘�¬�†£„ƒœz“vg{bPdKST5TU6WX9+/0+/0+/0+/0+/0+/0-/.-/...,--+--+--+-,(-,(-,(-,(,,,,,,+++******)))((((((*********************************))))))))))))(((******+++,,,,,,---.......*)/+*/+*/+*/+*.*)-)(-)(/+*/+*/+*/+*0,+1-,3/.3/.:20:20;31;31;30;30;1/;1/;2-;2-<1-<1-<1+=2,@3-A4.?61?61@72A83C84C84G6/G6/S5-T6.d4*h8.z8,~<0‰;.‹=0•>5”=4’?1“@2—@/™B1¥A1¦B2°@4²B6´D8³C7¨C7£>2–@1•?0“A,’@+“>*’=)“<+“<+“<+”=,‘;,‘;,Ž=,�<+‰<,ˆ;+…<+„;*…:'…:'ƒ:)ƒ:)�9*�9*~9*~9*~<0=1~>4€@6‚C:…F=…IA‡KCŽJ?ŒH=šGA¡NH¨UO¯\V£rc®}n——{���‹ªŠ‡¦†‚ž…{—~z…tp{jcoY\hRQ]IGS?BK:<E4;?19=/23+01)/.),+&-)&.*'1++2,,/+(/+(/+(/+(/+(/+(/+(/+(.-).-).-).-).-).-).-).-)7)85'60-$74+BM/\gIs�g„’x�š“‘ž—Ž� Œ›ž‰¤›—²©ŸÆ«§Î³¥Æ±¦Ç²¦Ê°£Ç­™¾Ÿ�²“‰®Œ�²��±�Ž²Ž�ª‹ƒ �˜{v�rcw^MaHPQ2QR3QR3+/2+/2+/0+/0+/0+/0-/.-/..0/.0/..,..,..,--+.-).-)------,,,+++******))))))******************************)))))))))((((((((()))******+++,,,---.......*).*)/+*/+*.*)-)(,(',('0,+0,+0,+0,+1-,2.-3/.40/:12:12;31;31;31;31;1/;1/;1/;1/;0,;0,;0,<1-?2,@3-96/:70983:94:94:94=82<71G4-H5.Y5)]9-o7(t<-<+�>-�;4�;4�?3Ž@4ŽB2�C3”C0”C0ž>.¡A1¨B6©C7¢B6ž>2•?2”>1‘@+�?*�=+�<*‘;*‘;*‘;,’<-�<,�<,‹<-Š;,‡;+‡;+„;*„;*ƒ:'ƒ:'‚9(‚9(�9*€8)}9,}9,}=1~>2~?6€A8�E=…IA†KEˆMGŒIAŒIA—IE¡SO¤]W­f`�€n¨‹yŠ¢€�§…€°Š|¬† ƒyš}�ˆvxmfx^`rXXgPN]FHT@AM9>D6:@257,35*2/*/,'0(&1)'3)*4*+2*(2*(2*(2*(0+(0+(0+(0+(0,)0,)0,)0,)0,)0,)0,)0,)9(28'10/65!<LUe6n‚Q€”c‹œz�¡Ž¢†Œ „Š«|–·ˆ›Íˆ¢Ô�¥Ó¬¤Ò«£Ò¨�Ì¢�¾��°‚­~ƒ±‚…¯�ˆ²„ˆ«ƒ} xz”qpŠg_sWJ^BKN1JM0IL/,03,03,01,01,01,01.0/.0//10/10//-//-//-..,/.*/.*......---,,,+++******)))))))))))))))))))))))))))******))))))((((((''''''))))))***+++,,,---.......*).*)/+*/+*.*)-)(,('+'&0,+0,+0,+0,+1-,3/.40/510:12;23;31;31;31;31;1/:0.;1/:0.;0,;0,;0,;0,>1+?2,96/96/872983983983<71<71C0)E2+U1%Z6*l4%q9*|9(~;*Ž92Ž92Œ>2�?3�A1�A1‘@-�?,›;+ž>.¥?3¦@4 @4ž>2•?2”>1�?*�?*�=+�<*�:)‘;*‘;,‘;,�<,Ž;+‹<-Š;,‡;+†:*„;*ƒ:)‚9&‚9&‚9(‚9(€8)€8)}9,|8+~>2~>2@7�B9‚F>†JBˆMGŠOIŒIAŒIA™KG¤VR¨a[±jd¢…s­�~�¥ƒ�¨†�±‹|¬† ƒz›~ƒŠx{‚pj|bew]_nWVeNO[GFR>AG9;A379.46+30+/,'0(&0(&3)*4*+1)'1)'1)'1)'/*'/*'/*'/*'/+(/+(/+(/+(/+(/+(/+(/+(8'18'11054 9IQa2kN~’a‡˜vŒ�{‹ŸƒŠž‚†§x’³„•Ç‚›Íˆ™Ç ™Ç ˜Ç�“˜ˆ·‰~­�¯€ˆ¶‡‚¬~…¯�†©�z�uv�mk…bZnRFZ>HK.FI,DG*-0'.1(.1(.1(01)01)01)01)33+33+63,52+52+41*30)2/(7/,6.+6.+5-*2-*1,)0,+/+*.,-.,---/--/+,0*+/()-()-/(//(//(//(/.)-.)-.*+.*+,+)-,*,-(,-(-.(-.(-.(-.(+-,+-,------/.,/.,2.+2.+1,&2-'5.&6/'8/&90';1(;1(;4*<5+<5+<5+=4-<3,<3,;2+;2-;2-;0,;0,;0.;0.<1/<1/E.&E.&86)86).:,.:,36-36-C+)B*(O%&R()V/(]6/\@2`D6vB,xD.…@0‡B2”>1“=0“9.’8-�<,‘>.ŽC0�E2�D/ŽC.�?+Œ>*›82š71�=1Ž<0‚A/�@.|B.{A-ƒB0ƒB0‡>/†=.‡;-†:,„;,ƒ:+u9.|@5s>.n9)s@-s@-t<+yA0};-�?1‡E9‹I=†L@‡MA€PB�QC•IK›OQˆh[�pct‡i{Žpx�{~£�Š¤‡‹¥ˆ‘¥Š�¤‰ƒ¤‡}ž�k�zg™vb�dZ‡\X|VSwQZnRSgKQUDHL;D=5?80:..5))5&+5&+2%,2%,$-*$-*&*)&*)*(++),0',1(-3(03(03'13'10)10)1-*1-*1.*+0,-.-(/.);>+VYFnyW~‰gŠ�p‹žq‰¤oŒ§r‡¨qˆ©rŽ´{˜¾…¡¿‹Ÿ½‰—¹†‘³€‰®{…ªwƒ©z„ª{ƒ¦|‚¥{€�~v“tqˆnf}c[jWL[HIP@AH8<C3.1(.1(.1(/2)01)12*12*12*44,44,63,63,63,52+41*30)7/,6.+6.+5-*2-*2-*0,+0,+.,-.,---/--/+,0*+/)*.()-.'..'..'..'.-(,-(,-)*-)*-,*-,*,-(-.)-.(-.(./)./),.-,.----.../.,/.,1-*1-*2-'3.(6/'70(90':1(<2)<2)<5+<5+=6,=6,>5.=4-<3,;2+;2-;2-;0,;0,;0.;0.<1/<1/E.&F/'86)86).:,.:,36-25,C+)C+)O%&R()U.'Z3,Y=/\@2s?)uA+ƒ>.†A1“=0“=0”:/“9.�:*�:*Š?,‹@-ŒA,�B-�A-�A-š71š71�=1Ž<0‚A/�@.|B.{A-‚A/‚A/†=.…<-†:,†:,ƒ:+ƒ:+t8-x<1p;+n9)r?,q>+q9(s;*~<.‚@2ˆF:‹I=†L@†L@€PB�QC“GI›OQŠj]“sfx‹m’t{ ~�¦„‹¥ˆŒ¦‰’¦‹‘¥Š…¦‰ ƒmŸ|i›xe’g^‹`\€ZX|V_sWZnRY]LPTCKD<E>6?338,,7(-6',3&-2%,$-*$-*&*)&*)+),+),1(-2).3(03(03'13'10)10)1-*1-*12./2./-,',+&8;(TWDmxVŠh„—j…˜k„Ÿj‡¢mƒ¤mƒ¤m‡­t�¶}š¸„˜¶‚‘³€Œ®{…ªwƒ¨u‚¨y‚¨y‚¥{�¤zœ}u’sp‡mf}c[jWL[HEL<=D48?//2)/2)03*03*12*23+23+23+55-55-74-85.74-63,63,52+81+70*50,4/+3.*2-)1-*0,)/.,/.,......--/,,.)*,)*,/)-/)-/)-/)-.*+.*+.*).*).-).-).-)/.*./)./)/0*/0*//////..,..,/.*/.*1.)1.)4/)4/)70(81);2)<3*>4*>4*>7->7-?6-?6-?6->5,=4-<3,;2-;2-;0,:/+:/-;0.<1/<1/@2)A3*N,+N,+W%.V$-L(,K'+=0*</):2';3(O.%S2)i/-k1/h:+j<-w;0z>3‰<6‰<6‹;4Š:3†=4†=4‚@2‚@2…A4†B5ŠB6‹C7‹?/‹?/ˆ?.‡>-†?-…>,‡;+‡;+Š;,‰:+ˆ;+‡:*<)<)w?(w?(}=1{;/u:,v;-w?0w?0v<.u;-?3‚B6ˆE<‹H?‰K@‰K@†MB‡NC”HHœPP�h\™qe‚ˆlˆŽr{Ÿ{€¤€‚ª…ƒ«†‹ªŠ‹ªŠ�£‡‰��†–|‚’xa™j\”eX‹`V‰^X…^S€YPsUIlNB[E;T>4F6+=-(5+%2($/'#.&4%(4%(1%'1%'0'*1(+/+,0,-+,.+,.(..(..%/0%/0$00$004.04.0.+&,)$58'QTCkvX}ˆj}’i~“j~šj‚žn~ m}Ÿl¦q†­xŠ¯y‰®x…¬w‚©t§s~¦r~¥v¦w¤{~£z}›yt’pp‡ke|`XjRJ\DEG9=?18:,14+14+14+25,34,34,45-45-66.66.96/96/96/96/85.74-92,92,61-61-50,4/+2.+2.+0/-10.000/////1--/+,.*+-0*.0*.0*.0*./+,/+,/+*/+*/.*/.*/.*0/+/0*01+01+01+00000000.00.0/+0/+2/*2/*61+72,:3+;4,=4+>5,@6,@6,@9/@9/A8/A8/@7.?6->5.=4-<3.;2-;0,;0,;0.;0.<1/=20B4+B4+O-,O-,X&/W%.M)-L(,>1+>1+;3(;3(N-$Q0'e+)g-+g9*i;,v:/z>3ˆ;5‰<6Š:3Š:3ˆ?6ˆ?6„B4„B4†B5†B5‰A5‰A5‹?/Š>.ˆ?.‡>-…>,…>,‡;+‡;+ˆ9*ˆ9*†9)†9)~;(~;(v>'v>'?3w7+u:,w<.s;,u=.z@2z@2�A5„D8‰F=ŠG>ˆJ?‰K@‡NC‰PE˜LLžRR�g[•ma|‚f‚ˆlvšv| |§‚�©„Š©‰Š©‰�£‡Šž‚‡—}ƒ“ydœm_—h^‘f]�e`�f]Šc\aVy[NgQF_I>P@3E5.;1(5+%0(#.&4%(3$'1%'1%'/&)0'*.*+/+,)*,)*,&,,&,,#-.#-.".."../)+1+-.+&.+&7:)QTChsUw‚dyŽez�fz–f€œl|žkzœiy k~¥p�¦p¤n}¤o|£n|¤p}¥q}¤u}¤u{ w{ wz˜vr�nn…ie|`XjRJ\DCE7;=/68*25,25,36-36-56.56.56.56.77/880;81;81;81;81:70:70<5/;4.94083/72.61-51.51.51.51.32032021/10.0/-/.,3*+3*+1++1++1++1++0,)0,)2.+2.+10+10+10+21,12*12*22022023.12-21,21,50*50*;4,;4,=4+>5,@6,A7-B8.C9/C;0C;0C;0C;0B90@7.?6->5,=4-<3,<3.;2-;2-<3.<3.=4/?61?61P-1P-1['3Z&2O+/N*.:6-:6-29)3:*B4'C5(Y.'Z/(b:2e=5r;6v?:ƒ;<ƒ;<…;:„:9;8€<9}A9C;‚C<�B;…@9„?8}E.}E.�C.€B-†=,…<+‹9+Š8*‰7)‰7)‚;)�:(x>(x>(qB(pA'‡4,,$�7,‡=2v4(v4({A5}C7zF9|H;K>€L?�M@ƒOB‡QEŠTHjkKnoOwsV|x[€|_ˆ„g‡’rŽ™y€£{ƒ¦~|¬‚~®„|¬‚w§}v zs�w„�n�Šk|Ši|ŠiyŠjwˆhpƒglcbpWZhOTYEHM9E?1=7):-$7*!6&&6&&1'%1'%+*%+*%(-'(-'&,(&,()+()+(+)*+)*0'*0'*,$!2*'51&95*?D0TYEbqRm|]qŠbqŠbp‘bv—hu›js™hqœgu kw¢lu jsŸju¡lx¤sz¦uz¤vy£ux�tx�tw–toŽll…gc|^UlOG^AFB7>:/95*47.47.47.58/67/67/780780991991<92=:3=:3=:3=:3<92?82>71<73;62:5194084173084184165365354243121/21/5,-5,-3--3--3--3--2.+2.+3/,3/,21,21,21,32-23+23+33133134/45054/65094.94.=6.=6.?6-@7.C9/D:0E;1E;1E=2E=2E=2D<1C:1B90@7.?6-?6/>5.=4/=4/=4/=4/>50?61?61@72R/3R/3])5])5R.2R.2=90<8/5<,5<,D6)E7*Y.'Z/([3+]5-k4/p94~6778�76�76|85}96y=5{?7~?8@9„?8„?8}E.}E.€B-A,†=,…<+Š8*Š8*‰7)‰7)‚;)‚;)x>(x>(qB(qB(†3+†3+™OD£YN…C7{9-�G;‚H<{G:|H;~J=K>�M@„PC‰SG�WKmnNqrRzvY~z]‚~aŠ†i‰”t‘œ|~¡y€£{{«�}­ƒ|¬‚x¨~w¡{tžx…Žo„�n€Žm€Žm~�o}ŽnwŠnt‡km{bft[afRUZFQK=GA3B5,</&;++:**4*(3)',+&,+&(-'(-'(.*(.*+-*+-*-+,-+,2),2),0(%:2/B>3IE:NS?]bNcrSixYk„\jƒ[iŠ[o�ao•dn”cl—bošetŸir�gpœgržiw£ry¥ty£uv rušqušqt“qmŒjk„fb{]UlOG^AEA6=9.84)58/58/58/690780891891891::2;;3>;4>;4?<5?<5?<5>;4?:4?:4>93>93=82<71<73;62:72;83;83:72:7285074/63.7/,7/,7/,7/,50,50,50,50,41,41,52+52+33+44,44,44,43/540650761:70;81=90=90?8.?8.A9.B:/D:.F<0H<.I=/G=3G=3G=3G=3E;1D:0A8/@7.@70@70?6/?6/>71>71?82?82=82>93>;4?<5@=6@=6C<2C<2E8/E8/H8+H8+F9(G:)D<'E=(a-"d0%q0*w60…73ˆ:6‹<8‹<8‹B;‰@9�>5~;2‚90ƒ:1Š;4Œ=6�A5€@4‚@4�?3�=2€<1}=1}=1x>0x>0w?0w?0{=0{=0€:0€:0›+'¨84Ïieá{wªZQ‘A8„OA�L>mN:nO;hT<iU=pU@tYD€]I„aM^{Kc€P�tY“x]½d^ÀgaÐfhÔjlÇ}tÊ€w¬”|®–~—¡~“�z‡�w„štŒ�p‹Žo„’o…“p}—r{•pq•oo“mhŒfb†``yYUnNT^CHR7BD.;='@3+?2*92(70&00&//%-0%,/$2/(2/(:+(:+(D%*D%*K"*K"*@8%LD1VQ;]XB]dEgnOgySi{UhƒVf�Te‡UlŽ\m”_m”_n–aq™držko›hm™hpœkvŸsx¡uxžwtšst–st–ss’pl‹ii„ca|[TnKF`=MD;E<3@7.6906906906908918919:29:2::2;;3>;4?<5@=6@=6@=6?<5A<6A<6@;5?:4>93>93=84=84<94<94<94<94;83:7296185091.91.91.91.72.72.72.72.52-52-52+52+44,44,44,55-540540761872<92=:3?;2@<3@9/@9/B:/C;0E;/F<0I=/J>0H>4H>4H>4G=3F<2D:0A8/@7.B92A81@70@70?82?82@93A:4=82>93>;4@=6A>7B?8E>4E>4F90F90I9,J:-H;*H;*F>)F>)i5*l8-{:4ƒB<’D@—IEšKG›LH›RK•LEˆE<€=4€7.�8/‰:3Œ=6�A5€@4‚@4�?3�=2€<1}=1|<0x>0x>0w?0x@1|>1|>1�;1�;1—'#®>:å{ü–’¼lc—G>‚M?{F8lM9mN:hT<iU=pU@uZE�^J†cOa~Ng„T“x]–{`½d^ºa[ÄZ\Å[]È~uÊ€w­•}°˜€˜¢•Ÿ|‰Ÿy†œv��qŽ‘r‡•rˆ–s™t~˜su™ss—qn’li�gh�a^wW]gLPZ?IK5AC-D7/C6.=6,:3)33)22(/2'.1&52+52+=.+=.+G(-G(-N%-N%-ME2YQ>c^HhcMfmNmtUj|Vk}Wj…XhƒVe‡UlŽ\n•`p—bq™dtœgq�jnšgl˜go›južrw tvœuq—pr”qs•rr‘okŠhi„ca|[TnKGa>SJAKB9F=48;27:17:17:19:2:;3;<4<=5<<4<<4?<5@=6A>7B?8B?8C@9E@:E@:D?9D?9C>8C>8B=7B=7@<3@<3B;3A:2A:2A:2A:2A:2A81@70@70@70>7/>7/=6.=6.<71;60:7096/85.74-74-63,66.66.85.96/;7.=90@9/A:0B:/C;0F<0H>2K?1L@2L@0L@0MA5L@4K?3J>2G=3F<2E;1D:0B90A8/?8.?8.<8/=90=90>:1;4,LE=P?8J92Z;6T50W,&d93a0)d3,e1&a-"e,!l3(s9.u;0�G8ŽF7šG9ŸL>£E;›=3�=1¢B6¥M?§OA«UD­WF©M>›?0�9*¡=.•78•78�55‹11ƒ83‡<7wC6r>1lE4iB1q=0q=0‰32‰32œ'0š%.ÑCAî`^ÅCAÕSQÞ{uµRL~I9‰TDcW=g[AafFbgGgbDgbDzeJ„oT¨WF»jYí`fâU[þ,Iö$AÿFÿ"Jÿ8Vÿ:XòTcÿapòtxóuyíy|ð|ÌyuÍzvÆ}vÇ~w»‚w»‚w³…v²„u¤zl£yk£qhœjaš[VŽOJˆA?‚;94;t)0m,0l+/T$"Q!N/*R3.J3-H1+D1*E2+F.*C+'H/+O62VK+aV6da>gdAhpIu}Vr„Zo�Wi…UlˆXnŽ\r’`t—au˜bwšdx›eq–ms˜ot–qr”ot“sy˜x{˜|y–zy“xs�rl†iiƒfg‚a`{ZSnKHc@SMAQK?OI=7:17:17:17:19:2;<4<=5=>6>>6>>6@=6A>7A>7B?8B?8C@9E@:E@:E@:D?9D?9C>8C>8C>8C?6B>5D=5D=5C<4B;3B;3B;3B92A81A81A81@91@91@91@91<71;60:7096/85.85.74-74-66.66.96/96/;7.=90A:0B;1C;0C;0F<0H>2K?1L@2L@0L@0L@4L@4K?3J>2G=3F<2E;1D:0B90B90@9/?8.=90>:1>:1?;2B;3B;3C2+F5.[<7`A<mB<vKEyHA|KD€LAK@†MB�TI‘WL�VK�E6ŒD5–C5™F8�?5–8.™9-�=1œD6˜@2—A0›E4¢F7£G8­I:³O@¬NO©KL�CC”::ˆ=8‰>9xD7s?2g@/rK:�M@yE8‚,+…/.°;DÆQZÚLJÝOMÂ@>Á?=Æc]¹VPzE5‡RBocIfZ@Z_?ejJrmOojL{fK�lQ·fU½l[âU[ÖIOú(Eù'Dÿ!Iÿ"Jþ/Mþ/MâDSéKZ×Y]ÖX\Ï[^Ò^aÂokÃpl½tm¿vo¶}r·~s±ƒt±ƒt±‡y°†x²€w¯}t³to°ql±jh°ig³ho¥Za”SWŽMQ}MK{KI]>9? L5/H1+E2+I6/N62O73T;7ZA=[P0dY9eb?ifCltMz‚[xŠ`t†\nŠZnŠZm�[o�]r•_t—at—at—ao”ks˜ow™tv˜s{šz€Ÿ‚Ÿƒ€��€šx’wpŠmj„gg‚a_zYSnKIdAVPDVPDTNB6906907:18;2:;3<=5>?7?@8@@8@@8B?8B?8C@9C@9C@9C@9EB;DA:DA:DA:EA8D@7D@7D@7HA7HA7H?6G>5F>3E=2D<1D<1C;0C;0C;0C;0C:1C:1B;1B;1<8/<8/<8/;7.96/85.85.74-96-96-96-:7.<8-=9.A;/B<0C;.D</I=-J>.L?.M@/M@/M@/M?2M?2L>3K=2I=1G;/E;1D:0C;0C;0?;0?;0?;0?;0?<3?<3ZB8N6,{*0�?EÆ?PÓL]ÚVaØT_œB:žD<—K;šN>¸KFÀSNàDOÜ@KŸD1�B/¤>/¤>/©9+¥5'¥7(©;,¬E2¢;(›7 �9"©<(­@,¶A0¹D3«.4´7=¿BHÄGM½LN·FH˜A:Š3,{9+x6(~0&�3)©.3¸=Bá@RëJ\çE@Õ3.Ï:6Å0,ÄOHÓ^W�N=•VEugJl^AdeCpqOƒnO~iJ–dM¢pYÿQ]ÿLXÿ<Nù4Fÿ/Fÿ2Iÿ0Jÿ,Fÿ2Jþ.Fú.Gü0Iÿ/Iü,Fÿ*Iÿ.Mÿ7Xÿ8Yÿ;\ÿ=^ÿDcÿFeüIgýJhøIh÷Hg÷EiöDhüClÿFoÿIsÿKuÿ`‚ÿXzïUqçMiÕVkÛ\q«PY�&/\&$a+)Q>0Q>0:?+6;'4F.AS;YO,aW4d`=ieBouS…c}Œkw†ew�iu‹grŠfrŠfvŒfxŽhz�hz�h€’x†˜~‹�‡�Ÿ‰“¡�™§–š¨›—¥˜’ ‘‰—ˆyŒxpƒofa^wYSnKKfCUVDUVDVWE58/58/6907:1:;3<=5>?7@A9AA9AA9C@9C@9DA:DA:DA:DA:EB;EB;EB;EB;FB9FB9FB9FB9JC9IB8I@7H?6G?4E=2D<1D<1C;0C;0C;0C;0D;2D;2C<2C<2=90=90<8/<8/:7096/85.85.96-96-96-:7.<8->:/A;/B<0D</E=0I=-K?/M@/M@/M@/M@/M?2L>1K=2J<1H<0G;/D:0D:0D<1C;0@<1@<1@<1A=2@=4A>5Q9/J2(Š9?­\bÛTeßXiÛWbËGR“91”:2Œ@0�D4±D?»NIÚ>IÕ9D¤I6¢G4¨B3¨B3®>0¯?1°B3³E6¯H5¨A.¢>'¡=&¨;'¨;'«6%ª5$µ8>·:@¶9?µ8>³BD½LN³\U´]V‰G9|:,~0&Œ>4ÂGLÐUZâASÕ4FÜ:5Ï-(Î95Ë62½HAÎYR›\K”UDj\?tfIstRtuS~iJ€kL¤r[µƒlþJVøDPý8Jü7Iÿ7Nÿ:Qÿ4Nÿ-Gÿ;Sÿ5Mþ2Ký1Jÿ/Iþ.Hÿ.Mÿ3Rþ2Sÿ3Tû5Vü6Wö;Z÷<[ò?]ó@^óDcòCbò@dò@dø?hûBkÿEoÿGqþSuýRtôZvïUqÙZoâcxÈmv´Ybn86f0.K8*L9+=B.<A-4F.7I1TJ'^T1ea>okHx~\ˆŽlƒ’q|‹jxŽjz�ly‘mw�kw�gxŽh”m„™r’¤Š˜ª�Ÿ±›¢´ž¨¶¥¬º©¬º­¨¶©ž¬�“¡’€“q„pd}_\uWRmJLgDUVDVWEXYG47.58/58/7:19:2<=5>?7?@8AA9AA9C@9DA:DA:DA:DA:EB;FC<FC<FC:FC:GC:GC:IB8IB8KC8KC8JB5H@3H>2G=1F=.E<-D;,D;,D:.D:.C;.D</D<1D<1@9/@9/=90<8/<8/;7.:5/:5/97+97+;7,;7,>8,@:.C;.D</F=.G>/J?-K@.M@/M@/M@-M@-N>1N>1K=0J</H<0G;/D:.C9-B<0B<0@<0@<0?=1@>2??3@@4j/)q60½DMØ_hõEYì<PÚ5<Ä&¤4&£3%�<"’>$±=.¼H9Ý:;Ø56§C3§C3­?2¬>1´>4¹C9»G:¼H;ª?-©>,§<*¥:(¬7&°;*º:-»;.ÌF=ÌF=Ø:9Ó54á27ê;@ëGHîJKÍ@9ÓF?ÕOFØRIäFEÝ?>é19æ.6Ø4(Û7+Ï2)Ô7.¼91¿<4®_PžO@uP6…`F‹hL„aEŸTA°eRÛhaáng÷>Fò9Aá<@ã>BÙJFÛLHÛMKÜNLíNRæGKò9Gñ8Fú1Dú1Dÿ2Hÿ8Nÿ:Vþ9Uü:Uý;Vþ;[ÿ=]þ?aÿ@bÿBkÿBkýBmýBmüCoüCoúCqúCqÿClÿ=fÿFkÿInõHfõHfèZpèZp½Yc›7A]("Z%RB2[K;HM6>C,UH5dWDrkY€ygŠŒ~˜šŒ‘›�‡‘†�Ž„ˆ•‹Žš�‹—�‰�…Š�†™›�¥§œ³³§ºº®¿¿·ÁÁ¹ÅÁÀÇÃÂÃÁľ¼¿±³°¤¦£‹—‹y…yh|c^rYRmLNiHQZ?S\AU^C47.47.47.58/891:;3;<4<=5??7@@8B?8C@9DA:EB;FC<FC<FC<GD=GD;GD;HD;IE<KD:KD:KC8KC8JB5IA4I?3H>2G>/G>/F=.F=.F<0E;/D</C;.C;0C;0A:0@9/>:1=90<8/<8/;60;60:8,:8,;7,<8-?9-A;/C;.D</G>/H?0K@.LA/M@/M@/M@-M@-N>1N>1K=0J</G;/F:.C9-C9-B<0A;/?;/?;/?=1@>2@@4@@4q60‡LFÍT]ÓZcë;Oã3GÖ18É$+¤4&¤4&�<"�<"«7(³?0Ô12Ñ./Ÿ;+ <,¦8+¦8+°:0·A7ºF9¹E8ª?-ª?-¦;)¡6$ª5$³>-ÄD7ÈH;½7.ÈB9áCBãEDí>Cê;@Ý9:Õ12Ì?8ØKDÝWNÙSJÛ=<Ñ32å-5ì4<Õ1%Þ:.É,#Ó6-¸5-¶3+»l]ºk\�\B‚]C|Y=|Y=¨]J¾s`ÛhaÐ]VúAIô;Cá<@ß:>ÑB>ÔEAÝOMæXVíNRæGKò9Gð7Eú1Dø/Bÿ/Eÿ4Jý8Tý8Tû9Tü:Uþ;[ÿ>^ÿBdÿDfû<eü=fú?jú?jù@l÷>jó<jò;iÿDmÿ7`ÿ>cÿInûNlùLjçYoåWmÛw�Ã_i†QKp;5L<,J:*CH1LQ:]P=reR„}k”�{›��¥§™�§œ“�’’Ÿ•›¨ž¤°¦£¯¥¡§�£©Ÿ´¶«ÁøÉɽÎÎÂÒÒÊÑÑÉÓÏÎÓÏÎÎÌÏÈÆɼ¾»¯±®” ”‹kf`t[UpORmLV_DV_DW`E47.47.47.47.67/7809:2:;3==5>>6A>7B?8DA:EB;FC<GD=GD=GD=GD;HE<IE<JF=LE;LE;LD7LD7MC7LB6NB4MA3MA1L@0JA0I@/I@1G>/E=0D</C;.C;.A:0A:0@91@91=90<8/<71;60:8,:8,;7+<8,?9-A;/C<,D=-H?.H?.K@.LA/M@-M@-O@-N?,N>1M=0L</K;.H:-G9,C9-C9-@:.@:.>:.?;/?=1@>2??3@@4ƒ('¦KJÄPQ´@A»11¼22¹3(¸2'¤4§7™?–<”>—A �@!™<œ5,ž7.¤5,£4+­81¶A:µC9²@6®B6­A5§;.£7*¯9/¹C9ÄC=Á@:Ä>#ÇA&ã1%Ý+ûÿ $ÿ%+ÿ'-ó&#ô'$å0%æ1&ä2&ä2&ë/&ë/&Õ6!Õ6!Ð2&Ò4(Á80Ä;3ÀbVØzn£fS’UB�K8žZGØ]XÞc^øJSî@IõGIî@BÖC<Ó@9ÁC7ÀB6ÓF?ÝPIöEMñ@Hÿ4Gÿ4Gû5Fù3Dñ9Eö>Jú8Pù7Oû6Rû6Rÿ5Xÿ8[ÿ9cÿ;eÿ=lÿ>mÿ@qÿ@qÿBvÿAuúBvùAuûHföCaúGeþKiýLiÿTqÿ^yûXsòi{òi{Ónv»V^†B?s/,m>6~OGte`‹|wŸ•“¬¢ ®®°µµ·°µ»©®´°·Áµ¼Æ¾ÀÌ¿ÁÍÄÁÌÈÅÐÖÌ×ÞÔßãØÔçÜØèÜÞæÚÜæÖàæÖàßÓÝÙÍ×ÉÄʼ·½ § Š‘Šr„nfxb[tVXqSZiJZiJZiJ47.36-36-36-56.67/780891;;3<<4@=6A>7C@9EB;GD=GD=GD=GD=HE<HE<JF=JF=MF<MF<ME8ME8ND8ND8PD6OC5OC3OC3MD3LC2KB3I@1G?2E=0D</D</B;1A:0A:2@91=90<8/<71<71:8,:8,;7+<8,?9-A;/C<,D=-H?.I@/LA/LA/M@-M@-O@-N?,N>1M=0L</K;.H:-G9,C9-B8,@:.@:.>:.>:.><0?=1??3@@4‘65º_^ËWX«78³))·--²,!´.#ª:"¯?'¤J(¡G%›E$šD#�@!™<Ÿ8/¢;2§8/¦7.¯:3¸C<¶D:±?5¨<0¨<0§;.©=0ºD:ÃMCÄC=¸71½7Å?$ç5)æ4(ÿ"&ÿ"&ÿ &ÿ$ø+(ù,)é4)é4)æ4(ä2&ê.%ê.%Ü=(Ð1Û=1Ò4(Â91Æ=5¤F:Ìnb­p]–YF™UB³o\äidÕZUì>GñCLë=?æ8:Ñ>7Ñ>7½?3¶8,À3,Å81ñ@Hí<Dý2Eÿ5Hþ8Iü6Gô<Hø@Lÿ?Wÿ>Vÿ:Vþ9Uÿ6Yÿ7Zÿ7aÿ8bÿ=lÿ=lÿ=nÿ=ný>rý>rø@tø@töCaÿLjÿSqúGeð?\ÿQnÿhƒÿdëbtêasÕpxØs{ºvsªfcŒ]U…VN��|§˜“»±¯Äº¸ÂÂÄÇÇÉÃÈοÄÊÅÌÖÆÍ×ËÍÙÏÑÝÚ×âàÝèêàëìâíïäàòçãóçéïãåïßéîÞèèÜæã×áÏÊн武�–�vˆri{e_xZ\uWbqRapQ_nO00.11/3315536727838938939:2:;3<>3>@5@B5BD7CE8DF9GI>FH=GH8GH8HJ4IK5LL2MM3ML8IH4ED?GFAJHSKITECX@>SOF7LC4JA2KB3LD7KC6F>3B:/D=5B;3@93?82>95=84;63:52C84C84@85@85;:8<;9<<<<<<?@;?@;D@5EA6LA/K@.O?(O?(AC.@B-K:2I80N45N45H8;H8;;B;3:3,?,4G49C+1;#B<$QK3ÒaQºI9£8&¢7% ;'�8$�:#£@)ž:#¡=&¬=)­>*³:)°7&²3"²3"Ÿ8'Ÿ8'Ÿ:(¢=+¦A-©D0ªC0¨A.¨=)©>*°;)³>,ÀB3ÄF7Á=0¶2%Ì5*Í6+Ê;+Ì=-Ð<.Ñ=/á6.á6.ñ.,ñ.,ô.+õ/,å8*ä7)Ó>(Ó>(Ü:%Õ3Ì4Ï7"Ä8'À4#Ä@3ÕQDÓHAÙNGñSR÷YXýHMð;@ù4=ÿ:CàDEÞBCÏD?É>9´>2±;/¨?,©@-ÀO?ÄSCÖI@ÒE<é<>ðCEÿ<Fü8BùFLøEKúDPüFRüATø=Pü>XÿD^ÿGkÿBfö=fó:cò>k÷CpùIxüL{ÿMyÿKwïTtìQqÞPhßQiþVwÿf‡ÿZƒÿ@iüTxýUy¹aoÊr€vi`ˆ{rª��À¦—ÍźÞÖËØÜÛÙÝÜÞÚèàÜêíÓîîÔïøÖïûÙòøâïûåòñíîòîïøõð÷ôïõñîóïìïééêääæàâãÝßÕÑÒÍÉʺº¸¤¤¢‹�‰{€yr|qpzo~…sxmu|j//-00.11/3315616727827829:2:;3;=2=?4?A4AC6BD7CE8FH=FH=GH8HI9HJ4JL6LL2MM3KJ6KJ6KJEQPKWU`YWbUShQOdRI:MD5H?0G>/G?2H@3F>3C;0C<4A:2?82>71=84>95<74;63=2.>3/=52>63:97<;9<<<===?@;?@;D@5EA6LA/K@.O?(O?(>@+>@+J91J91N45L23D47C369@96=61D14G48B*8B*OI1`ZB¼K;«:*�2  5# ;'�8$œ9"¢?(Ÿ;$£?(®?+¯@,µ<+³:)¶7&¶7&¢;*¡:) ;)¢=+¥@,§B.¨A.§@-ª?+¦;'±<*¼G5ÅG8¾@1»7*»7*Í6+Í6+Ê;+Ì=-Ð<.Ñ=/á6.á6.ñ.,ñ.,ô.+ô.+ã6(ã6(Ñ<&Ñ<&Ý;&Ø6!Ñ9$Ô<'Ç;*À4#¿;.ËG:ÐE>ÕJCëMLðRQùDIð;@û6?ÿ=FàDEÞBCÎC>Ç<7³=1¯9-¦=*§>+µD4¿N>ÙLCÖI@é<>ì?Aÿ<Fÿ=G÷DJöCIúDPüFRýBUù>Qû=WÿB\ÿFjÿCgù@iø?h÷CpûGtûKzüL{ÿLxÿHtêOoîSsí_wðbzýUv÷OpÿBkÿAjíEiêBf¼dr¾ft^QHŒv°–‡Ì²£ÝÕÊìäÙåéèæêééåóçãñùßúûáüÿâûÿåþÿîûÿðýýùúýùúýúõüùôû÷ôøôñöððòììïéëíçéãßàÚÖ×ÇÇŲ²°› ™Ž“Œ†�…†�…Ž•ƒ‰�~†�{--+..,/0+12-23.4505606718919:2:<1<>3>@3@B5BD7BD7IG;IG;JH9JH9LJ5MK6NL7NL7KI=NL@SRW_^cggll„il�fiŒ`XmXPeNGWG@PD>JD>JB=C?:@@<=>:;;83:72=90>:1>:.=9-74+85,:7.<90=;/><0></></E?1E?1J>.K?/O?/N>.P=,P=,5?$6@%;<*<=+>:/<8-961850=82@;5R42O1/j"&s+/œ1;«@J¨:) 2!�2 £8&¢<&ž8"œ8!¡=&¥;%¨>(²?,´A.¼=,»<+¿:+¿:+¯>.¬;+ª9)ª9)«:*­<,­<*­<*´<+¯7&»=.ÉK<ÌG8»6'½1$Ç;.Ô4(Õ5)Ô8+Õ9,Ù9+Ú:,æ3,æ3,ô.+ô.+ò/)ñ.(á6%ß4#Ï:"Î9!à7$Ý4!Ú9'Þ=+Ò<.Ë5'Å8.Í@6Æ;4Ê?8ØD@àLHèJGäFCçB@éDBÖF>ÔD<ÉC8Ã=2³;-¯7)©:)ª;*®6(¾F8×LE×LEã?@ä@AöAHýHOòAGòAG÷CLüHQþEUù@Pú>WþB[ÿEfÿDeýDmþEnþIvÿKxþM{ýLzÿCuÿDvÿJrÿKsÿYzÿ_€ÿUrïEbøKiÿSqÐEZÑF[Ê‚…¿wzUH8‹~n¶Œ€Û±¥òÜÑÿêßöòñúöõ÷ôÿðíøûêýýìÿÿëÿÿîÿÿõÿÿ÷ÿÿýüÿýüÿþùÿýøÿû÷þùõýôõûòóûðôúïóïæçæÝÞÓÏÌÀ¼¹¬¯¦¢¥œ�£—�£—£«œ�¥–š¢“--+--+./*/0+01,23.34.45/7808919;0;=2=?2?A4@B5AC6HF:IG;JH9KI:MK6NL7OM8OM8KI=SQE]\anmryy‘��™‚¥}€£}uŠsk€f_o[TdSMYMGSFAGB=CC?@?;<;83961;7.<8/<8,<8,:7.;8/<90>;2><0><0=;.=;.E?1E?1J>.K?/O?/N>.P=,P=,7A&7A&<=+<=+>:/=9.<94<94=82@;5P20M/-n&*|48¢7A¨=G¦8'¤6%¥:(©>,¥?)Ÿ9#�9"¡=&¤:$¨>(²?,´A.½>-½>-Â=.Ã>/´C3°?/¬;+«:*«:*¬;+¬;)¬;)³;*µ=,ÀB3ÆH9Ä?0¹4%¿3&È</Õ5)Ö6*Ô8+Ö:-Ú:,Ú:,æ3,æ3,ô.+ô.+ñ.(ð-'ß4#Ý2!Ì7Ë6Ú1Ù0×6$Ü;)Ô>0Ð:,Ë>4ÑD:Á6/Á6/Í95ØD@èJGçIFå@>Ý86Ð@8Ï?7Å?4À:/²:,°8*«<+­>-«3%¹A3ÑF?ÔIBâ>?ä@AøCJÿKRï>Dï>DöBKüHQÿGWüCSû?XýAZþBcÿCdþEnÿIrÿMzÿN{þM{üKyÿBtÿL~ÿT|ÿIqõMnþVwÿVs÷MjòEcõHfË@Uê_tÕ��PC3]P@¤znЦšîØÍÿêßùõôÿýüþûÿöóþþíÿÿîÿÿíÿÿïÿÿõÿÿöÿÿûúÿûúÿþùÿýøÿû÷þùõÿö÷þõöþó÷þó÷õìíëâãÙÕÒÈÄÁ¸»²²µ¬¯µ©°¶ª¶¾¯¯·¨«³¤10+0/*0/*0/*0/*21,33+44,77/88099/;;1==3??5@@4AA5JD8KE9NG7PI9QJ7RK8QK;QK;NJGXTQcdxwxŒƒ†±Œ�º‹’É‹’ÉŒ‹Çƒ‚¾yv­nk¢ed�[Z†QPpJIiHGYCBT=<A87<76165086*75)2713828:/9;0>;*?<+B<$B<$I?&I?&J?+K@,JA2I@1I?5I?5P=7N;5G:4D719:48931>41>46>15=0J2(N6,~43�CB¬;A¨7=§<*¨=+©@+«B-¨?*¤;&¥;%¨>(ª7$­:'¸;)º=+Ã9,Ä:-Ê:/Ì<1Ä@4À<0¼8,º6*º6*»7+»7*º6)½5)È@4ÏA5Æ8,Ã/%Æ2(Í6+Í6+á0(â1)á3(â4)è3*è3*ï0(ï0(õ.)õ.)ð/&ï.%Þ4#Ü2!Ì7Ê5×0Ö/Õ2!Ù6%Ö8,×9-Ó?5ØD:ÔE=Î?7Å<2ÊA7ÑMAÕQEËI<¿=0Á=.Á=.¼=,º;*´9)´9)¶<-¸>/¸8-¼<1É>9ÎC>ØBCÜFGéJNðQUì?Cì?CòCJùJQýKYùGUøBXùCYøA_úCaûFiÿLoÿNwÿOxÿMwýJtÿKwÿT€ÿU|ýIpùJkÿStÿVpøNhæG[åFZã[gÿ“Ÿÿ·¹ý™›‘FAt)$©jaÓ”‹ñÈÂÿßÙÿóðÿûøÿþÿù÷üûøÿüùÿÿøÿÿùÿÿûþÿûþÿýúÿýúÿüöÿûõÿú÷ÿùöÿ÷øÿö÷ÿôöÿóõüòñòèçäÜÙ×ÏÌÍÊÁÊǾÉ˽Ê̾ÆÐȽǿ·Á¹54/43.21,10+10+10+33+44,66.66.88.::0<<2>>4??3@@4IC7KE9OH8QJ:SL9SL9RL<QK;OKH[WTij~~“‹Ž¹–™Ä—žÕ˜ŸÖ™˜Ô“’ÎŽ‹Â‡„»€«wv¢mlŒfe…]\nUTfLKPCBG>=8:94:8,97+27127157,68-<9(?<+D>&E?'I?&I?&J?+K@,JA2I@1I?5I?5N;5L93F93C60671560-:0-:05=008+F.$S;1ˆ>=‘GF©8>¢17§<*¨=+¨?*©@+¨?*¨?*ª@*¬B,ª7$­:'¸;)º=+Ã9,Ä:-Ê:/Ì<1ÅA5Â>2¿;/½9-¾:.¾:.½9,¼8+Á9-ÍE9ÏA5À2&Á-#Í9/Ó<1Ì5*â1)ã2*â4)ã5*è3*è3*ï0(ï0(õ.)õ.)ð/&ï.%Þ4#Ü2!Ì7Ë6Û4"Ú3!Ö3"Ö3"Ó5)Õ7+Ñ=3Ó?5èYQáRJÒI?ËB8ÊF:ÍI=ÇE8¿=0½9*½9*º;*¸9(´9)¶;+¹?0¼B3ÁA6¾>3Æ;6ÌA<ÙCDÝGHçHLëLPí@Dì?CñBIùJQþLZüJXûE[üF\÷@^ùB`úEhÿJmÿMvÿOxÿNxþKuÿO{ÿO{ÿNuÿLsÿVwÿ\}þTnîD^ôUiïPdðhtÿ’ž÷“•ë‡‰´idœQL­neÎ�†è¿¹ÿÜÖÿóðÿùöþüÿüúÿýúÿýúÿÿùÿÿùÿÿúýÿúýÿûøÿûøÿüöÿüöÿüùÿûøÿùúÿö÷þòôýñó÷íìðæååÝÚÞÖÓÙÖÍÜÙÐÜÞÐÞàÒÙãÛÎØÐÄÎÆ::288055-33+22*22*33)33)55+66,77-99/;;1==3>>4??5KA5MC7RF6UI9VK9WL:TJ>TJ>PKQ]X^kl‹�‚¡�“Í™ŸÙœ¥ðŸ¨óž£ôšŸð—›å“—á‘•ÕŒ�І‰¾€ƒ¸xz¡np—ccVVrNM_HGYECPB@M<>M:<K;:@98><90=:1C>*D?+CA*CA*AB0BC1?E;>D:;DA;DA=<N?>P?BQADS>HQ>HQ=IG<HFFH=<>3O8(`I9…N:…N:�@,‰<(§<(¨=)©>*©>*¬?*®A,±>+°=*³6$¶9'Á7*Â8+Ê6*Ë7+Ò8.Ô:0Ï:3Í81Ì70Ì70Ï81Ï81Î7.Ì5,Ô<1Õ=2Ô7.Í0'Ñ3(Ù;0Û:0Õ4*è0&è0&ì0'í1(ó.(ó.(ö,(õ+'÷.'÷.'í3'ì2&Ü7$Ú5"Ï9!Î8 Ò:%Ñ9$Ï6$Î5#Ñ5(Ó7*Ñ7-Ñ7-ãNGêUNäZPÛQGÉI>ÃC8»C5·?1µ=$µ=$¶<%µ;$¸9&¹:'À</Ä@3ÉC:Å?6É@:ÐGAÓIGÒHFÓIG×MKèDEæBCêDHóMQúPYùOXøL\øL\õD^õD^÷EeûIiÿNrÿQuÿPwÿOvë\xéZvñWsõ[wÿ\vÿ[uúSgñJ^ýXiøSdìKZûZiéFWñN_ôM_òK]¶kfÌ�|㬧ÿÐËÿíéÿñíýùøÿýü÷ýý÷ýýøþþ÷ýýûûùúúøûøóûøóÿûõÿüöÿüùÿüùÿ÷÷ÿóóüîîùëëïáàêÜÛâÙÒâÙÒåßÓíçÛïîÜòñßèóïÚåáÍØÔ==5;;388044,22*22*33)33)44*55+66,88.::0<<2>>4??5KA5MC7RF6UI9WL:WL:TJ>TJ>PKQ]X^lmŒƒ„£�•Ï�£Ý¢«ö¦¯úª¯ÿ¥ªû¢¦ðŸ£íŸ£ãœ à˜›Ð”—Ì‘“º‡‰°zz–jj†`_qWVhSQ^PN[KM\GIXDCI?>D>;2=:1A<(B=)CA*CA*AB0BC1?E;>D:;DA;DACBTHGYORaX[jZdm[enZfdYec\^SPRGbK;r[K�YE‡P<�@,ŽA-«@,«@,«@,«@,¯B-°C.°=*¬9&µ8&·:(Â8+Ã9,Ë7+Ì8,Ó9/Õ;1Î92Í81Í81Î92Ò;4Ò;4Ð90Î7.ÛC8Ñ9.Ï2)Ô7.Û=2Û=2Ú9/Ø7-è0&é1'í1(í1(ó.(ó.(ö,(õ+'÷.'÷.'í3'ì2&Ý8%Û6#Ð:"Ï9!Ï7"Ï7"Ï6$Î5#Ò6)Ô8+Ñ7-Î4*É4-âMFðf\ëaWÐPEÁA6µ=/²:,¶>%¶>%·=&¶<%¸9&¹:'À</Ã?2ÉC:ÈB9ÎE?ÕLFÕKIÏECÎDBÒHFéEFçCDêDHòLPûQZúPYúN^úN^÷F`öE_÷EeûIiÿNrÿQuÿRyÿRyçXtê[wø^zù_{ûTnóLføQeÿZnÿ\mÿctîM\öUdñN_ùVgóL^òK]°e`¿toÔ�˜þÇÂÿçãÿìèúöõÿþýúÿÿúÿÿúÿÿúÿÿüüúûûùýúõüùôýøòÿúôÿú÷ÿú÷ÿõõûïï÷ééóååêÜÛçÙØâÙÒæÝÖîèÜøòæüûéÿÿíðû÷àëçÒÝÙ@>2@>2><0<:.:8,86*74+63*52)52)52)74+96-=:1@=4B?6I?3LB6RF6TH8VI8WJ9YMA[OCPIPXQXddˆ}}¡‰’Ó–Ÿàš©ü¡°ÿ¨´þ¨´þª´ü©³û©²÷¦¯ô§¬î¥ªìŸ¡Þ—™ÖŒ�Å~·sp¥gd™\ZŒUS…YPYPVOqMFhC?N=9H<<:??=<?4?B7?G<>F;9B?:C@=FKAJO:YxJiˆ^q�bu“xrŒ}w‘Ži{‚]ošR]šR]­PK®QL±G7¦<,£7¤8©<(ª=)­>*¯@,´?-´?-¹<*¸;)½5'¾6(Ç5(É7*Ò5,Õ8/Ú7.Û8/Ø61Ú83Ú83Ù72Ù6/Ø5.Û81Ý:3çC:à<3Ù6-â?6ý[Pÿmbÿ]RãA6è0&ê2(ó0*ò/)ö)&ô'$ø(&ú*(ô/)ô/)è3(å0%Õ2#Ó0!Ñ6$Ô9'Å=%Æ>&Ê<&Ë='Ð:)Ð:)Ñ7+Ñ7+Ú@6Ð6,ÜGA÷b\äWPÍ@9Æ?9¹2,µ;#¶<$º<%º<%½:(¾;)Ã;-Å=/Æ=3ÌC9ÕPIÝXQÐQJ¿@9½@:ÈKEáGEàFDåKKëQQîRVëOSìLVîNX÷RcõPaôLeóKdúMkÿRpÿStÿQrïYvîXuöWsú[wÿ[tÿYrÿWjþUhÿcqýWeÿ\kÿ[jôEVóDUÿOeÿLb³ph¬iaÆ‘‹é´®óÔÑÿèåøðíýõòõþûùÿÿ÷ÿÿõÿýûýúýÿüÿûøûöóýúóþûôÿøõþöóÿ÷õÿöôöèçæØ×ãÕÒÞÐÍàÖÌðæÜüõãþ÷åÿýèÿÿì÷ÿýçðíØáÞB@4A?3?=1><0<:.:8,85,74+52)52)52)74+96-<90?<3A>5H>2KA5QE5SG7UH7VI8XL@YMARKRXQXbb†zzž†�ГœÝ˜§úŸ®ÿ©µÿª¶ÿ­·ÿ®¸ÿ®·ü­¶û¯´ö®³õ¨ªç¡£à—˜Ð‹ŒÄ�~³ur§ig™b`’`W†`W†]VxVOqNJYFBQBB@AA?;>3@C8AI>?G<9B?<EBFOTQZ_PoŽ^}œq„¢r…£‡�›Š„žšu‡Ži{¤\g¥]h³VQ«NI¬B2§=-©=#«?%¥8$¦9%¨9%©:&®9'®9'´7%´7%»3%¾6(É7*Î</Ù<3Ú=4ß<3Þ;2ß=8Þ<7Ý;6ß=8äA:çD=æC<äA:ä@7åA8â?6Û8/Ó1&×5*êH=þ\Qö>4í5+ì)#ì)#ù,)ý0-ü,*ø(&ö1+÷2,ì7,é4)Ù6'Õ2#Ð5#Ñ6$Ä<$Å=%É;%É;%Î8'Í7&Î4(Î4(Ó9/Ï5+Ò=7åPJãVOÚMFÉB<¶/)·=%·=%»=&º<%½:(½:(Á9+Â:,Æ=3ËB8ÓNGÜWPÑRKÂC<¾A;ÆICâHFáGEåKKìRRîRVëOSìLVíMWùTeùTeúRkúRkþQoÿSqÿRsÿRsèRoîXuü]yÿa}ÿ[týUnûReûReüVdøR`ÿYhÿYhúK\ÿPaÿRhüE[ªg_�ZRµ€zܧ¡éÊÇþßÜóëèýõòôýúøÿþ÷ÿÿõÿýûýúýÿüÿüùýøõüùòýúóÿ÷ôþöóÿ÷õÿöô÷éèéÛÚäÖÓÝÏÌÞÔÊïåÛüõãÿùçÿýèÿÿìöÿüçðíÙâßEA5D@4C?3A=1?;/=9-;7,;7,73(73(73*84+:6-<8/?:4@;5E>4HA7NE6PG8SF3UH5WJ:YL<ULOYPS_^}vu”�‰Ç�˜Ö”¤òœ¬úª´ù«µú®¸ý°ºÿ²»ÿ²»ÿ³ºÿ²¹ÿ®²û¨¬õ ¢ì–˜â�ŒØ�€ÌwtÁol¹j_£g\ bZ“]UŽSSwKKoCDV=>P86;?=BO?@O?@W:>`CGwRY‡bi§iš±s¤¾w£¼u¡Ém”Çk’ÕZyÊOnÓ?UÙE[ÝDIÐ7<½5'½5'µ="³; ®;(®;(²:)³;*¸;)º=+À;,Á<-Å6(È9+Ô:0Ù?5á>5á>5â;3à91Û81Ø5.Ø3-Þ93èA9ìE=éB:ã<4â;2Û4+×5(Ù7*Ò2$Ë+Ò6'åI:õK<íC4î3,ê/(ø**ú,,û*-ù(+ë+(í-*ã5.â4-Ù5+Ö2(Ô2'Ô2'Ê9&Ê9&Ì9'Î;)Ñ:)Ñ:)Ô8+Ô8+Ð4(Ø<0Ó:4Ô;5æNMïWVÙEEÄ00À;(À;(¿:'¿:'Á9)À8(À8*¿7)Ä>2Æ@4ÎL?×UHÏSGÃG;¾D7ÂH;ØKB×JAÛNGáTMãUSßQOßNQàORðXdó[gú[mú[müVnøRjúPküRmÿCkÿLtÿZyÿYxÿXoýTkúWhþ[løZh÷Ygôblò`jácgèjnÜnmÊ\[ŒfYwQD•of¿™�Ùº·ñÒÏïáàÿóòöû÷úÿûöÿþõÿýüþýþÿÿÿûþÿ÷úùúòùúòûøñüùòÿùöÿøõøïêíäßåÛÒÛÑÈÛÔÄìåÕü÷áÿüæÿþåÿÿçöýöéðéÝäÝHD8GC7FB6D@4B>2@<0?;0>:/:6+95*84+95,:6-<8/>93?:4D=3G@6LC4NE6RE2SF3UH8WJ:VMPXOR[ZypoŽ|„ÂŒ”Ò‘¡ï™©÷¦°õ¨²÷«µú­·ü°¹þ°¹þ±¸ÿ±¸ÿ®²ûª®÷¤¦ð�Ÿé–•áŒ‹×‚Ì|yÆwl°pe©g_˜aY’XX|PPtFGY>?Q:8=?=BO?@TDEdGKqTXˆcj•pw®p¡³u¦¼u¡¶o›¿cŠ¸\ƒÅJiº?^Æ2HÔ@VÞEJÓ:?Â:,Æ>0ºB'±9ª7$ª7$­5$®6%µ8&¸;)À;,Â=.ÐA3ÑB4Ù?5Ú@6ß<3Ü90Ü5-Ú3+Ò/(Ð-&Ó.(Ø3-à91â;3ß80Û4,Ü5,Ñ*!Ì*Ø6)áA3Ý=/Ð4%É-íC4óI:ÿE>ø=6÷))ñ##ö%(ý,/î.+î.+á3,ß1*Ö2(Ô0&Ò0%Ò0%Ì;(Í<)Ð=+Ò?-×@/ØA0Û?2Û?2Ñ5)èL@ãJDÎ5/ßGFòZYÞJJÕAAÂ=*Á<)À;(¿:'Â:*Á9)À8*¿7)Á;/Â<0ÈF9ÐNAËOCÃG;½C6½C6ÓF=ÒE<ÖIBÝPIÞPNÛMKÚILÜKNíUaó[gþ_qÿ`rýWoôNfõKfùOjÿLtÿPxÿVuÿQpýTkÿXoÿbsÿfwÿesþ`nôblóakãeiâdhÒdcÊ\[‡aTkE8‚\S­‡~Ê«¨æÇÄéÛÚÿñðôùõøýùõÿýôÿüüþýþÿÿÿüÿÿùüúûóúûóüùòýúóÿûøÿûøþõð÷îéêà×ÞÔËÜÕÅìåÕýøâÿþèÿþåÿÿæôûôéðéßæßKH7JG6IF7GD5EB3C@1B>2A=1<8-<8-;7.:6-;60<71>93?:4A>5DA8JD4LF6PE/QF0VG2WH3XMIXMIXVlnl‚z�µ‹’Æ�Ÿä˜§ì§­÷¨®øª°ú¬²ü®´þ®´þ¯³ý¯³ýª®ø¨¬ö£§ñŸ£í™�瑕߉�׃‡Ñ�€Æxw½imª`d¡[a‘TZŠTQpLIh^CV^CVz<G‚DO©EMµQYÎRZÒV^ôFg÷IjíOfèJaÛGUÓ?MÏ5?Ç-7Ú6=æBIåGFÚ<;È@4ÎF:¾I5´?+·<,·<,»<+½>-Ä?0ÇB3ÑC5ÓE7ÚC8ÙB7Û=2Ù;0Ú6-Ø4+Ù2)Ù2)Ó2(Õ4*Ù6-Û8/Ü8.Û7-Ü5,Ü5,Ò0#Ø6)Õ8)Ï2#Ê3 Ñ:'Ñ>*Ï<(Ô=,ßH7÷I@öH?ü46ó+-ô),ø-0ñ86î53Þ50Ú1,Õ0,Õ0,Ú.,Û/-è50è50å9/æ:0â<.â<.Û>/Û>/Ï7*ïWJöaZÐ;4Ï;9äPNÛFHäOQÏ<2Ë8.Ç7,Æ6+Ä:-Å;.À<-¾:+º;*»<+¿B0ÅH6ÆI7ÂE3½@.¼?-ÇG:ÆF9ÊJ?ÑQFÓRLÐOIÑMKÓOMæ[`ë`eüfrÿiuü]oïPbòLbøRhÿSwÿQuýWoùSkëblôkuõnuðipøfqñ_jçajîhqÐxvÁig˜m]™n^pnYPN9jXL�~r¼žœÞÀ¾ëÕ×ÿëíúöõþúùùÿÿøþþýýÿÿÿÿÿúÿÿ÷ÿûÿ÷øþôùüóüÿöÿÿøÿÿøÿúòÿøððèÝãÛÐßÚÇîéÖýûâÿÿæÿÿáÿÿáõùëíñãæêÜNK:NK:LI:JG8HE6FC4EA5D@4@<1?;0=90<8/<71=82>93?:4B?6DA8JD4LF6PE/PE/UF1WH3WLHVKGVTjkiy€´Š‘Å�žã–¥ê¦¬ö§­÷©¯ù«±û¬²ü«±û­±û¬°ú¨¬ö¦ªô£§ñ¡¥ïž¢ìšžè•™ã‘•ß’‘׉ˆÎ{¼sw´nt¤ioŸkh‡da€z_rv[nŽP[“U`·S[¾ZbÎRZËOWé;\ê<]àBYßAXÕAOÐ<JÏ5?Ë1;éELêFMßA@Õ76Å=1ËC7½H4·B.ÀE5ÀE5ÄE4ÅF5ËF7ÍH9ÕG9ÖH:Õ>3Ó<1Õ7,Ó5*Ö2)×3*Û4+Ü5,Ù8.Ü;1ß<3ß<3Ü8.Ù5+Û4+Ü5,Ñ/"Ü:-Ú=.Î1"Å.Î7$Ð=)Ì9%Ï8'Ñ:)å7.î@7ÿ=?ÿ<>þ36ö+.æ-+ã*(Ô+&Ô+&Õ0,Ú51ä86æ:8ì94ì94ç;1å9/Þ8*Ü6(Ó6'Ò5&Í5(çOBÿkdÛF?È42ÖB@Ó>@êUWÖC9Ð=3È8-Æ6+Ä:-Æ</Á=.À<-¹:)»<+½@.¿B0¿B0½@.»>,º=+ÃC6ÃC6ÈH=ÏODÒQKÐOIÑMKÓOMé^cíbgýgsÿnzÿcuôUgöPfüVlÿRvÿPtýWoþXpõlvüs}ñjqßX_ò`kò`kéclõox㋉Õ}{¢wg uepnYPN9eSGƒqe±“‘ܾ¼ëÕ×ýçéúöõþúùùÿÿøþþýýÿÿÿÿÿúÿÿ÷ÿüÿøùÿõúýôýÿ÷ÿÿøÿÿøÿýõÿþöùñæíåÚèãÐôïÜþüãÿþåþÿàÿÿâ÷ûíðôæíñãQN=PM<NK:MJ9KH9IF7GC7FB6C?4B>3@<3>:1>93>93?:4@;5?@:BC=IF7JG8PE/PE/UF/WH1UKASI?TR`kiwz�­‹’¾��Ú•£à¥¦ù¦§ú¨©üª«þ¬­ý¬­ý«®ûª­ú§«ô¥©ò¡¨ì §ë ¨éŸ§èœ¥æš£ä•žß�™Ú‰•Ñ„�Ì��½}‰¹…£�{Ÿ–i€’e|·Q_¸R`âENãFOø9@õ6=è?,è?,ÖI-ÖI-ÅN,ÁJ(¾G)¼E'âUDÙL;Û96×52Û57Ý79Õ6:Ð15»<+»<+¾:+¾:+Ã7*Â6)Å3&Ä2%Î4(Î4(Ó2(Ó2(×3)Ø4*Ú6,Û7-Ü>3Û=2Û;/Ú:.Ú8-Ø6+×5(Õ3&Ô4&Ó3%Í4"Ï6$Î=(Ð?*Æ:!¾2Å:'Ã8%Ö/&ß8/ü:;ÿ?@ÿ8=ø16æ/1â+-Ô/-Õ0.Ü44á99é9<ê:=þ-2þ-2õ/.ô.-ä2(á/%Ó3#Ñ1!Å4#Ï>-økaâUKÆ84ÏA=Ê;=áRTæIDÞA<Ð92Ë4-Å8.Ç:0À>.¾<,·<*º?-»A,¹?*º;(º;(½<'¾=(¹?0¹?0¾D7ÆL?ÊPEÉODÊMGÌOIæabçbc÷gpÿpyÿhxö[köUgü[mù]rù]rÝhnálr܃}匆ãzwÍdaôfrünzñdtôgwì–—ôžŸ·šŠ¦‰y[rV?V:XUFnk\¥‹ŠÙ¿¾ðÖÙûáäÿôøÿøüýþÿûüÿþûÿÿýÿÿùÿÿõÿùÿøõÿô÷ýóúÿöÿÿøüýõÿýôÿÿöþúîóïãïíØøöáþþâýýáýþÞÿÿâýþîøùéö÷çRO>QN=PM<NK:LI:JG8HD8HD8EA6C?4A=4@<3?:4?:4@;5@;5@A;BC=IF7JG8PE/QF0UF/WH1SI?RH>SQ_jhvz�­Œ“¿��Ú•£àŸ ó¡¢õ¤¥ø¦§ú¨©ù©ªú©¬ù¨«ø¦ªó¤¨ñ §ë §ë¡©ê¢ªë¡ªë ©ê— á–Ÿà” Ü’žÚ‘�Í�™É–�´’Œ°—j�•h»Uc¸R`Ý@IÝ@Iõ6=õ6=êA.è?,ÕH,ÕH,ÃL*¿H&»D&»D&áTCÕH7Ü:7á?<éCEæ@BØ9=Ò37ÅF5ÅF5ÇC4ÅA2É=0Å9,Å3&Ã1$Ê0$Ë1%Ò1'Õ4*Ø4*Ù5+Ù5+Ù5+áC8Û=2×7+×7+Û9.Ü:/Ø6)Ó1$Ø8*Ð0"Ê1Ð7%Ë:%Å4Æ:!ÐD+Ã8%Æ;(Û4+Ü5,ð./ô23ý6;ÿ;@ÿKMýFHêECæA?å==â::â25Þ.1ù(-ù(-ó-,ó-,å3)ä2(×7'Ö6&Ã2!¿.ñdZê]SÎ@<ÓEAÉ:<ÝNPîQLãFAÓ<5Ì5.Ä7-Æ9/¾<,½;+¸=+¼A/¼B-¸>)·8%·8%½<'¿>)²8)²8)¸>1ÁG:ÅK@ÄJ?ÅHBÈKEä_`ä_`óclþnwÿgwö[köUgü[mù]rú^sÞioàkqÞ…ì“�ð‡„ßvsünzÿz†ðcsâUeæ�‘ÿ°±Í° ¯’‚WnR>U9URCgdUŸ…„Ù¿¾ñ×ÚøÞáÿõùÿùýýþÿûüÿþûÿÿýÿÿøÿÿôÿùÿøôþóõûñùÿõýþöúûóÿüóÿÿöÿÿóùõéõóÞüúåÿÿãüüàýþÞÿÿâÿÿñüýíüýíXO@WN?VM>UL=TK<RI:QH9QH9LC4LC4JA2I@1H?0G>/G>/H?0CC9DD:IF5KH7QI2RJ3SK4SK4UNFTMEVUgji{y}­ŠŽ¾�˜Û–Ÿâ¦¢÷¥¡ö¡¢þ¡¢þž¤ÿ ¦ÿŸ«ÿ ¬ÿŸ¬ò�ªð¢§ç£¨è­§ï¯©ñ³¦ø²¥÷—¬ÿ’§ú’¢Ý� Û¯ ¿¯ ¿Î|Š¹guÜ?HÜ?Hô5=ô5=é;=é;=ÔE?ÔE?â;Cá:BÙEEÌ88ÊC?ÊC?¾/+ÚKGìFJéCGî;Aì9?ä;>ã:=Ö>;Ô<9ÌA:È=6ÈC:ÑLCÉK=»=/³8)¶;,º8*»9+Å5,Å5,Ð.)Ò0+Ü0.ß31åB1Ý:)Ü1 Þ3"å8'è;*Ý8$Ö1Ë5Ê4Ä6Ä6Ç4$È5%Î4(Ï5)Í/&Ï1(Í7)Î8*Ê9&Ê9&Ò7#Ó8$à2'ã5*ï31î20í,1ë*/é,3ë.5ö.1ó+.ë.*ì/+å7,ä6+Õ8%Ñ4!À/Ã2æYHè[JÊ<2ÝOEÄ3.ÙHCÝQP×KJÑC?Ê<8Ç7.Æ6-È9)Ê;+Ç<)Å:'¾:%½9$¹<(º=)µ>*´=)±9+´<.¸>/¹?0½?0½?0Â@2ÃA3ïk_ôpdèc\óng÷qnèb_ðjiútsÿmzúhuðelìahòadøgjñpkñpkúxvÿ�ÿTkÿ\sÙerÿŸ¬’»›n—w]sLF\5MQ6lpU¥•ˆÈ¸«êÕÒÿîëÿ÷÷ÿööúøùüúûûÿÿüÿÿûÿÿúþÿåÿïßÿéãüçéÿíøÿôøÿôþÿñýþðÿüìÿûëÿùèÿ÷æþøâþøâýùàþúáüþûþÿýþÿýXO@XO@WN?VM>TK<SJ;RI:RI:NE6MD5LC4JA2I@1I@1I@1I@1DD:EE;JG6LI8RJ3SK4SK4SK4UNFUNFVUgji{x|¬‡‹»Œ•Ø’›Þ œñ œñœ�ù�žúœ¢ÿž¤ÿžªþ ¬ÿŸ¬ò�ªð¡¦æ ¥å¨¢ê¨¢ê«žðª�ý“¨û”¤ß‘¡Ü«œ»¥–µÃq®\jÜ?HÜ?Hô5=ô5=é;=é;=ÔE?ÔE?á:BÞ7?×CCÎ::ÊC?Ç@<Á2.áRNèBFæ@Dì9?é6<à7:ß69Ò:7Ñ96ÐE>Ç<5Â=4ÌG>ËM?ÁC5¶;,´9*¹7)º8*Å5,Æ6-Ó1,Õ3.à42ä86â?.Þ;*à5$à5$â5$â5$Ú5!×2Ï9 Î8Æ8 Ç9!Í:*Í:*Ï5)Í3'Ò4+Ó5,Ï9+Í7)Ç6#Æ5"Í2Ï4 Ý/$Þ0%é-+é-+ë*/ë*/é,3ê-4ð(+ò*-ì/+ì/+á3(à2'Ô7$Ô7$Ñ@+º)ÒE4Í@/ÕG=çYOÔC>ÔC>âVUÞRQÙKGÑC?Í=4É90Ç8(Ç8(Ç<)Å:'¿;&¿;&»>*¼?+·@,¶?+±9+³;-·=.¸>/»=.¼>/Â@2ÄB4ìh\ònbæaZðkdõolça^ðjiûutÿp}ÿn{õjqê_fêY\ñ`córmûzuøvtý{yýOfÿVm×cpÿš§‹´”eŽn]sLKa:SW<quZ¬œ�ÔÄ·ôßÜÿòïÿúúÿøøüúûüúûúþÿúþÿùýþøüýæÿðßÿéâûæçÿëõÿñöÿòþÿñÿÿòÿüìÿüìÿùèþõäùóÝøòÜùõÜüøßøú÷úüùýÿüYPAYPAXO@WN?VM>UL=TK<TK<PG8OF7NE6MD5LC4LC4LC4LC4GE9HF:MH5OJ7RJ3SK4VK5VK5VMDVMDYVglizx{¦…ˆ³ˆ‘ÌŽ—Ò˜™Ú˜™Ú—›å™�ç™ ðœ£óž§ò ©ôŸªè�¨æ›¤Ý™¢Ûœ ß›ŸÞœ›áš™ß•›×–œØ±�®Œ¿Ôx�Ên“ÙQeÉAUÜ@DÜ@Dá?=á?=â>?â>?ã<Cã<CÞ=B×6;ÑA@Ï?>ÇC>À<7Ã83åZUâ?Bâ?Bç:>ä7;Ø88Õ55Ê;3Ë<4ÒH>Å;1»7*ÅA4ÊL=ÅG8»>,µ8&¹5&»7(È6)Ê8+×4-Ú70æ95é<8Ó5)Ô6*Ù5+Ú6,Ú6*Ú6*Õ<*Ö=+¾7!¼5´7!µ8"¸:+¸:+¹5)¶2&è<8ç;7Ü>2×9-É;'Æ8$Æ8 Ç9!Ð7%Î5#Ô0$Ô0$×0*Ø1+Õ2-Ô1,Ö/'Ú3+Ù9-Ø8,Î7$Ì5"Ë:%Ï>)ÑC/À2çSEÏ;-Ü?:ßB=ëHIõRSêVTçSQãPIÜIBÖC9Ï<2È9)Å6&Á8&À7%º9&º9&¶=*·>+²?-°=+¬;-®=/±=.±=.µ=-·?/½B3ÀE6Ú\Pæh\âa[îmgôpnåa_êehñloâQ^î]jôipñfmò_eöci÷soýyuùstørsûE^ÿMfÛ^nÿŽž‡¨‰^`VoEMf<Y`?szY®£‘ÞÓÁþíåÿ÷ïÿþøÿý÷úÿøûÿùùÿûùÿû÷ÿùõÿ÷âÿèÚÿàÙ÷ÛÜúÞéûãëýå÷üèûÿìýúçþûèÿ÷æùðßòê×ïçÔóëØøðÝüùôÿü÷ÿþùZQBYPAYPAXO@WN?VM>VM>UL=SJ;RI:QH9PG8OF7OF7PG8PG8IG;JH<NI6PK8SK4TL5VK5VK5ULCVMDYVgkhyux£�„¯„�È‹”Ï’“Ô“”Õ’–à”˜â”›ë—žî˜¡ìš£îš¥ã—¢à•ž×’›Ô”˜×’–Õ“’Ø��Õ€†Â‚ˆÄ ~±Ÿ}°Äh�»_„ÑI]È@TÝAEÝAEâ@>â@>ã?@ã?@ä=Dä=Dà?DØ7<Ï?>ÓCBÅA<º61Ç<7æ[Vâ?Bâ?Bé<@æ9=Ù99×77Ì=5Î?7ÎD:Â8.¸4'¿;.ÄF7ÃE6»>,¶9'»7(¾:+Ë9,Í;.Ú70Ý:3ç:6ê=9Ò4(Ó5)Ø4*Ú6,Û7+Û7+Õ<*Õ<*Â;%Â;%»>(»>(º<-»=.Á=1Ã?3îB>ì@<ßA5Ù;/Ê<(Ç9%Ç9!É;#Ó:(Ï6$Ò."Ò."×0*Ø1+Ó0+Ñ.)×0(Ø1)Õ5)Ô4(Ì5"Ì5"Ë:%Î=(Ê<(Ì>*ñ]OÜH:Ð3.Î1,ã@AùVWÿvtÿrpÿleó`YäQGÔA7Ä5%½.Â9'À7%º9&º9&µ<)¶=*±>,¯<*¬;-­<.°<-¯;,³;+¶>.¼A2ÀE6ËMAÞ`Tâa[ñpj÷sqä`^â]`ã^aØGTèWdóhoóho÷djüioùuqûwsørsðjkö@YüF_Ú]mÿ†–�¢ƒYz[RkAPi?_fEv}\²§•æÛÉÿóëÿ÷ïÿûõÿûõúÿøüÿúùÿû÷ÿùòüôîøðÖûÜÍòÓÌêÎÍëÏÚìÔÝï×ëðÜñöâöóàùöãüóâôëÚêâÏæÞËëãÐñéÖùöñûøóýúõ\P@[O?[O?ZN>YM=YM=XL<XL<VJ:VJ:UI9TH8TH8TH8UI9UI9KG;LH<QJ8RK9WL6WL6WL6WL6TLAUMBYV_ifost“~žƒ‡¶‹�¾�‘¾Ž’¿�’Ë’”Í’•Ú”—Ü”™Ý•šÞ•žÙ‘šÕ‹•È‡‘ăŒÃˆ¿{ƒÁx€¾‡sŽˆt�½_ƒ½_ƒñBië<cõ4Iô3HåCAåCAÕK@ÕK@àDEàDEõ9Jõ9JÞEGÖ=?ÉB>ÏHD»A6¯5*ÇB;ÜWPÝCAÝCAâ?@à=>Õ@:Õ@:ÉG:ÊH;Æ</À6)»6%½8'¿>+¿>+½<'¼;&Ã:(Å<*Ñ;,Ò<-Þ7.ß8/æ81è:3é77æ44æ0/ç10ê20é1/Ý2+Ø-&Ë1)Î4,È91Å6.Â00Ä22Ï9;×ACõ47ó25â7/Ý2*È7$Æ5"À:!Â<#È=&Å:#È5#Ê7%Ñ9+Ó;-Ñ;-Ï9+ÈA.Ä=*¿:'¾9&À<'Â>)Å='Ä<&Ï>-Ë:)Ú<1àB7Ù-+å97ã/2æ25Ó61Ø;6ÞD<åKCéQFëSHåVFäUEÈ@0Å=-»>,¹<*³@.³@.¯@/®?.¨=-©>.«=,ª<+«<+®?.µA2¹E6ÅLAÛbWàc]ípjóqsâ`bå_få_föjuûozôkré`gñ^dþkqÿyyÿzzýlsñ`gü;Xÿ>[àYjÿz‹‹ �i~_XvDZxFhwN|‹b´°•æâÇÿúçÿüéýÿïýÿï÷ÿó÷ÿóñÿðêýéáòßÛìÙÁìÁºåº¹ß¶»á¸Åà½ÉäÁØçÊÞíÐìîØðòÜøñßòëÙìÞÑçÙÌìÛÑñàÖñçÞòèßóéà[O?[O?ZN>ZN>YM=YM=YM=XL<XL<WK;VJ:VJ:VJ:VJ:WK;WK;MI=NJ>RK9TM;XM7XM7XM7XM7TLAUMBVS\c`ijkŠst“y}¬‚†µ…‰¶†Š·ˆŠÃ‹�ÆŠ�ÒŒ�Ô‹�ÔŒ‘ÕŒ•Ð‡�Ë€Š½yƒ¶s|³mv­go­ck©€l‡k†´Vz¶X|ì=dç8_ö5Jø7LæDBæDBÖLAÖLAáEFáEFö:Kö:KÞEGÙ@BÉB>ÏHD·=2®4)ÌG@ÑLEÚ@>Ø><Ý:;Ý:;ÖA;×B<ËI<ÊH;À6)Á7*¾9(½8'º9&º9&¼;&¾=(Ä;)Æ=+Ò<-Ó=.Ý6-Ý6-ã5.å70å33ã11å/.è21ï75ð86å:3à5.Í3+Ð6.Ê;3Ç80Ä22Æ44Ñ;=ÙCEð/2ð/2à5-Ý2*É8%È7$Â<#Ä>%Ã8!Ä9"Ë8&Ï<*Ö>0ÙA3ÙC5ØB4ÏH5ÉB/À;(¾9&À<'Â>)Ä<&Â:$Ì;*Ê9(Í/$ãE:Ü0.ì@>ã/2Ü(+Ð3.Ò50Ò80Ó91Ñ9.Í5*Â3#¿0 ÌD4È@0º=+¶9'­:(­:(ª;*©:)¨=-©>.«=,«=,«<+¬=,²>/µA2ËRGßf[àc]èkeïmoåceïipônuûozþr}ömtìcj÷djÿrxÿzzýssúipí\cþ=Zü;XÝVgùrƒ�¤…x�nlŠXm‹Yz‰`Š™p¸´™àÜÁùòßþ÷äÿÿñÿÿñõÿññÿíçúæÜïÛÑâÏÉÚÇ­Ø­¨Ó¨ªÐ§­Óª¸Ó°»Ö³ÊÙ¼ÑàÃâäÎçéÓòëÙñêØïáÔéÛÎêÙÏëÚÐæÜÓäÚÑâØÏZN>ZN>YM=YM=YM=XL<XL<XL<XL<WK;WK;VJ:VJ:WK;XL<XL<QI<RJ=TK:VM<XM7YN8YM7YM7WMAVL@VPR_Y[c`sjgzqp�zy˜| ~�¢�‚²„…µ‡…Ɖ‡È‰ˆÌ‰ˆÌ‡‡Ãƒƒ¿|~±vx«qq¥ll fežcb›’]Y�[WÃGQÊNXý9Mù5Iÿ1>ÿ5Bï@;ï@;ÜIAÜIAãDHãDHõ:Mõ:MÕFBÕFBÃD;ÇH?°</­9,ÎPDÄF:Ñ<8Ì73Í2.Ð51Í@7ÑD;ÂI8¾E4À9&Á:'À;(¾9&¼:$¼:$¾:%¿;&Æ8$É;'Õ:(Ö;)ß5(ß5(å3)æ4*ï'*ñ),û*/ÿ.3ÿ/6ÿ07ÿ17ÿ17ý4:ý4:ù4=ù4=ÿ2Cÿ3Dÿ/Gÿ/G÷(0ø)1ê20è0.Õ7,Ô6+Î;+Ï<,Ë3%Ï7)Ü8.à<2ä<3å=4ç<4è=5ÒM<ÍH7ÉA1Ã;+Â9'Â9'Æ9(Æ9(É1$Ñ9,â=7èC=ß//Ý--ë46ë46Ö5+×6,Ö8-Ø:/Ô</Ò:-Ë<.É:,ÒM>ÌG8½?1¶8*¬6(¬6(©8*©8*«;-¬<.­>-­>-«<+ª;*¬=,­>-ÀKAØcYÝc^ælgîpsçilôqyúwígpöpy÷qxñkrÿktÿvÿu|ùhoüblðV`ÿ=]ÿ6VáPaøgxœ§‡�›{z›dx™b‚•g� r­±ŽÈÌ©ÝàÃçêÍéøÙäóÔÕðÏÌçÆÂ༹׳³Íª­Ç¤”ÅŽ’ÃŒ˜Ä�žÊ•§Ë›©Í�µÎ¤»ÔªÎÙ¹Ôß¿åãÎéçÒîàÕéÛÐçÒÍäÏÊÝÏÄÙËÀÕǼYM=YM=YM=XL<XL<XL<XL<XL<XL<WK;WK;WK;WK;WK;XL<YM=RJ=SK>UL;VM<YN8YN8ZN8YM7YOCWMATNPZTV[Xk`]pgf…poŽtw˜vyšz{«~¯‚€Á…ƒÄ…„È…„È~~º{{·vx«rt§pp¤mm¡ji¢gfŸ‡RN†QM¼@JÇKUþ:Nû7Kÿ3@ÿ5Bï@;ï@;ÜIAÜIAãDHãDHõ:Mõ:MÑB>ÔEAÁB9ÅF=®:-¯;.ÒTH¿A5Ë62Ä/+Ä)%È-)É<3ÎA8¾E4¹@/Â;(Ã<)À;(¿:'¾<&À>(À<'½9$Ä6"Æ8$Ó8&Õ:(ß5(ß5(å3)æ4*ù14þ69ÿ7<ÿ5:ÿ,3þ'.û(.ý*0÷.4ó*0í(1ð+4û.?û.?ø":ð2ò#+ô%-ç/-æ.,Ó5*Ñ3(Ê7'Ë8(Ï7)Õ=/ã?5ä@6å=4â:1ã80ã80ÓN=ÑL;ÏG7Ç?/À7%¾5#Ä7&É<+Í5(Ð8+çB<Ý82à00Ïî79é24Ó2(Ó2(Ñ3(Ñ3(Í5(Í5(Æ7)Å6(ßZK×RCÇI;½?1³=/²<.¯>0°?1«;-¬<.®?.­>-«<+ª;*ª;*ª;*°;1ÏZPÛa\êpkôvyëmpôqyõrzÿy‚ÿˆÿy€ðjqúdmÿpyÿu|þmtø^hîT^ÿ>^þ5UÞM^óbs «‹›¦† iz›d‚•gŠ�o¢¦ƒ´¸•ÈË®ÕØ»ÊÙºÃÒ³°Ëª¦Á �»—˜¶’—±Ž”®‹…¶…¶�¹„”À‹�Á‘žÂ’ªÃ™°ÉŸÁ̬ÈÓ³ÛÙÄâàËëÝÒå×ÌáÌÇÛÆÁÑø˽²Å·¬ZM<ZM<ZM<ZM<ZM<ZM<ZM<ZM<WJ9XK:XK:XK:YL;YL;YL;ZM<XL>YM?ZM<ZM<[O9[O9[O9\P:YM=YM=VMHXOJXQX\U\`Zfb\hhe‚ol‰roœurŸzu·‚}¿‰‚ÈŠƒÉ‰}»Š~¼Œx­…q¦�o£�r¦ŽiŸ�\’¤C3£B2»?3ÃG;Þ?9Þ?9ô=9ýFBö=;õ<:ï<?ð=@êAHëBIæBMåALÈC<Ä?8·A5´>2­>-­>-ÁC7ÍOCÊ5/Ñ<6Î1,Î1,ÑD:Ê=3°7&¸?.¼5!À9%Ã<(Á:&À9%À9%Å:%Æ;&Î9%Ï:&×5&Ô2#Ú-Þ1#ì6+ô>3ó8/ñ6-÷0)ô-&û'$ú&#ü$%ü$%ü%*ø!&ü ,ÿ'3ÿ%;ÿ!7ÿ4ÿ1ú*7ñ!.â#*ç(/Û-.Ö()Õ,)Ý41è21ê43õ34õ34÷/1ô,.ò*,ð(*Ø;2Ô7.Ô70îQJ¼"Ä*"áD=Ê-&Ì-'Ü=7çD=ß<5Ö/)Ö/)á6/æ;4Ø@2Ô<.Ì6(É3%É7*Í;.È</Å9,êbVáYMÇC7º6*¼9/º7-²2'´4)³/#³/#±1$²2%°6'²8)±<+³>-°:.¹C7Ö\Wêpkðswôw{úx‚óq{þ|„úx€÷tzøu{ÿr}ÿr}ÿmxþhsý_mëM[ÿ9[ÿ8ZÛDUîWh›¢€—ž|} fx›a~”c‚˜g”Ÿuœ§}�©ƒ›§��¦~‰¢z}¡sxœnvœkxžm¡o‚¤r¯qƒ³u‰¹{�½’¾ƒ”À…œ¿‰�ÀŠ¦»’ª¿–¾Â©Ë϶ßÓÇßÓÇÞÆÄÕ½»É¼¶¿²¬¸«¥YL;YL;YL;YL;YL;YL;YL;YL;WJ9XK:XK:XK:YL;YL;YL;ZM<XL>YM?ZM<ZM<[O9[O9[O9\P:[O?[O?WNIXOJWPWZSZ\Vb^Xd_\yfc€jg”nk˜to±z¼†Ň€Æ��¿Š~¼Šv«„p¥‘s§–x¬’m£‚]“ ?/œ;+±5)»?3Ú;5Ý>8ñ:6ö?;õ<:ô;9ï<?ð=@êAHêAHã?Já=HÂ=6¿:3²<0°:.©:)«<+¿A5ÊL@ÚE?ØC=Ô72Ò50ÐC9ÎA7µ<+®5$¾7#Á:&Ä=)Ã<(Ã<(Ã<(Æ;&Æ;&Î9%Ì7#Ô2#Ö4%â5'ç:,ï9.ï9.é.%é.%ñ*#ñ*#û'$ý)&ÿ)*ÿ+,þ',ú#(ý!-ÿ&2ÿ#9ÿ3ÿ0ÿ.ð -û+8ò3:è)0×)*Ü./Û2/×.+ð:9ï98ö45ô23÷/1÷/1ù13û35Ò5,Ù<3Õ81âE>àF>Ë1)Ï2+À#Ó4.Þ?9çD=à=6Ù2,Ø1+à5.â70Ò:,Ï7)Ë5'Ê4&Ê8+Í;.È</Æ:-ëcWàXLÅA5¸4(¾;1¿<2¸8-»;0º6*º6*·7*¶6)±7(±7(­8'­8'²<0¹C7ÓYTçmhíptõx|ý{…÷uþ|„úx€ùv|ùv|ÿr}ÿp{ÿjuûepü^lëM[ÿ9[ÿ9[ßHYíVg“šx�”rt—]q”Zw�\{‘`‰”j�˜nŠ–p†’ln‡_m†^fŠ\g‹]j�_p–exšh|žl~®p‚²t‡·y‹»}�¼�‘½‚˜»…š½‡�²‰Ÿ´‹°´›»¿¦ÐĸÔȼ׿½Ñ¹·Ä·±º­§²¥ŸYL;YL;YL;YL;YL;YL;YL;YL;WJ9XK:XK:XK:YL;YL;YL;ZM<[K<\L=\L<\L<]N;]N;[O9\P:_RA^Q@\PB[OAYNHZOI\QO]RP^RjeYqf^‡kcŒpk¥}x²†}¼‡~½›„º—€¶¢sŸŸpœºo˜ÀužÀc�®Q}Ë62Å0,¹5(¿;.ÄE4ÈI8ÙB7ÛD9ô<>ô<>û7Aü8BìAGé>DÐDCÌ@?»=1¸:.¬;+«:*¨9(©:)¾>1ÉI<ÜA<Õ:5Ú64Ý97ÙE;áMCÊK:³4#À9&À9&Â;(Å>+É>+È=*Ë:)Ê9(Ò6'Î2#Ö.!Ý5(î<2ò@6ñ91ê2*×8%×8%Þ3"Þ3"è-$é.%ì,'í-(î.-ê*)ï'1ò*4ú%9õ 4ú3ù2ë(6ñ.<ï7Að8Bí<Fí<Fê5>â-6ó/;ó/;ö-7ö-7ô-2õ.3ñ10ñ10Ö1-á<8Ú53à;9ÿkjåA@Ñ/-Ð.,Ù:6Þ?;ßB;Ù<5Ò8,Ð6*Ð7)Ï6(À9%¿8$¿7'Á9)Á<-Â=.Ä;1Ä;1ê`VÞTJÊ:2À0(Í81Ò=6Ò91Ô;3Ý:3Ý:3Ø;2Ö90Ê;-Æ7)»7(¹5&·;/¼@4ÏTMãhaëoqöz|ÿ€‡ú{‚û|ƒùz�ûx€ûx€ÿrÿn{þfsø`müYjíJ[ÿ7Wÿ9YâKZèQ`‚�ey„\e†Od…NmƒRq‡V{ˆ\{ˆ\t…[p�W^{Ma~P^†Re�Yl•]sœdz g}£j�­n„°q‡µtŠ¸w‰ºxŠ»yŽº{�»|’°|�®zœ§…¤¯�º³¡Â»©Ë¸²É¶°²­±¦¡¥œ—›YL;YL;YL;YL;YL;YL;YL;YL;WJ9XK:XK:XK:YL;YL;YL;ZM<[K<\L=\L<\L<]N;]N;[O9\P:^Q@]P?[OAZN@XMGYNHZOMZOM\Ph`TlaY‚f^‡mh¢zu¯ƒz¹ƒz¹šƒ¹�†¼¯€¬¬}©Äy¢Äy¢Ãf’±T€ÜGCÙD@ËG:ÊF9ÈI8ÈI8ÙB7ÚC8ô<>õ=?ý9Cþ:DíBHé>DÎBAÉ=<½?3¸:.¬;+«:*©:)ª;*½=0ÆF9Ö;6Ð50ß;9åA?ÛG=èTJØYH»<+Á:'¾7$¾7$Ã<)É>+È=*Ë:)Ë:)Ï3$Î2#Ù1$á9,ñ?5ñ?5í5-å-%Ø9&Ø9&ß4#ß4#è-$è-$ë+&ë+&ê*)è('í%/ò*4ú%9÷"6ÿ9ÿ <ú7Eè%3ç/9ÿOYÿakÿPZò=Fï:Cì(4î*6õ,6ø/9ö/4ò+0è('ã#"Õ0,Ú51Õ0.æA?ÿqpøTSÞ<:èFDÛ<8Ú;7Ö92Ô70Ñ7+Ð6*Í4&É0"¼5!¼5!¾6&Á9)À;,¿:+Â9/Ä;1ãYOØNDÈ80Â2*Ò=6ØC<Ö=5×>6â?8ã@9ßB9Ý@7ÑB4Í>0Á=.¾:+¶:.¸<0ÊOHßd]ëoqø|~ÿƒŠü}„÷x÷xûx€ýz‚ÿrÿmzýer÷_l÷TeëHYÿ4Tÿ7WãL[áJYozRfqIVw@XyBcyHiNs€Tu‚Vq‚Xo€Vj‡Yn‹]m•atœh{¤l¨pƒ©p„ªq‚®o„°q‡µtˆ¶u†·u†·u‹·xŒ¸y“±}Ž¬x• ~™¤‚¯¨–»´¢Èµ¯É¶°°«¯¢�¡—’–ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8YJ7ZK8ZK8ZK8[L9[L9[L9\M:]J;^K<]L:]L:]N;]N;]N9^O:\O<[N;[N;[N;^N>_O?_O@`PAbNZeQ]aTnfYshe�ur�~x¤}w£š|ž¦ˆªÎ†œÌ„šèq…âkéTjÚE[ÿ;Nÿ>QïJPèCIÌG@ÉD=ÏD=ÒG@êBBëCCô?FöAHæHGáCBÈE;Â?5ºB4µ=/«<+ª;*©:'©:'¾8-Æ@5æB@à<:î@?í?>Ü:5éGBåYLÖJ=ÆA0½8'º2"¿7'È;*È;*Í7)Î8*Ð0$Ô4(â5.è;4ð93í60ë.*ç*&á/-â0.ë/0í12ô/6ö18ö3;÷4<è-8è-8é.?ì1Bò-Ið+Gù+Oÿ2Vë=Gâ4>ìBMÿalÿlyÿYfôCSñ@Pë5Dê4Cæ7>æ7>Ø=8Ô94Â9)½4$Ç4*Â/%Ì91î[Só^ZîYUØC?ÞIEÙD@Ñ<8Ë7-Ë7-Ì:+Ì:+È7$Å4!®<$­;#°;'´?+µ;,²8)»8.Â?5ÜMEÖG?Ñ4/Î1,ä;4è?8ç60ç60ó-.ô./í4/í4/á90ß7.Î:,Ì8*¼6*¼6*ÈIBÞ_Xênlù}{ÿ†ˆû�óv|õx~úx€üz‚ÿq€ÿl{üds÷_nóRdëJ\ÿ4Tÿ8XãQ\ÙGR^nAWg:Vr?ZvCfzIkNt�Vw„YwŠ]yŒ_t–dw™gv¡i{¦n«n€¬o„ªm„ªm†¬oˆ®q‡´sˆµt„¶o…·p†¸sˆºuŽ¶zˆ°tŽ£x’§|¬¬”»»£ÎÀµÒŶºÆ§«·˜œ¨ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8YJ7ZK8ZK8ZK8[L9[L9[L9\M:]J;^K<]L:]L:]N;]N;]N9^O:ZM:[N;[N;\O<_O?`P@`PA`PAaMYbNZ]PjaTndaŒqn™ysŸxrž‘s•£…§ÐˆžÎ†œçp„Þg{åPfØCYü3Fÿ;NðKQèCIÊE>ÇB;ÏD=ÒG@éAAêBBó>Eô?FäFEßA@ÆC9Á>4¸@2²:,¨9(¨9(¨9&©:'½7,Ã=2ëGEåA?ê<;ä65Ö4/á?:é]Pôh[×RAÆA0½5%¿7'Ç:)Æ9(Ì6(Ï9+Ñ1%×7+æ92è;4ê3-æ/)é,(ë.*ê86ë97ó78ó78ø3:ø3:ö3;ö3;î3>ï4?ð5Fñ6Gõ0Ló.Jÿ1Uÿ:^ã5?ûMWÿ`kÿ^iÿVcûQ^õDTè7Gõ?Ní7Fß07Ú+2Ì1,Í2-Â9)Á8(É6,Ã0&ÜIAüiaßJFäOKÏ:6Ð;7àKGÕ@<Ë7-Ë7-Í;,Í;,Ë:'É8%°>&­;#¯:&²=)²8)®4%»8.ÈE;ÕF>ÔE=Ó61Ò50å<5æ=6å4.å4.ñ+,ò,-ë2-ì3.â:1á90Ò>0Ñ=/½7+º4(ÂC<×XQåigø|zÿ†ˆü€‚ôw}öyûy�üz‚ÿo~ÿixúbqö^mñPbîM_ÿ6Vÿ9YæT_ÕCNUe8Td7YuB`|In‚Qs‡Vz‡\~‹`’eƒ–i{�k}Ÿmz¥m|§o~ªm«nƒ©lƒ©l…«n‡­p…²q…²q‚´m‚´m„¶q†¸s‹³w†®rŽ£x•ª³³›ÄĬÙËÀÞОÂέ±½œ ¬ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8YJ7ZK8ZK8ZK8[L9[L9[L9\M:]J9^K:^K:^K:^M;^M;]N9^O:^O:^O:_P9`Q:aQ:aQ:aQ:`P9fLMgMN]Q_aUcdd€rrŽzw’xu�•u„¥…”Ø�‡Ø�‡ýknödgÿKQö=Cÿ#Gÿ,Pÿ;Tü6OÛ>GÚ=FÓCCÒBBÜCEÜCEÞCAßDB×G?ÓC;ÄB5À>1´>0®8*¦9%¨;'¬;)®=+Ä:/Ê@5ä::â88æ/1ä-/Ý52ß74ãOCÿk_îiZ×RCÆ</Ã9,É:,Ç8*Í5*Ò:/Ö3*Ú7.ç83æ72é0-ç.+í0.ñ42ÿ,Gÿ+Fÿ(Eÿ&Cÿ"@ÿ<ÿ<ý:÷&Eø'Fõ-Qö.Ró*Vó*Vÿ1cÿ=oçJSþajÿitôXcëJZíL\éDU×2CäBOÙ7D¿10·)(£6"§:&™F&šG'¸?,µ<)Ù[MÜ^PÊG=ÈE;¾71Ã<6àUNÑF?È:.Æ8,È;)Ç:(Ç9#É;%§A(£=$¥<'§>)©8(¦5%¼<3ÎNEÖE@×FAÜ73Ø3/í62ì51î-&ò1*ÿ(,ÿ(,û--û--ì5-ì5-Ý=1Ý=1Ç;.¿3&½=2ÏODßb^öyuÿ‡‡ÿƒƒø{�ù|‚ý{ƒûy�ÿl~ýdvö]oòYkíL^íL^ÿ5Tÿ7VàS[Ë>FJ`/Oe4ZuBdLv‡Z}Ža„�e†‘g†™lŠ�p„¦t„¦t}¨p|§o}©j}©j‚©jƒªk…©l†ªm„®n„®n±j€²k€¶kƒ¹n‡´sƒ°o�§x•¯€´·œÄǬØ̾ÜеÀÔ¢­Á�›¯YJ7YJ7YJ7YJ7YJ7YJ7YJ7YJ7YJ7ZK8ZK8ZK8[L9[L9[L9\M:]J9^K:^K:^K:^M;^M;]N9^O:`Q<`Q<aR;aR;aQ:`P9_O8_O8gMNhNO^R`cWeggƒvv’~{–}z•™yˆ¥…”Õ~„Ö…ÿmpýknÿRXüCIÿ'Kÿ-Qÿ;Tý7PàCLáDMÖFFÐ@@ÚACÙ@BÛ@>ÜA?ÔD<ÑA9ÃA4¿=0´>0®8*§:&ª=)±@.´C1Ê@5ÐF;Ý33à66ã,.æ/1ã;8Ö.+Ê6*éUIÿzkã^OÍC6Ç=0Ê;-Ç8*Í5*Ó;0Ú7.Û8/ä50ã4/é0-ê1.ò53õ86ÿ!<ÿ!<ÿ =ÿ =ÿ;ÿ9ý:ü9ì:î<ì$Hï'Kí$Pð'Sÿ0bÿ?qÿfoý`ióWbíQ\ðO_íL\ã>OÕ0AÝ;HÓ1>½/.¸*)¤7#§:&–C#•B"¯6#²9&ÓUG¹;-Â?5¹6,¹2,ÊC=ãXQÓHAÇ9-Ä6*Ä7%Â5#Ã5Ç9#§A(¡;"£:%¥<'§6&¦5%À@7ÕULÙHCÚIDß:6Ø3/ê3/é2.î-&õ4-ÿ(,ÿ'+ø**÷))ç0(æ/'×7+×7+Ì@3¿3&¹9.ÈH=Ù\Xóvrÿˆˆÿ††ü…ý€†ÿ}…ûy�ÿj|úasóZlïVhçFXêI[ÿ2Qÿ3RÜOWÄ7?CY(Lb1To<a|Ix‰\ƒ”g‹–lŒ—m‹žq�¢uŠ¬zˆªx~©q{¦nz¦gz¦g¦g¦g‚¦i„¨k�«k‚¬l|®g~°iµj�·l‰¶u…²q�©z—±‚³¶›ÁÄ©ÒƸÔȺ±¼Ðž©½‹–ª[J6[J6[J6[J6[J6[J6[J6[J6\K7\K7\K7\K7\K7\K7\K7\K7[I;\J<ZM=ZM=XM;YN<ZR=\T?^S?^S?^Q@_RAcQCbPBbOA`M?kLIjKHaQ\jZeglŒv{›{†³€‹¸Žƒ­–‹µ¾‡ž¾‡žåywäxvùlZôgUïMZåCPã=IçAMîBNê>Jé;Dé;Dâ<@à:>Ø=;×<:Í@6Ê=3¿=/»9+ :$ªD.¬;)©8&Ã?3ÊF:Ò;4Î70Ú70Ú70Û7+Ú6*Û:&Ü;'Ù;"Õ7ß82öOIòZOÛC8Å8'Ë>-ÓB1Î=,Ö1+íHBä+0î5:õ3>î,7ÿ?Pì'8ü&(ÿ*,ÿ*,ö "ù%$ý)(û)(ðí!$ø,/é!+î&0ï/>ä$3ÿYmé,@þajü_hÿZaúU\ôPQìHIØA:Ï81ÁD2ÁD2­D.¢9#”5–7!š;)›<*¨71ª93ÄDA¼<9À00Å55Ç31ÖB@ÛLFÙJDÑD:Æ9/Å2(Ç4*Ì4)Ë3(¸<0³7+²,!­'¸* Å7-ÔA7ÙF<æODÜE:Û7.Ö2)é-+î20ÿ.1ÿ.1ò12ñ01ó/0ó/0ð0+ð0+ã7+â6*Ñ?0Â0!º@1·=.Ò_Xæslÿ�Ž÷…„ùˆŒùˆŒÿs…ÿhzÿ[wÿYuÿRnûJfýAbó7Xï<Zí:XÃgj•9<BP/N\;Uw:e‡Jt˜[zžax d|¤h~§k¨l~§k~§k|¦g{¥f|£d|£d~¢b£cz¥m|§o~ªk«l~®d~®d€±b‚³d†´k‡µl‰±uŠ²v•´Š¡À–«Ä¤¬Å¥¤²³‹™šw…†\K7\K7\K7\K7\K7\K7\K7\K7]L8]L8]L8]L8]L8]L8]L8]L8\J<]K=\O?\O?ZO=ZO=ZR=[S>^S?]R>]P?\O>`N@aOAcPBcPBkLIkLIcS^m]hkp�y~ž}ˆµ‚�º •¿š�¹³|“±z‘áusé}{ÿr`öiWüZgðN[ëEQîHTöJVòFRí?Hê<Eá;?ß9=×<:Ö;9Í@6Ê=3¿=/¼:,¤>(¦@*©8&¬;)Ã?3Â>2Ì5.Ô=6Ö3,Ö3,Ø4(Ø4(Ø7#Ú9%Ú<#Ù;"Ü5/èA;æNCáI>Í@/Å8'Ê9(Ï>-Þ93ðKEç.3ï6;÷5@ð.9ÿ>Oë&7ù#%ú$&ý')ÿ*,þ*)ú&%ù'&û)(ï#&ð$'ì$.ÿ9CÚ)ò2Aü?SÞ!5ÿluÿclÿZaüW^óOPæBCÕ>7Ô=6¸;)ÀC1´K5¨?)’3‹,Š+‹,�,&­<6ÒROÉIFÄ44Å55Ð<:êVTÖGAÕF@ÎA7Å8.Å2(È5+Ï7,Ï7,¼@4¹=1¼6+¼6+Ä6,É;1ÔA7ÚG=áJ?ÙB7Ú6-×3*ê.,î20þ-0þ-0ò12ñ01ó/0ó/0ð0+ð0+ã7+â6*Ò@1Â0!¸>/¶<-Ñ^Wévoÿ��÷…„ö…‰ö…‰ÿqƒÿgyÿZvÿXtÿRnûJfû?`õ9Zò?]ó@^´X[�47AO.N\;Vx;e‡Jr–Yw›^vžby¡e|¥i}¦j|¥i{¤hz¤ey£d{¢c{¢c}¡a}¡ax£kz¥m|¨i}©j|¬b|¬b~¯`€±b‚°g„²i‡¯s…­qŠ©Ž­ƒ�¨ˆ‹¤„w…†aopO]^\K7\K7\K7\K7\K7\K7\K7\K7]L8]L8]L8]L8]L8]L8]L8]L8]L8_N:`Q>`Q>[P>ZO=YPAZQBZRGZRGZOKZOK]OO_QQcSVeUXdVVfXXa_ukimv¯y‚»€‰Ø„�Ü™ŽÚ—ŒØ®€¯®€¯ÓŒÚ†“ëräxkÿbgôUZîGOíFN÷GQöFPô@Ið<Eâ8;á7:Õ<7Õ<7ÇA5Å?3ºA.·>+ª?-¨=+ª9)µD4ÅI=¹=1º1'Ä;1Ë7-Ì8.Ó7*Ô8+Ø7%Ø7%Û9$Ý;&Ý4/Ü3.ÞB6êNBÛH8Ç4$É2!×@/å=4òJAé/0ð67ù5?÷3=ÿDP÷3?ì*+ç%&ç%%ô22ñ//ê((ç'&ò21æ*+ô89á,1Þ).å5?ÿS]ê>JÿZf÷acîXZîOSíNRéEFÝ9:Ó84×<8¿6,Ä;1¼B5¶</¬6,¨2(¥*%Ÿ$š%¦1*ÈD@À<8½++Å33×A@ô^]Ð@8ÒB:ÐB8Ê<2Í9/Ò>4Ü?6Ý@7á=<Þ:9ä:;è>?è<<ä88æ>;ïGDâA9Ü;3Ù5,×3*æ0,ê40õ20ô1/ó12ò01õ/0õ/0ñ0+ñ0+ä6+ã5*ÓA2Ä2#´:+µ;,Í\Tëzrÿ’‘ù‡†ú�ˆø†ÿoúgwÿYtÿVqÿLlÿEeû<^ø9[æJbëOg�JE{61CQ0R`?[z?h‡Lp”Wt˜[s›_wŸcy¢fz£gy£dx¢cw¡bw¡bxŸ`y a{Ÿa{Ÿay¢fz£gz§d{¨ey©]zª^|­\~¯^„²g‡µj‹±t†¬o…¥v€ qx’onˆeSb]@OJ2A<]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8]L8\K7^M9_P=`Q>[P>ZO=XO@YPAXPE[SH_TPbWSgYYi[[l\_n^aoaarddnl‚wu‹u~·ˆÁƒŒÛ‡�ß‘†Ò—ŒØµ‡¶¯�°Èt�Ìx…é}pî‚uÿlqÿbgúS[óLTöFPóCMò>Gï;Dá7:à69Ô;6Õ<7ÈB6Æ@4»B/¸?,§<*ª?-¯>.¹H8ÍQEÅI=½4*¹0&Ë7-Í9/Ö:-Ù=0Û:(Ø7%Ù7"Ü:%Ý4/Ú1,Ü@4êNBàM=Ï<,Ï8'ÛD3Þ6-êB9é/0ì23ö2<û7AÿNZÿJVÿ>?ô23å##ï--í++ò00é)(í-,ô89Õæ16à+0ç7AóCMÿVböJVÔ>@ØBDàAEÝ>BÞ:;Ü89Õ:6Ó84Ä;1Á8.¶</½C6ÇQGÏYOÌQLÀE@®92°;4ÏKGÔPLãQQìZZèRQïYXÆ6.Í=5ÐB8Î@6Ò>4ÖB8ßB9ßB9ä@?à<;æ<=ëABê>>ä88ç?<õMJÛ:2Ø7/Ø4+Ø4+ç1-ê40ô1/ó0.ó12ò01õ/0õ/0ñ0+ñ0+ä6+ã5*ÔB3Æ4%°6'³9*ÇVNî}uÿ’‘úˆ‡ú�ˆö}„ÿl|÷dtÿUpÿQlÿFfÿBbø9[ú;]èLdìPh|72w2-HV5XfE_~CiˆMp”Ws—Zq™]vžbx¡ex¡ev av auŸ`uŸ`wž_xŸ`{Ÿa{Ÿax¡ez£gz§dz§dx¨\y©]z«Z|­\„²g‡µjŒ²uˆ®q…¥v}�ns�jg�^Q`[CRM7FA]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]N9]L0^M1_O6_O6ZN>YM=VOGWPHUPT\W[f_omfvog~og~pf�qg‚ho‚ls†nw¢t}¨w€Ë~‡Ò„‰å‡Œè•�â•�â¡…¸“wªŸmv£qz¸yg½~lùniúojÿhgùa`ûQTòHKð;@ê5:æ68å57Ø;6Ø;6ÈC4ÇB3ºC/·@,­7+µ?3¯;.°</ÄRGÐ^SÄNB³=1º:-º:-É:,ÐA3Ý=/×7)Ý3&à6)à1,ã4/á?4åC8áE6ÞB3ß?/à@0ß.&è7/ï/.ï/.ð-1ù6:ÿPXÿZbÿ]_þQSÝ02Þ13Ú..òFFã77Ù--×/.ç?>×45Ö34ÿehÛ?BÆ-0È/2À('Ó;:ã@CÚ7:Ú14á8;Ü89Ñ-.Ë32Ò:9ÔFEàRQï^cøglùamñYeÁB;ÅF?ç\YõjgÿonÿkjäJH×=;Ç4,Ð=5ÕE:ÓC8Õ>3Ô=2Ú91×6.û/;ö*6÷)6ø*7ó*4ò)3ó9:ÿHIÞ3,Ý2+Ø4*Ù5+â4+ä6-î3.î3.ö/2õ.1ø-0ø-0ô.+ô.+æ5+å4*Ô@2Ê6(¯4%³8)¿NFí|tÿ‘�û‰ˆÿ|‹ÿu„ùk{òdtÿSmøLfÿ:aÿ8_ö7Yü=_ÌVbÈR^]9)a=-Q\>`kMbIl‰Sq’Yt•\s™^xžcy¡cx btž^tž^u�^u�^wž_xŸ`{Ÿa| bx¢bz¤dz¨`z¨`y¨Zy¨Z{ªY|«Z«^ƒ¯b‡®m„«j‡¥sƒ¡oz•lp‹bevfZk[QbR^O:^O:^O:^O:^O:^O:^O:^O:]N9]N9]N9]N9]N9]N9]N9]N9^M1_N2_O6^N5ZN>ZN>YRJ[TL^Y]gbftm}{t„}uŒ{sŠzp‹yoŠov‰sz�t}¨x�¬x�Ì|…Ђ‡ã†‹ç•�⒊ߟƒ¶��´²€‰­{„¦gU“TBÏD?ØMHêRQîVUøNQóILñ<Aè38è8:ç79Ú=8Ú=8ÊE6ÈC4»D0¸A-°:.·A5ª6)¢.!¼J?Øf[ÙcWÊTH»;.µ5(Â3%Ë<.Û;-×7)Þ4'á7*Þ/*å61ß=2Ü:/Û?0áE6ãC3Ú:*á0(æ5-ñ10î.-ì)-ô15ÿDLÿRZÿ^`ÿZ\ä79ä79×++õIIä88Ú..Ü43Ï'&Ê'(ùVWÚ>AÕ9<Ð7:¸"Â*)Ö>=åBEÙ69Õ,/Ú14Ö23Ì()Í54ãKJòdcøjiõdiòaf÷_kø`lÏPIÏPIç\YãXUæNMàHGÍ31Í31ÕB:ÞKCáQFÛK@ØA6Ó<1Ö5-Ó2*ú.:ù-9ø*7õ'4ò)3ø/9ù?@ÿHIÝ2+Ý2+Ù5+Ù5+â4+ä6-î3.ï4/ö/2õ.1ø-0ø-0ô.+ô.+æ5+å4*Ó?1Í9+®3$²7(¶E=êyqÿŽ�üŠ‰ÿz‰ûq€õgwïaqüPjôHbÿ7^ÿ6]÷8Zþ?aÈR^´>J\8(b>.S^@`kMbIkˆRq’Yt•\u›`xžcx bu�_rœ\rœ\s›\s›\uœ]wž_zž`{Ÿax¢bz¤dy§_y§_x§Yx§Yy¨W{ªY}©\�­`…¬k„«j‰§uˆ¦tƒžu{–mv‡wm~nevf^O<^O<^O<^O<^O<^O<^O<^O<]N;]N;]N;]N;]N;]N;]N;]N;`O1aP2_P;^O:XOFZQH[V]_Zahf|qo…yyŸ~~¤}{¬yw¨wt©wt©my³q}·v€»yƒ¾…Áƒ‰Å��Â’”Ç›”µž—¸­ž£·¨­Ç±œ¾¨“¥‹fˆnIŸD/¡F1²B4·G9ÖIBàSLîHHçAAê<=é;<ã;:ã;:×@7Õ>5Ç=0Ä:-º:1¾>5®5,©0'·K>ÎbUÐiZÎgXÁP@²A1¹3'À:.Ö8-Ô6+à2+ä6/à2+ã5.Þ8,Þ8,à>/åC4å>.Û4$é3(ç1&ò0.ï-+í-,ñ10ð67ú@AïPTøY]æGKñRVÕ49çFKÙ9;âBD×9:Í/0Ì24âHJÏ9:Í78µ#$Ï=>â57æ9;í:>é6:ã+3à(0â19ë:BíLRý\bÿksÿhpú^kõYfúWhûXiòe^è[TìTQ×?<Ï0-Ó41Ò3/áB>âKBèQHèTJßKAÛ>5×:1Û60Û60ú*7ÿ0=ÿ/<ü*7û-7ÿ8BÿDEýABÞ3,Þ3,Ü5,Ü5,á3*ã5,î3,ð5.÷/2ö.1ù-0ù-0õ.+õ.+ç5+æ4*Ó=/Ò<.±3%´6(±<5étmý‰‰ý‰‰ÿq…ÿg{ðbrì^nüMjóDaÿ5^ÿ6_ò?_øEe¨ZZ„66KG.KG.SY=]cG`zJiƒSrŽ[w“`wš`{ždwŸ`rš[o™Wo™Wp™WqšXr›Yt�[wž_xŸ`w \y¢^y¤\y¤\v¥Uu¤Tw¦Ux§V}©\�­`†ªj„¨h†¤p…£o€šmx’ep„hj~bbvZ_P=_P=_P=_P=_P=_P=_P=_P=^O<^O<^O<^O<^O<^O<^O<^O<cR4bQ3`Q<^O:YPG]TK_Zac^esq‡{y�€€¦ƒƒ©}®zx©yv«zw¬r~¸v‚¼{…À~ˆÃ„ŠÆ‰�Ë”–É™›Î«¤Å±ªËÀ±¶Å¶»Ó½¨Õ¿ªÓ¹”ɯŠÁfQ´YD±A3¬<.ÎA:âUNøRRòLLì>?ê<=å=<ä<;ØA8Õ>5È>1Ä:-»;2¾>5´;2·>5½QDÅYLÅ^OÍfWÔcS¿N>¾8,À:.Õ7,Ó5*ß1*ä6/ä6/á3,Ý7+ã=1åC4äB3á:*Û4$ê4)ä.#ð.,ð.,ñ10ó32ç-.é/0âCGöW[õVZÿosäCHèGLÞ>@û[]ÿqrÝ?@ÿ~€ÿnpèRSÓ=>Ã12Ï=>ôGIé<>ì9=ó@Dð8@æ.6ñ@Hÿ]eÿntÿjpýck÷]eù]jýanÿ\mùVgí`YÞQJáIFÐ85Ð1.Õ63Ð1-Ø95ÜE<áJAßKAÕA7Ò5,Ò5,Ü71ß:4ñ!.ý-:ÿ0=û)6ü.8ÿ:Dü@Aï34ß4-à5.Ý6-Ü5,á3*â4+î3,ñ6/÷/2ö.1ù-0ù-0õ.+õ.+ç5+æ4*Ò<.Ó=/±3%³5'¬70æqjû‡‡þŠŠÿm�ücwí_oë]mûLiôEbÿ6_ÿ8aôAaúGg¥WWu''NJ1IE,PV:W]A]wGg�QpŒYv’_wš`z�cu�^p˜Yl–Tl–Tn—Uo˜Vp™Wr›Yuœ]v�^vŸ[w \w¢Zw¢Zt£Ss¢Ru¤Sv¥Tx¤W|¨[€¤d}¡a}›gy—crŒ_h‚U`tXZnRRfJYL;[N=^Q@^Q@]P?\O>]P?^Q@_RA\O>ZM<[N=^Q@`SB_RA^Q@cT?_P;\PD^RFWQUZTXb`vrp†tw x{¤}€·‚¹{¿uy¹or·lo´rvÑvzÕ€„ĈŒÌ™™±¤¤¼µ²¡½º©ÌÈ›ÒΡØÒ˜ØÒ˜ÛÒ™ÛÒ™×͘ÐƑͪŒ©†hš`Hº€hÔtf®N@ÖQLìgbéIIà@@â89ã9:ã99Þ44Ü41Þ63É83Ä3.ÅB:½:2»K?ÇWKµO@ËeVÏdTÅZJÆF9½=0Î4*Ò8.â5/á4.Þ1-Þ1-Þ7.ã<3à<0Ú6*Þ4%â8)æ.$ç/%ì/)î1+ë4.ë4.ã6/ã6/Ñ8:äKMóX\êOS×8=Õ6;øW_ÿ~†ÿ}…ÿowÿgnÿhoÿgjõ\_ïY[ô^`ÿQWÿBHþ6@ý5?ù1>ò*7ê0=ô:GúT^ÿ^hÿhpÿemýalþbmÿ_lý[híRMßD?Ù53Ö20Ù0-×.+×2.Ý84ÞD:ÜB8×?4Ô<1Ù80Ù80â62ã73î7;î7;ð69ñ7:ñ78ï56æ40ã1-Û4+Ü5,Ý7+Ý7+æ1(ç2)ò1,ô3.÷03ö/2ù.1ù.1ö/,ö/,ê5,é4+ÚB5Ï7*º8+±/"·<7ÓXSÿ—™ú~€ÿh€ÿc{ø\rîRhùDcøCbÿ;bÿ7^æTkÞLcsI=T*6@%?I.JL4[]EWlAbwLo‰Yx’bw–]u”[p—Vp—Vl•Ql•Qm–Ro˜TqšXr›Yt›\uœ]rœTs�Ut Uu¡Vv¢Uu¡Tt SsŸRz¤Zy£Yy�]s—WmŠTbIWqBQk<?X0=V.:S+YL;[N=]P?^Q@]P?]P?]P?_RA`SB^Q@\O>]P?`SBaTC`SB^Q@fWBaR=\PDaUI`Z^f`dmk�xvŒvy¢z}¦}€·z}´quµlp°lo´nq¶swÒy}؇‹Ë“—ק§¿µµÍÈÅ´ÐͼÚÖ©ÞÚ­äÞ¤ãÝ£äÛ¢âÙ ÛÑœÓɔڷ™¾›}±w_Å‹s×wiµUGßZUöqlöVVéIIæ<=ç=>é??å;;ß74Ý52Ï>9È72ÅB:¹6.°@4¼L@°J;Ã]NÚo_ÎcSÇG:¶6)È.$Ð6,à3-Û.(á40å84ä=4â;2Û7+Ù5)ß5&à6'í5+ì4*í0*ì/)ç0*ç0*à3,à3,Ñ8:×>@äIMòW[û\aû\aý\dÿ`hòRZíMUïRYü_fÿknÿgjñ[]èRTó06ï,2ó+5õ-7ï'4ç,â(5ï5B÷Q[ÿYcþdlÿemÿdoÿcnüZgóQ^áFAÙ>9Ù53Ú64ß63Û2/Ø3/Ú51Ú@6Ø>4Õ=2Ó;0Ù80Û:2ä84æ:6í6:ì59ï58ï58ð67î45å3/á/+Û4+Ü5,Ý7+Ý7+æ1(ç2)ñ0+ô3.÷03ö/2ù.1ù.1ö/,ö/,ê5,é4+Ø@3Õ=0½;.µ3&³83×\Wÿ�’ówyÿbzÿ]uôXnïSiúEdöA`ÿ9`ÿ:aâPgÑ?VlB6V, 3="9C(EG/PR:Rg<^sHj„TtŽ^u”[s’Ym”Sm”Si’Nj“Ok”Pm–Rn—Up™Wr™Zsš[s�Us�UsŸTsŸTsŸRržQpœOo›Nm—Mj”JhŒL_ƒCXu?Li3D^/>X)=V.>W/?X0XK:[N=]P?^Q@]P?]P?^Q@`SB`SB_RA_RA`SBaTCaTC_RA]P?`US[PNXQX`Y`gczqm„sužy{¤uz²x}µt|½ltµbi±`g¯gm·pvÀ}|Ú…„◖Ч¦à½»ÀËÉÎÜÙ¬äá´éåœíé îê¡ëçžçà©áÚ£ÖÏ¥ËÄš¯²—� …–„nšˆr¥p`�XHº]VÏrk÷usåcaåIJßCDì?Aë>@ë7:ç36Ý>:Ø95ÒB:Æ6.¶8,»=1­<,·F6ÚiYÜk[ÜZLÅC5È1&Í6+Ú6-Ö2)Û64çB@éH@Þ=5Ô4(Ó3'Ü1'Û0&è3*ç2)å0)ä/(ß4,â7/Ý?4ßA6Â.*È40ÙCBìVUø]aö[_óS]ðPZçGQèHRîNXü\fÿmsÿntóX^Û@Fð,0ò.2ú1;ü3=ó+8ï'4ñ7DÿIVñJRöOWõY]û_cÿchÿ`eùPWêAHß82Þ71ä73é<8í;9é75à5.ß4-Ù;0Ø:/Ö:.Õ9-Û7.Þ:1è94ê;6ß<7Þ;6á96â:7æ:6ä84Ý6.Ù2*Ø6+Ù7,Þ8*Ý7)ç0(è1)õ.)÷0+ó23ò12ö01ö01õ/,õ/,ê5,é4+Ø<0ÞB6Ä:0½3)´0.ã_]ÿƒˆónsÿ_xûWpÿMmýKkÿ@eø8]ê?_ðEeÀ\d£?GXE4P=,7@%7@%HC-KF0N^9ZjEg}OrˆZr�WqŽVk�Mk�Mg‘Ih’Ji’Lj“Ml•Qn—So—Xo—XtœVtœVrœTq›SnšOl˜Mj•Mi”LeŒI`‡D\~BRt8Lh5D`-AX*>U'B\,D^.Ic3WJ9YL;\O>]P?]P?]P?_RA`SB_RA_RA`SBaTCaTC`SB^Q@]P?YNL\QO_X_jcjqm„xt‹xz£{}¦v{³ty±jr³bj«\c«`g¯lr¼v|Ɖˆæ’‘藍ݳ²ìÈÆËÓÑÖâß²çä·ìèŸîê¡íé èä›ãÜ¥ÚÓœËÄš½¶Œ¢¥Š—š˜†p”‚l¢m]™dT¿b[Ïrkÿ„‚õsqóWXæJKîACì?Aï;>ì8;Û<8Ú;7ØH@ÒB:»=1¸:.©8(¤3#»J:ÕdTëi[×UGÎ7,Ê3(Ø4+Ø4+Ø31òMKü[SíLDÙ9-Õ5)Þ3)Ý2(á,#á,#à+$á,%Û0(Þ3+Ø:/Ú<1Á-)Å1-Ï98ÙCBßDHßDHåEOëKUðPZõU_÷WaúZdÿhnÿmsóX^×<Bò.2ò.2õ,6ó*4í%2ï'4÷=JÿQ^èAIç@HäHLïSWü[`úY^ðGNà7>Þ71à93é<8îA=ò@>î<:å:3â70Ø:/Ø:/Ö:.Ö:.Ü8/Þ:1ç83é:5ß<7Þ;6á96â:7æ:6ä84Ý6.Ù2*Ù7,Ú8-Þ8*Ý7)ç0(è1)õ.)÷0+ó23ò12ö01ö01õ/,õ/,ê5,é4+Ö:.âF:Ä:0¾4*².,ëgeÿzñlqÿ`yùUnüJjüJjþ>cô4Yè=]óHh¾Zb›7?[H7^K:AJ/=F+LG1GB,GW2Sc>`vHl‚Tn‹SoŒTiŽKh�Jf�Hf�Hg�Ji’Lk”Pl•Qm•Vn–WtœVs›UpšRm—Oi•Jf’GcŽFb�E_†CZ�>Vx<Oq5Mi6Kg4Ne7Ne7Ke5Nh8Sm=SG7VJ:YM=[O?[O?\P@]QA_SC^RB^RB_SC_SC^RB^RB^RB^RBRK[^Wgjftp‰utœxwŸuz±v{²qyºiq²_g¯Zbª^f¯iqºu}Å|„Ì“–ÛœŸä­¯ßº¼ìÉËØÐÒßÚÛÍÝÞÐäàÃåáÄãÝÃÝ×½ØÏÀÎŶ¾³¯¯¤ “›�ˆ�…ˆ†z�s�ul‘vm§ne­tkä‰„à…€ñtpæieñWWéOOîDEé?@ä92å:3áC:âD;Ë=3É;1ÃA4²0#­/ ÇI:ÜWHÓN?Ì=-É:*Ð7)Í4&Ó54ôVUÿkgø]YâE<Ù<3Þ:0Þ:0á3*à2)Þ1*Ü/(Ï1&Í/$¿1#¾0"Â:*¼4$½3(Ç=2ÕDA×FCÙBGØAFÞDNêPZðV`ñWa÷`gþgnð[_ÙDHé:7ç85ë46ë46í4:ô;AýIRÿU^á8=Ý49Þ58é@CýILýILø:<ì.0ç60ë:4ò;7ó<8ò;7ð95ç83ä50Þ;2Ý:1Ù;0Ø:/ß6/à70è51é62ã;8â:7æ95ç:6é:5ç83ß7.Û3*Ù7*Ú8+Ü9*Û8)æ1&æ1&ñ.(ó0*î53í42ó11ó11ó0,ó0,ê5,é4+Ú9/âA7Ç7/À0(½12÷klÿqzøhqþ[vôQlÿ>eÿ=dÿ7`ÿ2[×NbàWk•`ZxC=WQ;_YCQO:KI4QI6JB/CQ0M[:ZlBgyOk†OmˆQiŒHh‹Ge�Ef�Fg‘Ih’Jj“Ol•Qm•Vm•VršTp˜Rm•Oj’LeŽHb‹E`‰E_ˆDZ�BV}>St=No8Pj:Qk;WnBYpDTq;Ur<Vs=PD4SG7WK;YM=YM=ZN>]QA^RB^RB_SC`TD_SC^RB^RBaUEcWG]Vfleuxt�~z“zy¡yx sx¯sx¯iq²bj«[c«\d¬go¸u}Æ€ˆÐ…�Õ˜›à £è°²â»½íÇÉÖÌÎÛÔÕÇ×ØÊÜØ»Û׺×Ñ·Ñ˱ÍĵĻ¬´©¥¥š–…�‚t|qxvjqoc…ja‹pg—^UšaXÈmhÔytø{wù|xÿmmýccúPQðFGïD=é>7Ü>5àB9Ë=3ÑC9ÙWJÀ>1´6'¹;,Â=.Ä?0ÐA1ÔE5×>0Ð7)Ð21êLKú_[òWSáD;×:1Ü8.Ý9/æ8/æ8/å81ä70Ø:/Õ7,Ç9+Å7)°(¹1!É?4ÒH=ÔC@Ì;8Ì5:Ð9>Ñ7AáGQîT^òXbô]d÷`géTX×BFë<9í>;ö?AýFHÿLRÿPVÿT]ÿU^Ý49Ø/4Ù03â9<ò>Aò>Aó57ï13è71ë:4ð95ì51é2.ê3/ã4/â3.à=4ß<3Û=2Ú<1á81á81è51è51ã;8â:7ç:6ç:6é:5ç83ß7.Û3*Ú8+Û9,Ý:+Û8)æ1&æ1&ñ.(ó0*î53í42ó11ó11ó0,ó0,ê5,é4+Ü;1Û:0É91À0(ÎBCüpqûktôdmòOjîKfÿ9`ÿ7^ÿ6_ÿ9bÛRfÜSg�LFt?9VP:]WAUS>NL7OG4KC0?M,GU4Rd:_qGe€Ij…NgŠFgŠFdŽDdŽDe�Gg‘Ii’Nj“Ok“Tl”Uo—Qm•Oj’Lg�Ib‹E`‰C^‡C^‡C[‚CX@Vw@Rs<Uo?Vp@\sG^uIXu?Wt>Vs=NB2QE5TH8WK;XL<YM=[O?]QAaUEaUEbVF`TD_SCaUEfZJj^NifƒtqŽyyŸyyŸuwªuwªot´jo¯_g¯^f®_g°em¶pxÀ}…͈‘Ö�–Û—£É�©Ï©³×±»ß¹¿á»ÁãÀÃâÁÄãÄÁÞ¿ܿ¹Õ¹³Ï¸®É²¨Ã¥˜´–‰¥‡†‹qputpokgfwhc€ql‚`WŠh_�ZQ­jaØskäwÿ{vþytÿheô\YÿPJ÷F@å>6êC;Ö<2ÞD:ê\PÇ9-Ã;-¾6(¾6&Æ>.ÐC1ÐC1ÓB/ÔC0Ð79Ö=?ÚEAØC?Õ<6Ñ82Ø5.Û81å:2å:2â;3á:2×@5×@5ËF5ÌG6«/¸<"ÌK6ÎM8Ä>5·1(¿02É:<ÔCJÞMTéV^íZbî]dëZaÝNRÏ@DÍ5*Ñ9.Þ<7äB=îDEðFGðCGì?Câ59ß26æ25ì8;û47÷03ÿ+/ÿ.2ë4.ï82ò72ë0+ç+)ê.,ä3-ã2,á=3á=3Þ=3Þ=3ä92ä92ë52ë52í57ì46ï34ð45ï64í42â4-Þ0)Û7+Ü8,Ý:+Ü9*ã3&ã3&ì0%î2'ì63ë52ð31ð31ò1,ò1,ê5,é4+ß<3Ô1(Ð;5Ç2,çUXÿqtøbmðZeæHaèJcÿ4`ÿ/[ÿ4aÿ;hÉVc¼IVV?/]F6QO8RP9\O>UH7LA/MB0;H*AN0J\4WiA^yBf�JeˆDf‰EbŠAc‹Bc�Ee�Gf�Kh‘Mi‘Rj’Sl“Pj‘NgŽKd‹HaŠH_ˆF^†G^†G_…J\‚G^|H\zF]wH^xIbwLdyNYv@Yv@Yv@L@0OC3SG7UI9WK;XL<[O?]QAcWGcWGcWGaUE`TDcWGi]MocSur�yv“vvœss™qs¦su¨jo¯af¦Ya©]e­em¶lt½u}Ňϋ”Ù“œá˜¤Ê�©Ï¥¯Óª´Ø¯µ×¯µ×±´Ó²µÔ´±Î²¯Ì®¨Ä¨¢¾©Ÿº¤šµ˜‹§Š}™zy~bafhdc_[Zj[Vte`xVMˆf]ŽKBšWNÂ]UÒmeùtoÿzÿvsÿliÿZTÿOIìE=óLDÜB8ßE;æXL¹+»3%½5'Æ>.ÎF6Ë>,À3!È7$ØG4ÝDFÒ9;È3/Ë62Ó:4Õ<6Ý:3à=6æ;3ã80Ý6.Ù2*Í6+Í6+Â=,Ã>-ÑU;ÄH.½<'½<'Â<3Á;2Ã46Â35ÜKRÝLSáNVæS[èW^æU\ÛLPÒCGÌ4)Ï7,Ø61Ø61Ý34Þ45à37ß26å8<ä7;í9<ñ=@û47õ.1ÿ*.ÿ04ê3-ð93ó83ì1,è,*ë/-æ5/å4.à<2à<2ß>4ß>4æ;4æ;4í74í74ë35ë35î23ï34í42ë20à2+Ý/(Ü8,Ü8,Ý:+Ü9*ã3&ã3&ì0%î2'ì63ë52ð31ð31ò1,ò1,ê5,é4+à=4Ï,#Ô?9Ê5/ödgÿqtó]héS^äF_êLeÿ7cÿ/[ÿ4aÿ;hÁN[ª7DH1!]F6PN7LJ3\O>VI8I>,NC18E'=J,EW/Rd<Zu>c~Gc†BeˆD`ˆ?a‰@a‹Cc�Ed�If�Kg�Pg�Pj‘Nh�LeŒIcŠG`‰G_ˆF_‡H_‡H\‚G[�F_}I^|H`zKa{Lf{Pg|Q[xB\yC]zD?C5BF8IF=KH?RH?TJA\L<_O?`R8`R8\T?^VA^ZWfb_jiwqp~uo›zt {w©zv¨op©gh¡`a¤]^¡_a®eg´qoÃ{yÍ…�ØŽŠá—‘뜖ð ŸÙ ŸÙ¤ Û¥¡Ü¥¡Ü¥¡Ü¤ŸÙ¤ŸÙ¤œÕ¢šÓ ™ÍŸ˜Ì�•Æ–Ž¿‹ƒ²‚z©potgfkd_\b]ZoXRnWQ‚TG‡YL™OD¥[P¶XNº\RÕmdä|sð…}ú�‡ðxwòzyça`ØRQØDBîZXÌ8.Æ2(Ä7&Ã6%À7$À7$Å8&É<*Ð=-Ò?/Ó7(×;,×@/Ó<+Í6%Ì5$Ù9+áA3à:.ß9-Ú8+Ø6)Ë:'Ë:'ÁA(ÃC*Ã:2Ç>6ÌC;ÏF>ÌC;Ä;3¹3*³-$Ã=4ÍG>×QHÝWNá[RßYPÕOFÊD;Í<!Ï>#Þ7'Þ7'ï+-í)+ø$3ù%4ÿ(?ÿ(?ÿ(@ÿ(@ÿ$Aÿ$Aÿ"Aÿ"Aÿ :ÿ"<ÿ(=ÿ&;ð,6ï+5ã77æ::Û94Û94Ü;3Ý<4å:3ä92è62è62ì23ì23é62é62ã80á6.Ù7*×5(Ú;(Ú;(Û:&Û:&ß9#ß9#â7#â7#ë4.ì5/ï4+ð5,ï5)ï5)å8(ä7'Ü<0Ø8,Õ:8Ú?=ÿjzÿaqÿQlõGbÿFh÷5Wÿ4\ÿ:bèEbîKh”WV_"!35 :<'EF4MN<MJ9EB16E.9H1EHHK JR-U]8ZjEdtOcNe�Pa†Cb‡DaŠ<cŒ>e�Dg�Fl‘Nm’Oi’Fg�De�Fc‹DeŠGd‰Fc…HaƒFdƒJb�HbEbEd€Fd€FcDa}BX}8[€;^ƒ>;?1>B4FC:HE<OE<QG>ZJ:]M=_Q7_Q7\T?_WB_[Xgc`lkyqp~sm™uo›tp¢plžfg `aš\] \] ce²ik¸trÆ}{Ï…�ØŒˆß“�痑똗љ˜Òœ˜Ó�™Ô�™Ôœ˜Óœ—Ñ›–Ð�•Î›“Ìš“Ç™’Ƙ�Á‘‰º†~­~v¥onsfejc^[a\YnWQmVP�SF…WJžTI©_T¸ZPº\RÏg^Úriávnêwý…„ÿ‡†útsìfeæRPíYWÇ3)Â.$Ä7&Ä7&À7$À7$Å8&Ç:(Î;+Ð=-Ô8)Ô8)Ò;*Ñ:)Ñ:)Ð9(×7)×7)à:.ß9-Û9,Ù7*Ì;(Ê9&½=$¼<#¾5-Á80Å<4È?7É@8Ç>6Â<3À:1»5,Ä>5ÏI@ÕOFØRIÕOFÉC:¾8/Ê9Í<!Þ7'ß8(ñ-/ð,.ü(7ý)8ÿ*Aÿ*Aÿ(@ÿ'?ÿ"?ÿ!>ÿ=ÿ<ÿ7ÿ9ÿ%:ÿ&;ò.8ó/9å99ç;;Û94Û94Ü;3Ý<4å:3ä92è62è62ì23ì23é62é62ã80á6.Ù7*×5(Ù:'Ù:'Ú9%Ú9%Þ8"Þ8"á6"á6"é2,ê3-í2)î3*î4(í3'ã6&ã6&Ý=1Ô4(Ñ64ñVTÿeuÿ\lõGbúLgÿ@bþ<^ÿ1Yÿ5]ñNkã@]|?>a$#79$79$:;)@A/DA0A>-5D-9H1@CMP%Zb=goJfvQhxScNe�P`…Ba†C`‰;b‹=e�Dg�Fk�Ml‘Nj“Gh‘Ee�FdŒEf‹HeŠGd†Ib„GdƒJc‚IbEbEd€Fe�GcDb~C\�<_„?b‡B4:,7=/??7BB:JA:MD=TH:WK=[P<[P<ZSC^WG`^ahfilm�pq…qn™nk–ig˜ca’[\•Z[”Y\Ÿ\_¢gkµmq»wxÈ~φ„׊ˆÛŽŠÞ�Œà‘‘Í‘‘Í“‘Г‘Ð’�Ò‘ŽÑ‘ŒÏ‘ŒÏ’ŒÌ‘‹Ë’‹Á’‹Á’‹·Œ…±‚|¢ztšƒbk{Zce\Wd[V]^L\]KuVDxYG¬UN¹b[ÓWWÓWWÝc`ãifÙqhàxoú‹„ûŒ…ÿ„~û~x÷leìaZÂ6)¿3&Á:'À9&½;%¼:$Á8%Ã:'É:)Ê;*×9-Ó5)Í5'Ð8*Ñ>.Ñ>.Î6(Ç/!Ù9-Ø8,Õ7+Õ7+Î;+Ì9)À9%¾7#¾4)¿5*Á7,Ä:/Ç=2É?4ËA6ÌB7Å;0Ê@5ÎD9ÐF;ÑG<ÑG<Ê@5Á7,Ð5!Ó8$â4+ä6-ô-2ô-2ü)8ý*9û(;ú':ù%;ø$:ù!8ø 7û4ú3ÿ6ÿ6ú&5ý)8ò5<ó6=å;;ä::Û94Û94Ý:3Þ;4å:3ä92é62é62í34í34é54é54ã81á6/Ø6)×5(Ù9)Ù9)Ù:%Ù:%Û9$Û9$Þ7$Þ7$Ý5(Þ6)à6'á7(ß8&ß8&Ù:'Ø9&Ú>2Î2&Ö7;ÿlpÿ[nýReó;]ÿJlÿ2[ÿ:cð8\ô<`Ò^q°<OK,)L-*0;*-8'19$6>)B='A<&A="D@%>HWa/pT}Œat‰^i~S`~JaK_„?`…@`‰;cŒ>e�Dg�Fk�Ok�Ok“Ji‘Hh�GgŒFgŒIf‹HfˆKd†Ie„Ic‚Gc€Fc€Fd‚FeƒGd€CcB`ˆBbŠDe�G06(39+;;3>>6G>7KB;RF8UI;XM9YN:XQA^WGa_bigjlm�op„ol—jg’`^�XV‡ST�WX‘[^¡ad§ko¹qu¿z{Ë�‚Ò‡…؉‡ÚŒˆÜ�‰ÝŽŽÊŽŽÊŽŒËŽŒË�ŠÍ‹ˆËŒ‡ÊŒ‡ÊŒ†Æ‹…Å�†¼Ž‡½Ž‡³ˆ�­yŸwq—�`iyXad[Vd[V]^L[\JsTBvWE¨QJ¶_XÔXXÖZZàfcåkhÚriáypë|uë|uÿ‚|ÿ‡�ÿ€yõjcÉ=0Ä8+Â;(Á:'½;%½;%Á8%Â9&Ç8'È9(Ø:.Ô6*Í5'Ï7)Ð=-Ð=-Í5'Ç/!Ô4(Ô4(Ò4(Ó5)Í:*Í:*Ä=)Ã<(Ã9.Ã9.Ã9.Ã9.Ä:/Æ<1Ç=2È>3ÐF;ÐF;ÌB7Ç=2È>3ÌB7ËA6È>3Ð5!Ó8$ã5,å7.ô-2ô-2û(7ü)8ù&9ù&9ù%;ù%;ü$;ý%<ÿ#:ÿ#:ÿ 7ÿ6ú&5þ*9ó6=ô7>ä::à66Û94Û94Ý:3Þ;4å:3ä92é62é62î45î45ê65ê65ã81á6/Ø6)×5(Ø8(Ø8(Ø9$Ø9$Ú8#Ú8#Ý6#Ý6#Ü4'Ü4'ß5&à6'Þ7%Þ7%Ù:'Ø9&Ö:.Ð4(èIMÿrvÿVióH[õ=_üDfÿ1Zÿ5^÷?cûCgÄPc—#6@!K,)*5$,7&5=(;C.C>(@;%B>#HD)S]+r|J�žs˜§|‚—ll�V^|H_}I^ƒ>a†Ab‹=f�Ah�Gj’Il�Pl�Pl”Kj’Ih�GgŒFgŒIgŒIg‰LfˆKe„IdƒHd�Gd�Gf„Hf„Hf‚Ed€CdŒFfŽHj’L,4'/7*692:=6A>9DA<KD:NG=QK=QK=ROFYVM__iggqij†klˆih”ba�UWŠOQ„MQ�UY˜^b«ei²mrÃrwÈ|Ђ…Ö†‡×‡ˆØ‰ˆÔŠ‰ÕŠ‹ËŠ‹Ë‰‰ÏˆˆÎ‡„Ó‡„Ó‡‚Ô‡‚ÔŒ„ÍŠ‚ËŒ„½�…¾�‡«‡�¥}z�ur‡‘S^‹MXgYPhZQRfKOcHeYAh\DžG@¯XQÚOVÝRYè]bíbgßliçtqáuiÝqe÷|tþƒ{ÿ‡|òlaÈ@0À8(¿='¿='½=$½=$Â9&Â9&É8'É8'Û7-Ú6,Ð8*Ð8*Ç<)Ç<)Æ;(Å:'Ì9)Í:*Ñ8*Ð7)Í7(Í7(É<*Ë>,È<-È<-É;-É;-Ç9+Æ8*Æ7)Å6(Ï@2ÐA3Ï=0Ê8+Ë7+Ï;/Í9-É5)Ô-%Ø1)è00ë33ö-7ö-7ú+;ú+;ù/?ù/?ö0=÷1>ü0<ý1=ÿ/<ÿ/<ú*8ù)7ò-7õ0:í9<ì8;Þ97Ù42Û94Û94Þ93ß:4æ95å84é54é54ñ35ò46í76ì65ã81á6/Ø6+×5*Ô8)Ô8)Ô9'Ô9'×8%×8%Ø7#Ø7#Ñ:%Ñ:%Ï=$Ð>%Ð@(Ð@(Ï>+Ï>+Ð50×<7ÿ\gÿalÿMlö;Zÿ@ký6aÿ2]ÿ*UßUnáWp|KQX'-40-730(2).8/5A+7C-;<<=VF"fV2t‰N�¤i¦¾Š¦¾ŠŠ¦uo‹Z_€G^F]…>`ˆAeŽ@i’Dl”Km•Lo“Sn’Ro•Ll’Ih�GgŒFgŒIh�Ji‹NhŠMd†Jc…IdƒGdƒGf†Gg‡Hg…EeƒCfŽGi‘Jl”M*2%-5(4707:3=:5@=8G@6JC9NH:NH:NKBVSJ]]geeoef‚ef‚a`Œ[Z†PR…MO‚PT“[_žei²lp¹sxÉw|Í‚Óƒ†×…†Ö…†Öˆ‡Ó‰ˆÔ†‡Ç…†Æ„„ʃƒÉƒ€Ïƒ€Ï…€Ò…€Ò�…΋ƒÌŒ„½Œ„½‹…©„~¢yv‹qnƒŽP[ˆJUgYPi[RSgLQeJfZBh\DžG@°YRÛPWÝRYæ[`ë`eÝjgåroç{oßsgö{sîskþxmã]RÃ;+½5%¿='¿='¾>%¿?&Ä;(Ä;(Ê9(Ê9(Ø4*Ú6,Ó;-Ð8*Å:'Ä9&Æ;(É>+Ï<,Ó@0ÚA3Ö=/Î8)Ê4%Æ9'Ê=+Æ:+Ç;,É;-É;-È:,Ç9+Æ7)Å6(Æ7)Ë<.Ñ?2Ñ?2Ó?3Ò>2Ë7+Ã/#Ò+#Ö/'è00ì44ù0:ù0:ý.>þ/?ú0@ú0@ö0=õ/<ù-9ù-9ü*7ü*7ý-;ü,:õ0:õ0:ê69é58Û64Ù42Û94Û94Þ93ß:4æ95å84é54é54ò46ó57î87í76ä92á6/Ø6+×5*Ó7(Ó7(Ó8&Ó8&Ö7$Ö7$×6"×6"Ï8#Ð9$Î<#Ï=$Ï?'Ï?'Ï>+Ï>+Ð50åJEÿgrýS^ÿHgõ:YÿAlø1\ÿ3^ÿ)TÝSlÓIbf5;S"(>:7730-7.0:12>(/;%56AB"m]9Œ|X›°uª¿„µÍ™®Æ’’®}w“bd…L`�H]…>a‰Bg�Bl•Go—Np˜Oq•Uo“So•Lk‘HgŒFf‹Ef‹Hh�Ji‹Ni‹Ne‡Kc…IdƒGe„Hg‡HhˆIh†Fg…Eh�Ij’Km•N%0"(3%/4-16/664997?<7B?:HE<HE<HIDQRMY[h`bo_c€]a~\]�WXˆRSŒTUŽZ\¦fh²nsÇsxÌy~Ù|�Ü€„Ý�…Þ‚„ׂ„׃…Òƒ…Òƒ‡Ç‚†Æ�ƒÍ€‚̃ÓƒÓ…Õ†€Ö�…ЋƒÎ‹„·‰‚µ‡ƒš{’usxljo„TROMeYIh\LYhKWfIc_BeaDŒQAžcSÌXXÌXXäY^è]bê_dóhmózoërgúulÝXOçYMÑC7Ä5$È9(Ã;%Ã;%Ä<&Å='Í:(Í:(Ô6*Ô6*Ò/&Õ2)Î:,Î:,Â=*À;(¿<(À=)Ç;*ÑE4àJ<ÝG9Ò<.Ê4&È6'Í;,Æ9'Æ9'È9(È9(Ê9(Ê9(Ë8(Ë8(Î6(Ò:,×>0Ø?1Ü@3Ü@3×9-Ï1%ß--ä22ð39ó6<ú2?ù1>ö0?ö0?í/9í/9ê17é06ë.2ë.2ï+/ï+/ñ4:ò5;î7;ì59â66à44Ú72Û83Ü94Ü94à85á96ç77æ66é56é56ó49ó49ï77î66ä84á51Õ7,Ó5*Ï9*Ï9*Í:(Í:(Ï9(Ï9(Ð9&Ð9&Å;!Æ<"Ä>#Æ@%Ä@)ÅA*Ë=/Ë=/Û6:úUYÿ[rÿH_ÿ<dÿ9aÿ5fÿ1bñ>\ôA_¨[a‹>D<;6:94=;>758B44A336;%5:$4FK]-€ƒJ£¦m©Ç‡­Ë‹®Í”¤ÃŠ�°{x™dfŠM]�D\„=`ˆAg�Bm–Hp˜Op˜On•Tm”Sn”Kj�Gf‹Ed‰CeŠGgŒIhŒNhŒNe‡Jd†Ic†Fd‡Gf‰EgŠFfŠDe‰Cg’Jj•Mm˜P#. &1#,1*.3,331553;83>;6FC:FC:FGBOPKWYf^`m\`}Z^{]^ŽZ[‹WX‘\]–df°qs½x}Ñ|�Õ„ß�†áƒ‡à‚†ß�ƒÖ�Ô€‚Ï�ƒÐ„ˆÈƒ‡Ç‚„Î�ƒÍ„€Ô†‚Öˆ‚؉ƒÙŒ„ÏŠ‚͉‚µ‡€³„€—{wŽpnsgej‚RP~NLeYIi]MZiLXgJeaDfbE†K;˜]MÅQQÆRRÞSXâW\åZ_îchîujépeõpgÊE<ÐB6Á3'Ä5$ÒC2Ã;%Ã;%Å='Æ>(Î;)Î;)Õ7+Õ7+Ï,#Ñ.%Ë7)Î:,Ä?,Â=*½:&º7#¿3"ÏC2äN@äN@×A3Í7)Ê8)Ï=.È;)Ç:(Ç8'Å6%È7&É8'Ë8(Í:*Ò:,Ó;-Ò9+Ï6(Ô8+Ù=0Û=1×9-é77í;;÷:@÷:@û3@÷/<ò,;ð*9ì.8í/9ê17ë28ï26ð37ö26ö26ð39ò5;î7;ì59á55ß33Û83Þ;6Ü94Ü94à85á96ç77æ66é56é56ó49ô5:ï77î66ä84á51Õ7,Ó5*Î8)Î8)Ì9'Ì9'Î8'Î8'Ï8%Ï8%Ã9Ä: Â<!Ä>#Ã?(Ã?(É;-É;-ß:>ÿ_cÿQhÿKbÿ6^ÿ>fû+\ÿ7hí:XÿMkœOUp#)=<7CB=:8;97:B44A338='>C-EW'`rB”—^µ¸­Ë‹ªÈˆ¤ÃŠ›º�‹¬wx™de‰LY}@Z‚;_‡@f�Al•Gp˜Op˜On•Tl“Rm“Jj�Gf‹Ed‰CeŠGgŒIhŒNhŒNe‡Jd†Ic†Fd‡GgŠFh‹Gg‹EfŠDi”Lk–Nn™Q(0!'/ (-&).'.0-1306729:5<>3?A6CF?ILEOR[UXaY[r\^uYX„[Z†]^—dežkmºvxÅ~܃„ቋ酅傂ØÕ‚�σ‚Ѐ„À„ÃɃƒÉ…‚φƒÐŠ„Ί„Ή�½ˆ€¼ˆ�£„}Ÿ�z‚vowkg^c_Vg_Lf^Kf^If^Ie_Gf`Hg`DhaElQ6w\A¥]N°hYÖVSÔTQ÷MVÿW`øiaçXPÚA9Ï6.Ö5+Ú9/Ù<-×:+Ï8%Ð9&Ï:&Ï:&×7)×7)Þ3+ß4,Ê0$Ì2&È;*Ê=,ÃB-Á@+¾=(»:%Ã7&½1 êVHßK=Á/ Ï=.Ë>-Å8'È;)È;)Ç8'Æ7&È5%È5%Ë3%Ì4&Ò6)×;.Ý=1Ü<0Ú6,Ú6,ä=4ëD;ù@Eñ8=ñ2:ó4<õ3>ð.9è.9ì2=æ9=ã6:ß76à87è;7ê=9ð95î73ë=<ê<;ç;9å97à85ß74Ü73Ü73Ý84Þ95ä88æ::é9;é9;ì8;ì8;ô5:ô5:î87î87ä=7á:4Ò:/Ï7,Æ9(Æ9(Ã:(Ä;)Ã<)Ã<)Ä;)Ä;)½9 Â>%º<#³5¹;%ÅG1Ð=6Å2+ÿ[hÿWdÿGeÿ;Yö2Xõ1W÷8cý>iÌgm«FLV>4K3)5E87G:E56>./Q-/R.0=< GF*Qs7zœ`—»q£Ç}©ÈŒ�¼€—µ“±{�žnfƒSVw>Xy@]�;b†@hŽAm“Fl”Kk“Ji’Pi’Ph�GeŠDa†Aa†Ad‰Ff‹HfŠJdˆHg‹Kg‹KgŒIeŠGc‹DdŒEh�Gk“Jm›Po�Rr U(0!(0!(-&).'-/,02/450783;=2=?4AD=GJCLOXSV_VXoY[rZY…]\ˆab›jk¤rtÁ}Ì…†ãŠ‹èŠŒð‰‹ïˆˆè„„䀀Ö}}Ó€Í�€Î‚†Åƒ‡Æ††ÌˆˆÎ‰†Ó‰†Ó‹…ÏŠ„Έ€¼†~º„}Ÿ~w™{t|rkshd[b^Uf^Kf^Ke]He]He_Ge_Gg`DhaEx]Bx]B›SD¤\MÑQNÕURùOXÿV_çXPÝNF×>6Ñ80Ø7-Ú9/Ø;,Ö9*Ï8%Ð9&Ï:&Ï:&×7)×7)Þ3+ß4,Ë1%Ì2&Æ9(É<+ÂA,Á@+¾=(»:%Ã7&È<+äPBÜH:Ç5&Ë9*Ç:)Â5$Ê=+É<*Ê;*É:)Ì9)Ë8(Î6(Î6(Ò6)Ô8+Ø8,Õ5)Ô0&Ö2(Þ7.ã<3úAFñ8=î/7î/7ñ/:ð.9ê0;í3>ã6:Þ15Ú21á98ñD@÷JFûD@õ>:ç98æ87ä86ã75à85à85Þ95Þ95Ý84Þ95å99æ::é9;é9;ì8;ì8;ó49ò38í76í76â;5à93Ñ9.Í5*Æ9(Æ9(Ã:(Ã:(Â;(Â;(Ã:(Ã:(½9 Â>%¸:!³5¹;%¼>(Ë81Ó@9ÿ[hÿR_ÿ>\ü7Uù5[û7]ú;fû<g¦AG™4:V>4O7-1A41A4I9:N>?R.0X46PO3gfJp’V�¯s�Áw¡Å{™¸|™¸|˜¶€ŒªtqŽ^WtDQr9Z{B\€:b†@hŽAl’Ek“Jj’Ih‘Oh‘Of‹EcˆBa†Ab‡BeŠGgŒIg‹KfŠJhŒLi�Mh�JgŒIdŒEfŽGi‘Hm•Lm›Po�RqŸT-1",0!,/&,/&-/,/1.25,36-8<-:>/?C4EI:KMJQSPUUaWWc\Zr`^vedŒpo—yy¹„„Ä‹‹á��æ�ŽìŽ�늊䄄Þ�~Í~{Ê~€Ã…‡Â†ˆÃŠŠÆ‹‹Ç�ŠÃ‹ˆÁ‹„·Šƒ¶Š€£…{ž�v†zoujfod`j^PfZLWdJWdJd^Hd^HoXFpYGp\Dq]EwbEq\?‰Q8”\CÑRIØYPÿIOÿKQâ?:ß<7ß6/Þ5.â6,â6,Ý7)Û5'Ò7#Ó8$Ñ8&Ñ8&Ù5)Ù5)à2+á3,Ã7(Ã7(½:(À=+¿@-¿@-À;*¾9(Å3&ÛI<ÚF:Ô@4Ë=/Ä6(¿<*»8&Å:'Å:'Ç:(Ç:(Ê9(É8'Ì4&Ì4&Ö8,Ô6*×3)Ö2(Û3*Þ6-å:2ç<4ñ:>ì59é06ë28ì4<ì4<ä9?ç<BÕ:8ÝB@éPKóZUüZUôRMîC<ã81ß82ß82ß82ß82à93á:4ã:7ä;8ã77ä88æ89ç9:é9<é9<ë8<ë8<ï58î47æ66ç77Ý;6Ú83Ê:/Æ6+¾:+¾:+¸=+¸=+¶=*¶=*¶=*¶=*µ;&»A,¯:&¬7#¶;,®3$Ç12çQRÿWjýJ]õ:Uñ6QèD]íIbÚOdÕJ_cF8^A3FE1CB.5>)1:%B4'H:-G-U;*aa;„„^‹­p˜º}�Ãx�Ãx”°}›·„™³†€šm\uMHa9Nh8\vF_~;dƒ@g�@k‘Di“Ih’Hc�Kc�KdˆBc‡Ac†Bd‡CgŒIiŽKjŽNjŽNh�Li�Mg�Jf�Ie�Eg‘Gk–Gn™Jq¢Ss¤Uu¦W04%04%/2)/2)/1.02/14+25,8<-:>/>B3CG8HJGMOLQQ]SS_WUm[Yqa`ˆml”vv¶‚‚‰‰ßŽŽä’‘ï��파慅߂Î~{Ê~}Á~Â�ƒ¾„†ÁˆˆÄ‰‰Å‹ˆÁ‡„½…~±‚{®„z�~t—xm}peumb^i^Zh\NeYKWdJWdJd^Hd^HpYGqZHq]Er^FoZ=lW:ŠR9™aHÖWN×XOù@Fõ<BÖ3.Ø50Þ5.ß6/á5+á5+Ü6(Ü6(Ò7#Ó8$Ñ8&Ñ8&Ù5)Ù5)à2+á3,Ä8)Â6'»8&¼9'½>+¾?,Á<+¿:)Æ4'çUHÑ=1Ì8,ÏA3À2$¿<*¿<*Ã8%Ä9&Æ9'Ç:(Ê9(É8'Ì4&Ì4&Ó5)Ñ3'Õ1'Ù5+â:1ç?6ê?7é>6ê37ë48î5;î5;ê2:è08á6<å:@ñVTýb`ÿmhÿhcõSNãA<á6/ß4-Þ71Ý60Þ71Þ71ß82à93ã:7ä;8ã77ä88æ89ç9:é9<é9<ê7;ê7;ï58î47ç77ç77Ý;6Û94Ê:/Æ6+¼8)¼8)¶;)¶;)´;(´;(´;(´;(³9$¸>)­8$¬7#´9*¨-Ç12øbcÿPcùFYô9Tô9TêF_íIbÖK`ÏDYiL>`C5FE1ED0=F17@+@2%>0#G-Y?.kkE��iŽ°s‘³v†ºo‰½r�¹†�¹†Ž¨{iƒVHa9>W/Jd4WqA^}:c‚?fŒ?j�Ch’Hf�Fb�JaŽIb†@c‡Ad‡Cf‰Eh�Jk�Mm‘Qm‘Qi�Mi�Mh‘Kg�Jg‘Gi“Im˜Ip›Lt¥Vu¦Ww¨Y78&67%57,57,56056057*68+;?(<@)?C*DH/HJ<LN@POKRQMVVLZZP``hjjrttš€€¦‡…Æ�‹Ì”‘Þ“�Ý��Õ‰ˆÎ…�¼€|·€|®€|®µ��·†„µˆ†·‰…¨…�¤ƒ{’wŽu}yowtgam`Zl\MjZKkZHjYGa_Ja_Jg\Jg\JmZIo\KwZHx[IƒS=ƒS=¥N;°YFÚMCÖI?è62ä2.Ú1*Ý4-ã5.ã5.á3(à2'Ú7&Ü9(Ï9!Ð:"Í;$Í;$Ô9'Ô9'Û5)Ü6*¹?*µ;&²8#³9$½<)À?,É:,Æ7)Ñ7-õ[QÖ>3Ë3(ÊB2º2"°<%¶B+Á9)Â:*Ä;+Ä;+É:,É:,Î7,Î7,Ð2)Ð2)Õ0*Û60ç;7ê>:è;7ä73ç85ë<9î<<é77à35ã68ãEDëMLÿ{uÿrlïb[ÙLEÐ71Í4.Ü73å@<Þ:1Ý90ß80Þ7/â62â62ä65å76æ68æ68ç7:è8;è8;è8;è7=ç6<î79í68æ:8æ:8Ù?7Ö<4Æ>2Â:.´;*´;*­=)­=)ª=(ª=(©<'©<'¦=(¨?*¡:)¡:)©9.£3(Î=Bûjo÷L^óHZèGYçFXÉV]ÇT[£]U�WONX@GQ9LG1JE/HD'D@#7H:K!C[)Tl:s„Nˆ™c�£k“¦n�²r—ºz¢´Ž“¥u†fRcC?N1BQ4Oa7Xj@_{>d€CgŒ?k�Cf’Ee‘D`�G_ŽFb…Ad‡Cf‰Eh‹GiŽKk�Mk’Qm”Si’Nj“Oj”Li“Kh”Ej–Gn›JqžMs¦Ws¦Wt§X<=+<=+;=2;=2;<6;<6;=0;=0>B+@D-BF-FJ1IK=MOAPOKRQMWWMZZP]]effnmm“yyŸ�À†„ÅŽ‹ØŽ‹ØŒ‹Ñ‡†Ìƒº|x³zv¨yu§xx®zz°}®€~¯�} }yœ|t‹xp‡tjroemm`Zi\VjZKk[Ll[Il[Ib`Kb`Kh]Ki^Lo\Kq^My\Jz]K‘aKŒ\F¦O<§P=Ë>4Ç:0Þ,(à.*Þ5.ß6/ã5.á3,à2'ß1&Ú7&Û8'Ï9!Ð:"Í;$Í;$Ô9'Ô9'Û5)Ü6*º@+´:%¯5 °6!»:'À?,É:,Ç8*Ø>4õ[QâJ?Ð8-ÉA1»3#©5²>'¿7'¿7'À7'À7'Å6(Ç8*Í6+Î7,Ú<3Ù;2Þ93á<6é=9é=9å84á40è96ê;8ì::é77ê=?úMOÿdcÿnmøicãTNÉ<5¾1*Ì3-Õ<6á<8à;7à<3ß;2à91ß80â62â62ä65ä65ç79ç79è8;è8;è8;è8;ç6<ç6<î79í68æ:8æ:8Ù?7×=5Æ>2Â:.²9(²9(¬<(«;'¨;&¨;&§:%§:%¥<'¢9$ 9( 9(¤4)±A6áPU÷fkóHZòGYêI[èGYÆSZÂOVŸYQ�WODN6EO7MH2E@*:695=N$Rc9pˆVz’`‡˜b‰šd‰œd�¢j�°p‘´tƒ•omYO`@;L,8G*AP3N`6Wi?`|?e�DgŒ?k�Cf’Ee‘D`�G_ŽFd‡Cf‰Ej�IkŽJj�Lk�Mj‘Pl“Rk”Pl•Ql–Nk•Mj–Gm™JqžMt¡Px«\x«\x«\CB0CB0BB6BB6BB:BB:AD3@C2CG,DH-GJ+JM.MO7OQ9SQDUSF[]8[]8\]MbcShgwsr‚zy¥€«„ƒ½…„¾‡„»„�¸~|¤wu�sp�qn�omŸqo¡tršus›vsˆsp…rlnoikmcYj`Vl\LjZJn\Ho]Iq\Iq\ItXMtXMk^Nk^NjbOkcP~]L^MªZO¡QF»<5¹:3Ì2(Ì2(Õ2#Ù6'à8/ß7.á5+à4*ß5&ß5&Õ9#Ô8"É<!Ê="Æ>$Æ>$Í<'Í<'Ô8)Õ9*µE/­='©4 ©4 º6'À<-Ï8/Í6-Ú72ðMHïUMÚ@8ÇC4À<-¡4¬?(À;,¿:+¿7)¿7)Ä6*Ç9-Ñ:1Ó<3âC=áB<ä?;ã>:è<:è<:è88æ66å:3å:3ä84è<8ñLJÿ\Zûfb÷b^Æ?9¾71º4+À:1ÔA:ÚG@Þ<9Ö41à<3ß;2â92á81ä73å84è66é77ê69ê69ê7;ê7;è7=è7=æ7<æ7<ë78ê67â96ã:7Ö?8Ó<5Â>1¾:-­:'­:'¦;'¦;'¢<&¡;%Ÿ;$Ÿ;$š?,’7$–=+”;)™2)ºSJðgnç^eìN]éKZÖMWÒIS¨[Q§ZPlN~kM?Q9<N6O>*H7#@8ME!Gx7d•Ts·ju¹l…£cv”T‡‹X‘•b‹•c€ŠXWdHGT8;E-8B*:D,?I1JX4XfBd|Bi�GhŒBl�Fg“Fe‘D^‘F]�Ef‰EiŒHm�Lm�Lk�Mj�Li�Mj‘No˜Rp™So›PnšOmœLožNs£Ow§Sz­^y¬]x«\FE3FE3EE9FF:FF>FF>EH7DG6EI.FJ/IL-LO0NP8QS;USFVTGZ\7Z\7XYI\]Mbaqlk{sržyx¤{z´}|¶�~µ|³ywŸqo—mj‰if…ec•fd–ig�jh�kh}hezicef`bg]Se[QiYIk[Ko]Iq_Ks^Kr]JuYNuYNk^Nl_OkcPldQ^M€_N£SH™I>´5.¶70Ï5+Ñ7-Ù6'Û8)Þ6-Ý5,ß3)à4*à6'à6'Õ9#Ò6 É<!Ê="Æ>$Æ>$Í<'Í<'Ô8)Õ9*µE/­='¨3¨3º6'À<-Ï8/Í6-Ò/*ã@;õ[SáG?ÊF7ÉE6£6¯B+ÊE6ÉD5È@2Ç?1Í?3ÑC7ÛD;ÞG>ß@:Þ?9ß:6Ý84á53â64ç77é99è=6å:3â62ç;7ïJHñLJÓ>:»&"´-'¼5/Å?6ÉC:Ò?8Ò?8ß=:â@=ß;2ß;2á81â92å84æ95ê88ë99ê69ë7:ê7;ê7;è7=è7=æ7<å6;é56è45á85á85Õ>7Ò;4À</½9,­:'¬9&¦;'¥:&¡;% :$ž:#ž:#š?,Ž3 —>,“:(–/&Æ_Výt{àW^ðRaèJYÐGQËBL¤WM§ZP‚oQ‚oQFX@9K3F5!I8$XP,woKv§f‘Â�}Áty½p~œ\g…Ey}J„ˆUt~LZd28E)0=!4>&<F.<F.9C+CQ-Vd@e}Cj‚Hi�Cm‘Gg“Fe‘D_’G^‘Fh‹GkŽJo’No’Nk�MiŽKgŽKi�Mr›Ur›UržSq�RpŸOr¡Qw§SzªVs¦Wr¥Vp£TRK;RK;RK;RK;RK;RK;RK;RK;UN>UN>VO?WP@YRBZSC[TD\UEb[IaZH`XMbZOf\Zlb`rgmvkqznx|pzs}€t~|qyujrmbhh]cbbZaaY__U]]S^ZO^ZO`ZLa[Me]Je]Jg\Fg\Fi]Ei]Ek]Cl^DbfOgkTwVEqP?›ND¥XN°TG²VI©M8¢F1¦;'§<(Ã6,È;1ß12Þ01Û7-Û7-Ü8.Ü8.Ú8-Ù7,×7+Ö6*Õ7+Õ7+Ô8+Ô8+Ò9+Ò9+Ò9+Ò9+¶=,¶=,®9'¦1¨3!±<*¿=-½;+È5+É6,Ó;0ÛC8ÜJ=ØF9Å<*¹0àMEÛH@ÔD;ÔD;ÚE>ÜG@â@;ß=8å33é77ï98î87ä84â62Ü;3Þ=5ÌL?ÆF9¼>0´6(¯6%²9(¹>.½B2Â:,Æ>0×=5Ú@8è96è96ñ56ò67å6;å6;ä79ä79â96â96â92â92Þ5.ß6/à70â92å95å95å97å97é;=ç9;Ý;9Ý;9ÐC<ÎA:»C5·?1¢=+¡<*š>)˜<'”<(’:&‘9%‘9%‡F2�@,‰,$�0(ÀJJännámpÊVY­RO¬QN™YO˜XN’\P”^R™_Tš`ULa 9N
2IMdv’H–²h »t »t˜°rƒ›]k~QZm@WcKQ]EHNB<B6;<,@A1<@2:>0:B3:B3>O/RcCZv<f‚Hm‘Gn’Hi’Fi’Fh’Jg‘Ii�Rk‘Tl“Pl“Pj”Jj”Jj—Fl™HnžHnžHo¡Jq£Ls¥Pu§RvªXx¬Zx¤im™^bŽSVO?VO?VO?VO?VO?VO?VO?VO?WP@WP@XQAYRB[TD\UE]VF]VFc\Jb[IaYNbZOdZXh^\lagodjpdnrfpthrthrqfnlaif[aaV\[[SYYQWWMUUKWSHXTI[UG\VHbZGc[HeZDf[Ei]Ej^Fl^Dm_E_cLgkT{ZIwVE SI¦YO®RE°TG§K6£G2¨=)¨=)Ã6,Ç:0ß12Þ01Û7-Û7-Û7-Û7-Ú8-Ù7,×7+Ö6*Õ7+Õ7+Ô8+Ô8+Ò9+Ò9+Ò9+Ò9+·>-¶=,­8&¦1§2 ¯:(¾<,½;+É6,Ç4*Ì4)Ò:/ÙG:àNAÛR@ÙP>ÚG?ÔA9Í=4Í=4Ò=6Ô?8Û94Ø61é77ê88í76ì65ä84ä84ß>6á@8ÃC6¿?2·9+²4&®5$±8'¸=-¼A1Ç?1ÉA3Ø>6Ù?7ç85ç85ñ56ò67æ7<æ7<å8:å8:ã:7ã:7ã:3ã:3á81â92ã:3ã:3æ:6å95å97ä86æ8:æ8:Þ<:Þ<:ÏB;Ê=6¶>0±9+¡<* ;)™=(˜<'“;'’:&‘9%�8${:&|;'�2*š=5ÄNNÙccÑ]`¾JM©NKªOL™YOšZP”^R•_S˜^S˜^SRg&Rg&\s-y�J–²h£¿u—²k†¡Zi�C_w9Sf9H[.CO7=I18>239-:;+@A1<@2:>0:B39A2>O/QbB[w=gƒIn’Ho“Ij“Gj“Gi“Kh’Jk‘Tl’Um”Ql“Pj”Jj”Jj—Fl™Hm�Gm�Gn Iq£Lu§Rw©TvªXvªXj–[_‹PV‚G[TD[TD[TD[TD[TD[TD[TD[TDZSCZSC[TD\UE]VF_XH_XH`YId^Fc]Ec\Ic\Id\Oe]Pf]Vg^Wh_Zi`[ja\ja\h_Xd[T`WN^ULXVJVTHSQDQOBQN?RO@TQ@VSB\WD]XE_XEaZGd^Hf`Jh`IiaJ`bMgiT}[O}[OªWQ«XR»EA¹C?À=5¾;3¼8+»7*¼8)¿;,Ã>-Ä?.Ù7*Ù7*Ú8+Ú8+Ù9+Ø8*Õ8)Ô7(Ô8)Ô8)Ô8)Ô8)Ò9)Ò9)Ò9)Ò9)¸@0¶>.©:&¡2ž4¥;%³>*µ@,Ã>/¿:+Ä2%Æ4'Î</ØF9ÛN=ÝP?ß55Ü22Õ1/Ö20Ü86à<:é;<è:;ï6;í49ì25ë14å55ç77à=8ß<7¶<-³9*¯7'¬4$«6$®9'µ<+¸?.ÉD5ÊE6Ø@5Ø@5ä84ã73í55ï77ê69ê69é77é77æ93æ93å:2å:2â;2â;2á<6á<6ß<7Þ;6Ý98Ý98Ø:9Û=<ÔC>ÒA<ÂB7º:/¨:+¤6'œ=)›<(”=)“<(�;&�:%�8$�8$u8#t7"ˆ9,˜I<¶SN»XS±RN©JF›QH�SJ’]O•`R“bT“bT•_U”^TfwC{ŒX“¥q¤¶‚ª¾‰ž²}y�^Xn=?Q)@R*AP3=L/:C24=,58/58/:;+@A1;?19=/9A29A2=N.QbB\x>i…Kp”Jq•Kl•Il•Il–Nj”Lm”Qn•Ro—Pm•Nk•Ij”Hk˜GmšIm�Im�IpŸNt£Rx©Xx©Xt¤Xp T_ˆNU~DNw=`YI`YI`YI`YI`YI`YI`YI`YI^WG^WG_XH`YIaZJb[Kb[Kc\Ld^Fd^Fe^Ke^Ke]Pe]Pd[Td[Tg^Yg^Yf]Xf]Xe\Ud[TcZQbYP[YMXVJTREPNAOL=OL=QN=RO>WR?XS@[TA^WDa[Ec]Gg_Hh`IacNegR{YM€^R±^X°]WºD@´>:À=5À=5À</¾:-¼8)½9*Á<+Ä?.Ø6)Ù7*Ù7*Ú8+Ø8*Ø8*Õ8)Õ8)Ô8)Ô8)Ô8)Ô8)Ò9)Ò9)Ò9)Ò9)¹A1¶>.©:&¢3�3¢8"°;'´?+À;,¿:+Æ4'Ç5(É7*Ì:-Ë>-Ë>-Þ44Ü22Ö20Ø42Ý97á=;ê<=ê<=î5:ë27ê03ê03å55å55Ù61Ö3.°6'¯5&®6&­5%¬7%­8&²9(³:)Å@1ÇB3×?4ÙA6å95ä84í55î66ê69ê69é77é77æ93æ93å:2å:2à90à90ß:4ß:4Þ;6Ý:5Ü87Û76Ø:9Û=<ÓB=Î=8»;0³3(£5&¢4%š;'š;'“<(‘:&�:%�8#Œ7#‹6"|?*s6!ƒ4'•F9¯LG®KF©JF¬MI›QHœRI‘\N“^P’aS“bT–`V–`V|�Y˜©u¬¾Š¦¸„�£nr†QLb1,B,>1C8G*9H+9B16?.9<3:=49:*?@0:>08<.9A28@1<M-PaA^z@j†Lq•Ks—Mn—Ko˜Ln˜Pm—Op—Tp—Tp˜Qn–Ok•Ik•Il™Hn›JnžJoŸKs¢Qx§Vz«Zv§VlœPd”HZƒIR{AMv<b[Kb[Kb[Kb[Kb[Kb[Kb[Kb[KaZJb[Kb[Kc\Lc\Ld]Me^Ne^Nd^De_Ef`HgaIhbLgaKf_Lf_LibOhaNgaKgaKf`Hf`HgaGgaGa^M^[JXUDSP?OL;NK:LK9ML:ON:PO;PR=SU@VXCXZE[]H\^Ib`Qb`QpTH{_S±f`³hbÊHJÄBDÙ69Ù69×52Ô2/½<)¼;(ªG(­J+Ô7&Õ8'Ö9(Ö9(Ö9(Ö9(Ô9'Ô9'Ò9'Ò9'Ò9'Ò9'Ñ:'Ñ:'Ñ:'Ñ:'»A2·=.§<(¡6"•7–8 >%§E,±<(²=)Â8+Ä:-Ì8,Í9-Ò;0Ô=2ñ4;ï29ç27æ16á46á46à35à35è59ç48å58æ69ß;:Û76Ë3.Å-(«:(«:(ª;(ª;(¨;&¨;&ª:&ª:&·:(¼?-Î>3ÒB7à=6à=6è96è96î64î64í74í74ê93ê93æ;3æ;3Ý:1Þ;2Ú=6Û>7×?:Ö>9Ô>=Ó=<ÇD<ÇD<¿E:¸>3¦:-¡5(˜9'š;)’=)‘<(‹='Š<&‡:&…8$…7#…7#}C/o5!x3#ŽI9¡SG�OC˜PDŸWK‘YJ‘YJ‹[MŒ\N�]R‘_T™`WšaX‚Žf˜¤|›©…€Žj\kNDS62?+$1,8*.:,4<17?4<?6;>5;;199/89)>?/:>07;-8@17?0<M-O`@^z@k‡Mr–Lt˜Np™Mp™Mo™Qn˜PpšPq›QpšNn˜Lm˜Im˜Io›Lq�Nq Pt£Sx§Y{ª\y§^pžUb‘IX‡?]†LU~DPy?d]Md]Md]Md]Md]Md]Md]Md]Md]Md]Md]Me^Ne^Nf_Of_Of_Of`FgaGhbJhbJicMicMhaNhaNhaNhaNgaKf`Jf`HgaIhbHhbHc`O`]L[XGVSBQN=MJ9IH6HG5KJ6LK7KM8MO:PR=RT?SU@TVAZXIZXIiMAvZN¬a[±f`ÑOQÔRTà=@Ý:=Ú85×52À?,¾=*ªG(©F'Ó6%Ô7&Õ8'Ö9(Ö9(Ö9(Ô9'Ô9'Ò9'Ò9'Ò9'Ò9'Ñ:'Ñ:'Ñ:'Ñ:'º@1¶<-§<(£8$–8”6œ:!¥C*µ@,µ@,Ã9,À6)Æ2&É5)Ò;0ØA6ï29ð3:ë6;ì7<ç:<æ9;æ9;å8:å26æ37ç7:è8;ß;:Ø43È0+Á)$¬;)¬;)«<)«<)ª=(¨;&©9%¨8$±4"¶9'É9.Í=2Ý:3Þ;4è96é:7î64î64í74í74ê93ê93æ;3æ;3á>5á>5Ý@9Ý@9Ø@;×?:Ô>=Ó=<Ã@8À=5¶<1°6+¡5(Ÿ3&˜9'š;)�;'�;'Š<&ˆ:$…8$„7#„6"ƒ5!v<(n4 }8(”O?¦XLŸQE—OCœTH“[L’ZK‹[M‹[M�[P�^S˜_VšaXq}Ux„\n|XSa=8G*.= /<(0=)1=/1=/5=28@5>A8>A8;;177-78(=>.9=/7;-7?06>/;L,N_?^z@k‡Mr–Lt˜Np™MqšNpšRo™Qq›Qq›Qq›Oo™Mn™JošKržOu¡Rw¦Vx§Wz©[w¦XqŸVf”K[ŠBU„<_ˆNW€FR{Ae^Ne^Ne^Ne^Ne^Ne^Ne^Ne^Ne^Ne^Nf_Of_Of_Og`Pg`Pg`PibOibOhaNhaNhaNhaNhaNhaNgaKgaKgaIf`HfaDgbEgbEgbEf`Jd^H`[E\WAUS>PN9KJ6IH4EH5FI6EI8GK:FM=GN>FO>GP?OOCRRFWP@^WGƒYI‡]MµOJÅ_ZâFGÝABã77à44Ë=/É;-­D'ªA$Ð7%Ð7%Ð9&Ñ:'Ò;(Ò;(Ñ:'Ñ:'Ñ:'Ñ:'Ñ:'Ñ:'Ð;'Ð;'Ð;'Ð;'º>2¶:.¨=)¦;'’; �6‘:šC'¯B-±D/¿;,¼8)Å.%Æ/&Ò5.Ù<5ë2:î5=è>Aê@CÞDBÜB@ÖB@ÕA?Æ52É85Ë=9Ì>:¿@9¹:3«9.¨6+¢?*¢?*£@+£@+£@)¢?(£='¢<&ª7$¬9&»7*¾:-Ë8.Ï<2Ú=6ÞA:ê93ê93è:3è:3æ;1æ;1à>3à>3ÝC9ÜB8ÓE;ÑC9ÉB<Æ?9À?:¿>9§?2¤</›:*™8(’9'’9'‹=)Š<(†=*†=*ƒ<(‚;'~:'}9&~7%}6$v3"x5$…B2—TD›YK“QC�RD”YK�[M�[M�[N�\O”[R•\S˜[V™\W`hAX`9IR3=F'4@,2>*4>55?64=86?:=B;AF?EG9BD6?>*;:&78(=>.8<.6:,6>/6>/:K+N_?^z@j†Lr–Lt˜Np™MqšNq›SpšRp�Lp�Lp�Ln›Jn›Jp�Lt Sx¤W~ª_~ª_x¦^o�Ud‘N\‰F[ˆE\‰FaˆQY€IRyBg`Pg`Pg`Pg`Pg`Pg`Pg`Pg`Pf_Of_Of_Of_Og`Pg`Pg`Pg`PkdQjcPibOg`Mf_Lf_Lf_Lf_LgaKgaKgaIgaIgbEgbEgbEfaDgaKf`Jc^H`[EZXCUS>ON:LK7DG4DG4CG6DH7BI9BI9BK:BK:JJ>PPDVO?UN>oE5oE5¦@;¿YTæJKßCDä88á55Í?1Ë=/­D'§>!Ï6$Ð7%Ð9&Ñ:'Ò;(Ò;(Ñ:'Ñ:'Ñ:'Ñ:'Ñ:'Ñ:'Ð;'Ð;'Ð;'Ð;'¹=1¶:.©>*§<(“<!Œ5�9™B&¤7"«>)Á=.Â>/Í6-Î7.Ø;4ÞA:ó:Bõ<DìBEê@CØ><Ð64Å1/Á-+¾-*Â1.Ä62Ä62¶70²3,¨6+¨6+ =(¡>)¢?*¢?*¢?(¢?(¤>(¤>(¬9&­:'¸4'¸4'Å2(Ë8.Ù<5ßB;ë:4ë:4é;4é;4ç<2ç<2á?4á?4ÜB8ÛA7ÑC9Î@6Ä=7À93¹83·61 8+œ4'•4$–5%’9'’9'‰;'†8$†=*…<)ƒ<(�:&~:'|8%}6$|5#v3"}:)‹H8•RB’PBŒJ<�RDš_Q�YKŽZL‘]P’^Q–]T•\S–YT•XSV^7EM&5>8A";G3:F23=4-7./835>9@E>EJCGI;AC5>=);:&67'<=-8<.6:,6>/6>/:K+M^>]y?j†Lq•Kt˜Np™MqšNq›SpšRp�Lp�LoœKn›Jn›JqžMv¢Uz¦Y‚®c«`v¤\h–NZ‡DVƒ@\‰Fc�McŠSY€IRyBhaQhaQhaQhaQhaQhaQhaQhaQg`Pg`Pg`Pg`Pg`Pg`Pg`Pg`Ph_Zh_Zh_Xh_Xh_Vh_Vh`Sh`ShaOhaOhaOhaOhaNhaNhaNhaNg_Hf^Ge_Id^Ha\H\WCRQ=ML8GJ7DG4>E5=D4=F5>G6<G7;F6EL<FM=>H0CM5XS6RM0ƒI5¬r^àe]ÕZRçECãA?á?:Ü:5Ë;0Ê:/Ç4$È5%Ê7'Ê7'Ë8(Ì9)Î;+Ð=-Ï<,Ï<,Ï<,Ï<,Ñ;,Ñ;,Ñ;,Ñ;,·2)º5,®=-©8(—8"“4•7™;"°@,®>*¼4(½5)Í4,Õ<4âC=çHBãEFåGHÔFBÈ:6°:.°:.¤=.Ÿ8)¡=. <-š=.˜;,�=-�=-ˆ?,ˆ?,•>+•>+–?+—@,—@,˜A-œA,œA,Ÿ<)Ÿ<)§8'¨9(³7+·;/Á=1Ã?3Ú=4Û>5ÝA5ßC7ÜD7ÚB5ÔB5ÓA4ÓI<ÐF9ÁC7º<0ª8.¦4* 3,Ÿ2+�=-�<,‹<+‰:)†;(†;(�=*�=*|;'{:&x9'x9'x9'x9'u6%t5$z, �?3—OA•M?ŒM>‘RC‘VH�RD™ZQ™ZQ’SL’SL™ZU–WR“VQža\QZ-HQ$AK&AK&<I/5B(1<+3>-1<,EP@X_MU\JLP7CG.@A"<=9:*=>.9=/9=/8@14<-8I)PaA`|Bk‡Mq•Kr–Ln—Kp™MpšRpšRnžHoŸIoŸKoŸKpŸOs¢Rz¦[~ª_‚­fw¢[i“Q_‰G\†F`ŠJdŽNg‘QcŠUZ�LSzEhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQi`[i`[i`Yi`Yi`Wi`WiaTiaTibPibPibPibPibOibOibOibOh`Ig_Hf`Je_Ic^J^YETS?ON:HK8EH5?F6>E5=F5>G6<G7;F6@G7>E5=G/CM5NI,ID'†L8±wc×\TÛ`XóQOèFDá?:á?:ÒB7Ê:/Í:*Î;+Ï<,Ï<,Î;+Í:*Í:*Í:*Î;+Î;+Î;+Í:*Ï9*Ï9*Ï9*Ï9*Å@7Å@7¶E5®=-˜9#’3’4”6¥5!ª:&Ã;/ÌD8ÞE=ãJBéJDéJDßABàBCÏA=Ä62¯9-°:.¥>/¢;,¢>/ <-š=.˜;,�=-�<,‡>+‡>+‘:'’;(’;'“<(”=)•>*™>)™>)¢?,¡>+¨9(§8'°4(³7+»7+½9-Ó6-Õ8/Ø<0Û?3ÙA4ÚB5ÕC6ÔB5Ê@3Ç=0»=1µ7+¨6,¥3) 3, 3,�<,Ž;+Š;*‰:)†;(…:'�=*�=*|;'{:&y:(x9'x9'w8&u6%t5$€2&‘C7—OA“K=ŠK<�PA�TF�RD‘RI–WN”UN–WP›\W•VQ�RM—ZUT]0Zc6^hC[e@GT:6C)4?.<G6LWGWbR^eST[IIM4AE,?@!;<89)<=-8<.9=/8@14<-9J*PaA`|BlˆNr–Ls—Mo˜Lp™MpšRpšRp JoŸInžJnžJq Pu¤T|¨]~ª_{¦_ošSa‹I\†D_‰Ic�MdŽNbŒLb‰TY€KRyDhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQi_^i_^i_]i_]i`Yi`YiaViaViaTiaTiaTiaTiaTiaTiaViaVjbMiaLh`Kh`Ke`La\HYVCTQ>KL:HI7CF5AD3?F6?F6>E5=D4:H14B+?G/DL4H=!NC'ŒXB«wa¹SEÎhZòc_åVRèEFéFGèBDÝ79Ç9+Ç9+È:,È:,È9+Ç8*Æ7)Å6(Ê8+Ê8+Ê8+Ê8+Ë7+Ë7+Ë7+Ë7+Á2.Â3/¶6-±1(¡0 Ÿ.£/ ¦2#Á=1ÇC7ÙD=àKDéLGçJEåFBáB>ÐE@ÎC>ºD:±;1›@.�B0“H3’G2ŽG3ŒE1ˆE2…B/~D0|B.wE.vD-‡:*‡:*ˆ;+ˆ;+‰<*Š=+Š=+‹>,“@.’?-•<,•<,›:*�<,¢<.£=/µ;0¶<1º>2¼@4ÀB4ÀB4¾C4¾C4°<-®:+¥;-¢8*—9-–8,’:.’:.‡>-†=,„=+‚;)€;+:*}<*~=+z;*z;*x:+w9*s8(r7'p7&p7&‰;1”F<’NA�I<‡I<ŒNA’RH‘QG“LF™RL˜SNšUP—^W“ZS‡\S“h_v„S}‹Z|Œ_iyLIY25E;J)L[:crUapS[fHLW9CJ+>E&=@!:=78(;<,7;-9=/9A25=.9J*O`@`|BlˆNs—Mt˜Np™Mp™MpšRpšRq£Ln Il�KnŸMt£Uy¨Zz¨`z¨`o™WdŽL[„B[„BcŒJh‘Od�K^‡Eb‡SY~JRwChaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQhaQi_^i_^i_]i_]i`Yi`YiaViaViaTiaTiaTiaTiaTiaTiaViaVkcNjbMjbMjbMgbNd_K]ZGYVCNO=KL:FI8CF5?F6?F6=D4<C3;I26D-BJ2BJ2I>"_T8˜dN™eO¤>0½WIï`\ï`\òOPìIJêDFä>@È:,Æ8*Å7)Å7)Ç8*É:,Ê;-Ê;-Ð>1Ð>1Ñ?2Ñ?2Ò>2Ò>2Ó?3Ó?3ÐA=ÑB>ÇG>ÅE<ºI9¼K;ÃO@ÇSD×SG×SGáLEàKDãFAáD?àA=Þ?;ÑFAÌA<·A7¯9/›@.œA/“H3“H3�F2ŒE1‡D1„A.|B.{A-uC,uC,…8(…8(†9)†9)‡:(‡:(‡:(ˆ;)Œ9'Œ9'�7'‘8(™8(›:* :,¢<.°6+±7,³7+´8,¶8*µ7)²7(±6'©5&¨4% 6(ž4&”6*”6*‘9-‘9-…<+„;*‚;)�:(~9)~9)|;)|;)z;*z;*x:+v8)r7'q6&p7&q8'�A7•G=�K>‹G:†H;�OB”TJ•UK–OI›TN—RM—RM•\U•\U‘f]¢wn‹™h…“br‚UYi<BR+>N'O^=crQdsVZiLNY;AL.>E&=D%=@!;>78(:;+6:,8<.:B36>/9J*N_?_{Ak‡Mr–Lt˜Np™MqšNq›SpšRp¢KmŸHmžLr£Qy¨Z{ª\w¥]r Xa‹I\†D[„B^‡EeŽLh‘Od�K_ˆFa†RX}IQvBhaQhaQhaQhaQhaQhaQhaQhaQibRibRibRibRibRibRibRibRja\ja\jaXjaXjbUjbUjcQjcQjcPjcPjcQjcQjbUjbUjbWjbWldQkcPkbQkbQjcQhaOb]J_ZGURARO>LK9IH6EF4CD2AB2?@09L.7J,SC,P@)t6)–XK¹XQ�<5¥0)µ@9ØWRílgôfeäVUãJOçNSÕK@ÐF;Ê@5È>3Í@6ÑD:ÔF<ÕG=ÚJAÚJAÚJAÛKBÝJBÞKCÞKCÞKCìRRëQQãTPàQMÚQKÚQKßRKáTMäIEãHDçCBåA@ãA>äB?àE@áFAÕJGÍB?µ@7¯:1™B1™B1ŽG3�H4ˆE4‡D3†A2„?0~?0}>/y>0y>0�9-�9-€:.€:.;.€</€=-€=-~;+~;+‚:+ƒ;,…<-†=.Š>.‹?/”<0”<0˜</˜</œ;+œ;+ž:*�9)š9(™8'’9)‘8(Š;*‰:)…<-…<-�<-€;,:+}8)|9){8(z8*z8*x9*y:+v;-t9+p6(o5'p8+r:-‹I;ŒJ<…L;ƒJ9‚M=‰TD”UL–WN—PL›TP•TP•TPŠdYˆbWylY†yfgvMZi@HY-AR&GZ-Wj=bxJg}OXjBK]5=L-7F':D,;E-;?08<-78(9:*59+8<.;C47?08I)L]=\x>i…Kq•Kt˜Np™MqšNq›SpšRmžLmžLp¡Pv§V|ªay§^p�Zg”QYƒC[…E`‰GcŒJgŽKgŽKe�GdŒF`…RW|IPuBhaQhaQhaQhaQhaQhaQhaQhaQibRibRibRibRibRibRibRibRja\ja\jaXjaXjbUjbUjcQjcQjcPjcPjcQjcQjbUjbUjbWjbWldQkcPkbQkbQkdRibPe`Mc^KZWFVSBPO=LK9GH6EF4BC3@A17J,8K-QA*SC,ƒE8�_R±PI’1*±<5µ@9ÓRMñpkÿtsóedðW\ô[`àVKÙODÑG<ÍC8ÑD:ÔG=ÖH>ÖH>ÙI@ÙI@ÚJAÚJAÝJBÞKCÞKCßLDáGGàFFÙJF×HDÓJDÒICÕHAÕHAÝB>ÞC?æBAçCBåC@äB?ÝB=Û@;ÖKHÌA>´?6°;2›D3˜A0‹D0ŒE1†C2…B1„?0ƒ>/}>/}>/z?1z?1‚:.‚:.€:.€:.€</€</€=-€=-€=-€=-‚:+‚:+‚9*‚9*ƒ7'ƒ7'�7+�8,”8+•9,›:*œ;+ <, <,™8'™8'’9)‘8(‰:)ˆ9(ƒ:+ƒ:+:+~9*}8)|7(z7'z7'y7)y7)u6'w8)u:,s8*o5'n4&q9,u=0�K=‹I;„K:„K:…P@‹VF”UL•VM”MIšSO–UQšYUŒf[�[PdWDgZGBQ(?N%>O#K\0^qDk~QcyKVl>M_7@R*6E&5D%<F.<F.:>/6:+89):;+59+8<.;C47?06G'HY9Ws9e�Gn’Hr–Lp™Mr›OrœTq›Sl�KnŸMs¤Sx©Xz¨_s¡Xg”Q]ŠGX‚B\†FcŒJf�Mf�JeŒIe�GfŽH_„QV{HOtAhaQhaQhaQhaQhaQhaQhaQhaQibRibRibRibRibRibRibRibRjcSjcSjcPjcPjdLjdLjdJjdJjdJjdJjdLjdLjcQjcQjbUjbUlcTkbSjaRkbSlcTkbShaQf_OaZH]VDVQ>QL9NI6KF3HC0FA.6I)7J*a:)pI8´FGµGHÂ0:À.8Ì@CÈ<?ÇKIÞb`íspêpmãcdÝ]^ÚQIÔKCÎE=ÌC;ÏD=ÐE>ÐC<Í@9ÑB<ÒC=ÓB=ÔC>×C?ØD@ØD@ÙEAß>Cà?DÝCCßEEáIHáIHæFFåEEê@Cê@Cî@Bî@BåA@â>=Ø?9Õ<6ÚFFÎ::¶93µ82¢B4�=/Ž?0Ž?0Š</‰;.‰90ˆ8/‡81‡81†93‡:4:3:3~;2~;2|<2|<2{=0{=0y>0y>0y>.x=-w<,v;+w:(v9'};/~<0‚:,ƒ;-Š;,Œ=.’<+“=,�9(�9(Š;*Š;*‚>+�=*~=+}<*€8*7)~6(}5'z4(z4(y5*y5*s3'u5)t8-r6+m5(m5(q<.vA3‡R@ƒN<|O<R?„VF†XHŽTIŽTI“NI—RM�VO“ZSjYoZIFP5AK0?N/JY:VhB_qKd{MbyKUm=E]-AV-5J!/?"3C&;F89D67;:598:;+;<,59+8<.;C46>/5F&EV6So5a}Ck�Eq•Ko˜Lr›OrœTq›Sl No£Qr¥Xt§Zq Xi˜P_ŒKY†E]‡G_‰Ib‹Id�KfŽHe�GdŒCc‹B^ƒPUzGNs@haQhaQhaQhaQhaQhaQhaQhaQjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdQkdQkeMkeMkeKkeKkeKkeKkeMkeMkdRkdRkcVkcVlcTkbSjaRkbSlcTlcTibRg`Pd]K_XFXS@SN;OJ7LG4ID1FA.7J*7J*d=,XGÁST«=>¹'1ÕCMÉ=@Ä8;·;9ÂFDÖ\YàfcÚZ[ÈHIÒIAÎE=ËB:ÍD<ÑF?ÒG@ÐC<Ë>7ÑB<ÒC=ÓB=ÔC>×C?ØD@ÙEAÙEAß>Cß>CÝCCßEEßGFÞFEà@@Þ>>ê@Cé?Bê<>é;=â>=ã?>ÞE?ßF@ÖBBË77³60´71¢B4œ<.�>/�>/ˆ:-ˆ:-ˆ8/ˆ8/‡81‡81‡:4‡:4:3:3~;2~;2|<2{;1z</z</u:,u:,v;+w<,w<,w<,y<*y<*};/|:.€8*7)„5&…6'‹5$‹5$�7&�7&ˆ9(‰:)�=*�=*~=+}<*7)~6(}5'|4&y3'y3'x4)x4)q1%t4(s7,r6+m5(n6)r=/xC5‰TB…P>~Q>�TA†XH†XHŒRGŠPE—RM—RMŠQJ�TMzeTlWFCM2>H-HW8ZiJgySewQXoAQh:Ld4H`08M$-B*:1A$:E79D68<;7;:;<,;<,59+8<.;C46>/3D$CT4Pl2_{AjŽDp”Jo˜Lr›OrœTrœTn¢Pq¥Ss¦Yp£VkšRb‘I[ˆGW„CbŒL_‰I_ˆFaŠHfŽHg�Ic‹B^†=^ƒPTyFMr?h_Pi`Qi`Qi`QjaRjaRjaRkbSjaRjaRjaRjaRjaRjaRjaRjaRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTn]Mq`PqdTqdTlcTi`QgaShbTebQebQa_JYWBVR9TP7OI/HB(NE(I@#dH2ˆlV²zm¬tg³[W­UQ´DCµEDÃCDÆFGÎFFÑIIÔJHÕKIÏCDÏCDÔDCÕEDÔE?ÑB<ÍE9ÎF:ÃE6ÄF7ÇG<ÈH=ÓCBÓCBÛ>EÛ>EàBCàBCáCBáCBàECàECáFBáFBáC@áC@áB>àA=â@=â@=â@=â@=ÌF=Å?6±;1¬6,™=0˜</†A2„?0|A3|A3}?4|>3„:1„:1Š71Š71…<-„;,ƒ;,ƒ;,‚:+‚:+€;,€;,:+:+~:-~:-~:-~:-};/};/|8-|8-|8-|8-|:.|:.|:.|:.y9-y9-y9-y9-w9,w9,w9,v8+u9.t8-t8-s7,r6+q5*p6*p6*j2%o7*m5(k3&m8*j5'n9+~I;€TI€TI‡PI‡PI‘MJ’NK”OJ•PK�TL�RJ~_K�bNdcEVU7GY3M_9g�T_yLWpFQj@Qf?Mb;HW8BQ28B)4>%38$5:&=>.@A1A?2?=00;-0;-39+5;-9=,;?.=B+>C,S_9coIn‡PuŽWp•Ps˜Sq�PržQv�^~¥f~¥fq˜YcŠK^…F^…F]„E_†GaˆId‹LeŒMeŒMb‰J_†G]„E^€MSuBKm;h_Pi`Qi`Qi`QjaRjaRjaRkbSjaRjaRjaRjaRjaRjaRjaRjaRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTraQraQobRqdTpgXpgXlfXicUgdSkhWmkVjhShdKc_F[U;RL2SJ-H?"X<&lP:‹SF†NA–>:™A=ÀPOÀPOÊJKÊJKÎFFÎFFÐFDÐFDÚNOÏCDÉ98Ê:9ÑB<ÔE?ËC7Æ>2ÄF7ÅG8ÇG<ÉI>ÔDCÕEDÝ@GÝ@GàBCàBCáCBáCBàECàECàEAáFBáC@áC@áB>àA=á?<á?<â@=â@=ÍG>ÈB9¶@6¯9/™=0–:-„?0‚=.|A3|A3}?4|>3„:1„:1Š71Š71„;,„;,ƒ;,ƒ;,‚:+‚:+€;,:+:+:+~:-~:-~:-~:-};/};/|8-|8-|8-|8-{9-{9-|:.|:.z:.z:.y9-y9-x:-w9,w9,w9,t8-t8-s7,s7,r6+q5*p6*o5)j2%o7*m5(k3&l7)j5'n9+~I;€TI€TI‡PIˆQJ’NK“OL•PK–QL‰NFŠOG|]I€aMmlNkjLasMfxR[uHUoBMf<Ib8J_8H]6FU6BQ2;E,6@'5:&6;'<=-?@0@>1></1<.1<.4:,5;-9=,;?.<A*=B+P\6`lFjƒLr‹Tn“Nr—RržQržQwž_{¢cxŸ`k’S`‡H]„E]„E\ƒD_†GaˆIcŠKd‹Ld‹Lb‰J_†G]„E\~KQs@Ik9g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTrbSo_Pj^Pl`RmeXogZieYeaUecVom`zwh~{lzgzubojVe`LMW<?I.=?'CE-NC/H=)V:,_C5d<0f>2qA5xH<„NB�WK–^QšbUµ`Y­XQ²IE°GCÇCAËGEÓB?Ï>;ÑA9ÒB:ÒC;ÔE=ÖE@×FAÚDCÚDCÞBCÞBCÞDDÞDDÝEBÞFCÞFCÞFCÛF@ÚE?ÙD=ØC<ØC<ØC<ÙD=ÙD=ÓF=ÑD;¿C9·;1�;.—5(„;,„;,~@1~@1~?0}>/ƒ;/ƒ;/ˆ8/ˆ8/ƒ;,ƒ;,ƒ;,‚:+‚:+‚:+:+:+~:-~:-~:-~:-};/};/};/};/{8/{8/{8/{8/{8/{8/y9/y9/y;0y;0y;0x:/v:/v:/u9.u9.t8.s7-r8-q7,p6+p6+n5*n5*j1&o6+j4(i3'l6*i3'm7+}G;‡QG‡QG‹PHŒQI‘RK’SL�UM‘VN…YN‡[PxaOv_MkkQqqWhwXfuVPdAJ^;DX5BV3DU5FW7ES9DR8?I1:D,6>)5=(9=,:>-;=/9;-3=42<36<26<29=.:>/<A*<A*LX4[gCe~Gn‡Pl‘Lr—RsŸPt Qz¡bwž_n•Vb‰J\ƒD[‚C\ƒDZ�B_†G`‡Hb‰JcŠKcŠKaˆI^…F]„EX}JMr?Di6g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTscTqaRl`Rk_QiaTkcVieYieYigZwuhˆ…v��~•�}’�zˆƒo~ye[eJKU:DF.DF.LA-G<(T8*[?1Z2&Z2&_/#`0$g1%j4(p8+s;.™D=£NG¹PL·NJÅA?ÅA?ÔC@ØGDÑA9ÒB:ÓD<ÔE=×FA×FAÛEDÛEDÞBCÞBCÞDDÞDDÝEBÝEBÝEBÝEBÛF@ÚE?ÙD=ØC<ØC<ØC<ØC<ØC<ÒE<ÒE<ÃG=¼@6 >1˜6)…<-…<-~@1~@1~?0}>/ƒ;/ƒ;/ˆ8/ˆ8/ƒ;,ƒ;,‚:+‚:+‚:+�9*:+:+~:-~:-~:-~:-};/};/};/};/{8/{8/{8/{8/{8/z7.x8.x8.y;0y;0y;0x:/v:/u9.u9.u9.s7-s7-q7,p6+p6+o5*m4)m4)j1&o6+j4(h2&k5)h2&k5){E9†PF‡QGŒQI�RJ’SL“TM‘VN‘VN‰]RŠ^Sv_MnWE``FggMYhIN]>CW4>R/:N+9M*>O/AR2ES9FT:BL4>H0:B-7?*9=,9=,9;-79+3=43=46<26<29=.:>/;@);@)HT0Wc?azCk„Ml‘Lt™Tu¡Rv¢Sz¡bq˜Yd‹L[‚CY€A[‚C\ƒDZ�B^…F_†GaˆIaˆIaˆI_†G]„E\ƒDW|ILq>Ch5g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTrfXsgYph]ldYfbYie\qngzwp~}x�Œ‡Ÿœ•§¤�ª¦›¨¤™¢œ�š”ˆx{hcfSMR>EJ6>J6;G3:G3=J6BQ<?N9;H66C16>/4<-7;-7;-K8)^K<ˆQJ�VO°HI³KLÎCHÎCHØ?AØ?AÚB?ÛC@ÜD?ÜD?ÜD?ÜD?ÞDDÞDDÜDCÜDCÜDCÜDCÙEAÙEA×G?ÖF>ÔG>ÓF=ÐF;ÐF;ÐH<ÐH<Ö?8ØA:ÊF:ÅA5ª@3¢8+Ž=,�<+‚?.‚?.~?.}>-�<,�<,†:,†:,‚:+‚:+‚:+‚:+:+:+~9*~9*~:-~:-};/};/};/};/};/};/z:0y9/y9/y9/w9.w9.v8-v8-v:/v:/v:/u9.t:.t:.s9-s9-q7,q7,p6+p6+n5*m4)l3(l3(g1%l6*h4'f2%i5(f2%i5(yE8�LD�ME�PI‘RK‹WLŒXMƒ[Oƒ[Os\Lt]Mc\I\UBUZDY^HLXB?K5:F28D05A-5A-8D0<H4@L8CO;@L8=I59E17C/5A-3?+2>*0<(3<72;65:35:38;09<1:?);@*FR.T`<^vBj‚Nn“NwœWw£Tw£Txžak‘T]ƒFW}@YB\‚E\‚E[�D^„G^„G_…H_…H_…H^„G]ƒF\‚EVzJKo?Bf6g`PhaQhaQhaQibRibRibRjcSibRibRibRibRibRibRibRibRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdTmaSpdVnf[kcXhd[qmdƒ€y’�ˆ¤£ž²±¬ÃÀ¹ÈžÈĹÄÀµ¿¹­¸²¦œŸŒ�„qafRMR>?K7:F27D08E17F16E07D27D2;C4=E6BF8CG9D1"M:+p92|E>±IJÀXY×LQÍBGÛBDÜCEÜDAÝEBÝE@ÝE@ÜD?ÛC>ÞDDÞDDÜDCÜDCÜDCÛCBØD@ØD@×G?ÖF>ÓF=ÒE<ÐF;ÐF;ÏG;ÐH<Õ>7×@9ÊF:ÇC7®D7§=0�?.�<+‚?.‚?.~?.}>-�<,�<,†:,†:,‚:+‚:+‚:+�9*:+~9*~9*~9*~:-~:-};/};/};/};/};/};/z:0y9/y9/x8.w9.v8-u7,u7,u9.u9.t8-t8-s9-r8,r8,r8,p6+p6+p6+o5*m4)l3(l3(k2'g1%l6*h4'f2%h4'e1$h4'xD7�LD‘NF�QJ‘RK‹WL‹WL�YM€XLs\LpYI_XEYR?OT>LQ;>J44@*1=)/;'.:&.:&0<(3?+7C/9E1=I5<H4<H4:F28D05A-2>*0<(0940943814927:/8;09>(:?)BN*Q]9\t@j‚No”Ox�Xw£Tv¢Sp–Yc‰LW}@V|?YB[�D[�D[�D]ƒF]ƒF^„G^„G]ƒF\‚E\‚E[�DUyIJn>@d4f`PgaQgaQgaQhbRhbRhbRicShbRhbRhbRhbRhbRhbRhbRhbRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdThbVicWhd[kg^qpk…„��›¯¯­ÀÂÁÐÒÑààÞååãæâßáÝÚÛÖÒÔÏËÔ¸­½¡–˜…v}j[_ZFPK7>F/<D-=J6<I5=E6;C4<=5:;3;74:636A37B4K81Q>7†IH¡dcÄad¾[^ÊRSÊRSÕNKÔMJàHEÞFCåA?ä@>ÞDFÞDFÝCEÝCEÜBBÜBBÛCBÚBAÙEAØD@ÔE=ÓD<ÐF<ÐF<ÍG<ÍG<Û>7Û>7ÐC9ÎA7·F8°?1•>-�8'…>,…>,€?->,€=*€=*ƒ<*ƒ<*€;,€;,:+:+~9*~9*}9,}9,};/};/};/};/{;1{;1{;1{;1w;1w;1v:0u9/t8.s7-r8-q7,q8-q8-q8-p7,n8,n8,m7+m7+o6-o6-l6,k5+j4*j4*h4)h4)f2'k7,f4)c1&e3(b0%e3(uC8”KE•LFŽRJ�SK�YM€XLp]No\MdaP[XGNRAJN=AJ99B12:+08)25.25.06,/5+.6)/7*2:-4</9D4:E5<I7<I79H35D/1@+/>)-63-6316027158/7:19=,:>->I)MX8Zr@i�Oo”Qx�Zu¡VržSiŒT]€HUx@WzB[~F[~F[~F]€H_‚J_‚J_‚J^�I^�I]€H]€H]€HSwKGk?>b6f`PgaQgaQgaQhbRhbRhbRicShbRhbRhbRhbRhbRhbRhbRhbRjcSjcSjcSjcSjcSjcSjcSjcSkdTkdTkdTkdTkdTkdTkdTkdToi]mg[lh_tpg„ƒ~ Ÿš¼¼ºÎÎÌÍÏÎÞàßññïøøöúöóöòïðëçéäàêÎÃÜÀµ¿¬�¢�€{vb`[GGO8AI2=J6=J6?G8>F7@A9>?7?;8>:74?18C5H5.D1*o21�RQÃ`cÍjmÍUVÍUVØQN×PMâJGàHEçCAæB@ÞDFÞDFÝCEÝCEÜBBÜBBÚBAÚBAØD@ØD@ÔE=ÓD<ÏE;ÐF<ÍG<ÍG<ÞA:Ü?8ÐC9ÏB8¹H:±@2”=,Œ5$…>,…>,€?->,€=*€=*ƒ<*ƒ<*€;,:+:+:+~9*~9*}9,|8+};/};/};/};/{;1{;1{;1{;1w;1w;1v:0u9/t8.s7-q7,q7,p7,p7,p7,o6+m7+m7+l6*l6*o6-n5,l6,k5+j4*i3)h4)g3(f2'k7,e3(c1&e3(a/$e3(tB7”KE•LFŽRJ�SK�YMWKo\MmZKWTCKH7?C2?C29B109(/7(5=.47047039/17-/7*/7*08+19,6A19D4<I7>K9:I46E02A,0?*+41,52/4.16047.7:19=,:>-;F&KV6Xp>h€No”QwœYsŸTnšOb…MX{CRu=VyAZ}EY|DY|D[~F^�I^�I^�I^�I]€H]€H]€H\GQuIEi=<`4icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUkdTkdTkdTkdTee]ee]stn‘’Œ­¯¬ÃÅÂ×ÛÜéíîñõöôøùûýüþÿÿÿÿýýüúúöó÷óðûàÍóØÅÜʲƴœ§ ƒŠƒfbcDGH)EH-CF+ID1ID1I:3D5.H32M87;@<<A=B:7B:7_GC†nj½–�Ù²«ÊŽ„·{q½_W¸ZRÖNN×OOëEIìFJØFGÜJKÚDEÒ<=Ø>@àFHàDEÖ:;Ù??Ú@@×B>×B>ÓB=ÒA<ÏB9ÎA8àA;Ý>8ÏA7Í?5¹C7³=1™=.’6'‡>-‡>-�>-€=,~=)~=)<)<):+~9*}9,}9,|8+|8+{9-z8,{9-{9-y9/y9/y9/z:0z:1z:1v=2v=2u<1s:/r9.q8-n8,m7+m7+m7+l8+l8+l8+l8+j9+j9+g3(k7,o;0n:/j6+g3(d2'e3(c1&h6+b2&]-!_/#^."a3&oA4‹NI�PKƒRK‚QJu[Nw]P`ZJSM=>G4<E26C23@/2=/1<.39/39/77566457246116/05..6+.6+,8*0<.5D19H58J45G11C-.@*)2-)2--2,.3-25,58/8<+:>-5@"FQ3Vm?i€Rq•Xxœ_qœWi”OZ}GVyCSv@Sv@VyCY|FY|FX{E]€J`ƒM`ƒM\I[~H]€J_‚L]€JSqOEcA9W5icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUleUleUleUkdTee]oog…†€¤¥ŸÀ¿×ÙÖéíîöúûöúû÷ûüûýüýÿþÿþüÿþüÿüùÿûøÿíÚÿæÓîÜÄÜʲÁº�¨¡„‚ƒdijKMP5FI.FA.GB/K<5H92L76O:98=9?D@PHE[SPgc¥�‰Õ®§êüÿÒÈí±§Û}uÁc[ØPPÙQQêDHã=AÕCDÕCDÜFGßIJÞDFÖ<>×;<Ü@AØ>>Ù??ÖA=ÖA=ÓB=ÒA<ÏB9ÏB9ß@:Ý>8ÐB8Î@6ºD8´>2š>/“7(‡>-‡>-�>-€=,~=)~=)<)<):+~9*}9,}9,|8+|8+{9-z8,|:.|:.z:0z:0z:0z:0z:1z:1u<1t;0t;0r9.q8-p7,m7+l6*l6*l6*k7*k7*k7*k7*i8*i8*j6+l8-n:/n:/k7,h4)e3(d2'b0%d2'^."]-!b2&^."_1$n@3„GB‹NI‡VO„SLpVIjPCOI9C=-8A.7@-3@/2?.2=/1<.4:04:077577557246116/16//7,/7,,8*/;-3B/7F37I36H23E/1C--61,50-2,-2,03*36-6:)8<+5@"FQ3Ul>j�Sr–Yxœ_ošUe�KX{EUxBRu?Ru?UxBWzDX{EWzDZ}G]€J^�K[~HZ}G\I]€JZ}GPnLB`>5S1gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUleUleUleUleUaedw{z•™˜²¶µÉÑÓáéëñùû÷ÿÿùÿÿùÿÿûÿüúÿûþÿùÿÿúÿÿøÿÿøøÿÞôûÚéóÐÜæÃÍÙ±ºÆž›ª„“h^hCNX3CG,AE*HE4IF5I@7H?6C;9SKIia_wu™”�¹´°ÓÑÅáßÓ÷ëÝÿúìÿåÛ౧Áfc¥JGÄFJÓUYÔYRÊOHÆE@ÇFAÐFCÑGDÕCCÕCCÖ>=Ö>=Õ?>Ö@?ÕA=ÕA=ÑB<ÑB<Ù@:Ø?9ÎD:ÍC9»F=´?6œ>2–8,‰=/‰=/�>.€=-~=+~=+~=+~=+~:-}9,}9,}9,{9-{9-{9-z8,{;1{;1{;1{;1z;2z;2y:1y:1q;1q;1o;0n:/m9.l8-i7,h6+i7,i7,h8,h8,h8,h8,h8,h8,k90k90j8/j8/j8/i7.d4*b2(b3)`1'[,"_0&a5*Y-"\0%pD9yRK~WPwZRoRJYMAOC7;=/46(/:*0;+0<.1=/4<14<17:169068368349338227027008-08--8*/:,2?.5B17F38G46G46G43=20://5+,2(-0%03(37(6:+5?&FP7WmGmƒ]u—dy›hk”Z^‡MVvDTtBRr@Rr@TtBVvDVvDVvDWwE[{I\|JZzHZzH]}K[{IXxFOeN@V?2H3gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSibRjcSjcSjcSkdTkdTkdTleUleUleUleUleUleUmfVmfVmfVgkjƒ‡†£§¦º¾½ÍÕ×åíïòúüôüþùÿÿùÿÿûÿüúÿûýþøýþøÿÿ÷ÿÿø÷þÝõüÛîøÕæðÍÝéÁÎÚ²³Â—ž­‚}‡bdnIMQ6DH-KH7LI8KB9G>5OGEia_Š‚€¢š˜µ°¬ËÆÂÝÛÏèæÚýñãÿýïÿòèÿÝÓÿ¯¬éŽ‹Ù[_¹;?ÈMFÌQJÎMHÆE@ÍC@ÕKHÙGGÑ??Ô<;Õ=<Ô>=Õ?>ÕA=ÕA=ÒC=ÑB<×>8Ø?9ÏE;ÎD:»F=´?6�?3—9-‰=/‰=/�>.€=-~=+~=+~=+~=+~:-}9,}9,}9,{9-{9-{9-z8,|<2|<2{;1{;1z;2y:1y:1y:1o9/o9/m9.l8-k7,j6+h6+g5*i7,i7,h8,h8,h8,h8,h8,h8,l:1i7.g5,h6-j8/i7.d4*`0&c4*`1'].$`1'`4)V*^2'yMB�ZS|UNjME]@8H<0D8,9;-8:,-8(.9)/;-1=/4<14<169058/68368349338238127019.19.-8*.9+/<+2?.4C07F37H58I69C84>317-,2(,/$-0%15&37(3=$GQ8\rLs‰czœixšgd�SS|BRr@Qq?Pp>Pp>Qq?SsASsATtBUuCYyG[{IZzH[{I]}KYyGTtBH^G8N7*@+gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSibRjcSjcSjcSkdTkdTkdTleUkdTleUleUleUmfVmfVngWngWksu…��¢ª­·¿ÂËÕ×ãíïðüüòþþ÷ÿþ÷ÿþúÿùùÿøûÿñøüîøùé÷øèñÿçñÿçðýáëøÜéóÑÞèÆËÑ­·½™œ¢€|‚`X^DGM3CL9CL9=H::E7d]W…~x«¢�Á¸³Ê¿×ÏÌâáÜîíèö÷ïõöîúòçÿøíÿöíÿëâᬦ¥pj¤M<«TC»[K¿_OÁRG·H=ÀA:ÆG@ÈA=ÉB>ÍB=ÎC>ÎC>ÏD?ÎC<ÎC<Ò?8ÔA:ËF?ÊE>·F>°?7›>6—:2‰<2‰<2�=0€</~<.~<.~<.~<.~:-}9,|:.|:.{9-{9-y9/x8.|<2|<2z;2z;2y:1y:1x92x92k9.k9.j8-i7,i7,h6+f6*f6*f8+f8+f8+f8+d8+d8+d8+d8+j:0g7-e5+e5+h8.h8.c4*`1'a2(b3)\0%\0%\0%V*a7+~THg]TZPGHA7>7-77+88,8<.9=/19,2:-4</4</7:/69.68-57,2:/2:/2:-2:-19,19,/:,/:,-8(-8(.9)0;+2?.5B18E4:G6;F57B14;+/6&.2$.2$04%15&09$HQ<avWx�nzšqs“j[‚SIpANn=Nn=Oo>Oo>Oo>Oo>Pp?Qq@TtCXxGZzI[{J]}L]}LWwFOo>DQG5B8(5+gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSibRjcSjcSjcSkdTkdTkdTleUkdTleUleUmfVmfVngWohXohXgoqz‚„“›ž¬´·ÃÍÏÛåçë÷÷òþþóþúõÿüùÿøùÿøúþð÷ûíõöæòóãïÿåïÿåðýáîûßîøÖåïÍÓÙµÀÆ¢¬²�Š�nbhNLR8CL9EN;EPBGRD€ys ™“ż·×ÎÉÜÔÑäÜÙíìçø÷òîïçúûóÿþóÿþóÿóêÿðçÿàÚÿÕÏä�|ºcR¨H8³SCÃTI´E:¼=6ÊKDÈA=ÉB>ÍB=ÎC>ÎC>ÎC>ÍB;ÍB;Ó@9ÔA:ËF?ÈC<³B:­<4š=5˜;3‰<2‰<2�=0€</~<.~<.~<.~<.~:-}9,|:.|:.{9-{9-y9/x8.{;1{;1y:1y:1x90w8/w81w81j8-j8-j8-i7,i7,h6+g7+g7+f8+f8+f8+f8+d8+d8+d8+d8+h8.f6,e5+f6,g7-g7-f7-d5+a2(d5+^2'[/$`4)`4)g=1{QERH?G=4;4*;4*99-88,48*37)3;.3;.3;.3;.58-47,46+46+19.19.19,08+08+08+.9+.9+.9)-8(-8(.9)/<+2?.4A06C2:E48C28?/4;+37)15'04%/3$/8#JS>cxYv‹lr’ig‡^QxIAh9Jj9Ll;Mm<Mm<Ll;Ll;Mm<Nn=TtCXxGZzI[{J]}L\|KSsBIi8<I?.;1"/%feSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSibRjcSjcSjcSkdTkdTkdTleUkdTkdTleUmfVngWohXohXpiYbkpluzƒŽ’¡¬°½ÈÌÒÝáâððíûûòý÷óþøøÿõùÿöûÿîùýìøúåöøãøôëùõìýóçýóçÿñáúéÙêÕÀØîÀ¯›¡�|tm[\UCMPESVKXc[bme•™Š¯³¤ÕÎÆæß×õáâûçèÿðóÿô÷ÿýüþúùòüóñûòóÿïøÿôûÿóûÿóÿãÉé­“·pZ YC¬SA­TB¸L@¹MAÂG?ÂG?ÆF=ÇG>ÈE=ÈE=ÇD<ÆC;ÐC<ÑD=ÇF@ÂA;¯@9¨92–<4–<4‰<4‰<4�=4€<3~<0~<0~<0~<0};/|:.|:.|:.{9-{9-y9/x8.z:1z:1x90x90w81v70s7/s7/j8-j8-i9-h8,h8,h8,g9,f8+f8+f8+d8+d8+d8+d8+d8+d8+f5.g6/h70h70g80g80i:2k<4d8/i=4b6-`4+nC:vKBsH?tI@7>6180/7,4<16<228.14)25*58-47,46)46)35(35(46)46)-8(-8(-8(-8(-8(-8(-8(-8(.9).9)-8(-8(.9)/:*0;+1<,7C-7C-:A/9@.9=.59*15&.2#09&KTAbv[oƒhd„_YyTGmD=c:If6Li9Nk;Nk;Li9Kh8Mj:Ol<UrBXuE[xH\yI^{K\yIQn>Eb27@;*3. )$feSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSfeSibRjcSjcSjcSkdTkdTkdTleUkdTkdTleUmfVngWohXpiYpiYgpuluz€‹�£®²ÁÌÐÓÞâàîîîüüòý÷óþø÷ÿôøÿõüÿïûÿîýÿêüþéû÷îýùðÿ÷ëÿøìÿ÷çÿñáòÝÈáÌ·Á°œ¥”€{tbd]KVYN_bWkvn{†~¨¬�¼À±ÜÕÍìåÝþêëÿðñÿòõÿðóüø÷üø÷õÿöøÿùôÿððýìõþíûÿóÿðÖÿðÖÿůÃ|f¦M;«R@¼PD°D8ÂG?ÂG?ÇG>ÇG>ÈE=ÇD<ÆC;ÅB:ÑD=ÒE>ÇF@À?9¬=6¦70•;3•;3‰<4‰<4�=4€<3~<0~<0~<0~<0};/|:.|:.|:.{9-{9-y9/x8.z:1y90x90w8/v70u6/r6.r6.j8-j8-i9-i9-i9-h8,g9,g9,e7*e7*c7*c7*c7*c7*c7*c7*f5.i81k:3k:3h91i:2l=5pA9l@7pD;g;2g;2~SJ‹`W~SJpE<5<42912:/5=24:0.4*25*8;025*25*13&13&13&24'46)57*,7',7',7',7',7'-8(-8(-8(.9).9)-8(-8(,7'-8(-8(-8(4@*6B,<C1=D2=A28<-26'.2#2;(LUB_sXg{`YyTNnIAg><b9Gd4Kh8Mj:Mj:Kh8Jg7Kh8Nk;TqAWtDYvF[xH]zJZwGMj:@]-3<7'0+(#heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgfThdYhdYkd\kd\ldWmeXe`JjeOmlXmlXef`de_nqx~�ˆ‡‘’•Ÿ «³µÁÉËØÜÝéíîöööùùùþýùÿþúþÿùýþøúýôøûò÷ýñøþò÷úñ÷úñ÷ùëøúìûýåõ÷ßãçÆÑÕ´·½›•›ylwY\gI_mVqh�’‚�žŽ·¸³ËÌÇâãÞîïêôõðøùôùúõøùôþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúèÿÿíÿÿÿÿýóòðÕ¨¢’e_œD:ªRHÄK@ÀG<ÇC7ËG;ÎL?ÎL?ÈJ>ÆH<ÁJ2ÁJ2µL6±H2žI4™D/‰F5ˆE4D4~C3€@4~>2„:/ƒ9.‡7.‡7.}:1}:1|90{8/x8.w7-v6,v6,u7,v8-v8-w9.t8-r6+p4)o3(h9/g8.g8.g8.f7-f7-g7-f6,f6,f6,h8.h8.i7.h6-f4+e3*h1,k4/l7/m80i9+j:,f?.hA0`D/gK6^L8r`Lun[‚{hƒ‚pNM;3;.3;.2:-2:-19,19,19,19,08+/7*/7*/7*.6).6).6)-5(.5-.5-.5-.5-.5-.5-.5-.5-.5-.5--4,-4,-4,-4,.5-.5-/;12>45A77C97C56B43@/2?.2?+DQ=WeN\jSSaGGU;DR8FT:Hd>Ie?Jf>Ie=Hd;Ie<Kh<Mj>SoFWsJZvN\xP\wVVqPFaB7R3&.1$,/!),heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgfThdYhdYkd\kd\ldWmeXidNlgQlkWlkWfgaklf{~…�’™§±²¸ÂÃÎÖØßçéïóôøüýÿÿÿÿÿÿýüøþýùþÿùýþøúýôøûòöüðöüðöùðöùð÷ùë÷ùëøúâñóÛßãÂÌЯ³¹—‘—uhsUXcE^lUt‚k‰šŠ˜©™¿À»ÒÓÎçèãòóî÷øóúûöûü÷úûöþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúíÿÿìÿÿÿþüÿÿýÿéãíÀºÐxn¢J@¾E:ÀG<ÌH<ÌH<ÆD7À>1¾@4ÃE9ÀI1ÁJ2·N8³J4žI4˜C.ŠG6ŠG6€E5D4€@4~>2„:/„:/‡7.‡7.}:1}:1|90{8/x8.w7-w7-v6,u7,u7,v8-v8-s7,r6+p4)o3(h9/g8.g8.g8.f7-f7-g7-g7-c3)d4*e5+g7-i7.j8/j8/j8/l50l50m80m80k;-l<.hA0iB1dH3lP;dR>weQ}vc‹„q„ƒqLK92:-2:-2:-19,19,08+08+08+08+/7*/7*/7*.6).6).6)-5(-4,-4,-4,-4,-4,-4,-4,-4,.5-.5--4,,3+,3+-4,.5-.5-,8..:01=34@65A36B45B14A05B.BO;P^GQ_HHV<AO5BP6FT:Gc=Ie?Jf>Ie=Gc:Fb9Gd8He9QmDVrIZvN[wOZuTRmLA\=2M.&.1$,/"*-heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgfThdYhdYkd\kd\keWlfXkiRkiRikVikVileuxqŽ‘˜¨«²½ÇÉÒÜÞêòôõýÿüÿÿüÿÿÿÿýþþüýüøÿþúÿÿøýþöøûðõøíñ÷ëñ÷ëó÷éôøêõøåô÷äô÷ÜìïÔÛܽÈɪ®±’“–wuyblpYt{k‰�€�§œ¬¶«ÍÎÉÞßÚïðë÷øóûü÷ýþùþÿúüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúûÿÿûÿÿÿþüÿþüÿûóÿûóùÔÄÊ¥•�P=•XE¤M:›D1¶H9Ê\MÏUH»A4ÆB5ÉE8ÃJ?¾E:©B9£<3”A9•B:†A:…@9@9~?8{?7z>6z>6z>6w=2v<1u<1t;0t;0s:/p:.p:.o9-o9-n:-n:-m9,l8+k7*j6)l7/l7/j8/i7.i7.h6-h6-h6-e5+e5+e5+f6,f7-h9/i:0j;1i=4g;2f<0g=1g@/iB1fE2fE2bL5kU>fZDvjT‚€k‘�z}€mCF319,19,19,08+08+/7*/7*/7*08+/7*/7*/7*.6).6).6)-5(-4,-4,-4,-4,-4,-4,-4,-4,.5--4,,3+,3+,3+,3+-4,.5-)5+*6,,8./;12>24@46C27D37D2?L:ES<BP9;I08F-=K2DR9G_=Ia?Kc?Kc?Ha:F_8F_8F_8QjCWpI\tR]uSZqTOfI=T:,C)&.0$,."*,heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVfeSgfThdYhdYkd\kd\keWlfXkiRljSjlWlnYnqj€ƒ|ž¡¨º½ÄÃÍÏÛåçóûýùÿÿúþÿúþÿüüúúúøþýùÿþúþÿ÷úûóô÷ìîñæèîâçíáíñãïóåñôáðóàðóØèëÐØÙºÇÈ©¯²“ž¡‚�“|�”}™ �§®ž³½²¼Æ»ÜÝØéêåö÷òûü÷üýøþÿúþÿúüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúöûþûÿÿÿÿýóòðÿñéÿûóÿöæÿòâù¼©¼l¡J7¤M:ÁSDºL=»A4ÁG:Ã?2ÈD7ÃJ?¾E:¨A8¢;2“@8•B:†A:…@9€A:~?8{?7z>6z>6z>6w=2v<1u<1u<1t;0t;0q;/q;/p:.p:.n:-n:-m9,l8+k7*k7*l7/l7/j8/j8/i7.i7.h6-h6-k;1j:0h8.g7-e6,f7-g8.h9/k?6i=4f<0h>2iB1kD3fE2eD1aK4kU>i]GuiS…ƒn�Žyor_:=*08+08+08+08+/7*/7*/7*.6)08+/7*/7*/7*.6).6).6)-5(-4,-4,-4,-4,-4,-4,-4,-4,-4,,3+,3++2*+2*,3+,3+-4,)5+)5+)5++7--9-0<03@/5B15B09F4;I27E.2@'1?&7E,>L3BZ8E];H`<Ia=Ha:F_8F_8F_8PiBWpI^vT^vTXoRKbE7N4%<"%-/#+-!)+heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheThdYhdYkd\kd\keWlfXjhQkiRkmXoq\qvoƒˆ� §­¼ÃÉÍ×Ùåïñùÿÿùÿÿùýþúþÿÿÿýýýûÿÿúÿÿúüþó÷ùîîòäæêÜàç×ÞåÕæëÕéîØìðÙìðÙîîÔèèÎÚØ¿ËÉ°Á¼¨¸³Ÿ±®Ÿ¶³¤À½¶ÊÇÀÒÑÏÚÙ×èéäñòíúûöüýøüýøþÿúþÿúüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿúýÿùüüûùûúøõÿø÷ÿúõÿôóÿòÿÿìæåÐÀ�‡ŒiS“J7šQ>´Q>³P=Â=6ÆA:ÅF@ÁB<±>;ª74�99ž::�>=Ž=<€@>~><sB;sB;lD:lD:p>5p>5p>5p>5o?5n>4n>4n>4l=3l=3i=2h<1g;0f:/f:/f:/q60q60o6/o6/m6/l5.i7.i7.j;1i:0f:/d8-a9-b:.c;/d<0_C5_C5_D3aF5aH2bI3^J2]I1ZN4bV<gcJrnU€…nƒˆqXaL1:%08+08+08+08+/7*/7*/7*.6)08+/7*/7*/7*.6).6).6)-5(,3+,3+,3+,3+,3+,3+,3+,3+,3++2*+2**1)*1)+2*+2*,3+*6,*6,)5+)5+*6*,8,/;-1=/0=,3@/5B03@.0=)/<(3A*7E.=O5@R8EX<I\@I\>J]?J]=K^>SfH[nPbtZasYYkUJ\F5F3$5"#,+!*) )(heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheThdYhdYkd\kd\keWlfXigPkiRlnYoq\pun�†œ£©·¾ÄÕßáí÷ùùÿÿùÿÿøüýüÿÿÿÿýýýûÿÿúÿþùúüñóõêéíßâæØÜãÓÛâÒãèÒæëÕéíÖèìÕëëÑççÍÜÚÁÐεÒ͹ÌdzÇĵÍÊ»ÕÒËÞÛÔçæäïîìòóîøùôýþùüýøüýøþÿúþÿúûü÷þÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿùüÿõøÿþüÿÿý÷ÿúôÿ÷ïÿîôÿóõôßÿÿìÿéÓʧ‘®eR¢YF¶S@¯L9Ã>7Å@9ÃD>ÁB<³@=®;8Ÿ;;�99�>=Ž=<�A??=tC<sB;mE;mE;p>5p>5p>5p>5o?5o?5o?5o?5n?5m>4j>3i=2h<1h<1h<1h<1q60q60p70o6/m6/m6/i7.i7.f7-f7-e9.e9.d<0e=1f>2g?3^B4aE7eJ9gL;dK5cJ4_K3_K3]Q7cW=lhOvrY„mv{dEN94=(19,19,19,08+08+/7*/7*/7*08+/7*/7*/7*.6).6).6)-5(,3+,3+,3+,3+,3+,3+,3+,3++2*+2**1)*1)*1)*1)+2*+2*+7-*6,)5+(4*)5)*6*+7)-9+,9(/<+2?-3@.3@,2?+2@)3A*4F,8J0>Q5BU9EX:GZ<J]=L_?ReG[nPbtZ_qWVhRFXB2C0!2$-,"+* )(heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheTjcYjcYkd\kd\keWlfXkkSkkSinXjoYjqiy€x“œ¡­¶»ÑÛÝéóõùÿÿ÷ÿÿöúûüÿÿÿÿýøøöÿÿøþþöøúíðòåèìÝãçØßæÔßæÔäëÌæíÎèìÑæêÏéçÐæäÍáÚÈ×оÚÐÄÙÏÃÝÒÎæÛ×íãäóéêùðóÿöùùúõýþùÿÿûýþùüýøÿÿûÿÿûüýøþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿøöÿù÷ÿÿýÿÿýðÿÿëþúíÿÿíÿÿôÿúéûïÿüêÿÿíóϹ |fŠR9šbI³L;³L;²K<²K<±I@¬D;¢?9œ93”<8“;7…@9ƒ>7uC8tB7lG7lG7p>7p>7p>7p>7o>7o>7p?8p?8pA9o@8k?6j>5i=4i=4j>5j>5q60q60p70p70l7/l7/h9/g8.a9-a9-a=/b>0aA2aA2_B2_B2RG3ZO;bWAcXB_U:ZP5XQ5ZS7YX:ZY;jnSrv[x‚icmT0?(4C,2:-2:-2:-19,19,08+08+08+08+/7*/7*/7*.6).6).6)-5(,3+,3+,3+,3+,3+,3+,3+,3++2**1))0()0()0()0(*1)+2*+5-*4,*4+*4+*4++5,,6+,6+,7).9+2=-5@07B26A15@/4?.0?*4C.9H1=L5@P6DT:HX>K[ARbH[kQap[^mXSaPDRA1?0!/ &/,$-*"+(heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVgdSheTjcYjcYkd\kd\keWlfXmmUmmUhmWglVeldszr�–›§°µÊÔÖâìîõýÿôüþ÷ûüüÿÿÿÿýøøöÿÿ÷ýýõ÷ùìïñäèìÝåéÚãêØåìÚæíÎèïÐèìÑåéÎçåÎåãÌáÚÈØÑ¿×ÍÁÜÒÆéÞÚøíéþôõÿõöþõøÿ÷úýþùÿÿûÿÿûþÿúýþùÿÿûÿÿûýþùþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿüúÿüúûûùûûùòÿÿòÿÿëÿÿæÿúîÿôôÿúÿÿíþ÷åÿïÙÿâÌá©�¨pW·P?´M<²K<³L=´LC±I@¤A;�:4”<8“;7…@9ƒ>7uC8tB7lG7lG7p>7p>7p>7p>7p?8p?8p?8p?8qB:pA9l@7k?6j>5j>5k?6k?6r71q60p70p70l7/l7/h9/h9/d<0e=1d@2d@2aA2`@1\?/[>.RG3]R>h]Gi^HbX=[Q6YR6]V:XW9XW9imRrv[t~eWaH$35D-3;.3;.2:-2:-19,19,19,19,08+/7*/7*/7*.6).6).6)-5(+2*+2*+2*+2*+2*+2*+2*+2**1)*1))0()0()0()0(*1)*1)(2*)3+)3**4++5,,6-,6+,6+,7).9+2=-6A18C39D47B15@/-<'0?*4C,7F/;K1?O5DT:HX>QaGYiO_nY[jUP^MAO>.<- .(1.&/,#,)heVheVheVheVgdUgdUgdUgdUifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYnpXkmUgoXjr[gsinzpƒ‘”œª­»ÄÉÏØÝçïòóûþúþÿûÿÿþÿúýþùÿÿøýýõøúìõ÷éò÷ãïôàêòÛçïØïöÔëòÐëîÑéìÏçåÐÞÜÇÜÔÉÜÔÉáÓÐçÙÖñããúììþòöÿöúÿùýÿûÿþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿ÷ýÿ÷ýÿüýÿüýþÿýþÿýþÿýþÿÿþÿÿýÿÿûóÿüôÿÿìÿýêÿ÷ÛûðÔ¸ƒdœgH�R6–Y=¤\F�U?›L;�N=Œ?/Œ?/ƒ@/�>-v?+v?+sD0uF2r=7r=7s>8s>8u@:u@:r@9q?8r@9r@9q@9p?8p?8o>7n?7m>6j8/k90j:0k;1g=1g=1b>0b>0]@0\?/_H6_H6Q?+L:&I8$A0QS;MO7HK0JM2Z[<fgHabCTU6UX9]`AntXycanT=J0,:#1?(3;.3;.3;.3;.2:-2:-2:-19,2:-19,19,08+08+/7*/7*/7*/6./6./6..5-.5-.5--4,-4,+2*+2*+2*+2**1)*1)*1)*1)+5-+5-*4,*4,*4++5,+5,,6-(2)*4+.8-2<17A6:D9=G<>H=5B04A/1>,/<*1?(7E.AO8HV?N\EUcL[hVXeSKWI:F8.:0(4*&/*%.)%.)heVheVheVheVgdUgdUgdUgdUifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYlnVjlThpYks\htjp|r†”— ®±¾ÇÌÒÛàéñôôüÿúþÿûÿÿþÿúüýøÿÿøýýõùûíöøêóøäðõáéñÚæî×éðÎäëÉäçÊäçÊãáÌÜÚÅÜÔÉÞÖËçÙÖíßÜõççýïïÿôøÿ÷ûÿùýÿûÿþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúýÿ÷ýÿ÷ýÿüýÿüýþÿýþÿýþÿýþÿÿþÿÿþÿÿüôÿýõÿÿìÿÿìÿùÝÿôØÿçÈå°‘¯rV–Y=žV@ XB¡RAŸP?’E5‘D4ˆE4†C2|E1yB.pA-o@,t?9s>8t?9u@:t?9q<6p>7sA:q?8q?8o>7o>7n=6n=6m>6l=5m;2m;2m=3n>4j@4i?3eA3d@2dG7]@0]F4_H6O=)G5!G6"D3IK3HJ2DG,BE*LM.Z[<abC_`AWZ;cfGrx\rx\UbH7D*.<%6D-5=05=05=04</4</4</3;.3;.4</4</3;.3;.2:-2:-2:-19,18007/07/07//6./6./6./6.,3+,3+,3++2*+2**1)*1))0((2*'1)'1)'1)'1((2))3*)3*+5,+5,,6+.8-2<17A6<F;?I>9F48E35B02?-1?(4B+:H1?M6JXAQ_HWdRUbPJVH;G90<2*6,(1,(1,'0+gdUgdUgdUgdUheVheVheVheVifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYkmUikSgqYku]hukq~tˆ—š¤³¶ÀÉÎÔÝâêòõõýÿúþÿûÿÿþÿúüýøÿÿõüüòùüëöùèò÷áîóÝçï×ãëÓáçÅÜâÀÝÝÃÞÞÄàÚÌÝ×ÉßÖÑãÚÕîâäóçéùîòþó÷ÿ÷øÿøùÿûûÿüüþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúöÿüöÿüûÿþûÿþÿûÿÿûÿÿ÷ÿÿ÷ÿÿùÿÿùÿÿúøÿüúÿÿôÿÿôöÿîòþêÿôÛÿüãÿï×Ù맅l]D�U<‹_F’`IˆV?|H2s?)q>)s@+tD.vF0o?5l<2n>4sC9rB8n>4pA7wH>k?4k?4j>3j>3i?3h>2h>2h>2cB3cB3cC4cC4`E4`E4[F3ZE2^M9SB.VK5^S=OI1B<$B>%EA(=G,?I.=F)8A$=B"KP0\_@fiJ^aBmpQuy^gkPHP84<$4=(=F17?27?27?26>16>15=05=05=06>16>16>16>15=05=05=05=03:23:23:2291291180180180/6..5-.5-,3++2**1)*1))0((/('.''.''.'(/()0)*1**1*/6/-4-+2++2+.5.4;4<C<@G@=K:=K::I66E22A*0?(2A*3B+CR;IXAP_LO^KGUF:H91>4+8.+5-*4,)3+gdUgdUgdUgdUheVheVheVheVifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYkmUjlTgqYjt\fsio|r‡–™¤³¶¿ÈÍÓÜáéñôóûþùýþúþÿþÿúýþùþþôüüòùüë÷úéò÷áíòÜåíÕàèÐÛá¿ÕÛ¹ÖÖ¼××½ÜÖÈÞØÊäÛÖëâÝøìîüðòÿõùÿøüÿùúÿùúÿûûÿüüþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúöÿüöÿüûÿþûÿþÿûÿÿûÿÿ÷ÿÿ÷ÿÿøþÿ÷ýÿúøÿûùÿÿôÿÿôøÿðõÿíÿýäÿüãÿðØÿðØÿîÕôÒ¹ºŽu}Q8yG0xF/zF0~J4€M8|I4p@*f6 {KAsC9qA7tD:qA7i9/f7-j;1i=2i=2i=2i=2h>2h>2h>2h>2bA2bA2bB3bB3_D3^C2YD1XC0XG3Q@,^S=qfPc]EOI1KG.NJ1:D)=G,>G*9B%;@ DI)UX9adEjmNuxYsw\Z^C=E-4<$7@+;D/8@37?27?27?26>16>16>16>17?27?27?27?27?27?27?27?26=56=55<45<44;34;34;34;329118007//6.-4,,3++2**1)*1*)0))0))0))0)*1*+2+,3,.5.,3,*1*)0),3,1817>7;B;>L;?M<?N;<K87F/2A*0?(/>';J3AP9GVCHWDBPA9G81>4,9/,6.+5-*4,fcTfcTgdUgdUheVheVifWifWifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYjnUimTfrZgs[aqfiyn�•–Ÿ³´¾ÇÌÑÚßèðóòúýøüýúþÿÿÿûþÿúÿÿôÿÿóüÿìúýêöûäïôÝåîÑàéÌÙܽÒÕ¶ÑѹÓÓ»ÚÓÉßØÎéàáòéêýôùÿ÷üÿûÿÿýÿþÿúýþùüÿöýÿ÷þÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúúÿÿúÿÿýÿþýÿþÿþúÿþúÿý÷ÿý÷ÿûõÿûõÿü÷ÿýøþþüÿÿýûÿþúþý÷óòÿþýÿþøÿþøÿüïÿùìõè×âÕÄȶ ®œ†�v`qXBfI7fI7mN<pQ?hG4dC0cB/gF3iJ6kL8kL8lM9_C._C._C._C.^C._D/_D/_D/XG3XG3XI4XI4TI3SH2OI1OI1KI0KI0beJ|dntXSY=FO2DM09F*:G+>G*>G*@E%BG'LM.ST5wvXzy[nnTTT:@B,?A+>A.:=*7?27?27?27?26>16>16>15=06>16>16>17?27?27?27?27?28?78?77>67>66=56=56=56=54;34;33:229107//6..5--4,+2++2+*1**1**1**1*+1-,2.+1-+1-*0,+1--3//511733958I6;L9>P:=O9:L45G/1C)0B(3E-8J2=O9?Q;<M;5F4.>1*:-+5,+5,*4+fcTfcTfcTgdUheVifWifWifWifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYjnUimTfrZfrZ_odgwl�•–Ÿ³´ÁÊÏÓÜáéñôòúýøüýûÿÿÿÿûÿÿûÿÿôÿÿôÿÿïþÿîùþçò÷àæïÒàéÌÖÙºÏÒ³ÌÌ´Í͵ÕÎÄÞ×Íëâãöíîÿ÷üÿùþÿýÿÿýÿÿÿûýþùýÿ÷ýÿ÷þÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúúÿÿúÿÿýÿþýÿþÿþúÿþúÿý÷ÿý÷ÿûõÿûõÿü÷ÿü÷ýýûþþüûÿþûÿþýùøÿüûÿúôúõïÿøëÿþñÿýìÿýìÿøâÿñÛúáËéкո¦º�‹¡‚p�q_{ZG}\ItS@dC0Y:&Z;']>*[<(`D/`D/aE0aE0`E0aF1aF1aF1\K7\K7[L7[L7WL6VK5RL4QK3MK2RP7ilQ}€elrVSY=DM0>G*9F*9F*?H+BK.DI)BG'GH)LM.xwYsrTddJSS9IK5HJ4CF3:=*8@38@38@37?27?26>16>16>16>16>16>16>17?27?27?27?29@89@89@88?78?77>67>67>67>66=56=55<44;33:2291180-4--4-+2+*1*)0))0))/+)/+)/+*0,,2.-3/.40.40-3/-3/0A.5F3;M7=O9=O79K37I/5G-2D,5G/9K5;M7:K94E3.>1*:-+5,*4+)3*ebSfcTfcTgdUheVifWifWjgXifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYimTimTfrZgs[_qehznƒ—˜¤¸¹ÇÐ×Úãêíõøõýÿùýþûÿÿÿÿúþÿùÿÿóÿÿóýÿíüÿì÷ýãîôÚâëÎÛäÇÕÕ¹ÌÌ°ÇıÇıÏÈÂÚÓÍéàåöíòþ÷ÿÿùÿÿþÿÿþÿûÿùúÿøøÿóùÿôþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿûÿÿûÿýþÿýþÿ÷ÿ÷÷ÿ÷óÿòóÿòóÿñôÿòøÿ÷øÿ÷þüÿÿýÿÿúÿÿúÿùôÿýøÿÿýÿÿýÿýþÿûüÿöüòóùïþÿóþÿóÿÿïÿÿïÿýîÿ÷èüîáøêÝöåËùèÎÝ̲¡�vq`FbQ7^N4YI/ZJ0ZJ0ZL1ZL1[M2[M2YN2ZO3UQ8UQ8UQ8UQ8QQ7OO5LO4KN3JP4NT8]hJdoQQ^@BO1>K-9F(=J.<I-BK.FO2GL,EJ*ON0YX:qmPjfI`\CZV=VT?TR=LI8EB1:B5:B5:B59A49A48@38@38@36>16>17?27?28@38@38@39A49@89@89@89@88?78?78?77>68?78?78?77>67>66=56=56=538227105/.3-,1-+0,+0,+0,+0,,1--10-10.21-10,0/,0/':$-@*5I0:N5;O4:N39M18L04H-5I.8K5:M79L95H5.@0*<,+5*+5**4)ebSebSfcTgdUheVifWjgXjgXifWifWifWifWifWifWifWifWjdTkeUkdZkdZle]le]jgXkhYgkRhlSfrZht\`rfj|p‡›œ¨¼½Î×Þßèïòúýøÿÿûÿÿûÿÿþÿùýþøùùíùùíøûè÷úçñ÷ÝéïÕÜåÈÔÝÀÒÒ¶Éɭ¿¬Â¿¬ÊýÕÎÈæÝâóêïþ÷ÿÿùÿÿýÿÿþÿûÿùúÿøøÿóùÿôþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúþÿúÿûÿÿûÿýþÿýþÿ÷ÿ÷÷ÿ÷óÿòóÿòõÿóõÿóùÿøùÿøÿýÿÿýÿÿúÿÿúÿÿûÿÿûÿþüÿþüÿüýÿùúþùÿõüÿø÷ûìøüíüûéüûéÿøéÿúëÿüïÿüïÿïÕÿúàìÛÁŸŽt`O5RA'XH.XH.[K1\L2[M2\N3\N3]O4[P4[P4SO6SO6SO6SO6OO5MM3JM2IL1IO3KQ5Q\>OZ<<I+9F(>K-=J,AN2AN2HQ4IR5IN.IN.ZY;jiKgcFa]@]Y@_[B][FXVAPM<LI8<D7;C6;C6;C6:B5:B5:B5:B57?27?28@38@39A49A4:B5:B5:A99@89@89@88?78?78?77>69@89@89@89@89@89@89@88?78=77<65:4271051/40.3/-2..3/-2.-10,0/,0/,0/,0/,0/#6 )<&1E,8L3:N3:N39M19M16J/7K09L6;N8:M:6I60B2+=-,6+,6++5*YgMYgM[hN]jP`jRblTfkUfkUfgUefTheTgdSjdVjdVldWmeXjgVjgVjfZjfZjf]ie\geVgeVdiSfkUco[bnZ`qks„~”§®µÈÏÓÛÞáéìïø÷ôýü÷þ÷øÿøüÿòûÿñõøå÷úçø÷ãñðÜïìÙìéÖáÞËÔѾÚͼ̿®¿³§¾²¦Â·µÌÁ¿ÛÒ×êáæøóúü÷þýüÿþýÿýþÿûüÿùýüúþýýÿþýÿþýÿþýÿþþþüþþüþþüþþüÿþúþýùÿüùÿüùÿýøÿýøÿþùÿÿúÿýúÿüùþýûÿþüÿÿÿÿÿÿýýýûûûÿýüþúùÿûõÿþøÿýôÿúñÿøíÿúïþüÿþüÿþþüþþüýÿúýÿúúþýúþýýþÿýþÿÿýÿÿýÿÿýñÿýñÿÿäÿÿäõõÛûûáâáÅ‹ŠnSN0TO1RL,QK+TN,TN,SN.SN.QP2QP2PP4PP4ON2ON2OO3PP4LP7JN5EN3DM2GQ6DN3AJ+=F'?F$CJ(FL&FL&QU2PT1KO.HL+TT8bbFbbHZZ@YY?ZZ@[[A[[AZ[<WX9UT5SR3ID.HC-DB-B@+@>/@>/=?2>@3:=2:=28@38@37B47B47D37D37B47B46A36A36A36A37B47B47B47B47B47B48C59D6;F8;F87C97C96C25B16?.4=,2:-08+-4-.5.,8,,8,*=.A#0J4N6Y!7Z"8W.7V-:R:9Q9<O<=P=9K1@R8DY:EZ;BZ:=U51I/&>$+5**4))3(YgMYgM[hN]jP`jRblTfkUglVghVfgUifUheTkeWkeWmeXmeXjgVjgVjfZjfZjf]ie\geVgeVfkUglVdp\fr^gxryŠ„™¬³¶ÉÐÑÙÜàèëïø÷õþýøÿøøÿøûÿñøüîöùæöùæôóßìë×éæÓåâÏÚ×ÄÎ˸ʽ¬Á´£¹­¡»¯£Á¶´ÌÁ¿ÙÐÕçÞãùôûü÷þýüÿþýÿüýÿûüÿùýüúþýýÿþýÿþýÿþýÿþþþüþþüþþüþþüÿþúþýùÿüùÿüùÿýøÿýøÿþùÿþùÿýúÿüùýüúþýûÿÿÿÿÿÿýýýûûûÿþýÿþýÿý÷ÿøò÷íäïåÜõçÜþðåþüÿþüÿþþüþþüýÿúýÿúúþýúþýýþÿýþÿÿýÿÿýÿÿþòÿýñÿÿäÿþãõõÛûûááàĉˆlQL.RM/RL,QK+TN,TN,TO/TO/QP2QP2PP4OO3ON2ON2OO3OO3LP7JN5FO4EN3FP5DN3AJ+>G(AH&EL*IO)JP*NR/OS0VZ9]a@aaE[[?VV<TT:UU;VV<WW=WW=WX9VW8VU6UT5PK5NI3IG2FD/CA2A?0>@3>@3:=2:=28@38@37B47B47D37D37B47B46A36A36A36A37B47B47B47B47B47B48C5:E7;F8<G9;G=;G=;H7:G6;D39B18@37?26=66=61=11=11D&8K-?Y*F`1Mp8Lo7Ih?Dc:@X@;S;:M:9L9<N4DV<I^?LaBJbBE]=9Q7-E+*4)(2''1&YgMYgM[hN]jPakSblTglVglVijXhiWjiWihVlfXlfXmgYmgYjgVjgVjfZjfZie\ie\hfWgeVfkUejTbnZgs_l}w~�‰™¬³²ÅÌÍÔÚÞåëï÷ùöþÿøÿø÷þ÷÷ûíôøêñôáîñÞéèÔàßËÜÙÆ×ÔÁÏÊ·ÆÁ®À³¢½°Ÿ¼°¤Ã·«ËÀ¾ÖËÉáØÝìãèúõüýøÿýüÿþýÿüýÿûüþúþýúþýýÿþýÿþýÿþýÿþþþüþþüþþüþþüþýùþýùÿüùÿüùÿýøÿýøÿýøÿýøÿýúÿüùýüúýüúþþþÿÿÿþþþüüüÿþýÿþýÿþøúóíêà×àÖÍëÝÒùëàþüÿþüÿþþüþþüýÿúýÿúúþýúþýýþÿýþÿÿýÿÿýÿÿþòÿýñÿþáÿýà÷÷ÝûûáàßdžjOJ,QL.RL,RL,TN.TN.TO/TO/QP2QP2OO3OO3NM1ON2NN2OO3KO6KO6GP5FO4FP5DN3CL-BK,FM+IP.OU/RX2OZ/U`5grJv�YovTW^<KP2LQ3NP8NP8QQ9QQ9TR9US:VU9WV:VQ;TO9OM8KI4GE6DB3?A4>@3;>3;>38@38@37B47B46C26C28C57B47B46A36A37B47B48C58C57B47B48C58C5:E7;F8<G9=I?=I?=J9=J9?H7>G6<D7;C6=D=;B;5A54@46I+BU7Mg8WqB]€HZ}EUtKKjABZB8P85H52E2<N4EW=MbCQfGPhHLdD?W=2J0/9.,6+*4)XfLYgM\iO]jPakScmUhmWhmWklZjkYkjXkjXmgYnhZnhZnhZkhWkhWjfZieYie\ie\hfWhfWdiSbgQ_kWfr^l}w}Žˆ“¦­§ºÁÈÏÕÙàæìôöôüþöýöôûôòöèíñãçê×áäÑÛÚÆÒѽÍÊ·ÈŲ½ª¼·¤Ä·¦Çº©ÌÀ´ÖʾÞÓÑéÞÜòéîúñöü÷þþùÿýüÿþýÿüýÿûüþúþýûÿþýÿþýÿþýÿþýÿþþþüþþüþþüþþüþýùþýùÿüùÿüùÿýøÿü÷ÿü÷þûöÿýúÿûøüûùüûùýýýþþþþþþýýýÿûúÿþýÿþøüõïíãÚãÙÐðâ×ÿñæþüÿþüÿþþüþþüýÿúýÿúúþýúþýüýÿýþÿÿýÿÿýÿÿþòÿüðÿýàÿüßùùßþþäáàć†jPK-RM/SM-SM-TN.UO/UP0UP0RQ3QP2NN2MM1NM1NM1MM1NN2KO6KO6HQ6HQ6HR7FP5FO0HQ2LS1OV4V\6]c=grGtT‡’j�›s†�kmtRZ_ARW9LN6KM5LL4LL4OM4QO6TS7VU9XS=VQ;RP;NL7JH9FD5AC6@B5;>3;>38@38@37B47B46C26C28C58C57B47B47B47B48C58C58C58C58C58C59D6:E7;F8<G9<H><H><I8<I8>G6=F5<D7<D7=D=:A:5A54@47J,DW9Pj;[uFY|DWzBRqHGf=<T<1I1.A.,?,6H.@R8J_@QfGQiIMeE?W=2J00:/.8-+5*XfLYgM[iO]kQblTdnVfnWgoXkn[jmZlkYkjXliZliZnhZoi[liXkhWjfZieYhd[hd[hfWigXejTbgQ_kWfr^k|vz‹…ŒŸ¦ž±¸ÂËÒÔÝäçñóðúüóùõðöòíðåçêßàãÒØÛÊÑмËʶÉİþª½¸¤¹´ Ë¾®ÓƶÜÏÆæÙÐíâàöëéýôùÿùþýøÿÿúÿþýÿþýÿüýÿûüþúÿûûÿüýÿþýÿþýÿþýÿþþþüþþüþþüþþüýüøþýùÿýúÿýúÿýøÿü÷þûöýúõÿüùþú÷ûúøûúøüüüþþþþþþþþþÿûúÿüûÿüöÿøò÷íäðæÝøêßÿôéþüÿþüÿþÿúþÿúýÿùýÿùúþýúþýüýÿüýÿÿýþÿþÿÿÿðÿýîÿüßþúÝüûæÿÿêäâÉŠˆoRM0TO2UN1UN1TN.UO/UP0VQ1RQ3QP2NO0LM.ML0ML0MM1MM1KO6KO6IR7IR7KU:HR7JS4NW8RY7T[9\b<flF}”^�§qœ²�—­|‘¢x…–lq|ZYdBKO6JN5KJ5JI4NI5PK7SN:UP<UP:TO9QO:OM8KI:HF7CE8BD7<?4<?49A49A46A36A35B15B19D68C57B47B47B47B48C59D68C58C58C59D69D6;F8<G9=H:;G=;G=<I8<I8>G6=F5<D7;C6:A:8?84@45A57J,BU7Lf7Uo@Ru=Ru=NmDCb97O7,D,*=*)<)1C);M3G\=OdEPhHLdD>V<2J0+5*(2'%/$
\ No newline at end of file
diff --git a/testimgp.jpg b/testimgp.jpg
new file mode 100644
index 0000000..968a90e
--- /dev/null
+++ b/testimgp.jpg
diff --git a/testorig.jpg b/testorig.jpg
new file mode 100644
index 0000000..9816a0c
--- /dev/null
+++ b/testorig.jpg
diff --git a/testorig.ppm b/testorig.ppm
new file mode 100644
index 0000000..2a5d1e9
--- /dev/null
+++ b/testorig.ppm
@@ -0,0 +1,4 @@
+P6
+227 149
+255
+0/-0/-10.21/51.51.62/62/83/83/:2/:2/:3-:3-:3-:3-:2/:2/91.80-80-91.:2/:2/80-80-80-80-80-80-80-80-6.+6.+6.+5-*5-*5-*4,)4,)4,)4,)4,)4,)4,)4,)4,)2-)/*$/,%0-&0-&1.'2/(30)30)63,63,74-85.96/96/:70:7.A:0B<0D>2F@4IA4JB5KC6KC6NE6MD5OC3NB2OC3OC3PD4RE5R?1Y?2b@4nB5}E6‹H8™G9£F7°H;¸F;¿F;ÅF=ÇG>ËH@ËH@ÐEBçFLíCLìEMëEIîCIïBDò?Cô=Aø;A÷:@ô:?ð<?é?Bâ?@×?<ËA7»=/µ@.µ@.´?-´?-³@-²?-¯@-­?.ªA.¦A-¢B, A-›@+™A+–A,”>-’?/’?/‘>.‘>,‘>,’<+’<+”>-”>-”=*”=*•>+•>+–?,–@/–?6•>5—=4Ÿ?3©B3³D3¼D4¿D4¹?0¶B3¬F:žH;‡G;oA2U9+C3&=52:659548437116005//5//72/72/72/61.61-61-50,41,//-.0-//-//-0/-0/-2.-2.-5,-4+,4*+3)*7(+=.1E69P:<jAE|HJŽNO•OQŸW[ªdnªoƒŸt�”{£‡®€†º~ˆ½tz®`a‘TKvQEiSJgPH_MH^TQbfdo|}‚�‘�™ž˜£©Ÿ¢¨šž “�‘ƒ{|lfgWYXFQNEUR[UQbUQb0/-0/-10.10.40-51.62/62/83/83/:2/:2/:3-:3-:3-:3-:2/91.80-80-80-80-91.:2/80-80-80-80-80-80-80-80-6.+6.+6.+5-*5-*4,)4,)4,)5-*5-*5-*5-*5-*5-*5-*3.*1,(0-&0-&1.'2/(30)41*41*63,63,74-85.85.96/:70:7.@9/B<0C=1E?3H@3IA4JB5JB5LC4LC4MA3MA1MA1NB2OC3QD4P>0U?1^A3jC4xD6…F5’E5œC3§C4¯A4µ@6¼B7ÀD:ÄE<ÅF=ÍC@áEIçBIèCIêDHíDGðCEó@Cö?Cø;A÷:@ô:?ð<?é?@à@@Õ@<Ê@6¹>/µ@.´?-´?-´?-²?,°?-¯@-­?.ªA.¦A-¢B,Ÿ@,›@+˜@*–A,”>-’?/’?/‘>.‘>,�=+’<+’<+”>-“=,”=*”=*”=*•>+–?,–@1•A6–?6˜>5¡?4«A3µD4½C4¿D5»A2·C6¬F:œH=…G:l@3S9*B4)>63:65:6584382271160060072/72/72/61.61-61-50,41,//-.0-//-//-0/-0/-2.-2.-4..5,-5+,3)*5)+<-0C47N8:d=>vDC†JI�MNšTV¤aj¥l}�rŽ‘{¢†€®…¹{„»ou©[[�RIvOCiOFePH`PH_RN_[Yfnot…†ˆ”™•™ž—š ”™���‘ƒ~qjk[][LVSJXSZVRaXQa/.,0/-0/-10.40/40/51.51.72.72.72.72.92,92,92,92,91.80.80.7/-7/-80.80.91/80.80.80.80.80.80.80.80.6.,6.,5-+5-+5-+4,*4,*4,*6.,6.,6.,6.,6.,6.,6.,4/+2-)1.)2/*30+30+41,52-52-63.63.74/74/850961961:70?8.@:.B<0D>2G?4H@5H@3H@3J@4I@1K?1K?1K?1L@2MA1NB2MA3QA2YB4dC4qC4|C2‡B2’A0˜<-¡;,§;.¯=2µ@6ºD:¿F=ÅD>ÙCEá@FãBGèCGêDFðCEôADø?Dú;@ù:?õ;@ð=@è@@ÝB>Ñ@;Æ@5·=.³@-³@-³@-²?-°?-¯>,­@,ª?-§@-¥@.¡@-�A,›@+˜@*•@+”>-’?/‘>.‘>.‘>.�=-�=+�=+‘>,‘>,’<+’<+“=,”>-•?.•?0•A6–?5š>3¤?3¯A4¹C5¿D5ÁC5ÀD8¹G<®I=™J=�G;h@4Q:,B5,?74=77<66;5594183072/72/62/62/62/51.52-52-41,21,/1.-2./1./1.00.00.10.3/.5//4..5,-4*+4*+9-/>24I56[97l?:}FA†ID�OM˜[`›fv•n‰Œwžƒ}­}‚¹u~·gl¤UU‰MEvLAkMAeOFcQHcNI_NK\[[esty‡‰ˆ‡Œ†Š�ˆ��…†Š|xzlfhZZ[MVSLZU[ZT`[S`.-+/.,/.,0/-3/.40/40-51.61-61-61-61-81+81+81+80-80.7/-6.,6.,6.,6.,7/-80.80.80.80.80.80.80.80.80.5-+5-+5-+5-+4,*4,*3+)3+)6.,6.,6.,6.,6.,6.,6.,4/,30+30+30+41,41,52-52-52-52-63.63.74/85096196196/>7-?9-A;/B<0E=2E=2F>3F>1G=1G=1H<.I=/I=/J>0L@0JA0LE5NE4VE5^D3iD2sB1~A/†?-Œ9)”9'�9*¤=.¬@3³E8¸H<ÁF>ÒDCÚACÞBEâDEèDEìBCó@C÷?Aú;@ù:?ö<Aî>@åA@ÚB=Í@9Â@3¶>.°@,°@,°@,¯>,®?,®?,¬?+©@-¦?,£@- ?,œ@+˜@*–@)”?*‘>,‘>.‘>.‘>.�=-�=-�<*�<*�=+�=+�=+�=+’<+‘>,”>-’?/•A6—@6œ>2¦@4²B6¼C8ÁC7ÂB7ÂF<ºJ?¬KB—J@|F:b@4L:.A7-@85>88=77<66:5294183083062/62/62/32.52-32-21,12--2.-2./1./1.00.00.10.10.5106005//5,-4+,6,-:01D22T71c;3rB8{E;ƒIE‰RU�_l‹i‚ƒs˜}y«x}µowµae¢SRŒMDyL@pL@hPEgQFfLC^HCWNLZ^^fjnquyxy~xz€vwzokoa`bUWYLTTL]WY]V]]V^------/.,/.,0/-10.3/,40-40-40-50,50,50,50,7/,7/,4/,4/,3.+3.+3.+3.+4/,4/,50-50-50-50-50-50-50-50-3.+3.+3.+2-*2-*1,)1,)1,)4/,4/,4/,4/,4/,4/,4/,4/,41,41,41,41,52-52-52-52-52-52-63.74/74/85096196/<5-=6,?8.@9/B90C;0C;0C;0E;/D:.F:.G;/H<.I=/J>0I@1JG6MH5RG5YF5bE3jD1uB/|?,‚;)‹:)“:*š=,£B2¬F8²J=¼J@ÌGBÔDCØDDÝEDãCCéAAð=@ô<>ù:?ù;=ô<>í?>áB>ÓC:ÅA5¹?0²?-¯@-®?,®?,®?.¬>-¬>-ª?-¨>.¤?- ?,ž?+š?,—?+•>*”?+‘>,�?.�?.�>-�>-Ž=,Ž=,Ž=,Ž=,Ž=,Ž=,Ž=,�<,�>-‘>.‘@/”B4—A4ž@4¨@3¶A7¿C9ÅB:ÄA9¾C;·H?¦LC‘KCtE;Z>2E9-=6,A96@86?75>64=53<4294183062/43/43/23.32.23.12-02--2.,2.-2.-2./1./1.00.10.3205105104..2,,5,-7./>0/N5.Y9.e=1oA4tC<yKK€Ze„h�p—zxªu{·ltµ_d¦TT”OGƒLBwNAmNBhMAeJA`GBYHEXKKWMPU^bc`fbbia`f\Z`VWZOUXMXXP^YV`WX`WZ,,,,,,.-+/.,/.,0/-3/.3/,2.+2.+3.*3.*3.*3.*5-*5-*3.+3.+2-*2-*2-*2-*3.+3.+3.+3.+3.+3.+3.+3.+3.+3.+3.+2-*2-*2-*1,)1,)1,)0+(4/,4/,4/,4/,4/,4/,4/,4/,40-41,41,41,41,41,41,41,41,52-52-63.74/85085096/;4,<5+=6,?8.@7.A8/A8/A9.C9/C9-E9-F:.G;/I=/J>0HA1JG6JI7NG5VF6\E3dC2lA0t?-|=,ƒ<*Œ;*”=,œ@1£F5ªJ:´J=ÄH@ÌEAÑFAÖE@ÞCAåA?ì>?ò;=÷;<ô:;ð<=é@=ÜC=ÍC8¾@2²?-®?,«@,«@,ª?+ª?-©>,©>,¨?,¥>-£@- ?,œ?-—?+•>*“>)‘?*�?,�>-�>-�>-Ž=,Ž=,Ž=,�<+Ž=,‹<+�<+‹<+‹<-Œ=,�>/Ž?0”B4˜B3¡A3¬B5¹C9ÂC:ÅB:ÃB<»B:±HB£NG�NEpH>T@5A;/96-@85A75?75>63=5394194173043/43/34/23.23.13.02-.3--3/-3/.3/.3/02/02/11/11/21/32032040/2.-1-,4..8.,G4-O4)X8+`<0e?6mGFyYd‚k…€uŸ||²w|¼nu»dh¯[[¡SL�LB~OArL@hI=cH>`HB^ECX@BO<AGCHKDMJJQJJQIIPHKQGOUKVWO^YS`YS`XU+++,,,,,,---/-.0/-0/-0/-1-*1-*1-*1-*2-)2-)2-)2-)2-*2,,1++1++1++1++2,,2,,1++1++1++1++1++1++1++1++2,,2,,2,,1++1++1++0**0**3--3--3--3--3--3--3--3.+40-40-40-40-3/,3/,3/,3/,40-40-51.62/730730841850:3-;4,<5-=6.?6/?6-?6-?6-C90C9/E8/F90G:1I=1J>2H@3HE6GE6KE5QD4YD3_B2g@/n=,v=,|:*…9+Œ:,“=.›B2¢F7¬F8¼G=ÂF>ÉF>ÐE>ÙD@âC?ê@@ð>>ò::ñ;:ì<<äA<ÖC;ÆD6µ@/ª=)ª?-©@-©@-©@-¨>.¨>.§=-¥>-£=. ?.ž?-š?-–?,”?+‘?*�>)�>+�>-�>-Œ=,Œ=.Œ=.‹<-‹<-‹<-Š=-Š;,‰<,Š</‹>.Œ>1Œ?/’C4˜B3¡A3®B6¼C:ÃD=ÄC=ÀC=ºGB²QK¦YS�XQsRIWI>CC7<?6>93@72>63=60:5194083/63.43.43.34/23.13.13.02-.3--3/-3/.3/.3/.3/.3/02/02/00.11/22021/10./.,2.-4/,?0+D0)K3)T8-Z<4eGGu]jƒs�‰€«…„¾~ƒÇtzÆmp½ee¯VSšLC‚K?qJ=hG;cE>_FB]DBW?AN;?H:BE>HGDMHGQIGQHJRGNVKUXM^ZOaYNaXO++++++,,,---.,-/-.0/-0/-1-,1-*1-*1-*2-)2-)2-)2-*2,,1++1++0**0**1++1++2,,0**0**0**0**0**0**0**0**2,,2,,2,,1++1++0**0**0**2,,2,,2,,2,,2,,2,,2,,2,,3/,3/,3/,3/,3/,3/,3/,3/,40-40-51.62/62/73084185092,:3+;4,<5->5.>5.>5.>5,B8/B8/E80F90G:1I<3J=4I?5FB6FB6JB5OA4UB3\@2c?1j<-q<.w9*}8)…7*Œ:,–>0›B4¤B5²F:ºE;ÁF>ÊG?ÔG@ÞFAçCAîB@í;;ë;;ç>;ßB;ÑD:¿D4°A.¤>(¦A-¦A-¦A-¥@.¥@.¤?-¤?-¤>/¢>.Ÿ@.œ?.˜?-•>+‘?*�>)�>+�>-�>-�>-Œ=,Œ=.‹<-‹<-Š=-Š=-‰=/ˆ<.ˆ<.ˆ</‡=0ˆ>1‹?1‘D4–C3¢B4­C6ºC;ÁD>ÁD@»EA¹PL²[T¦f]‘f]u_T[UIHNBCI?<92?82>71;6094.74-63.43.43.43.34/23.13.13..3-.3-.3/-3/.3/.3/.3/.3/02/02///-11/22022010.0/-0/-3/,8,,<-*C0*K70S<6^HJtbn‡z”�Š¶Œ�Æ„ˆÏz€ÌtwÆjl·YW ID„E=nG<dD<aC>^CAY@CV@DP>EKGQRKWUQ^WU`XT`VS]TT^SY_S^[LaZJaZJ,-/,-/--/--/------.,-.-+/.,/.,1-*0,)0,)0,)/+(/+(/+*/+*/+*/+*/+*/+*0,+0,+/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*0,+0,+1-,1-,2.-2.-2.-1-,1-,1-,1-,1-,1-,0,)1-*2.+3/,3/,3/,3/,3/,3/,3/,3/,3/,40-51.62/73080-92,:3-;4.=4/>5.>5.>5.@5/@6-B5-C6.D7/F91H;3G=4G>5H@5J@6P?5T>3X<1^90c7,m9.t8-|8-ƒ9.�;/“=0˜?1ž>0§A3­A4µC8¾E:ÊG=ÔG>ÞE?åC@è@?êBAæDAÚE>ÈD8·B1ªA.¢B,¢A.¡@-¢?,¢>.¡=-¡=-¢>. ?/œ<,š=,˜<-•>-“=,�=+Ž=*Œ=,‹<+‹<+Š=-‰<,‰<,‰<,ˆ:-‡;-‰=/‡=0‡=0‡=0ˆ>3ˆ@4‰A5‹A4‘E5—D4£E9±I>ºG@»D>»EA¸MGµ[S¯f_£qf‘sh~rdjj^V^SJRGLLBJF=B>5=90:6-74+63,33+54/34.34/23.02-/1,.0-,1--2.-2.-2.-2./1./1./1./1.02/02/11/11/11/11/11/40/4+0;/3A32C41J;8]NQym{‹…Ÿ“�»”–Ï�•Ùƒ‰ÓtzÆjn·`c¨Z[”LItHBdA>]>>X?BUAIVLU\U`bbqno~yv†|s€vlyohth_k_W_P^]Ib\Fc]G,-/,-/--/--/------.,-.,-/.,/.,0,+0,)0,)/+(/+(/+(.*).*).*)/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*/+*0,+0,+1-,1-,1-,1-,1-,1-,1-,1-,1-,1-,1-,0,+1-*2.+2.+3/,3/,3/,2.+2.+2.+2.+2.+3/,40-62/62/80.91.:2/;4.=4/>50>50>5.?4.?4.B5/B5-D7/E80G:2H;3H>5H>5L=6O>6R>5V;2Z90_7/i81p7.x8.�8/Š:/�<1–<1›=1¢@3§A3­C6´D8¾E:ÉF<ÔE=ÛC>ßD@àEAßGBÔG>ÄF:³D3¥B/žB-ŸC.žA/Ÿ@.ž?-ž?-ž?-ž>.Ÿ?/š=,™>,–=-”=,�=+�>+Œ=,Œ=,‹<+Š=+‰<,‰<,‰<,‡;+‡;-…<-†</…=/…=1ƒ=1…?3†@6ˆB8ˆB6“G9˜F8£G<¯JB¸HD¸GC·KH³TN±d\ªqfŸ~o�o|mmseZfZNXMLNCKI=EC7A>5=:188.44,11)23-23-12-01,/1,/1,.0-.0-/1.-2./1./1./1./1./1./1.02/02/11/11/11/11/11/2015+49-7<15?54I?=^UVys}Šˆž““¹”—Ê�–Ô„‹Ïy�Âqy¸kt­hnž]`�XZqSUjRWjT^hZgmfvvr‚tˆ~’‡ƒ•‰~Ž�w‡zr€qftgZeT[ZE`ZBb\D-.0-.0-.0-.0-.0-.0.......,-.-+.-+-,*/+*.*).*'.*',+),*+,*+,*+,*+,*++)*+)*-+,-+,-+,-+,-+,-+,-+,-+,,*+-+,-+,-+,-+,.,-.,-.,-.,-/-./-./-./-./-./-./.,0,+0,+1-,2.-2.-2.-2.-1-,1-,1-,1-,1-,2.-3/.40/51.80.91/:20<41=31>42=31=4/?40?4.A4.A4.C60D71F93G:4H;5J;6K<7N=6P;6S:5W83[6.c60k6.t5,}7/‡9/�;0”<2—=2ž@6 @4¢@3¨@3±C6ºD8ÅE<ÍD<ÕF@×HBÔIBÌI?¾E:®C3¡B0œA.�B/œA/œ?.›>-›>-›>-›>/›?0˜<-–=-”<.“=.�>-Œ=,Œ=.Š=-‰<,‰<,‰<,ˆ<,‡;-…<-„:-ƒ;-„<0‚<0‚<2‚>3ƒ?4…A8‡C:ˆD9”J=—H; H>¬KD²KF³LG²SM®`V­sg¦�q�Œz‘Ž{‚‰ws€ocqbXcUNRDMN@HI;DD8@@4:</46+/1&12*01+/0+/0+./*/0+//-//-//-.0-//-//-//-//-//-//-00.00.00.00.00.00.00.1/26+98,:8/4;63HE@_^Yzy~‹�š�”±’—¿�•Ç„ŽÃ}‰»{‰¶|‹²}Œ«}ˆšyƒ�q~‡o~ƒn~~pƒyŽ…ƒ™�‹¡”‘¨˜�¥”ˆ�Œƒ•…|�{j{k\hTXX@]Y>_[@-.0-.0-.0-.0-.0-.0.......,-.,--,*-,*/+*.*).*'.*),*+,*++)*+)*+)*+)**()*(),*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+-+,-+,-+,-+,.,-.,-.,-.,-.,-/-./-./-./+*0,+1-,1-,1-,1-,0,+0,+0,+0,+0,+0,+1-,2.-3/.40/91/:20;31<42=31=31=31=31>3/>3/@2/@3-A4.C60D71E82F93H94I:5J;6L:6N94Q83T50^72e60o6/x8/‚90‹;2�<2”=3š@7›?4›?2Ÿ?1¥A2®B5¸C9¿E:ÈH?ËH@ÊJAÃH@¶F:ªB5žA0™@.šA/™@.˜?-—>,—>,™>,™=.˜?/–=-”=,“=.�=-Ž=,Š=+Š=-‰<,‰<,ˆ;+‡;+‡;+„;,„;,ƒ;-€;,�;/€<1�=2�>5ƒ@7ƒC:‡D<ˆE<”KB–H>žG@§JE®LI®QL­]Vªj^§€o¡�yšš‚�›ƒ†•€z‹xm{lbm_SZJQUFKO@EI:@D6;=057,13(01)/0(./*.-).-)/.*0/-0/-0/-0/-0/-0/-0/-//-0/-//-10.00.10.00.00.00.00.3.27,:6*83-1961HJ?bfX{€z‹�“�“£�–°Œ•¶ƒ�²��¯…™²�¤¶’¨³”§«‘££Œ ž‡ž˜‚™�}˜‰�œ‹ˆ£�”°š˜±›”­—‹£�…›†~‘}k|iXfOSU=ZV;^Z?+/2+/2-.2-.2-.0-.0..0..0------.-+-,*-,*-,*,+),+),*+,*+,*++)*+)**()*()*(),*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*+,*++)*+)*-+,-+,-+,.,-.,-/-./-./-./+*0,+0,+1-,0,+0,+/+*/+*0,+/+*/+*0,+0,+2.-3/.40/:12:12;31<42=32=32<20<20>31=2.?1.?1.@2/A30B41C52D63C84D95E:6G96H94K84N50X72_60i70s80}:1†<1Œ>2�>2—@6—?5—?5—?3œ@3£C5«C6³E8ºE;½G=¾H>¹G=°D8¦A5œ@1—@/—@-—@-–?.•>-”=,–=-–=/–=/•<.“;-’</Ž<.‹<-Š=-‰;.ˆ<.‡;+‡;+†:*†:*„;,ƒ:+‚:,€:.€:0€<1€=4?6‚B9ƒD=‡F@ˆH?‘KC’H?šGA£KG¨PL©YR©f]¦vh¢Œw›™€•¤‡Œ£‡†�ƒ~‘}t‚qjue\eTV_LNUCEL:?F6<@27:/58-12*/0(.-)/+(/+*0,+1-,2.-1-,1-,1-,1-,1-,/.,1-,/.,2.-0/-2.-0/-0/-0/-0/-2-18,:5)51++66,GL8aiRzƒpˆ’‡�—–�š �™¥ˆ–£‡›¢�ª©›¹±¡¿³š¶¨š¶§›·¨˜·¥‘±œ‰«’‰«��±”’³–•´•�­�ˆ¢…„›�{’xgzdTbIQT9VU9YX<,03,03./3./3./1./1//1//1....../.,.-+.-+.-+-,*-,*.,--+,-+,,*++)*+)**()*(),*+,*+,*+,*+,*+,*+,*+,*+,*+,*++)*+)*+)**()*()*(),*+,*+,*+-+,.,-/-./-./-./+*/+*0,+0,+0,+/+*.*)-)(0,+0,+0,+0,+1-,2.-3/.40/:12;23;31<42=32=32<21<20=20=20>0->0-?1.@2/A30?40@51@72@93A:4B94C84F74H5/Q51X5/a6/l8-v:/€</†>0‹=1“?5•>5“?5“?4•B4šB4 C4¥D4¬B5°D8´E:±E;ªB7¢@5š>1–>0•?.•@,”>-“=,“<+“<+”<.”<.“;-’<-�<.�;-‹<-‰<,‡;-‡;-†:*†:*†:*ƒ:)ƒ:+�9+9-9-€<1<3?6€A8‚C<…F?ˆIBŠICŽIB�G@—HD OK§VS§`Z©pe¤ƒrœ”}– …�¨Šˆ§ˆƒ „~–~z‡uq|kdp\]iUR^JJS@BK:>E5<@29<134,22*1-*/+(/))0**1++2,,1++1++1++1++1++0,+0,+0,+1-,1-,1-,1-,/.,/.,/.,2,.8*75(13+(56&EK1\gGu‚d†“yŽ›Š‘ž”’žš�ž˜�¥——´ž¢Å¥©Î­¥É­§É°§Ë±¤È¬—¾Ÿ‹³‘ˆ°��²�Ž²Ž�²��¬Š„¡‚€™{wŽrdx]R`FNR7QQ5SR6,03,03,03,03./1./1./1./1000//////////.,/.,/.,.-+/-./-..,--+,,*++)*+)**()+)*+)*+)*+)*+)*+)*+)*+)*,*++)*+)*+)**()*())'()'(+)*+)*,*+-+,.,-/-./-.0.//+*/+*/+*/+*/+*.*)-)(,('0,+0,+0,+0,+1-,3/.40/510:12;23<34<34=34<21<21;10<1/<1/>0/=/.>0-?1.@2/>3/=52;62;83<94=:5>93@72C60G4.O4-Y4,d5+n8,x:-;.…;.�<4‘<5�>3�@3�A2“B1—A2šA1 >1¦@4ªB7ªB9¦A7Ÿ>5˜>3•?2’?-’?-‘>,�=+‘;,’<-’<-’<-‘;.‘;.�<.�;-Š;.ˆ:-‡;-„;,…9)…9)ƒ:)ƒ:)‚9*�9+~8,~8,�=2€=4€@7�B9„E>…HCˆKFŒMFŒICŽGA˜JH¡SO¨]Z©hb©{n¤Žy™œ�’§ˆ‰¬‹‚ªˆ€¡„€˜€~‹wx�njwcdr[ZgSQ]IKTCEL<@D6<?467/44,3/,1,)1()1()3*+4+,2)*2)*2)*2)*2)*0**0**0**1++1++0,+0,+0,+0,+0,+1+-6)25)-4-%46!BI'Wc;q€W‚’m�œ}� †’¡Œ� Š�¨‰•¸Ž È”¤Ð�©Ó«¨Ò®§Ñ­¡Ë¥“½—…¯‡‚¬„‡°†ˆ®…Œ±ˆ‹«†�Ÿ{{”tqˆk_sWM]BIO3JM0KK/./1,01./1./1.0/.0/.0////00.00.00.10.0/-0/+/.,1-,2.-1-,1-,0,+/+,,*++)*+)*+)*+)*+)*+)*+)*+)*+),+)*+),+)*-(,*(),())'()'()'(+)*+)*,*+-,*.,-.-+/.,0/--,*-,*/+*/+*/+*.*),(',('1-,2-*2-*2-*3.+4/,50-61.;31;31<42<42=32<21;10;1/<1/<1/=/.=/.=/,>0-?1.=2.=4/=60;81::29:49:4;81?61C2+J1,T1+^3,g7-o9-u=.{=0‡=4‹=3‹?2Œ@2Ž@3�@1’?1•=/˜</œ>2¢B6¡C7žB7™?4–>2•?2”>/“=.�=-�<,Ž<.Œ=.Œ=.�>/Ž<.Ž<.‹</Š;.‰;.ˆ:-†:,„;,‚:+‚:+€;+~;+~;+|:,}9,|:.€>2�>5€@7‚C:ƒG?†KCˆOH‹OGŒHEŽHF“OL˜[Vžg`�uiž‡už˜€”¡…�ª‹‰¬‹ƒ¨‡€¡„~›|�ytˆoj~ccz^]qXWfOO\HIRAAH8>@399/85.7/,3+)2()2()3)*4*+0*,/+*0*,0**0*,0*,2),2),3*-1+-1+-1+-0,-0,-0,-1+/5*05+,4-%46!?F%T`8n}R��g‰™t�žz�¡~ˆ¡zŠ§{‘¶‚›ÄˆŸË�ŸÉ™žÇ��Æš˜Á•Œ·Š‚­€…®‚Œµ‰„«‰®…‰©‚~œxv�olƒfZnSJZ@GM3FJ1DF./0+.0+/0+01,01,12-12-21-43/43/43/62/51.41,3/,4/,50-50-4/,3.+3--1-,0,+0,+.,-.,-..0--/,,.++-*).))+.)/.)-/(/.)-/)-.)-.*+.*+/+*/+*-,*.-)--+./*./*./*------.,-/-./.,0/-2.+2.+2-*4,)5-*6.+90+:1,;2+;2+=4-=4->50>50>50=4/<3.<3.=2.<1-<1/;0.=/.>0/>0/@1.A0)@1*;4*77-39/39/560:3-?-+F*'L)'S*&Z/(`5,d<0k@0yA2€A0†A2‹B3�@3“=0“;/’8-“;/”>/”A1•B2”C2’A0‘>.‘;.—:2—:2’;1�>1ˆ?0„?/‚?.ƒ>.ˆ@1ˆ?0Š>1Š<0‰;/ˆ:.†:-ƒ;-{9+~@3w<.q7)w=/w=/v;-}?2{;/‚@4ˆE<ŠJAˆLB†MB„PE‡NE‘KI—SR”b[‘ocŒ}j‰‰q‰–|Šž‚�¤‡�¦‰Ž§Š�§Œ‡¤ˆ~Ÿ‚uš{o—ui�ib„_[zXZsU[mSWeNPWEJK=C?6@93;0.6**4(*3'+3'+1&*,*-)+***,*(),'+.(,1(-2'-3(.3(.3(02(00)00)0/*1/*0/)-1++0-(//#<?,V\BpzX‹g�œu‹�sŒ¡vŒ¦wŠ¨v‰ªuŽ´{™¿†ŸÁ��¾�–ºŠ�³ƒˆ®}…«|ƒª{…ª~ƒ¨}ƒ¦~€ž|y”ur‰mh|cYkUN[GIP>CH4?B/01+01+01+01+12,12,21,32-43.43.74/74/63.52-50,50,7/,7/,6.,6.,3.+2-*1-,0,+/-./-./-0..0-,1+*/)*.)(-.)//(//(//(//)-/)-.*+.*+/+*0,+.-).-)./*./*./)/0+.....0....../.,/.,2.+3.+5-*5-*7.)8/(:/)<1+<2)=3*>5,>5,>5.?6/>5.=4-=4/<3.=2.<1-<1/;0.=/.>0/>00@1.C0*C0)A2+>3-:5/94.:2/<1/?-+D*)I*(N+'T/)Z5-`;2e=1pA1x@/€A0ˆ@2�?4’>4”;3“;1�9.‹</‹=0Œ@0�A1�A1�B2’@2•;2•;2�<2‹=0‡?1ƒ@/�?/‚?.…@1†>/‰=0ˆ<.‡;.„;,ƒ;-€;,x8,x<1t:.n9+t?/s>.r8*u:,}=1ƒA5‰E<‹H?ŠKB†MB„PCˆOF‘JH—SR’f]�wi„†p€“w�ž„¥†ˆ§ˆ�§Š�§��§�Š¤‰�Ÿƒwœ}o˜ve’iaŒaZƒ[Y{X\wXZnSSaJNUCFH;C@7<737/-3*+2)*1(+.(*,*-**,+),+),-(,/)-2(02(04)13)13)21*20+2.+2.+4.+22-13/.0,)--%8:-SXDox]€‹i‡–o†™l†žnˆ£p†¤p„¥pŠ­w“¶€™¹‡–¶„“³��¯|†«x‚¨w‚¨yƒ¨|ƒ¦|„¤�{x’uqˆnh{eYjWN[IEL<@D5;=/12,12,12,23-23-23-43.43.54/65085085085074/72.72.80-80-7/,7/,4/,3.+1-,1-,1-.0./0.1/-0..0,,.+*/+)./(/1'//)-/)-/)-/)-.*+.*+0,+0,+/.*/.*0/*0/*/0*/0+//-///0./0./0/-/.,1-*2-*6.+7/,90+:1*<1+=3*>4+?5,?6-@7.@7.@7.?6/>5.=4/=4/=2.<1-<1/;0.;0.<1/<1/?1.C2+E0+H/+K--L,/K+.I*/E+.A-,@.*A.(F/)N/*X1*b3-g5.j:,o;-x<1�=4‡<6Œ<5�:4�:4Š;4‡=2†>2…?3†@4ŠB6ŒB5�C6Ž?2�=/Œ>1Š>.‡>/†?-†=.ˆ<,‰;.‰<,ˆ:-‡;+ƒ;,<)~=+{<+}=1z<1v:/v<0x@1x@1v<.v;-?5ƒC9‰F=‹H?‰JAˆLB‡NEŠNF’KI˜TQ–f\‘wh‡…n‚‘t�{�¤ƒ…¨‡ˆª‰�ªŒ�©ŒŽ¥‰ˆŸƒ�š}y˜xi–ma’e\‹a\†`^ƒa[|]UrVQgPHYGBPA:D93:2.3,+0**,)**(1(+1&*1&*1&*0'*1(+2).1+//*.-+.,+0+,0)-0(-0(-1)-01/23/./+(**"57*QVBmu]|Ši€�k~“j™lƒŸo€ o~ m‚¤q‰«x�¯|‹­zˆªx„©v‚¦v€¦u€¦w€¥y€£y‚¢}}›yw‘tp‡mh{eYkUN[IDH9>@399-23-23-34.34.45/45/54/54/761761:72:72:72961:51940:2/:2/91.80-50-50-3/.3/.3/03/01/01/20.1..0--/-+.0)02).0*.0*.0*.0*,/+,/+*1-,2.-0/+0/+10+10+01+12-11/1111/010.10.2.+3.+3.*91.92,;2+<3,?5,@6-A7.A7.B8/A8/B90A8/A81@70>50>50=2.=2.<1/<1/<1/<1/=20=2.B3.E2.M//R+0W(0U&.P'/I).C/.<1+;2+?2)H1)S0*_.*d/)i9/k;/u<3}<6†<9‰;9Š;7ˆ:6‰>9†?9†B9…B9…B9†B9‡A7ŠB6ˆ@1ˆA/ˆA/‡@.‡>-ˆ<,‰<,‰:+‹9+ˆ9*‡:*ƒ:)<){=(x>(x>*‚>3{7.z7.z<1v;-w=/|A3{@2€B7ƒE:†H=ˆI@‡KA‡MBˆOFŠQHŽSK�\QŽh[‹tb…g€‹m~–vž|ƒ§ƒ„©‡ˆª‰‹ªŠ‹¦‡‡ ‚ƒš~}˜yq•oj“if�gfŒeg‹gd†e_~_\v[PhRK^KBPA8D62:/.4*,/(+*%2&(5%(4%(2&(1&*/)+/+,.,-++-*+-*+-(,-(,/',/',/*+-/+,1++0+(/,%99-PTCiqYvƒe{Œhy�h|•kœn}žoyžk{ m€¥r‚§t€¥q}£p}£p~¤s~¥v~¥x}¤x|¡x}Ÿzz™wu�qn‡je|bXlSN[GDF9?=1:8,45/45/45/560560671761761872983;83<94<94<94<73<73<41<41;30;3083072/61.61.5106216213123121/00./1-.2,04+.4+.4+.2,.2,,2,,2,,3/.3/,3/,40-21,21,32-32.22022032032051051.61.61-;30<5/>5.?6-A7.B8/E8/C9/E;2E;2E;2D:1C90B8/A60@5/>3/>3/=2.=2.=20=20>31>31@51F42M02T,4X)3W(2R)1K,1B30:6-77-:6*B4)M2)X/)^/)f:1j;3s<7z=:‚<:†<;‡;;†::‚;7�>8�@:‚C<ƒC:ƒC:…@9„@7‚C2‚C1ƒB0„?/‡>/ˆ<.‰:-‹9-Š8,‡8+…9+‚:+~=+x>*v?*x=+�9-|/'‚8/„>4w4+t6+}A6}C7E:€G<‚I>„KB„KB†MDˆRH…WJ}`N{iQ€pY�u]‚|bƒ†i†“uˆ�|„¢~‚¦‚‚ª…‚¬†�©„¥€{ž}z™wz’p{�lyŒlwŒku�msŠmm„gkd`rZ[hTR[JIM>@@4;7,70&5*$6('5''3''1'&.((,+)++)+-*(,+(,+*,+*,+++-,*-,*-.*+,#$3+(50*85,BC5UZFfpWn}^tˆeqŠbu�ex˜ivšjs™htšiwŸkz¢nx lwŸmx n{¢s{¥u{¥w|£wyžuzœww–tsŽom†hd{_WkPN[GCC7>:195,560671671671782782872983983:94=:5>;6>;6>;6>95>95?74?74>63>63;63:5294194184195495484384343151240/6-06-.6-06-.4..4..4..4/,40-40-40-51.32-32-43.43.43/442542540841850:51:5/>71>7/@7.A8/C90D:0G:1H;2G=4G=4F<3F<3E;2C90B71A60@51@51?40?40?42?42@53?53@72D63I35P16T/6S.5P05J22C52=90<:.=9-C7)I6(Q3)W2)]2+d3,l50v64}77‚87ƒ77ƒ77~75}:4}<6~?8€?9ƒ@8†?;…A8�C4€C1�B1ƒ@0…=/‡;.ˆ:.‡9-…9,ƒ9,‚:,<,|=,y>,x?,|=,‡4,‹2,¡LE¨XQ‹A8|90‚F;€K=yH:zJ<{M@|NA}OB�QE‡UJ‚_LrkOptS|uX‡w]�yb˜iŸ‹s ”z—–z‘œ|‹£ˆ¨ƒ…§‚€¤~{Ÿy~™vƒ‘p…Žo„�p�Žp�r{ŽpwŒmtˆlj~cfv\_hSV[GOM>GA3A7+=0';,'9+(6+)3+(/,',-'+-().().(+-(-,*/+*3)*4(*7'*7'(3($<3,E>4IG:QR@^cMgqVjyZoƒ^k„\l‰]p‘bq•eo•do—eršfuŸmrœjq›itžnx¢rz¤vy¢vyŸvvštw™vu”ro�kk„fc|^UlON\ECC7@91;4,782782782782893893:94:94:94;:5>;6?<7@=8@=8@;7@;7B:7B:7A96@85=84=84<73<73<73<74<74<74;639529338308/09/.8/080.80.80.61.61-61-61-52-63.63.63.74/54/540651841952;74<94=84@93@70A8/C90D:0G:1H<0I=1J>2J=4J=4J=4I<3F<3D:1C90B8/A81A81@72?61?61@72A83A83?74@85B86E:8G96I:7H96H94E80E8/E9-E9+G9,I9*K9+Q7*Z/&d/'n3-z63ƒ98‰;9‹;:‹=;‹A>‡@:ƒ>7<4:3‚<4ˆ=8‰@9ƒA5�B3‚@2ƒ?2ƒ=1„<0�;/€</=/}>/|>/|>/|>/}>/=/†9/�1+£;8ÎkfÛ~y­ZT�D=ˆLAN@tJ<rN>pQ?qR@tSB{WG‚[J~eOmsOpzU„y[˜u_©oc¶mf¾qk½wo»‚w±Œz§—~žž‚–¡��¡~…�y…™vˆ‘r‹�rˆ‘r…“r�”t}•sx“pt’pm‹ii„edx]]kRV^GMP;ED0B;+@3+?2,;0,70*30)00(./)./)01+30+7-,;+,?),D',F%,D'+F5-LC4VP@[XE`bLgmSjxWj{Wl‚[g‚Wg†Zl�^o“cn”an–bršfs�kpšjo™irœnvŸsy¢xxžwvšvs—su—vs’rn‹lk„db{[UmMN\CGH:E<5@707827828938938939:4:94:94;:5<;6?<7@=8@=8@=8A<8A<8D<9C;8C;8B:7?:6>95>95=84>95>95>95>95=85<73:52<41:0/:0.91/91.91/91.91.72.61-61-63.63.63.74/74/74/540651952;83<94?:4B;5C<4A8/B:/D:0E;/H<0I=1J>2J>2K>5K>5K>5J=4F<3E;2C90B8/B92B92A83A83A83A83B94A:4@85A96B;5D=7F=8G<6K<5N;4M6.O7-Q6+Q6+R8+P9+P9)W7(f6*r6,~;3‰@9•D@›HDŸJGŸLF QJ™LDŽD;†<1„7-†8.Œ91�=6ˆ>5‡>7†=6…<5…<5�=4}=3z>3x@3vA3x@3z>3<3ƒ:3‰84’42˜(&³=;í{zþ•’Åhc–G@‰K@xH:nM<jQ=fT>hV@lWBt[E`L€hP{rQ…vUžt\´l^É__ÓV\ÙQ[×T\äouÙzx̆~À�€µ•€«•}£’xŸ�vœŠrœŠt™�u•�u‘‘u‹‘u†‘s‚‘r|‹ny†ju{anpZgbN_TBUE5R</O4-N2.I0,D/*>/(9/&7.'6/'92*;0*>/,B-,G*,I),L'.I)*TB6YQ<d^HgeLilOnuVm{Xl~Xn…[hƒVg‡Xm�]p•bq–br™dvœit›lq˜ko–jr™mw�vzžxxšyt–uu”uu”us�roŠkjƒeb{[UmMM]BMN@KB;F=69:49:49:49:4:;5;<6=<7=<7=<7=<7@=8@=8A>9B?:D?;D?;E@<E@<E@<D?;D?9C>8E>8E>8B;5B;5B;5B;5B:7B;5B:7A:4A75A83A75@72@64>71>71>71<71<71;60:5/96185074/74/961961961:72<73>93A:2B;1C:1D<1F<2J>2L@4L@2N@3N@5N@7N@7L?6K>5J=4H;2F<3E;2B90B90@91?80?82?82@93@93<5/MD?M@:K:3T?:S73O0+a<6\3-`5.`4+^/'^/%f6,m=1q=0ŠJ>�G;—I?ŸKAŸF>›>7›>7žD;¦ND¥QF©WK«YK¥OB™A5—;0›>6“98’:9Œ65†52†84„?8|@6r>1rB4oA2q=0w;1‚72Œ43’-1š',ÈHIèbaÅEDÍWUáyv«VO†F<„VFnS>i[AibFjeHj`En^D|dL�jT’bL´l]ÑnhÛX]à;Lç-Eò+Hõ0Lö@YìI\ä[eånrâ{|Û€}Ûƒ؉‚Ã~w½�wº„x¶†x²ˆx­‰y¨‹y¥Œxš€o™~m˜xi“oc�bY„UM{IBxA>u:<i.2d04b14P&'G%#L0-M51G4.E2,D1-F1.F.,G+*K-/Q51YH6`V;f_CgdEkoNu|Zu„]o�YlƒWl‡Zp�_r’at”cu—dw™fxšhq•is˜ot–qr”ou”ty˜x{˜|y–zx•ys�rm‡lk„gi‚dazZTmMO_DNOASJAQH?9:49:49:49:4:;5;<6>=8?>9>=8>=8A>9A>9B?:C@;D?;D?;FA=FA=E@<E@<D?9D?9E>8E>8E>8E>8E>8D=5D=7C<4D;4C:3B92B92B92B92B94A83@93@93>71<71<71;6096185085074/961961961:70<71>:1A:0B<0C;0E;/I=1J>0L@2L@2N@3N@3N@5N@7M?6K>5I<3H;2E;2E;2C:1B90A:2@91@93@93A:4B94E:4J71L.,V..j68u9;{;;…EC„E@…IA…IA‡H?ŽIB™NH¡PL¢OG“D7”B4šB6 D9 >3œ7-œ7-Ÿ=2£C5œ@1™@0 D5¤G6§F6®H:¶NE´LM¯HL¥@D™9;‘98�>:ƒ@8x>3o>0zI;…LA„?8…-,’).³9DÒLWÜKNßMMÊ>=À@?Ð`\³YQ„@5ƒUEyaIe[@_^@ihIslOvgJ~dK–cN¼dXÖ_[æUZëANò,Eÿ&Eÿ&Hÿ'Iÿ.Nù3Nò<RðI[ëP^åQ]ãS]àXbÛbkÕflÔioÒmqÏqrÍutÉxwÇywÊ||ÉyzÈvxÇqtÆkrÅfnÅakÃ_kÅ`n³Sa¦LXšGQŽEL†EIm69HT1-M0*I2,J6/L92K:2O>7UE8[O7cZ;hbBifEnrO{ƒ^z‰bt†^r‰_p‰_p‹`r�`u’dw”dw”dw”fs‘kx–t|—x{•x—}„œ„…�‡ƒ›…�™ƒz’|rˆqm„jh�d`y[TmMP`ERSCVPDUOC8938938939:4;<6=>8?>9@?:@?:@?:C@;C@;C@;C@;E@<E@<FA;FA;FA;FA;E@:E@:G@8G@8JA:I@9I@9H?6F=6E<3F<3E;2D:1D:1D:1D:1D;4D;4D;4C<4?80?80<71;60;60:5/94.94.96/96/:5/;7.>7/@9/C:1C;0E;1F<0I=1K?1M?2NA1P@1P@3P?5M?4L>3K=2I<3H;2E;1E;1D;2C:1B;1A:0?;2@<3A=4D;4VE>U4/j25ŽAG­LWÃUbË]fÆ_c›B>—I?–LAœND¬QL¾URÌOSÄLM¥@6�B0 @0¥?1¥;.£7+¤8+§=/«E6¡=-˜7$�:'¥>-¬A1²C2·E;§02¯:B¹DL½LRºMR­KL–A>‚71{8/v6,x2*„1-œ37¹=GÕGWèM[âFIÏ53Ð:9½31ÈNKÆc]•K@Œ[J}cLk`DgdEtoOpS�gL�hN°hYäbbõQZôDQö7Iþ2Kÿ2Nÿ/Mÿ+Iÿ2Oü0K÷1Jõ5Nó5Nð2Kó3Lô8Qò@ZïC[îF_íIaëNcêQfèTjèVkåSjäRiäPjäNiçNlëPoòSsóVuþg†ñ`}å[uÕUlÎZmÉdr¨R]r-2^&'Y0,W:4Q?5C;.:8)>B3LN9VO3aV6e_?heDquT†e€‹ixˆdxŒiu‹etŠdr‹cvŒey�hz�iz�j€•v„›Šž…� Š’£�—¨–˜©™–¦™‘¢’‡˜†z�ypƒmhc_vYTmMPcGTWDYVGYVG7827828939:4;<6=>8@?:A@;BA<BA<DA<DA<DA<DA<E@<E@<GB<GB<GB<GB<FA;FA;HA9HA9LC<KB9JA8I@7H>5G=3F<2E;1D:0D:0E;1E;2D;2E<3E<5E<5@91?80=82<71;60;60:5/:5/:70:70;7.;7.?8.@9/C;0D</F<0G>/J>0K?/M@0NA1P@1O?0O?2O>4L>3K=2J<3H;2E;1D:0D;2D;2B;1@<1@<3A=4B>5G=4S81[*&Š;@¾T`ÙTeæTgãUcËJO 41—;0’?1™A3®F=ÅIGÕBHÎ?C±B7§F5©B3¬@3¯?3°>3³B4µE7²G7©B1£<)¤;(©;*«:*®6&­3(¶97º9=½7>¼7<¿<BÆJLÄTS¾XS™?7Œ4*‰,%�61ÃHKÜOWãBRÛ3@Þ:9Í.*Ó84È74ÆE@È]U©WK‘XGwW@veI~pS€qR„fL�cL©o[ÕuiîSWüDP÷;Jù:Jþ=Nÿ@Qÿ;Où4Hÿ?Sÿ9Pû4Ký3Ký1Jÿ/Iÿ0Nÿ6Sÿ3Tý4Vý6Wú7Wù:[ø=\÷>`õ@aøCf÷Agõ?eõ?gø?hüCmÿErÿIsÿOwÿQuþUvòSqçTpåaxÖfv·Wb~17j0/Y2+P8,G>/>A.;D/?G/QL/_T8g`CokN||`ˆŽr‡‘v~‹q|Œq|Žt|�sz�qzŠm}Œoƒ’uˆ–|•£Œœ©•£®�¦±¡¬´©±¸°°·°«µ­¡«¢•¡•ƒ‘‚s„qh|c]tXSlNQdHRZCWYDYZH671671782893:;5<=7?>9@?:BA<BA<DA<DA<EB=EB=FA=FA=FC<GD=HC=HC=HD;HD;JC;JC;LC:KB9KA8J@6K>5I=1H<0H<0E;/E;/E;/E;1E;1E;1D;2D;2A:2@91@91?80<71;60;60:5/;7.;7.;7.<8/?8.A:0C;0D</G=1G>/K?1M@0NA1NA0P@0O?0O?2O>4N=3K=2I;0G;/D:0D:0D<1D<1B;1@<1A=2A?3B?6K=4\50w32­LSÛ]iæM_å?S×6E·&+¨1+�7(“;'—<)¬@3ÃD=Ô>?Í;<±?5¨B4ª@3­>3²@5¸C9¼G=»I>¬>/ª?/¨=+¦;)ª9)°:,·;/¼<1ÊFAÏEBÐ>?Ð79Ö7<ÞBEâLNÞSPÄC>ÇLDÍTLÔSMÙKJÜ@Cà5>á27Û4.Ú7.Ð1+Ï:4¿84³@9µ[P–TD{N9‚bI†kP…aG’[F®gUËrdçmhêEKò:Dê9Cç<DçBHçFKèGMêFMõKTóAOó8Iö5Hù2Gþ0Gÿ3Mÿ8Rÿ9Vÿ9Xÿ:Xý:Zþ;]ÿ=_ÿ?dÿAeÿCiþDkþDmüCmüCoüCoýCrþCpÿGpÿ@hÿGmÿMpòIjéOkå[rÙcs½YeŽ?D`&$Q+"V@3VO=IL7@D-RJ3eYCvkW€{gŽ�y™œ‹•œŒˆ“ƒ‚Ž€Š–ˆ�›�Œ˜Š‡’‚Š“‚—žŽ¤¨™²¶¨¸¼®¿Á¶ÀÁ¹ÂþÅÅÃÃÃþ¾¾°µ±¢©¢�—Œy‡vi{e\sYTmORfJQ[BUZCW\F560671671782893:;5=<7>=8@?:@?:C@;DA<EB=EB=GB>GB>GD=GD=HC=ID>IE<IE<LE=LE;LC:LD9LB8KA5L@4K?3J>2J>2G=1G=1F<0F<0E;1E;1D;2C:1A:2A:2@91@91=82<71<71;60;7.;7.<8/<8-@9/A;/D</E=0H?0J>0L?/M@0NA0NA0O?/O?/O?2N=3M<2L;1I;0G;/F90C9/C;0B<0B;1@<1@>2A?3B?6N;4n95—FEÄY_ÛWbßCQÞ8FÔ3;½++­1'¡8%˜:!—9 ¤;&¶>.Ã:2¾71§7,¢:-¥9/§7,¯:1¸A9ºE<·E:®>2­?0§<,¤6'ª6)¶>0ÃD;ÊG=¾8-Î@6âE@ìBBô=Aó9>ë27Þ.0Û75âGBèSLåNGÜ=9Ù10ã/2ê67Ù1(Þ</Í,"Ð71¾3,°7/Ég\¸l^ŽWC‡ZCƒV?‰R=©\LÎlaágbåUUðEKñ=Fä;@Ù=>ÒA>ÒGDÝOMëWWíNSïDMò9Gö5Fø1Fø.Dû1Gþ7Nþ8Sý8Tþ9Wÿ:Xÿ;\ÿ>aÿAeÿCkÿ<fý<hý>jý?mú>nø>oô=mö:jÿHrÿ;aÿ?dÿJpÿKn÷NmïUoãYpèr‚Ä`j˜INt:8\6-P9+PC2UM:^QAreUˆ|n•�‚Ÿ�‘¦§Ÿ¡¦Ÿ–œ˜•ž›ž§¤¦¯¬¦¬ª¢§£¦¨£µ´°ÂÁ½ÌÉÂÑÎÇÕÐÌÕÐÍÔÎÐÕÏÓÑËÏËÆÊ¿½À°²¯˜Ÿ˜€Œ~m~kau\XoSVjNS`FV`EX`H560560560671782893:94;:5>=8>=8A>9C@;DA<FC>HC?HC?HE>HE>ID>JE?JF=JF=MF<MF<MD;ME:MC9MC7NB6NB4N@3MA3JA2JA2I@1H?0G=1F<0D<1D<1C:1C:1A:2@91?82?82<71<71<8/<8-<8-=9.@9/A;/D</E=0H?0K?1M@0NA1P@0P@0O@-O?/O=1O=1M<2L;1I;0F:.F90C9/B:/A;/A:0?;0?=1@>2@@4Q:2r1/©KLÁRX»>D½06Â03Á//º1)ª2"£9#�>"™;›< ¡=#¥;%¤6%ž6) 8-£6/¥4,®92·@:¸C<²@6±B7­A4¨</¦8+¯:0¼C8ÄE>Å@7Ã>-Ï>-Ù5+á*&í"%ú%+ÿ+2ÿ.4í)*ç.+á3,ß4,à3,ä3-é0-ç2+Ú2%Ö6&Ð2&Ï7,Å6.½>7È`UÒq¦eS�VBŠM:£WIÈf[ßd_çRTèCIõFMîAEß??Ô@<É@8Æ?9ÑFAãMLïJPô@Kù9Hý7Hü5Hø3E÷6Gú;Mû9Qü7Rû6Rý6Uÿ6Xÿ9]ÿ;aÿ<gÿ>kÿ?mÿ@qÿBtÿBuþBuüBtùBpÿFmúAaþEeÿJlþKkÿVtÿ_zõ]vòj~æp~Ómx¯]aƒCAg6/hB7sUJye\Œ}v£–�­£¡²®«¶¶¶²¶¹ª¯³°·½¶½Å¾ÁÊÂÂÊÄÃËÉÆÍÕÎÖÝ×ÛâÚØçÝÛéÝÝæÚÜåØßä×ààÔÞÙÏØÊÅ˺ºº¢§£ˆ’‡t‚qexb\sY[oS[jM\iK]gL561561561560561671983983<;6=<7@=8B?:DA<FC<GD?HE>HE>HE>HE<JF=KG<KG;KG;NH:MG;MG;NF;MD;MD;MD=NC?NC=MC7MD5LC4JA2H>2G=1E=2D;2C:1C:3A:2A:4@93?74<73<71>7/>7/<8/=90>:1B;3C<2D=3H@5I?3L@2MA1NA1P@0O@-O@-L?/L>1M<2L;1J91I81E80C90@9/>:/;;/<<0=>0>?1A?0S8-‡<7Ä\[ÊZY®86«.*²1+¯/&­2#¬9&¬B*ªF,¦D)£A&£A&£=$Ÿ9#¡:+£;0¦:0¦7.­<4¶C<·E;±A6«=0©=0©;.­=1»E9ÇMBÅE<¼8,¿6$Î;)Ý;.æ3,ñ+*ù(+ý&+û%'ø**õ.+ï2,ë4,è3*å3)ä2(à3%à:*Ñ2Ü?0Ï7*È8-Â?5³A6Ñl`¿gY¥RB¨NCËcZçhbæRRêAFôCKë>Dá;=Ö<<Í@9Á>4¶8,¹6.Ä94âHHé@Eð:F÷;Iú:Iø8G÷;Iû?NþAUÿ?Vþ<Tþ9Tÿ8Uÿ9Zÿ:^ÿ;cÿ>iÿ?mÿ>oÿ>pþ?rú@rø@t÷Aq÷CjÿMmÿSs÷JhìAaýTsÿg…ÿe‚÷]wï_xèg{ÜpÆqx§gg�^X{ZQ—~wª™’¾°­Ä¼ºÆÂÃÈÇÌÇÇÏÃÃÏËÊØÍÉØÑÊÚ×ÎßßÔäæÛéêàëíãëïåæòèæôèèðäæîáèíàéêÝçäØâÑÊÑÁ¿Â««©‘–�z†xlzibu_`rZcqWcpTcmT21/320542651875984984984;:5<;6==5??7AA9CC9EE=EE=HH@HH>HH<JH;KJ8LK7ML8OL9NK:JF:HE<JFCNIMNIPICOE>ENE>LC4J@4KA5MC9LB8F=4B90E<5C:5@93@85@85@86<74<42C97B94A96>95?:7>:7?;8@<9B?:D@7G@6J@4L@0O@-O@+L?,G@.G@0I;0K81M53L76J88F;9C@;6904</9D4:A/69$?=(bC1¼j^¹JA¦7,¥5)§7+¥5'¥7&¨=+£8$¥<&ª@*­@+¬=)­:'®6%«6%¤6'¡7)£9+¦<.ª@2¬B4¬B2¬A1©>.«=.­<,³=/¾D7ÅG;Á?2º2&Ç9+Ì8*Ð8+Ö8-Ü8.â5.é3/î1-ò/-ô.-ó0,ñ1,ê5,ä8,Ý:)Û:(Ü9&Ô3Ð3"Ï8'Æ7'¾6(ÂB5ÏSIÈMEÐTLá[Xî\\ðPRí>Cò9?÷?GêAFßCDÖBBÈ?9»<3±;/°</²<0ÊKBÔKCÜFEá>?ë<A÷?Gý?Iù;GþDQüCQþDRÿDUý@Vü<Uÿ<ZÿBdÿFkÿAjú;gõ:g÷=nùCtýH{ÿK{ÿMyÿMtøPsðOpçMiêMjüYxÿgˆÿ_ƒõGl÷Vxè_{Ã^n½z�‡a^�xp§‘„»§œÑÄ»ÝÖÐÝÜÚÜÛàÞÜçâÛëæ×ìëØîòØñöÝó÷âó÷çôõëôôîòùôñùôðùñïöíîóèìîãééÞäæÝâÖÐÔÍÉÊ»º¸£¥ ‹�‰yƒxr}op{k}†sx�lv}k0./1/0320542653873873984;:5;:5==5??7AA7CC9DD:EE;HH@HH>HH<JH9KJ6LK6ML7NM8LK9KI=KJEPPRWU`YXjVTjSO`QHKMD?H?:F=8G>9H?:F=8B;5D:8A96?74?74@85@85=85<73=31=31=52=85=96>:9>=;>=9B?:D?9G@6J@4M@0O@-P?+L@*B?,B?.F<0H:1J65I56F35@65A=::=6;>5@@4E<-J9)^B4€L>®QB«;- 2#¡6&£:'ž7$ž9%£>*¡;%§<(®?,³@.¶=,·9*¹7)¶8*ª9)¥:*¥:*¦<,©?/«A1«B/«@.®?.«:*²=,ÀF7ÆH:Á?2¼8+À6)Ë7+Ï7,Ð9.Ñ;-Ô</Ø:.à8/ç4-î1-ô.-õ/,ò1,ê3+á7*×:'Õ:&Þ;(Ù6#Õ8%Ô=,Ì;*Á5&Ä:/ÏF<ÑF?ÚIFéOOóPQôGKñ<Aõ:AøCHåBEÛEDÒBAÅ>8·;1®:-ª<-«=.¶E5ÆL?ÔNEÜGCä@Aï?Bû@Iþ@JùDM÷CLúDPÿFTþCTú?Tü>XÿBaÿGiÿCiü?iù@jùBpüFvþJzÿK|ÿJzÿHtõKpñQsõ\zù_{úWvüNqýEkúDlïEjÚKiÈ^r±nukLI�~t°–‰Ì²¥äÓÉìãÜëçæêéîêçòíâóöãùûáúÿâýÿçÿÿìÿÿñýÿöþÿùûþú÷ýúõü÷ôúõòøïòôëîñèíïæëæÝàÜÖØÊÆÅ´³±ž ›Ž“Œ‰�…‡�‚�•‡‰�€†Œ~/.,0/-10.21/43/651762761:94:94<<4>>6@@6BB8CC9DD:IF=JH<JH;LI:MJ7NK8OL9OL9KI:NKBTRS_^fihxmm…lkŠkhƒ`YiYQ\OGRH@KH>GG=EE<AA;=D;<?:7=85<73=82>93>:1>:194.:5/;60<92>;4?<5?<5@<3E>4G?4I?3L@2O?0P?/P?-L?,<;&:=(?<+A;-B71A62>42;30=84B;5H94N2.Y,)l/.ˆ<>¢GF£;0¢2$Ÿ1"¤9)¥<)Ÿ8%ž7$¢;(¤;&«>*±A-¶A/¹>.¼=.¾:-¾</²<.¬<.ª:,©9+«<+­>-­>-­<,²?-°8(¹>/ËK>ËG:¼6*º2&É9.Ð6,Ô6+Ô8,Ö:.Ø:/Ý9/ã6/é4-ð0-ô.+ô.+ð0+ç2'Ý6&Ô7$Ò7#Ý8%Ü5#Ú9'Ú>/Ó=.È6)Å9,ËA7É<5Î?9ÙEAåKIêJJéCCí@BëCCÞCAÕD?ÍB=Â=4¶:.®8*«:*¬;+¬8)ÀD8ÕNHÞJHâ?Bê=A÷BKÿHRô@Iô@IùCOýGTþEUú@Uû?XþA_ÿEgÿCiÿCkÿFpÿIwÿK{ÿL|ÿJzÿFxÿHwÿKtûOuÿZ|ÿa€ýUvñCdûIkÿTvÜ?\ÍI`Ø{†¸||aC;�{p®Ž�Õ³§óÛÑûìåúòðù÷ú÷ôýôìûúëÿÿêÿÿìÿÿîÿÿóÿÿ÷ÿÿûÿÿýüÿþûÿýùÿúøÿøöÿôøýòöûðôùðóðçêåßáÓÏξ½»­®©¢¥ž�£™�£—£«žž¦—›£–/.,/.,0/-10,21-43/54/650983983;;3==5??5AA7BB8CC9IF=JF;LI:MJ9PK8OL9QL9OL=LJ>QPL^]bmlzzz’�‚¡€‚¨€€¦{u—rkŠe_{YSmTLdMEZG@PB<HD?F@;?<87;63<71=90=90;8/:70991;;3>;2?=1@<1?;/A;-F?/H?0K?1M@0O?0O?/P>0L?/@=*?>,@=.?;/?;2>93=:5:94<94E:6M51V-+j)-†18ž8C¬=C§8/¥7(¥:*ª?/¨?, 9&¡8%¤;&¥:&«<)´?-¹@/¾?0Á=0Å=1Ä>2½?3¶=2²9.®8*¯9+°:,²:,³9*¶;,º<.ÄB5ËE9È>3¿2(À2(Í9/Ö5+Ù5+Ú6-Û7-Þ7.â7-ç4-ì1*ó0,õ/,ô.+ï0(ä2&Û4$Ò5"Ð5!×2Ú0Ú4&Ü<.Ö=/Ñ:/Í=2ÒD:Ä5-Å60Ê;5ÙEAãNJåKIÞC?Ö=8Ó@9ÎA8È?5¿=0µ9-±9+¯9+±;-¬4&½?3ÑHBÙHEßABæ@DôGMýNUð?Gð?GöBMþHUþHWüBWú@XüA^ÿAcÿCfÿFmÿIrÿMyÿO|ÿL{ÿIyÿDvÿN~ÿR}þKrøMoÿVuÿUu÷MjñGdðLgÒ>VádtþŸ§Ó�Žd92lJ@¥ypÐ¥œõÔËÿéãÿôòÿýÿÿüÿùóÿüïÿÿïÿÿðÿÿòÿÿôÿÿ÷ÿÿûýÿüûÿýúÿüøÿúøÿù÷ÿö÷ÿôøÿôøþó÷öëïìãäÚÔÔÉź¹´³´¬±´©°¶ª¶¾³°¸­¬´©10,10,0/+0/+10,21,43.54/77/880::2;;3>>4??5AA7DA8IE:LF:NH:PJ:RK9RM:SL<RL@OJDVTUfdqwx�‡‡©��¹�“Ã��ÆŽ‹À‡ƒ¶{w©pmšid�_[~UPnNJaKGXFCN?=B<8995296/85,85.671783891;;1></@=,B=*C<)G?,J?-K@.L@0M@0M?2L@4K>5M@7K>5F<3B92=82:946;47<59<5>:1H4-Y2-x89”?D¦>E©8<ª;0«=,«@.¬A/©@-¥<)¦;'«>*§8%®;(¶=,¼=.Á<-Å;.Ê<0Ê=3Å?6¾>3º:/·7,¸8-¸8+º8+º8+»7+ÇA5ÍC8Æ9/Á1(Æ3)Ì8.Ï5+Û4+ß3)á4-â6,ä6-ç4-í2-ð/*õ/,ö/*ò/)í1(ã3&Ù4!Ð5Ï4×2Ù/Ø2$Ø6)×9-Õ;1Ö?6ØE=ÖF>Í@7É<5Ë@9ÖKDÙNGÒE>Å;1Å=1Ã=1À</º;,·9+´:+¶<-¹?0µ9-¿<2É@:ÒC?ÛBDãDHíJOöOVí@Fí>EòBLûIUÿJYýFXúCYûC]û@_ûBdÿElÿJsÿNzÿO|ÿLyþIvÿJwÿT�ÿT}ýJqùImÿStÿVtøNiëD^äF]êXkÿ–¢ÿµ¼ó�žŽHFk.)¦keÍ–�óÈÁÿãÞÿñðÿüýÿýÿü÷þþöÿÿöÿÿøÿÿøÿÿúÿÿüÿÿýÿÿýüÿüùÿûøÿùøÿø÷ÿöøÿõ÷ÿôøÿôøýñóóéêäÜÚÖÑÍÍÊÃÉÉ¿ÊÊÀÊÍÂÉÐȾȿ¸Â¹65154/32.21,10+21,32-43.66.77/880::2<<2>>4@@6CA5JD8ME8OH8RK;TK:TM=SK>RKAQLIZW^li|~���¸–™Ì™žØšžÞ™˜Ú”‘ÔŒŠÉ†…¿�~µxv§nk–fe‡^[xWTiMKYDAJ><?;:8;76875557457664872<:.?<+C>*F@*G?*H@+IA.IB0IA4HB6HA9G@:K=<I;;C9:=77875384/83/917<574+C1']80~C=–GCž>?ž52¨<0ª<+ª?-ª?-ª?-©>*¬?+°?-«8%²9(¹:+¿9-Ã9,È:.Í:0Ï<4Ë>5Æ;4Â91Á8.Á8.Â9/Â8.Ã6,Æ8.ÑC9ÑA8Ã0&Ã,#Ð7/Ö<4Ó2*à3,ä1*ç2+è3,ê3-í2+ð/*ó-*ö/*ö/*ò/)ì1(á4$Ú5"Ñ6 Ð5Ø7#Ù3#Ö3$Õ3&Õ5)Ô8,Ô<1Ö?6ìYQáRJÔIBÌC;ËF=ÍJ@ÈE;À>1¼:*½;+¼:*¹:)·9*¹;,¼>0¿A3ÁA6Â=4Æ=7ÏA=×EEÝGHäIMìMQèBFê@CðBKùKTþN[üJZüF\üF_øA_ùBaýEiÿJpÿNwÿOzÿNxÿLvÿQ{ÿR{þOvýOtÿVyÿ]{ÿSoîD^úTlîQd÷dtÿ’žÿ‹•ôƒ‰Ç`d§MMµjgÎ�Šï¼¸ÿÞÙÿïîÿúúÿüþýûÿýúÿþûÿÿûÿÿûÿÿüÿÿûýÿûûÿûøÿûøÿüùÿûúÿûúÿøúÿöøÿó÷ýñóùíïòææçÝÛßØÒÛ×ÎÝÛÏàÞÒßáÖÜãÛÏØÓÅÎÉ=:5;8185063,41*41*52+63,74-85.:70;81>;2?<3A>5C?4KC8MC7PG8SJ;WK;UL=UK?SJCSJK]Wcomƒ‚‚¦’”Å› ÚŸ¦ê¢©ñ¡¥ïž ë™›å•˜Ý”•Ö�‘ʉ‰½‚ƒ±zy¡rq‘fc~ZWlPN\IGRFCNCAL??K==G;:@;9:=:5?;/C=-E@,F@*FA+EB/EC4CC7BC;AC>@ACB<FD>LD@NEEOCHNAIL?IJ?HECHAA=2L:.gF7ƒOAŽJ=�?4”8+¥>/©>,«=,¬>-­@,¯@-°?-²=+±8'¶8)½9*Ã9,È8-Ì8.Ò91Ô;5Ï:4Í:3Ë81Ë81Ì92Í:2Ì70Ë6/Ô=6Õ>5Ò91Ì2(Î4*Ù<3Ú<3Ø4+ä1*é1)ì1*î1+ð0+ò/)ô-(ö,(ø.*ö/*ñ0)ê2(â5'Ù6%Ó6#Ï7"Ô9%Õ8%Ó6%Ò5&Ò4(Ó7+Ò8.Ð7/ãNGèXPçZSÚQIÌG>ÄD9¿@7¹>/¹<*¹<(¹<(¹:'¸9(º;*À>0ÃA4ÉD;Æ@7É@:ÑFCÕIHÖHGØGJßIKæFHèBDíDKõLSüP\ûO]ûL]ûK`öD^÷DaùFfýIlÿNuÿOxÿOxÿPy÷UzóUxôTvüYxÿ]{ÿZwûQkòI`ýWköUgêM^ö^mâKZæTaåTc×XcÁefȃ~嫧ÿÔÑÿëéÿóôÿøùÿþÿûüÿûüÿûüÿüûÿüúýýùúý÷÷ü÷ôÿú÷ÿûøÿüûÿûúÿøúÿóõýîñúëîðáäéÝÝäÙ×âÙÔæßÕëçÛñíáññåíòìÜåâÏØÕ@=6>;4:7074-52+52+52+63,74-85.96/;81=:1?<3@=4B>3JB7MC7SG9VJ<WK;XL>WK?TICSJM^Xfpmˆ„ƒ«“–Ëœ¤ã¤¬õ§°ý«°ÿ¦¬ú¡§ó ¤îŸ£ê�¡á™›Ö•–Ì’’ĉˆ´zxŸji‹`]|XUpRPhOMeOKdKG^FBS@=H?:>?:6@<1A>/C@/CB0BC3BD7?D=>D@<ED;BJ>CVDHaMRhV[n\br`dobbje`dfXWaJDlG>„TF•XF“J7‘@-™>+¨A0®@/®@/¯@-°A.´A/²=+±8'µ7(¼8+Â8+È8-Í6-Ñ7/Õ81Ø;4Ò72Ñ83Ð72Ò93Ô94Ô94Ó83Ó6/ÞA:Õ8/Ñ3*Ö8/Ü>5Þ=5Ú91Ý6.ç2+ë0)ï/*ñ0+ô/)÷-)÷,(÷,(÷-)ô/)ð1)ê4)á5'Ú7&Ô7$Ð8#Î7"Ï8%Ï8'Ï6&Ð7)Ò9+Ñ7+Ï5+Ì5,ãNGód\ë`YÕOFÄA7º;2µ9-¸>)¸>'º='¹;%º9&½9*À</Ä>2ÊD;ÊA9ÎE?ÕKHÔMJÎGDÍEEÖHGãIIæCDéFIòMSøSZùR\÷O^úOaôH^õG`øGdûIiÿNrÿQxÿRyÿTzðUuïZxù^}ý^|ùTrõMgùOiÿYnÿ[mÿctðM^øUfôO`üUgõL_äR_À^_»vqל˜úÉÅÿåãÿííÿôôÿÿýüÿÿûÿÿûÿÿûÿÿýýýýûüþøøýøõÿøõÿùöÿúùÿúùÿöøüðòøéìôææëÝÝèÚÙäÙÕèÝ×ðèÝùóåÿúìþþòõúöãìëÔÝÜB>5A=4@<3>:1<8/:6-84+84+62)62)72,83-;60>93A<6E>6I@7MC7RF8UI;WJ:XK;ZLA[NHTIMXR`gd�|~§�’Ê— ã ©ö¥°ÿª³ÿ«²ÿ¬´ÿ¬²üª±ùª¯ó¨¬ì©«èŸ¡Ú™šÒ��Á�±ur¡he’]Z‡XR~XQ{YQxVOpNI`C?M?:@?;:B?8>>4AC6BE:?D=<B>:CB>FH?LUEWoSf„bqŽisŽtv��x�„o€�_mŠ[c“XZžWU¦VM¦M?Ÿ@.�:% :$¨=)¬=*®?,±@.³@.´?-¶=,¹;,º6'¿7)Ä6*Ë7-Ñ7-Ö8/Ú91Ú83×84Ø95Ù:6Ø93Ø61Ø61Ú83Þ;4çD=à=6Ú70ã@7þ[Rÿlcÿ]Tç?6å3)í2)ñ2*ô/)ô*&õ(%ø((ø**÷-+ó0,í2+ã1'Ú0#Ô1"Ô4$Ó:(Ì;&Ê<(Í<)Ï<*Ò:,Ò9+Ò8,Ñ7-Ú@8Ï81ÜGAôc^åXQËB<ÆA<·4*·:&¶<$¹<&º<&½:(¿:+Ä<0Æ>2Ç>6ÌC;ÙPJßXTÔNKÃ?;À?:ÏHDßIHâGEæJKïPTñRWïNVîLYñL\ùRføOf÷Jf÷JhüMnÿQuÿRvýRvôWvîXuõXwû[wÿ\xÿZsÿWnÿVkÿbuýVgÿ\mÿ[lñFXôDXÿQeîVeÆfgªieÊ�‹ç¶²øÒÏÿêéýïîüø÷ùýüùÿÿùÿÿúÿþüþýÿýþÿúúþöôþùõÿû÷ÿùöÿöôÿööÿõõöèèæØØãÕÔÞÐÍáÖÐñçÞüôçþøèÿüéÿÿóúÿùèñðØáàC?6B>3A=4?;2=90;7.:6-95,73*73*72,83-:5/>93A<6D=5H?6LB8OE9TH8VI8WJ:YK@ZMGUJNXP_fay{¤Š�É”�⛦õ£®þ«´ÿ¬µÿ®¶ÿ¯·ÿ°·ÿ¯¶ü¯´ø¯²õ©ªë¢£ã˜˜ØŒŠÉ€»uq®jf¡d_™`WŽ`VŠ]TVPtMIbFCTBAIAAC<<:AB=FEACD?A?@DBEOMRWVdch†quš}¢ˆ}Ÿ”|œ |˜¥pŠ¤`u¬Yi¶VaºQU¸HF°@4¬;)¬<&®>(¨8$¨7%¬9'®9(±9(²9(µ7(¸6(»3%Â4(Ê7-Ó:2Ù<3Þ=5ß<5ß<5á>9ß<7ß<7á>9åB=èC=èC=æA;å@:æB9ä@7Ü8/Ô1(Ø5,ëH?ÿZQñ@6î6,ë,$î)#ø-)ÿ//þ,-ö((ú0.õ4/ï61ç6.Þ3)Ö2&Ô2%Ï6&Ë:'È;'Ê<(Í:(Ï9*Ð7)Ñ5)Ñ5)Ô:0Ï5-Ô=6åPJåTQÛLHÍB?·1(¼;(»=&½<'¾;'¾9(Á9+Â:.Ä:/È>4ËB:ÕNHÝXQÓRLÃD=¿C;ÌHCÛLHßJFåMLëSRíTWëPVéNVìOZ÷VføUhùSkúRlþQoÿRsÿStÿRuòMmøUtÿ\|ÿ_}ÿ\vÿUoûRiûReüUg÷RbÿZjÿZjóN^øScþYißS^¶`_—^W¶�}Ù©¥îÈÅýáÞùééýøõúüùùÿÿøÿÿ÷ÿýúþýÿÿÿÿûûÿö÷ýøõþùõþùõþöóÿ÷öÿõõöêêêÜÛäÖÓÝÐÊßÕÌïåÛüõåÿùçÿýéÿÿïúÿøèñîÛáßFB7EA5C?4A=2?;0=9.<8-;7,84+84+83-83-:5/=82?:4A=4F?5JB7ND8RF6TG6WG7WI<XKCWKKXP]b^ysuœ…ŠÂ�™Ü˜¤ð «úª³þ­µþ¯·ÿ±¹ÿ³ºÿ´»ÿµ¹ÿµ¹ÿ¯²ù¬¬ô££ë™—àŽ�Óƒ€Çxu¼sm³k`¤i]›cZ‘]W…VRwNKhEDV@@L85<A<@KABL@BQ>B[FKpW]‚cu—s—¤{§®«²{£¸vœÀp“ÂcƒÀSpÂF^ÐK\ÑHRÄ<<º5.¸9*µ<'°<%°;)°;*²:)´;*·<,»<-À<-Ä<.Ä6*Ë8.Ó<1Ù?5ßA8á>7á=4ß:4Ý84Ù40Ù40Þ93çB<ìE?éB<ä=5ã<4Ü5-Ø4+Û7-Ô2'Ì, Ö6*éG:õK>ò@6ï4-ð-)ö,*û-/ü+.ø*,ð*)ì/+é3/ã5.Ü4+Ù2)Õ1'Ñ3'Í7(Ë:)Í:(Ð:)Ò;*Ó:*Ô8+Ô8,Ñ4+Ö<4Õ:5Õ<7åONíYWÙEEÂ3/Á<-¿<(À;*À;*Á9+Á9+À8,À8,Æ=3ÇA6ÏJAØUKÓSJÆG>ÀD:ÇG>ÖKDÙJDÝLIäSPåSTâPSâMSåNWòXdöZhý[pþZrýUoûQnýPnÿQrýGmÿPvÿZ}ÿ[{ÿVrýSmþVmÿ\mûXiøZhû`nö`lí]gïfnìfmÉ\_ ]W{OF˜mfÁ˜”ݺ¸ðÔÓôàßÿõôùù÷ùÿýùÿÿ÷ÿÿúþýÿÿÿÿýÿÿøùüøõüùôüùôýøôÿùöÿøöùîìïäâåÚÖÛÐÊÜÒÈîäØýöäÿûèÿýçÿÿíúýôêðìÞãßIE9HD8FB6EA5B>3A=2?;0>:/:6-:6-94.94.:5/<71?:4@;5C?6HA7LD7OF5SF5UE5UH8VJ>YNLXNW^Zqpp–€†ºŒ–Ô•¡é�©õ¨°ù©±ù­³ý°·ÿ±·ÿ²¹ÿ´¸ÿ´¸ÿ¯±ü«­ø¥§òŸžê—–⌋׃€Í}yÅum¶pgªf`ž_Z�YX„QQuGGcB?R=5DH8BT=C`@EnCJ€OU—^gªgx½k�ÉnšËo˜Éi�È`ƒÉWxÆIgÄ;UÆ2HÙAPàDOÑ;=Ç82Ä?0½@,±:$­5$­5%¯6%³5&¶7(½9,Ã;-È</Ó@6×@7Ú@6Ý@7Þ=5Þ93Ý6.Û4,Ó.(Ò-'Ó.(Ø3-á:4ã<4á81Ü5-Ü5-Ð,"Î* Ù7,âB6Ü>2Ñ5&Ë.éF7óI<ýH?ü<9ô,,ñ#%õ')û/2ñ-.ì0/ä2.Þ3,Ú1*×0*Ö/)Ó0'Ó7+Ò9+Ö:-Ø<-Ú>/Û?0Ü@3Ü@3Ò6*æLBåJEÎ50ßGFóZ\ßIKÕA?È</Ä;+Ã:*Ã:*Â:,Â:,Á9-¾8,Â<1Â>2ÈF9ÏOBÎPDÆH<¿C7ÁC7ÎH?ÑF?ÕJEÛPKÝROÙMLÚKMÝLQëX`ó[güaqÿauûWoõNhõMgÿMmÿOuÿRyÿUwÿTqüTmý[pÿcuÿizÿguúboõamídnãflÙhjÎghµfa�^TeG<^U©ˆ�Í«©æÈÈïÚÙÿóóùøöùýüøÿÿ÷ÿÿûÿÿÿþÿÿüÿÿùüþú÷ûúõûúõþûöÿýùÿûøþôòøíéêßÙÞÔËÝÓÉíæÖÿøæÿþèÿÿæÿÿëøúïêïèãæßNH:MG9LF8JD6HB6F@4D=3D=3?80>7/=60=60=60?82@85?:4C>8EA8LD9NE6RE4RE2VF6VH;ZMGVMR[Whlk‹~‚±Š“Δžãœ¦î§­÷©¯ù­°ý¯³ý°³ÿ°´þ°³ÿ¯³ý«®û¨«ø¤§ô ¢ïšœé“•âŒ‹Ù†…уÉ|v¼mk¬dež_a’Y[„RQsNHbWFY_CQmBL‚FP˜LV¬U^¼[dÊZhÙRpâRußTuÙPlÒJbÍBWÇ8JÅ.?Ô8EãCMæGLØ>>Í>8ÌF:ÃH6·@,·>-¸=-º<-½>/Ã?0ÊB4ÐD7ÕE:ÙE9ÚB7Û>5Ù;2Ù6-Ù5,Ú3+Ù2*Õ0*×4-Û60Ý90ß80Þ7/Ý6.Ý6-Ö/&Ú6,Ø8,Ï2#Î2#Ò9)Õ>-Ô;+Ø<-èD8øJAýDAù56õ*-ö*-÷/2÷67í55ã41Ú1,×0*×/,Ù0-Ú1.ã81ä91ã;2â;2á=1ß=0Ü?0Ú>1Ñ7+ïWLøaZÓ;6Ñ;:æPQÜFHæPQÏ<4Ê:/Æ8,Å9,Ä:-Ä<.Ä<0À:.¾:-¾<.ÃA3ÇH9ÇI;ÂD6¾@2¼>0ÉF<ÊE<ÎIBÕPI×RMÕNKÔLL×LOéZ`ð^hüetÿhzü]qóOgóMeýNkÿVzÿTxÿRrøTmö]qûhxûjw÷erüetó_mì`kêktßruÃjfªd\žla}fVYI:kXJ”{t½ œÞÂÁîÕØþîïûõõûûûûÿÿùÿÿýþÿÿþÿÿûÿÿúýÿþúûþõûüöÿþùÿÿúÿþúÿûõÿ÷òñèáäÜÑáÙÌðé×ÿùæÿþèÿþåÿÿê÷ùëíðçèéáQK=PJ<OI;MG9KE9IC7G@6G@6B;3A:2@93?82?82@93A96@;7B?8EB9KE9MF6PE3RE2VG4UH7YLDUKLZScii…|�«‰“È“žÞš¤ë¦¬ø©¬û¬­ý­°ÿ®±ÿ®±þ­°ý­°ý©¬ù§ª÷¤¨ò¢¦ðŸ£í›Ÿé–˜ã’”Þ’’Ú‰‰Ï|€Àty³ou©mp�igŽj_}y_x‚WjŽP_£O\»S^ÍT]ÖPWÛHRå@Qè?RäCRÜDPÖCKÏ@DÊ9<É46äJLçIJáAAÖ87Î95ÎA8ÈD8¾>1ÃE6ÃE6ÅF7ÉE8ÎF8ÑG:×G<ÚG=Ö>3Õ;1Ö8/Ö5-Ö3*Ø4+Û4,Û7.Û81Þ;4à=4à<3Ý90Ü5-Ü5,Û7-Ò.$Ý;0Ú=.Í1"Ç.Î7&Ò<+Í:(Î8'Ô8)â:1ó=9ÿ@@ÿ<>þ37ó,/ê,.á+*Ø*)Ô+(×/,Þ44ä88é9;ð88ð75î73è71ã7-Ü6*×5(Ñ5&Í5(çPEÿmcÛHAÈ42×CCÔ?CêVVÙA<Ò>4Ë80Ç7,Æ:-Æ</Ã=1Á=0¼:,»<-¾?0ÀA2¿A2½@.¼=.»<-ÄD7ÄD9ÈH?ÐPGÓRLÑPJÒNLÕONç_cîaiûitÿm|ÿdwõVjõQiÿUoÿWvÿVtúXpñ^pônyöw€íksÞXañ`mñ`mìboïs}숊Ì�|ªth™yjskXON<_VG„rh¯–’ÜÀ¿ïÕØüéëþó÷ýûüýþÿúþÿþýÿÿýÿÿûÿÿùþÿÿûùÿõúýöÿÿúÿÿúÿÿúÿý÷ÿþöùðçíåÚéáÔôïÜÿûçÿþåÿÿãÿÿèùúêòôçîðåSM=SM=QK=OI;MG;KE9JC9IB8E>6D=5B;5A:4@93A:4B:7@;7A@;CC;IE:MG7OG4RF0VG2UH5WJASIHXR^jg‚}�§‹“Ä’žÚ˜¢ç¤§ö¦§ú«©ü«¬þ¬­ÿ­®þ¬®û¬®û¨¬ö¦ªô¤¨ñ£§ð£§î §íŸ¤è�¢æ™žâ”™Û‹“Ò†�ʃ�„ˆ·ƒ�¨ˆyš�mˆšcz©Xi»Q_ÐNZàIRèAIê=Aé=;ä?9ÝC9ÖH:ÎI8ÆG4ÂC0Å@1äTIàG?Ú;7Ø64Ø88Ù;<Ñ98È74À:/¾<.¿;.À:.Â8+Ã7*Æ3)Æ2(Ï5+Ð3*Ñ3(Ô3)Ö3*Ù5,Ý6.Ü8/ß>4Ý<2Ü;1Û8/Û7.Ú6-Ø4*×3)Õ3&Ó3%Ð4%Ð7'Ò=)Ñ>*Ê9$À2È:&Ê4%Õ1(ç51ú<<ÿ=@ÿ8>÷37ê-1á-.Ù--Ø0/Ü43ã9:é;=ð9=ö26ø03ô01í1/å2+Þ2(Õ1%Ï3$É1$Ñ=1új_äTLÈ95Ð@?Ë:=äRSäJHÜC=Ñ94É6.Ç9/Å;0Â<0¿=/º;*¼?-¾A/¼?-º;*º;(½;+¾<,¼>0¼>2ÀD:ÈLBÌPHÊNFÌLIÐNLç`dìahøhrÿo{ÿhx÷Zk÷UjþZrÿYsû]tîaræjtë}€ñ‡‰èwyÚ_dñgtþm|óeuëlwö‘•ì¢ŸÄ”Š¢�|ikVHQ<WUFtg^¤ŒˆÛ¾ÀòÕÙùãæÿóøÿúüþþÿüýÿþûÿÿüÿÿúÿÿøýÿÿûùÿõ÷üõýÿùÿÿúÿýøÿý÷ÿÿöÿùïõïáòìÜúõáÿýçÿýäÿþâÿÿçýþìùúì÷÷ëVO?TN>SK>PJ<OG<MG;LD9JC9H?6F?5E<5B;3B92A:2C:5A<6C@9DD:JF;MG7OG2QF0TH2UH5UH?QGFWQ]jh€~‚¨Œ”Å“�Ø™¢çŸ ò£ ÷¦£ú§§ý¨ªÿ©«þ©¬ý§­û¥«õ£ªò£§î¢§ë¤©í¦©î¥¨í£§î˜ è– å”Ÿß”Ÿ×•›Íš•¾ Œ¯ªƒ ¥e´]pÃRdÐIWÜBNå>Fë<Aì<<ë@9ãA4ÜE4ÔI4ÊI3ÂF.¿B,Â?-âSEÞD<Ý97æ>>êDFå@DÙ8=Î65ÍB;ÉE9ÉC8ÇA5Æ>2Ã9.Â5+Â2'Æ2(Ë3(Î4*Ó5,Ø5.Ù4.Û4.Û4,äA8Þ>2Ù7,Ú6,ß8/à:.Û5)Õ1%Ø8(Í2 Ê1!Ï9(Ï9(Ç4"Ê7%ÔC0È7&Î8*Ø7-Þ5.ç40î53ö:9û?>ÿLKúHFòBBì>?ê<=è8;å26æ+2ô+1ø)/ô,.ï/.è1+à4*Ù5)Ó7(Æ0!À.ôdYë]SÏ@:ÔDCË;;ßMMêRQàHEÔ<9É6/Æ8.Æ9/Ã;/¾:+¾<,Á@-ÀA.¼=*·8%¸9&½<)¾?.µ7)´8,¸>3ÁG<ÅK@ÅIAÆJBËJEåa_æ^`ñdjþnwþkuó]iðXeø]mú^sõbtèdqãlrì~ù�‹ö†…éqsøryÿy†õaqàXföˆ‘ÿ´´Ð®¢ ›‡]jPDR9QTAjbUœ‰‚ÚÀ¿òØÛöàãÿô÷ÿúýÿþÿýüÿþüÿÿýÿÿûÿþùýûÿúôÿôòþòùÿöüÿöûüôÿýöÿÿöÿÿóûõç÷ñáþùåÿýçþüãþýáÿÿêÿÿñýýóüüòXNBWOBWMAUM@TJ>RJ=RH<PH=MC9KC8KA7H@5H>4G?4H>4F?5GC:GE9JF:NI6RJ5UJ4VK5VK9XNEULMZTbkhƒ{~§‰�љטŸç£¤ö¥¢ù¤¡ü¡¢ü¡¥ÿ¢¨ÿ£«ÿ¤­ü¡ªõ¡©ñ¡¨ì¥¨íª©ï­©ó±©ö¬©ø ©ø—§ò’¡â–ŸÖ¦£Î¶ž¾¹…œ¸gz¿L]ÑERß>Mç;Gç<Dæ=Bà@@à@@â>?ß<=ÝEBÉ:4ÊD;ÊE<º1)ÜKFèJIéCEë>Bé<@æ<?á=>Ú<=Ô<;ÏB;È?7ÊD;ÑMAÍK>¿=0¸6)½9,¿7+Ã6,È5-Ë2*Ñ/*Õ0,Û2/Þ50ç@7ß9-Ú2%Ý3&å9+ç;-ß8(Õ2Ï2Ì4Ê5!Ç4"Ç4"Ê4%Í5(Ï5)Ê0&Í3)Ñ5)Ò6*Ó5)Õ3&×3'Ü4)à2)æ3.ë52ë33é/2ç-2è-4í.5ö-3ö+/ï+,í0.ê5.ã9,Ù7(Ð4%Ã-Ä3"çXHéYNË;3ßNIÅ41ÚHHáOPÜJJÔBBÍ<7È80Æ8.È:.È<-Æ=+Â;(¾9&¼;(»<)»>,º=+·<,´8,µ;0¹=1»?3»?3¾@4À@5ÄA7ïjaõpiêc_ômjûqqìadôgmÿpyÿlzûhxðdoébiëdjîlnîqoîqoûy{ÿ}…ùZlücwßbpû¬¯¶ªš}‘u]rQJY:LR8snZ¥•ˆË·°íÕÓÿîðÿöùÿ÷ûýøüüüþÿÿÿÿþÿÿþÿûÿþðÿôåþëæûêïÿðøÿöüÿôþÿóÿþñÿýñÿûîÿúêÿùæÿùãÿùáÿùáüúåÿÿõÿÿûÿÿûYOCYOCXNBVL@UK?TJ>SI=RH<ND8ND8LB6KA5J@4J@4J@4IA6GC8HD9LF8OH6RJ5SK4VK7TK:XNEULMYUckhƒ{}¦†ŒÀ�•Ó–šáŸžì žò �ôžŸù�¢ýŸ¦ÿ ªÿ¡¬û ¬ô «íŸ§è¢¥è¥¤ê©£í« î¤Ÿï §÷š¦ðœ¡á£œÐ±˜À¿�ªÂq„ÄTbÑCOâ<Fì8Aï7?ê:=å<?Ü@AÛACà<CÛ:@ÚADÌ<<ÍCAÊ@>Á31åOPçDGé@Eê;Bç8?â8;Ü9:Õ:8Ï;7ÔE?Æ=5Æ=5ÌG>ÌLAÂD6¸:,·9+¹7*¿7+Æ6-Ì5.Ò3/Ø3/á55ä84å>5á;-Þ6)ß5&â6(á5'Û6#Ó4Ñ9$Ë9"Ç9%Ç:&É<+Ë<,Ì8.Î4,Ó4.Ö5-Ó6-Ï7)Ë5&È5#Ë4!Ï4"Ø2$Þ2(ã0+æ.,æ,-æ,/æ-2é/4ì+.í,-ì0/è2.á5+Û5'Õ8'Ò9'Õ?.»*ÖD5Ñ?2ÙF>ëVPÚB?ØBAçUVâRRÜJJÔC@Î>6Ê:/È8-Ç8*É<+Ä;(¿:'½<)½>-º?-¹@/¶>.²:,´;0·=0¸>1º>2½?1ÁA4ÅC5ëi\òmdçb[ðkfõolèbañklütxÿq~ÿo|öhtí]fêY`ñbf÷qpþzxûvwÿx~ùRdþ[pä]nö¦©©¦“jŒk[sOO_;SZ;us\©��ÖĺöáÜÿôóÿúûÿúûÿûüüüüüþýûÿÿúþÿõÿúíÿóáÿéáýæéÿëóÿñøÿòýÿòÿÿñÿýïÿýíÿúéÿöåûóàúòÝüôßü÷äýúóüüúÿÿý\PB[OA[OAZN@XL>WK=WK=VJ<SG9RF8QE7PD6OC5OC5OC5MC7KD:JF;MG9PI7SK6VK5VK7TK:WMCVMNZWbki�z|¢…‰¹‹‘Ë‘–Ö˜˜Þ™™ãš™éš�î› ô�£÷Ÿ§ø ©ô ªïž§èœ¤ãœ¢àžŸàž�៙㜘▛ߚšÚ¤•Ð±ŒÀ¿‚¬Çr“ÉYqÈCT×ALß>Dã=?å==å<?ã=?á>Cá=Dà<C×8=Ö@BÏA@ÊC?Ã<8Ä96éWWàBCä>Bè;Aã8>Ü68Ö66Ñ96Í<7ÕHAÄ;1¾5+ÅA5ÍK>ÆH:»=/·8)º6)¿7)Ç7,Ï6.×50Þ71æ87é=9×3*×5*Ú6,Ü6*Ý5*Ý7)Ú:*×>,Å4!¿4¹4!º7#¼8)¾9*¾4*À-%ã?=å>8ß>6Õ<.Î8)È7$È7"Í8$Ò7%Õ3&×/&Ù.'Ü/+Ý/.Û//Ü./Û,)ß0+à5.Ü5,Õ3&Ð4%Ð9&Ñ>*ÔA/Â1 çSEÒ;0ÚA;ÞC?éKJðTUíTVèTTãOMÝIEÖC<Ï<2Ê7-Å6(Å8'À7%¼7&¹:'º=+·>-µ=-²=,°:,±;/µ<1µ=/·=0¹?0¾C4ÃE7Ü\Qêg]äaYðlgöpoçabìfgójqàR^ï^kõgsòemñ`gõfjùqqþxwútuÿnuóJ]øRhä[mñœ¡£œŠb`VoHRd>W`Axx^­¥’áÑÄÿîæÿøñÿüøÿþúÿþúýÿüûÿýûÿÿùÿýóÿ÷èÿîÜþãÚøÞáùáéüæïüèöýëýþîýúëÿúêÿøçúñàôèØòçÕöëÙøñáþúñýüøÿþû\PB\PB[OA[OAZN@YM?XL>XL>UI;TH:SG9RF8RF8RF8RF8QG;LE;KG<OI;RK9TL7WL6WL8WK;VLBUMK\V`jh~yy›‚…²Š�Â�”Γ•Ò•”Ö••Û–˜ã˜›è˜žìš¢í›£ëœ¥è™¡ß–ŸÚ•›Õ”˜Õ•–Ö’“Ö‘�Ò…„½‘ƒ¶¥}±·s¢Èf�ÓW{ÕGaÔ<KÞAJàBCÞCAàB?áAAä?Cç=Fç=Fã?FÖ:>Ñ?@ÑEDÆA<º61Å>:êZYÝABã@Cæ=Bã:?Û89Ó97Ð<8ÎA8ÏE;Â:.»2(À</ÇE8ÄE6½>/¸9(½8)Â9)Ì9/Ò:/Ú91à91è;7ë>8Ù2,Ù2*Ü3,Ý5,à5-ß7.Ü:/Ö=/É7(Æ:)Á<+¾<,¿;/Á;0Å<6Î95ëCBîA=ä@7Ú<1Ï9*É8%È:&É;%Ð=)Ð7'Ï1%Ñ1%Ô1(Ö3,Ô1,Ñ/*Ó2*Ó5,Ó7+Ð6*Í5'Ê7%Ì;(Ì>*Ì;(Ï<,ô^PàF<Ò3/Ó/-æ@@ûWXÿtuÿrqÿjgö^YçPIÕA7Ç5(½. Ä8'À8(»8&¸;)·<,µ=-²=,°<-¯;.®=/°</°</±;-¶>0¼B3ÀF9ËMAß`Wâa[ôpløstæ_cä]aæ]dÖHTçYeógrôgoöelükpþsvþvvûpsûenñDXôKbåXkë’– •ƒ_zYRkCSh?\gEy|_°ª”çÜÊÿõèÿùðÿûñÿýôýÿ÷ûÿúùÿúöÿøòüôêúíÚøÞÎóÒÊëÌÏìÎ×îÔàïØéòÝò÷ãöõãûöãüõãõìÛíáÑéÝÍíáÑñèÙüõíü÷ñþùó\PB\PB[OA[OAZN@YM?YM?YM?WK=VJ<VJ<UI;UI;UI;UI;TJ>OG<OI=QI<SL:UM:XM9YL9XL<UKAVLJ\U]hexut“|~¥…‡·Œ�ÂŽ�Ã��È�’Ï’“Ô’•Ü”–à”˜ß•šÞ—œÜ”šÖŽ”Ή�Ç…‹Å�‡Ã}ƒÁ~¸}tŸŽp”¨g�Á\„ÖNvâAcè:Uê9KæAGàECÜGAÛGCßEEæAGî<Jí=JæBI×>AÐ@?ÐIEÀ?9´3-ÈC<àUPÜDCáACä@Aá>?Ú?=Ô@<ÑD;ÏF<Ç=2À8*½5)½9*À>0¿=-½;+½<)Â:*Ç;*Ï;/Õ;/Û8/á90ç83é:5ç:6å63ä20æ21è43æ40Ý2+Ô0'Ì2(Ë7-É;1Å8/Á3/Ã40Í;<Ý?@ë8;í76ä71Ø5,Ì4&Æ5"Ä7#Å:%Ë=)Ë8&Ì4&Î5'Ô8,Õ;1Ó9/Ð8-Ð>1Ê;-Å8'Å8'Ç:(Ç<)Ç<)È;)Í>-Í:*Ø>2áC:Ô/-ä::à24à67×53Ö;7ÞC?æKFêQIëTIèTHäUGË?0Ä?.¾<,¸=-·=.´>0°?/®>0¬<.¬<0­=/¬<.­<,±=.·A3ºD6ÇMBÝbZàc]ïolôrrå`cå`eç`g÷kvûoz÷kví`hð_fýlqÿy|ÿy|úmsø^jóAWõD^èUhæ‡�¤“ƒh€^[tJ_vJhwP‚ˆf³²–èâÌÿúéÿýíÿþïþÿòûÿôøÿôôÿñíûêåñãÚíÙÆèǻ伺޺¿Þ¼ÇàÀÍâÃØæÌãëÓìîØóðÝ÷òßóêÛìàÒèÚÍêÜÏïáÖòèßñèáòéâ[OA[OA[OAZN@ZN@YM?YM?YM?XL>XL>WK=VJ<VJ<WK=WK=XL>QI>RJ?SK>UL;XM;YN:ZM:YL;XL@ULGYSWe^nnjƒut–~}¥„…±†‰´‡‰¹Š‹Á‹�ÈŽ�ÏŽŽÔ��Õ��Ó�”Ó‹�Ê„‰Á|�¸u|²ov­io«ljœ€kŠ”dz®YvËPoãBdð7Vö4Lø9Iï@EåFBÝJCÛJEáGGèCJò=Nð>NäCI×BDÌA>ÎID»<5°4,ÊJAÕLFØBAÚ>?Û;;Û=<ÙA>ÔE?ÍG<ËG:À8*Á7*À8*¾9(¼8)¼9'½:(À;(Ä;)Ê=,Ò<.×;.Ü8.á6.ä6/è50é32è./ê-1î02õ47õ79î45æ21Ú.,×4/Õ63Ñ32Ï/1Ò/4Ü7>ê:Dí06î02æ21Ü3,Ñ5)Ê7'È9(É<(Æ8$É6$Ï7)Ô;-Ú>2ÝA5Ý@7ÙB7ÓG8ÉA1Ã<)Á:'Ã<)Æ=*Å<*Å8&Ì:+Î8*Í0'æC<Û2/ï?Aã.3Ú*,Õ1/Ó6/Õ83Ö92Ò8.Ì5*Æ2&À1#ÎE5ÆA2¼=.µ:+±9+¯9+«:,ª:,¬<0«=0¬>/¬<.«<+®=-³?0µA4ÉSIàg_ßd_èkiînoæchðjqõoxúp}üs}ömwîbködnÿrzÿx~ÿquúipøWfùAYô@[êQfâ}ƒ­–†zŽkn‡]o‰ZyŠ`�—r¶·˜ßÞÂùôÞüúåÿÿïûÿïõÿïïÿëçúäÜïÙÑãÍÅÞÁ±Õ¯¨Ó¦ªÐ§°Ó«¶Ô®¼Ö±ÉÚºÒßÁàæÌèêÔòíÚóêÛîâÖìÛÑêÙÑìÛÓéÚÓæÙÑä×Ï[N>[N>[N>[N>ZM=ZM=ZM=ZM=ZM=YL<YL<XK;XK;YL<YL<YM?SI?TJ>UL=XL<YN<ZM:ZM:ZM<ZN@VKEXOR_Wbd_sjf�so�zzž|~¤~€©�‚²ƒ„¼‡…ƈ‡Ë‰ˆÌ‰ˆÊˆˆÈƒƒ¿|}µvw­qr¨kl¤fg l`Ž‚_s•V_³N^ÒJ^ì@Xø7Lþ3Fþ9Có?BéD@áHBÞICâFIéBJò=Nï>NÝAEÔFDÈA=ÉHB´:/±7,ÏPGÊC=Ï>9Ð64Ð21Ð53Ó>8ÒE<ÉG:ÃD5À8*Ã:*Â:*¿:'½8'¾9&¿:'Á:&Å:'Ë:)Ó:,Ù9+Ý7+á5+å4,ê3-ë*+ñ*-ø-3ü/4ÿ17ÿ39ÿ49ý58ù8;õ8<ò9?ò9Aô9Dø8Gû6Hþ5Gô+5ô-2í12ã4/Ù6-Ð8+Î:,Î=,Ë5&Ñ8*Ù;/à>3â>4ä=4å>6ßA6ÙJ<ÐH8ÉA1Ä<,Ã:*Â9'Æ8*È9)É2'Ó9/á@8ëD>Þ0/à..ì59è58Ø3/Õ7.×82×:1Õ;1Ò;0Í:0É;/ÖL?ÌF:¿@1´9*°6)¬6*«7*©8*¬;-­=/®>0­=/¬=,«<+­<.±=0ÁLBÚc[Þc^èkiðosêinõrzüv�îfr÷oyùq{öjsþlvÿwÿu|ùhoùcn÷RbþB[÷<YêKaÝt{µš‰�žz}˜k{—f‚•h‘�u¬²�ËÌ­ààÄéìÑîöÞåóÙÙíÑÏæÉÅÞÀ¼Õ·´Ë®¬Ç¤šÁ”–‘™Ã“¡È™§Ë�¬Ë¡·Í§ÁÒ°ÎÙ»ÙÝÄåâÏìåÕïáÖëÚÒçÒÍãÐÊàÏÈÚÌÃÕǾ[N>[N>ZM=ZM=ZM=ZM=YL<YL<ZM=YL<YL<XK;XK;YL<ZM=ZL?VJ>VJ>XL>YN<ZM<[N;ZM:ZM<[OAWMDWML\RZ^Xfb]qjfsp�wv–xxœ|{§�~µ„¿‡‚ň‚ȉ„Ç‚}¿y·zu­xq§vo¥tk¢pg za‹„P\™HG³EHÓHOîBNø:Fþ6Aÿ:Aõ==îA=æDAãEBäCHèAKî=Më?MØ?BÑGDÅA<ÄH@²9.³:/ÑUKÃ>7É83Ç/,Å*(È/*Î;4ÍC9ÃC6¼>/Á<-Ä;+Ã<)Á:'Â;(Ã<)Â;'Ã8%Ä5$Ê7%Ô8)Ù7*Þ6+á3*å3)ë3+õ33ÿ58ÿ8<ÿ4:ÿ-2ÿ(-ÿ(/þ+1ü-3ö)0ò&1÷)6ÿ,?ÿ*Aý!:ó/ö"1ó'0í,1æ0/Û2-Ô3+Ñ4+Ð6*Ô6+Ý90æ=6ë>7ë:4è71è51â92ßH=×K<ÓE7Ë=/Ä6(Â3%È5+Ï8-Î4,Ô6-æC<à93à21Ò  ì89å55Õ0*Ò4+Ò4+Ï5+Î6+Ê6*È5+Å7+ãYNØRFÊH;¾@4·;/µ;0µ<1µ<1¯9-±;/²>1±=.¯;,­9*­9,®:-±<2ÐYQÜb]ìppôvyîmrõq|ör}ÿz…ÿ‰ÿx‚ôhq÷epÿq|ÿuÿkvø^jöOaÿA^÷8WéG^Ûntº�‹™©…‚žn}™f€–g�œq¡©„´º˜ÈÍ­ÕܽÌÙ»ÁÔ´²Ì©§ÂŸž¼˜š¶�–²Œ�°‡‰³�†µ�¹„•À‹œÂ�¡Ã‘¨Å™²È¡¾ÎªËѵÚÙÄæßÍêÞÒéÖÏáÌÉÛÆÃÔÁ»Î½µÈ·°ZM=ZM=ZM=ZM=ZM=ZM=ZM=ZM=XK;XK;YL<YL<YL<ZM=ZM=ZL?ZLAZLA[M@[N=]M=^O<\O<\O>YM=YLCXMIXNO[QY]Ub`Yia\rjf�okŽsp›vq§{u³„}ÁŠƒÉŒ‚LjÀŠ~¼ˆy²ƒr¨‡s¨�t«‡l£†\„�GN¤B9¶@<ÍC@Þ@?è:;õ=?ÿDE÷><ô>=í?@ê@CëBIêCMëANãALÑ??Å@9¾?6·>3°</²<0ÀD8ÒMDÈ72Ô:8Í2.Ì3.ÖC<É?4µ6'»>,»6%Á:'Ã<)Â;(À9%Â9&Æ;(Ç:(Ë:'Ð9(Ö6(Ø2$Ú."á0&ì7.õ=5õ82÷40÷0-ø**û&(û%'û%'û$)û&,÷!)ø".ÿ'7ÿ(>ÿ!;ÿ7ý1ÿ'8ï#.æ#+ã+-Ý--Ô+(Ó.*Ü41â64è64ï77ò65ó23ï./î,,å/,Ü92Ñ:1Ò91ìSK¼#Ä+#ßF@È.&Í.*Þ<7èE@â:7Ø/*Ú.*â51å<7ß>4Õ;/Ï5)Ë3&Ì5*Í;.Ë;0Å8.ìbXàZOÈC:¹6,º:/¸8-±3'´6*±1&°2&®2&°5&°6'²:*´<.³=1°:0ºC;Ö\Wìppñsvöx|ûyƒõq|ÿ|„ýw€út}üs{ÿr~ÿr~ÿmyÿgtü`nôI\ÿ>\ÿ>]ä@XÔelµ”�’ |ƒ�m}™f~”c‡—j”Ÿwž¦� §…œ§…’£Š£|€Ÿvzšqyšo|�r� vƒ¤u‚­wƒ³y‹¸‘¼„•¾†—¾‡œ¾Œ¢½�§¼“±¼š¾Â§Ï̹ßÓÇãÒÊÜÇÄÔ¿¼Ìº¶Á²«¹ª¥ZM=ZM=ZM=ZM=ZM=ZM=ZM=ZM=XK;XK;YL<YL<YL<ZM=ZM=ZM=ZL?ZL?]M>]M=]N;^O<\O<\O>]P@[OC[NFYNJZPQ\SX_U^_Wfd\tibƒng‘qj�xp¬€y¼‰Ä‹€Ä��¿�~º�t«Œn¢—q¤¡u¨›kŸ™W}�<Cª5+²6,Â<3Ó@8Ý>8è?:óA=õ=;õ;<ô;@ñ>Dî@IèAIâ@KÛBGÇ<9½=4¶<1±;/«:,®:-½A5ÐJAÚFBÜA=Ô63Ð51ÖA:ÍC8¼:,²5#½8'Â;(Ä=*Ä=*Ä;(Æ;(È;)Ë:)Ï9(Ï6&Ô2%Ú2%ã5*ê8.ð91ð91ç1&è0&ê.%ï,&ö+'û+)þ,+ÿ-.ü)/ú$,ú$0ÿ'7ÿ$;ÿ8ÿ3ü0õ0ö.9õ3<æ+2Û(,Ü.0Þ02Û+.ñ:>ô7=÷4:ö26÷05ø03ù25ô87Û2-Ú;5×82äE?äEAÎ/+Ñ2.Á"Ô51á?<èE@à=8Ø3-Ù2,Ý60Ý:1Ô:.Ï9*É5'Ç5(È9+Ê<0Ê<2Æ9/ícYáXNÉ@8º4+Â91Ã:0¿6,Â9/¿5+¿5+½4*»5)¸6)µ5(²7(°6)´;0¼C:ÓYTèljïqtöx|þ|„øv€þ|„üy�ùv~þuÿsÿp}ÿkwüdqý_nöI]ÿ>\ÿ>\êDZÓdj®Žy‹—qz”du’\xŽ]�cŠ•k�˜p‹•p…”mr‡`mˆ]iŠ]g�^l�bq•ey›izŸk�¬t„±vŠ·|�»}�¼“¼‚—¼†›»‰›µˆ¢²�­¶—¾½¨ÏÆ·ØÇ¿ÖÁ¼Ï¼¸Æ·´»­ª²¤£\L<\L<\L<\L<\L<\L<\L<\L<ZJ:ZJ:[K;[K;[K;\L<\L<\L=\L=\L=^L>]M=]N;^O<^O<\O<_RA]QA\PD[NFZOK[PN]QS]RZ_SgcXvh]…kb‘sk¤vµ‡|¾‰}»—…¿™€·›v©žpž®u¢ºv¥µi˜°Rv¸<D¾5/º4+À<0ËC7ÓE9ÛC8ä?9ñ=<÷:>ù:Aö=Eï@GæAHÙ@EÏ@BÀ;6¸90±8-­9,ª9)­9*»?3ÏF>ÙD>Ø:7Ù75Ú;7ÜC=ÞPFÎH<µ3#¾9(Á:'Ä;)Æ=+É=,È<+É:)Ì9)Ï7)Ð3$Ó/#à5+î=5õ@9ò:2é4+à4&ß5&â2%å/$ì-%î+%ñ+(ó,)ó+-ð(+ò&/ø(6þ%:ý7ü3ø3ô%7ò0;ó5Aò8Cñ<Eð<Eê6?ä,6ð2<ó/;ö.9ö-7õ.3õ.1ó/1í42Ü0,â=9Ú64ß;9ÿmkåCAÒ0.Î/,Ú;8ß@<àA;Û<6Õ7.Ö5-Ö5+Ò6*È6'Ä7&Â6'Ä8)Å;.Æ<1Æ<1Å;1ë`YáTMÉ<3¿0(Ë;3Ñ>6Î;3Ò=6Ø=8×<7Õ>7Ð=3Ê<2Â:.»7*¶6)¹;/¼@6ÏTMãgeíorøz~ÿ�ˆû|ƒû|ƒûy�ûx€ÿw�ÿs�ÿn}ÿgvû`pü[mõG^ÿ<Zÿ>\ìFZÎ`cž~gy…]i„Qh…OmƒRt…X|‰^}ˆ`x„\q‚XazP_Pa…UfŽZn”auœg} j~£mƒ¬r„²tˆ¶x‹¹y‹¹y�¹z‘º~”º�“±•¬€›©†©­”ºµ¢Ç¹®Ë¸²È¶´¸¬¬ª¡¢ —š[K;[K;[K;[K;[K;[K;[K;[K;ZJ:ZJ:[K;[K;[K;\L<\L<\L<]K=]K=^M=^M=^M;^O<^O<^O<^Q@^Q@]OB[OC\NE\OG]PJ]OO^P_bTmdY{i^‰pgœ}u®…{·ˆ{³™…¸¥…´¯€ª¸z¡ÆxŸÑuœÊdŒÆMlÜFRßA@ÕA=ÏE;ÌH;ÍG;ÓE9ÞA:î@Aõ<Aú=Dø?GñBIãBGÒ@AÆ?<¿?6¶<1°:.¬;-ª9)¯9+¼>2ÎD:Õ<7Õ31à;9åA?áD?èSLßUJ¿:+Â:*¿8%À7%Ä;)É=,Ê=,Ë9*Î8)Î5'Ò2&Ù2)å9/ñ>7õ>8ð50ä/&à5+Þ6)à4*ã1'ç.)ê-)í+)í++ê),è',ë'1ò)9÷'=ø"<ü>û%Aû6Hä(6ç.<ÿP\ÿ`lÿP\ò<Ió9Gè)8í*8ò.:ô0:ñ27ê01á+*Ø)&Ø0-Ú72Ô2/åC@ÿspõVSÝ==åGFÛ=<Ú<9Õ:5Ò80Ñ7-Ð6*Í4&Æ3#¾5#¹6$»8&¿;,À</¿9.À:/Å<4åXQÜMGÊ70Ç0)×<7ÞA:Û<6ß<5ç>9é=9æ?9á@8Ù?5Ï<2È:.Á9-»9,»<3ÌOIàc_ìnoú|ÿ„ˆý~…øy€øy€üy�ÿx‚ÿt‚ÿn}ýetù^n÷VhóE\ý8Vÿ<ZîH\ÇYZ‹mScrG]uC\yCdzIm~Qt�VvƒXs‚Yo�Wm‡ZlŽ\o•bu�h{£n�¨q„ªq…«p…­q…±tˆ´u‰·v‰·v‰·v‹·x�¸|‘´~�«{’¤~›¤‡­«–¾²¤È·¯Ç¸µ´«°¡ž¥—’™[K;[K;[K;[K;[K;[K;[K;[K;ZJ:ZJ:[K;[K;[K;\L<\L<\L<]K=^K=_L=^M=^M;_N<^O<^O<^O<^N>^N>\O?_OB]OB^PE^OJbNYcRecTqdZ}le‘xr¢w¨ƒv¤”}§¬†«Â‰©Î‚žÚw”ál‰ÛXwÚD_ôCUûEQïJPáIHÑF?ËE<ÐE>ÚEAéCCð@Cõ@GõBHíDIßCDÏA?Ã@8¿@7µ=/®:+«:*ª9)®8*»;.Ì?6áFBâ=;îBBëAAÞ;6äKCæXLÕK>ÉA3À8(¼3#Á8(È<-É:*Ë7)Ð8+Î0%×3)à8/é;4í:5í62ë/-å,)ã0,ã1-ç10ì31ó25ö37÷48÷6;ì-5ê,6í.=ñ0Có-Fö)Hý+Nÿ3Rô9Lå3?ð@Mÿ_lÿkyÿWfõCSõ>Pî3Dí3Aí3@ê7=ã99Û83Î4*Ç/$Ë2*Ç.(Ð72òYTö\ZòXVÛAAâHFÜB@Ô<7Î70Í6-Ð9.Ï9+Ê7'Â5#¹8%³8&µ:(¹>.¹;-¶6+¼7.Ç>6ÛNGÙEAÎ50Ð1-á>9ç@:ã81ç51ë20î21ë52ç83á:2Ø:/Ð9.È:.¾5+º7-ÈIBÜ_Yënlû}€ÿ…‰ýƒõv}öw~üy�ÿx‚ÿsƒÿk}ýdvù^pôRgôF_ý8Vþ=ZïL]ÀTR{`CWh<Yq?[vCfyKn~Qt�Vx…Zyˆ_wŒay•exšhz m|§q�ªr„«r…«p†©o‡­rˆ°t‰³tˆµt‡µt…·r‰·v‹¸w�¶{Ž¬xŽ£x™¤‚­­•À¹§ÎÀµÒÄÁ¼¹À¨ª¶œœ¨[K;[K;[K;[K;[K;[K;[K;[K;ZJ:ZJ:[K;[K;[K;\L<\L<]L<^K=^K=_L=_L=^M;_N<^O<^O<]N;]N;^O<_O?`P@`P@aQAbPFbNPcN]_Pe`Usga…to—{t�~r˜�v–®ƒŸÌ‹¡Û‚–ær…ëduéPdé;Rù3Lÿ<QûERéDKÕ?@Î@>ÑCAÖEBãEFèBDîAEíCFçDGÚDCËB<À@5»?3±;-©8(©:)ª9)°8*¼:-Ê;3èIEê@@ë=>å78Ú2/ßB;îZPôh[ÛQDÉA3¾4'Â6'È:,È9+Ë7+Ò8,Ñ0&Ú6-æ;4é;4ê40è/,ê-+ë--ó49õ3;ø3=ü3=ÿ1>ÿ1>ý1=ù1>õ0Aò1Dô4Kõ5Nõ/Pö-Qÿ2Yÿ>_æ3HùO\ÿanÿ_nÿWfûP`ôDXé8JóAQê:GÝ2;Ò/2Ê2-Â8-½;+½;+Ä:/À2(ÙJBúkcÝLGáPKÏ;9Ï;9àLJÔ@<Ë81Ê7-Î</Î=,Ë:)Ä;(´;&®;&¯<)³>-±9+¯3'º:1ÊC=ÖGAÙD>Ô72×2.ç;7ì:6è2.í2-ò,-ô,,ó/0ï31è71â:1Ú<1Ò>2Ã6,º4)ÅB:ØYSçjhú|}ÿ†‰þ€„õv}÷xýz‚ÿx‚ÿq�ÿi{ûbtø]oòPe÷Ibÿ:Xþ@\ðO_ºQNqX:Sf8]uEa|In�Sv†Y{ˆ]~‹`��g‚—l€œl}Ÿm|¤o}¨p€©o‚ªnƒ©l…©lˆ«qˆ®qˆ°r†³r„³oƒµp…·r‰¸t‹µv‰¬tŽ¥wš¨„°´™ÈïØÌÀÛÑÏÂÂÌ«±Á›¡±ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8ZK8[L9[L9[L9\M:\M:]L:^K<`J<_L=_L=^M=_N>^O<^O<^O<_P=aP>aP>aP>aP>aP>bOAeMMdMW`O_aUmgb€uq’|v˜€s‘�uŽ­€”Ά”ß~‡îqyùclúP[ú;Mÿ*Fÿ2Qÿ<Sñ<Mà<E×@EÕCDÔBCÜCEàBCäADãCEÞDBÔC>É@8¾>3¸>1¯9+¨7'©:)­<*´<,Á=1Ï?7à>;å99ã03â/2à42Û94çNFülaòh]ÙQCÆ</Æ:-È:.É7*Í6+Õ8/Õ4,Þ7/æ93é73é1/è/-î1/õ15ÿ2Cÿ0Eÿ-Fÿ*Bÿ&Aÿ#=ü :ö :÷&Cõ*Gõ.Mõ/Pó+Rô)Tÿ2`ÿAhóCXÿ`mÿhuøVeíKZñL\ëBUÚ2Cë@RÝ7CÉ,3º((±.$«8&¥?&¨@'¸?.¼:,ÝYMá]QÍG>ÍD>Á63É;7äUQÕF@É91È8-Ê;-É:)È9(Ä=)±>)¦<&¨;'¬=*«7(¬3(¼=4ÓLFÖGAÝE@Û96Ü30é73í62ë0+ô1-ý,/ÿ+-ú./ô1/í4/æ81ß;1Ø>2É9.½3(À;2ÏNHáa^÷yzÿˆŠÿƒ‡ùz�û|ƒÿ{†ÿx„ÿn~þew÷^põZlëL`õIaü9Wû=YîM]³KHhQ1Qd6\tDf~Nv‡[Žc„�h†’jˆ—n‰žsˆ¤t„¦t§r}¨p¨n�©mƒ©l…©l…¨n‡«n…­o…¯o�°l€²k‚¶n…·r‡³t‡­t�§wœ¬…²¹šÈƯ×οÙÒÌ»¿Ê¤¬¿’š­[J8ZK8[J8ZK8[J8ZK8[J8ZK8[J8ZK8\K9[L9\K9\M:]L:]L:^K<^K<_L=^M=^M=^N>^N>\O<aR?aR?aR?bQ?bQ?aP>aN=bL?hOKgNRbP\cWkjfwv•}|œ�x—‘z–«�•Æ„’Ú}…ðuzýklÿY[ýDLÿ0Jÿ5Tÿ<Uô=Qé?LãCMÝBHÖ=@ÛACÜ@CÞ@AÛA?ØC?ÐA9Ç>4½=0µ?1¬;+©8(­<,´?.»@1ÊA7ØC=Ù74á85ß//â30ç;7Ó0'Ñ5)ëUGÿvlë[PÑC7Ê<0Ë;0Ê8+Í6+Ô:0Ú70ß82å63ç32ê01í12ô36ý4<ÿ&7ÿ#9ÿ"9ÿ"9ÿ!8ÿ8þ6ø6ï4î9ð%Bð)Hî&Jñ'Mÿ3^ÿEhÿbuÿ`l÷WcñQ]òP]ïMZâ@MÔ2=Ú=FÌ7;»/.®/(¦5'¢;(œ@'œ>%ª9'²8+ÒVJ»;2Ä?8¾41½2/ÎC@åWSÕHAÈ91Ä6,Æ7)Ä5%Æ5$Ã:(±<(§:&¨7%­9*¬4&¯1%Â?5ÛRJÚIDáHCÞ;6Þ2.ê41í1/ï-+ü22þ,/ý+.ù+-ó++í-*æ1*ß4,Ö8,Ñ=1¿2(¼7.ÈIBÙ\Xówwÿ‰‹ÿ‡Šû€…þˆÿ}ˆÿv„ÿj|þ_s÷XlöTiëE]õD^õ6Uñ:VàLZ¥HC^K-Mc5Wo?b}Ju‹\‚•g‡˜lˆ›nŠ r‹¥vŒ¬z†«w�ªr~§m|¥i}¥g�¥h‚¦i‚¥k‚¨mƒ«m‚¬l®h~°i�³j„¶o‰¶s‰¯rŽ¬x™°„­»˜¿Å©ËʵËͶ½Åžªº‹—§]K7\K7]K7\K7]K7\K7]K7\K7^L8]L8^L8]L8^L8]L8^L8]L:]J<]K=^L>]M>\L=[N=]P?^SA`SB_RA`P@bQAcRBdQBdNAdLBjMIhKMdPYf[lnk†x{ž�„­‰ˆ²�ƒ©¥‡©¾ˆ Ð‚’ây~ðsqújbÿa^ùJWñ>Që:Lî?PïCQê@Mç=Hå>Fâ=Cß<AÚ<=Ö<:Ï>9Ê=4Á;0¹;-¨7'¯A0­<,¬6(¿A5ÌF=Í>6Ð72Ú85Ü71Þ5.Þ6-Þ8*Ý<*Ù:'Ó6%Ù;2ðSLòYQÙE9Ç8*Í>.ÔB3Ò:-Ô3+ðGBá//î5:ó6=î,7ÿBMñ'3ú&2ÿ)2ÿ(2ù'ý!+ÿ'0ÿ(0ñ'ð)ù+7ê .ì&7ñ-Cç#;ÿXsè.Eÿ]lÿ_iþZcùU\õPWëHMÝ?@Î95Ë>7ÄA7·?1¦7&š3"™6#š;'Ÿ<)§8-¯80ÂE?¿;7½31Å75Å54ÔDCÛLHÚKEÑD;È:0Ä4)Ç5(Ì5*Å5*¼</³9,­/#«(³-$Ã9/ÒC;ØH?äOHßE=Ú70Ý0*ç/-ò21ú24ý14÷/2ö/2ö01ó0.ò0.î3.æ5-Ý9/Õ=0¾2%À>1¹=3Ö_Yésqÿ��û„†û†ŒÿƒŽÿxˆÿh}ÿ_wÿ[uÿUoøLføEcñ:Yô;[æ?YÛ]i‹@;VF-L^6ZuBh‡Nw–]}œc�g€¡jƒ¦p‚§q€§n¦m~¥l}¥i}£h£f£f£f~¤k¦m�ªn¬k~­g®f�±gƒ³i‡µm‹´r�°v�®|š³‰¨¼™±Á¤²Á®¦²°�˜žz…‹]K7]K7]K7]K7]K7]K7]K7]K7^L8^L8^L8^L8^L8^L8^L8^K:]K=^L>`N@^N?[N>[N>\P@^RB_SC]QC]OB_NDaOEbPFcOHeNHiMJgNQeS_h_ton�y~¨€‡»‡ŠÁž•Ê£‹»«£¹x�Óyƒê|{ôvjöi`ø]aóOZìFRòHUùJYöFSï?Jë;Eä9Aà:>Ú::Õ=8Î?7È?5À</·<,«<+ª?-ª9)²:,ÁA6Æ=5È72Ô<7Ò50Ö5-×5*Ù5)Ú7&Û:&Û;%Ú9'Ý6.æA;ëMDâJ=Ñ?0Ç8(Ë9*Ô</Þ;4öJFå12ð6;ö6Añ.<ÿ@Nï'4ó$,÷%(û),ý+.ý+.ø((ö(*÷+,ï$'í%(ë&-þ<EÙ)ð4CüATØ&6ÿjuÿbjþZaüW]ôOSåBCØ=9Ñ>6¾8,¿D5¹H6©>,™/�)�(’(ž-%±;7ÑTRÎHGÃ54Ç56Ñ==ëWUØGBÖGAÐA9Ç9/Æ3)Ê6,Ð7/Ï81Ç=3Â91Â5.Ä3.É4.Ñ83Ø?9ßF@äJBÞA8Ú6-Ý0*ç/-ò12û03û03ô04ó12ô02ó/0ó0.ï2.ç4-ß9-×=1¿3$¾</·=2Ó^Wévsÿ��ù…ˆø…Šÿ�Œÿt…ÿf|ÿ\vÿXtÿQoþIhýAbó:Zñ?_âJaÁS\|@8SG1M_9[vCg†Ks•Yxœ_w�b{¡f}¤k}¦l|¥i|¥i|¤h{£g{¡d}¡d~ c~¢e|£j}¦n}©l}ªi|«e}­c¯c‚°e…°h‡²kˆ¯pˆ«s‹«|�¬„�ª‡Œ¢‹x‡€dppR^^]L:]L:]L:]L:]L:]L:]L:]L:^M;^M;^M;^M;^M;^M;^M;^M;^K<_N>aP@`P@]P?\O?\P@[RC\RF[QG\OG\OI_PMcQOfTTgUUhVVhW]h]nmh†sv£|ƒº‚ŠËŠ�Ô˜‘× ŠÊª‚´¹}ŸÏ€“䄆îvósjüfh÷S\íGQðFQöGT÷EQò@Lî=Gå8>à7:Ù99Ô<7Í@7ÅA4¾?0¶>-­>-ª<+«:,¸B6ÇH?À;2¼1*Ê:2Í6/Ð6.Õ7,Ø8,Ù7(Ù8&Ü9&à:*Ý4-Û4.âA9éOCÞH:È5%É3$Ù@2â>5õHDè13ò5;ù7@ø4@ÿESø2?ð)0ë#&ì$'ù13õ-/ï''ì&'õ12ë)*ö8:è+1â)1ì4@ÿR^ï=Mÿ\iþ^fñV\îOTñNSêEIÝ:;Õ74Ô?9À6,Á?2¼B5³=1¬7-¥3)¡,%œ' š%ª0+ÇECÂ::½--È35×ABô^]Ñ@;ÒC=ÐA9Ë=3Ì9/Ô=4ÛA9Û@;Û@<Ø=;Ý==ã@Aä>>à::ã?=ìJGàA;Ý<4Ù5,Ý2+ç1-ð31õ12ø02ô04ô02ô02ó/0ó0.ï2.ç4-ß9-Ø>2À4%º8+µ;0Ð[Tízwÿ“’û‡Šúƒ‰ý~‰ÿr‚ÿdyÿ[tÿTqÿNmÿFgû=aô<^ïFcßTg¢AHp;3SH2Qa<^yFh‡Nr“Zu˜^tš_xžc{£g|¤hy¢fy¢fy¡cx byŸb{Ÿb}Ÿb| cz¡h{¤j{§h{¨e{©a{©`}«`­b‡²jŠµn‹²s‰¬t†¦w‚žuy“np‡mTdYCOM4@>^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^M;^L8`N:aP<`Q>]P?ZN>YOCXPEXOH[QO_UTbXYeX_iYcl\fk^glaimerok‚ttšy~µ†Ê†ŒÚ�ŽáŽ‡×Ÿ‹Ó­ŠÄ³€©½yŽÌ{�Þ�zò‚wÿsrÿcgùU\ôMTõGQôCMò>Iï;Då6;à88Û97Ô<7Í@6ÆB5¼A1·?.«:*­>-±=0¼F:ÍSHÇH?º5.º1)È91Í81Ô<1Ù=1Û;-Ù6'Ü6&à8+à5-Û2+á>5ìNBäL>Ò:,Ñ8*àB6ß7.ð>:é/0ð16÷4<ü8BÿNZÿLUý>Eò38â$&í/1ê,,ð22è**ë/.ò::Òæ37Ý,2è8BðENÿXcòNWØ=CÛBEáBGá>Cß:>Ý8<Ø88Ó97É83Â91½:2ÀA:ÍNHÖVSÐNNÃCB²72¶95ÏKIÙMLäRSïYZëSRñYVÉ51Î>6ÒB9Ð@7Ó?5ÙB9àC<ãA>é?@ç8=ì9?ò?Cð;@è58ì>?öMJÞ93Û60Ú3+Þ3,å4.í41ò21ô01ô02ö/2ö/2õ/0õ/.ñ1.é4-à8-Ù?3Â6'·5(´:/ËVOï|yÿ“’üˆ‹ý�‰þz‡ýoücwÿXqÿOlÿFhÿ@cù9^ô?`éKdØZh…35g;0WO8XhDa{Kj‰Ps’Yt•\u˜^w�by¡ez¢fw dv awŸavž`xžaxža{Ÿa| cz¢fz£g|¦fz§dz¨`z¨]|«]~­_†²g‹´nŒ³tŠ­s†¦u€�qtŽii€dRbUFSL:G@_N<_N<_N<_N<_N<_N<_N<_N<^M;^M;^M;^M;^M;^M;^M;^M9]K5_M5`P9_P;\O?ZNBXOFXQKYPQ`W\g`hnfsshysg{sg}ri~pm~pq†sv—x|«|�Á€†Ò…‰àŒŒæ•�ã�ŒÚ „¿žuŸ l‚¬nq¼vlÒviòsmÿllÿgiþ^`ûRYõFMï;Dê7=ç6<â89Ý98Ö=7ÎA7ÆB5½B0¸@/­7)µA4°;1°;1ÈRHÕ\SÉME»;2¾90Á7-Ì9/×?4Û=1Ú6*Ü4'â6*à3,â70ä=5åE9âF9ÞB3ÞA2â>2Û0&ê72ì0/ï.1ñ.4ú7?ÿPXÿ]dÿZ_ÿOSá.1ã03à,/÷DGè66Ü,,Û-.ë>@Ý36Ù37ÿbhà<CË*2Ê-4Á&*Õ;=á@EÛ6:Ù37ß9=Ü7;Í/0Ì43Í;;ÕGFÞRSì`cõhnôekê]cÄ@>ÆE@å^[øhgÿopÿjjåKK×=;É40Ñ>7ØE=ÖC;Õ>5Õ<4Ø93Ü43ï6<ñ-9ð,6ò.8ð,6ë,3ó9>þJKß3/Ü3.Ú3+Ü3,ã5.ê40ï4/ò21ô02ö/2÷/2ö.0õ/.ñ1.ê3-â7-Ù?3Æ8*´2%²8-ÃNGï|yÿ’‘ÿˆŒÿ‰ÿw…úl|øatûTnýHgÿ>bÿ:_ú6\ôBbàNeÂWaq0,_?0[W>^lIf~Nk‰St’\u–_v™_xžcy¡ex dvž`vž`u�_vž_wž_xŸ`{Ÿa| bz¢d{¥e|¦d|§by§^y§\{ª\}¬^€¬a…¯g‡®m‡ªp…¥sƒ r{•nrŠjfwe[k^RbU_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N<_N:^L4_N4`N8^O<[N>YOEZSMZTT_Zahbprm�zuŒ}t‘{q“zp“vp’qt“rxšu|¨y€¶{ƒÄ…φ‰Ú�‹Þ–�Ü™ŠÍ›…¶¢ƒ£ª„‘«~yŸl[›R?¼MBÔQIâWRëWUñSTñKOí>Cæ5;é9<æ9;à;9Ù>9Ó@8ÊB6ÁB1º?/´:-ºA6«5+£.$¾I?Ùg\ÙdZËUK¹=3¸5+À3)Î:0Ù;0Ú6,Þ3)ä6-ß1*å82ã;2Ý9/Þ>0ãF7äB5ß9-ß1(ê5.ð31ð./í*.ô17ÿEJÿV[ÿ]bÿ\aâ9<ã9<Õ+.ôJMâ89Ø00Û57Ì()Ì'+øW\Ý>CÖ:>Ò7=º#Å),Ü<>çBFÝ49Ù,2Þ17Ú16Í(.Ò37äKNöaeûhnùcl÷_kú`løbkÖLLÕNJê\ZêVTçMMãHFÏ42Î3/ÙA<àKDäOHßJCÙ@8×:3×50Ü.-õ1;û,<ø*7õ'4õ)4ø0:ÿ>CÿJLà1.Þ3,Ü3,Ý5,á4-ç4/í4/ñ42ô02÷/2ù.2ø-0ö..ò0.ë2-â7-Ù=1É;-´2%±7,»D>ìyvÿ��ÿˆ�ÿ|ˆÿq�öhxö_rùRlùDcÿ9^ÿ6]ú8]ñGdÑN`¢IMe5+VD0[Y@^lIe|NkˆRt‘[w•_x™`{ždz evž`s›]s›\s›\tœ]v�^wž_{Ÿaz¡bz¡b{¤b|§b{¦_x¦[x§Yz©Y{ªZ~ª_ƒ­e…¬k‡©m‰§s‰¥u„�u|”rvˆpnlfwe_O?_O?_O?_O?_O?_O?_O?_O?^N>^N>^N>^N>^N>^N>^N>_N:aO7aP6aO9^O<[OC[RK\VV_Zakgvso†}z—‚~¡�|¥}w£{u¥xv§sy©s}°z�¸}ƒ¿€†Ä…‰É�ŽÎ”“Íš“Ç •½«›µ¼¦²É­©Å¦”ªˆmœfJœE2¯A4·A5ÃC:×IEéOOîJKì?Cë=?é;=ä<;ß=:Ù@:Ð@7È>1Á=0½:2¼@8¯4,§1'¼J@ÑbWÖg\ÕeYÃOBº>2º4)Ç7.Ö90Ù5,ß2+ç60ã1-ä50â7/ß8/à>1æD7ä>0Þ4'æ4*é0+ð3/ð..ï-.ò14ó78öBEöMRüY^éFKöQWÙ48ëFJÝ8<æAEÚ7:Ï/1Ñ05äHKÓ8<Ð7:¹#%Õ:>Û8;å;>ë<Aè7=â-6Ý)2á1;ç<DðLSý]eÿksÿjsú`löZhùXh÷[hôddèZVèWTØ@=Ï10Ô44Ó33áC@ãJEéRKêSLáJAÙ?7Ø93Û81ã33ó.8ÿ0>ÿ0>ú*8ü.:ÿ;EÿCHûADã1/ß4-Ý4-Ý5,â4-æ3,í4/ò53ô02÷/2ù.2ø-0ö..ò0.ë2-ã7-Ø:/Í>0µ1%±7,²;5ésqÿŠŠÿ‡�ÿv†þi}óató[p÷PjúCbÿ8`ÿ8aù;aìLh¿O[~86]=0NF1VX@ZeEcxMj…RrŽ[v”^yša{ždxžatš]p˜Zp˜Yq™ZqšXt›\uœ]y�_xŸ`y _y¢^z¥^y¤\x¤Yx¤Wy¥V{§X©_‚¬d„«h†¨l‡¥o†¢q€™ox‘jr…il~ddv^_O?_O?_O?_O?_O?_O?_O?_O?^N>^N>^N>^N>^N>^N>^N>_N<cQ9cQ9aP<]P@ZPG\TR_Zaa_mtrŠzy™�€¨ƒ�°}¯{x¯zv±wwµw|¾x€Á}…Ä‚ˆÄ‡ŠÁ��¿—–¾ �¼¬§»¶¬µ¾³±È¸«Ð¾¦ÖÁ¢Ñ»–Ьˆ¯oT¯]H¥H6¨>0ÂG?àWQïWVñNOé@Cê=?æ<=ã><ß=:Ø=8Ñ:1É90Á80À<7¹:3·>5ÂPFÇXMÇ]PÑeXÓcUÃM?¼9/Å8.Ò91×4-ß2,ç60æ40â3.à5.ã<3çC7æB6â:-ß3'ê5,æ.&ð0-ð0-ð31ñ54ç-.ã35çBF÷X]öW\ÿouåDJëHMá=Dþ[`ÿouß>Cÿ}‚ÿmqêQTÔ>@Æ02Ö:=óILî9@ñ8@ø=Fò7Bè-8ô>Jÿ]hÿmwÿjrÿcn÷]gú^kÿapÿ]nùYeò\]ãOKâJGÕ74Ó01Ù54Ò0.Ø95ßD?ãJDãJBÙ@8Ó6/Ö4/Ü71é77ê'/ü0<þ2>ø,7ù0:ÿ>Eý>Cì44â3.à5.Þ6-ß4,á3*è3,ï4/ô43ö/2÷/2ù.2ø-0ø..ó0.ë2-ã7-×9.Ñ?2¶0%²6,°61çonÿ‡ˆÿ†Žÿp‚ücxñ\pôXn÷OiûBbÿ9aÿ<döAdäRi°QWd/'WD3GG/RT<U^?cuMi�Qq‹[v’_z˜b|�dxœ_r™Zn–Wn—Un—Uo˜Vr™Zt›\w›]wž]wž[x¡[y£[w£Xv¢Wv¢Uw£Ty¥Xz¤Z~¦_�¦c¡d~œf{—frŒ_k�Z`vR[nPSfH\L<^N>`P@`P@_O?_O?`P@aQAbRB_O?]M=^N>aQAcSCbRB`Q>dS?aP>]P@^QI[QPZSZeapsq‰yxšz|¥~€±€�¹~~ºyy·sr´nn¶twÈv{Ì€ƒÊ‹�Æ–—䤾²°»¼¹´ËƳÓÌ°ÙϬÛϧÜÏ£ÝУÙ˜٘Ѩˆ¸€g¡_IËzgÖtg½H?×QN÷ccéKLå?Aâ89ä:;â:9Ü75Ù53Ö;6Ë73À50ÇB;¹=3¿I?ÆXK¹OAÎdVÑcTÌXIÅG;Ä:/Í6-Ø7/á6/ã4/ß0-ß2,à70ä=4á=1Ý7+ß5(å7,å0'ê/(ì/+î1-î20ë40è51ã75×79æJNõY]îOT×8=×6<ùX`ÿ~„ÿ~†ÿouÿgoÿipÿflõ\añX[ûZ_ÿU\ÿDLü8Bû7C÷3Aï,:ì0>ñ=IÿS_ÿ`iÿgrÿepþbmÿbpÿ`oý]gíSSßFAÖ85Ö41×11×/.Ø31Û94áD=ÜB:Ù?7Ö<4×90Û81à93æ66ñ7<ô5<ö4<ö4<÷5=ò59ë35ã1/Þ3,Þ5.ß7.á6.ä3+é2,ï2,ö31÷03ø03ú/3ù.1ù//ô1/ì3.å7.à?5Ì8,¿6,¯/&¸;7ÒXWÿ˜šþ{ƒÿm€ÿe|ø\tðPjõHfüCeÿ;dù9`õNlÓRd‰=?R,!C;&CG.KM5Z`DZjEcxMqˆZx’bx–`t•\r–Yq˜Yl•Sm–Rn—So˜Tr™Xt›Zuœ]uœ[t›Xt�Wv Xw¡Wv¢Wv¢Uu¡Tt S{¥[{£]zž^u—[m‹Ub~MXrCSj>AW0?U/<R,\L<^N>`P@aQA`P@_O?`P@aQAcSC`P@_O?`P@bRBcSCbRB`P@fVG_QF]PH^TSaZad`omkƒvw–xz£{}®}¸x|¹rv¶lo´mp·oq¾uvÐ|~Õ‰‰Ó˜˜Î¨§Ç¹¶ÁÉźÓαÛתâÛ¥çÞ¥æÝ¢äÛ¤âØ£ÛÑ ÔǛӼœ»œ€©|eÅ‹wÊ}m´VJÒa[örpìZ[çKLã?@è>?ë>@è<<â66Û55Ö=8É83Ê?8¹6.¸>3¾LA¶H;È]MÝo`Ô`QÆH:½5)Å1'Ô6-Ü5-Û0)â51å95ä=7à<3Ü8.Û5)ß4*ã5*ì7.í5-í2-ë0+ç0,ã2,á4.Ý52Ö87Ø>@åJNóW[ú]dü\dþ]eÿaiôS[ïNVñQ[ÿ_gÿjqÿflõZ`óNTí4:ñ,5ô*6÷-9ñ'5å!/æ'6í7DþO\ÿZdÿblÿdmÿenÿcmý[hõQZåEEÞ<7Û64Þ63ß55Ü30Ú2/Û62à>9Û>7Ù<5×:3Û81ß82å95ç:6ë8;ë7:í6:ï6;ï79ì57å31ß2,Ý4-Ý6-ß7.â6,å2+ê1,ò1,õ20÷03÷03ú/3ù.1÷//ô1/ì3.å7.ß<3Ô=2Â8.¶1*·74ØZ[ÿ�“ûs}ÿfzÿ^vùVq÷NkúGgü@cü<cõAdåOhºLWy<9L2#<:#<B(EG/PT9VdA`rJnƒXu�_v’_t“Zq“Wp”Tk’Ok”Pl•Qm–Rp—Vr™Xsš[sšYuœYu�WuŸWuŸWtžTržQq�Rq›Qo—Pl”NiŒLb�FXu?Ni6F]/@W+@W-AX.AX.YL<[N>^QA^QA^QA^QA_RB`SCaTD`SC_RB`SCbUEbUE`SC^PCbUM\QM[QRcYbkdtqm„vu•z|¥y{¬y~¶x|»ot¶ei°bf¯jn¸ruÆ}}ׇ„Ý—”Û©§Ø»¹ÏÍÉÈÜØ¿åá´éäªíè¦ðé¥íå¤çá§âÛ§×Ï¡ËÞ¸¯’©›�›‚l¨ƒp§o`œRGº]UÝmi÷uuï__äJLæAEëADì>@è8;ã68ß=;Ô;6ÔA:Ã81·7,·>3®=/·G9Ûj\ßi[Ú\NÊB6Æ3)Ñ4+Ù6-Ù2,Þ52éA>ìGCà=6Õ2)Õ1'Û3(Þ0'ç4-è1+ç.)æ/)ã2,ã80á=4àB9Æ-(Ê43ÜCEðVXú^b÷ZaóS[ñQ[éGRêHSïO[ü\fÿmvÿntôY_ã>Dé.5ô-4û2<ü2>õ+9í'6ö6EÿKW÷ISõPWøW_þ^fÿckÿ`i÷RYéBIÞ97Þ71ã75é<8ì<<æ95á51Ý60Ý:3Ù;2×90×90Ü71â94è;7é<8ã99ã99æ87è88è88æ66á40Û2+Ú6-Û7-Þ7.â6,ç2+ë0)ó0,÷10õ13õ13ø03÷/1÷//ó1/î3.ç6.Þ:1ÜD9È91»4.µ/.ä_bÿƒˆølwþbxþWsúPmþLlÿBh÷:`ô<`ìIfÔSg�DJm<5O>,=<':?(ED/FH0Q\<YkEg|Qr‰[t�\q�Wm�Rl�Pi�Mi‘Kj“Mk”Pn•To–Up—Xq˜YuœYu�WtœVrœTpšRm—Mk•Mj”LfŽHaˆE\~AUt;Li3E`-AX*>U'D[/E_2Ic6WJ:ZM=]P@^QA^QA^QA_RBaTD`SC`SCaTDaTDaTDaTD_RB]QEZOM[RWbXcjcsqm†vu•xz£z|­w|´sy·mr´dh¯]d¬ag±lr¾x{ʉ‰ß•‘夡䷴áÇÄÙÕÑÎáÝÄçä¹ëæ¯îê­îè¬éâ«ãÛªÚÓ©Ë¡»¶™¥¥�›˜…“ˆv˜�q™rc›bW³i`Ïrmû‹‰õutê[]æLNéDHí@Dì;Aé;=ß:8Û<8ßF@ÓC;Ã:0º:/°6)§1#¿I;ÛaRêj]ÛSGÊ:/Ì4)Ô6+×4-Ù42ñMKþZXíKFÚ91×4+Û4+Þ2(ß.&â-&á,%Þ-%Ü1)Ú6,Ù;0×?4Â/'Ä4,Î:8ÙCBßFHàEIäGNêMVñQ]öVbøXdù\gÿgrÿmuôYaÞ;@ë28ñ.4ó.7ò*4ë%2í)5û<KÿS_í?IèAIêGLôQVÿY`ÿX]ôGMå79á53å84ë<9ò@>ô@?ï=;æ95à72Ý82Ú91Ù80Ù80Ü71ä84è96ê;8â:7à;9ã:7æ:8ç98å84à5.Ú3+Ú6,Ü8.Þ8,â6,ç2)ë0)ò/+ö0/ô23ô23÷03ö01ö0/ó1/î3.ç6.Ü8/àF<É91¿4/¶,,îfjÿy€øgtÿbyýTqþKkÿGjÿ=eó6\ë=^ãQhÄXeˆBBgC7[N;GF1AC-JH3EE-LU6Sc>btLlƒUp‹XoŒTmŒPiŽKiŽIh�Jh‘Kj“Ml“Rn•To•Xp—Xt›Xs›Uq™So—Qk”Nh‘KeŽHd�G`‡D]�AWy=Rq8Mi6Lf6Oe7Of8Ne7Oi9Sm>UH8XK;[N>\O?]P@]P@_RBaTD_RB`SCaTDaTD`SC`SC_RB^QIVMR_Xhnf{uqŠxt—xw¡xz«y{´tx·lq³bf­\c«`f°kq»v|Æ€„Γ•àŸžä®­ç½¼æËÉßÔÓØÝÛÎáßÆåá¾æâ¼äß¹à׶ÚеÎƱ¾µ¦­§›™™�ŒŽ�‹…yŠ|q‘vk›si¨meºoi扄î�~ósrñedñWYðMNîDGì?Aã:7à;7äB=ÞE?Ì<3Æ=3ÄB5°2$®0"ÈJ<ÜXKÖN@Ì=/Î7,Ñ7+Ñ3*Ö42õUUÿkhú]XãD>Ü;3ß;2ã80â4-ã2,â1+Ý0)Ô0'Í/$Æ/$Â0#Å9,À4'Á2*Ì;6×CAÛEGÛBGÚ@HàDOìP[òVcòVaú^iÿgoóYaáBGê:<í57ï28ï28ò2=÷9CÿHUÿVaå7@Ý4;ß6;ëADùJOüIMõ<Bé13é54ë95ð=9ó=:ó;9ð:7é73ã81ß;2Ü;1Û:2Ü92ß82ä73è64é75æ:8ä;8æ:8è:9è;7æ95á6/Ü3,Û7-Ü8.ß9-á6,å3)é1)ñ0)ô1-ó23ñ33ô23ô01ô1/ó1/î3.ç6.Þ7/àC:Ê70À1-À13økqÿqzúftû]túPmÿBhÿ>gÿ9cõ8^æEdÙZm«V[wE>dJ=aYFTO;KI4PI7GE0GP3M]9ZlFf{Pm…SmŠRlŒMh�JiŽIg�Hg�Ji’Lk”Rl•Sn”Wo–Wt˜Xs˜Un•Rk’Of�KcŒHaŠF_ˆF[‚CY}@St;Pn8Ok:Rl=Xn@ZqCWo?Vq>WqARE5UH8XK;ZM=[N>\O?^QA`SC`SCaTDaTDaTD`SC`SCbUEbWQ^Vckf}yu�~z›zy¡xy©vy°txµlq³cj°\c«^d®io¹w}Ç�‡ÑˆŒÓ™�Ý£¥à°²ã¼¾åÆÇãÏÎÞÔÓØ×ÖÑÙ×ËÙÖÇØÐÅÒÉÀÌþŹ¹´¨¬£š�Š‹†yztyvquleleŽphŽa[�`[ÀsmØyuð~}ü|yÿrqýccöSTðHHðDBê>:á<6áB<Ñ:3ÓC:ÞUKÂ>2¸6)½9,Â=.È?/ÐB4×C5Ö@1Ñ7-Ñ30êLMú_]òWSßD?Ù:4Û81á81æ93ç92ç92â:1Û:0Ñ9,Ê8+Å9*±)¸3"È@2ÓI>ÒD@Ë;:Ì7;Ï:@Ò8BâHRíUaðXdõ^g÷`iêU[ÜAEå@>ì?;ó@CûGJÿLTÿQYÿT]ÿV`á3<Ú/5Ý04è8;õ>Bø;?÷48ó04í42ï96ñ97î53ë20ê20ç40á4.ã<4à=4ß<5Þ;4á83ä73è43è43é99ç98è88ë99ë97é75â5/Þ3,Ü8.Ý9-Ý9-ß7*ä3)è1)ï0(ò1,ñ33ð43ó23ó11ó1/ò2/ì3.ç6.à91Ù<3Ì71Á0-Ñ@CÿnuýkuùbqðRiõHfÿ;cÿ5aÿ6bü?eäNiË]j‹GFjF:^M=^WDYR?RJ7OF5ID0DJ.GU2Sc>^sHgMj‡OjŠKgŒGgŒFfŽGf�Ih‘Ki’Nk”Rm“Vm”Uq•Up•Rk’OgŽKcŒH`‰E_ˆF_†G^‚E[~DYwAVr?Uo@YpB]rG^uG[sCYtAXs@NB2QE5UI9WK;XL<ZN>\P@^RBaUEbVFbVFaUE`TDaUEfZJi^ZngwupŽ|x™{y wx¦vw­pu¯jo¯ch¬`g¯`f°em¶qxÀ~…͉�Ø�—Ø™ Ö¡¨Öª±Û´¹ßº¿ßÀÁÝÃÃÛÇÆØÆÃÔÅÁкɼ³Ä»°Áµ¨¼¦˜¯–‹œˆ…Œqqquqppfewhe†nj„_Y•d_›ZT¹f`Ùsoðzvÿ{yÿvuÿhhüZXþRNõHBæ=8èE>Ù<5ÛG=ê\RÅ;0Ä<0½7+¿7)Ç>.ÑC5ÔB3ÔA1ØB4Ò86Ø>>ÝCCÜB@Ö=7Õ81Ø7/Þ7/å:3æ93å:3â;3Û>5ÖB6ÑE8ÌG6®-¼;%ÍJ8ÑL=Æ=5»0+À00Ê9>×AJâLWëU`ðZeð]eìYaßLTÔ>@Ï4/Ø7/à;9ê@@òCHõDJñBIî?Fâ5;ß26ã36ï8<ö8:ø14û/2ü14ï20ð95ñ85ì0.è,*ç0,æ3/á4.ä=5á>5à=4à<3ã:5æ95ê65ì65ì57ì57í57î68î66ë54ä2.Þ1+Þ7.Ý9-Ý9-ß7*á3(æ1&í1(ñ1,ð43î53ñ33ò21ó1/ò2/ì3.ç6.â;3Ò4+Ò:5Æ20èVYÿpxùcnòZiçI`óEfÿ8bÿ0^ÿ5bÿBhÜLe±PYk60^G7XK:UN;[O?VI9M@0JC0@F,AO.L\7Xj@bxGf�JhˆIf‹FeŠDc‹BdŽFeŽHg�Lh‘Oj�Sk’Sn’Rm‘Qh�NeŒKb‰HaˆG`‡H`‡Ha…H_‚H_}I]yH]wH`wKcxOczN]uC[vA[vAJC3MF6RI:TK<WK=YM?\O?^QAdWFeXGdXHbVF_UIbXNg^Ulcfwr‰zv›zužtsŸss¥tu­lp¯bg©\`©ae¯fl¸ntÀx{ȃ†Ó�“Ý—œÞ›¡Õ §Ó©­Ú®³Ý±´Ý²´Û´´Ú¶´Ûµ±Ö³­Ñ®§É«¡ÄªžÂ¥™½™‹¯ˆ~™|x†dbgiefdZ[j[X{c_yVP’c]ŠNF¢UM½`YÖkeñwrÿ{ÿzwÿnjþa\÷TOéGDîOKÝB=ÞG@êWOº,"¾1'¾4'È?/ÑE6Ë>-Á2"Ê7'ÛE7ÞE?Ó97Ì3.Î61Ô;5×=5Ý<4á=4ä;4ä92Þ5.×4+Ï5)Ê8+Å<,Â?-ÕT?ÈG2½<)¿;.Ã=2Ä;5Â74Â35ÜKPÞMRáPUåTYèY]åVXÜMQÓCBÉ5+Ñ7+Ù80Ý52â24å26æ18ç/9ì4>ì3;ó5?û9Bý4>ú+5ÿ(3ÿ0:ø-3ù59ù59ð./ë*+ê1/è43â51ã<6à=4à=4á=4ä;6ç:6ì87ï77ì46ë35ì25í34í55ê41ã1-Û0(Þ8,Ý:+Ý:+Þ8(à4&ã3$é3&ì4,î52î53ð42ð3/ò2/ð3-ì3.ç6.ä=4Í/&×>9Ì35ùbgÿoy÷[iðNcêC]ùEhÿ:eÿ0]ÿ6aöFjÉJ]“BHS-$WJ9SL:OH6[O?UI9F?/HF1>B'@J(IU1Sd:_rEe}Kg„JeˆHcˆCa‰Bc‹Dc�EeŽJgŽMh�Pi�Ql�Pk�Mf�JcŠIaˆGb†Fb†Hc‡J_€G^H_}I`|Ka{Kd{Mf|Nf~N]xC]z@^{CDB6FD8KG<PG>RH>XJ?]M>`P@`Q>aR?`SB_VGcZSgb_ojnso~wq“zt ~x¨yw©pp¦gg£`a¢\_¤`b­fhµopÀyzÌ„‚ÖŽŠß–’眙栞ݡ¡×££Ù¤¤Ú¦£Ø¦¡×¥¡Ô¦ŸÓ¥žÑ¤œÍ¢šÉ¢™ÈŸ–Ù�»Ž‚°ƒ|žrn}igld^`g[[kYUrWP}WNŒWO•RI©[QµXPÁZSØkdìyrù‚|ÿ‹†ùuqÿusí_]àQMÙECïZTÌ71Å2*Å7+Ã7(À7'Ã7&Æ9(Ê;+Ñ=/Õ?1Ò8.Ö<2Ø@5Ô<1Í5(Î4(Ø:/ãA6à90á90Þ7.×7+Ï7)É:*Æ?,ÄA/Â=.Æ@4ËE:ÍG<ÌC;Å<4¼3-¶+&Ç<9ÐEBÚPMàVSãYVáXRØNKÐC:Ì;*Ó:(Ü8,â5.è./ì*2ð(3ô'8ú->û,@û*?ý)?ÿ(?ÿ&?ÿ%=ÿ%=ÿ#:ÿ':ÿ+<þ,9ô,6í.5è45ç;9Ý84Ý:3Þ;4à;5ã:5æ95é73ë54ë35ì46ë54é54ç53â5/Ý5,Û5)Ü8,Û9*Ü9(Þ9&ß8&â7%ã6%æ4&ì5/í41ð50ñ4.ñ4.í5+ë6-å7,à90Ø7/Õ:6Û?BÿmwÿbsûSjõIcÿKjö6[ÿ6_ÿ<gø>cçNj¯IV^$#E-!?:'HE4PM<PI9EB1>B1?G/BH&EK%KS.S]8[kDavKd~NdƒJc†F`ˆBa‰@c‹Be�Fj�Jl�Pn’Rl‘Lj�IgŒGf‹HfŠJf‰Ic…HaƒGe„Kd�IbIdJgKg€Ie~Ga}C\|=\=_‚B=?4@B7FC<JE?PE?TG>[K>^N>aP>`Q<^SA^VIc[Xhcinlyso†uo•vpžtp¢pm¢gfŸaaŸ]^¡]]¥ed°kjºusÆ}{φ‚ÙŽˆà”Žæ˜“å›–Ùœ˜Õ�™Öžš×ž˜Ö�—Õž•Ô�•Ñž–Ñœ”Íœ“Ê›’Çš�Ä“‰»ˆ®‚x�tk|mdif]^f[YiZSoWM{UJŠVKšVK±]S»ZSÁXRÓfaàqjãvoí~wý‡…ÿ‡…÷uuîgdãUSëZUÇ2,Â/%Å7)Ã7&Á8&Á8&Ä8'È;*Ï;-Ó=/Õ9,Õ9,Ó:,Ò:,Ñ9+Ñ9+Ô8+Ø8,ß;1à90Ü8.Ö8,Ï9*È9(Á:&¾9&¾6*Á80Æ=5É@8ÊA9È?7Å<4Ã:2¿4-Ç>6ÒIAØOGÚQI×NFÌC;Ã7*Ë8&Ó8$Ý7)å4,í//ô+3ù)7þ);ÿ+Bÿ*Aÿ)Aÿ'Bÿ#@ÿ!<ÿ=ÿ;ÿ:ÿ ;ÿ%;ÿ);÷-9ð19ê7:å;;Ü94Ü:5Þ;4à;5ã:5æ95é73ë54ì46ì46ì65è64ä71ß6/Ü5,Ù5)Û9*Û9*Ü9(Ü9&Þ9&ß8%á8%ã6&ç5+ê3-ë4,ë4,ë5*ê5*ä7)â8+à<2Õ3.Ó54óTYÿesÿ[pöHcÿIhÿAdÿ;`þ4\ú<bòMmÍLb‰8?S*&A5)7:);<,B@1F@0@?-=B.?H-=CKP'\b<fpKhvRf{Rf}Qe�NaƒFa†Aa‰@cŒ@gŒFj�Jk�Mm’Om’Lj�Gh�HgŒGgŠJgŠJf…Je„IfƒKd�IdHdHg€Gg€GfEb~A^�?_„?b‡B7:/;=2@@8EB;KB=OE<VH=YL<^N>]P?\SD^WMc]]ifmom{sq‰so”qk™jf˜c`•\[”ZZ˜\] \_¦hjµno¿wxÊ�Ò‡ƒØ‹‡Ü�Šà’�Ý“�Ó“‘ÑÑД�Ñ“ŽÐ“ŒÏ’‹Î•ŒÍ“‹Ç”‹Ä”‹À“‹¾Œ„³ƒz¥€r•|ewu^fi[[e\Wd[RgYLrVH„TH£ZQ½`YÉ\YÐZXÛcbákißokãvqûŠ†ÿŠ†ÿ…ƒÿ}y÷nhîaXÅ5,À2&Å9*Á:'À9&À9%Â9'Æ:)Ì:+Ð:,Ö:-Ô6*Î5'Ð8*Ô>/Ô>/Ï7)Ë/"Ú:.Ú8-Ù7,Ö8,Ð:+É:)Â9&¾7$¾6*À6,Â8.Å;1Ç=3Ê@6ËA7ÌB8Ç:1ËA7ÏE;ÐF<ÒH>ÑG=Ê@6Å7+Ì6%Ö6&ß7,è50ñ03ö-7û+9ÿ*<ü'=ü'=û%=ù#;û!:ú7ü7þ6ÿ8ÿ6ÿ$8ý*;÷2<ñ6=ë8;ä::Ü94Ü:5Þ;6à;7å95ç85é75ë54í36î47í55ê65æ72á6/Ü5,Ù5)Ú8)Ù9)Ú9'Û8%Ý8%Þ7$à7$á6%á5)â4)ã5*ã6(ä7)á7(ß7*Û7+Ý<2Ð1+Ø88ÿkrÿ]nþPiñ<[ÿHlÿ7]ÿ>eù5[î>bæVqª?Qa#(L/+;7+18(37&:;)B=*A<(@>)BB&?FXa4s~T€Œdw‡`i€Tb|MaI`‚E`…@a‰@d�Ah�Gj�Jk�Ml‘Nn“Mk‘HiŽIgŒGh‹Kh‹Kh‡Lg†KfƒKe‚Jc€Hd�If‚Hf‚He�Fb€Bc†DeŠEh�H25*69.;<4A>7G@:KB=SF>VJ>XL>YM?YQD^WOc^bigron€roŒrm•jg”a_�WW‹TUŽWX˜\_¢ac­ln»rsÃ{|΂ƒÕˆ†Ù‹‡ÛŒˆÜ�Š×��Ò��Î�ŒÏŽ‹Î�‰ÏŽˆÐŽ†Ï�…Î�…ÊŽ…Æ�…Á�‡¼�‡¶‰‚¬€x�oŒ€_p|Y`lXYd[T^_Q_\IjWF}SEžUL¾\YÑZ\ÙX\äcgçkkáplåxsì|xízuþƒ~ÿ†�ÿ�{õkaÉ<2Ä8)Â;(À;(¿;&¾:%Á:'Å9(È9)Î8*Ù;/Ô6*Î5'Ï9*Ñ>.Ñ>.Ë8(Ç/!Ò6)Ó5)Ó5)Ó7*Ñ9+Í<+È=*Ä;)Ä:-Ä:/Æ90Æ9/Ç:1È;1É<3Ë=3ÔE=ÓE;ÐA9Ë=3Ì=5ÏA7Ï@8Î;1Ï6(Ù5)ã6/ë31ò/5÷+6ú*:ü)<ù&9÷&;ø'<ø'<û&<ý%:ÿ$:ÿ#:ÿ#8þ#7û&8ù,;÷5>ò9?è8:à87Ü94Ü:5Þ;6à;7å95ç85é75ë54î47ï58î66ê65æ72á6/Ü5,×5*Ù9+Ù9)Ù9)Ú9'Û8'Ý8%Þ7%Þ7%Û5%Ü6(Ý7'Ü9(Ü9(Û:(Ù9)×:+×:1Ñ4/ìILÿoyÿVjùE`ø<_ÿAgÿ3\ÿ8bø>cèMlÅOe€0=J C1-11'/7(5<,>A.B@+B<$E>$IF%U\0p~M�Ÿr—¨|ƒ˜mj„W_{J]~E_ƒCa†Ac‹Bf�Ck�Jm’Mm’Om’Oo”Nl‘KiŽIh�HiŒLiŒLj‰NiˆMg„Le‚Jd�Ie‚HgƒIh„IgƒHd‚DgŠHiŽIl‘L.4(36+891<;6B=9H@=MD=QG=SI=SK>UNDZUQa^eigumk�lkŠki‘b`�XVˆPP†QQ�XYš`cªfj´orÁuxÉ|Ѓ„Ö†‡Ù‰ˆØŠ‰Ù‹ŠÖŒ‹Ï‹ŠÌŠ‰Í‰‡Ð‰…ш„ÑŠƒÑ‰‚ÐŒ„Ï‹„È�…Á�†º�†°‡�¥~w–m…„Yj€SZnVTc\RZaOZ_HeZDxUB“ME¸TRÐSWÛRYå^dèejãklæuqäunâoh÷~vÿ‚{ÿ‡~ôl`É?2Á8(À;(¿<(¿='À<'Â;'Å:'È9)Í7)Ø:/Ù8.Ò8,Ï9+Ì:+É:*È9)É7(Ï9+Ñ7+Ò8,Ñ7+Ï7*Ì8*Ë<,Ë>-Ê<0É=0Ê<2É;/È:0Ç9-Å7-Å5*Ð@7Ð@5Í=4É9.Ë80Î;1Í:2Ì5,Ð/%Ù0)ä20ì25ô/8ù-9û+;ü+>ü-?ü/@ü/@ü/@þ/Aÿ/?ÿ.@ÿ.@ÿ)9ü)8ø,8õ0:ñ6=í8=ã77Ü43Ü86Ü:7Þ;6â:7å97ç77ê65ì44ñ48ñ48ð67í76æ74á6/Ú6,×5*Ø8*Ö9(Ö9(Ø9&Ø9&Ù8&Ú7&Ù8&Õ8%Õ:&Ö;'Õ=(Ô=(Ô=*Ô=,Ô<.Ñ7/Û<9ÿ^eÿaqÿPi÷<[ÿAiÿ4_ÿ5aû1YôKlÛYq•>NZ%-C),:20.1*08-8A0=B,<<"A<UH(`[5|†T�¢nª¾�¨¾�Ž¥wo‹[aK^F`„DcˆBg�Dl’Go”Np•Po“So“So”Ol‘KiŽIgŒGiŒLiŒLi‹OhŠNg†Ke„IdƒHe„Ih†Jh†Hg…Ge…FiŽKi‘Kl”N+1%.4*470894>:7B=:HA;KD<NG=NG=PKEWSR_^ffeugggf†ba‰[YŠRR†NO‡RS“[^£ei³lo¾uwÊz|Ñ�Öƒ…؆‡Ù†‡×ˆ‡Õ‰‰Ó‡ˆË†‡È†…Ë…ƒÎ…€Ð…€Ò‡Ò‡€Ð�…ÒŒ…Ë�…ÁŒ…¸‹…«„�{v�}k{„Vc�PVnUQd^P[eMYcJb^EsXC�OE´WRÐUZÛT[ç\cèagågjèpoìyræqh÷|tósjþxmæ\QÄ;+¾5#Ã<)Á=(Á=(Ã<(Ä;(È;)Ë9*Ï7*×6,Ø7-Ô:.Î:,Ç:)Ä8'Ç;*Ê=,Ï=.Ö@2ÙA4Ö>1Î8*É5'Ê8)Ë<,É:,È:,É;/Ê<.Ê:/È9+È5+Ç5(Ê6,Ï;/Ò>4Ó?3Ö?6Õ>3Î7.É,#Ò+%Ü-*ç02ð39ö1;ú0>ý0Aþ1Bú1A÷1@÷1@ö0=÷/<÷-9ú,9ú*7ü0<ù/;ô0:ñ4;í6;æ69ß55Ú53Ü86Ý97ß:6â:7å97ç77ê65ì44ò59ò59ð67í76æ74à72Ú6-Ö6*Õ7+Ô8)Ô8)Ô9'Õ8'Õ8%×7'Ô9%Ð9$Î<%Ï=&Ï?'Î@(Î@*Ð?.Ó<1Ñ61ëHIÿfrÿOdÿHfû7[ÿAmý/]ÿ7aò4XäQk¼Ufm19E(,E697325426926=-5:#38CCja:…�TŸ®w©Á‡µÍ™¯Ç•’®}v”be†Q`‚F`…BcˆBi�Fo•Jr˜Or—Rq•Up”To”Ol‘Kh�Hf‹Fh‹KiŒLjŒOi‹Og†Kf…Je„If…Ii‡IjˆJi‡Ig‡Fj�Jk“Mn–P(0#+1'14-561764<87@<9C>8IE<HE<KHCRPQ\Zecbtbb|``‚_^ˆZXŠTTŠTT�\] gi³psÄvxÍ{Ø~‚Ûƒ„Þ„†Ý„…؃„Ö…„Ô†…у†Ë‚…ȃƒË‚�̓€Ñ…€Ôˆ€Õˆ€Ó�…Ò‹„È‹ƒ¾‹ƒ²‰‚¤�|“wp€wfp~U[|PQnUNf_O^fNZdIaaGq]E‰TF­]TÈZ[ÔVZâ[aæ]déaeîllõzsírj÷xoÞYPå[PÒD8Ã5'Ç:)Å9(Ä;(Å<)È=(Ê;*Î;+Ñ9,Ô6+Ó0'Õ4*Ò8,Í;,Ç;*Â;(À;(Ã<)È;*ÕC4áK=ÞH:Ó=/Ë5'Ê6(Í;,È9)È9)È9+È9)Ê8+Ë9*Ì8,Ë7)Í6+Ò<.×?4Ø@3ÛA7ÛA5×:1Ò1)Ü0.å14ï5:õ6>ù4>ú2?ú0@ù0@ò.<ï/<í/9í07ï-5ï,2ò+2ò+2ö3;ó4;ò5;í6:å57á55Þ65Ü75Ü86Ý97ß:8ã99æ87é77ê67ì46ô5:ô5:ñ7:î79ç85à72Ú6-Ô6+Ó7+Ñ8*Ð8*Ð9(Ð9(Ð9&Ñ8(Ð9&Ë9"È:"È=&Ç?'È@(È@*Ì=-Ï;1×96üU\ÿ^qÿGcÿ?aÿ;bÿ7gþ3bû8`íFdÅNa‰@IT15A55@:<:46?56>6389+6<"9CSZ.‚�Q¢§q±ÅŠ¯ËŽ°Ë”¦Ã�’®}{™eh‰R_�E^ƒ@c‰@i�Do•Js™Ps˜Sq•Uo“So’Nl�Jh‹Gf‰EgŠJiŒLi‹Ni‹NfˆKd†Id†Ie‡Ji‰JjŠIi‰Hf‰Ej’Ll–No™Q&.!)/%-2+13.333764:97=<7GD=DD<HGCPPRZZd_ap``z\\~`_‰][�YX‘]]›gg­rtÁz|Ñ€Ü�„ႅ䅆ㄅჃÛ�‚Õƒ‚Òƒ‚΄‡Ìƒ†É„„̃‚Î…‚Ó‡‚ÖŠ‚׌ƒÔ�…ÐŒƒÄ‹‚·Š‚«†€œxˆsnuqbewXVtSLjVKg`NbfO_eIcaHp^F{R@ž\N·ZSÆTS×VZâW\êY`òce÷qnðme÷qhÑD;ÔA7Å1%È4&ÔC2Ç8'Ç:(É;'Í<)Ñ;,Ó:,Ø7-Ù6-Ï,#Ï1&Î6)Ë<,Ç@-Â?+¾;'¼7$À4#ÑB2ãOAåOAØB4Ë7)Ë7)Ð>/Ê;+É:*Ç8(Æ7'È6'Ê6(Î8*Ð8+Ó;.Õ;/Õ9-Ò6*Õ7,Ú<1Ý<2Ü71è88ñ8=ø;Bû9Bú4Aö0=ò,;ï,:í/;ê19ê2:ë48í49ñ48ô36ó57ð37ï6;í9<é69â45Þ44ß76à;9Ý86Þ97á98ã99æ87é77ê67ì46ô5:ô5:ñ7:ì89æ95ß82Ø7/Ó7+Ð8+Í9+Í9+Ì;*Í:*Í:(Í:*Ì;(Å9"Ã;#Ã=$Â>'Ã?(Æ?,Ê<.Ð92ß<?ÿ\fÿQhÿIfÿ9]ÿ=hù.]ý<hé;\å\p¦LV^-0G85BC><89?48F37C41=:)>D(GV-bs?�›a­½±ËŒ«É‹¦Ã‹�º„Ž¬z{™ef‡PZ|@^�?a‡>i�Do•Jr˜Or—Rp”Tn’Ro’Nl�JgŠFeˆDf‰Ih‹Ki‹Ni‹NfˆKe‡Jd†IeˆHgŠHh‹GjŠGg‹Ek“Lm—OpšR)/#)/%(-&+-(///3317759:4==5?@8DE@KKKQQYVXe[[s^]}YX‚[Y‹^]–ee£nn¶wxÈ�Ø„…⌌ðŠŠì†‡ä‚‚Ü€€Ö‚€Óƒ‚΂‚ȃ„ǃƒË„„ΆƒÐˆƒÓ‹„ÒŒ„ÏŠ€ÅŠ€¼‰€¯‡}Ÿ€x�wozkegf]Xm[Qm[Mi\Lf^Kd_IeaHh`Io_FoP;…UA¦^PÀbZÑYXÝQTîS[ø\`ùggæYRÔD<Î70Ñ7/×:1Ö<0Ô;-Ï9*Í:(Ð:)Ñ:)Ô8+Ø8,Ü5,Ü5-Ð/%Í3)Ì8,Ê>/Å@/ÂA.¾=*½:(Â9)¾0"éWJàL@Ã/#Ñ=1Î</Ç8(Ê;+É:)È9(Ç8'È7&É6&Ë5&Ì4&Ð7)×;.Ü>2Ü<0Ù7,Ú8-â>4ìC<õBEò9?ñ3=ö4?ö2@ñ-;í-:í2=ì7@æ5;ä5:å78ë8;ð9;ò89ï77ò;=î<<ê::æ87á77ß76Ü75Ü75Þ97á98ã99æ::é9;ì9<í9<ï8<ô7;ó6:ð8:í9:ç;9à;5×90Ï7,Ê8+È9+Ç9+Æ:)Ç:)È;)È;*Ç<)Á9#Ä='¾:#¶4¼9%ÉD3Ï?6Ê0.ÿ^eÿVgÿIbÿ<\ø1X÷0Yû6cõBiçYo¬FQo33N3,B?6>C<A89@-1L/3K2.B9(DI)Zm?~š_ž¸w¨Å‚«É‹Ÿ¾‚˜¶€”²~‚ nf„RYwCY{?^�?c‡Ai�Fn”Im•Ll”Nk’Ql�Pj�IfŠDc†Bc†Bf‰Ih‹Kg‰LfˆKhŒNhŒNg‹Kf‹HeŠGgŒGk�Im’Lq›Sr�UtŸW*0$)/#).',.)//-220561782;<4>?7CD>IIGOPTSUaXYk[[u\Z^]‰db”kj¤utº~·‡ß‹Œé��ð‰‰ë…„ ؀~Ò€Ï��Ë…†É†‡È‡ˆË‰ˆÎ‹ˆÏŒ†Î�†ÌŽ…ÈŠ€¼‰±†|¡‚x‘{s€tknlb`e]Re_Oe^Lf^Kf^Kh]Ij^Hk_Io_Fw_G[C’YF­ZLÇVPÞRSõQXÿX_ðUSãKFÙ>9Õ60Ù6/Ü90Û9.Ö9*Ó7(Ñ8(Ò9)Ô8)×7+Ü5,ß4,Ý4-Î0'É5)È9+Ç>.ÄA/Á@-¿<*¾9(Ä8)Ë<.äREÜH<È4(Ì:-È:,Â4&Ê=,Ë<+Ê;*Ê;*Ë:)Ì9)Î8)Ð7)Ó7*Ö8,Ù7,Ø4*Õ1'Ø1(Þ7.ç<5øDGñ8>í/9ï/:ð0=î.;ë1<ë6?å6=Ü36Ø24ß:8îDDùIIøDCó?>é:7ç98å97â96á77à87á98á98à87â88å99ç9:é9;ì9<í9<ï8<ó6:ò59î68ê88ä;8Ü:5Ó9/Ë7+Æ8*Ã:*Â:*À;*Á<+Á<)Á<+Á<)½9$Â@*»:%³5º;(Á;/Ê94Ø<?ÿ]iÿPdÿA[ù8Wö6[õ;`ô>dãIe«?Lƒ?>`;3G;/:>08>2F<:R>=N0.S8/VK7ciEx�[�°p�Â}¡Æ�š¹}š¹€š¶ƒ�©xrŽ^YuETp=\{B_@b†@i�Em“Hl”Kk“Mj‘Pi�OgŒGe‰Cc†Bd‡EgŠJiŒLi‹NhŠMi�Oi�MiŽKh�HgŒGh�Gl’Im•Lp›SqœTtŸW-1#-1#,/&,.)//-11/34/56.:<1<>1@B7FGAKMLQRWVVbXWi\Yt^]fdŒpn zx·…ƒÎ�‹ß��ê�ŽïŽŒíŠ‰ç‡„ß�Ó|Ë}È€ņ…LjˆÆŠŠÊŒŠÉŽ‰É�‡ÅŒ„¿‹‚·‰€­†|¡€v�{p€ujpoeci_Vc]M_aL_aKc^Jg\Jl[In\Ho]Ir^FwaIxX?‡S=£VDÇXOãUSøOTþMSâ>=Þ<7Ý84Þ71á6/â7/Ý7+Ú7(Ô7&Ó8&Ô9'Õ8'Ø6)Ý5*à3,Ý4-Ë3(Ä6*Â8+Á<-Â>/Â>/Á<-À8*Ã5)ÜI?ÚF<ÕA7Ï<2Ä6*Ä:-¾6&Æ:)Ç:)È;*É:*Ë9*Ë7)Ì6(Î4(Õ9-Õ7,×4+×3*Ú3+ß6/å:3é<8ð<=í49ë07í29î3<í4<ë7@ê=CÝ7;áAAìNMøYUýYWúQNïB>å84â62à72à74à74á85â96ä::ä::á77ä88æ8:é9<ë8<ì9=ë8<ì8;ï58ï58ì57ç77â:7Ù:4Ï8/Ç7,Â8+¿:+½;+»<+»<+»<)»<+»<)¸9&½@*³9$±6$¶;+µ0'Ã40ìPTÿ[iÿJ`õ:Uð7Vñ@]ïHfçHdÍNa‚89e?6R@2DA09=,57)?7,L:0F.$R=,g]B�ˆ_“ªt—»{–Áz–¿{”³zœ¸‡š´…€šm]vLHb5Nh8\wB_}?cƒ@hŒDl’Gk“Ji‘Jg�Lg�Nd‰Dd‡Cc†DeˆFh‹KkŽNk�Ok�Ok�Ok�Oi�Mh�Jg�Hi‘Hl”Ko™Mu¡Vt¢Wv¤Y24&04&12*01+12-23.34.46+9;-;>-?A3DF9JKENPORQWTT^XVd\Zoda~nl”zw¬„‚ÃŒ‰Ö‘�㔑ê“�뎋≅قÌ|Ã|¿€~¿„‚Á†…Á‰ˆÂ�ŠÃŒ‡¿‹„¸‡®…}¤…{�€vŽyn~rgmmaak^Vi[Pc\J^aL]aJc^Jh[JmZKqZJtZIv[FsWAxR;ŠQ=«ZGÍ[PãTPðFIñ>AÛ2/Ü5/á51ã60ã5.á5+à6)Û8)Ö6&Ó8&Ò9'Ô9'Ø6)Ü6*à3,Ú6-É7*À8*¼7(½9*¿=-Â>/Ã;-Ä8+Æ3)éUKÔ=4Ì8.ÑA6¾4'À<-À<-À8(Ä8'Å9(Ç:)È9)Ë7)Ì6(Í3'Ò6*Ô3)Õ2)Ü5-ã:3é>7ì?9ì=:è45ì59ï6<î5;ê2:æ39ä7=â<@óUVübbÿmjÿidóVQäA<ß63Þ50ß61Þ71Þ63ß74á85â96å99æ::å78æ89è8;é9<ë8<ë8<ë8<ì7<ï6;î68ê67å97ß<7Ö<4Ì9/Å9,¾9*º;*·<*µ<)µ<+´=)´<+³<(³:'·@,­8&­8'³;-«+"Ã54ý`iÿTdúG]ð<Uë=XèH`áOdÒO_´SZvGAXH9KD2FE1BE0=?*:8#;3?3PG*nkHŠ“f—«v“³tŽ·s‘ºxžº‡¡¸Œ�¦l‚[K_:BW0Mb7Yq?`|Adƒ@iŠCl�Fi‘Hg‘Ie�KeŽJb‡Bc†Bd‡EgŠHj�Ml�Om‘Qn’Rl�Pj‘Nj’Li‘Ji‘Hi“Gm—KpœOv¤Yw¥Zy§\78(68*77-56056167267/68+;>-=@-@C.DG4IK>MNFQPNRRRVUSYX]a_lki�us›€}²ˆ†ÇŒŠÓ•’ᔑâ��Ü‹‡Ó„�Ä�}º|µ|³�~·„�º‡„¹‰…·Š„²†¨�z›€x�€tˆznzrfjm_^j[Ti[PjZKfZJb_Lb_Le^Li\Kn[Lr[Kw\K|[HVB‡S>›RA¶VFÏSIÚG@â88å33Ý1-ß3/â51ã4/á3,ß3)Þ6)Ü9(Ó8$Ð9$Ð;'Ñ:'Õ8)Ù7*Ý5,Ø7-Æ:+»9)µ6%·8'½;+Ã>/Ç;.Ç7,Ð90õ\TÖ=5É5+ÏA5¹4%·:(¼?-À;*Ã:*Æ:+È:,Ê;-Ì:-Ï8-Ñ7-Ð3*Ò1)Ö1+Þ71å<7ê>:é<8æ74é77ï;<ð<?ë6;å28ä5:èBFíMOÿxwÿqnñ`[ÝJCÑ83Ñ2.Ü75ç?<à93à93à72à72â64ã75å76å76ç79ç79ê7;ë8<ë8>ë8>è7=ê7;ï8<î7;ê7:ä::ß=:Ö=7Ë;2Ã;/º8*µ:*³;*¯<)¯<*¬<(¬;)¬;)«<)­>+§9(§9*ª8-¬/)Ë@CÿhrôM^óH[îEZåH[ÙO^ÉT]·TWœXUdM?PM:LG4KF2FD-@C(>E$?I$KV.\g<x�TŒ˜h‘¤m�¨n�²r—¹}Ÿ¶ˆ’§€u†dRcC?P0BS1Oa9Wm>b{Be‚BiŠCk�Ei‘Hf�Hb�HcŽGb‡Bf†CgŠHiŒJkŽNm�Pn’Ro“Sk’Ok’Ok“Mi“Ki“Ik•Io™MržQv¤Yv¦Zw§[?=.=>0==3==5=<7<=7;=2;=/?B/@D-CG.FJ3KL:NOAQQIRRHYWJZZN`_]gerqn‰{y¡ƒ€·‡…Ä�ŒÑ�ŒÓ�ŠÏŠ…ǃº~y¯zw¦yv¥{v¬|y®€|®ƒ}©ƒ}£x—|t‹zr�vksrffn`]j\Sk[NlZLl[Kk[Ki]Mg^Mi]Mj]Mo^Ns^Mz]M‚[J“`O›VG§N@¸J=Å@7Ï60Ø1+Þ1+á51â70â5/â4-à2)Þ4'Ü6&Ú;(Ñ9$Í;$Í<'Î;'Ñ:)Ö9*Ù7,Ó9-Â=,µ:(°5#³6$»9)Ä<.É;/Ë7-Ù?7ø[TãIAÎ:0Ì@3¸6&¬3 µ<)½:(¿7'À7'Ã7(Å7)É7*Î7,Ñ7-Ù<3Ü;3ß:4å<7ê>:ë>:ç85ã41è96ë;;ì::è8:ì>@øORÿbeÿppùheãUQÊ=6À1+É40Ö;7ß=;à;9â;5â;5â94á83ã75ä65å55æ66é69ê7:ê7;ë8<ë8>ë8>è7=ê7=ï8=ì8;è8:ã;:Ý>:Ó>7Ê=3À</¶8)²:)®;)«<)©<(¨;'¦;'¥:(¥<) 9( ;) 9*¢6*µ>8ÛTXûepîK\ïI]éJ^ÛM[ÉQZ´VV ZRŠ`RPI7IN:NI5H@+;578>O%Lf6lŠTw”\„›e‰šd�›f‘¢l•¬t˜°|ƒ•mn}\Q`C<J09F,AP1O_;Wj=d{EfƒCkŠDl�Fi‘He‘Fa�FcŽGcˆCi‰Fj�Kl�Ml�Ol�Om‘Qn’Rm”Qm•Om–Pl–Ll–Ln˜LržOs¢R{¬]{«_zª^EB1DB3DB6CC;CC;CC;BB6BC3DF0EH-GJ-JM2ON9RQ?TREUSD\ZC\\D\]Oaaaihxrq�{z¦�~µ†‚¿ˆƒÃˆ„Á„�º{®ys¡to—ql•qmŸrn¡vpžwršwr’to…skzrinmb`l_Wj\Qk[Lm\Lo\Mp]Lq[Mq[No[Pm]Nl_Om`Os`O{`O†ZMŸ_S¥QG®A:º;4Æ6.Î1*Ö2)Ý5,â70á6/á4.á4-à4*Ý5(Ù8&Ô9%Ï:"Ê="Ê>%Ê>'Í<)Ò;*Õ9,Ï;-¿@-±<(¬5!®3!º8(Ä<.Ê:/Î5-Ø93ðNIñTMÙB9ÌB5¾?.¦1´=)¿=-Á9+Á7*Ã7*Å7+É9.Ï;1Õ<4âE>ãA<ä?;ç>;é=;é;:è88ç85è96æ95æ66é;<ôJKÿ\]ÿdeøbaÌ=9Á82½4.Ä92ÕA=ÞE@ß=;Ù42â;5â;3â94â94ä84å84ç77ê88ê69ë7:ì7<ë8<ë8>è7=è7?é6<ë6;ê7:æ89à;9Û>9Ñ>6Æ<2¾</³8)®;)ª;(¦;'¤;(£:'¡:'Ÿ:&¡>+–7%œ=+š9)š2'ÃPKíkmî\fìP^éJ\ßHYÎLV¹RS¤[R’cQ€jUJM8DL7JB/H9$B7FH#Up=g“V~²p‚³qƒ¥h}’Yƒ�YŽ–cŠ–d|Œ_XfCHU9:F.7C+:D,=J0JX7Wh>e|FiƒDl‹Em‘Gi“Gd’Ga‘Ga�Gf‹FlŒIm�Nn‘Ol�OkŽNk�Om‘Qp—Tq™Sp™SpšPo›Pq�Pv¢Sw¦V|­^z­^y¬]IC3JD6IE9JF;IF=IG;HF9HG5IH3JJ2LL2NN4RP;TQ>XRDYTA]Y>[Z>\YH_\Udahnk~xs“}x �{§ƒ}«…­ƒ|¦}w�vo�oi…lf‚ieˆjf‹mhˆmi‚mh|lfrkbgj`^g]Th\Nh[Kk\Io^Lq_Kq^Mq^Mp]Op]Op]Os]Ow]N~]N‡\L’XL¢VI¤F<«9/º7/Ê7/Ô7.Ú6-Þ6-ß6/à5.ß4,à5-ß7,Þ8*Ø6'Ñ6$Ï:$Ì<$Ë='Ë='Î;)Ñ;*Ó:,Î<-¾A/¯<'©4"¬3 ¶9'Á=.É;/Ì8.Í3+àA;ø[TàI@ÐD7ÊF7«2¹>,ÊF7ÌB5Ê@5Ë?2Î@4ÔA7ÜE<ãF?â@;â=9ã:7ã75ã54ä65æ:8å<9ä?9à=8Ü:5ß@<êLIêQLÓ?;¸)#¸-&¿6.É>7Ñ@;Ö>9Ú;8ã;;ç==â96â96â96ä86å95ç:6é99ê::ê88ê88ë8;ë8;é9<è8;ç6<ç6<é6:æ68â89Þ:8Ö=8Í>6Â<1¹;-®8*ª;*¦;)¢;( ;'Ÿ<)ž;(š;'˜?-Œ5"™=.˜8*—/&Í\X÷y|ßX_æXdÛQ^ÊKT¹LO©VRž`U‘hV€nVPW8?J*<;C=UQ+lvD„¢f”Á~‡¼v†¸s|žbn‚MvR‚‡^u}V[b@;D)4<%5=(=D2<E29C+CP2Vf?g{HiƒFmŒHm‘Gj“Gf’Gb�GcŽGh�Jl�Mo’Po’Pl�MiŽKj�Jj’Ls›Us�Us�SržQržQs¢Ry¥V{ªZv¦Zu¥[s£YRK;RK;RJ=RJ=RJ=RJ=RJ=RK;UN<VO=WP>XQ?YR@[TD\UE\UEc\Ib[IaYLbYRe][kaiqftuj{zn„|pˆs‹s‰|p„uj{mbrf^kc`k`_g_\c_Z^^ZY^YUaZRbZOe[Og[Ki\Kk\Il]Hm]Fm\Hj^HibOqfTqZJuOB‹WJœ\P¤ZM¬ZL¥M?¦E5§<,±7(Â8-Ñ80Ü41Ý1-ß6/Þ7.Þ7.Þ7.Þ7.Û7+Ù7*×7)Õ8)Ô8)Ò9)Ò9)Ò9+Ò9+Ô8+Í;,½;+¶>-®9(©1 «2!¸:+À</Ã9.É6,Î5-Ô:2ÛB:ÝK>ÔH9Â=,·2#ÝPFÙKAÔE=ÕE<ÚG?ÞG@àC>á=;à74æ87í;;ì::ç77ã75à85Ü=7ÙF>ÏB9Ä:0¼3)·1&º6*¿;/Ä@4Ã:0Ê=4Ó@8Û@;â=;ç;;ì8;í9<ç7:å8:å8:å99å97å97å95å95á51á51ä65å76æ89ç9:ç9;ç9;ë;>æ9;ß9;Ü=:ÖA=ÍB;ÁA6·?1©9+£<+ =*›<(™:&–9'•8&�9&�A0‹<-‡/#’0'¼LJämoçloÍVZ½LNµNOªPO¢UO›XP—]R•aTŠjSWZ/:M7FTa)~ŽP›°o¡¼y�¼y’µs~�ah€N\mCYdBT]BHP9?D0:=,@C2=A2;?1<C3:C0AN2Rd>]sBgƒHp�Ko“Kl’Gj’Ii“Kg�Ll�Pm‘So“Sn“Pm’Mk“Jm–Jm˜IqœMq�NrŸNr¡Pt£Rv§Vz©Y{©^y¦cm™Zc�PWN?WN?WN?WN?WN?WN?WN?WN?XO@YPAZQB[RC\SD]TE^UF_VGd[Jd[LcYMcYOf[Uj_]nbdpdhreltgpvirvirsfmm`gg[_bXY^[VZZRZWPXULXTIXTH\VH^WGcZKcZIf[Ig\Hi]Gj^Fk_IiaJcbMngTv\M�SF›WN«VO°RJ¸RF¯G:¬B2¬=,±9)¾8,Ë80Õ62Ù40Ý6.Ü8.Ü8.Ü8.Ü8.Ú8-Ù7,×7+×7+Õ7+Õ7+Õ7+Ô8+Ô8+Ô8+Î:,¾<.¶>.®9(¦1 §2 ²:)¹>.¾<.Æ:-È5+Ë4+Ò;2ÜE:áOBßQCÝOCÛF?Ø@;Ó;6Ò:5Ö=7Ú=8Ü86Þ44è8:ë78î79ë78ç77å97â=9ÛB<ÊA7À@5¸:,±6'¯5&²8)¸=-¿@1ÅA4Í@6×@7ß=8ã:7é77ï58ñ7<é6<ç6<ç7:æ89æ87å95å95ã:5â92ã:3ã:5ä;6ä;8ä;8æ::å8:ä:=ã:=ß<=Ù?=ÐA;Å@7·=2­<.¥;+ =*œ=+—<)–;(’:&‘8&Œ9'€9'€8)‰5*�<5¿QP×eeÑ^a¸NN¬NL¤SO�XQ˜\R—^S˜_T˜_VŒfQ`_/Sg*^r3z�O™¯n¥¾|™²pˆ bj‚HbvCTe;KY8FP7@I69?158-:</AB4=?4:=2>B4:C0@M3Rd>^tEh„Iq�Mp”Ln’Hm“Jk“Lj‘Nm‘Sn’To“So”Ql”Mk”Hl—Hm˜Hp�Jp�Jp Ls¢Qu¦Tx©Xx©Xx¨^m—Xb‹QY‚H]TE]TE]TE]TE]TE]TE]TE]TE\SD\SD]TE^UF_VG`WHaXIaXGe]Je]Jd[Je[Oe[Qf[Uh]Yi^\j^^j^`k_ak__i^\f[YbWU_VQZUOWTKUQHRNCQMBSOCWQCXRB^WG_XFaYFc[Fg\Hi^Hk`LicMbaMngUy_PˆXN¢[U±VQ²IE¸E@¹@8¸>3·;/¸8+¼8+Ã9.É;/Ñ:/Ù8.Û7-Ü8.Û9.Ú8-Ú8-Ù7,×7+×7+Õ7+Õ7+Ô8+Ô8+Ô8+Ò9+Í;,À>0·?/­:(£2 £2 ª9'³>,º?/Ä?0Ä8+Å2(È4*Ð<0ØF9ÝN@âNDÙ:6Ø43Ô20Ö42Û97á=<æ<=é;<í9<î5:ë27ê37è58å99á<:Ù@:¾8-µ:+±6'®4%­5%°8(µ<+»=.ÈD7ÐC9ÖB8Ý>8â96è66î66ï79ë6;ê7;ê7:è88è88ç:6æ:6æ;4ã:3ä;4ä;6ã<6ã;8â:9á99Þ88Ý9:Ü<<Ú@>ÓB=È?9º;2­7+¥7(¡;,�<+™<+”;)“:(�9&�8%Š9&z6#y6%ˆ9, F=¹SOÀWT¸PO«IF¢MHœSLšZQ—^S—aW—aW˜_V‹ePsrFz�V“¦n¤¹€«À‡ž³{|�[[o<@S%AS+BQ0?L2:C06<.69058/9;.@A3=?4:=2=A3:C0@M3Qc=`vGi…Js’Or–Np”Jo•Lm•Nl“Pp”Tq•Uo–Sn–Pl”Ml•Il—Hn™Ip�Lp�LpŸNu¤Sx©Xyª[u¦Wq X_‰JVGOx@aXIaXIaXIaXIaXIaXIaXIaXI_VG`WH`WHaXIbYJcZKd[Ld[Je]Hf^If^Kg^Mg^Of\Pf\Rf\Sh]Wh]Wh]Wh]Wg]Tf\RdZPc[P]YNZVKVRFRNBPL@PM>TN@UO?XRBYTA[VC]XDaZGd]Jf_LfaMdcQleUv\OŠ[Q©b\¸ZX¶FD¹<:Ã<9Æ;6Ä92Â8-¾8,½;+¾?.Ç?/Ó9-Ù7,Ú8-Ù9-Ú8-Ù9-Ø8,Õ7+Õ7+Ô8+Ô8+Ô8+Ô8+Ô8+Ò9+Í;,Á?1·?/­:(¢3 ž3¢7#¬=)µ@.¼?-¿:+Ã7*Å5*Ë7+Î:.Î</Ó:2Þ66à24Û12Ý34à88æ<=ê=?ì<?î7<ì38é06æ25å58á77Ú65Î61·5(¯7'®6&¬7&¬7&®9(²:)µ:*ÃA3ÊB6ÔA7Ü?8ã;8æ87ë76î79í6:ë7:ë78ê86ê86è94ç:4æ;4â92á:2ß:4à;7Þ:8Ý;9Ü:8Ù99Ö<:ÖA=ÒC?ÊA;¼<3¯6+¤6'ž7(œ;*˜;)”;)’;(�:&Ž9%Œ7#‡9%€=*w6$�5'™E:ªPH¬MG©LG¦QL�PJ˜UL”ZO”^R•aV–bW—aW�gT‰…_™©z­½Ž¦¸ˆ�¡ts†YN`8/A->4C&:F.=F3;A3:=2:=4<=5::.?@2<>39<1<@29B/?L2Qc=awHj†Kt“Ps—Or–Lq—No—Pn–Pr—Ts˜Up˜Ro—Pm•Lm–Jn™Jp›KqžMržOt£Sy¨X|«]w§[m�Se’M[„HS|DNw?d[Ld[Ld[Ld[Ld[Ld[Ld[Ld[LcZKcZKd[Ld[Le\Mf]Nf]Ng^Mf^Ig_Hh`IiaLiaNi`Oh_Ph_Pj`Tj`Ti`Qh_Ph_Nh_Nh`MhaOd^N`ZL[UGVPBPM>NK:NK:NK:QN=RO>TQ@VS@XWC[ZF]\H^]Kb`Qf^QmWJ†[R­jd¾c`ÁMMÈBCÐ;=Ô89Ó84Ì70Á9-º=+µB-¼B+Î;+×7)Ø8*×:+Ù9+Ö9*Ö9*Ô8)Ô8)Ô8)Ô8)Ò9)Ò9)Ò9)Ò9)Í;,Á?1¶>0«:(¡6"š4›7 £=&¯B-³;*º;,Â:,Ç9-Ë8.Î:0Ñ=3Ú;5é9<ì4<æ39ã28â38â59â38â17é6<é49ç4:ç7:ã9<Û76Ð21À/*´6*­9*­:(«:(«:(«:(¬9&¯7&¸:+Á;/Î>3Ø?7á>9ä;8è96ì87î68î66î66í74ê84è:3ç:3ä<3á:2ß;2ß<5Ü=7Û=:Ù><×=;Ô>=Î@<ÊC=ÅA<º>6­7-£5(ž7(š:*–;)“<)‘;*�:(Š9&‰8%ˆ7$ƒ8%ƒ@-u4"{3%“G:¥RJ¡NFžNG¡WN™VM“YM�ZNŽ]O�]R”`U—`Y�hW�ˆh˜¤|�©ƒ€Žj]kHDS42@&$2-:&0;*5=.9?3<?4=>6;<4::099-?@2;=28;0;?18A.>K1Pb<bxIk‡Lu”Qt˜Ps—Ms™Pq™Rp˜RršSršSršSp˜Om—Mm—Kp›LsžOt Qv¢Uz¦Y{©^z¨_qŸVd’JZ‡D]†JW~GRyBe\Me\Me\Me\Me\Me\Me\Me\Me\Mf]Nf]Nf]Ng^Og^Oh_Ph_Nh`Kh`IiaLjbMjbOjbOjaPjaPjaPiaNh`Mh`Kh`Kh`IiaJjbKf_Mc^K^XHXRBSM=MJ9KH7IH6LK9LK9LM;NO=PQ?QTCSVETUE[YL^VIcQEzXN¡ha¶eaÄVU×RSÙ@Cß:>Ü87Õ83Ç;.¼?-±C*µA(Ê;*Õ8)Õ8)Õ9*×:+Õ9*Õ9*Õ9*Ô8)Ò9)Ò9)Ò9)Ò9)Ò9)Ò9)Í;,Á?1¶>0«<)£:%™6•5›;#§A*²B.¹@/¿;,Ã7*Å2(Ë4+Ò;2ß=:ê5:î3<ë6=é8>ç:>ä;>ã:=â9<ß58á7:â9>á=>Û=>Ñ96Ã2/¶0'¯9+«=,«=,©>*ª=)©<(©:'«8%¯6%¹7)Å;0Ð=3Ù<5ß<7ä;8é:7î87ï75ï75î85ë:4é;4ç<4ä=4â>5à?7ÞA:ÛB<ÙA>ÕA?ÒA>ÍB?Á@:»@8³=3ª8-¡5(›5'™9)•<*‘;*�<)Š<(‰;'†9'„7%ƒ6$€7&}:)t3!~9*—OA£YNœRG•OE™WK•YN�[MŽZMŒ[MŽ\Q“_T™`Y‘gY~wZyƒ`r|ZVaC;E,/;%0;*0<.3=25=27=3<?6@A9?A6<<277+89+>?1:<17:/;?17@->K1Oa;bxIk‡Lu”Qu™Qt˜Ns™PršSq™Rs›RsœPr›Op™Mn˜LpšNsžOv Ty¥X{§\|¨]z¥]qŸWg”O]ŠEX‚BaˆOY€ISzCg^Og^Og^Og^Og^Og^Og^Og^Og^Og^Og^Oh_Ph_Ph_Pi`Qi`OjaPjbOjbOjbOi`Oi`OiaNiaNiaNh`Kh`Kh`Ih`IhaGhaGhaGg`Me^Kb[I\WDWR?PM:MJ9IH6IG8GH8GJ9GK:HL;IM>JN?KM?PPDXRF[OCiRD„ZN–VL°SLÔZYÛIJâ?Bá99Û75Ï;1Ã@.´B*³<$É:)Ó7(Ô8)Ó:*Õ9*Ó:*Ó:*Ó:*Ò9)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Í;,Á=0µ=/¬=,£<)˜8"�4•9 ¡A)­C-¶A/¿=/À6+Â/%Ç0'Ð6.Þ:8è59î6>ì;Cê?EæAEàBCÜ@AÚ@@Í12Ð66Ó:<Ñ=;É;9¿82µ4.­4)ª<-¥>+¦?,¥@,¥>+¥?)§<(©:'ª7%²7'¼8+Å8.Ì70Ô94Þ<9æ>;ë97ì95ì95ë:4é;4è;4å=4â>4àB7ÝC9ÚC<ÕB;ÏA=Ê?<Ç=;¿>9±<3©:/¢8+œ6(˜7'”8)‘:)�:(‹<+‰<*‡<)ƒ:'‚9(€7&6%}6$x3#x5%†C3˜UE�WK”PC�PD•ZL“YM�ZN�\Q�\Q’[T•\U™\WŽbUkbEWa>JS4=G,6?*3>-4>35?66=69@9=B;CF=EG<DE7@>/;9*78*=>0:<17:/:>07@-=J0N`:awHk‡Lu”Qu™Qt˜NtšQršSq™Rq›Or�NqœMp›LošKr�Nu¡Ty¥X«`~©az¥^qœWf‘L_‰G]‡E^ˆIb‰PY€KSzEh_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Ph_Pi`Qi`Qi`QmdUlcTj`Ti`Qh_Pg^Oh_Nh`MiaNiaLiaLiaLiaJh`IhaGh`IgaKg`Me^Kb[H\WDUR?QN=ML:EF6EH7DH7DH7CJ:CJ:CI;EI;IK>SQDSPAYM=eK:tD6–I?ÂYSÙQQàDEà::Ý86Ö<4Ê>/º=)µ8"É:)Ð9(Ñ:)Ò;*Ò;*Ò;*Ò;*Ò;*Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Ñ:)Í;,Á;0¶:.­<,§>+™9#�3“7Ÿ?'¢9$¯<)À>0Ç=2Ì8.Ð7/Ù<5ã?=ì=Bî?DêDHâDEØ@?Ê;7À50»1.¹/,¼51¾:6»<6³:2ª8.¤8,¢;,¡=-Ÿ>+ ?,¡@-¡@-¡A+¤?+¦=*©;*­:(²7(¹5)¿5+É90Ö>9àC>å<7è;5è;5ç<4å<5ä=4á>5Þ@5ÜD9ØD:ÒC;ËB:Ä?:½<7¸85°93¢7-™8(•5'“6'‘:)�<,‹:)…8&ˆ=*…<+ƒ<*€<)~9)}9&{6&{6&z2$€8*ŒG8–SC“QC‹L=�RE™_S‘WL’YN”[R–]T—^W—\V—ZU�_RaX;EM(6@8A&>G2=F53=2-7.1817>6CF?GJ?HI;BC1?>*;:&78(=>09;069.:>06?,=J0N`:awHj†Ku”Qt˜Pt˜NtšQršSq™RqœMqœLqœLp›Kp›Ls�Qw£V{§\‚­e€«dx£^j•P]‡EYƒA^ˆHe�PcŠSZ�LSzEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi_Sj`Wj`Wj_Yj`Wj`Vj`Vj`TjaRjaRjaPjaPjbOjbOjbOjbMjbMg_Jg_Jf^Ie_Ib]IZXCSP=ML:HI9EH7AE6@D5?E7?G8?G:?E9GK=IK=BF5KJ6\Q;aE0ƒJ9Àj]Üg`ãUSäFEåA@àA=Õ>5É=0Æ=-Æ4%Ë5&Ì6'Í7(Í7(Ï9*Ñ;,Ó=.Ò<-Ò<-Ò<-Ò<-Ò<-Ñ;,Ñ;,Î<-»2(¶7.±;/¨:+š7"”5–7!œ9"¬A-²=,º6)Á4*Í6/×<7âC?êHEèDEéFIÝCEÉ98»61´80®8.©4+«9/©9.¥9/ 8-›9,–:+“<+’<+–=+˜=*™>+š?,›@-�@. A/£@-¢;*¥:(«7(°6)·7,¾8-Ä;3Ì<4Ú=6à=6â?8âA7âA9ÞB6Ú@6×@5ØH=ÒE;ÇA8»;2²5/ª3-¦0,Ÿ2+™9-’;*‘9+�:*‹:)ˆ;+‡;+…<+�:(~9)|9(|9(|9){8'y6&x3$w- ‹A6˜PD“MA�MA’RF“UJŽRG˜\R˜\R‘UM�TL—ZU•XS”WT“gZ]T5IQ*BL)BK,>H/7@+2=-4?14<1HPCZ]RX\MMO:DF.AA'==%:;)>?1;=2:=2<@24=*;H.Pb<cyJlˆMu”Qs—Or–Lr˜OršSršSr�NqžKrŸNrŸNsŸPv¢U{§\ªb‚­fx£^j“O`‰G^†Ga‰JfŽPi‘Ud‰S[€LTyEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`Qi_Sj_Yj_]j_]j_[j_Yj`Wj`Vj`Tj`TjaRjaRjaRjaRjaRjaPjaPh`Mh`Kg_Jg_Jd]J^YEUR?ON<IJ:FI8BD6@D5?E7?G8>F9=E6@F8>E5>F1HL5MI0U?(P<¼sdËcXà_ZëUTçGGâ@>àA>Ø?9Î:0Í9-Î<-Ï=.Ï=.Ï;-Î:,Î:,Î:,Î:,Ð:,Ð:,Ð:,Ð:,Ð:,Ð:,Í;.É?5ÅB:¼B7®=/ 7$—2–1›2¦5#²7(Ã:0ÑA8ßF@æIDéJFêKHÞCAÛEDÐB@¼;5¯9/©=0¥?1Ÿ=0Ÿ?1œ>2˜?1”>/�=/Š>.‡@.ˆ?,Ž=*‘;*’<+“=,”>-–?,—>,š?-Ÿ@.¡=-¤:,¨8,¬8+±8-¶:0¼9/Ê;3Ð;4Ó?5ÖB8×C9ÖC9ÔD9ÑE8ÊB6ÃA4¹=3°:0©6/¢5. 5/™7.’</Ž=,Š;,ˆ;+†:*„;*ƒ;,�<,~;*|:*{9)y:)x9(w8'v7(w3&€3)�C9˜NC’L@�K?�QD’TI�RG“SJ˜XO–UO˜WQœ]X”WRŽSOŠcT]W5Xc9_jBZfBIT66C)4@*<H4OXGYbQ_fTW\HJN7BE*??#<<"9:(=>0:<1:=2<@25>+;H.Pb<dzKlˆMu”Qt˜Pr–Ls™PršSq›QsŸPrŸLp�Lq�Nt Sx¤Y|§_ªc{¦aošUcŒJ]†Da‰Je�Ne�Nc‹Od‰S[€LTyEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QjaRjaRjaRjaRjaRjaRjaRj`Vk`\k__k__k`^k`\k`ZkaXkaWkaWkaUkaUkaUkaUkaUkaWkaUlaOk`LiaNiaNf_Ma\IYVESR@NL=IJ:DF8AE6AE6?F6>D6<D5=F57@-@H1IK3F?%W?'‡ZE·rb¶VJÙbZðebìTSæFHëGHæCDØ:7Ë7-É:,É:,É:,Ë9,Ê8+É7*È6)Ê8+Ë7+Ë7+Ì8,Ì8,Ì8,Ì8,Ë8.Â3-¾5/¶6-­3(¡0"ž. ¡2!§3$¼@4ÇC7ÖG?áLFèMIéJGåFBßD@×CAÏDAÁ@:±;1¤</¡A1œE4˜E3“D3‘D4�B2ŠA0…@0‚A/~A.@.†:*‹9+‹:)Š;*‹<+�<+Ž=,�<*–?.—>.›;+�9* 8+¥9-©:/®9/¸90¾90Á;0Ä>3Å?4ÅA5ÄB4ÁA4¶:.²9.ª8-£7+ž6+š8-™8/“;1‹</‰=-‡;-„;,‚:+€;+€;,<,};+{<-z;,w9*v8)u7(t6)v4&‰<2•G=–MDŽJ?ˆH<�OB’TI’RI’OI™TO™SQ˜WSš]Z“YU�[T�l[�}X~‰^€‹alxRLX46D#=J.LZ@eqYeoW]eMPU>EI0?C(=@#:<$89';<.9;09<1=A36?,<I/Pb<dzKm‰Nv•Rt˜Ps—Ms™PršSpšPu¡Rp Ln�LpŸOw£V{§\}¨a|§bp™Uf�M\„E\„EeŒMj‘Rf�N`†IcˆRZLSxEi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QjaRjaRjaRjaRjaRjaRjaRj`Vk`\k__k`^k`\k`\kaXkaXkaWkaUkaUkaUkaUkaWkaWkaXkaWnbRmbPjbOjbOibPe^L_YIYVERO@NL=GH:DF8CE7BD6@D5<C3<H4:C.FJ3MG1P;&qL9šbQª^P¡A5ÆSLçc_ð^^íSUëJOéDJáACÌ92Ç9-Å7-Å7+È8/Ê:/Ë;2Í:0Ñ>6Ñ>4Ò=6Ò>4Ó>7Ó?5Ó>7Ó>7Ô@<ÑB>ËD>ÆF=ÂF<ÃI>ÇM@ÏQEÙSJÞQJâMGäIEäFCâC@àA=ÛB=ÕFBÉD?º?7«=0 ?/šC2–H4”I6ŽG5ŒE3ˆC3…B2€A0~A/zA0}@.‚:,†8,…9+†:,†:,‡;-‡;+ˆ;+‰:+‹9+�9*‘9+•9,™;/�=1¢<0ª:/®9/°:0°:.°:.°:.¯9+®8*¨7)¥7*ž6)›6*•7+’8-�9/Ž<0ˆ<.…<-„;,‚:+:+~9*~;+};+|:,z;,y;,w9,s8*q6(r7)u5)ŒC:“I>‘KAŠH<‡I<‹OD“TK—TL–QL›TP—QO”SO–\X�_X�h^—~j””p…’fu‚WYh?CR+?O*P`<crQetU\iKOY>CL1?E+?C*=A(;='78(:;-79.9<1=A37@-<I/Oa;bxIlˆMv•Ru™Qt˜NtšQršSpšRsŸPožMožMs¢R|¨]~©ay¤_tŸZcŒJ^‡E[ƒD_‡HgŽOj‘Pf�L`‡Hb‡QY~KRwDi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QjaRjaRjaRjaRjaRjaRjaRj`Tk`Zk`\k`ZkaXkaXkaUkaUkbSkbSkbSkbSkaUkaWkaWkaXkaWocUnbRkbQlcRkbShaQd]M`ZJXRBSP?NK<IG8GE6DE5BC3>B1=I1?G/LG1S>+g=-•ZL­_Sœ>6ž3+¶?9ØXUóigödeëRUæJNéOQÚIDÓF<Í@7Ë>4Î?7ÒD:ÕF>×G>ÚJBÚJAÝJCÝJBÞKDÞKCàKEàKEëSRèTTãSRÝSPØSLØSLÛTNàUPàLHáIFãEBâC@äB@âC@äEBáGEÙIHÊDAº?8¬=2¡?2˜B1’E3�G6ŒE3‰D4‡B3„A1�?1~?0z?1|>/€:.ƒ9.ƒ9.„:/ƒ;-ƒ;-ƒ;-ƒ;-�9*ƒ:+…9+‡9,Š;,�;-�<.“;/›:1ž9/Ÿ9- :.¡9,£9,¢8*¢8*Ÿ8)œ8)˜8(”8)�9*Œ:,‰:-‡;-„:-‚:,�9+€8*}7+}7+|8+{9+{9-y9-x:-v8-q5*o5)q7+u:,ŒH=ŽJ?‰I=„I;†L@ŒRF“WM˜XO˜QMœUQ—SP“VQ’`YˆbWƒg[‡{enrQZi@JY0BS)IZ0WiAdvNj|VXiEL[:?L.:D)<D,>C-<A-:=,89+:;-79.9<1>B47@-;H.L^8_uFi…Jt“Pt˜Pt˜NtšQs›Tq›Sn�MmžLp¡Pw¨Y|ªaz¨`r�Xj”RZ‚C\„EaˆIeŒMgŽMgŽKh�Jg‹Mc†PZ|JRtBi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QkbSkbSkbSkbSkbSkbSkbSkaUlbYlbYlbYlbVlbVlcRlcRldQldQldQlcRlcTlbVlbXlbXlbXocUnbRkbQlcRlcTkbSg`Pd]M^WGYSCTN@MJ;KH9GE6FC4@C0;H.@I.OC-a=-„F;¬YS´NJž,+²:9¹=;ÓSRòpnþvvñehëX^ð^_ãUQÚPFÔG@ÐC:ÑD=ÔG>ÖGAÖG?ÙHCÚJBÜHDÝJCÞJFÞKDàKGàKGâHHáGIÜHHÙGGÖHD×HDØGDÛEDàBAäABæBAèBBçCBäB@ßA>ÛA?ÛIIÊC@¹<6®>3¢@3˜@2‘B3�D3ŠA2ˆ@1‡?1…=1‚<2€<1=1=1�;1�;1ƒ:3�;1�;1€</€<1=/€</€=-�;/€;,‚:,�9*‚8+ƒ7*Œ:/�8.‘9-”:/—;.š=.ž</Ÿ>.™9+—:+“:*�:+Š;,‡;+„;,‚:+‚:,�9+€8*~6(|6*{5){5){7*x6*x8,x:/s7,p4)o5)r:-w=1‹M@‰K>„J<‚M=…QC‹VH’XM—WN“NI™TO—VR–]V�e\|_QgWGa_HEO->N)?O(J\4_qGk}SewMXj@N`8@Q-6E&7D*=F1=D4:>07;-8:,:;-68-9<1?C57@-9F,I[5[qBe�Fr‘Ns—Ot˜NtšQs›Tq›SmœLo Os¤Ux¨\z¨`s [i“Q`ŠHY�B^†GeŒMgŽMgŽKe�GgŒGhŒLa„NXzHQsAi`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QkbSkbSkbSkbSkbSkbSkbSkbSlbXlbVlbVlcRlcRldOldOldMldMldMldOldQlcTlbVlbVlbVocUnbTkbSlcTmdUlcTjaRf_Ob[K^WGXQARL<OI9JG6IC3AC.:G+AG+U@+xG9£PLµII¶8<º39ÈAEÄ?@ÊJIßb`ñqpðnpçbeâZ\ÛQNÕLDÐE@ÍB;ÏD?ÐE>ÐB>ÎA:ÒC?ÒC=ÔC@ÕD?ØDBØD@ÙECÚDCÝADÝADÞDFÝGHßIHàJIãIIåGHçACêADìBEìBEçAAâ@>Û=:Õ=:ÚFFÉ=<¹83±<3¦@4˜?1�>0ŽA1‹=0Š<0‰;1ˆ:0‡81†93…:4„;4�;3€<3�:4€<3<3}=1}=3}=1|<0|=.|<0|=.};-{9)z8*{7*‚:.ƒ9.‡9-‰9.�9.‘;.•<.–=-’9+‘9+Ž;+Š;*‡;+ƒ<*€;+<+€8*€8,~6*}5)|3*z4*z4*y5*u2)v6,w9.s7,p4*o5*s:/x@3‹QC†N?�N=‚Q@…VF‰XIŽUJ’SJ“PJ—SP”TR‘\VŠe\n[LPK7EK1AN0KZ9WgC`rLgyQdyNWm?G\1AV-7H$1@!6B*=F5<D98=67:3:</;<.79.9<1?C57@-7D*FX2Vl=b~CoŽKq•Ms—MtšQtœUrœTožPq¢Qu¦Wv¦Zs¡Yj—RbŒJ[…E^†H`ˆId‹Lf�Lh�Jh�Hf‹Fe‰Ia„NXzHPr@i`Qi`Qi`Qi`Qi`Qi`Qi`Qi`QkbSkbSkbSkbSkbSkbSkbSkbSlbVlcTlcTlcRldQldOldOldMldMldOldOldQlcRlcTlbVlcTnbTmaSmaSmaSmdUlcTjaRg`Pd]M_YIZSCTN>PK8MH5ID1DC.@F*DD([B,ˆTG¯\X«>A¬/5ÎHOÆ?EÀ;>»;<ÄFGØZ[ãefÛY[ÌGHÑJGÐEBÍB?ÐB>ÔFBÔFBÐB>Ê?8ÏD?ÏD?ÒD@ÔEAÖEBÙECÚDEÝDFß@Dß@DÞBEßEGáGGàFFàBCá>?èBDé@Cè>Aç=>ä>>ãA?áC@ßEC×CCÅ98¶50°;4¦?6š>3�=/�?2ˆ</‡:0‡:0‡81‡83ˆ94ˆ94ˆ;5ƒ:3�;3€:2;2~<0~<0|<0|<0x8,y9-z:.z:.{<-{<-{<-z;,};/9/€7.�7,„6*…5*†7*‡8)Š8,ˆ9*ˆ:-…<-„<-€=,~<,~<,}9,}7-|6,z6+z4*x4)y5,x5,r2(u5+v8-r6+n4)n5*t;0xB6ŠUG„PB€O@„SDˆWIŒVJŒRG�NG•TN”UP‹RK„XOcUf[GGK2=J,IX9ZkIhyUcxQZoFSh?Pb:L^6;M'0?,95>)=C5<B8:<79<5;=2;=079.9<1?C58?-8B)EU1Uh;azCnŠMr“Nr–PtšQtœUrœTt Uu£Xv¤[t¢ZošUf�N^ˆH[ƒDdŒNa‰K`‡HcŠIh�JiŽKeŠEa…E`ƒMWyGPr@i`Qi`QjaRjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbQkbQkbQkbQlcRlcRlcRlcRlcTlcTlcTnbTk]PqaRtdWrdWnbVj`TiaThbTgaSebQc]M[VCVQ=TO9NH2GA)KG,I?$bI3‘gW¶wn»lg¸YWºPR¶BE½BEÂDGÈFHÍHKÐJKÒLMÓKMÏEEÐDEÓEDÖFEÕFBÒC?ÏD=ÏF<ÈC:ÉD;ÊE<ÌE?ÒD@×CCÚADÜ@DáBFâBDáCDàDEàDEáEFãEDåEEãCCäB@ä@?ä@?ã?>â@=ãA>ÞC?ÓB?Æ?;¶93ª70Ÿ:0—=2Ž>3†>2ƒ?4�?3�=2‚<4„93‡83‰92‰92…;0„<.„<.ƒ;-�;/�;/€:.€:.~:-~:-~:/~:/};/};/};/};/}90}90~80~8.€7.€7.�8/9-€7.~8,}9.|8-{9-z8,x8,x8,x8.x8.w7-v8-v6,t6+s4+q5+m1'r6,n4)m3(o6+m4)q8-I=…SH…SJ‡RJˆQJŒQI�QJ”QK•RL“TM�TK‰[N‚cQo_HYU:MV7O_;j~YbxQZpIUjCRg@Pa=IY5DR17D&4>#2:#6;'<>0@B5@@6<>358-39-4:.7:/9=/;?.=B,=F)S_;arFo†Rt�Ut”St™Ss�Ss�UužZ~§e~§ep˜YbŠK^†G^†H^„G`†IbˆKdŠMf�NeŒMcŠK`‡H^„G_‚LTvDKm;i`Qi`QjaRjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTnbTrbSqaRrbUrdWrfZogZmeXjdVgcWliZmjYjgThcOc^H[U=RL4PJ0I?&Q?)kP=XI„NDˆE=”EA±WV¸TT»QQÀPOÀNMÁMMÃMKÃKJÐTTÈHIÂ=>Ä=:ÐB@ÓEAÏB;Ç>4ÊD9ÊF:ËF=ÎH?ÔFBØDDÜAEÝAEáBFáCDáCDâDEâDEáEFáFDáFDàEAßD@àB?ßA>àA=àA=àA=ÞC?ÔE?ÈC<º>6­;1 90”:/Š<0ƒ=1€@4~@5~>4€=4ƒ:3†91‰92‰90‡;.…<-…<-„;,ƒ;-‚:,‚:,‚:,9-9-9/9/~:/~:/~:/~:/}90}90}90}90}90}90|90|90|90z:0z:0z:0y9/y9/x:/w9.w8/w8/t8.s7-r6,r6,p6+p6+l2'q7,m4)l3(o6+l3(p7,~H>…SJ…SLˆQLŒQM�QL”PM•QN”SOŠOG…SH€[K{ePsiPilMfqQgxT^tMXnGPf@Ma<M^<L[:HU7CP4<F-8@)6;'7:)<=/?@2>>2;=04:04<15;16<0:</<?.>A,>D*Q[9^mDn�SsŒUr’Ss˜Ss�SsžVxŸ^|¢exžak‘T`†I]ƒF]ƒF\‚E`†Ia‡JdŠMe‹Ne‹NbˆK`†I^„I]LRtBIk9i`Qi`QjaRjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTnbTqcVo_Rm\Rm_Tqd[og\ldYhbVfbWqma{xi~{j€zj{vcqjWe`JTT<FF.C>(MB.S@/U8*]7*m>4n70u91~;5‡@:’IB›PJ¥XP­ZT»^YºRO·GE¼B?ÈDBÑEDÓD@Ï@:ÐA;ÐC<ÒE>ÔG@ÖGCÚFFÛEFÞDFßCDßCDßCDàDEàDEßEEßECàFDßECÞDBÜD?ÛC>ÜC>ÜC>ÜC=ÛC>ÖE@ÏF@ÃB<³>5¢:1”8-‰9.ƒ;/‚@2?3~>2€>2ƒ:1†91‰92‰90‡;.…<-„:-„:-ƒ;-‚:,‚:.‚:.9-9-~:/~:/~:/~:/~:/~:/}90}90}90}90|90|90|90{8/{;1{;1z:0y;0y;0x:/x:/x:/t8.t8.s7-s7-r6,q5+p6+o5*j1&p7,m4)k2'l6*i3'm7+}G=‡PIˆQJŒQKŽQL‘RM“TO”UP‘VPŒWO‡[P~^OtbNoiQorUlwYfwUQeBK_:EY6DU3EV6GV9GT8ER8@J2<D/9>*7;*:<.<>0;=/9;.6<25=26<26<0:</<>0=@-=C)MW5Zi@h{Mn‡Pp�Qs˜RuŸUtŸWz¡`w�`n”Wc‰L\‚E\‚E\‚E[�D_…Ha‡Jc‰LdŠMc‰La‡J_…H]ƒH[}JPr@Gh9i`Qi`QjaRjaRjaRkbSkbSkbSjaRjaRjaRjaRjaRjaRjaRjaRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTnbTrdWpbUm_Tl_VkaWlcZle[jf]jf]yujˆ„x‘��•��’Œ~‰‚r}zg`cNOT>DG2GF2HC0F=,L<,VC4P9+R6*V4*Y4+\6+a9/e=3n>4ŠKB�RL®UQ·OL¼FFÆABÒDCÙGGÔ@>ÕA?ÖB>ÖE@ÙECÚFDÛEDÝEDÞDDßCDßCDÞDDÞDDÝEDÞFCÞFCÛFBÛFBÚE?ØE>×D=×D=ÙD=ØE>ÔD<ÓF?ÉF>ºA8§;1˜8,Œ8-…;.ƒ@0@1~?0€>0ƒ:1„:/ˆ:0ˆ:.„:-„;,„:-„:-‚:,‚:,�9-9-9-9-~:/~:/~:/~:/~:/~:/}90|90}90{8/{8/{8/{8/y9/{;1z<1y;0y;0x:/x:/x:/v:/t8.s7-s7-r6,q5+o5*o5*m4)j1&o6+l3(h2&k5)h2&l6*|F<ŠOGŒOJ�PK�RM‘TO‘VP�WPŒZQ‰]R…aUy_Pj[Hd_IdhO\fKN]@EU8AQ4>M0=L/?M3DQ7FR:GS;BK6>G2:A/9=,9=.9=.8<-7;-7=36=56;47=3:<1;=/<?,<A*JS4Ve>dwIl…Lp�QušTw¡Wx¢Z{¢ar˜[e‹N\‚EZ€C\‚E\‚EZ€C_…H`†Ia‡JbˆKa‡J`†I^„G]ƒHZ|INp>Ef7haQhaQibRibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlcTrfZth\sf]ndZjaXle]snh{xq�~w��† �”¨¥œ«§œ¨¤™£�‘™–‡z{kdgTOR?GJ7EH5BE2BE2EH5KL<JH9GD5D>0B:-A7+@6*G5)S5*qB8ŒPHŸPK¯HI¿GIÌEIÓBEÖ@BÙ@BÚBAÜDCÚEAÛFBÝEBÜDAÞDDÞDFÞDFÞDFÞDDÝEDÝEDÛEDÚFB×FAÖF>ÔE=ÔE=ÓF=ÔE=ÕF>ÔA:ÓC;ÎE=ÁB9®>3Ÿ:.‘;.‹=0„?0�?/€>.�>.�;/„:-‡;.‡;.„:-ƒ;-ƒ;-‚:,‚:,�9+9-9-9/~:/~:/~:/}:1}:1}:1}:1|91z:1|91y90y90x8/x8/w8/y:1y:1x90v:0v:0u9/u9/t:/r7/q6.q6.p5-o4,o4,n3+m4+j1(m7-j4*h2(j4*f2'j6+}D;�NG�NH�QJ�TL�UOŽWPŠYR„\R{YMw]PkZJ]UBYXDY^HOYA?K3:H/7E.5A+5A+8D.?H3CL7EN9BK8@I6<E4:A19@07>.5<,4:,5<45<56;56;49;0:</;>-;@)GP1Sb;buGk„Kq‘RwœVy£Yy£[xŸ`n‘W`ƒIZ}C\E^�G^�G]€F`ƒIa„Jb…Kb…Kb…K`ƒI_‚H^�IY{INo@Ef9haQhaQibRibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlcTkaUoe[pf\lcZkd\rmg„�|‘�Œ§£ µ±®ÄÁ¼ÉÆÁÉľÅÀºÀ¹±¹²¨¤œ�ˆ‚rgdUPQ?FI8>E3;D1:F28E38E38C3:B3<B6@C8CD<HE<>4*P:/e?6€EA¡PO¿X[ËRWÊCIÕFJ×EFÚFFÜFEÞFCÞDBÞE@ÞCAÞDDÞDFÞDFÞDFÜDCÜDCÜDCÚDCÙEA×FAÕF@ÓF=ÓF=ÑG=ÑG=ÔE=Ô?8Õ@9ÐC:ÆC9¶A7¥=0–=/Œ=.†>/‚?/�>.�>.�<-ƒ;-…;.…;.ƒ;-‚:,‚:,‚:,�9+:+9-~8,~:/~:/~:/~:/}:1}:1}:1}:1z:1z:1z:1y90x8/v7.v7.t8.w8/u9/u9/t:/t8.s9.s9.r8-q6.q6.p5-o4,o4,m4+l3*j4*h2(m7-i3)f2'i5*e1&i5*{B9ŽMG’MH‘PJ�TL�VO‡XN‚ZP|\Qs\Nl\M`WFVSBQRBJQ?@I64@,3<+2;*09(09(2;*5>-9B1<E4?H7?H7>G6;F5:C25@/3>-1<,3:33954954928919;.;>-;@*CL/O^7`sEk„Mr’SyžYy£Yw¡Yq˜Yf‰OZ}CX{A\E^�G^�G^�G`ƒI`ƒI`ƒI`ƒI`ƒI_‚H^�G^�KXzHLm@Cd7haQhaQibRibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlbVj`VjaXkdZlh_uplˆ„�Ÿžš°°®ÃÂÀÓÒÐâáßçæäçãàâÞÛÜ×ÓØÏÈμ²¸¥——‡zym]_YIKL:AF2<E0?K7>I8>F7;C6;>5:;3983872==3C<2N71a84ˆHH¯]_Å`dÊWZÍQSÓNO×MKÛKJÞHGáFDãDAâBBßCFÞCGÞDFÝCEÝCEÝCEÜBBÛCBÚDCØD@ÕD?ÓF?ÒE>ÐE>ÐG=ÔE=Ù@:Ù@:ÑB:ÉD;¼C:¬@4™;/Œ9+‰=/ƒ>.�>-�>-�<,ƒ;,„<-„<-‚:,‚:,‚:.�9-9-9-~8.}9.~:/~:/}:1}:1}:1}:1}:2}:2{;2y:1y:1x90w8/t8.s7-r8-t8.s9.r8-q8-r8-p7,p7,p7,o6-o6-o6-n5,m4+l3*l3*i3)h2(k7,h4)f2'h4)b0%f4)zA8�MG“NI‘PJŽULˆXN�YOxZOr\Nk^N_XHRPAKN=EI:9B13;,08)17+17-06,06,06*17+2:+4<-9D4;F5=H7<I7;F56C13@,1>-.800621622717829;0:<.:?)?H+LZ6]pCjƒLs’Vyž[w¡YsœVg�P^�IVyAX{C[~F[~F[~F]€H_‚J_‚J_‚J_‚J_‚J^�I]€H]€JUvIJk@Ab7haQhaQibRibRibRjcSjcSjcSibRibRibRibRibRibRibRibRkbSkbSkbSkbSkbSkbSkbSkbSlcTlcTlcTlcTlcTlcTlcTlbVqh_ng_mhbtqj‡ƒ€¡ ž¼¼ºÍÏÎÏÏÏàààòòòúøùû÷ø÷óòñëëîãßèÐÆÝÁµÁ«�¢‘�}ub^\GKM7CH2AI2AH6@G7BD9BB:B=9C97?:6:=6>?7C82J.+e77�QR´hjÆnmÀ\\ÇZWÐVS×QNÝMLâHHæCDåBEàDGÝDGÝDFÝCEÝCEÜBDÜBBÜBBÜDCÙCB×C?ÓD>ÒE>ÐE>ÑG=ÔE=ÜC=ÛA9ÒB:ÊD;¿E:°A6œ:-Œ6'Š</…=.‚=-�>-�<,�<,„<-„<-‚:,‚:,‚:.9-9-~8,}9.}9.~:/~:/}:1}:1}:1}:1}:2{;2z;2w;1x90v:0u9/s9.r8-r8-r8-q8-p7,n8,p7,m7+o6+o6+o6-o6-n5,m4+m4+j4*i3)h4)f2'k7,h4)c1&f4)b0%e3(xB8�NH’OI�RM‹VN„XOzZOq[Mj]MZTDIJ:@D5>D69A219,/7*6<247058157247025.06,08-2:-6A3:E5=J9>K9<I77F33B-1>-,6..400511606718:/:<.:?+<E*IW4[nAi�Mr‘Vxœ\užXp™U`†IX{CSv>WzB[~FZ}EY|D\G_‚J_‚J_‚J^�I^�I]€H]€H]LStIHhA@^8icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSjcSkbSkbSlcTlcTlcTmdUmdUmdUmdUmdUmdUmdUmdUmdUmdUldWjc[gd]vsn“’Ž¯¯­ÄÆÅÚÜÛêîïóôöö÷ùüüþÿÿÿÿþÿÿûúüööýòî÷áÔòØÇàɷƵ¡ª ‡ˆ„ibbFEH-EI0CG0FE3GC7F<3C60G53I;:>=9>?:B:7I75cGF’jjÅ“’è­©Ò‹‡ÇtnÀ]XÆSPÕONàKMçFKéFKßDHßIKÛEGÕ<?Ø>@áGIàDGÖ:=Ú@@Ú@@ÙA@ÖB@ÔC>ÒC=ÐA;ÑA9ÞC>ÛA9Ó@9ÊA9¾B8°@5Ÿ:.�8*Œ=0‡>/ƒ>.�>-€=,€=,�<,�<,:+9-9-~8,~8.}9.|8-|8-|90|90|90|90z:1z:1z:1z:1y=5w<4w;3u:2t91q8/p7.p7.o6-m7-m7-l8-m7-l8-m7-m7-i2+m6/p92p92l5.g2*g2*h3+e0(i7.d2)^.$`0&_/%b2(s>6ŒOJ�QLˆQL�RJ}XOx]RhXITN>DE5@D5:B37?24</3;04:049279478366457246116//6./7,.9+2=-6C2:G6:I68G44C.1?.*4+,2.-2./4.45/79.:</:?+6>&ES2YkCj‚Ru“]y�_r›Yk”R[~DXyDTu@Tu@XyDZ{F[|GZ{F^JbƒNbƒN^J\}H_€K`�L_NTsJFd@<W6icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSkbSkbSlcTlcTlcTmdUmdUmdUmdUmdUmdUmdUmdUmdUmdUldYgd]onj†…ƒ¤¤¢ÁÃÂÖÚÛêîï÷ûü÷ûþøüÿüýÿþþþÿþÿÿþüÿüûÿúóÿðàüêÔíÞÇ×ͲÀ»ž¤£„‚ƒcgjKLQ3GJ/EC.GA1H>4I:3L95K<9?:7CB@NIF]RPyjg£Ž‹Î²®èÅ¿ÿ×Ñ벩͂|Àc^ÊVVØRSßHMÚADÖDEÕEEØHHÜJKÛEFÕ<>Ö<>ÛACÙ??Ù??ÙA@×A@ÖB>ÓB=ÒA<ÔA:ÜA<Û@;ÔA:ËB:¿C;±A6 ;1’8-Œ=0‡>/ƒ>.�>-€=,€=,�<,�<-9-9-9-~8,}9.}9.|8-|8-|90|90|90|90{;2{;2{;2z;2w<4u<3u:2s:1r90n8.p7.m7-m7-l8-l8-l8-l8-j8-l8-l8-k6.n70p92n91l7/i4,g2*e3*c1(d4*_/%^.$b2(^.$`0&p>7‚KF‡RN‡XR~XOrVKdRFRI:A?0<@18@15@23>04<14<15:37:379677577557246116/07//7,.9+1<.5B19F59H58G44E24B1.8//6/.3--2+23+46+8:-9=,7?(DQ3YkEj�Uv”`yœbp˜Zg�QY|DWxCTtBTu@VvDYzEYyGYzE[{I_€K`€N\}H[{I^J^~L^{MSnKF_A9R5icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSkbSkbSlcTlcTlcTmdUmdUmdUmdUmdUmdUmdUneVneVneVlfZed_y{x—™˜²¶·ÍÑÒäéìôùüúÿÿûÿÿüÿÿýþÿþþþÿþüÿÿûÿÿúÿÿôþüçúùÝðïÑáåÄÒØ´½Å  ¨ƒ‡‘l`gEQX9EH-DC.JD4LD9KA8J?;C;9SJKia_�wvœ’�¾³¯ÙÎÈêÜÓþèÝÿóéÿäÛð©£ÃecµBEÇDIÚRVÛUTÑMKÊDAÍCAÔFE×EFØDDÙCDÖ=?×>@Ø@?Ö@?ÖB@ÓB?ÓB?ÕA=Ù@;Ù@:ÓC;ËE<ÀD<²B7¡<2“9.Ž<0‰=0ƒ>/�>.=-=-€=-€=-9-9/~:/}9.}9.}9.{8/{8/}:1}:1{;2{;2{;2{;2z;4x<4t;2q;1r90o9/n8.l8-l6,k7,l8-j8-j8-j8-j8-i9-j8-j8-n91m82m82m82m82i70f4-c1*e3,a0)],%_0(c4,\-%_0(sD<}PK�VP~YQpSK^J?OC7@=.46(3:*1<,1<.2=/5=26<27:37:379679668368357227018008-.9+0;-2?.6C27E48G49H59F54>33:205.,2(/0(13(57*7;*6>'ER6[lJm„Zy–fyœfl“Z`‡NWwEWuCUrBUsAWtDXvDYvFYwEZwG^|J_|L]{I]zJ_}K^{KZwKPgJAW@3I2icSicSicSicSicSicSicSicSicSicSicSicSicSicSicSicSjcSkbSlcTlcTlcTmdUmdUmdUmdUmdUmdUneVneVneVofWmg[jkfƒ‡ˆ£§¨¹¾ÁÐÕØåíïóûýõýÿûÿÿûÿÿüÿÿüþûýþùþýøÿþ÷ÿÿóùýæöýÞî÷ØçñÎÞèÃÎÚ´¶À›¡«†€‡efmLNR7FH0JI7KI<HD9D@7OGDj`_‹�€¤š™·¯¬ËÆÂÞÛÔêæÝúñèÿûñÿ÷íÿÝÕûµ³å�ŽÊcf°@?ÃPKÉTMÉPHÄGAÈGBÔMJÖKHÐ@?Ò>>Ó??Õ?>ÕA?ÕA?ÔC@ÓB?ÓB=Ö>9Õ@:ÒC=ÌE?ÀE>²A9¢=5•;2Ž<1‰=0ƒ>/�>.=/=-€=-€</9/9/~:/}9.}9.}9.{8/{8/~;2~;2|<3|<3{;2{;2z;4x<4r90o9/o9/m9.l8-k7,j6+h6+i7,i7,i7,h8,h8,h8,h8,h8.m;4l71j5/i70j81j81f4-a0)e4-a2*].&a2*b3+Y*"a2*yNG�\VvWRjOHXD;I<3@:.:;-8<.08)/:*3;.4</6<26<27:169279479468349238139/19.19..9+/:,1<,2?.5B17E4;H6<I89C87?428.-3'./'/1&24'48'4<'ER8_pPtŠd}šny›if�XU|GSsBTqASp@Sp@TqAUrBVsCVsCXuE[xH]zJ]zJ^{K_|L\yIWrIK_F<M;.?-gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSkbSlcTlcTlcTmdUmdUmdUmdUmdUneVneVofWofWofWng]qqo‡Œ�¤©¬·¿ÂÍÕØãíïòüþôþÿøÿÿûÿÿûÿüýÿúüýõúüñøøìöøê÷ÿèóÿäòüáîùÙêôÒÞèÅÊѯ¶½›�£�}‚bY]BIM6GJ9FJ;CF;AB:c^X…}z«¡ŸÁ·µÌÂÁ×ÏÍæáÝñîéú÷ðüõíÿñèÿòéÿóíÿåß騤¸id§MBµPDÁYNÈ\RÆQH¿D=ÂA<ËDAÌA>Î@>Ð@?ÑA@ÒC?ÑC?ÑC?ÒC?Ó?;ÒA<ÐE@ÉE@½C>¯@9 =7•<4Ž<1‰<2ƒ=1�=0=1=/=/=/~:/~:/~:/}9.|90|90{8/{8/|<3|<3|<3{;2z;4y:3y:3w;3n70m80m80l7/k6.i7.h6-h6-i7.h8.h8.g8.g8.g8.g8.h8.k92h6/f4-g5.i70h70e4-b1*c2+c4,_0(]1(]1(W+"c7.{ULtXTaNHN?8A7.=6,;9-9;.9<139-3;.5;/6<06<07:/68-47.69049238139/39/2:/19.19,08+.9+/:,1<.3>.6C2:E5<G7=E8:B55;/06(02'02%13%26%2:%GS;dtWy�k~šrs”g]ƒRJp?Oo>Qn>Qn>Qn>Qn>Ro?Sp@TqAWtD[xH]zJ]zJ_|L`}MZwGSmFCU?6D7(6)gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSkbSlcTlcTlcTmdUmdUmdUmdUmdUneVneVofWpgXpgXoh^lnmz‚…“š «µ·ÄÎÐÜæèí÷ùôþÿõþýöÿüúÿúûÿùûþóøúíõöèñôãõüêôýèôüåóùßòöÛéìÏØ×¹ÅĦ±°’�ŽrggONP;GK:FL>HPEKQG{{sŸš”ž¸ØÎÌßÓÓçÛÛóëéü÷ôðïëüûöÿÿöÿýôÿøñÿóêÿçÞÿ×ËØ“ƒ¸eS¤K;´TF½WIµF;·@8ÈKEÇC>ÊC?ÍC@ÎC@ÎC>ÎC>ÍD>ÍB=ÑB<ÒC?ÎDAÅD?¸A=ª=8ž;6•<6Ž;3‰<4ƒ=3�=2=1=1=1=/~:/~:/~:/}9.|90|90{8/{8/|<3|<3{;2z:1y:3x92w81u91n70m80l7/j8/i7.i7.h6-g7-h8.h8.g8.g8.g8.g8.g8.g8.h70h6/g5.g5.h70h70g6/f5.a2*e6.^2)\0'`4+a5,i=4uSJXG@E>6<5/:6-;8199/57,36+7:/5;/7:/7:/69.58-57,36+28,19,28.19,19.19,19,19,08+08+.9+/:,0;+3>.6A18C3<D5:C28>05;-57*35(13%04%18&HT>gw\wŒmv’lhˆ_RwKBh9Kk<Nk;Pm=Pm=Ol<Nk;Pm=Qn>WtDZwG]zJ^{K`}M_|LVsCNg@<J90;3%/'gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSkbSlcTlcTlcTmdUmdUmdUmdUmdUneVofWofWpgXqhYoh^hikmtz…Œ”£¬±¿ÈÍÓÞâåðòðüüòüûõþùøÿ÷úÿöüÿòúüîøùçö÷çööêù÷ëú÷èüöæýõâ÷ìØçØÃÕƯÀ±šž“}vn[ZWFPQCSVK\c[fme•˜�²²¨ÕÎÈèÝÛóãäúêëÿñôÿ÷ùÿýýþýùùúôøùñüüòÿÿóÿÿñÿûêÿæÑñª”¼o[©VD®RCµPD»LA¿JAÂG@ÆE@ÉE@ÊE@ÊE>ÉD=ÈC<ÉB<ÑC?ÐE@ÌEAÂC=³=9¦;5œ;5•<6Œ<5‡<6…<5�=4<3=1=1=1~:/~:/~:/}9.|90|90{80{80{;2{;2y:3x92w81v70v70t80n70k90j8/j8/j8/h8.h8.h8.g7-f7-f7-f7-d8-d8-d8-f7-g6/h70i81i81h70i81j;3l=5g80l=5c7.a5,oC:xLCtI@nMDD:167/45/891:;556.14+25,58/58/57,46+46+46+46)48*08).9)08+08)08+08)08)08)19,08+08+08+08+2:-3;.4<-9B1:C0;B2:A1:<.68*35'/3$29'IU?eu[oƒghƒbYyTInE?d8Hg;Li;Ol>Nk=Mj<Li;Mj<Ol>UrDYvH[xJ]zL^{M\yKQn@G_;6B4+4/#)%gdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSgdSicSjcSjcSkcVlbVlbVmcWmdUmdUmdUleUmfVngWogZoi]oh^mjekormtzƒŠ�¥®³ÃÌÑÕßáæîñôüþôýú÷ýùùþ÷úÿöýÿòýÿñüÿîýþðûøïþúñÿùïÿúíÿùéÿòßïàËÝηÁ²›¢—�|ubb_NXZL^dXmwn|†}¦¬¢¾¿·ÚÕÏîäâüîîÿôôÿô÷þôõþøøûúøûýøüÿúøÿõôüñ÷ýñÿÿñÿùèÿôãÿÌ»½pœRE©SF¶TI²C:ÀIAÄH@ÇF@ÉF>ÉF>ÈE=ÆC;ÆC;ÍG>ÌG>ÇG>¼C;­>5 ;3˜;4‘>6Š=5†=6ƒ=5�=4€<3€<1�;1�;1~:/~:/~:1}90|90|90{80{80z:1z:1x92w81v70v70u6/q6.m80k90k90j8/i9/i9/i9/i9/g7-f7-g7-f7-f7-f7-f7-f7-f5.j81m;4k:3j92j:0m>4oC8mA6oD;e=3d>3zVJ†dZyWMbLA>:13814927<569214-25,9<336-25,24)03(13(25*46+47,/7*-8*/7*/7*/7*/7*/7*/7*19.08-08-/7,/7,/7,08-08+7?09B1<E4=D4;A38<.26'-4$2;*JVBcr[i{a^uXRmLGeAAa:Hd;Kh<Nk?Nk?Kh<Jg9Li=Nk=TqEXuIZwK[xL]zNZvMNjAC[;2>4*00!''heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjdVjdVkeYldYlcZmeZmeXmfVg`NkfSnkXol]he\feasqr�‚†Š�“–ž¡«³¶ÃÈËØÝàêîïö÷ùùûúüþýþþüÿÿúýþöûüôúüñûýòüþóùúòúúòùùíüúëþüçù÷àéåÊÕÔ¶º¹›˜˜|qtY_dMdkYr}l„�‚‘›�´¹²ËÌÇãâÞòîë÷óòû÷öýùøüø÷ÿþüþþüþþüýÿúýÿúýÿúýÿúýÿúúü÷ÿÿúÿûöÿêåا £`XœE>µNG¿MCÂF>ÄD9ËF=ÎI@ÎKAËH@ÇG<ÆG8ÃH8½J8³H8¦E5šB4‘B5‹C5†B7„B6ƒA5‚>3‚<2„:1†91†91€:2~:1|91|91{80z7/w7.w7.w7.w7.w8/w8/v7.u6-q5+o4,l7/i81i81h70h70g6/g6/g6/f5.g6/i70i81i81g6/f5.e3,h3-j5/m80k90j:0i;.j>1kA3gB2jI:dH:v_O~k\ŠxjŒoRJ=79.4:04:039/39/39/28.28.17-17-36-06,25,06,14+/5+/4-.5-/4-/4-/4-/4-/4-/4-05.05./4-.3,.3,/4-05./6.1;23=26@59C89D67B45@03>-4?.DQ=XfO[kQRbHEX<BU9BW6Jd?Ke>Ke>Ke>Je<Je:Mh?Oj?UpGYtK\vO]wP]wRWqNHaA;N8&1+&+.$),heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVkdZkdZlc\md[meZmfVidQjhSmjWjk[gg_lll~„�“š¨¯µ¸ÂÄÑÖÙâçêòóõûüþÿþÿÿþÿþýûÿþúÿÿúÿþùûüôùúòøûòùüóøùñøùñøúí÷úéùûåòõÚßâÅÌÏ°´·˜‘–vkqUZbJbkVu€oŒ˜Š�§œ¿Á¼ÓÓÑççåóóñ÷÷õûûùüüúûûùþþüþþüþþüþþüþþüþþüþþüûÿþõÿÿõÿÿþÿÿÿûúÿíêð¿ºÂ~uŸKA¯MB¹J?ÂL@ÅK>ÃE9¼=4¾?6ÃE9ÄF7ÃJ7½L:³I9¦F6™C4‘C6�E7…C7‚B8�A7�>5‚<4ƒ:3†91„93:3}:2{;2z:1y90x8/v7.v7.v7.v7.u9/u9/t8.r6,p6+m4+l7/i81j81h70i70h70g6/g6/e3,f4-g5.h6/j81j81k92k92k60l71k90k;1l<0j>1j@0iD2gH6lQ>gRAvgT�ue‹†s†ƒrKL<69.39/39/39/28.28.28.28.17-17-17-06,06,06,/5+/5+/4./4./4./4./4./4./4./4.05//4./4..3-.3-/4./4./6/.80/:22>44@66B66B66B45B16C1CP>P]IR`IIW@AO6BP7FV;Jb@Ke@LfAKe@Ic<Hc:Id;Je<SnEXrK\vQ]wT\uUTmOC\?6I5&1-',0%*.heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVkdZkdZlc\md[meZmfVmhUljUmjWjk[kkcwww‘’—§¬²¿ÆÌÒÜÞìñô÷üÿþÿÿþÿÿÿþÿÿþÿþýûÿþúÿÿúþýøúûóö÷ïóöëóöëõ÷ìöøíöøêõøåõ÷áîîÔÛÛ¿ÉÊ«°±’“–yuy`kpZt{iˆ��Ÿ§œ®µ­ÍÏÊÞÞÜïïíøøöûûùþþüþþüýýûþþüþþüþþüþþüþþüþþüþþüýÿþûÿÿûÿÿþþþÿûùÿúôÿ÷íýÑÆØž�•M?¥RD©L=§@1·G;ÏYMÏUJ¼@4ÄD7ÆH:ÄK@¹I=«C8�?5—@7“D=‰@9…B:�@:>8~>5}=4<4}=4z;4x<4x<4w;3u:2t91t91t91r90r90r90r90o9/n8.l6,k5+m80j81l71j81k60i70i70h6/g5.g5.f5.g6/h70i81k:3l;4l;4j;3i:0h<1k?2jC4iD2fE2gJ8lVAjYGujV†m‘Ž{€mDG658-39/28.28.17-17-17-17-17-17-17-06,06,06,/5+/5+/4./4./4./4./4./4./4./4./4./4..3--2,-2,.3-/4..5.+5-+6.-9//;12>25A56B47D38E3?L:ER>CQ:;I28F->L3DT9J_@Jb@LdBKc?Ia=G`9F_8G`9RkDXpL]uS^vVZqTPgK=T8.A.&1-',0%*.heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVkdZkdZlc\md[meZmfVniVljUlkWlm]pqi€‚�ž¡¦¹¾ÄÄËÑÛåçõúýûÿÿüýÿüýÿþüýüûùÿþüÿÿûÿþùüüôõöîïñæëîãéìáîðãñóåñôãñôßòñÜêêÐØؼÇÇ«²±•¡¡‰‘“}�“€œž‘ª­¢¶»´ÀÅ¿ÜÜÚééçööôûûùýýûÿÿýÿÿýýýûþþüþþüþþüþþüþþüþþüþþüþþþûûýÿþÿÿþÿ÷òïüóìÿüóÿúíÿôäì²¹�p•P@¢O?¸XJ·M?µE9ÀG<ÁA6ÈE;ÅIA½F>¯@9¡;6š=8—B?�?;‡@<�@<}@;z?9y@9y@9y@9v=4v=4v=4u<3s=3r<2r<2r<2o;0o;0o;0n:/k9.k9.j8-i7.m82n72n72l71l71k60i70i70l:3j92h70f7/f7/g80h91g;2j?6f>4e=1f>2hC3iE5fG3cG2cJ6hV@i\IskV…‚o��zor_9=,28,28.28.17-17-06,06,06,17-17-17-06,06,06,/5+/5+.3-.3-.3-.3-.3-.3-.3-.3-.3-.3--2,-2,-2,-2,.3--4-*4,)4,*6,+7-.:01=14@46B46C2:G5<I78E12?+2@)8F/>N4DX<E]=K`AJb@K`?G_;H^:F^:ShGXpN`uV^uX[oTLcI9M4):(%.+%*.$)-heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVkdZkdZlc\md[meZmfVlgTljUnmYop`tum…‡†¤§¬¿ÄÊÎÕÛæðòûÿÿûÿÿûüþýþÿÿþÿþýûÿÿýÿÿûþýøøøðïðèèêßâåÚàäÖçëÜêïÛíðÛìðÙîîÖèèÐÚØ¿Ìʱ½§·µ ±®�¶´§Á¾µÉÈÃÒÓÎÙÙ×èèæòòðúúøüüúýýûÿÿýÿÿýüüúþþüþþüþþüþþüþþüþþüþþüÿþüÿüÿÿüÿÿûúüû÷ýÿùþÿøÿÿôÿþíÿþìøßËÅœŠ�bP–J:¥L<µQA¼L>¿@7ÆC;ÅF@¿D?±>;¦:8ž::™>=‘=;Š?<„@=}@=xA<tB;rC;rC;q?6s>6s>6r=5r=5r=5p>5p>5n>4m=3l<2k;1j;1i:0i:0j:0n72p62p62n72m61l71i70i70l;4k:3h91e90e90e90f;2d<2fB6dB6cC4dD5dG5dI6bI3_I2^L6eV?jbMrmW„ƒo„‡r\aM3:(28,28.28.17-17-06,06,06,17-17-17-06,06,06,/5+/5+.3-.3-.3-.3-.3-.3-.3-.3-.3--2,,1+,1+,1+,1+-2,-4-,6.+5-*4,*4+,6-.8-0:/2=/2=-5@/7B15A-2>*1=)5A-8F/=O7@T9FX>H\@J]?I^=K^>K`?TgI[oScvZauZZlTJ]G6G4'5&$-*$),"'*heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVkdZkdZlc\md[meZlfVjgTkjUlnYorasvm�…„Ÿ¤¨¸¿ÇÖßäí÷ùûÿÿûÿÿûüÿþÿÿÿþÿþýûÿÿûÿÿúüüôôôêëíâäæÙßãÕÞâÓãèÔçìÖéíÖêìÔëëÓèæÏßÚÄÓθÓκÌƶËøÐÉÁ×ÒÎàÛØêæåðïíóòðùù÷ýýûýýûýýûÿÿýþþüüüúþþüþþüþþüþþüþþüþþüþþüÿýüÿûüÿøùÿþýþÿýùÿýôÿøðÿõ÷ÿöð÷çÿÿíÿîÛͧ”¤kX¢[GªWE­L;¹D:ÀD<ÀE>¼E?´A>ª=:¢:9š:;•=<�?=…?=~A<wB<sD<oD;oD;p?8q?6q?8q?6q?8q?6p?8o?5o@8n?5m>6k<2i=4h<1h<3l;4o83r73p62p62n72l71j81i81h70g80e90e:1d<2e?4f@5dB6_C5aG8cL:dM;cM8`K6^L4]M4^R:bX?mhRss[€ƒnv{eFO:4=*39-39/28.28.17-17-17-17-17-17-17-06,06,06,/5+/5+.3-.3-.3-.3-.3-.3-.3-.3--2,,1+,1++0*+0*,1+,1+,3,/6/,6.+5-*4,*4++5,-7,.8--8*0;+4?/5@/4?.3?+4?.3@,7F19K3@O8CU;GW<H[?M]BM`DUeJ\nTdtZ`rZXgRGXE4B1$2%$-*$),"'*heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjdVjdVkdZkdZlc\md[lfZlfVliVlkVkmXlo^lrhz€~•� ®·¾ÒÛàêóøûÿÿùþÿùúþþÿÿÿþÿúù÷ÿÿûþýøùùñòòèéëÞäæØáåÖâçÓäéÒçíÓèìÓçéÑèèÐçåÎàÛÇ×оÚÓÃÙÑÆÞÔËæÝØîäãóéêúñôÿùûúù÷þþüÿÿýýýûýýûÿÿýÿÿýüüúþþüþþüþþüþþüþþüþþüþþüÿýüÿùøÿúùÿþýüÿý÷ÿþîÿùñÿÿòÿý÷ÿúõùëÿýíÿúèôλ«wb�Q:ª\H²N>¸J=¹I>¸I@´G@¯D>¥=:œ:7–=9�>:ˆ?9€A:yB;tD:pF:pE<p?8q?8q>:q?8q>:q?8p?:p?8qB<pA9n?9l=5j=7j>5j=7n=6o83r73q73p62n72l71j81i81d8/d90d<2d>3d@4cA5cA5_C5ZE4_N<gVBhWCcS<^O8[O5\R7]V<]Y>nlUtv^|�kcmU2>(6B.3;.39/39/39/28.28.28.28.17-17-17-06,06,06,/5+/5+-2,-2,-2,-2,-2,-2,-2,-2,,1+,1++0*+0*+0*+0*,1++2+-4-+5-*4,*4,+5,+5,,6--7,,6+/:,2=/6A17B27B15@04?.2?-4C.:G3=L5AO8DT:JX?L\BTbI\kTcp\^mZUbQDRC2>0#/%'0-&,,$**heVheVheVheVheVheVheVheVheVheVheVheVheVheVheVheVjdVjdVkdZkdZlc\md[lfZlfVolYmlWjlWgkZgmcr{xŽ˜š¨±¸ÊÓØãìñøýÿ÷üÿúûÿþÿÿÿþÿúùõÿÿúýýõøøîññåéëÝæéØåéØçìÖæìÒéíÒêíÒèèÎèæÏèãÏáÚÈØÐÃØÎÄÝÒÌéÞÚøíëÿôöÿõùÿõûÿùüÿýþÿÿýÿÿýþþüþþüÿÿýÿÿýýýûþþüþþüþþüþþüþþüþþüþþüÿþúÿýûÿýûýüúùýüøÿÿõÿÿóÿÿíýúóÿûüÿúÿÿôÿöæÿóßÿãÍ×­•©oY¨WD®Q@¬O>¯OA±OB­K@¤C<™<5•>7�?6‰@9�A7{B7uD6qE8qE:p?8q>:s>:q>:q>:q>:r?;q@;sB=qB<p?:m>8l=7k>8m>8o>9o83q73o83m82m82j92h91f:1f;2d>3d@4bB5`C5]A3\@2WB1TG4\T?f^Gj_IbX?ZS9ZS7\W:[W<XW;kmUsw^u}fUaI&28D03;.4:04:039/39/39/28.28.17-17-17-06,06,06,/5+/5+-2,-2,-2,-2,-2,-2,-2,-2,,1++0*+0**/)*/)+0*+0*,1++2+)3+*4,+5-,6-,6--7.-7.-7./9.2<16A39D69D47B46A10;+2?-6C19F2<I5AO8FT=JX?R`I[hTan\]jYQ]OBN@0<0",#+1/(..&,,ifWheVheVheVheVheVheVheVifWifWifWifWifWifWifWifWkeWkeWle[le[md]ne\mg[mgWqn[kmWknYkr`kshpyv‡‘“Ÿª°»ÄÉÐÙÞêïóöûþüýÿþÿÿÿþÿþýùÿÿúþþöúúð÷÷ëôöèñôãìðßéîØñõÚíòÔëîÓììÔèæÑáÜÉÝÕÈÞÔËáÔÎèÚÙòääúëîþòôÿ÷ûÿùÿÿüÿÿýþþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþÿúÿþúÿþúþþüýÿþûÿÿúÿÿúÿÿûÿÿþÿÿÿÿýÿûõÿûñÿüìÿûåÿóÛÿêѺ‚i¦bK—O9ŸT?¨ZF£RAžK; M?Ž<0‹=0†>2=/x=/t?/sB3uE7r=7s>:t=:t?;u@<u@<t?;q>:s@<rA<r?;q@;p?:o@:o>9o>9k92m82l:3l;4l;4h<3g<3d>3c?3^>1dG9cG9T=-N9(M8'@1XR<PN7KI0OK2]Y>ieJeaDVU7XX<^aDpt[z€fdlU>J2-9#2>*4</5;15;14:04:039/39/39/39/39/39/28.28.17-17-06,16016005/05/05//4./4./4.,1+,1+,1+,1+,1+,1+,1++0*.5..5.-4--4--4--4-.5./6.+2*-4,1805=29A6=E8@H=?J<7B26A03>-1=)3?+9E/BN8IU?O[GWcO]hWZeULWI<F;/90)3+)/+(.,'-+heVheVheVheVheVheVheVheVifWifWifWifWifWifWifWifWkeWkeWle[le[md]ne\mg[khWonZjlVknYlsajtiq|x‰”–£®´¾ÇÌÒÛàëðô÷üÿýþÿþÿÿÿýþþýùÿÿúþþôûûïøùëô÷æñôáìñÝèíÖêïÑæëËäçÊææÎãàÍßÙËÞÔËßÔÎçÙØíßßöçêûïóÿôøÿ÷ûÿúþÿüÿÿþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿúýÿúýÿüýþÿþýÿÿüÿÿüÿÿûÿÿüÿÿüÿÿûùÿüõÿþñÿÿíÿûåÿòÙÿëÑà±—§u\’\D–ZBšYCšVCšSAŽG5‹F6ˆE4ƒD3|C2wB0s@/o>/u@8t?9u>;u@:t?;p>7p=9tB;q>:p?8p?:o>7o>9m>6n=8m>6k<4l=5m>6k?6j?6gA6eA5bB5dG9[A2^G7_J9O>,F7$G8%?7"LL4HL3DG,CF+MM1[[?baC`aBZZ>dgJsw^sy_X`I:D,/;%7C/6>17=36<26<26<25;15;15;15;15;15;14:04:04:039/39/27127127116016016005/05/.3-.3--2,-2,,1+,1++0*+0**1**1*)0))0)*1**1*+2+,3,.5..5-.5-07/4;3:B7?F>BJ?;F8:E57B14?.3?+6B.<H2@L6LXDS_KYdTWbRLVK=G<1;2+5-+1-*0,)/-heVheVheVheVheVheVheVheVifWifWifWifWifWifWifWifWkeWkeWmd[md[md]ne\mg[khWnmYimVjoYktaiuiq}yŠ˜™¥²¸¿ÊÐÔÝâíòö÷üÿüýÿýÿþÿýþþýùÿÿ÷ýýóúúîøùëô÷äðóÞéîØåëÑâçÉÞáÂÝÝÁÞÞÆÞÛÈÝ×ËàÕÏäÙ×îâäóçëúîòÿôøÿ÷ûÿùüÿúýÿýýÿþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿüúÿüúÿüûÿþþþþÿüÿÿúÿÿùÿÿøÿÿùÿÿúÿÿúúÿýøÿÿöÿÿóýÿïÿúæÿóÝÿøàÿìÕ㿧¬‚j‡YB†R<’[F—]I�S?F2u>)r>)tA,tD0vE6q?4n<3q<6tB9tB;o?5q@9xH>n=6m>4m>6m>4l=5j>3l=5j>5g?5gA6hB7gC7eC7cC6aD6^D5bK;UB1ZI7`S@RG3D<'E=(FA+BF-BH.AE*<@%?B%LO2_`AijKaaEnqTvzagmSJO95=&5=(=F39?38>48>48>47=37=37=37=38>48>48>47=37=37=36<26<25:45:449349338238238238205/05//4..3--2,,1++0*+0*).*).*).*).*).**/++0,,1-162/4.,1+,1+05/6;4=B<AH@@K=>K:<I78E34A-2?+3A*5C,DQ=KXDQ^MQ^MHTH<H<2>4.8/-4--4-,2.gdUgdUheVheVheVifWifWifWifWifWifWifWifWifWifWifWkeWkeWmd[md[md]mf\mg[jiWkmXimVjoYjs`hthn}xˆ˜˜¦³¹¿ÊÐÓÜãëðööûÿüýÿýÿþÿÿýÿþúþþöýýñúûíøùéô÷äïóÜçìÕãçÎÜßÀ×Ú»ÖÖ¼Ù×ÂÝ×ÇÞ×ÍåÛÙìààøìðûðöÿõûÿùþÿúýÿûûÿûûÿýüÿþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿüúÿþøÿþûÿþþþþÿüþÿúÿÿùÿÿøþÿúýÿùüÿùùÿýúÿÿúÿÿøûÿõüþðÿÿïÿûëÿñßÿðÝÿïÚñÒ½´�zyS>tJ4uH3vI4zM8~O;yL7nA.e7'{M@tE;qB8sD:oC8g;0d8-g=1i?3i?3i?3g?3g?3f>2f>2e?2cA5bB5`C5`D6_C5[D4ZE4WD3XI6OB/\T?oiSc^HMK4IG0IK3>D*?H-@F*<B&<A#EJ,XY:deFmmQxx\tv^[_FAD/6;%:?+>B19?39?59?58>48>48>47=37=39?59?58>48>48>48>48>48>47<67<67<66;56;56;55:45:438238227105//4.-2,,1+,1++0,+0,+0,+0,+0,,1--2.-2.051.3/+0,+0*-2.3828=9;B:AL>@M;@M;=J68E14A-1?(1?(=J6BO;IVDJWFDPB;G;2>4.:0,6..5.-4-gdUgdUgdUheVheVifWifWifWifWifWifWifWifWifWifWifWkeWldWmd[md[md]mf\mg[jiWlnYinWiqZgs_drejyt…••£²·½ÈÎÒÛâêïõôùýûüÿýÿþÿÿýÿÿúÿÿøÿÿóþÿñüýë÷úåñõÞèíÖãçÌÙÜ¿ÕÕ¹ÒйÔѾÚÔÈàÙÑêàßôéíþóùÿ÷üÿûÿÿýÿÿþÿÿþüÿþúÿþúþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüýÿþúÿÿúÿÿûÿÿþþþÿýþÿüüÿûüÿûúÿúøÿù÷ÿúøÿüùÿÿûÿÿûýÿúýþùùôîÿþöÿýôÿüóÿûîÿ÷çûæÕéÒÀÏ´¡¶š…”uawVClI6mH6rM=uP@mF7jC4gB2kF6nI9nJ:oK;pL<eA1cB1cB1cB1cB1cB1bC1`C3]F6\G6]H7ZG6ZG6WH5TG4RG3PH3OI3fdM~~frrZVX@JL4GM3<E*<F+?H+AG+AF(CH*MN/TU6xw[{z^nnVUU=CB.?A,?@.;>-;>39?59?58>48>47=37=37=38>48>48>48>48>48>48>48>49>89>89>89>88=78=77<67<66;56;55:438227105//4./4.-2.-2.,1-,1-,1-,1--2.-2.-10,0/,0/-2..211623764;4;G9=L9@O<@O:<K67F/4C,2A*6E.:I4@O<BP??M>8F71?2-9-,6--4,,3+fcTgdUgdUheVheVifWjgXjgXifWifWifWifWifWifWifWifWldWldWmd[md[md]mf\mg[jiWkmXinWiqZgs_bpcgxr‚””£²·ÀËÑÔÝäëðöõúþûüÿýÿþÿÿýÿÿúÿÿøÿÿôÿÿóÿÿïúýèó÷àéïÕãçÌ×Ú½ÑѵÍË´Î˺ÖÏÅÞ×Ñìâã÷îóÿöþÿúÿÿýÿÿþÿÿÿýþÿúþÿùþÿúþÿúþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþþýþÿýþÿýþÿýÿþýÿüþÿúþÿùÿÿ÷ýýóýýõþýøþýùÿþüþþþÿýÿÿþÿþùÿÿüÿÿûüûöóÿøòÿýôÿþñÿþíÿúéÿóà÷äÓçÒ¿Ôº©ºŸŽŸƒu�rd|]K}^LtUCbE3X;)Y<*[>,X=*`E2`E2`E2`G3`G3`G3`G3^I4\K7\M:\M:ZM:XM9UM8SM7QL6NL5QQ9jlT|€gmqXSY?FL2>G,<F+<F+?H+DJ.EJ,DG*HI*LM.yx\tsWfdMTR;LI6KJ6FE3<=-<?4:@69?59?58>48>48>48>47=37=38>48>48>49?59?59?5;@::?9:?9:?99>89>89>89>89>88=77<66;55:4493382382/40.3/-2.,1-+0,+0,+0,+0,+/.,0/-10.21/32/32/32.5.4@27F3=L7@O:?N7<K49H18G05D-8G0<K6>M:<J97E61?2-9-+5,-4,,3+fcTfcTgdUheVifWifWjgXjgXifWifWifWifWifWifWifWifWldWldWmd[md[md]mf\mg[jiWjlWhmVhrZfs_cqdhys…——¥·»ÇÒØÚãêðõûøýÿüýÿýÿþÿÿýÿþùÿÿ÷ÿÿóÿÿðýþìøûæðôÝäêÐÞâÇÕÕ¹Î̳ÉÄ°ÊÄ´ÐÉ¿ÚÒÏéàãöíòÿ÷ÿÿúÿÿýÿÿþÿÿÿýýÿúüÿøüÿøýÿúþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþþÿüÿÿüÿþýÿýÿþúÿúøÿùøÿõøÿóøÿóøÿõüÿøýÿúÿýþÿüÿÿüÿÿûÿý÷ÿÿúÿÿüÿÿýÿÿþÿþýûûûóúøìÿÿñÿÿïÿÿïÿþíÿýíÿöçÿîâýêÛúãÑÿæÐãÊ´¥�xt^GeO8cM6\H0^J2\J2\J2]K3]K3^L4^L4\M6YN8XP;XP;VP:UO9SN8PN7NN6MO7PT;`fLfoTU^CEO4AK0<F+?I.?I.CL/HN2HM/GJ-OP1YX:qmRkgLa\F[V@YTATQ>MJ9BC3>A6;A7;A7;A7:@6:@6:@69?58>48>48>49?59?5:@6:@6:@6;@:;@:;@::?9:?99>89>89>8:?9:?9:?99>88=78=78=77<6495384162/40-2.,1-,0/+/.,0/-10-10.21.23.23-12,2.+9*/@-6H2<N8=O7<N6:L2:L25G/7I1:L6<M:;L:6G70@3-;.,6--4,,3+cdTcdRdeUgfThfWigXigXjhYhfWifWifWifWifWifWifWkeWkeWkeWmeZle[md]mf\kg[jiWikVhmWis[gt`dqgk|v‰›�¬»ÀÍØÞàçíôùÿúÿÿûÿÿýÿüþÿúþþöúúðúûíúûëøùçôöáëí×ßãÊÙÜÁÓѸÌDZž¬Æ¾±ÌüÖÎËæÝàôêòþ÷ÿÿúÿÿþÿÿþÿþþþýÿúüÿúüÿúýÿüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüþþüÿþüÿþüÿþüÿýþÿüÿÿüÿÿýÿýÿþûÿüøÿù÷ÿ÷÷ÿõøÿöùÿöüÿøÿÿûÿþýÿüÿÿûÿÿúÿÿúÿÿûÿÿüÿÿýÿþýÿûýüüþûþÿúùúòûûñýúñþúïÿùìÿûìÿýïÿýìÿðÛÿûâìÜÞŽu`P7RB)XH/XH.[M3[M3ZN4ZN4[O5[O5\P6[Q8UO7TO9TO9QO8PN7NN6KM5IM4JN5LR8S\APY>@I,<E(AL.?J,DM0EN1JP4MR4KN1LL0YY=kjNhdIb^E^ZA_ZD`[GZUAQO:KJ8AB4@B5@B5>A6>A6>A6=@7=@79?59?59?5:@6:@6:B7:B7;C8;@9;@9;@::?8:?9:?89>89>7;@:;@9;@:;@9:?9:?8:?9:?98?87=94;4382160/4..3/.3/.21.21-2.,2.,3,,3+,3+)6%$6*>"2F+8L3;O6<N6;M5;M58J29K3;M5=O7<N87H51B0.<--7..5--4,]eP^fO_gRbgQdiUfiVgjWijXfgUhgUheVheVjdVkeWlfZlfXlfXkhYmg[kg\mf^jf]ieZgeVghVfkUenYbo]dqht„�–¨¬¸ÇÎÒÛâäéïò÷û÷üÿøüûüþùüÿøüþó÷÷ëùúêø÷åòñßîíÙëêÖáßÊ×Ò¾×ϼËÀ®½´¥½³©Â·±ÌÂÁÛÒ×êàèúóûüöÿþûÿÿýÿþýÿüýÿüþýüþýþþþþþþþþþþþþþþþþþþÿþüÿþüÿþüÿþüþýùþýùÿüùÿýúÿþûÿþûÿþÿÿýþÿüýÿýþÿÿÿþÿýýÿüûýøÿÿûüý÷ÿýøÿþúÿýøÿúöÿ÷ôÿùùÿúÿÿüÿÿýÿÿþÿÿÿÿþþüýÿþüþýÿÿÿÿÿÿÿþüÿþùÿþõÿþðÿýêÿýæúôÜÿúàçàÆ�ˆnUN2WM2TK.SJ+WN1WN1UN1UN1TO2TO2TN4SO4QM4RN5QO6PP6MO7KO6HM6GL5JP6FO4AJ-=F)@G(DK*FM+EL*PV4PS4KN/IL/TT8bbFccIZZ@[Y@\ZA][B][B[Y@XV=US:SQ8HF1GD1EB1B@1A?2>>2??5>@5;>59?59?59A69A48C58C59C8:B79A69@88@58?79A69@8:B7:A9:B7:A9:B7;B:<D9=D<>E=8B:8B:8B98@56>14</39/271.40/51070.8-.;)0@&6F)7M';V+:W+;U0;T4;R8=P<>O<>O<:L6@R8EW=FY=DX=@T;5H2+<*,6+-4,+2*ZfN[gO^hP_iQbjSckTglXhkXhiWghVifWifWkeWlfXmeZmgYmgYkhYkg[kg\je_jf]hfZfgWhiWglVfo\erajwnzŠ‰œ«²ºÈÑÐÙàãèìò÷ûøþþùþúúÿøûþõùýïøùë÷øèõôâíìÚéæÓåâÏÜ×ÄÒ˹ʿ­Á´¤¹­¡¼¯¦Á¶²ËÁÀÛÐÖæßæûóþý÷ÿþûÿÿýÿþýÿýüÿüýÿüýÿþþþþþþþþþþþþþþþþþþÿþüÿþüÿþüÿþüÿýúÿüùÿüùÿýúÿýùÿþúÿþýÿýüÿüýÿýþÿþÿÿþÿÿýþýûüÿþýÿþûÿüúÿøóøíçòåÝöèßûðêÿúýÿüÿÿýÿÿÿýÿÿýþþüýÿþüýÿþýÿþýÿÿýÿÿþüÿý÷ÿÿðÿþèÿÿäúöÝýûâäàÇ‹‡lRM0SN1RK.QK+UO/UO/TO1TO1SO2SO2QP4QO6QM2PO3PN5NQ6MO7JP6HM6EM5HQ6FO4BK.?H)CH(GM+JP,KQ-OS0PT1VZ9^aB_bCZ]@VV<UU=VV>WU>XV=YW>YW>XW;WV:XT9QK3PJ4LG4ID1DA2B@3A?3>@5<=59?59?59A67B48C57C57C5:B7:B79A68@58@59A6:B7:B7:B7:B7:B7:B7;C8<D9=E:<F;=G?<H><F;;F6<D5:B39?38=67>76=63=52>04C,9L,BY/Ga1Qo;Mm;Li=Fa>BY?<R=<M=:K9=O9DV<K^BNaCL`DH\A:Q71D0+5*+2*)0(ZfN[gO\hP`jRbjSdlUglXilYijXijXkhYkhYmgYmgYog\nhZmgYkhYkg[jf[je_jf]hfZfgWhiWejTdmZgtco|s�Ž�¬³¶ÄÍÍÖÝÞæéñöúøþþúþýùþøøûòôøêóôæðñáêé×áàÎÜÙÆ×ÔÁÐ˸ÇÀ®¿³£½° ¼°¤Ã¶®ËÀ¾ÕËÌãØÞëäëüôÿý÷ÿþûÿÿýÿþýÿýüÿüýÿýÿþþþþþþþþþþþþþþþþþþþÿþüÿþüÿþüÿþüÿýúÿýúÿüùÿüùÿüøÿüøÿýüÿüûÿûüÿüýÿþÿÿþÿÿýþþüýÿþýÿþûÿýûüóîêßÙâÕÍëÝÔ÷ìæÿúýÿüÿÿýÿÿÿýÿÿýþþüýÿþüýÿþýÿþýÿÿþÿÿÿýÿþõÿþíÿýçÿþãû÷ÞþüããßƉ…jPK.RM0QJ-QJ-UO/UO/UP2UP2TP3SO2QP4PO3QM2ON2PN5MP5MO7JP6IN7GO7HQ6FO4CL/CL-GL,KQ/OU1SY5QY2X`9jpLy]qvVY^@LO4MP5QQ9QQ9RP9SQ8TR9VT;YU<ZV=XR<VP:RM:NI6HE6DB5CA5?A6<=59?5:@69A67B48C57C57C5:B7:B79A69A69A69A6:B7:B7:B7:B7:B7:B7;C8<D9>F;<F;?IA>J@?I>?J:@H9>F7>D8=B;>E><C<7A95A39H1BU5Pg=XrB`~J\|JXuINiFD[A9O:6G74E3<N8EW=NaESfHSgKNbG@W=6I5/9./6.-4,ZfN[gO\hP^jRblTemVhmYinZklZklZlj[liZnhZnhZph]oi[liZkhYkg[jf[id^ie\hfZghXfgUbgQajWerao|s~Ž�–¥¬«¹ÂÇÐ×Úáçîó÷öüüøüûõúôóöíîòäéêÜãäÔÜÛÉÓÒÀÎ˸ÉƳþ«½¶¤Ã·§Ç¹¬ÍÁµÖÉÁßÔÒèÞßóèîøñøýõÿþøÿÿüÿÿüÿýüÿýýÿüþýýÿþþþþþþþþþþþþþþþþþþþÿþüÿþüþýûÿþüÿýúÿýúÿýúÿüùÿû÷ÿû÷ÿýüÿüûþúûÿûüÿýþÿþÿÿýþþüýÿüûÿþûÿýûýôïíâÜæÙÑðâÙüñëÿúýÿüÿÿýþÿÿýÿÿûþÿúýÿþüýÿþýÿþýÿÿþÿÿÿýÿþõÿþíÿüæÿüãþùãÿþçäßÉŠ†mPJ0RM0SL/SL/UO/UO/UP2VQ3TP3SO2PO3ON2PL1ON2OM4LO4MO7KQ7JO8HP8JS8GP5GP3IR3NS3PV4W]9\d=grHt�U‡’j�šu‡ŽlnuTZaBSW<MO7NN6MM5NL5OM6RP9WR<YT>ZT>XR<TO<QL9KH9GE8EC7AC8=>6:@6:@69A67B47B46B46B4;C8:B7:B79A69A6:B7:B7;C8;C8:B7:B7;C8<D9=E:>F;=G<=G?=I?>H=>I9?G8>F7>D8=B;=D=;B;6@85A3:I2DW7Ri?[uE\zFXxFTqEIdA>U;2H3/@0.?-6H2@R8L_CReGSgKOcH@W=6I51;007/-4,YgNZhO\hP^jRblTemVinZjo[mn\lm[mk\mj[oi[oi[oi]oi[liZkhYjfZieZid^ie\ig[ghXghVchRajWerao|s{‹Š�Ÿ¦¢°¹ÁËÔÔÝäéñôòúüôúúñöòíðçèìÞâãÕÚÛËÒÑ¿ÌɸÉıþ«¿¸¦¼³¢Ê¾®ÓŸÜÐÄæÙÑíâàöìíÿôúÿùÿÿ÷ÿÿúÿÿüÿÿüÿýüÿüüþüþýþÿÿþþþþþþþþþþþþþþþþþþÿþüÿþüþýûþýûÿýúÿýúÿýúÿüùÿúöþùõÿýüÿûúýùúþúûþüýÿýþÿþÿÿýþÿüûÿýúÿüúÿøó÷ìæòåÝøêáÿôîÿúýÿüÿÿýþÿÿýÿÿûþÿúýÿþüýÿýüÿþýÿÿþÿÿÿýÿþõÿýìÿûåþúáÿüæÿÿéçâÌŒˆoRL2TO2TM0UN1UO/UO/VQ3VQ3UQ4SO2ON2NM1PL1NM1NL3KN3LN6KQ7LQ:JR:LU:IR7JS6OX9TY9U[9]c?fnG‚’c’¦sŸ°ƒ›ª” zˆ”pr{\\bFLP7LN6LK6LJ5LJ5QL8TM:VP:WQ;VP:SN;QL9MJ;IG:GE9CE:=>6:@6:@69A67B47B46B46B4;C8;C8:B7:B7:B7:B7;C8;C8;C8;C8;C8;C8<D9=E:>F;=G<=G?<H>=G<=H8?G8=E6=C7=B;:A:9@96@85A39H1BU5Ne;Vp@Vt@SsAPmAE`=9P6-C.+<,+<*1C-<N4H[?PcERfJNbG@W=5H4,6++2*'.&
\ No newline at end of file
diff --git a/transupp.c b/transupp.c
new file mode 100644
index 0000000..5c31416
--- /dev/null
+++ b/transupp.c
@@ -0,0 +1,1613 @@
+/*
+ * transupp.c
+ *
+ * Copyright (C) 1997-2009, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 2010, D. R. Commander.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains image transformation routines and other utility code
+ * used by the jpegtran sample application.  These are NOT part of the core
+ * JPEG library.  But we keep these routines separate from jpegtran.c to
+ * ease the task of maintaining jpegtran-like programs that have other user
+ * interfaces.
+ */
+
+/* Although this file really shouldn't have access to the library internals,
+ * it's helpful to let it call jround_up() and jcopy_block_row().
+ */
+#define JPEG_INTERNALS
+
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "transupp.h"		/* My own external interface */
+#include "jpegcomp.h"
+#include <ctype.h>		/* to declare isdigit() */
+
+
+#if JPEG_LIB_VERSION >= 70
+#define dstinfo_min_DCT_h_scaled_size dstinfo->min_DCT_h_scaled_size
+#define dstinfo_min_DCT_v_scaled_size dstinfo->min_DCT_v_scaled_size
+#else
+#define dstinfo_min_DCT_h_scaled_size DCTSIZE
+#define dstinfo_min_DCT_v_scaled_size DCTSIZE
+#endif
+
+
+#if TRANSFORMS_SUPPORTED
+
+/*
+ * Lossless image transformation routines.  These routines work on DCT
+ * coefficient arrays and thus do not require any lossy decompression
+ * or recompression of the image.
+ * Thanks to Guido Vollbeding for the initial design and code of this feature,
+ * and to Ben Jackson for introducing the cropping feature.
+ *
+ * Horizontal flipping is done in-place, using a single top-to-bottom
+ * pass through the virtual source array.  It will thus be much the
+ * fastest option for images larger than main memory.
+ *
+ * The other routines require a set of destination virtual arrays, so they
+ * need twice as much memory as jpegtran normally does.  The destination
+ * arrays are always written in normal scan order (top to bottom) because
+ * the virtual array manager expects this.  The source arrays will be scanned
+ * in the corresponding order, which means multiple passes through the source
+ * arrays for most of the transforms.  That could result in much thrashing
+ * if the image is larger than main memory.
+ *
+ * If cropping or trimming is involved, the destination arrays may be smaller
+ * than the source arrays.  Note it is not possible to do horizontal flip
+ * in-place when a nonzero Y crop offset is specified, since we'd have to move
+ * data from one block row to another but the virtual array manager doesn't
+ * guarantee we can touch more than one row at a time.  So in that case,
+ * we have to use a separate destination array.
+ *
+ * Some notes about the operating environment of the individual transform
+ * routines:
+ * 1. Both the source and destination virtual arrays are allocated from the
+ *    source JPEG object, and therefore should be manipulated by calling the
+ *    source's memory manager.
+ * 2. The destination's component count should be used.  It may be smaller
+ *    than the source's when forcing to grayscale.
+ * 3. Likewise the destination's sampling factors should be used.  When
+ *    forcing to grayscale the destination's sampling factors will be all 1,
+ *    and we may as well take that as the effective iMCU size.
+ * 4. When "trim" is in effect, the destination's dimensions will be the
+ *    trimmed values but the source's will be untrimmed.
+ * 5. When "crop" is in effect, the destination's dimensions will be the
+ *    cropped values but the source's will be uncropped.  Each transform
+ *    routine is responsible for picking up source data starting at the
+ *    correct X and Y offset for the crop region.  (The X and Y offsets
+ *    passed to the transform routines are measured in iMCU blocks of the
+ *    destination.)
+ * 6. All the routines assume that the source and destination buffers are
+ *    padded out to a full iMCU boundary.  This is true, although for the
+ *    source buffer it is an undocumented property of jdcoefct.c.
+ */
+
+
+LOCAL(void)
+do_crop (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	 JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	 jvirt_barray_ptr *src_coef_arrays,
+	 jvirt_barray_ptr *dst_coef_arrays)
+/* Crop.  This is only used when no rotate/flip is requested with the crop. */
+{
+  JDIMENSION dst_blk_y, x_crop_blocks, y_crop_blocks;
+  int ci, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  jpeg_component_info *compptr;
+
+  /* We simply have to copy the right amount of data (the destination's
+   * image size) starting at the given X and Y offsets in the source.
+   */
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      src_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	 dst_blk_y + y_crop_blocks,
+	 (JDIMENSION) compptr->v_samp_factor, FALSE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	jcopy_block_row(src_buffer[offset_y] + x_crop_blocks,
+			dst_buffer[offset_y],
+			compptr->width_in_blocks);
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_flip_h_no_crop (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+		   JDIMENSION x_crop_offset,
+		   jvirt_barray_ptr *src_coef_arrays)
+/* Horizontal flip; done in-place, so no separate dest array is required.
+ * NB: this only works when y_crop_offset is zero.
+ */
+{
+  JDIMENSION MCU_cols, comp_width, blk_x, blk_y, x_crop_blocks;
+  int ci, k, offset_y;
+  JBLOCKARRAY buffer;
+  JCOEFPTR ptr1, ptr2;
+  JCOEF temp1, temp2;
+  jpeg_component_info *compptr;
+
+  /* Horizontal mirroring of DCT blocks is accomplished by swapping
+   * pairs of blocks in-place.  Within a DCT block, we perform horizontal
+   * mirroring by changing the signs of odd-numbered columns.
+   * Partial iMCUs at the right edge are left untouched.
+   */
+  MCU_cols = srcinfo->output_width /
+    (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_width = MCU_cols * compptr->h_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    for (blk_y = 0; blk_y < compptr->height_in_blocks;
+	 blk_y += compptr->v_samp_factor) {
+      buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, src_coef_arrays[ci], blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	/* Do the mirroring */
+	for (blk_x = 0; blk_x * 2 < comp_width; blk_x++) {
+	  ptr1 = buffer[offset_y][blk_x];
+	  ptr2 = buffer[offset_y][comp_width - blk_x - 1];
+	  /* this unrolled loop doesn't need to know which row it's on... */
+	  for (k = 0; k < DCTSIZE2; k += 2) {
+	    temp1 = *ptr1;	/* swap even column */
+	    temp2 = *ptr2;
+	    *ptr1++ = temp2;
+	    *ptr2++ = temp1;
+	    temp1 = *ptr1;	/* swap odd column with sign change */
+	    temp2 = *ptr2;
+	    *ptr1++ = -temp2;
+	    *ptr2++ = -temp1;
+	  }
+	}
+	if (x_crop_blocks > 0) {
+	  /* Now left-justify the portion of the data to be kept.
+	   * We can't use a single jcopy_block_row() call because that routine
+	   * depends on memcpy(), whose behavior is unspecified for overlapping
+	   * source and destination areas.  Sigh.
+	   */
+	  for (blk_x = 0; blk_x < compptr->width_in_blocks; blk_x++) {
+	    jcopy_block_row(buffer[offset_y] + blk_x + x_crop_blocks,
+			    buffer[offset_y] + blk_x,
+			    (JDIMENSION) 1);
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_flip_h (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	   JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	   jvirt_barray_ptr *src_coef_arrays,
+	   jvirt_barray_ptr *dst_coef_arrays)
+/* Horizontal flip in general cropping case */
+{
+  JDIMENSION MCU_cols, comp_width, dst_blk_x, dst_blk_y;
+  JDIMENSION x_crop_blocks, y_crop_blocks;
+  int ci, k, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JBLOCKROW src_row_ptr, dst_row_ptr;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  /* Here we must output into a separate array because we can't touch
+   * different rows of a single virtual array simultaneously.  Otherwise,
+   * this is essentially the same as the routine above.
+   */
+  MCU_cols = srcinfo->output_width /
+    (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_width = MCU_cols * compptr->h_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      src_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	 dst_blk_y + y_crop_blocks,
+	 (JDIMENSION) compptr->v_samp_factor, FALSE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	dst_row_ptr = dst_buffer[offset_y];
+	src_row_ptr = src_buffer[offset_y];
+	for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks; dst_blk_x++) {
+	  if (x_crop_blocks + dst_blk_x < comp_width) {
+	    /* Do the mirrorable blocks */
+	    dst_ptr = dst_row_ptr[dst_blk_x];
+	    src_ptr = src_row_ptr[comp_width - x_crop_blocks - dst_blk_x - 1];
+	    /* this unrolled loop doesn't need to know which row it's on... */
+	    for (k = 0; k < DCTSIZE2; k += 2) {
+	      *dst_ptr++ = *src_ptr++;	 /* copy even column */
+	      *dst_ptr++ = - *src_ptr++; /* copy odd column with sign change */
+	    }
+	  } else {
+	    /* Copy last partial block(s) verbatim */
+	    jcopy_block_row(src_row_ptr + dst_blk_x + x_crop_blocks,
+			    dst_row_ptr + dst_blk_x,
+			    (JDIMENSION) 1);
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_flip_v (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	   JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	   jvirt_barray_ptr *src_coef_arrays,
+	   jvirt_barray_ptr *dst_coef_arrays)
+/* Vertical flip */
+{
+  JDIMENSION MCU_rows, comp_height, dst_blk_x, dst_blk_y;
+  JDIMENSION x_crop_blocks, y_crop_blocks;
+  int ci, i, j, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JBLOCKROW src_row_ptr, dst_row_ptr;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  /* We output into a separate array because we can't touch different
+   * rows of the source virtual array simultaneously.  Otherwise, this
+   * is a pretty straightforward analog of horizontal flip.
+   * Within a DCT block, vertical mirroring is done by changing the signs
+   * of odd-numbered rows.
+   * Partial iMCUs at the bottom edge are copied verbatim.
+   */
+  MCU_rows = srcinfo->output_height /
+    (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_height = MCU_rows * compptr->v_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      if (y_crop_blocks + dst_blk_y < comp_height) {
+	/* Row is within the mirrorable area. */
+	src_buffer = (*srcinfo->mem->access_virt_barray)
+	  ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	   comp_height - y_crop_blocks - dst_blk_y -
+	   (JDIMENSION) compptr->v_samp_factor,
+	   (JDIMENSION) compptr->v_samp_factor, FALSE);
+      } else {
+	/* Bottom-edge blocks will be copied verbatim. */
+	src_buffer = (*srcinfo->mem->access_virt_barray)
+	  ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	   dst_blk_y + y_crop_blocks,
+	   (JDIMENSION) compptr->v_samp_factor, FALSE);
+      }
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	if (y_crop_blocks + dst_blk_y < comp_height) {
+	  /* Row is within the mirrorable area. */
+	  dst_row_ptr = dst_buffer[offset_y];
+	  src_row_ptr = src_buffer[compptr->v_samp_factor - offset_y - 1];
+	  src_row_ptr += x_crop_blocks;
+	  for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks;
+	       dst_blk_x++) {
+	    dst_ptr = dst_row_ptr[dst_blk_x];
+	    src_ptr = src_row_ptr[dst_blk_x];
+	    for (i = 0; i < DCTSIZE; i += 2) {
+	      /* copy even row */
+	      for (j = 0; j < DCTSIZE; j++)
+		*dst_ptr++ = *src_ptr++;
+	      /* copy odd row with sign change */
+	      for (j = 0; j < DCTSIZE; j++)
+		*dst_ptr++ = - *src_ptr++;
+	    }
+	  }
+	} else {
+	  /* Just copy row verbatim. */
+	  jcopy_block_row(src_buffer[offset_y] + x_crop_blocks,
+			  dst_buffer[offset_y],
+			  compptr->width_in_blocks);
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_transpose (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	      JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	      jvirt_barray_ptr *src_coef_arrays,
+	      jvirt_barray_ptr *dst_coef_arrays)
+/* Transpose source into destination */
+{
+  JDIMENSION dst_blk_x, dst_blk_y, x_crop_blocks, y_crop_blocks;
+  int ci, i, j, offset_x, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  /* Transposing pixels within a block just requires transposing the
+   * DCT coefficients.
+   * Partial iMCUs at the edges require no special treatment; we simply
+   * process all the available DCT blocks for every component.
+   */
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks;
+	     dst_blk_x += compptr->h_samp_factor) {
+	  src_buffer = (*srcinfo->mem->access_virt_barray)
+	    ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	     dst_blk_x + x_crop_blocks,
+	     (JDIMENSION) compptr->h_samp_factor, FALSE);
+	  for (offset_x = 0; offset_x < compptr->h_samp_factor; offset_x++) {
+	    dst_ptr = dst_buffer[offset_y][dst_blk_x + offset_x];
+	    src_ptr = src_buffer[offset_x][dst_blk_y + offset_y + y_crop_blocks];
+	    for (i = 0; i < DCTSIZE; i++)
+	      for (j = 0; j < DCTSIZE; j++)
+		dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_rot_90 (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	   JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	   jvirt_barray_ptr *src_coef_arrays,
+	   jvirt_barray_ptr *dst_coef_arrays)
+/* 90 degree rotation is equivalent to
+ *   1. Transposing the image;
+ *   2. Horizontal mirroring.
+ * These two steps are merged into a single processing routine.
+ */
+{
+  JDIMENSION MCU_cols, comp_width, dst_blk_x, dst_blk_y;
+  JDIMENSION x_crop_blocks, y_crop_blocks;
+  int ci, i, j, offset_x, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  /* Because of the horizontal mirror step, we can't process partial iMCUs
+   * at the (output) right edge properly.  They just get transposed and
+   * not mirrored.
+   */
+  MCU_cols = srcinfo->output_height /
+    (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_width = MCU_cols * compptr->h_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks;
+	     dst_blk_x += compptr->h_samp_factor) {
+	  if (x_crop_blocks + dst_blk_x < comp_width) {
+	    /* Block is within the mirrorable area. */
+	    src_buffer = (*srcinfo->mem->access_virt_barray)
+	      ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	       comp_width - x_crop_blocks - dst_blk_x -
+	       (JDIMENSION) compptr->h_samp_factor,
+	       (JDIMENSION) compptr->h_samp_factor, FALSE);
+	  } else {
+	    /* Edge blocks are transposed but not mirrored. */
+	    src_buffer = (*srcinfo->mem->access_virt_barray)
+	      ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	       dst_blk_x + x_crop_blocks,
+	       (JDIMENSION) compptr->h_samp_factor, FALSE);
+	  }
+	  for (offset_x = 0; offset_x < compptr->h_samp_factor; offset_x++) {
+	    dst_ptr = dst_buffer[offset_y][dst_blk_x + offset_x];
+	    if (x_crop_blocks + dst_blk_x < comp_width) {
+	      /* Block is within the mirrorable area. */
+	      src_ptr = src_buffer[compptr->h_samp_factor - offset_x - 1]
+		[dst_blk_y + offset_y + y_crop_blocks];
+	      for (i = 0; i < DCTSIZE; i++) {
+		for (j = 0; j < DCTSIZE; j++)
+		  dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+		i++;
+		for (j = 0; j < DCTSIZE; j++)
+		  dst_ptr[j*DCTSIZE+i] = -src_ptr[i*DCTSIZE+j];
+	      }
+	    } else {
+	      /* Edge blocks are transposed but not mirrored. */
+	      src_ptr = src_buffer[offset_x]
+		[dst_blk_y + offset_y + y_crop_blocks];
+	      for (i = 0; i < DCTSIZE; i++)
+		for (j = 0; j < DCTSIZE; j++)
+		  dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+	    }
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_rot_270 (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	    JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	    jvirt_barray_ptr *src_coef_arrays,
+	    jvirt_barray_ptr *dst_coef_arrays)
+/* 270 degree rotation is equivalent to
+ *   1. Horizontal mirroring;
+ *   2. Transposing the image.
+ * These two steps are merged into a single processing routine.
+ */
+{
+  JDIMENSION MCU_rows, comp_height, dst_blk_x, dst_blk_y;
+  JDIMENSION x_crop_blocks, y_crop_blocks;
+  int ci, i, j, offset_x, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  /* Because of the horizontal mirror step, we can't process partial iMCUs
+   * at the (output) bottom edge properly.  They just get transposed and
+   * not mirrored.
+   */
+  MCU_rows = srcinfo->output_width /
+    (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_height = MCU_rows * compptr->v_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks;
+	     dst_blk_x += compptr->h_samp_factor) {
+	  src_buffer = (*srcinfo->mem->access_virt_barray)
+	    ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	     dst_blk_x + x_crop_blocks,
+	     (JDIMENSION) compptr->h_samp_factor, FALSE);
+	  for (offset_x = 0; offset_x < compptr->h_samp_factor; offset_x++) {
+	    dst_ptr = dst_buffer[offset_y][dst_blk_x + offset_x];
+	    if (y_crop_blocks + dst_blk_y < comp_height) {
+	      /* Block is within the mirrorable area. */
+	      src_ptr = src_buffer[offset_x]
+		[comp_height - y_crop_blocks - dst_blk_y - offset_y - 1];
+	      for (i = 0; i < DCTSIZE; i++) {
+		for (j = 0; j < DCTSIZE; j++) {
+		  dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+		  j++;
+		  dst_ptr[j*DCTSIZE+i] = -src_ptr[i*DCTSIZE+j];
+		}
+	      }
+	    } else {
+	      /* Edge blocks are transposed but not mirrored. */
+	      src_ptr = src_buffer[offset_x]
+		[dst_blk_y + offset_y + y_crop_blocks];
+	      for (i = 0; i < DCTSIZE; i++)
+		for (j = 0; j < DCTSIZE; j++)
+		  dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+	    }
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_rot_180 (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	    JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	    jvirt_barray_ptr *src_coef_arrays,
+	    jvirt_barray_ptr *dst_coef_arrays)
+/* 180 degree rotation is equivalent to
+ *   1. Vertical mirroring;
+ *   2. Horizontal mirroring.
+ * These two steps are merged into a single processing routine.
+ */
+{
+  JDIMENSION MCU_cols, MCU_rows, comp_width, comp_height, dst_blk_x, dst_blk_y;
+  JDIMENSION x_crop_blocks, y_crop_blocks;
+  int ci, i, j, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JBLOCKROW src_row_ptr, dst_row_ptr;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  MCU_cols = srcinfo->output_width /
+    (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+  MCU_rows = srcinfo->output_height /
+    (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_width = MCU_cols * compptr->h_samp_factor;
+    comp_height = MCU_rows * compptr->v_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      if (y_crop_blocks + dst_blk_y < comp_height) {
+	/* Row is within the vertically mirrorable area. */
+	src_buffer = (*srcinfo->mem->access_virt_barray)
+	  ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	   comp_height - y_crop_blocks - dst_blk_y -
+	   (JDIMENSION) compptr->v_samp_factor,
+	   (JDIMENSION) compptr->v_samp_factor, FALSE);
+      } else {
+	/* Bottom-edge rows are only mirrored horizontally. */
+	src_buffer = (*srcinfo->mem->access_virt_barray)
+	  ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	   dst_blk_y + y_crop_blocks,
+	   (JDIMENSION) compptr->v_samp_factor, FALSE);
+      }
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	dst_row_ptr = dst_buffer[offset_y];
+	if (y_crop_blocks + dst_blk_y < comp_height) {
+	  /* Row is within the mirrorable area. */
+	  src_row_ptr = src_buffer[compptr->v_samp_factor - offset_y - 1];
+	  for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks; dst_blk_x++) {
+	    dst_ptr = dst_row_ptr[dst_blk_x];
+	    if (x_crop_blocks + dst_blk_x < comp_width) {
+	      /* Process the blocks that can be mirrored both ways. */
+	      src_ptr = src_row_ptr[comp_width - x_crop_blocks - dst_blk_x - 1];
+	      for (i = 0; i < DCTSIZE; i += 2) {
+		/* For even row, negate every odd column. */
+		for (j = 0; j < DCTSIZE; j += 2) {
+		  *dst_ptr++ = *src_ptr++;
+		  *dst_ptr++ = - *src_ptr++;
+		}
+		/* For odd row, negate every even column. */
+		for (j = 0; j < DCTSIZE; j += 2) {
+		  *dst_ptr++ = - *src_ptr++;
+		  *dst_ptr++ = *src_ptr++;
+		}
+	      }
+	    } else {
+	      /* Any remaining right-edge blocks are only mirrored vertically. */
+	      src_ptr = src_row_ptr[x_crop_blocks + dst_blk_x];
+	      for (i = 0; i < DCTSIZE; i += 2) {
+		for (j = 0; j < DCTSIZE; j++)
+		  *dst_ptr++ = *src_ptr++;
+		for (j = 0; j < DCTSIZE; j++)
+		  *dst_ptr++ = - *src_ptr++;
+	      }
+	    }
+	  }
+	} else {
+	  /* Remaining rows are just mirrored horizontally. */
+	  src_row_ptr = src_buffer[offset_y];
+	  for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks; dst_blk_x++) {
+	    if (x_crop_blocks + dst_blk_x < comp_width) {
+	      /* Process the blocks that can be mirrored. */
+	      dst_ptr = dst_row_ptr[dst_blk_x];
+	      src_ptr = src_row_ptr[comp_width - x_crop_blocks - dst_blk_x - 1];
+	      for (i = 0; i < DCTSIZE2; i += 2) {
+		*dst_ptr++ = *src_ptr++;
+		*dst_ptr++ = - *src_ptr++;
+	      }
+	    } else {
+	      /* Any remaining right-edge blocks are only copied. */
+	      jcopy_block_row(src_row_ptr + dst_blk_x + x_crop_blocks,
+			      dst_row_ptr + dst_blk_x,
+			      (JDIMENSION) 1);
+	    }
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+LOCAL(void)
+do_transverse (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	       JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+	       jvirt_barray_ptr *src_coef_arrays,
+	       jvirt_barray_ptr *dst_coef_arrays)
+/* Transverse transpose is equivalent to
+ *   1. 180 degree rotation;
+ *   2. Transposition;
+ * or
+ *   1. Horizontal mirroring;
+ *   2. Transposition;
+ *   3. Horizontal mirroring.
+ * These steps are merged into a single processing routine.
+ */
+{
+  JDIMENSION MCU_cols, MCU_rows, comp_width, comp_height, dst_blk_x, dst_blk_y;
+  JDIMENSION x_crop_blocks, y_crop_blocks;
+  int ci, i, j, offset_x, offset_y;
+  JBLOCKARRAY src_buffer, dst_buffer;
+  JCOEFPTR src_ptr, dst_ptr;
+  jpeg_component_info *compptr;
+
+  MCU_cols = srcinfo->output_height /
+    (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+  MCU_rows = srcinfo->output_width /
+    (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    comp_width = MCU_cols * compptr->h_samp_factor;
+    comp_height = MCU_rows * compptr->v_samp_factor;
+    x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+    y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+    for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+	 dst_blk_y += compptr->v_samp_factor) {
+      dst_buffer = (*srcinfo->mem->access_virt_barray)
+	((j_common_ptr) srcinfo, dst_coef_arrays[ci], dst_blk_y,
+	 (JDIMENSION) compptr->v_samp_factor, TRUE);
+      for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+	for (dst_blk_x = 0; dst_blk_x < compptr->width_in_blocks;
+	     dst_blk_x += compptr->h_samp_factor) {
+	  if (x_crop_blocks + dst_blk_x < comp_width) {
+	    /* Block is within the mirrorable area. */
+	    src_buffer = (*srcinfo->mem->access_virt_barray)
+	      ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	       comp_width - x_crop_blocks - dst_blk_x -
+	       (JDIMENSION) compptr->h_samp_factor,
+	       (JDIMENSION) compptr->h_samp_factor, FALSE);
+	  } else {
+	    src_buffer = (*srcinfo->mem->access_virt_barray)
+	      ((j_common_ptr) srcinfo, src_coef_arrays[ci],
+	       dst_blk_x + x_crop_blocks,
+	       (JDIMENSION) compptr->h_samp_factor, FALSE);
+	  }
+	  for (offset_x = 0; offset_x < compptr->h_samp_factor; offset_x++) {
+	    dst_ptr = dst_buffer[offset_y][dst_blk_x + offset_x];
+	    if (y_crop_blocks + dst_blk_y < comp_height) {
+	      if (x_crop_blocks + dst_blk_x < comp_width) {
+		/* Block is within the mirrorable area. */
+		src_ptr = src_buffer[compptr->h_samp_factor - offset_x - 1]
+		  [comp_height - y_crop_blocks - dst_blk_y - offset_y - 1];
+		for (i = 0; i < DCTSIZE; i++) {
+		  for (j = 0; j < DCTSIZE; j++) {
+		    dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+		    j++;
+		    dst_ptr[j*DCTSIZE+i] = -src_ptr[i*DCTSIZE+j];
+		  }
+		  i++;
+		  for (j = 0; j < DCTSIZE; j++) {
+		    dst_ptr[j*DCTSIZE+i] = -src_ptr[i*DCTSIZE+j];
+		    j++;
+		    dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+		  }
+		}
+	      } else {
+		/* Right-edge blocks are mirrored in y only */
+		src_ptr = src_buffer[offset_x]
+		  [comp_height - y_crop_blocks - dst_blk_y - offset_y - 1];
+		for (i = 0; i < DCTSIZE; i++) {
+		  for (j = 0; j < DCTSIZE; j++) {
+		    dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+		    j++;
+		    dst_ptr[j*DCTSIZE+i] = -src_ptr[i*DCTSIZE+j];
+		  }
+		}
+	      }
+	    } else {
+	      if (x_crop_blocks + dst_blk_x < comp_width) {
+		/* Bottom-edge blocks are mirrored in x only */
+		src_ptr = src_buffer[compptr->h_samp_factor - offset_x - 1]
+		  [dst_blk_y + offset_y + y_crop_blocks];
+		for (i = 0; i < DCTSIZE; i++) {
+		  for (j = 0; j < DCTSIZE; j++)
+		    dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+		  i++;
+		  for (j = 0; j < DCTSIZE; j++)
+		    dst_ptr[j*DCTSIZE+i] = -src_ptr[i*DCTSIZE+j];
+		}
+	      } else {
+		/* At lower right corner, just transpose, no mirroring */
+		src_ptr = src_buffer[offset_x]
+		  [dst_blk_y + offset_y + y_crop_blocks];
+		for (i = 0; i < DCTSIZE; i++)
+		  for (j = 0; j < DCTSIZE; j++)
+		    dst_ptr[j*DCTSIZE+i] = src_ptr[i*DCTSIZE+j];
+	      }
+	    }
+	  }
+	}
+      }
+    }
+  }
+}
+
+
+/* Parse an unsigned integer: subroutine for jtransform_parse_crop_spec.
+ * Returns TRUE if valid integer found, FALSE if not.
+ * *strptr is advanced over the digit string, and *result is set to its value.
+ */
+
+LOCAL(boolean)
+jt_read_integer (const char ** strptr, JDIMENSION * result)
+{
+  const char * ptr = *strptr;
+  JDIMENSION val = 0;
+
+  for (; isdigit(*ptr); ptr++) {
+    val = val * 10 + (JDIMENSION) (*ptr - '0');
+  }
+  *result = val;
+  if (ptr == *strptr)
+    return FALSE;		/* oops, no digits */
+  *strptr = ptr;
+  return TRUE;
+}
+
+
+/* Parse a crop specification (written in X11 geometry style).
+ * The routine returns TRUE if the spec string is valid, FALSE if not.
+ *
+ * The crop spec string should have the format
+ *	<width>x<height>{+-}<xoffset>{+-}<yoffset>
+ * where width, height, xoffset, and yoffset are unsigned integers.
+ * Each of the elements can be omitted to indicate a default value.
+ * (A weakness of this style is that it is not possible to omit xoffset
+ * while specifying yoffset, since they look alike.)
+ *
+ * This code is loosely based on XParseGeometry from the X11 distribution.
+ */
+
+GLOBAL(boolean)
+jtransform_parse_crop_spec (jpeg_transform_info *info, const char *spec)
+{
+  info->crop = FALSE;
+  info->crop_width_set = JCROP_UNSET;
+  info->crop_height_set = JCROP_UNSET;
+  info->crop_xoffset_set = JCROP_UNSET;
+  info->crop_yoffset_set = JCROP_UNSET;
+
+  if (isdigit(*spec)) {
+    /* fetch width */
+    if (! jt_read_integer(&spec, &info->crop_width))
+      return FALSE;
+    info->crop_width_set = JCROP_POS;
+  }
+  if (*spec == 'x' || *spec == 'X') {	
+    /* fetch height */
+    spec++;
+    if (! jt_read_integer(&spec, &info->crop_height))
+      return FALSE;
+    info->crop_height_set = JCROP_POS;
+  }
+  if (*spec == '+' || *spec == '-') {
+    /* fetch xoffset */
+    info->crop_xoffset_set = (*spec == '-') ? JCROP_NEG : JCROP_POS;
+    spec++;
+    if (! jt_read_integer(&spec, &info->crop_xoffset))
+      return FALSE;
+  }
+  if (*spec == '+' || *spec == '-') {
+    /* fetch yoffset */
+    info->crop_yoffset_set = (*spec == '-') ? JCROP_NEG : JCROP_POS;
+    spec++;
+    if (! jt_read_integer(&spec, &info->crop_yoffset))
+      return FALSE;
+  }
+  /* We had better have gotten to the end of the string. */
+  if (*spec != '\0')
+    return FALSE;
+  info->crop = TRUE;
+  return TRUE;
+}
+
+
+/* Trim off any partial iMCUs on the indicated destination edge */
+
+LOCAL(void)
+trim_right_edge (jpeg_transform_info *info, JDIMENSION full_width)
+{
+  JDIMENSION MCU_cols;
+
+  MCU_cols = info->output_width / info->iMCU_sample_width;
+  if (MCU_cols > 0 && info->x_crop_offset + MCU_cols ==
+      full_width / info->iMCU_sample_width)
+    info->output_width = MCU_cols * info->iMCU_sample_width;
+}
+
+LOCAL(void)
+trim_bottom_edge (jpeg_transform_info *info, JDIMENSION full_height)
+{
+  JDIMENSION MCU_rows;
+
+  MCU_rows = info->output_height / info->iMCU_sample_height;
+  if (MCU_rows > 0 && info->y_crop_offset + MCU_rows ==
+      full_height / info->iMCU_sample_height)
+    info->output_height = MCU_rows * info->iMCU_sample_height;
+}
+
+
+/* Request any required workspace.
+ *
+ * This routine figures out the size that the output image will be
+ * (which implies that all the transform parameters must be set before
+ * it is called).
+ *
+ * We allocate the workspace virtual arrays from the source decompression
+ * object, so that all the arrays (both the original data and the workspace)
+ * will be taken into account while making memory management decisions.
+ * Hence, this routine must be called after jpeg_read_header (which reads
+ * the image dimensions) and before jpeg_read_coefficients (which realizes
+ * the source's virtual arrays).
+ *
+ * This function returns FALSE right away if -perfect is given
+ * and transformation is not perfect.  Otherwise returns TRUE.
+ */
+
+GLOBAL(boolean)
+jtransform_request_workspace (j_decompress_ptr srcinfo,
+			      jpeg_transform_info *info)
+{
+  jvirt_barray_ptr *coef_arrays;
+  boolean need_workspace, transpose_it;
+  jpeg_component_info *compptr;
+  JDIMENSION xoffset, yoffset;
+  JDIMENSION width_in_iMCUs, height_in_iMCUs;
+  JDIMENSION width_in_blocks, height_in_blocks;
+  int ci, h_samp_factor, v_samp_factor;
+
+  /* Determine number of components in output image */
+  if (info->force_grayscale &&
+      srcinfo->jpeg_color_space == JCS_YCbCr &&
+      srcinfo->num_components == 3)
+    /* We'll only process the first component */
+    info->num_components = 1;
+  else
+    /* Process all the components */
+    info->num_components = srcinfo->num_components;
+
+  /* Compute output image dimensions and related values. */
+#if JPEG_LIB_VERSION >= 80
+  jpeg_core_output_dimensions(srcinfo);
+#else
+  srcinfo->output_width = srcinfo->image_width;
+  srcinfo->output_height = srcinfo->image_height;
+#endif
+
+  /* Return right away if -perfect is given and transformation is not perfect.
+   */
+  if (info->perfect) {
+    if (info->num_components == 1) {
+      if (!jtransform_perfect_transform(srcinfo->output_width,
+	  srcinfo->output_height,
+	  srcinfo->_min_DCT_h_scaled_size,
+	  srcinfo->_min_DCT_v_scaled_size,
+	  info->transform))
+	return FALSE;
+    } else {
+      if (!jtransform_perfect_transform(srcinfo->output_width,
+	  srcinfo->output_height,
+	  srcinfo->max_h_samp_factor * srcinfo->_min_DCT_h_scaled_size,
+	  srcinfo->max_v_samp_factor * srcinfo->_min_DCT_v_scaled_size,
+	  info->transform))
+	return FALSE;
+    }
+  }
+
+  /* If there is only one output component, force the iMCU size to be 1;
+   * else use the source iMCU size.  (This allows us to do the right thing
+   * when reducing color to grayscale, and also provides a handy way of
+   * cleaning up "funny" grayscale images whose sampling factors are not 1x1.)
+   */
+  switch (info->transform) {
+  case JXFORM_TRANSPOSE:
+  case JXFORM_TRANSVERSE:
+  case JXFORM_ROT_90:
+  case JXFORM_ROT_270:
+    info->output_width = srcinfo->output_height;
+    info->output_height = srcinfo->output_width;
+    if (info->num_components == 1) {
+      info->iMCU_sample_width = srcinfo->_min_DCT_v_scaled_size;
+      info->iMCU_sample_height = srcinfo->_min_DCT_h_scaled_size;
+    } else {
+      info->iMCU_sample_width =
+	srcinfo->max_v_samp_factor * srcinfo->_min_DCT_v_scaled_size;
+      info->iMCU_sample_height =
+	srcinfo->max_h_samp_factor * srcinfo->_min_DCT_h_scaled_size;
+    }
+    break;
+  default:
+    info->output_width = srcinfo->output_width;
+    info->output_height = srcinfo->output_height;
+    if (info->num_components == 1) {
+      info->iMCU_sample_width = srcinfo->_min_DCT_h_scaled_size;
+      info->iMCU_sample_height = srcinfo->_min_DCT_v_scaled_size;
+    } else {
+      info->iMCU_sample_width =
+	srcinfo->max_h_samp_factor * srcinfo->_min_DCT_h_scaled_size;
+      info->iMCU_sample_height =
+	srcinfo->max_v_samp_factor * srcinfo->_min_DCT_v_scaled_size;
+    }
+    break;
+  }
+
+  /* If cropping has been requested, compute the crop area's position and
+   * dimensions, ensuring that its upper left corner falls at an iMCU boundary.
+   */
+  if (info->crop) {
+    /* Insert default values for unset crop parameters */
+    if (info->crop_xoffset_set == JCROP_UNSET)
+      info->crop_xoffset = 0;	/* default to +0 */
+    if (info->crop_yoffset_set == JCROP_UNSET)
+      info->crop_yoffset = 0;	/* default to +0 */
+    if (info->crop_xoffset >= info->output_width ||
+	info->crop_yoffset >= info->output_height)
+      ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+    if (info->crop_width_set == JCROP_UNSET)
+      info->crop_width = info->output_width - info->crop_xoffset;
+    if (info->crop_height_set == JCROP_UNSET)
+      info->crop_height = info->output_height - info->crop_yoffset;
+    /* Ensure parameters are valid */
+    if (info->crop_width <= 0 || info->crop_width > info->output_width ||
+	info->crop_height <= 0 || info->crop_height > info->output_height ||
+	info->crop_xoffset > info->output_width - info->crop_width ||
+	info->crop_yoffset > info->output_height - info->crop_height)
+      ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+    /* Convert negative crop offsets into regular offsets */
+    if (info->crop_xoffset_set == JCROP_NEG)
+      xoffset = info->output_width - info->crop_width - info->crop_xoffset;
+    else
+      xoffset = info->crop_xoffset;
+    if (info->crop_yoffset_set == JCROP_NEG)
+      yoffset = info->output_height - info->crop_height - info->crop_yoffset;
+    else
+      yoffset = info->crop_yoffset;
+    /* Now adjust so that upper left corner falls at an iMCU boundary */
+    info->output_width =
+      info->crop_width + (xoffset % info->iMCU_sample_width);
+    info->output_height =
+      info->crop_height + (yoffset % info->iMCU_sample_height);
+    /* Save x/y offsets measured in iMCUs */
+    info->x_crop_offset = xoffset / info->iMCU_sample_width;
+    info->y_crop_offset = yoffset / info->iMCU_sample_height;
+  } else {
+    info->x_crop_offset = 0;
+    info->y_crop_offset = 0;
+  }
+
+  /* Figure out whether we need workspace arrays,
+   * and if so whether they are transposed relative to the source.
+   */
+  need_workspace = FALSE;
+  transpose_it = FALSE;
+  switch (info->transform) {
+  case JXFORM_NONE:
+    if (info->x_crop_offset != 0 || info->y_crop_offset != 0)
+      need_workspace = TRUE;
+    /* No workspace needed if neither cropping nor transforming */
+    break;
+  case JXFORM_FLIP_H:
+    if (info->trim)
+      trim_right_edge(info, srcinfo->output_width);
+    if (info->y_crop_offset != 0)
+      need_workspace = TRUE;
+    /* do_flip_h_no_crop doesn't need a workspace array */
+    break;
+  case JXFORM_FLIP_V:
+    if (info->trim)
+      trim_bottom_edge(info, srcinfo->output_height);
+    /* Need workspace arrays having same dimensions as source image. */
+    need_workspace = TRUE;
+    break;
+  case JXFORM_TRANSPOSE:
+    /* transpose does NOT have to trim anything */
+    /* Need workspace arrays having transposed dimensions. */
+    need_workspace = TRUE;
+    transpose_it = TRUE;
+    break;
+  case JXFORM_TRANSVERSE:
+    if (info->trim) {
+      trim_right_edge(info, srcinfo->output_height);
+      trim_bottom_edge(info, srcinfo->output_width);
+    }
+    /* Need workspace arrays having transposed dimensions. */
+    need_workspace = TRUE;
+    transpose_it = TRUE;
+    break;
+  case JXFORM_ROT_90:
+    if (info->trim)
+      trim_right_edge(info, srcinfo->output_height);
+    /* Need workspace arrays having transposed dimensions. */
+    need_workspace = TRUE;
+    transpose_it = TRUE;
+    break;
+  case JXFORM_ROT_180:
+    if (info->trim) {
+      trim_right_edge(info, srcinfo->output_width);
+      trim_bottom_edge(info, srcinfo->output_height);
+    }
+    /* Need workspace arrays having same dimensions as source image. */
+    need_workspace = TRUE;
+    break;
+  case JXFORM_ROT_270:
+    if (info->trim)
+      trim_bottom_edge(info, srcinfo->output_width);
+    /* Need workspace arrays having transposed dimensions. */
+    need_workspace = TRUE;
+    transpose_it = TRUE;
+    break;
+  }
+
+  /* Allocate workspace if needed.
+   * Note that we allocate arrays padded out to the next iMCU boundary,
+   * so that transform routines need not worry about missing edge blocks.
+   */
+  if (need_workspace) {
+    coef_arrays = (jvirt_barray_ptr *)
+      (*srcinfo->mem->alloc_small) ((j_common_ptr) srcinfo, JPOOL_IMAGE,
+		SIZEOF(jvirt_barray_ptr) * info->num_components);
+    width_in_iMCUs = (JDIMENSION)
+      jdiv_round_up((long) info->output_width,
+		    (long) info->iMCU_sample_width);
+    height_in_iMCUs = (JDIMENSION)
+      jdiv_round_up((long) info->output_height,
+		    (long) info->iMCU_sample_height);
+    for (ci = 0; ci < info->num_components; ci++) {
+      compptr = srcinfo->comp_info + ci;
+      if (info->num_components == 1) {
+	/* we're going to force samp factors to 1x1 in this case */
+	h_samp_factor = v_samp_factor = 1;
+      } else if (transpose_it) {
+	h_samp_factor = compptr->v_samp_factor;
+	v_samp_factor = compptr->h_samp_factor;
+      } else {
+	h_samp_factor = compptr->h_samp_factor;
+	v_samp_factor = compptr->v_samp_factor;
+      }
+      width_in_blocks = width_in_iMCUs * h_samp_factor;
+      height_in_blocks = height_in_iMCUs * v_samp_factor;
+      coef_arrays[ci] = (*srcinfo->mem->request_virt_barray)
+	((j_common_ptr) srcinfo, JPOOL_IMAGE, FALSE,
+	 width_in_blocks, height_in_blocks, (JDIMENSION) v_samp_factor);
+    }
+    info->workspace_coef_arrays = coef_arrays;
+  } else
+    info->workspace_coef_arrays = NULL;
+
+  return TRUE;
+}
+
+
+/* Transpose destination image parameters */
+
+LOCAL(void)
+transpose_critical_parameters (j_compress_ptr dstinfo)
+{
+  int tblno, i, j, ci, itemp;
+  jpeg_component_info *compptr;
+  JQUANT_TBL *qtblptr;
+  JDIMENSION jtemp;
+  UINT16 qtemp;
+
+  /* Transpose image dimensions */
+  jtemp = dstinfo->image_width;
+  dstinfo->image_width = dstinfo->image_height;
+  dstinfo->image_height = jtemp;
+#if JPEG_LIB_VERSION >= 70
+  itemp = dstinfo->min_DCT_h_scaled_size;
+  dstinfo->min_DCT_h_scaled_size = dstinfo->min_DCT_v_scaled_size;
+  dstinfo->min_DCT_v_scaled_size = itemp;
+#endif
+
+  /* Transpose sampling factors */
+  for (ci = 0; ci < dstinfo->num_components; ci++) {
+    compptr = dstinfo->comp_info + ci;
+    itemp = compptr->h_samp_factor;
+    compptr->h_samp_factor = compptr->v_samp_factor;
+    compptr->v_samp_factor = itemp;
+  }
+
+  /* Transpose quantization tables */
+  for (tblno = 0; tblno < NUM_QUANT_TBLS; tblno++) {
+    qtblptr = dstinfo->quant_tbl_ptrs[tblno];
+    if (qtblptr != NULL) {
+      for (i = 0; i < DCTSIZE; i++) {
+	for (j = 0; j < i; j++) {
+	  qtemp = qtblptr->quantval[i*DCTSIZE+j];
+	  qtblptr->quantval[i*DCTSIZE+j] = qtblptr->quantval[j*DCTSIZE+i];
+	  qtblptr->quantval[j*DCTSIZE+i] = qtemp;
+	}
+      }
+    }
+  }
+}
+
+
+/* Adjust Exif image parameters.
+ *
+ * We try to adjust the Tags ExifImageWidth and ExifImageHeight if possible.
+ */
+
+LOCAL(void)
+adjust_exif_parameters (JOCTET FAR * data, unsigned int length,
+			JDIMENSION new_width, JDIMENSION new_height)
+{
+  boolean is_motorola; /* Flag for byte order */
+  unsigned int number_of_tags, tagnum;
+  unsigned int firstoffset, offset;
+  JDIMENSION new_value;
+
+  if (length < 12) return; /* Length of an IFD entry */
+
+  /* Discover byte order */
+  if (GETJOCTET(data[0]) == 0x49 && GETJOCTET(data[1]) == 0x49)
+    is_motorola = FALSE;
+  else if (GETJOCTET(data[0]) == 0x4D && GETJOCTET(data[1]) == 0x4D)
+    is_motorola = TRUE;
+  else
+    return;
+
+  /* Check Tag Mark */
+  if (is_motorola) {
+    if (GETJOCTET(data[2]) != 0) return;
+    if (GETJOCTET(data[3]) != 0x2A) return;
+  } else {
+    if (GETJOCTET(data[3]) != 0) return;
+    if (GETJOCTET(data[2]) != 0x2A) return;
+  }
+
+  /* Get first IFD offset (offset to IFD0) */
+  if (is_motorola) {
+    if (GETJOCTET(data[4]) != 0) return;
+    if (GETJOCTET(data[5]) != 0) return;
+    firstoffset = GETJOCTET(data[6]);
+    firstoffset <<= 8;
+    firstoffset += GETJOCTET(data[7]);
+  } else {
+    if (GETJOCTET(data[7]) != 0) return;
+    if (GETJOCTET(data[6]) != 0) return;
+    firstoffset = GETJOCTET(data[5]);
+    firstoffset <<= 8;
+    firstoffset += GETJOCTET(data[4]);
+  }
+  if (firstoffset > length - 2) return; /* check end of data segment */
+
+  /* Get the number of directory entries contained in this IFD */
+  if (is_motorola) {
+    number_of_tags = GETJOCTET(data[firstoffset]);
+    number_of_tags <<= 8;
+    number_of_tags += GETJOCTET(data[firstoffset+1]);
+  } else {
+    number_of_tags = GETJOCTET(data[firstoffset+1]);
+    number_of_tags <<= 8;
+    number_of_tags += GETJOCTET(data[firstoffset]);
+  }
+  if (number_of_tags == 0) return;
+  firstoffset += 2;
+
+  /* Search for ExifSubIFD offset Tag in IFD0 */
+  for (;;) {
+    if (firstoffset > length - 12) return; /* check end of data segment */
+    /* Get Tag number */
+    if (is_motorola) {
+      tagnum = GETJOCTET(data[firstoffset]);
+      tagnum <<= 8;
+      tagnum += GETJOCTET(data[firstoffset+1]);
+    } else {
+      tagnum = GETJOCTET(data[firstoffset+1]);
+      tagnum <<= 8;
+      tagnum += GETJOCTET(data[firstoffset]);
+    }
+    if (tagnum == 0x8769) break; /* found ExifSubIFD offset Tag */
+    if (--number_of_tags == 0) return;
+    firstoffset += 12;
+  }
+
+  /* Get the ExifSubIFD offset */
+  if (is_motorola) {
+    if (GETJOCTET(data[firstoffset+8]) != 0) return;
+    if (GETJOCTET(data[firstoffset+9]) != 0) return;
+    offset = GETJOCTET(data[firstoffset+10]);
+    offset <<= 8;
+    offset += GETJOCTET(data[firstoffset+11]);
+  } else {
+    if (GETJOCTET(data[firstoffset+11]) != 0) return;
+    if (GETJOCTET(data[firstoffset+10]) != 0) return;
+    offset = GETJOCTET(data[firstoffset+9]);
+    offset <<= 8;
+    offset += GETJOCTET(data[firstoffset+8]);
+  }
+  if (offset > length - 2) return; /* check end of data segment */
+
+  /* Get the number of directory entries contained in this SubIFD */
+  if (is_motorola) {
+    number_of_tags = GETJOCTET(data[offset]);
+    number_of_tags <<= 8;
+    number_of_tags += GETJOCTET(data[offset+1]);
+  } else {
+    number_of_tags = GETJOCTET(data[offset+1]);
+    number_of_tags <<= 8;
+    number_of_tags += GETJOCTET(data[offset]);
+  }
+  if (number_of_tags < 2) return;
+  offset += 2;
+
+  /* Search for ExifImageWidth and ExifImageHeight Tags in this SubIFD */
+  do {
+    if (offset > length - 12) return; /* check end of data segment */
+    /* Get Tag number */
+    if (is_motorola) {
+      tagnum = GETJOCTET(data[offset]);
+      tagnum <<= 8;
+      tagnum += GETJOCTET(data[offset+1]);
+    } else {
+      tagnum = GETJOCTET(data[offset+1]);
+      tagnum <<= 8;
+      tagnum += GETJOCTET(data[offset]);
+    }
+    if (tagnum == 0xA002 || tagnum == 0xA003) {
+      if (tagnum == 0xA002)
+	new_value = new_width; /* ExifImageWidth Tag */
+      else
+	new_value = new_height; /* ExifImageHeight Tag */
+      if (is_motorola) {
+	data[offset+2] = 0; /* Format = unsigned long (4 octets) */
+	data[offset+3] = 4;
+	data[offset+4] = 0; /* Number Of Components = 1 */
+	data[offset+5] = 0;
+	data[offset+6] = 0;
+	data[offset+7] = 1;
+	data[offset+8] = 0;
+	data[offset+9] = 0;
+	data[offset+10] = (JOCTET)((new_value >> 8) & 0xFF);
+	data[offset+11] = (JOCTET)(new_value & 0xFF);
+      } else {
+	data[offset+2] = 4; /* Format = unsigned long (4 octets) */
+	data[offset+3] = 0;
+	data[offset+4] = 1; /* Number Of Components = 1 */
+	data[offset+5] = 0;
+	data[offset+6] = 0;
+	data[offset+7] = 0;
+	data[offset+8] = (JOCTET)(new_value & 0xFF);
+	data[offset+9] = (JOCTET)((new_value >> 8) & 0xFF);
+	data[offset+10] = 0;
+	data[offset+11] = 0;
+      }
+    }
+    offset += 12;
+  } while (--number_of_tags);
+}
+
+
+/* Adjust output image parameters as needed.
+ *
+ * This must be called after jpeg_copy_critical_parameters()
+ * and before jpeg_write_coefficients().
+ *
+ * The return value is the set of virtual coefficient arrays to be written
+ * (either the ones allocated by jtransform_request_workspace, or the
+ * original source data arrays).  The caller will need to pass this value
+ * to jpeg_write_coefficients().
+ */
+
+GLOBAL(jvirt_barray_ptr *)
+jtransform_adjust_parameters (j_decompress_ptr srcinfo,
+			      j_compress_ptr dstinfo,
+			      jvirt_barray_ptr *src_coef_arrays,
+			      jpeg_transform_info *info)
+{
+  /* If force-to-grayscale is requested, adjust destination parameters */
+  if (info->force_grayscale) {
+    /* First, ensure we have YCbCr or grayscale data, and that the source's
+     * Y channel is full resolution.  (No reasonable person would make Y
+     * be less than full resolution, so actually coping with that case
+     * isn't worth extra code space.  But we check it to avoid crashing.)
+     */
+    if (((dstinfo->jpeg_color_space == JCS_YCbCr &&
+	  dstinfo->num_components == 3) ||
+	 (dstinfo->jpeg_color_space == JCS_GRAYSCALE &&
+	  dstinfo->num_components == 1)) &&
+	srcinfo->comp_info[0].h_samp_factor == srcinfo->max_h_samp_factor &&
+	srcinfo->comp_info[0].v_samp_factor == srcinfo->max_v_samp_factor) {
+      /* We use jpeg_set_colorspace to make sure subsidiary settings get fixed
+       * properly.  Among other things, it sets the target h_samp_factor &
+       * v_samp_factor to 1, which typically won't match the source.
+       * We have to preserve the source's quantization table number, however.
+       */
+      int sv_quant_tbl_no = dstinfo->comp_info[0].quant_tbl_no;
+      jpeg_set_colorspace(dstinfo, JCS_GRAYSCALE);
+      dstinfo->comp_info[0].quant_tbl_no = sv_quant_tbl_no;
+    } else {
+      /* Sorry, can't do it */
+      ERREXIT(dstinfo, JERR_CONVERSION_NOTIMPL);
+    }
+  } else if (info->num_components == 1) {
+    /* For a single-component source, we force the destination sampling factors
+     * to 1x1, with or without force_grayscale.  This is useful because some
+     * decoders choke on grayscale images with other sampling factors.
+     */
+    dstinfo->comp_info[0].h_samp_factor = 1;
+    dstinfo->comp_info[0].v_samp_factor = 1;
+  }
+
+  /* Correct the destination's image dimensions as necessary
+   * for rotate/flip, resize, and crop operations.
+   */
+#if JPEG_LIB_VERSION >= 70
+  dstinfo->jpeg_width = info->output_width;
+  dstinfo->jpeg_height = info->output_height;
+#endif
+
+  /* Transpose destination image parameters */
+  switch (info->transform) {
+  case JXFORM_TRANSPOSE:
+  case JXFORM_TRANSVERSE:
+  case JXFORM_ROT_90:
+  case JXFORM_ROT_270:
+#if JPEG_LIB_VERSION < 70
+    dstinfo->image_width = info->output_height;
+    dstinfo->image_height = info->output_width;
+#endif
+    transpose_critical_parameters(dstinfo);
+    break;
+  default:
+#if JPEG_LIB_VERSION < 70
+    dstinfo->image_width = info->output_width;
+    dstinfo->image_height = info->output_height;
+#endif
+    break;
+  }
+
+  /* Adjust Exif properties */
+  if (srcinfo->marker_list != NULL &&
+      srcinfo->marker_list->marker == JPEG_APP0+1 &&
+      srcinfo->marker_list->data_length >= 6 &&
+      GETJOCTET(srcinfo->marker_list->data[0]) == 0x45 &&
+      GETJOCTET(srcinfo->marker_list->data[1]) == 0x78 &&
+      GETJOCTET(srcinfo->marker_list->data[2]) == 0x69 &&
+      GETJOCTET(srcinfo->marker_list->data[3]) == 0x66 &&
+      GETJOCTET(srcinfo->marker_list->data[4]) == 0 &&
+      GETJOCTET(srcinfo->marker_list->data[5]) == 0) {
+    /* Suppress output of JFIF marker */
+    dstinfo->write_JFIF_header = FALSE;
+#if JPEG_LIB_VERSION >= 70
+    /* Adjust Exif image parameters */
+    if (dstinfo->jpeg_width != srcinfo->image_width ||
+	dstinfo->jpeg_height != srcinfo->image_height)
+      /* Align data segment to start of TIFF structure for parsing */
+      adjust_exif_parameters(srcinfo->marker_list->data + 6,
+	srcinfo->marker_list->data_length - 6,
+	dstinfo->jpeg_width, dstinfo->jpeg_height);
+#endif
+  }
+
+  /* Return the appropriate output data set */
+  if (info->workspace_coef_arrays != NULL)
+    return info->workspace_coef_arrays;
+  return src_coef_arrays;
+}
+
+
+/* Execute the actual transformation, if any.
+ *
+ * This must be called *after* jpeg_write_coefficients, because it depends
+ * on jpeg_write_coefficients to have computed subsidiary values such as
+ * the per-component width and height fields in the destination object.
+ *
+ * Note that some transformations will modify the source data arrays!
+ */
+
+GLOBAL(void)
+jtransform_execute_transform (j_decompress_ptr srcinfo,
+			      j_compress_ptr dstinfo,
+			      jvirt_barray_ptr *src_coef_arrays,
+			      jpeg_transform_info *info)
+{
+  jvirt_barray_ptr *dst_coef_arrays = info->workspace_coef_arrays;
+
+  /* Note: conditions tested here should match those in switch statement
+   * in jtransform_request_workspace()
+   */
+  switch (info->transform) {
+  case JXFORM_NONE:
+    if (info->x_crop_offset != 0 || info->y_crop_offset != 0)
+      do_crop(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+	      src_coef_arrays, dst_coef_arrays);
+    break;
+  case JXFORM_FLIP_H:
+    if (info->y_crop_offset != 0)
+      do_flip_h(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+		src_coef_arrays, dst_coef_arrays);
+    else
+      do_flip_h_no_crop(srcinfo, dstinfo, info->x_crop_offset,
+			src_coef_arrays);
+    break;
+  case JXFORM_FLIP_V:
+    do_flip_v(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+	      src_coef_arrays, dst_coef_arrays);
+    break;
+  case JXFORM_TRANSPOSE:
+    do_transpose(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+		 src_coef_arrays, dst_coef_arrays);
+    break;
+  case JXFORM_TRANSVERSE:
+    do_transverse(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+		  src_coef_arrays, dst_coef_arrays);
+    break;
+  case JXFORM_ROT_90:
+    do_rot_90(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+	      src_coef_arrays, dst_coef_arrays);
+    break;
+  case JXFORM_ROT_180:
+    do_rot_180(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+	       src_coef_arrays, dst_coef_arrays);
+    break;
+  case JXFORM_ROT_270:
+    do_rot_270(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+	       src_coef_arrays, dst_coef_arrays);
+    break;
+  }
+}
+
+/* jtransform_perfect_transform
+ *
+ * Determine whether lossless transformation is perfectly
+ * possible for a specified image and transformation.
+ *
+ * Inputs:
+ *   image_width, image_height: source image dimensions.
+ *   MCU_width, MCU_height: pixel dimensions of MCU.
+ *   transform: transformation identifier.
+ * Parameter sources from initialized jpeg_struct
+ * (after reading source header):
+ *   image_width = cinfo.image_width
+ *   image_height = cinfo.image_height
+ *   MCU_width = cinfo.max_h_samp_factor * cinfo.block_size
+ *   MCU_height = cinfo.max_v_samp_factor * cinfo.block_size
+ * Result:
+ *   TRUE = perfect transformation possible
+ *   FALSE = perfect transformation not possible
+ *           (may use custom action then)
+ */
+
+GLOBAL(boolean)
+jtransform_perfect_transform(JDIMENSION image_width, JDIMENSION image_height,
+			     int MCU_width, int MCU_height,
+			     JXFORM_CODE transform)
+{
+  boolean result = TRUE; /* initialize TRUE */
+
+  switch (transform) {
+  case JXFORM_FLIP_H:
+  case JXFORM_ROT_270:
+    if (image_width % (JDIMENSION) MCU_width)
+      result = FALSE;
+    break;
+  case JXFORM_FLIP_V:
+  case JXFORM_ROT_90:
+    if (image_height % (JDIMENSION) MCU_height)
+      result = FALSE;
+    break;
+  case JXFORM_TRANSVERSE:
+  case JXFORM_ROT_180:
+    if (image_width % (JDIMENSION) MCU_width)
+      result = FALSE;
+    if (image_height % (JDIMENSION) MCU_height)
+      result = FALSE;
+    break;
+  default:
+    break;
+  }
+
+  return result;
+}
+
+#endif /* TRANSFORMS_SUPPORTED */
+
+
+/* Setup decompression object to save desired markers in memory.
+ * This must be called before jpeg_read_header() to have the desired effect.
+ */
+
+GLOBAL(void)
+jcopy_markers_setup (j_decompress_ptr srcinfo, JCOPY_OPTION option)
+{
+#ifdef SAVE_MARKERS_SUPPORTED
+  int m;
+
+  /* Save comments except under NONE option */
+  if (option != JCOPYOPT_NONE) {
+    jpeg_save_markers(srcinfo, JPEG_COM, 0xFFFF);
+  }
+  /* Save all types of APPn markers iff ALL option */
+  if (option == JCOPYOPT_ALL) {
+    for (m = 0; m < 16; m++)
+      jpeg_save_markers(srcinfo, JPEG_APP0 + m, 0xFFFF);
+  }
+#endif /* SAVE_MARKERS_SUPPORTED */
+}
+
+/* Copy markers saved in the given source object to the destination object.
+ * This should be called just after jpeg_start_compress() or
+ * jpeg_write_coefficients().
+ * Note that those routines will have written the SOI, and also the
+ * JFIF APP0 or Adobe APP14 markers if selected.
+ */
+
+GLOBAL(void)
+jcopy_markers_execute (j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+		       JCOPY_OPTION option)
+{
+  jpeg_saved_marker_ptr marker;
+
+  /* In the current implementation, we don't actually need to examine the
+   * option flag here; we just copy everything that got saved.
+   * But to avoid confusion, we do not output JFIF and Adobe APP14 markers
+   * if the encoder library already wrote one.
+   */
+  for (marker = srcinfo->marker_list; marker != NULL; marker = marker->next) {
+    if (dstinfo->write_JFIF_header &&
+	marker->marker == JPEG_APP0 &&
+	marker->data_length >= 5 &&
+	GETJOCTET(marker->data[0]) == 0x4A &&
+	GETJOCTET(marker->data[1]) == 0x46 &&
+	GETJOCTET(marker->data[2]) == 0x49 &&
+	GETJOCTET(marker->data[3]) == 0x46 &&
+	GETJOCTET(marker->data[4]) == 0)
+      continue;			/* reject duplicate JFIF */
+    if (dstinfo->write_Adobe_marker &&
+	marker->marker == JPEG_APP0+14 &&
+	marker->data_length >= 5 &&
+	GETJOCTET(marker->data[0]) == 0x41 &&
+	GETJOCTET(marker->data[1]) == 0x64 &&
+	GETJOCTET(marker->data[2]) == 0x6F &&
+	GETJOCTET(marker->data[3]) == 0x62 &&
+	GETJOCTET(marker->data[4]) == 0x65)
+      continue;			/* reject duplicate Adobe */
+#ifdef NEED_FAR_POINTERS
+    /* We could use jpeg_write_marker if the data weren't FAR... */
+    {
+      unsigned int i;
+      jpeg_write_m_header(dstinfo, marker->marker, marker->data_length);
+      for (i = 0; i < marker->data_length; i++)
+	jpeg_write_m_byte(dstinfo, marker->data[i]);
+    }
+#else
+    jpeg_write_marker(dstinfo, marker->marker,
+		      marker->data, marker->data_length);
+#endif
+  }
+}
diff --git a/transupp.h b/transupp.h
new file mode 100644
index 0000000..7c16c19
--- /dev/null
+++ b/transupp.h
@@ -0,0 +1,210 @@
+/*
+ * transupp.h
+ *
+ * Copyright (C) 1997-2009, Thomas G. Lane, Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains declarations for image transformation routines and
+ * other utility code used by the jpegtran sample application.  These are
+ * NOT part of the core JPEG library.  But we keep these routines separate
+ * from jpegtran.c to ease the task of maintaining jpegtran-like programs
+ * that have other user interfaces.
+ *
+ * NOTE: all the routines declared here have very specific requirements
+ * about when they are to be executed during the reading and writing of the
+ * source and destination files.  See the comments in transupp.c, or see
+ * jpegtran.c for an example of correct usage.
+ */
+
+/* If you happen not to want the image transform support, disable it here */
+#ifndef TRANSFORMS_SUPPORTED
+#define TRANSFORMS_SUPPORTED 1		/* 0 disables transform code */
+#endif
+
+/*
+ * Although rotating and flipping data expressed as DCT coefficients is not
+ * hard, there is an asymmetry in the JPEG format specification for images
+ * whose dimensions aren't multiples of the iMCU size.  The right and bottom
+ * image edges are padded out to the next iMCU boundary with junk data; but
+ * no padding is possible at the top and left edges.  If we were to flip
+ * the whole image including the pad data, then pad garbage would become
+ * visible at the top and/or left, and real pixels would disappear into the
+ * pad margins --- perhaps permanently, since encoders & decoders may not
+ * bother to preserve DCT blocks that appear to be completely outside the
+ * nominal image area.  So, we have to exclude any partial iMCUs from the
+ * basic transformation.
+ *
+ * Transpose is the only transformation that can handle partial iMCUs at the
+ * right and bottom edges completely cleanly.  flip_h can flip partial iMCUs
+ * at the bottom, but leaves any partial iMCUs at the right edge untouched.
+ * Similarly flip_v leaves any partial iMCUs at the bottom edge untouched.
+ * The other transforms are defined as combinations of these basic transforms
+ * and process edge blocks in a way that preserves the equivalence.
+ *
+ * The "trim" option causes untransformable partial iMCUs to be dropped;
+ * this is not strictly lossless, but it usually gives the best-looking
+ * result for odd-size images.  Note that when this option is active,
+ * the expected mathematical equivalences between the transforms may not hold.
+ * (For example, -rot 270 -trim trims only the bottom edge, but -rot 90 -trim
+ * followed by -rot 180 -trim trims both edges.)
+ *
+ * We also offer a lossless-crop option, which discards data outside a given
+ * image region but losslessly preserves what is inside.  Like the rotate and
+ * flip transforms, lossless crop is restricted by the JPEG format: the upper
+ * left corner of the selected region must fall on an iMCU boundary.  If this
+ * does not hold for the given crop parameters, we silently move the upper left
+ * corner up and/or left to make it so, simultaneously increasing the region
+ * dimensions to keep the lower right crop corner unchanged.  (Thus, the
+ * output image covers at least the requested region, but may cover more.)
+ *
+ * We also provide a lossless-resize option, which is kind of a lossless-crop
+ * operation in the DCT coefficient block domain - it discards higher-order
+ * coefficients and losslessly preserves lower-order coefficients of a
+ * sub-block.
+ *
+ * Rotate/flip transform, resize, and crop can be requested together in a
+ * single invocation.  The crop is applied last --- that is, the crop region
+ * is specified in terms of the destination image after transform/resize.
+ *
+ * We also offer a "force to grayscale" option, which simply discards the
+ * chrominance channels of a YCbCr image.  This is lossless in the sense that
+ * the luminance channel is preserved exactly.  It's not the same kind of
+ * thing as the rotate/flip transformations, but it's convenient to handle it
+ * as part of this package, mainly because the transformation routines have to
+ * be aware of the option to know how many components to work on.
+ */
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jtransform_parse_crop_spec	jTrParCrop
+#define jtransform_request_workspace	jTrRequest
+#define jtransform_adjust_parameters	jTrAdjust
+#define jtransform_execute_transform	jTrExec
+#define jtransform_perfect_transform	jTrPerfect
+#define jcopy_markers_setup		jCMrkSetup
+#define jcopy_markers_execute		jCMrkExec
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/*
+ * Codes for supported types of image transformations.
+ */
+
+typedef enum {
+	JXFORM_NONE,		/* no transformation */
+	JXFORM_FLIP_H,		/* horizontal flip */
+	JXFORM_FLIP_V,		/* vertical flip */
+	JXFORM_TRANSPOSE,	/* transpose across UL-to-LR axis */
+	JXFORM_TRANSVERSE,	/* transpose across UR-to-LL axis */
+	JXFORM_ROT_90,		/* 90-degree clockwise rotation */
+	JXFORM_ROT_180,		/* 180-degree rotation */
+	JXFORM_ROT_270		/* 270-degree clockwise (or 90 ccw) */
+} JXFORM_CODE;
+
+/*
+ * Codes for crop parameters, which can individually be unspecified,
+ * positive, or negative.  (Negative width or height makes no sense, though.)
+ */
+
+typedef enum {
+	JCROP_UNSET,
+	JCROP_POS,
+	JCROP_NEG
+} JCROP_CODE;
+
+/*
+ * Transform parameters struct.
+ * NB: application must not change any elements of this struct after
+ * calling jtransform_request_workspace.
+ */
+
+typedef struct {
+  /* Options: set by caller */
+  JXFORM_CODE transform;	/* image transform operator */
+  boolean perfect;		/* if TRUE, fail if partial MCUs are requested */
+  boolean trim;			/* if TRUE, trim partial MCUs as needed */
+  boolean force_grayscale;	/* if TRUE, convert color image to grayscale */
+  boolean crop;			/* if TRUE, crop source image */
+
+  /* Crop parameters: application need not set these unless crop is TRUE.
+   * These can be filled in by jtransform_parse_crop_spec().
+   */
+  JDIMENSION crop_width;	/* Width of selected region */
+  JCROP_CODE crop_width_set;
+  JDIMENSION crop_height;	/* Height of selected region */
+  JCROP_CODE crop_height_set;
+  JDIMENSION crop_xoffset;	/* X offset of selected region */
+  JCROP_CODE crop_xoffset_set;	/* (negative measures from right edge) */
+  JDIMENSION crop_yoffset;	/* Y offset of selected region */
+  JCROP_CODE crop_yoffset_set;	/* (negative measures from bottom edge) */
+
+  /* Internal workspace: caller should not touch these */
+  int num_components;		/* # of components in workspace */
+  jvirt_barray_ptr * workspace_coef_arrays; /* workspace for transformations */
+  JDIMENSION output_width;	/* cropped destination dimensions */
+  JDIMENSION output_height;
+  JDIMENSION x_crop_offset;	/* destination crop offsets measured in iMCUs */
+  JDIMENSION y_crop_offset;
+  int iMCU_sample_width;	/* destination iMCU size */
+  int iMCU_sample_height;
+} jpeg_transform_info;
+
+
+#if TRANSFORMS_SUPPORTED
+
+/* Parse a crop specification (written in X11 geometry style) */
+EXTERN(boolean) jtransform_parse_crop_spec
+	JPP((jpeg_transform_info *info, const char *spec));
+/* Request any required workspace */
+EXTERN(boolean) jtransform_request_workspace
+	JPP((j_decompress_ptr srcinfo, jpeg_transform_info *info));
+/* Adjust output image parameters */
+EXTERN(jvirt_barray_ptr *) jtransform_adjust_parameters
+	JPP((j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	     jvirt_barray_ptr *src_coef_arrays,
+	     jpeg_transform_info *info));
+/* Execute the actual transformation, if any */
+EXTERN(void) jtransform_execute_transform
+	JPP((j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	     jvirt_barray_ptr *src_coef_arrays,
+	     jpeg_transform_info *info));
+/* Determine whether lossless transformation is perfectly
+ * possible for a specified image and transformation.
+ */
+EXTERN(boolean) jtransform_perfect_transform
+	JPP((JDIMENSION image_width, JDIMENSION image_height,
+	     int MCU_width, int MCU_height,
+	     JXFORM_CODE transform));
+
+/* jtransform_execute_transform used to be called
+ * jtransform_execute_transformation, but some compilers complain about
+ * routine names that long.  This macro is here to avoid breaking any
+ * old source code that uses the original name...
+ */
+#define jtransform_execute_transformation	jtransform_execute_transform
+
+#endif /* TRANSFORMS_SUPPORTED */
+
+
+/*
+ * Support for copying optional markers from source to destination file.
+ */
+
+typedef enum {
+	JCOPYOPT_NONE,		/* copy no optional markers */
+	JCOPYOPT_COMMENTS,	/* copy only comment (COM) markers */
+	JCOPYOPT_ALL		/* copy all optional markers */
+} JCOPY_OPTION;
+
+#define JCOPYOPT_DEFAULT  JCOPYOPT_COMMENTS	/* recommended default */
+
+/* Setup decompression object to save desired markers in memory */
+EXTERN(void) jcopy_markers_setup
+	JPP((j_decompress_ptr srcinfo, JCOPY_OPTION option));
+/* Copy markers saved in the given source object to the destination object */
+EXTERN(void) jcopy_markers_execute
+	JPP((j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+	     JCOPY_OPTION option));
diff --git a/turbojpeg-mapfile b/turbojpeg-mapfile
new file mode 100755
index 0000000..b2776f2
--- /dev/null
+++ b/turbojpeg-mapfile
@@ -0,0 +1,13 @@
+{
+	global:
+		tjInitCompress;
+		tjCompress;
+		TJBUFSIZE;
+		tjInitDecompress;
+		tjDecompressHeader;
+		tjDecompress;
+		tjDestroy;
+		tjGetErrorStr;
+	local:
+		*;
+};
diff --git a/turbojpeg.h b/turbojpeg.h
new file mode 100644
index 0000000..e382e71
--- /dev/null
+++ b/turbojpeg.h
@@ -0,0 +1,226 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005, 2006 Sun Microsystems, Inc.
+ * Copyright (C)2009 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+#if (defined(_MSC_VER) || defined(__CYGWIN__) || defined(__MINGW32__)) && defined(_WIN32) && defined(DLLDEFINE)
+#define DLLEXPORT __declspec(dllexport)
+#else
+#define DLLEXPORT
+#endif
+
+#define DLLCALL
+
+/* Subsampling */
+#define NUMSUBOPT 4
+
+enum {TJ_444=0, TJ_422, TJ_420, TJ_GRAYSCALE};
+#define TJ_411 TJ_420  /* for backward compatibility with VirtualGL <= 2.1.x,
+                          TurboVNC <= 0.6, and TurboJPEG/IPP */
+
+/* Flags */
+#define TJ_BGR             1
+  /* The components of each pixel in the source/destination bitmap are stored
+     in B,G,R order, not R,G,B */
+#define TJ_BOTTOMUP        2
+  /* The source/destination bitmap is stored in bottom-up (Windows, OpenGL)
+     order, not top-down (X11) order */
+#define TJ_FORCEMMX        8
+  /* Turn off CPU auto-detection and force TurboJPEG to use MMX code
+     (IPP and 32-bit libjpeg-turbo versions only) */
+#define TJ_FORCESSE       16
+  /* Turn off CPU auto-detection and force TurboJPEG to use SSE code
+     (32-bit IPP and 32-bit libjpeg-turbo versions only) */
+#define TJ_FORCESSE2      32
+  /* Turn off CPU auto-detection and force TurboJPEG to use SSE2 code
+     (32-bit IPP and 32-bit libjpeg-turbo versions only) */
+#define TJ_ALPHAFIRST     64
+  /* If the source/destination bitmap is 32 bpp, assume that each pixel is
+     ARGB/XRGB (or ABGR/XBGR if TJ_BGR is also specified) */
+#define TJ_FORCESSE3     128
+  /* Turn off CPU auto-detection and force TurboJPEG to use SSE3 code
+     (64-bit IPP version only) */
+#define TJ_FASTUPSAMPLE  256
+  /* Use fast, inaccurate 4:2:2 and 4:2:0 YUV upsampling routines
+     (libjpeg version only) */
+typedef void* tjhandle;
+
+#define TJPAD(p) (((p)+3)&(~3))
+#ifndef max
+ #define max(a,b) ((a)>(b)?(a):(b))
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* API follows */
+
+
+/*
+  tjhandle tjInitCompress(void)
+
+  Creates a new JPEG compressor instance, allocates memory for the structures,
+  and returns a handle to the instance.  Most applications will only
+  need to call this once at the beginning of the program or once for each
+  concurrent thread.  Don't try to create a new instance every time you
+  compress an image, because this will cause performance to suffer.
+
+  RETURNS: NULL on error
+*/
+DLLEXPORT tjhandle DLLCALL tjInitCompress(void);
+
+
+/*
+  int tjCompress(tjhandle j,
+     unsigned char *srcbuf, int width, int pitch, int height, int pixelsize,
+     unsigned char *dstbuf, unsigned long *size,
+     int jpegsubsamp, int jpegqual, int flags)
+
+  [INPUT] j = instance handle previously returned from a call to
+     tjInitCompress()
+  [INPUT] srcbuf = pointer to user-allocated image buffer containing pixels in
+     RGB(A) or BGR(A) form
+  [INPUT] width =  width (in pixels) of the source image
+  [INPUT] pitch = bytes per line of the source image (width*pixelsize if the
+     bitmap is unpadded, else TJPAD(width*pixelsize) if each line of the bitmap
+     is padded to the nearest 32-bit boundary, such as is the case for Windows
+     bitmaps.  You can also be clever and use this parameter to skip lines, etc.,
+     as long as the pitch is greater than 0.)
+  [INPUT] height = height (in pixels) of the source image
+  [INPUT] pixelsize = size (in bytes) of each pixel in the source image
+     RGBA and BGRA: 4, RGB and BGR: 3, Grayscale: 1
+  [INPUT] dstbuf = pointer to user-allocated image buffer which will receive
+     the JPEG image.  Use the macro TJBUFSIZE(width, height) to determine
+     the appropriate size for this buffer based on the image width and height.
+  [OUTPUT] size = pointer to unsigned long which receives the size (in bytes)
+     of the compressed image
+  [INPUT] jpegsubsamp = Specifies either 4:2:0, 4:2:2, or 4:4:4 subsampling.
+     When the image is converted from the RGB to YCbCr colorspace as part of the
+     JPEG compression process, every other Cb and Cr (chrominance) pixel can be
+     discarded to produce a smaller image with little perceptible loss of
+     image clarity (the human eye is more sensitive to small changes in
+     brightness than small changes in color.)
+
+     TJ_420: 4:2:0 subsampling.  Discards every other Cb, Cr pixel in both
+        horizontal and vertical directions.
+     TJ_422: 4:2:2 subsampling.  Discards every other Cb, Cr pixel only in
+        the horizontal direction.
+     TJ_444: no subsampling.
+     TJ_GRAYSCALE: Generate grayscale JPEG image
+
+  [INPUT] jpegqual = JPEG quality (an integer between 0 and 100 inclusive.)
+  [INPUT] flags = the bitwise OR of one or more of the flags described in the
+     "Flags" section above.
+
+  RETURNS: 0 on success, -1 on error
+*/
+DLLEXPORT int DLLCALL tjCompress(tjhandle j,
+	unsigned char *srcbuf, int width, int pitch, int height, int pixelsize,
+	unsigned char *dstbuf, unsigned long *size,
+	int jpegsubsamp, int jpegqual, int flags);
+
+DLLEXPORT unsigned long DLLCALL TJBUFSIZE(int width, int height);
+
+/*
+  tjhandle tjInitDecompress(void)
+
+  Creates a new JPEG decompressor instance, allocates memory for the
+  structures, and returns a handle to the instance.  Most applications will
+  only need to call this once at the beginning of the program or once for each
+  concurrent thread.  Don't try to create a new instance every time you
+  decompress an image, because this will cause performance to suffer.
+
+  RETURNS: NULL on error
+*/
+DLLEXPORT tjhandle DLLCALL tjInitDecompress(void);
+
+
+/*
+  int tjDecompressHeader(tjhandle j,
+     unsigned char *srcbuf, unsigned long size,
+     int *width, int *height)
+
+  [INPUT] j = instance handle previously returned from a call to
+     tjInitDecompress()
+  [INPUT] srcbuf = pointer to a user-allocated buffer containing the JPEG image
+     to decompress
+  [INPUT] size = size of the JPEG image buffer (in bytes)
+  [OUTPUT] width = width (in pixels) of the JPEG image
+  [OUTPUT] height = height (in pixels) of the JPEG image
+
+  RETURNS: 0 on success, -1 on error
+*/
+DLLEXPORT int DLLCALL tjDecompressHeader(tjhandle j,
+	unsigned char *srcbuf, unsigned long size,
+	int *width, int *height);
+
+
+/*
+  int tjDecompress(tjhandle j,
+     unsigned char *srcbuf, unsigned long size,
+     unsigned char *dstbuf, int width, int pitch, int height, int pixelsize,
+     int flags)
+
+  [INPUT] j = instance handle previously returned from a call to
+     tjInitDecompress()
+  [INPUT] srcbuf = pointer to a user-allocated buffer containing the JPEG image
+     to decompress
+  [INPUT] size = size of the JPEG image buffer (in bytes)
+  [INPUT] dstbuf = pointer to user-allocated image buffer which will receive
+     the bitmap image.  This buffer should normally be pitch*height
+     bytes in size, although this pointer may also be used to decompress into
+     a specific region of a larger buffer.
+  [INPUT] width =  width (in pixels) of the destination image
+  [INPUT] pitch = bytes per line of the destination image (width*pixelsize if the
+     bitmap is unpadded, else TJPAD(width*pixelsize) if each line of the bitmap
+     is padded to the nearest 32-bit boundary, such as is the case for Windows
+     bitmaps.  You can also be clever and use this parameter to skip lines, etc.,
+     as long as the pitch is greater than 0.)
+  [INPUT] height = height (in pixels) of the destination image
+  [INPUT] pixelsize = size (in bytes) of each pixel in the destination image
+     RGBA/RGBx and BGRA/BGRx: 4, RGB and BGR: 3, Grayscale: 1
+  [INPUT] flags = the bitwise OR of one or more of the flags described in the
+     "Flags" section above.
+
+  RETURNS: 0 on success, -1 on error
+*/
+DLLEXPORT int DLLCALL tjDecompress(tjhandle j,
+	unsigned char *srcbuf, unsigned long size,
+	unsigned char *dstbuf, int width, int pitch, int height, int pixelsize,
+	int flags);
+
+
+/*
+  int tjDestroy(tjhandle h)
+
+  Frees structures associated with a compression or decompression instance
+  
+  [INPUT] h = instance handle (returned from a previous call to
+     tjInitCompress() or tjInitDecompress()
+
+  RETURNS: 0 on success, -1 on error
+*/
+DLLEXPORT int DLLCALL tjDestroy(tjhandle h);
+
+
+/*
+  char *tjGetErrorStr(void)
+  
+  Returns a descriptive error message explaining why the last command failed
+*/
+DLLEXPORT char* DLLCALL tjGetErrorStr(void);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/turbojpegl.c b/turbojpegl.c
new file mode 100644
index 0000000..03833c5
--- /dev/null
+++ b/turbojpegl.c
@@ -0,0 +1,367 @@
+/* Copyright (C)2004 Landmark Graphics Corporation
+ * Copyright (C)2005 Sun Microsystems, Inc.
+ * Copyright (C)2009-2010 D. R. Commander
+ *
+ * This library is free software and may be redistributed and/or modified under
+ * the terms of the wxWindows Library License, Version 3.1 or (at your option)
+ * any later version.  The full license is in the LICENSE.txt file included
+ * with this distribution.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * wxWindows Library License for more details.
+ */
+
+// This implements a JPEG compressor/decompressor using the libjpeg API
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <jpeglib.h>
+#include <jerror.h>
+#include <setjmp.h>
+#include "./turbojpeg.h"
+
+
+// Error handling
+
+static char lasterror[JMSG_LENGTH_MAX]="No error";
+
+typedef struct _error_mgr
+{
+	struct jpeg_error_mgr pub;
+	jmp_buf jb;
+} error_mgr;
+
+static void my_error_exit(j_common_ptr cinfo)
+{
+	error_mgr *myerr = (error_mgr *)cinfo->err;
+	(*cinfo->err->output_message)(cinfo);
+	longjmp(myerr->jb, 1);
+}
+
+static void my_output_message(j_common_ptr cinfo)
+{
+	(*cinfo->err->format_message)(cinfo, lasterror);
+}
+
+
+// Global structures, macros, etc.
+
+typedef struct _jpgstruct
+{
+	struct jpeg_compress_struct cinfo;
+	struct jpeg_decompress_struct dinfo;
+	struct jpeg_destination_mgr jdms;
+	struct jpeg_source_mgr jsms;
+	error_mgr jerr;
+	int initc, initd;
+} jpgstruct;
+
+static const int hsampfactor[NUMSUBOPT]={1, 2, 2, 1};
+static const int vsampfactor[NUMSUBOPT]={1, 1, 2, 1};
+
+#define _throw(c) {sprintf(lasterror, "%s", c);  return -1;}
+#define _catch(f) {if((f)==-1) return -1;}
+#define checkhandle(h) jpgstruct *j=(jpgstruct *)h; \
+	if(!j) _throw("Invalid handle");
+
+
+// CO
+
+static boolean empty_output_buffer(struct jpeg_compress_struct *cinfo)
+{
+	ERREXIT(cinfo, JERR_BUFFER_SIZE);
+	return TRUE;
+}
+
+static void destination_noop(struct jpeg_compress_struct *cinfo)
+{
+}
+
+DLLEXPORT tjhandle DLLCALL tjInitCompress(void)
+{
+	jpgstruct *j=NULL;
+	if((j=(jpgstruct *)malloc(sizeof(jpgstruct)))==NULL)
+		{sprintf(lasterror, "Memory allocation failure");  return NULL;}
+	memset(j, 0, sizeof(jpgstruct));
+	j->cinfo.err=jpeg_std_error(&j->jerr.pub);
+	j->jerr.pub.error_exit=my_error_exit;
+	j->jerr.pub.output_message=my_output_message;
+
+	if(setjmp(j->jerr.jb))
+	{ // this will execute if LIBJPEG has an error
+		if(j) free(j);  return NULL;
+	}
+
+	jpeg_create_compress(&j->cinfo);
+	j->cinfo.dest=&j->jdms;
+	j->jdms.init_destination=destination_noop;
+	j->jdms.empty_output_buffer=empty_output_buffer;
+	j->jdms.term_destination=destination_noop;
+
+	j->initc=1;
+	return (tjhandle)j;
+}
+
+DLLEXPORT unsigned long DLLCALL TJBUFSIZE(int width, int height)
+{
+	// This allows enough room in case the image doesn't compress
+	return ((width+15)&(~15)) * ((height+15)&(~15)) * 6 + 2048;
+}
+
+DLLEXPORT int DLLCALL tjCompress(tjhandle h,
+	unsigned char *srcbuf, int width, int pitch, int height, int ps,
+	unsigned char *dstbuf, unsigned long *size,
+	int jpegsub, int qual, int flags)
+{
+	int i;  JSAMPROW *row_pointer=NULL;
+
+	checkhandle(h);
+
+	if(srcbuf==NULL || width<=0 || pitch<0 || height<=0
+		|| dstbuf==NULL || size==NULL
+		|| jpegsub<0 || jpegsub>=NUMSUBOPT || qual<0 || qual>100)
+		_throw("Invalid argument in tjCompress()");
+	if(ps!=3 && ps!=4 && ps!=1)
+		_throw("This compressor can only handle 24-bit and 32-bit RGB or 8-bit grayscale input");
+	if(!j->initc) _throw("Instance has not been initialized for compression");
+
+	if(pitch==0) pitch=width*ps;
+
+	j->cinfo.image_width = width;
+	j->cinfo.image_height = height;
+	j->cinfo.input_components = ps;
+
+	if(ps==1) j->cinfo.in_color_space = JCS_GRAYSCALE;
+	#if JCS_EXTENSIONS==1
+	else j->cinfo.in_color_space = JCS_EXT_RGB;
+	if(ps==3 && (flags&TJ_BGR))
+		j->cinfo.in_color_space = JCS_EXT_BGR;
+	else if(ps==4 && !(flags&TJ_BGR) && !(flags&TJ_ALPHAFIRST))
+		j->cinfo.in_color_space = JCS_EXT_RGBX;
+	else if(ps==4 && (flags&TJ_BGR) && !(flags&TJ_ALPHAFIRST))
+		j->cinfo.in_color_space = JCS_EXT_BGRX;
+	else if(ps==4 && (flags&TJ_BGR) && (flags&TJ_ALPHAFIRST))
+		j->cinfo.in_color_space = JCS_EXT_XBGR;
+	else if(ps==4 && !(flags&TJ_BGR) && (flags&TJ_ALPHAFIRST))
+		j->cinfo.in_color_space = JCS_EXT_XRGB;
+	#else
+	#error "TurboJPEG requires JPEG colorspace extensions"
+	#endif
+
+	if(flags&TJ_FORCEMMX) putenv("JSIMD_FORCEMMX=1");
+	else if(flags&TJ_FORCESSE) putenv("JSIMD_FORCESSE=1");
+	else if(flags&TJ_FORCESSE2) putenv("JSIMD_FORCESSE2=1");
+
+	if(setjmp(j->jerr.jb))
+	{  // this will execute if LIBJPEG has an error
+		if(row_pointer) free(row_pointer);
+		return -1;
+	}
+
+	jpeg_set_defaults(&j->cinfo);
+
+	jpeg_set_quality(&j->cinfo, qual, TRUE);
+	if(jpegsub==TJ_GRAYSCALE)
+		jpeg_set_colorspace(&j->cinfo, JCS_GRAYSCALE);
+	else
+		jpeg_set_colorspace(&j->cinfo, JCS_YCbCr);
+	j->cinfo.dct_method = JDCT_FASTEST;
+
+	j->cinfo.comp_info[0].h_samp_factor=hsampfactor[jpegsub];
+	j->cinfo.comp_info[1].h_samp_factor=1;
+	j->cinfo.comp_info[2].h_samp_factor=1;
+	j->cinfo.comp_info[0].v_samp_factor=vsampfactor[jpegsub];
+	j->cinfo.comp_info[1].v_samp_factor=1;
+	j->cinfo.comp_info[2].v_samp_factor=1;
+
+	j->jdms.next_output_byte = dstbuf;
+	j->jdms.free_in_buffer = TJBUFSIZE(j->cinfo.image_width, j->cinfo.image_height);
+
+	if((row_pointer=(JSAMPROW *)malloc(sizeof(JSAMPROW)*height))==NULL)
+		_throw("Memory allocation failed in tjInitCompress()");
+	for(i=0; i<height; i++)
+	{
+		if(flags&TJ_BOTTOMUP) row_pointer[i]= &srcbuf[(height-i-1)*pitch];
+		else row_pointer[i]= &srcbuf[i*pitch];
+	}
+	jpeg_start_compress(&j->cinfo, TRUE);
+	while(j->cinfo.next_scanline<j->cinfo.image_height)
+	{
+		jpeg_write_scanlines(&j->cinfo, &row_pointer[j->cinfo.next_scanline],
+			j->cinfo.image_height-j->cinfo.next_scanline);
+	}
+	jpeg_finish_compress(&j->cinfo);
+	*size=TJBUFSIZE(j->cinfo.image_width, j->cinfo.image_height)
+		-(unsigned long)(j->jdms.free_in_buffer);
+
+	if(row_pointer) free(row_pointer);
+	return 0;
+}
+
+
+// DEC
+
+static boolean fill_input_buffer (struct jpeg_decompress_struct *dinfo)
+{
+	ERREXIT(dinfo, JERR_BUFFER_SIZE);
+	return TRUE;
+}
+
+static void skip_input_data (struct jpeg_decompress_struct *dinfo, long num_bytes)
+{
+	dinfo->src->next_input_byte += (size_t) num_bytes;
+	dinfo->src->bytes_in_buffer -= (size_t) num_bytes;
+}
+
+static void source_noop (struct jpeg_decompress_struct *dinfo)
+{
+}
+
+DLLEXPORT tjhandle DLLCALL tjInitDecompress(void)
+{
+	jpgstruct *j;
+	if((j=(jpgstruct *)malloc(sizeof(jpgstruct)))==NULL)
+		{sprintf(lasterror, "Memory allocation failure");  return NULL;}
+	memset(j, 0, sizeof(jpgstruct));
+	j->dinfo.err=jpeg_std_error(&j->jerr.pub);
+	j->jerr.pub.error_exit=my_error_exit;
+	j->jerr.pub.output_message=my_output_message;
+
+	if(setjmp(j->jerr.jb))
+	{ // this will execute if LIBJPEG has an error
+		free(j);  return NULL;
+  }
+
+	jpeg_create_decompress(&j->dinfo);
+	j->dinfo.src=&j->jsms;
+	j->jsms.init_source=source_noop;
+	j->jsms.fill_input_buffer = fill_input_buffer;
+	j->jsms.skip_input_data = skip_input_data;
+	j->jsms.resync_to_restart = jpeg_resync_to_restart;
+	j->jsms.term_source = source_noop;
+
+	j->initd=1;
+	return (tjhandle)j;
+}
+
+
+DLLEXPORT int DLLCALL tjDecompressHeader(tjhandle h,
+	unsigned char *srcbuf, unsigned long size,
+	int *width, int *height)
+{
+	checkhandle(h);
+
+	if(srcbuf==NULL || size<=0 || width==NULL || height==NULL)
+		_throw("Invalid argument in tjDecompressHeader()");
+	if(!j->initd) _throw("Instance has not been initialized for decompression");
+
+	if(setjmp(j->jerr.jb))
+	{  // this will execute if LIBJPEG has an error
+		return -1;
+	}
+
+	j->jsms.bytes_in_buffer = size;
+	j->jsms.next_input_byte = srcbuf;
+
+	jpeg_read_header(&j->dinfo, TRUE);
+
+	*width=j->dinfo.image_width;  *height=j->dinfo.image_height;
+
+	jpeg_abort_decompress(&j->dinfo);
+
+	if(*width<1 || *height<1) _throw("Invalid data returned in header");
+	return 0;
+}
+
+
+DLLEXPORT int DLLCALL tjDecompress(tjhandle h,
+	unsigned char *srcbuf, unsigned long size,
+	unsigned char *dstbuf, int width, int pitch, int height, int ps,
+	int flags)
+{
+	int i;  JSAMPROW *row_pointer=NULL;
+
+	checkhandle(h);
+
+	if(srcbuf==NULL || size<=0
+		|| dstbuf==NULL || width<=0 || pitch<0 || height<=0)
+		_throw("Invalid argument in tjDecompress()");
+	if(ps!=3 && ps!=4 && ps!=1)
+		_throw("This decompressor can only handle 24-bit and 32-bit RGB or 8-bit grayscale output");
+	if(!j->initd) _throw("Instance has not been initialized for decompression");
+
+	if(pitch==0) pitch=width*ps;
+
+	if(flags&TJ_FORCEMMX) putenv("JSIMD_FORCEMMX=1");
+	else if(flags&TJ_FORCESSE) putenv("JSIMD_FORCESSE=1");
+	else if(flags&TJ_FORCESSE2) putenv("JSIMD_FORCESSE2=1");
+
+	if(setjmp(j->jerr.jb))
+	{  // this will execute if LIBJPEG has an error
+		if(row_pointer) free(row_pointer);
+		return -1;
+  }
+
+	j->jsms.bytes_in_buffer = size;
+	j->jsms.next_input_byte = srcbuf;
+
+	jpeg_read_header(&j->dinfo, TRUE);
+
+	if((row_pointer=(JSAMPROW *)malloc(sizeof(JSAMPROW)*height))==NULL)
+		_throw("Memory allocation failed in tjInitDecompress()");
+	for(i=0; i<height; i++)
+	{
+		if(flags&TJ_BOTTOMUP) row_pointer[i]= &dstbuf[(height-i-1)*pitch];
+		else row_pointer[i]= &dstbuf[i*pitch];
+	}
+
+	if(ps==1) j->dinfo.out_color_space = JCS_GRAYSCALE;
+	#if JCS_EXTENSIONS==1
+	else j->dinfo.out_color_space = JCS_EXT_RGB;
+	if(ps==3 && (flags&TJ_BGR))
+		j->dinfo.out_color_space = JCS_EXT_BGR;
+	else if(ps==4 && !(flags&TJ_BGR) && !(flags&TJ_ALPHAFIRST))
+		j->dinfo.out_color_space = JCS_EXT_RGBX;
+	else if(ps==4 && (flags&TJ_BGR) && !(flags&TJ_ALPHAFIRST))
+		j->dinfo.out_color_space = JCS_EXT_BGRX;
+	else if(ps==4 && (flags&TJ_BGR) && (flags&TJ_ALPHAFIRST))
+		j->dinfo.out_color_space = JCS_EXT_XBGR;
+	else if(ps==4 && !(flags&TJ_BGR) && (flags&TJ_ALPHAFIRST))
+		j->dinfo.out_color_space = JCS_EXT_XRGB;
+	#else
+	#error "TurboJPEG requires JPEG colorspace extensions"
+	#endif
+	if(flags&TJ_FASTUPSAMPLE) j->dinfo.do_fancy_upsampling=FALSE;
+
+	jpeg_start_decompress(&j->dinfo);
+	while(j->dinfo.output_scanline<j->dinfo.output_height)
+	{
+		jpeg_read_scanlines(&j->dinfo, &row_pointer[j->dinfo.output_scanline],
+			j->dinfo.output_height-j->dinfo.output_scanline);
+	}
+	jpeg_finish_decompress(&j->dinfo);
+
+	if(row_pointer) free(row_pointer);
+	return 0;
+}
+
+
+// General
+
+DLLEXPORT char* DLLCALL tjGetErrorStr(void)
+{
+	return lasterror;
+}
+
+DLLEXPORT int DLLCALL tjDestroy(tjhandle h)
+{
+	checkhandle(h);
+	if(setjmp(j->jerr.jb)) return -1;
+	if(j->initc) jpeg_destroy_compress(&j->cinfo);
+	if(j->initd) jpeg_destroy_decompress(&j->dinfo);
+	free(j);
+	return 0;
+}
diff --git a/usage.txt b/usage.txt
new file mode 100644
index 0000000..0519b94
--- /dev/null
+++ b/usage.txt
@@ -0,0 +1,602 @@
+USAGE instructions for the Independent JPEG Group's JPEG software
+=================================================================
+
+This file describes usage of the JPEG conversion programs cjpeg and djpeg,
+as well as the utility programs jpegtran, rdjpgcom and wrjpgcom.  (See
+the other documentation files if you wish to use the JPEG library within
+your own programs.)
+
+If you are on a Unix machine you may prefer to read the Unix-style manual
+pages in files cjpeg.1, djpeg.1, jpegtran.1, rdjpgcom.1, wrjpgcom.1.
+
+
+INTRODUCTION
+
+These programs implement JPEG image encoding, decoding, and transcoding.
+JPEG (pronounced "jay-peg") is a standardized compression method for
+full-color and gray-scale images.
+
+
+GENERAL USAGE
+
+We provide two programs, cjpeg to compress an image file into JPEG format,
+and djpeg to decompress a JPEG file back into a conventional image format.
+
+On Unix-like systems, you say:
+	cjpeg [switches] [imagefile] >jpegfile
+or
+	djpeg [switches] [jpegfile]  >imagefile
+The programs read the specified input file, or standard input if none is
+named.  They always write to standard output (with trace/error messages to
+standard error).  These conventions are handy for piping images between
+programs.
+
+On most non-Unix systems, you say:
+	cjpeg [switches] imagefile jpegfile
+or
+	djpeg [switches] jpegfile  imagefile
+i.e., both the input and output files are named on the command line.  This
+style is a little more foolproof, and it loses no functionality if you don't
+have pipes.  (You can get this style on Unix too, if you prefer, by defining
+TWO_FILE_COMMANDLINE when you compile the programs; see install.txt.)
+
+You can also say:
+	cjpeg [switches] -outfile jpegfile  imagefile
+or
+	djpeg [switches] -outfile imagefile  jpegfile
+This syntax works on all systems, so it is useful for scripts.
+
+The currently supported image file formats are: PPM (PBMPLUS color format),
+PGM (PBMPLUS gray-scale format), BMP, Targa, and RLE (Utah Raster Toolkit
+format).  (RLE is supported only if the URT library is available.)
+cjpeg recognizes the input image format automatically, with the exception
+of some Targa-format files.  You have to tell djpeg which format to generate.
+
+JPEG files are in the defacto standard JFIF file format.  There are other,
+less widely used JPEG-based file formats, but we don't support them.
+
+All switch names may be abbreviated; for example, -grayscale may be written
+-gray or -gr.  Most of the "basic" switches can be abbreviated to as little as
+one letter.  Upper and lower case are equivalent (-BMP is the same as -bmp).
+British spellings are also accepted (e.g., -greyscale), though for brevity
+these are not mentioned below.
+
+
+CJPEG DETAILS
+
+The basic command line switches for cjpeg are:
+
+	-quality N[,...]  Scale quantization tables to adjust image quality.
+			Quality is 0 (worst) to 100 (best); default is 75.
+			(See below for more info.)
+
+	-grayscale	Create monochrome JPEG file from color input.
+			Be sure to use this switch when compressing a grayscale
+			BMP file, because cjpeg isn't bright enough to notice
+			whether a BMP file uses only shades of gray.  By
+			saying -grayscale, you'll get a smaller JPEG file that
+			takes less time to process.
+
+	-optimize	Perform optimization of entropy encoding parameters.
+			Without this, default encoding parameters are used.
+			-optimize usually makes the JPEG file a little smaller,
+			but cjpeg runs somewhat slower and needs much more
+			memory.  Image quality and speed of decompression are
+			unaffected by -optimize.
+
+	-progressive	Create progressive JPEG file (see below).
+
+	-targa		Input file is Targa format.  Targa files that contain
+			an "identification" field will not be automatically
+			recognized by cjpeg; for such files you must specify
+			-targa to make cjpeg treat the input as Targa format.
+			For most Targa files, you won't need this switch.
+
+The -quality switch lets you trade off compressed file size against quality of
+the reconstructed image: the higher the quality setting, the larger the JPEG
+file, and the closer the output image will be to the original input.  Normally
+you want to use the lowest quality setting (smallest file) that decompresses
+into something visually indistinguishable from the original image.  For this
+purpose the quality setting should be between 50 and 95; the default of 75 is
+often about right.  If you see defects at -quality 75, then go up 5 or 10
+counts at a time until you are happy with the output image.  (The optimal
+setting will vary from one image to another.)
+
+-quality 100 will generate a quantization table of all 1's, minimizing loss
+in the quantization step (but there is still information loss in subsampling,
+as well as roundoff error).  This setting is mainly of interest for
+experimental purposes.  Quality values above about 95 are NOT recommended for
+normal use; the compressed file size goes up dramatically for hardly any gain
+in output image quality.
+
+In the other direction, quality values below 50 will produce very small files
+of low image quality.  Settings around 5 to 10 might be useful in preparing an
+index of a large image library, for example.  Try -quality 2 (or so) for some
+amusing Cubist effects.  (Note: quality values below about 25 generate 2-byte
+quantization tables, which are considered optional in the JPEG standard.
+cjpeg emits a warning message when you give such a quality value, because some
+other JPEG programs may be unable to decode the resulting file.  Use -baseline
+if you need to ensure compatibility at low quality values.)
+
+The -quality option has been extended in this version of cjpeg to support
+separate quality settings for luminance and chrominance (or, in general,
+separate settings for every quantization table slot.)  The principle is the
+same as chrominance subsampling:  since the human eye is more sensitive to
+spatial changes in brightness than spatial changes in color, the chrominance
+components can be quantized more than the luminance components without
+incurring any visible image quality loss.  However, unlike subsampling, this
+feature reduces data in the frequency domain instead of the spatial domain,
+which allows for more fine-grained control.  This option is useful in
+quality-sensitive applications, for which the artifacts generated by
+subsampling may be unacceptable.
+
+The -quality option accepts a comma-separated list of parameters, which
+respectively refer to the quality levels which should be assigned to the
+quantization table slots.  If there are more q-table slots than parameters,
+then the last parameter is replicated.  Thus, if only one quality parameter is
+given, this is used for both luminance and chrominance (slots 0 and 1,
+respectively), preserving the legacy behavior of cjpeg v6b and prior.  More (or
+customized) quantization tables can be set with the -qtables option and
+assigned to components with the -qslots option (see the "wizard" switches
+below.)
+
+JPEG  files  generated  with separate luminance and chrominance quality are
+fully compliant with standard JPEG decoders.
+
+CAUTION: For this setting to be useful, be sure to pass an argument of
+-sample 1x1 to cjpeg to disable chrominance subsampling.  Otherwise, the
+default subsampling level (2x2, AKA "4:2:0") will be used.
+
+The -progressive switch creates a "progressive JPEG" file.  In this type of
+JPEG file, the data is stored in multiple scans of increasing quality.  If the
+file is being transmitted over a slow communications link, the decoder can use
+the first scan to display a low-quality image very quickly, and can then
+improve the display with each subsequent scan.  The final image is exactly
+equivalent to a standard JPEG file of the same quality setting, and the total
+file size is about the same --- often a little smaller.
+
+Switches for advanced users:
+
+	-dct int	Use integer DCT method (default).
+	-dct fast	Use fast integer DCT (less accurate).
+	-dct float	Use floating-point DCT method.
+			The float method is very slightly more accurate than
+			the int method, but is much slower unless your machine
+			has very fast floating-point hardware.  Also note that
+			results of the floating-point method may vary slightly
+			across machines, while the integer methods should give
+			the same results everywhere.  The fast integer method
+			is much less accurate than the other two.
+
+	-restart N	Emit a JPEG restart marker every N MCU rows, or every
+			N MCU blocks if "B" is attached to the number.
+			-restart 0 (the default) means no restart markers.
+
+	-smooth N	Smooth the input image to eliminate dithering noise.
+			N, ranging from 1 to 100, indicates the strength of
+			smoothing.  0 (the default) means no smoothing.
+
+	-maxmemory N	Set limit for amount of memory to use in processing
+			large images.  Value is in thousands of bytes, or
+			millions of bytes if "M" is attached to the number.
+			For example, -max 4m selects 4000000 bytes.  If more
+			space is needed, temporary files will be used.
+
+	-verbose	Enable debug printout.  More -v's give more printout.
+	or  -debug	Also, version information is printed at startup.
+
+The -restart option inserts extra markers that allow a JPEG decoder to
+resynchronize after a transmission error.  Without restart markers, any damage
+to a compressed file will usually ruin the image from the point of the error
+to the end of the image; with restart markers, the damage is usually confined
+to the portion of the image up to the next restart marker.  Of course, the
+restart markers occupy extra space.  We recommend -restart 1 for images that
+will be transmitted across unreliable networks such as Usenet.
+
+The -smooth option filters the input to eliminate fine-scale noise.  This is
+often useful when converting dithered images to JPEG: a moderate smoothing
+factor of 10 to 50 gets rid of dithering patterns in the input file, resulting
+in a smaller JPEG file and a better-looking image.  Too large a smoothing
+factor will visibly blur the image, however.
+
+Switches for wizards:
+
+	-baseline	Force baseline-compatible quantization tables to be
+			generated.  This clamps quantization values to 8 bits
+			even at low quality settings.  (This switch is poorly
+			named, since it does not ensure that the output is
+			actually baseline JPEG.  For example, you can use
+			-baseline and -progressive together.)
+
+	-qtables file	Use the quantization tables given in the specified
+			text file.
+
+	-qslots N[,...] Select which quantization table to use for each color
+			component.
+
+	-sample HxV[,...]  Set JPEG sampling factors for each color component.
+
+	-scans file	Use the scan script given in the specified text file.
+
+The "wizard" switches are intended for experimentation with JPEG.  If you
+don't know what you are doing, DON'T USE THEM.  These switches are documented
+further in the file wizard.txt.
+
+
+DJPEG DETAILS
+
+The basic command line switches for djpeg are:
+
+	-colors N	Reduce image to at most N colors.  This reduces the
+	or -quantize N	number of colors used in the output image, so that it
+			can be displayed on a colormapped display or stored in
+			a colormapped file format.  For example, if you have
+			an 8-bit display, you'd need to reduce to 256 or fewer
+			colors.  (-colors is the recommended name, -quantize
+			is provided only for backwards compatibility.)
+
+	-fast		Select recommended processing options for fast, low
+			quality output.  (The default options are chosen for
+			highest quality output.)  Currently, this is equivalent
+			to "-dct fast -nosmooth -onepass -dither ordered".
+
+	-grayscale	Force gray-scale output even if JPEG file is color.
+			Useful for viewing on monochrome displays; also,
+			djpeg runs noticeably faster in this mode.
+
+	-scale M/N	Scale the output image by a factor M/N.  Currently
+			the scale factor must be 1/1, 1/2, 1/4, or 1/8.
+			Scaling is handy if the image is larger than your
+			screen; also, djpeg runs much faster when scaling
+			down the output.
+
+	-bmp		Select BMP output format (Windows flavor).  8-bit
+			colormapped format is emitted if -colors or -grayscale
+			is specified, or if the JPEG file is gray-scale;
+			otherwise, 24-bit full-color format is emitted.
+
+	-gif		Select GIF output format.  Since GIF does not support
+			more than 256 colors, -colors 256 is assumed (unless
+			you specify a smaller number of colors).  If you
+			specify -fast, the default number of colors is 216.
+
+	-os2		Select BMP output format (OS/2 1.x flavor).  8-bit
+			colormapped format is emitted if -colors or -grayscale
+			is specified, or if the JPEG file is gray-scale;
+			otherwise, 24-bit full-color format is emitted.
+
+	-pnm		Select PBMPLUS (PPM/PGM) output format (this is the
+			default format).  PGM is emitted if the JPEG file is
+			gray-scale or if -grayscale is specified; otherwise
+			PPM is emitted.
+
+	-rle		Select RLE output format.  (Requires URT library.)
+
+	-targa		Select Targa output format.  Gray-scale format is
+			emitted if the JPEG file is gray-scale or if
+			-grayscale is specified; otherwise, colormapped format
+			is emitted if -colors is specified; otherwise, 24-bit
+			full-color format is emitted.
+
+Switches for advanced users:
+
+	-dct int	Use integer DCT method (default).
+	-dct fast	Use fast integer DCT (less accurate).
+	-dct float	Use floating-point DCT method.
+			The float method is very slightly more accurate than
+			the int method, but is much slower unless your machine
+			has very fast floating-point hardware.  Also note that
+			results of the floating-point method may vary slightly
+			across machines, while the integer methods should give
+			the same results everywhere.  The fast integer method
+			is much less accurate than the other two.
+
+	-dither fs	Use Floyd-Steinberg dithering in color quantization.
+	-dither ordered	Use ordered dithering in color quantization.
+	-dither none	Do not use dithering in color quantization.
+			By default, Floyd-Steinberg dithering is applied when
+			quantizing colors; this is slow but usually produces
+			the best results.  Ordered dither is a compromise
+			between speed and quality; no dithering is fast but
+			usually looks awful.  Note that these switches have
+			no effect unless color quantization is being done.
+			Ordered dither is only available in -onepass mode.
+
+	-map FILE	Quantize to the colors used in the specified image
+			file.  This is useful for producing multiple files
+			with identical color maps, or for forcing a predefined
+			set of colors to be used.  The FILE must be a GIF
+			or PPM file.  This option overrides -colors and
+			-onepass.
+
+	-nosmooth	Use a faster, lower-quality upsampling routine.
+
+	-onepass	Use one-pass instead of two-pass color quantization.
+			The one-pass method is faster and needs less memory,
+			but it produces a lower-quality image.  -onepass is
+			ignored unless you also say -colors N.  Also,
+			the one-pass method is always used for gray-scale
+			output (the two-pass method is no improvement then).
+
+	-maxmemory N	Set limit for amount of memory to use in processing
+			large images.  Value is in thousands of bytes, or
+			millions of bytes if "M" is attached to the number.
+			For example, -max 4m selects 4000000 bytes.  If more
+			space is needed, temporary files will be used.
+
+	-verbose	Enable debug printout.  More -v's give more printout.
+	or  -debug	Also, version information is printed at startup.
+
+
+HINTS FOR CJPEG
+
+Color GIF files are not the ideal input for JPEG; JPEG is really intended for
+compressing full-color (24-bit) images.  In particular, don't try to convert
+cartoons, line drawings, and other images that have only a few distinct
+colors.  GIF works great on these, JPEG does not.  If you want to convert a
+GIF to JPEG, you should experiment with cjpeg's -quality and -smooth options
+to get a satisfactory conversion.  -smooth 10 or so is often helpful.
+
+Avoid running an image through a series of JPEG compression/decompression
+cycles.  Image quality loss will accumulate; after ten or so cycles the image
+may be noticeably worse than it was after one cycle.  It's best to use a
+lossless format while manipulating an image, then convert to JPEG format when
+you are ready to file the image away.
+
+The -optimize option to cjpeg is worth using when you are making a "final"
+version for posting or archiving.  It's also a win when you are using low
+quality settings to make very small JPEG files; the percentage improvement
+is often a lot more than it is on larger files.  (At present, -optimize
+mode is always selected when generating progressive JPEG files.)
+
+Support for GIF input files was removed in cjpeg v6b due to concerns over
+the Unisys LZW patent.  Although this patent expired in 2006, cjpeg still
+lacks GIF support, for these historical reasons.  (Conversion of GIF files to
+JPEG is usually a bad idea anyway.)
+
+
+HINTS FOR DJPEG
+
+To get a quick preview of an image, use the -grayscale and/or -scale switches.
+"-grayscale -scale 1/8" is the fastest case.
+
+Several options are available that trade off image quality to gain speed.
+"-fast" turns on the recommended settings.
+
+"-dct fast" and/or "-nosmooth" gain speed at a small sacrifice in quality.
+When producing a color-quantized image, "-onepass -dither ordered" is fast but
+much lower quality than the default behavior.  "-dither none" may give
+acceptable results in two-pass mode, but is seldom tolerable in one-pass mode.
+
+If you are fortunate enough to have very fast floating point hardware,
+"-dct float" may be even faster than "-dct fast".  But on most machines
+"-dct float" is slower than "-dct int"; in this case it is not worth using,
+because its theoretical accuracy advantage is too small to be significant
+in practice.
+
+Two-pass color quantization requires a good deal of memory; on MS-DOS machines
+it may run out of memory even with -maxmemory 0.  In that case you can still
+decompress, with some loss of image quality, by specifying -onepass for
+one-pass quantization.
+
+To avoid the Unisys LZW patent, djpeg produces uncompressed GIF files.  These
+are larger than they should be, but are readable by standard GIF decoders.
+
+
+HINTS FOR BOTH PROGRAMS
+
+If more space is needed than will fit in the available main memory (as
+determined by -maxmemory), temporary files will be used.  (MS-DOS versions
+will try to get extended or expanded memory first.)  The temporary files are
+often rather large: in typical cases they occupy three bytes per pixel, for
+example 3*800*600 = 1.44Mb for an 800x600 image.  If you don't have enough
+free disk space, leave out -progressive and -optimize (for cjpeg) or specify
+-onepass (for djpeg).
+
+On MS-DOS, the temporary files are created in the directory named by the TMP
+or TEMP environment variable, or in the current directory if neither of those
+exist.  Amiga implementations put the temp files in the directory named by
+JPEGTMP:, so be sure to assign JPEGTMP: to a disk partition with adequate free
+space.
+
+The default memory usage limit (-maxmemory) is set when the software is
+compiled.  If you get an "insufficient memory" error, try specifying a smaller
+-maxmemory value, even -maxmemory 0 to use the absolute minimum space.  You
+may want to recompile with a smaller default value if this happens often.
+
+On machines that have "environment" variables, you can define the environment
+variable JPEGMEM to set the default memory limit.  The value is specified as
+described for the -maxmemory switch.  JPEGMEM overrides the default value
+specified when the program was compiled, and itself is overridden by an
+explicit -maxmemory switch.
+
+On MS-DOS machines, -maxmemory is the amount of main (conventional) memory to
+use.  (Extended or expanded memory is also used if available.)  Most
+DOS-specific versions of this software do their own memory space estimation
+and do not need you to specify -maxmemory.
+
+
+JPEGTRAN
+
+jpegtran performs various useful transformations of JPEG files.
+It can translate the coded representation from one variant of JPEG to another,
+for example from baseline JPEG to progressive JPEG or vice versa.  It can also
+perform some rearrangements of the image data, for example turning an image
+from landscape to portrait format by rotation.
+
+jpegtran works by rearranging the compressed data (DCT coefficients), without
+ever fully decoding the image.  Therefore, its transformations are lossless:
+there is no image degradation at all, which would not be true if you used
+djpeg followed by cjpeg to accomplish the same conversion.  But by the same
+token, jpegtran cannot perform lossy operations such as changing the image
+quality.
+
+jpegtran uses a command line syntax similar to cjpeg or djpeg.
+On Unix-like systems, you say:
+	jpegtran [switches] [inputfile] >outputfile
+On most non-Unix systems, you say:
+	jpegtran [switches] inputfile outputfile
+where both the input and output files are JPEG files.
+
+To specify the coded JPEG representation used in the output file,
+jpegtran accepts a subset of the switches recognized by cjpeg:
+	-optimize	Perform optimization of entropy encoding parameters.
+	-progressive	Create progressive JPEG file.
+	-restart N	Emit a JPEG restart marker every N MCU rows, or every
+			N MCU blocks if "B" is attached to the number.
+	-scans file	Use the scan script given in the specified text file.
+See the previous discussion of cjpeg for more details about these switches.
+If you specify none of these switches, you get a plain baseline-JPEG output
+file.  The quality setting and so forth are determined by the input file.
+
+The image can be losslessly transformed by giving one of these switches:
+	-flip horizontal	Mirror image horizontally (left-right).
+	-flip vertical		Mirror image vertically (top-bottom).
+	-rotate 90		Rotate image 90 degrees clockwise.
+	-rotate 180		Rotate image 180 degrees.
+	-rotate 270		Rotate image 270 degrees clockwise (or 90 ccw).
+	-transpose		Transpose image (across UL-to-LR axis).
+	-transverse		Transverse transpose (across UR-to-LL axis).
+
+The transpose transformation has no restrictions regarding image dimensions.
+The other transformations operate rather oddly if the image dimensions are not
+a multiple of the iMCU size (usually 8 or 16 pixels), because they can only
+transform complete blocks of DCT coefficient data in the desired way.
+
+jpegtran's default behavior when transforming an odd-size image is designed
+to preserve exact reversibility and mathematical consistency of the
+transformation set.  As stated, transpose is able to flip the entire image
+area.  Horizontal mirroring leaves any partial iMCU column at the right edge
+untouched, but is able to flip all rows of the image.  Similarly, vertical
+mirroring leaves any partial iMCU row at the bottom edge untouched, but is
+able to flip all columns.  The other transforms can be built up as sequences
+of transpose and flip operations; for consistency, their actions on edge
+pixels are defined to be the same as the end result of the corresponding
+transpose-and-flip sequence.
+
+For practical use, you may prefer to discard any untransformable edge pixels
+rather than having a strange-looking strip along the right and/or bottom edges
+of a transformed image.  To do this, add the -trim switch:
+	-trim		Drop non-transformable edge blocks.
+Obviously, a transformation with -trim is not reversible, so strictly speaking
+jpegtran with this switch is not lossless.  Also, the expected mathematical
+equivalences between the transformations no longer hold.  For example,
+"-rot 270 -trim" trims only the bottom edge, but "-rot 90 -trim" followed by
+"-rot 180 -trim" trims both edges.
+
+If you are only interested in perfect transformations, add the -perfect switch:
+	-perfect	Fail with an error if the transformation is not
+			perfect.
+For example, you may want to do
+  jpegtran -rot 90 -perfect foo.jpg || djpeg foo.jpg | pnmflip -r90 | cjpeg
+to do a perfect rotation, if available, or an approximated one if not.
+
+This version of jpegtran also offers a lossless crop option, which discards
+data outside of a given image region but losslessly preserves what is inside.
+Like the rotate and flip transforms, lossless crop is restricted by the current
+JPEG format; the upper left corner of the selected region must fall on an iMCU
+boundary.  If it doesn't, then it is silently moved up and/or left to the
+nearest iMCU boundary (the lower right corner is unchanged.)
+
+The image can be losslessly cropped by giving the switch:
+	-crop WxH+X+Y	Crop to a rectangular region of width W and height H,
+			starting at point X,Y.
+
+Other not-strictly-lossless transformation switches are:
+
+	-grayscale	Force grayscale output.
+This option discards the chrominance channels if the input image is YCbCr
+(ie, a standard color JPEG), resulting in a grayscale JPEG file.  The
+luminance channel is preserved exactly, so this is a better method of reducing
+to grayscale than decompression, conversion, and recompression.  This switch
+is particularly handy for fixing a monochrome picture that was mistakenly
+encoded as a color JPEG.  (In such a case, the space savings from getting rid
+of the near-empty chroma channels won't be large; but the decoding time for
+a grayscale JPEG is substantially less than that for a color JPEG.)
+
+jpegtran also recognizes these switches that control what to do with "extra"
+markers, such as comment blocks:
+	-copy none	Copy no extra markers from source file.  This setting
+			suppresses all comments and other excess baggage
+			present in the source file.
+	-copy comments	Copy only comment markers.  This setting copies
+			comments from the source file but discards
+			any other data which is inessential for image display.
+	-copy all	Copy all extra markers.  This setting preserves
+			miscellaneous markers found in the source file, such
+			as JFIF thumbnails, Exif data, and Photoshop settings.
+			In some files, these extra markers can be sizable.
+The default behavior is -copy comments.  (Note: in IJG releases v6 and v6a,
+jpegtran always did the equivalent of -copy none.)
+
+Additional switches recognized by jpegtran are:
+	-outfile filename
+	-maxmemory N
+	-verbose
+	-debug
+These work the same as in cjpeg or djpeg.
+
+
+THE COMMENT UTILITIES
+
+The JPEG standard allows "comment" (COM) blocks to occur within a JPEG file.
+Although the standard doesn't actually define what COM blocks are for, they
+are widely used to hold user-supplied text strings.  This lets you add
+annotations, titles, index terms, etc to your JPEG files, and later retrieve
+them as text.  COM blocks do not interfere with the image stored in the JPEG
+file.  The maximum size of a COM block is 64K, but you can have as many of
+them as you like in one JPEG file.
+
+We provide two utility programs to display COM block contents and add COM
+blocks to a JPEG file.
+
+rdjpgcom searches a JPEG file and prints the contents of any COM blocks on
+standard output.  The command line syntax is
+	rdjpgcom [-raw] [-verbose] [inputfilename]
+The switch "-raw" (or just "-r") causes rdjpgcom to output non-printable
+characters in JPEG comments.  These characters are normally escaped for
+security reasons.
+The switch "-verbose" (or just "-v") causes rdjpgcom to also display the JPEG
+image dimensions.  If you omit the input file name from the command line,
+the JPEG file is read from standard input.  (This may not work on some
+operating systems, if binary data can't be read from stdin.)
+
+wrjpgcom adds a COM block, containing text you provide, to a JPEG file.
+Ordinarily, the COM block is added after any existing COM blocks, but you
+can delete the old COM blocks if you wish.  wrjpgcom produces a new JPEG
+file; it does not modify the input file.  DO NOT try to overwrite the input
+file by directing wrjpgcom's output back into it; on most systems this will
+just destroy your file.
+
+The command line syntax for wrjpgcom is similar to cjpeg's.  On Unix-like
+systems, it is
+	wrjpgcom [switches] [inputfilename]
+The output file is written to standard output.  The input file comes from
+the named file, or from standard input if no input file is named.
+
+On most non-Unix systems, the syntax is
+	wrjpgcom [switches] inputfilename outputfilename
+where both input and output file names must be given explicitly.
+
+wrjpgcom understands three switches:
+	-replace		 Delete any existing COM blocks from the file.
+	-comment "Comment text"	 Supply new COM text on command line.
+	-cfile name		 Read text for new COM block from named file.
+(Switch names can be abbreviated.)  If you have only one line of comment text
+to add, you can provide it on the command line with -comment.  The comment
+text must be surrounded with quotes so that it is treated as a single
+argument.  Longer comments can be read from a text file.
+
+If you give neither -comment nor -cfile, then wrjpgcom will read the comment
+text from standard input.  (In this case an input image file name MUST be
+supplied, so that the source JPEG file comes from somewhere else.)  You can
+enter multiple lines, up to 64KB worth.  Type an end-of-file indicator
+(usually control-D or control-Z) to terminate the comment text entry.
+
+wrjpgcom will not add a COM block if the provided comment string is empty.
+Therefore -replace -comment "" can be used to delete all COM blocks from a
+file.
+
+These utility programs do not depend on the IJG JPEG library.  In
+particular, the source code for rdjpgcom is intended as an illustration of
+the minimum amount of code required to parse a JPEG file header correctly.
diff --git a/win/config.h.in b/win/config.h.in
new file mode 100644
index 0000000..110af3c
--- /dev/null
+++ b/win/config.h.in
@@ -0,0 +1,3 @@
+#define VERSION "@VERSION@"

+#define BUILD "@BUILD@"

+#define PACKAGE_NAME "@CMAKE_PROJECT_NAME@"

diff --git a/win/jconfig.h.in b/win/jconfig.h.in
new file mode 100644
index 0000000..e1fc20e
--- /dev/null
+++ b/win/jconfig.h.in
@@ -0,0 +1,32 @@
+/* jconfig.vc --- jconfig.h for Microsoft Visual C++ on Windows 95 or NT. */
+/* see jconfig.txt for explanations */
+
+#define JPEG_LIB_VERSION @JPEG_LIB_VERSION@
+
+#define HAVE_PROTOTYPES
+#define HAVE_UNSIGNED_CHAR
+#define HAVE_UNSIGNED_SHORT
+/* #define void char */
+/* #define const */
+#undef CHAR_IS_UNSIGNED
+#define HAVE_STDDEF_H
+#define HAVE_STDLIB_H
+#undef NEED_BSD_STRINGS
+#undef NEED_SYS_TYPES_H
+#undef NEED_FAR_POINTERS	/* we presume a 32-bit flat memory model */
+#undef NEED_SHORT_EXTERNAL_NAMES
+#undef INCOMPLETE_TYPES_BROKEN
+
+/* Define "boolean" as unsigned char, not int, per Windows custom */
+#ifndef __RPCNDR_H__		/* don't conflict if rpcndr.h already read */
+typedef unsigned char boolean;
+#endif
+#define HAVE_BOOLEAN		/* prevent jmorecfg.h from redefining it */
+
+#define inline __inline
+
+#ifdef JPEG_INTERNALS
+
+#undef RIGHT_SHIFT_IS_UNSIGNED
+
+#endif /* JPEG_INTERNALS */
diff --git a/win/jpeg62.def b/win/jpeg62.def
new file mode 100755
index 0000000..3c33fbf
--- /dev/null
+++ b/win/jpeg62.def
@@ -0,0 +1,102 @@
+EXPORTS
+	jcopy_block_row @ 1 ; 
+	jcopy_sample_rows @ 2 ; 
+	jdiv_round_up @ 3 ; 
+	jinit_1pass_quantizer @ 4 ; 
+	jinit_2pass_quantizer @ 5 ; 
+	jinit_c_coef_controller @ 6 ; 
+	jinit_c_main_controller @ 7 ; 
+	jinit_c_master_control @ 8 ; 
+	jinit_c_prep_controller @ 9 ; 
+	jinit_color_converter @ 10 ; 
+	jinit_color_deconverter @ 11 ; 
+	jinit_compress_master @ 12 ; 
+	jinit_d_coef_controller @ 13 ; 
+	jinit_d_main_controller @ 14 ; 
+	jinit_d_post_controller @ 15 ; 
+	jinit_downsampler @ 16 ; 
+	jinit_forward_dct @ 17 ; 
+	jinit_huff_decoder @ 18 ; 
+	jinit_huff_encoder @ 19 ; 
+	jinit_input_controller @ 20 ; 
+	jinit_inverse_dct @ 21 ; 
+	jinit_marker_reader @ 22 ; 
+	jinit_marker_writer @ 23 ; 
+	jinit_master_decompress @ 24 ; 
+	jinit_memory_mgr @ 25 ; 
+	jinit_merged_upsampler @ 26 ; 
+	jinit_phuff_decoder @ 27 ; 
+	jinit_phuff_encoder @ 28 ; 
+	jinit_upsampler @ 29 ; 
+	jpeg_CreateCompress @ 30 ; 
+	jpeg_CreateDecompress @ 31 ; 
+	jpeg_abort @ 32 ; 
+	jpeg_abort_compress @ 33 ; 
+	jpeg_abort_decompress @ 34 ; 
+	jpeg_add_quant_table @ 35 ; 
+	jpeg_alloc_huff_table @ 36 ; 
+	jpeg_alloc_quant_table @ 37 ; 
+	jpeg_calc_output_dimensions @ 38 ; 
+	jpeg_consume_input @ 39 ; 
+	jpeg_copy_critical_parameters @ 40 ; 
+	jpeg_default_colorspace @ 41 ; 
+	jpeg_destroy @ 42 ; 
+	jpeg_destroy_compress @ 43 ; 
+	jpeg_destroy_decompress @ 44 ; 
+	jpeg_fdct_float @ 45 ; 
+	jpeg_fdct_ifast @ 46 ; 
+	jpeg_fdct_islow @ 47 ; 
+	jpeg_fill_bit_buffer @ 48 ; 
+	jpeg_finish_compress @ 49 ; 
+	jpeg_finish_decompress @ 50 ; 
+	jpeg_finish_output @ 51 ; 
+	jpeg_free_large @ 52 ; 
+	jpeg_free_small @ 53 ; 
+	jpeg_gen_optimal_table @ 54 ; 
+	jpeg_get_large @ 55 ; 
+	jpeg_get_small @ 56 ; 
+	jpeg_has_multiple_scans @ 57 ; 
+	jpeg_huff_decode @ 58 ; 
+	jpeg_idct_1x1 @ 59 ; 
+	jpeg_idct_2x2 @ 60 ; 
+	jpeg_idct_4x4 @ 61 ; 
+	jpeg_idct_float @ 62 ; 
+	jpeg_idct_ifast @ 63 ; 
+	jpeg_idct_islow @ 64 ; 
+	jpeg_input_complete @ 65 ; 
+	jpeg_make_c_derived_tbl @ 66 ; 
+	jpeg_make_d_derived_tbl @ 67 ; 
+	jpeg_mem_available @ 68 ; 
+	jpeg_mem_init @ 69 ; 
+	jpeg_mem_term @ 70 ; 
+	jpeg_new_colormap @ 71 ; 
+	jpeg_open_backing_store @ 72 ; 
+	jpeg_quality_scaling @ 73 ; 
+	jpeg_read_coefficients @ 74 ; 
+	jpeg_read_header @ 75 ; 
+	jpeg_read_raw_data @ 76 ; 
+	jpeg_read_scanlines @ 77 ; 
+	jpeg_resync_to_restart @ 78 ; 
+	jpeg_save_markers @ 79 ; 
+	jpeg_set_colorspace @ 80 ; 
+	jpeg_set_defaults @ 81 ; 
+	jpeg_set_linear_quality @ 82 ; 
+	jpeg_set_marker_processor @ 83 ; 
+	jpeg_set_quality @ 84 ; 
+	jpeg_simple_progression @ 85 ; 
+	jpeg_start_compress @ 86 ; 
+	jpeg_start_decompress @ 87 ; 
+	jpeg_start_output @ 88 ; 
+	jpeg_std_error @ 89 ; 
+	jpeg_stdio_dest @ 90 ; 
+	jpeg_stdio_src @ 91 ; 
+	jpeg_suppress_tables @ 92 ; 
+	jpeg_write_coefficients @ 93 ; 
+	jpeg_write_m_byte @ 94 ; 
+	jpeg_write_m_header @ 95 ; 
+	jpeg_write_marker @ 96 ; 
+	jpeg_write_raw_data @ 97 ; 
+	jpeg_write_scanlines @ 98 ; 
+	jpeg_write_tables @ 99 ; 
+	jround_up @ 100 ; 
+	jzero_far @ 101 ; 
diff --git a/win/jpeg7.def b/win/jpeg7.def
new file mode 100644
index 0000000..5ca227b
--- /dev/null
+++ b/win/jpeg7.def
@@ -0,0 +1,104 @@
+EXPORTS
+	jcopy_block_row @ 1 ; 
+	jcopy_sample_rows @ 2 ; 
+	jdiv_round_up @ 3 ; 
+	jinit_1pass_quantizer @ 4 ; 
+	jinit_2pass_quantizer @ 5 ; 
+	jinit_c_coef_controller @ 6 ; 
+	jinit_c_main_controller @ 7 ; 
+	jinit_c_master_control @ 8 ; 
+	jinit_c_prep_controller @ 9 ; 
+	jinit_color_converter @ 10 ; 
+	jinit_color_deconverter @ 11 ; 
+	jinit_compress_master @ 12 ; 
+	jinit_d_coef_controller @ 13 ; 
+	jinit_d_main_controller @ 14 ; 
+	jinit_d_post_controller @ 15 ; 
+	jinit_downsampler @ 16 ; 
+	jinit_forward_dct @ 17 ; 
+	jinit_huff_decoder @ 18 ; 
+	jinit_huff_encoder @ 19 ; 
+	jinit_input_controller @ 20 ; 
+	jinit_inverse_dct @ 21 ; 
+	jinit_marker_reader @ 22 ; 
+	jinit_marker_writer @ 23 ; 
+	jinit_master_decompress @ 24 ; 
+	jinit_memory_mgr @ 25 ; 
+	jinit_merged_upsampler @ 26 ; 
+	jinit_phuff_decoder @ 27 ; 
+	jinit_phuff_encoder @ 28 ; 
+	jinit_upsampler @ 29 ; 
+	jpeg_CreateCompress @ 30 ; 
+	jpeg_CreateDecompress @ 31 ; 
+	jpeg_abort @ 32 ; 
+	jpeg_abort_compress @ 33 ; 
+	jpeg_abort_decompress @ 34 ; 
+	jpeg_add_quant_table @ 35 ; 
+	jpeg_alloc_huff_table @ 36 ; 
+	jpeg_alloc_quant_table @ 37 ; 
+	jpeg_calc_jpeg_dimensions @ 38 ; 
+	jpeg_calc_output_dimensions @ 39 ; 
+	jpeg_consume_input @ 40 ; 
+	jpeg_copy_critical_parameters @ 41 ; 
+	jpeg_default_colorspace @ 42 ; 
+	jpeg_default_qtables @ 43 ;
+	jpeg_destroy @ 44 ; 
+	jpeg_destroy_compress @ 45 ; 
+	jpeg_destroy_decompress @ 46 ; 
+	jpeg_fdct_float @ 47 ; 
+	jpeg_fdct_ifast @ 48 ; 
+	jpeg_fdct_islow @ 49 ; 
+	jpeg_fill_bit_buffer @ 50 ; 
+	jpeg_finish_compress @ 51 ; 
+	jpeg_finish_decompress @ 52 ; 
+	jpeg_finish_output @ 53 ; 
+	jpeg_free_large @ 54 ; 
+	jpeg_free_small @ 55 ; 
+	jpeg_gen_optimal_table @ 56 ; 
+	jpeg_get_large @ 57 ; 
+	jpeg_get_small @ 58 ; 
+	jpeg_has_multiple_scans @ 59 ; 
+	jpeg_huff_decode @ 60 ; 
+	jpeg_idct_1x1 @ 61 ; 
+	jpeg_idct_2x2 @ 62 ; 
+	jpeg_idct_4x4 @ 63 ; 
+	jpeg_idct_float @ 64 ; 
+	jpeg_idct_ifast @ 65 ; 
+	jpeg_idct_islow @ 66 ; 
+	jpeg_input_complete @ 67 ; 
+	jpeg_make_c_derived_tbl @ 68 ; 
+	jpeg_make_d_derived_tbl @ 69 ; 
+	jpeg_mem_available @ 70 ; 
+	jpeg_mem_init @ 71 ; 
+	jpeg_mem_term @ 72 ; 
+	jpeg_new_colormap @ 73 ; 
+	jpeg_open_backing_store @ 74 ; 
+	jpeg_quality_scaling @ 75 ; 
+	jpeg_read_coefficients @ 76 ; 
+	jpeg_read_header @ 77 ; 
+	jpeg_read_raw_data @ 78 ; 
+	jpeg_read_scanlines @ 79 ; 
+	jpeg_resync_to_restart @ 80 ; 
+	jpeg_save_markers @ 81 ; 
+	jpeg_set_colorspace @ 82 ; 
+	jpeg_set_defaults @ 83 ; 
+	jpeg_set_linear_quality @ 84 ; 
+	jpeg_set_marker_processor @ 85 ; 
+	jpeg_set_quality @ 86 ; 
+	jpeg_simple_progression @ 87 ; 
+	jpeg_start_compress @ 88 ; 
+	jpeg_start_decompress @ 89 ; 
+	jpeg_start_output @ 90 ; 
+	jpeg_std_error @ 91 ; 
+	jpeg_stdio_dest @ 92 ; 
+	jpeg_stdio_src @ 93 ; 
+	jpeg_suppress_tables @ 94 ; 
+	jpeg_write_coefficients @ 95 ; 
+	jpeg_write_m_byte @ 96 ; 
+	jpeg_write_m_header @ 97 ; 
+	jpeg_write_marker @ 98 ; 
+	jpeg_write_raw_data @ 99 ; 
+	jpeg_write_scanlines @ 100 ; 
+	jpeg_write_tables @ 101 ; 
+	jround_up @ 102 ; 
+	jzero_far @ 103 ; 
diff --git a/win/jpeg8.def b/win/jpeg8.def
new file mode 100644
index 0000000..3fa6111
--- /dev/null
+++ b/win/jpeg8.def
@@ -0,0 +1,107 @@
+EXPORTS
+	jcopy_block_row @ 1 ; 
+	jcopy_sample_rows @ 2 ; 
+	jdiv_round_up @ 3 ; 
+	jinit_1pass_quantizer @ 4 ; 
+	jinit_2pass_quantizer @ 5 ; 
+	jinit_c_coef_controller @ 6 ; 
+	jinit_c_main_controller @ 7 ; 
+	jinit_c_master_control @ 8 ; 
+	jinit_c_prep_controller @ 9 ; 
+	jinit_color_converter @ 10 ; 
+	jinit_color_deconverter @ 11 ; 
+	jinit_compress_master @ 12 ; 
+	jinit_d_coef_controller @ 13 ; 
+	jinit_d_main_controller @ 14 ; 
+	jinit_d_post_controller @ 15 ; 
+	jinit_downsampler @ 16 ; 
+	jinit_forward_dct @ 17 ; 
+	jinit_huff_decoder @ 18 ; 
+	jinit_huff_encoder @ 19 ; 
+	jinit_input_controller @ 20 ; 
+	jinit_inverse_dct @ 21 ; 
+	jinit_marker_reader @ 22 ; 
+	jinit_marker_writer @ 23 ; 
+	jinit_master_decompress @ 24 ; 
+	jinit_memory_mgr @ 25 ; 
+	jinit_merged_upsampler @ 26 ; 
+	jinit_phuff_decoder @ 27 ; 
+	jinit_phuff_encoder @ 28 ; 
+	jinit_upsampler @ 29 ; 
+	jpeg_CreateCompress @ 30 ; 
+	jpeg_CreateDecompress @ 31 ; 
+	jpeg_abort @ 32 ; 
+	jpeg_abort_compress @ 33 ; 
+	jpeg_abort_decompress @ 34 ; 
+	jpeg_add_quant_table @ 35 ; 
+	jpeg_alloc_huff_table @ 36 ; 
+	jpeg_alloc_quant_table @ 37 ; 
+	jpeg_calc_jpeg_dimensions @ 38 ; 
+	jpeg_calc_output_dimensions @ 39 ; 
+	jpeg_consume_input @ 40 ; 
+	jpeg_copy_critical_parameters @ 41 ; 
+	jpeg_core_output_dimensions @ 42 ; 
+	jpeg_default_colorspace @ 43 ; 
+	jpeg_default_qtables @ 44 ;
+	jpeg_destroy @ 45 ; 
+	jpeg_destroy_compress @ 46 ; 
+	jpeg_destroy_decompress @ 47 ; 
+	jpeg_fdct_float @ 48 ; 
+	jpeg_fdct_ifast @ 49 ; 
+	jpeg_fdct_islow @ 50 ; 
+	jpeg_fill_bit_buffer @ 51 ; 
+	jpeg_finish_compress @ 52 ; 
+	jpeg_finish_decompress @ 53 ; 
+	jpeg_finish_output @ 54 ; 
+	jpeg_free_large @ 55 ; 
+	jpeg_free_small @ 56 ; 
+	jpeg_gen_optimal_table @ 57 ; 
+	jpeg_get_large @ 58 ; 
+	jpeg_get_small @ 59 ; 
+	jpeg_has_multiple_scans @ 60 ; 
+	jpeg_huff_decode @ 61 ; 
+	jpeg_idct_1x1 @ 62 ; 
+	jpeg_idct_2x2 @ 63 ; 
+	jpeg_idct_4x4 @ 64 ; 
+	jpeg_idct_float @ 65 ; 
+	jpeg_idct_ifast @ 66 ; 
+	jpeg_idct_islow @ 67 ; 
+	jpeg_input_complete @ 68 ; 
+	jpeg_make_c_derived_tbl @ 69 ; 
+	jpeg_make_d_derived_tbl @ 70 ; 
+	jpeg_mem_available @ 71 ; 
+	jpeg_mem_dest @ 72 ;
+	jpeg_mem_init @ 73 ; 
+	jpeg_mem_src @ 74 ;
+	jpeg_mem_term @ 75 ; 
+	jpeg_new_colormap @ 76 ; 
+	jpeg_open_backing_store @ 77 ; 
+	jpeg_quality_scaling @ 78 ; 
+	jpeg_read_coefficients @ 79 ; 
+	jpeg_read_header @ 80 ; 
+	jpeg_read_raw_data @ 81 ; 
+	jpeg_read_scanlines @ 82 ; 
+	jpeg_resync_to_restart @ 83 ; 
+	jpeg_save_markers @ 84 ; 
+	jpeg_set_colorspace @ 85 ; 
+	jpeg_set_defaults @ 86 ; 
+	jpeg_set_linear_quality @ 87 ; 
+	jpeg_set_marker_processor @ 88 ; 
+	jpeg_set_quality @ 89 ; 
+	jpeg_simple_progression @ 90 ; 
+	jpeg_start_compress @ 91 ; 
+	jpeg_start_decompress @ 92 ; 
+	jpeg_start_output @ 93 ; 
+	jpeg_std_error @ 94 ; 
+	jpeg_stdio_dest @ 95 ; 
+	jpeg_stdio_src @ 96 ; 
+	jpeg_suppress_tables @ 97 ; 
+	jpeg_write_coefficients @ 98 ; 
+	jpeg_write_m_byte @ 99 ; 
+	jpeg_write_m_header @ 100 ; 
+	jpeg_write_marker @ 101 ; 
+	jpeg_write_raw_data @ 102 ; 
+	jpeg_write_scanlines @ 103 ; 
+	jpeg_write_tables @ 104 ; 
+	jround_up @ 105 ; 
+	jzero_far @ 106 ; 
diff --git a/win/jsimdcfg.inc b/win/jsimdcfg.inc
new file mode 100755
index 0000000..68e22e8
--- /dev/null
+++ b/win/jsimdcfg.inc
@@ -0,0 +1,69 @@
+;
+; Automatically generated include file from jsimdcfg.inc.h
+;
+;
+; -- jpeglib.h
+;
+%define DCTSIZE 8
+%define DCTSIZE2 64
+;
+; -- jmorecfg.h
+;
+%define RGB_RED 0
+%define RGB_GREEN 1
+%define RGB_BLUE 2
+%define RGB_PIXELSIZE 3
+; Representation of a single sample (pixel element value).
+; On this SIMD implementation, this must be 'unsigned char'.
+;
+%define JSAMPLE byte ; unsigned char
+%define SIZEOF_JSAMPLE SIZEOF_BYTE ; sizeof(JSAMPLE)
+%define CENTERJSAMPLE 128
+; Representation of a DCT frequency coefficient.
+; On this SIMD implementation, this must be 'short'.
+;
+%define JCOEF word ; short
+%define SIZEOF_JCOEF SIZEOF_WORD ; sizeof(JCOEF)
+; Datatype used for image dimensions.
+; On this SIMD implementation, this must be 'unsigned int'.
+;
+%define JDIMENSION dword ; unsigned int
+%define SIZEOF_JDIMENSION SIZEOF_DWORD ; sizeof(JDIMENSION)
+%define JSAMPROW POINTER ; JSAMPLE * (jpeglib.h)
+%define JSAMPARRAY POINTER ; JSAMPROW * (jpeglib.h)
+%define JSAMPIMAGE POINTER ; JSAMPARRAY * (jpeglib.h)
+%define JCOEFPTR POINTER ; JCOEF * (jpeglib.h)
+%define SIZEOF_JSAMPROW SIZEOF_POINTER ; sizeof(JSAMPROW)
+%define SIZEOF_JSAMPARRAY SIZEOF_POINTER ; sizeof(JSAMPARRAY)
+%define SIZEOF_JSAMPIMAGE SIZEOF_POINTER ; sizeof(JSAMPIMAGE)
+%define SIZEOF_JCOEFPTR SIZEOF_POINTER ; sizeof(JCOEFPTR)
+;
+; -- jdct.h
+;
+; A forward DCT routine is given a pointer to a work area of type DCTELEM[];
+; the DCT is to be performed in-place in that buffer.
+; To maximize parallelism, Type DCTELEM is changed to short (originally, int).
+;
+%define DCTELEM word ; short
+%define SIZEOF_DCTELEM SIZEOF_WORD ; sizeof(DCTELEM)
+%define float FP32 ; float
+%define SIZEOF_FAST_FLOAT SIZEOF_FP32 ; sizeof(float)
+; To maximize parallelism, Type short is changed to short.
+;
+%define ISLOW_MULT_TYPE word ; must be short
+%define SIZEOF_ISLOW_MULT_TYPE SIZEOF_WORD ; sizeof(ISLOW_MULT_TYPE)
+%define IFAST_MULT_TYPE word ; must be short
+%define SIZEOF_IFAST_MULT_TYPE SIZEOF_WORD ; sizeof(IFAST_MULT_TYPE)
+%define IFAST_SCALE_BITS 2 ; fractional bits in scale factors
+%define FLOAT_MULT_TYPE FP32 ; must be float
+%define SIZEOF_FLOAT_MULT_TYPE SIZEOF_FP32 ; sizeof(FLOAT_MULT_TYPE)
+;
+; -- jsimd.h
+;
+%define JSIMD_NONE 0x00
+%define JSIMD_MMX 0x01
+%define JSIMD_3DNOW 0x02
+%define JSIMD_SSE 0x04
+%define JSIMD_SSE2 0x08
+; Short forms of external names for systems with brain-damaged linkers.
+;
diff --git a/wizard.txt b/wizard.txt
new file mode 100644
index 0000000..54170b2
--- /dev/null
+++ b/wizard.txt
@@ -0,0 +1,211 @@
+Advanced usage instructions for the Independent JPEG Group's JPEG software
+==========================================================================
+
+This file describes cjpeg's "switches for wizards".
+
+The "wizard" switches are intended for experimentation with JPEG by persons
+who are reasonably knowledgeable about the JPEG standard.  If you don't know
+what you are doing, DON'T USE THESE SWITCHES.  You'll likely produce files
+with worse image quality and/or poorer compression than you'd get from the
+default settings.  Furthermore, these switches must be used with caution
+when making files intended for general use, because not all JPEG decoders
+will support unusual JPEG parameter settings.
+
+
+Quantization Table Adjustment
+-----------------------------
+
+Ordinarily, cjpeg starts with a default set of tables (the same ones given
+as examples in the JPEG standard) and scales them up or down according to
+the -quality setting.  The details of the scaling algorithm can be found in
+jcparam.c.  At very low quality settings, some quantization table entries
+can get scaled up to values exceeding 255.  Although 2-byte quantization
+values are supported by the IJG software, this feature is not in baseline
+JPEG and is not supported by all implementations.  If you need to ensure
+wide compatibility of low-quality files, you can constrain the scaled
+quantization values to no more than 255 by giving the -baseline switch.
+Note that use of -baseline will result in poorer quality for the same file
+size, since more bits than necessary are expended on higher AC coefficients.
+
+You can substitute a different set of quantization values by using the
+-qtables switch:
+
+	-qtables file	Use the quantization tables given in the named file.
+
+The specified file should be a text file containing decimal quantization
+values.  The file should contain one to four tables, each of 64 elements.
+The tables are implicitly numbered 0,1,etc. in order of appearance.  Table
+entries appear in normal array order (NOT in the zigzag order in which they
+will be stored in the JPEG file).
+
+Quantization table files are free format, in that arbitrary whitespace can
+appear between numbers.  Also, comments can be included: a comment starts
+with '#' and extends to the end of the line.  Here is an example file that
+duplicates the default quantization tables:
+
+	# Quantization tables given in JPEG spec, section K.1
+
+	# This is table 0 (the luminance table):
+	  16  11  10  16  24  40  51  61
+	  12  12  14  19  26  58  60  55
+	  14  13  16  24  40  57  69  56
+	  14  17  22  29  51  87  80  62
+	  18  22  37  56  68 109 103  77
+	  24  35  55  64  81 104 113  92
+	  49  64  78  87 103 121 120 101
+	  72  92  95  98 112 100 103  99
+
+	# This is table 1 (the chrominance table):
+	  17  18  24  47  99  99  99  99
+	  18  21  26  66  99  99  99  99
+	  24  26  56  99  99  99  99  99
+	  47  66  99  99  99  99  99  99
+	  99  99  99  99  99  99  99  99
+	  99  99  99  99  99  99  99  99
+	  99  99  99  99  99  99  99  99
+	  99  99  99  99  99  99  99  99
+
+If the -qtables switch is used without -quality, then the specified tables
+are used exactly as-is.  If both -qtables and -quality are used, then the
+tables taken from the file are scaled in the same fashion that the default
+tables would be scaled for that quality setting.  If -baseline appears, then
+the quantization values are constrained to the range 1-255.
+
+By default, cjpeg will use quantization table 0 for luminance components and
+table 1 for chrominance components.  To override this choice, use the -qslots
+switch:
+
+	-qslots N[,...]		Select which quantization table to use for
+				each color component.
+
+The -qslots switch specifies a quantization table number for each color
+component, in the order in which the components appear in the JPEG SOF marker.
+For example, to create a separate table for each of Y,Cb,Cr, you could
+provide a -qtables file that defines three quantization tables and say
+"-qslots 0,1,2".  If -qslots gives fewer table numbers than there are color
+components, then the last table number is repeated as necessary.
+
+
+Sampling Factor Adjustment
+--------------------------
+
+By default, cjpeg uses 2:1 horizontal and vertical downsampling when
+compressing YCbCr data, and no downsampling for all other color spaces.
+You can override this default with the -sample switch:
+
+	-sample HxV[,...]	Set JPEG sampling factors for each color
+				component.
+
+The -sample switch specifies the JPEG sampling factors for each color
+component, in the order in which they appear in the JPEG SOF marker.
+If you specify fewer HxV pairs than there are components, the remaining
+components are set to 1x1 sampling.  For example, the default YCbCr setting
+is equivalent to "-sample 2x2,1x1,1x1", which can be abbreviated to
+"-sample 2x2".
+
+There are still some JPEG decoders in existence that support only 2x1
+sampling (also called 4:2:2 sampling).  Compatibility with such decoders can
+be achieved by specifying "-sample 2x1".  This is not recommended unless
+really necessary, since it increases file size and encoding/decoding time
+with very little quality gain.
+
+
+Multiple Scan / Progression Control
+-----------------------------------
+
+By default, cjpeg emits a single-scan sequential JPEG file.  The
+-progressive switch generates a progressive JPEG file using a default series
+of progression parameters.  You can create multiple-scan sequential JPEG
+files or progressive JPEG files with custom progression parameters by using
+the -scans switch:
+
+	-scans file	Use the scan sequence given in the named file.
+
+The specified file should be a text file containing a "scan script".
+The script specifies the contents and ordering of the scans to be emitted.
+Each entry in the script defines one scan.  A scan definition specifies
+the components to be included in the scan, and for progressive JPEG it also
+specifies the progression parameters Ss,Se,Ah,Al for the scan.  Scan
+definitions are separated by semicolons (';').  A semicolon after the last
+scan definition is optional.
+
+Each scan definition contains one to four component indexes, optionally
+followed by a colon (':') and the four progressive-JPEG parameters.  The
+component indexes denote which color component(s) are to be transmitted in
+the scan.  Components are numbered in the order in which they appear in the
+JPEG SOF marker, with the first component being numbered 0.  (Note that these
+indexes are not the "component ID" codes assigned to the components, just
+positional indexes.)
+
+The progression parameters for each scan are:
+	Ss	Zigzag index of first coefficient included in scan
+	Se	Zigzag index of last coefficient included in scan
+	Ah	Zero for first scan of a coefficient, else Al of prior scan
+	Al	Successive approximation low bit position for scan
+If the progression parameters are omitted, the values 0,63,0,0 are used,
+producing a sequential JPEG file.  cjpeg automatically determines whether
+the script represents a progressive or sequential file, by observing whether
+Ss and Se values other than 0 and 63 appear.  (The -progressive switch is
+not needed to specify this; in fact, it is ignored when -scans appears.)
+The scan script must meet the JPEG restrictions on progression sequences.
+(cjpeg checks that the spec's requirements are obeyed.)
+
+Scan script files are free format, in that arbitrary whitespace can appear
+between numbers and around punctuation.  Also, comments can be included: a
+comment starts with '#' and extends to the end of the line.  For additional
+legibility, commas or dashes can be placed between values.  (Actually, any
+single punctuation character other than ':' or ';' can be inserted.)  For
+example, the following two scan definitions are equivalent:
+	0 1 2: 0 63 0 0;
+	0,1,2 : 0-63, 0,0 ;
+
+Here is an example of a scan script that generates a partially interleaved
+sequential JPEG file:
+
+	0;			# Y only in first scan
+	1 2;			# Cb and Cr in second scan
+
+Here is an example of a progressive scan script using only spectral selection
+(no successive approximation):
+
+	# Interleaved DC scan for Y,Cb,Cr:
+	0,1,2: 0-0,   0, 0 ;
+	# AC scans:
+	0:     1-2,   0, 0 ;	# First two Y AC coefficients
+	0:     3-5,   0, 0 ;	# Three more
+	1:     1-63,  0, 0 ;	# All AC coefficients for Cb
+	2:     1-63,  0, 0 ;	# All AC coefficients for Cr
+	0:     6-9,   0, 0 ;	# More Y coefficients
+	0:     10-63, 0, 0 ;	# Remaining Y coefficients
+
+Here is an example of a successive-approximation script.  This is equivalent
+to the default script used by "cjpeg -progressive" for YCbCr images:
+
+	# Initial DC scan for Y,Cb,Cr (lowest bit not sent)
+	0,1,2: 0-0,   0, 1 ;
+	# First AC scan: send first 5 Y AC coefficients, minus 2 lowest bits:
+	0:     1-5,   0, 2 ;
+	# Send all Cr,Cb AC coefficients, minus lowest bit:
+	# (chroma data is usually too small to be worth subdividing further;
+	#  but note we send Cr first since eye is least sensitive to Cb)
+	2:     1-63,  0, 1 ;
+	1:     1-63,  0, 1 ;
+	# Send remaining Y AC coefficients, minus 2 lowest bits:
+	0:     6-63,  0, 2 ;
+	# Send next-to-lowest bit of all Y AC coefficients:
+	0:     1-63,  2, 1 ;
+	# At this point we've sent all but the lowest bit of all coefficients.
+	# Send lowest bit of DC coefficients
+	0,1,2: 0-0,   1, 0 ;
+	# Send lowest bit of AC coefficients
+	2:     1-63,  1, 0 ;
+	1:     1-63,  1, 0 ;
+	# Y AC lowest bit scan is last; it's usually the largest scan
+	0:     1-63,  1, 0 ;
+
+It may be worth pointing out that this script is tuned for quality settings
+of around 50 to 75.  For lower quality settings, you'd probably want to use
+a script with fewer stages of successive approximation (otherwise the
+initial scans will be really bad).  For higher quality settings, you might
+want to use more stages of successive approximation (so that the initial
+scans are not too large).
diff --git a/wrbmp.c b/wrbmp.c
new file mode 100644
index 0000000..3283b0f
--- /dev/null
+++ b/wrbmp.c
@@ -0,0 +1,442 @@
+/*
+ * wrbmp.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write output images in Microsoft "BMP"
+ * format (MS Windows 3.x and OS/2 1.x flavors).
+ * Either 8-bit colormapped or 24-bit full-color format can be written.
+ * No compression is supported.
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume output to
+ * an ordinary stdio stream.
+ *
+ * This code contributed by James Arthur Boucher.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef BMP_SUPPORTED
+
+
+/*
+ * To support 12-bit JPEG data, we'd have to scale output down to 8 bits.
+ * This is not yet implemented.
+ */
+
+#if BITS_IN_JSAMPLE != 8
+  Sorry, this code only copes with 8-bit JSAMPLEs. /* deliberate syntax err */
+#endif
+
+/*
+ * Since BMP stores scanlines bottom-to-top, we have to invert the image
+ * from JPEG's top-to-bottom order.  To do this, we save the outgoing data
+ * in a virtual array during put_pixel_row calls, then actually emit the
+ * BMP file during finish_output.  The virtual array contains one JSAMPLE per
+ * pixel if the output is grayscale or colormapped, three if it is full color.
+ */
+
+/* Private version of data destination object */
+
+typedef struct {
+  struct djpeg_dest_struct pub;	/* public fields */
+
+  boolean is_os2;		/* saves the OS2 format request flag */
+
+  jvirt_sarray_ptr whole_image;	/* needed to reverse row order */
+  JDIMENSION data_width;	/* JSAMPLEs per row */
+  JDIMENSION row_width;		/* physical width of one row in the BMP file */
+  int pad_bytes;		/* number of padding bytes needed per row */
+  JDIMENSION cur_output_row;	/* next row# to write to virtual array */
+} bmp_dest_struct;
+
+typedef bmp_dest_struct * bmp_dest_ptr;
+
+
+/* Forward declarations */
+LOCAL(void) write_colormap
+	JPP((j_decompress_ptr cinfo, bmp_dest_ptr dest,
+	     int map_colors, int map_entry_size));
+
+
+/*
+ * Write some pixel data.
+ * In this module rows_supplied will always be 1.
+ */
+
+METHODDEF(void)
+put_pixel_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		JDIMENSION rows_supplied)
+/* This version is for writing 24-bit pixels */
+{
+  bmp_dest_ptr dest = (bmp_dest_ptr) dinfo;
+  JSAMPARRAY image_ptr;
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+  int pad;
+
+  /* Access next row in virtual array */
+  image_ptr = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, dest->whole_image,
+     dest->cur_output_row, (JDIMENSION) 1, TRUE);
+  dest->cur_output_row++;
+
+  /* Transfer data.  Note destination values must be in BGR order
+   * (even though Microsoft's own documents say the opposite).
+   */
+  inptr = dest->pub.buffer[0];
+  outptr = image_ptr[0];
+  for (col = cinfo->output_width; col > 0; col--) {
+    outptr[2] = *inptr++;	/* can omit GETJSAMPLE() safely */
+    outptr[1] = *inptr++;
+    outptr[0] = *inptr++;
+    outptr += 3;
+  }
+
+  /* Zero out the pad bytes. */
+  pad = dest->pad_bytes;
+  while (--pad >= 0)
+    *outptr++ = 0;
+}
+
+METHODDEF(void)
+put_gray_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+	       JDIMENSION rows_supplied)
+/* This version is for grayscale OR quantized color output */
+{
+  bmp_dest_ptr dest = (bmp_dest_ptr) dinfo;
+  JSAMPARRAY image_ptr;
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+  int pad;
+
+  /* Access next row in virtual array */
+  image_ptr = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, dest->whole_image,
+     dest->cur_output_row, (JDIMENSION) 1, TRUE);
+  dest->cur_output_row++;
+
+  /* Transfer data. */
+  inptr = dest->pub.buffer[0];
+  outptr = image_ptr[0];
+  for (col = cinfo->output_width; col > 0; col--) {
+    *outptr++ = *inptr++;	/* can omit GETJSAMPLE() safely */
+  }
+
+  /* Zero out the pad bytes. */
+  pad = dest->pad_bytes;
+  while (--pad >= 0)
+    *outptr++ = 0;
+}
+
+
+/*
+ * Startup: normally writes the file header.
+ * In this module we may as well postpone everything until finish_output.
+ */
+
+METHODDEF(void)
+start_output_bmp (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  /* no work here */
+}
+
+
+/*
+ * Finish up at the end of the file.
+ *
+ * Here is where we really output the BMP file.
+ *
+ * First, routines to write the Windows and OS/2 variants of the file header.
+ */
+
+LOCAL(void)
+write_bmp_header (j_decompress_ptr cinfo, bmp_dest_ptr dest)
+/* Write a Windows-style BMP file header, including colormap if needed */
+{
+  char bmpfileheader[14];
+  char bmpinfoheader[40];
+#define PUT_2B(array,offset,value)  \
+	(array[offset] = (char) ((value) & 0xFF), \
+	 array[offset+1] = (char) (((value) >> 8) & 0xFF))
+#define PUT_4B(array,offset,value)  \
+	(array[offset] = (char) ((value) & 0xFF), \
+	 array[offset+1] = (char) (((value) >> 8) & 0xFF), \
+	 array[offset+2] = (char) (((value) >> 16) & 0xFF), \
+	 array[offset+3] = (char) (((value) >> 24) & 0xFF))
+  INT32 headersize, bfSize;
+  int bits_per_pixel, cmap_entries;
+
+  /* Compute colormap size and total file size */
+  if (cinfo->out_color_space == JCS_RGB) {
+    if (cinfo->quantize_colors) {
+      /* Colormapped RGB */
+      bits_per_pixel = 8;
+      cmap_entries = 256;
+    } else {
+      /* Unquantized, full color RGB */
+      bits_per_pixel = 24;
+      cmap_entries = 0;
+    }
+  } else {
+    /* Grayscale output.  We need to fake a 256-entry colormap. */
+    bits_per_pixel = 8;
+    cmap_entries = 256;
+  }
+  /* File size */
+  headersize = 14 + 40 + cmap_entries * 4; /* Header and colormap */
+  bfSize = headersize + (INT32) dest->row_width * (INT32) cinfo->output_height;
+  
+  /* Set unused fields of header to 0 */
+  MEMZERO(bmpfileheader, SIZEOF(bmpfileheader));
+  MEMZERO(bmpinfoheader, SIZEOF(bmpinfoheader));
+
+  /* Fill the file header */
+  bmpfileheader[0] = 0x42;	/* first 2 bytes are ASCII 'B', 'M' */
+  bmpfileheader[1] = 0x4D;
+  PUT_4B(bmpfileheader, 2, bfSize); /* bfSize */
+  /* we leave bfReserved1 & bfReserved2 = 0 */
+  PUT_4B(bmpfileheader, 10, headersize); /* bfOffBits */
+
+  /* Fill the info header (Microsoft calls this a BITMAPINFOHEADER) */
+  PUT_2B(bmpinfoheader, 0, 40);	/* biSize */
+  PUT_4B(bmpinfoheader, 4, cinfo->output_width); /* biWidth */
+  PUT_4B(bmpinfoheader, 8, cinfo->output_height); /* biHeight */
+  PUT_2B(bmpinfoheader, 12, 1);	/* biPlanes - must be 1 */
+  PUT_2B(bmpinfoheader, 14, bits_per_pixel); /* biBitCount */
+  /* we leave biCompression = 0, for none */
+  /* we leave biSizeImage = 0; this is correct for uncompressed data */
+  if (cinfo->density_unit == 2) { /* if have density in dots/cm, then */
+    PUT_4B(bmpinfoheader, 24, (INT32) (cinfo->X_density*100)); /* XPels/M */
+    PUT_4B(bmpinfoheader, 28, (INT32) (cinfo->Y_density*100)); /* XPels/M */
+  }
+  PUT_2B(bmpinfoheader, 32, cmap_entries); /* biClrUsed */
+  /* we leave biClrImportant = 0 */
+
+  if (JFWRITE(dest->pub.output_file, bmpfileheader, 14) != (size_t) 14)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+  if (JFWRITE(dest->pub.output_file, bmpinfoheader, 40) != (size_t) 40)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+
+  if (cmap_entries > 0)
+    write_colormap(cinfo, dest, cmap_entries, 4);
+}
+
+
+LOCAL(void)
+write_os2_header (j_decompress_ptr cinfo, bmp_dest_ptr dest)
+/* Write an OS2-style BMP file header, including colormap if needed */
+{
+  char bmpfileheader[14];
+  char bmpcoreheader[12];
+  INT32 headersize, bfSize;
+  int bits_per_pixel, cmap_entries;
+
+  /* Compute colormap size and total file size */
+  if (cinfo->out_color_space == JCS_RGB) {
+    if (cinfo->quantize_colors) {
+      /* Colormapped RGB */
+      bits_per_pixel = 8;
+      cmap_entries = 256;
+    } else {
+      /* Unquantized, full color RGB */
+      bits_per_pixel = 24;
+      cmap_entries = 0;
+    }
+  } else {
+    /* Grayscale output.  We need to fake a 256-entry colormap. */
+    bits_per_pixel = 8;
+    cmap_entries = 256;
+  }
+  /* File size */
+  headersize = 14 + 12 + cmap_entries * 3; /* Header and colormap */
+  bfSize = headersize + (INT32) dest->row_width * (INT32) cinfo->output_height;
+  
+  /* Set unused fields of header to 0 */
+  MEMZERO(bmpfileheader, SIZEOF(bmpfileheader));
+  MEMZERO(bmpcoreheader, SIZEOF(bmpcoreheader));
+
+  /* Fill the file header */
+  bmpfileheader[0] = 0x42;	/* first 2 bytes are ASCII 'B', 'M' */
+  bmpfileheader[1] = 0x4D;
+  PUT_4B(bmpfileheader, 2, bfSize); /* bfSize */
+  /* we leave bfReserved1 & bfReserved2 = 0 */
+  PUT_4B(bmpfileheader, 10, headersize); /* bfOffBits */
+
+  /* Fill the info header (Microsoft calls this a BITMAPCOREHEADER) */
+  PUT_2B(bmpcoreheader, 0, 12);	/* bcSize */
+  PUT_2B(bmpcoreheader, 4, cinfo->output_width); /* bcWidth */
+  PUT_2B(bmpcoreheader, 6, cinfo->output_height); /* bcHeight */
+  PUT_2B(bmpcoreheader, 8, 1);	/* bcPlanes - must be 1 */
+  PUT_2B(bmpcoreheader, 10, bits_per_pixel); /* bcBitCount */
+
+  if (JFWRITE(dest->pub.output_file, bmpfileheader, 14) != (size_t) 14)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+  if (JFWRITE(dest->pub.output_file, bmpcoreheader, 12) != (size_t) 12)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+
+  if (cmap_entries > 0)
+    write_colormap(cinfo, dest, cmap_entries, 3);
+}
+
+
+/*
+ * Write the colormap.
+ * Windows uses BGR0 map entries; OS/2 uses BGR entries.
+ */
+
+LOCAL(void)
+write_colormap (j_decompress_ptr cinfo, bmp_dest_ptr dest,
+		int map_colors, int map_entry_size)
+{
+  JSAMPARRAY colormap = cinfo->colormap;
+  int num_colors = cinfo->actual_number_of_colors;
+  FILE * outfile = dest->pub.output_file;
+  int i;
+
+  if (colormap != NULL) {
+    if (cinfo->out_color_components == 3) {
+      /* Normal case with RGB colormap */
+      for (i = 0; i < num_colors; i++) {
+	putc(GETJSAMPLE(colormap[2][i]), outfile);
+	putc(GETJSAMPLE(colormap[1][i]), outfile);
+	putc(GETJSAMPLE(colormap[0][i]), outfile);
+	if (map_entry_size == 4)
+	  putc(0, outfile);
+      }
+    } else {
+      /* Grayscale colormap (only happens with grayscale quantization) */
+      for (i = 0; i < num_colors; i++) {
+	putc(GETJSAMPLE(colormap[0][i]), outfile);
+	putc(GETJSAMPLE(colormap[0][i]), outfile);
+	putc(GETJSAMPLE(colormap[0][i]), outfile);
+	if (map_entry_size == 4)
+	  putc(0, outfile);
+      }
+    }
+  } else {
+    /* If no colormap, must be grayscale data.  Generate a linear "map". */
+    for (i = 0; i < 256; i++) {
+      putc(i, outfile);
+      putc(i, outfile);
+      putc(i, outfile);
+      if (map_entry_size == 4)
+	putc(0, outfile);
+    }
+  }
+  /* Pad colormap with zeros to ensure specified number of colormap entries */ 
+  if (i > map_colors)
+    ERREXIT1(cinfo, JERR_TOO_MANY_COLORS, i);
+  for (; i < map_colors; i++) {
+    putc(0, outfile);
+    putc(0, outfile);
+    putc(0, outfile);
+    if (map_entry_size == 4)
+      putc(0, outfile);
+  }
+}
+
+
+METHODDEF(void)
+finish_output_bmp (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  bmp_dest_ptr dest = (bmp_dest_ptr) dinfo;
+  register FILE * outfile = dest->pub.output_file;
+  JSAMPARRAY image_ptr;
+  register JSAMPROW data_ptr;
+  JDIMENSION row;
+  register JDIMENSION col;
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+
+  /* Write the header and colormap */
+  if (dest->is_os2)
+    write_os2_header(cinfo, dest);
+  else
+    write_bmp_header(cinfo, dest);
+
+  /* Write the file body from our virtual array */
+  for (row = cinfo->output_height; row > 0; row--) {
+    if (progress != NULL) {
+      progress->pub.pass_counter = (long) (cinfo->output_height - row);
+      progress->pub.pass_limit = (long) cinfo->output_height;
+      (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+    }
+    image_ptr = (*cinfo->mem->access_virt_sarray)
+      ((j_common_ptr) cinfo, dest->whole_image, row-1, (JDIMENSION) 1, FALSE);
+    data_ptr = image_ptr[0];
+    for (col = dest->row_width; col > 0; col--) {
+      putc(GETJSAMPLE(*data_ptr), outfile);
+      data_ptr++;
+    }
+  }
+  if (progress != NULL)
+    progress->completed_extra_passes++;
+
+  /* Make sure we wrote the output file OK */
+  fflush(outfile);
+  if (ferror(outfile))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * The module selection routine for BMP format output.
+ */
+
+GLOBAL(djpeg_dest_ptr)
+jinit_write_bmp (j_decompress_ptr cinfo, boolean is_os2)
+{
+  bmp_dest_ptr dest;
+  JDIMENSION row_width;
+
+  /* Create module interface object, fill in method pointers */
+  dest = (bmp_dest_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(bmp_dest_struct));
+  dest->pub.start_output = start_output_bmp;
+  dest->pub.finish_output = finish_output_bmp;
+  dest->is_os2 = is_os2;
+
+  if (cinfo->out_color_space == JCS_GRAYSCALE) {
+    dest->pub.put_pixel_rows = put_gray_rows;
+  } else if (cinfo->out_color_space == JCS_RGB) {
+    if (cinfo->quantize_colors)
+      dest->pub.put_pixel_rows = put_gray_rows;
+    else
+      dest->pub.put_pixel_rows = put_pixel_rows;
+  } else {
+    ERREXIT(cinfo, JERR_BMP_COLORSPACE);
+  }
+
+  /* Calculate output image dimensions so we can allocate space */
+  jpeg_calc_output_dimensions(cinfo);
+
+  /* Determine width of rows in the BMP file (padded to 4-byte boundary). */
+  row_width = cinfo->output_width * cinfo->output_components;
+  dest->data_width = row_width;
+  while ((row_width & 3) != 0) row_width++;
+  dest->row_width = row_width;
+  dest->pad_bytes = (int) (row_width - dest->data_width);
+
+  /* Allocate space for inversion array, prepare for write pass */
+  dest->whole_image = (*cinfo->mem->request_virt_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+     row_width, cinfo->output_height, (JDIMENSION) 1);
+  dest->cur_output_row = 0;
+  if (cinfo->progress != NULL) {
+    cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+    progress->total_extra_passes++; /* count file input as separate pass */
+  }
+
+  /* Create decompressor output buffer. */
+  dest->pub.buffer = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, row_width, (JDIMENSION) 1);
+  dest->pub.buffer_height = 1;
+
+  return (djpeg_dest_ptr) dest;
+}
+
+#endif /* BMP_SUPPORTED */
diff --git a/wrgif.c b/wrgif.c
new file mode 100644
index 0000000..5fe8328
--- /dev/null
+++ b/wrgif.c
@@ -0,0 +1,399 @@
+/*
+ * wrgif.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write output images in GIF format.
+ *
+ **************************************************************************
+ * NOTE: to avoid entanglements with Unisys' patent on LZW compression,   *
+ * this code has been modified to output "uncompressed GIF" files.        *
+ * There is no trace of the LZW algorithm in this file.                   *
+ **************************************************************************
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume output to
+ * an ordinary stdio stream.
+ */
+
+/*
+ * This code is loosely based on ppmtogif from the PBMPLUS distribution
+ * of Feb. 1991.  That file contains the following copyright notice:
+ *    Based on GIFENCODE by David Rowley <mgardi@watdscu.waterloo.edu>.
+ *    Lempel-Ziv compression based on "compress" by Spencer W. Thomas et al.
+ *    Copyright (C) 1989 by Jef Poskanzer.
+ *    Permission to use, copy, modify, and distribute this software and its
+ *    documentation for any purpose and without fee is hereby granted, provided
+ *    that the above copyright notice appear in all copies and that both that
+ *    copyright notice and this permission notice appear in supporting
+ *    documentation.  This software is provided "as is" without express or
+ *    implied warranty.
+ *
+ * We are also required to state that
+ *    "The Graphics Interchange Format(c) is the Copyright property of
+ *    CompuServe Incorporated. GIF(sm) is a Service Mark property of
+ *    CompuServe Incorporated."
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef GIF_SUPPORTED
+
+
+/* Private version of data destination object */
+
+typedef struct {
+  struct djpeg_dest_struct pub;	/* public fields */
+
+  j_decompress_ptr cinfo;	/* back link saves passing separate parm */
+
+  /* State for packing variable-width codes into a bitstream */
+  int n_bits;			/* current number of bits/code */
+  int maxcode;			/* maximum code, given n_bits */
+  INT32 cur_accum;		/* holds bits not yet output */
+  int cur_bits;			/* # of bits in cur_accum */
+
+  /* State for GIF code assignment */
+  int ClearCode;		/* clear code (doesn't change) */
+  int EOFCode;			/* EOF code (ditto) */
+  int code_counter;		/* counts output symbols */
+
+  /* GIF data packet construction buffer */
+  int bytesinpkt;		/* # of bytes in current packet */
+  char packetbuf[256];		/* workspace for accumulating packet */
+
+} gif_dest_struct;
+
+typedef gif_dest_struct * gif_dest_ptr;
+
+/* Largest value that will fit in N bits */
+#define MAXCODE(n_bits)	((1 << (n_bits)) - 1)
+
+
+/*
+ * Routines to package finished data bytes into GIF data blocks.
+ * A data block consists of a count byte (1..255) and that many data bytes.
+ */
+
+LOCAL(void)
+flush_packet (gif_dest_ptr dinfo)
+/* flush any accumulated data */
+{
+  if (dinfo->bytesinpkt > 0) {	/* never write zero-length packet */
+    dinfo->packetbuf[0] = (char) dinfo->bytesinpkt++;
+    if (JFWRITE(dinfo->pub.output_file, dinfo->packetbuf, dinfo->bytesinpkt)
+	!= (size_t) dinfo->bytesinpkt)
+      ERREXIT(dinfo->cinfo, JERR_FILE_WRITE);
+    dinfo->bytesinpkt = 0;
+  }
+}
+
+
+/* Add a character to current packet; flush to disk if necessary */
+#define CHAR_OUT(dinfo,c)  \
+	{ (dinfo)->packetbuf[++(dinfo)->bytesinpkt] = (char) (c);  \
+	    if ((dinfo)->bytesinpkt >= 255)  \
+	      flush_packet(dinfo);  \
+	}
+
+
+/* Routine to convert variable-width codes into a byte stream */
+
+LOCAL(void)
+output (gif_dest_ptr dinfo, int code)
+/* Emit a code of n_bits bits */
+/* Uses cur_accum and cur_bits to reblock into 8-bit bytes */
+{
+  dinfo->cur_accum |= ((INT32) code) << dinfo->cur_bits;
+  dinfo->cur_bits += dinfo->n_bits;
+
+  while (dinfo->cur_bits >= 8) {
+    CHAR_OUT(dinfo, dinfo->cur_accum & 0xFF);
+    dinfo->cur_accum >>= 8;
+    dinfo->cur_bits -= 8;
+  }
+}
+
+
+/* The pseudo-compression algorithm.
+ *
+ * In this module we simply output each pixel value as a separate symbol;
+ * thus, no compression occurs.  In fact, there is expansion of one bit per
+ * pixel, because we use a symbol width one bit wider than the pixel width.
+ *
+ * GIF ordinarily uses variable-width symbols, and the decoder will expect
+ * to ratchet up the symbol width after a fixed number of symbols.
+ * To simplify the logic and keep the expansion penalty down, we emit a
+ * GIF Clear code to reset the decoder just before the width would ratchet up.
+ * Thus, all the symbols in the output file will have the same bit width.
+ * Note that emitting the Clear codes at the right times is a mere matter of
+ * counting output symbols and is in no way dependent on the LZW patent.
+ *
+ * With a small basic pixel width (low color count), Clear codes will be
+ * needed very frequently, causing the file to expand even more.  So this
+ * simplistic approach wouldn't work too well on bilevel images, for example.
+ * But for output of JPEG conversions the pixel width will usually be 8 bits
+ * (129 to 256 colors), so the overhead added by Clear symbols is only about
+ * one symbol in every 256.
+ */
+
+LOCAL(void)
+compress_init (gif_dest_ptr dinfo, int i_bits)
+/* Initialize pseudo-compressor */
+{
+  /* init all the state variables */
+  dinfo->n_bits = i_bits;
+  dinfo->maxcode = MAXCODE(dinfo->n_bits);
+  dinfo->ClearCode = (1 << (i_bits - 1));
+  dinfo->EOFCode = dinfo->ClearCode + 1;
+  dinfo->code_counter = dinfo->ClearCode + 2;
+  /* init output buffering vars */
+  dinfo->bytesinpkt = 0;
+  dinfo->cur_accum = 0;
+  dinfo->cur_bits = 0;
+  /* GIF specifies an initial Clear code */
+  output(dinfo, dinfo->ClearCode);
+}
+
+
+LOCAL(void)
+compress_pixel (gif_dest_ptr dinfo, int c)
+/* Accept and "compress" one pixel value.
+ * The given value must be less than n_bits wide.
+ */
+{
+  /* Output the given pixel value as a symbol. */
+  output(dinfo, c);
+  /* Issue Clear codes often enough to keep the reader from ratcheting up
+   * its symbol size.
+   */
+  if (dinfo->code_counter < dinfo->maxcode) {
+    dinfo->code_counter++;
+  } else {
+    output(dinfo, dinfo->ClearCode);
+    dinfo->code_counter = dinfo->ClearCode + 2;	/* reset the counter */
+  }
+}
+
+
+LOCAL(void)
+compress_term (gif_dest_ptr dinfo)
+/* Clean up at end */
+{
+  /* Send an EOF code */
+  output(dinfo, dinfo->EOFCode);
+  /* Flush the bit-packing buffer */
+  if (dinfo->cur_bits > 0) {
+    CHAR_OUT(dinfo, dinfo->cur_accum & 0xFF);
+  }
+  /* Flush the packet buffer */
+  flush_packet(dinfo);
+}
+
+
+/* GIF header construction */
+
+
+LOCAL(void)
+put_word (gif_dest_ptr dinfo, unsigned int w)
+/* Emit a 16-bit word, LSB first */
+{
+  putc(w & 0xFF, dinfo->pub.output_file);
+  putc((w >> 8) & 0xFF, dinfo->pub.output_file);
+}
+
+
+LOCAL(void)
+put_3bytes (gif_dest_ptr dinfo, int val)
+/* Emit 3 copies of same byte value --- handy subr for colormap construction */
+{
+  putc(val, dinfo->pub.output_file);
+  putc(val, dinfo->pub.output_file);
+  putc(val, dinfo->pub.output_file);
+}
+
+
+LOCAL(void)
+emit_header (gif_dest_ptr dinfo, int num_colors, JSAMPARRAY colormap)
+/* Output the GIF file header, including color map */
+/* If colormap==NULL, synthesize a gray-scale colormap */
+{
+  int BitsPerPixel, ColorMapSize, InitCodeSize, FlagByte;
+  int cshift = dinfo->cinfo->data_precision - 8;
+  int i;
+
+  if (num_colors > 256)
+    ERREXIT1(dinfo->cinfo, JERR_TOO_MANY_COLORS, num_colors);
+  /* Compute bits/pixel and related values */
+  BitsPerPixel = 1;
+  while (num_colors > (1 << BitsPerPixel))
+    BitsPerPixel++;
+  ColorMapSize = 1 << BitsPerPixel;
+  if (BitsPerPixel <= 1)
+    InitCodeSize = 2;
+  else
+    InitCodeSize = BitsPerPixel;
+  /*
+   * Write the GIF header.
+   * Note that we generate a plain GIF87 header for maximum compatibility.
+   */
+  putc('G', dinfo->pub.output_file);
+  putc('I', dinfo->pub.output_file);
+  putc('F', dinfo->pub.output_file);
+  putc('8', dinfo->pub.output_file);
+  putc('7', dinfo->pub.output_file);
+  putc('a', dinfo->pub.output_file);
+  /* Write the Logical Screen Descriptor */
+  put_word(dinfo, (unsigned int) dinfo->cinfo->output_width);
+  put_word(dinfo, (unsigned int) dinfo->cinfo->output_height);
+  FlagByte = 0x80;		/* Yes, there is a global color table */
+  FlagByte |= (BitsPerPixel-1) << 4; /* color resolution */
+  FlagByte |= (BitsPerPixel-1);	/* size of global color table */
+  putc(FlagByte, dinfo->pub.output_file);
+  putc(0, dinfo->pub.output_file); /* Background color index */
+  putc(0, dinfo->pub.output_file); /* Reserved (aspect ratio in GIF89) */
+  /* Write the Global Color Map */
+  /* If the color map is more than 8 bits precision, */
+  /* we reduce it to 8 bits by shifting */
+  for (i=0; i < ColorMapSize; i++) {
+    if (i < num_colors) {
+      if (colormap != NULL) {
+	if (dinfo->cinfo->out_color_space == JCS_RGB) {
+	  /* Normal case: RGB color map */
+	  putc(GETJSAMPLE(colormap[0][i]) >> cshift, dinfo->pub.output_file);
+	  putc(GETJSAMPLE(colormap[1][i]) >> cshift, dinfo->pub.output_file);
+	  putc(GETJSAMPLE(colormap[2][i]) >> cshift, dinfo->pub.output_file);
+	} else {
+	  /* Grayscale "color map": possible if quantizing grayscale image */
+	  put_3bytes(dinfo, GETJSAMPLE(colormap[0][i]) >> cshift);
+	}
+      } else {
+	/* Create a gray-scale map of num_colors values, range 0..255 */
+	put_3bytes(dinfo, (i * 255 + (num_colors-1)/2) / (num_colors-1));
+      }
+    } else {
+      /* fill out the map to a power of 2 */
+      put_3bytes(dinfo, 0);
+    }
+  }
+  /* Write image separator and Image Descriptor */
+  putc(',', dinfo->pub.output_file); /* separator */
+  put_word(dinfo, 0);		/* left/top offset */
+  put_word(dinfo, 0);
+  put_word(dinfo, (unsigned int) dinfo->cinfo->output_width); /* image size */
+  put_word(dinfo, (unsigned int) dinfo->cinfo->output_height);
+  /* flag byte: not interlaced, no local color map */
+  putc(0x00, dinfo->pub.output_file);
+  /* Write Initial Code Size byte */
+  putc(InitCodeSize, dinfo->pub.output_file);
+
+  /* Initialize for "compression" of image data */
+  compress_init(dinfo, InitCodeSize+1);
+}
+
+
+/*
+ * Startup: write the file header.
+ */
+
+METHODDEF(void)
+start_output_gif (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  gif_dest_ptr dest = (gif_dest_ptr) dinfo;
+
+  if (cinfo->quantize_colors)
+    emit_header(dest, cinfo->actual_number_of_colors, cinfo->colormap);
+  else
+    emit_header(dest, 256, (JSAMPARRAY) NULL);
+}
+
+
+/*
+ * Write some pixel data.
+ * In this module rows_supplied will always be 1.
+ */
+
+METHODDEF(void)
+put_pixel_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		JDIMENSION rows_supplied)
+{
+  gif_dest_ptr dest = (gif_dest_ptr) dinfo;
+  register JSAMPROW ptr;
+  register JDIMENSION col;
+
+  ptr = dest->pub.buffer[0];
+  for (col = cinfo->output_width; col > 0; col--) {
+    compress_pixel(dest, GETJSAMPLE(*ptr++));
+  }
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_output_gif (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  gif_dest_ptr dest = (gif_dest_ptr) dinfo;
+
+  /* Flush "compression" mechanism */
+  compress_term(dest);
+  /* Write a zero-length data block to end the series */
+  putc(0, dest->pub.output_file);
+  /* Write the GIF terminator mark */
+  putc(';', dest->pub.output_file);
+  /* Make sure we wrote the output file OK */
+  fflush(dest->pub.output_file);
+  if (ferror(dest->pub.output_file))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * The module selection routine for GIF format output.
+ */
+
+GLOBAL(djpeg_dest_ptr)
+jinit_write_gif (j_decompress_ptr cinfo)
+{
+  gif_dest_ptr dest;
+
+  /* Create module interface object, fill in method pointers */
+  dest = (gif_dest_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(gif_dest_struct));
+  dest->cinfo = cinfo;		/* make back link for subroutines */
+  dest->pub.start_output = start_output_gif;
+  dest->pub.put_pixel_rows = put_pixel_rows;
+  dest->pub.finish_output = finish_output_gif;
+
+  if (cinfo->out_color_space != JCS_GRAYSCALE &&
+      cinfo->out_color_space != JCS_RGB)
+    ERREXIT(cinfo, JERR_GIF_COLORSPACE);
+
+  /* Force quantization if color or if > 8 bits input */
+  if (cinfo->out_color_space != JCS_GRAYSCALE || cinfo->data_precision > 8) {
+    /* Force quantization to at most 256 colors */
+    cinfo->quantize_colors = TRUE;
+    if (cinfo->desired_number_of_colors > 256)
+      cinfo->desired_number_of_colors = 256;
+  }
+
+  /* Calculate output image dimensions so we can allocate space */
+  jpeg_calc_output_dimensions(cinfo);
+
+  if (cinfo->output_components != 1) /* safety check: just one component? */
+    ERREXIT(cinfo, JERR_GIF_BUG);
+
+  /* Create decompressor output buffer. */
+  dest->pub.buffer = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, cinfo->output_width, (JDIMENSION) 1);
+  dest->pub.buffer_height = 1;
+
+  return (djpeg_dest_ptr) dest;
+}
+
+#endif /* GIF_SUPPORTED */
diff --git a/wrjpgcom.1 b/wrjpgcom.1
new file mode 100644
index 0000000..d419a99
--- /dev/null
+++ b/wrjpgcom.1
@@ -0,0 +1,103 @@
+.TH WRJPGCOM 1 "15 June 1995"
+.SH NAME
+wrjpgcom \- insert text comments into a JPEG file
+.SH SYNOPSIS
+.B wrjpgcom
+[
+.B \-replace
+]
+[
+.BI \-comment " text"
+]
+[
+.BI \-cfile " name"
+]
+[
+.I filename
+]
+.LP
+.SH DESCRIPTION
+.LP
+.B wrjpgcom
+reads the named JPEG/JFIF file, or the standard input if no file is named,
+and generates a new JPEG/JFIF file on standard output.  A comment block is
+added to the file.
+.PP
+The JPEG standard allows "comment" (COM) blocks to occur within a JPEG file.
+Although the standard doesn't actually define what COM blocks are for, they
+are widely used to hold user-supplied text strings.  This lets you add
+annotations, titles, index terms, etc to your JPEG files, and later retrieve
+them as text.  COM blocks do not interfere with the image stored in the JPEG
+file.  The maximum size of a COM block is 64K, but you can have as many of
+them as you like in one JPEG file.
+.PP
+.B wrjpgcom
+adds a COM block, containing text you provide, to a JPEG file.
+Ordinarily, the COM block is added after any existing COM blocks; but you
+can delete the old COM blocks if you wish.
+.SH OPTIONS
+Switch names may be abbreviated, and are not case sensitive.
+.TP
+.B \-replace
+Delete any existing COM blocks from the file.
+.TP
+.BI \-comment " text"
+Supply text for new COM block on command line.
+.TP
+.BI \-cfile " name"
+Read text for new COM block from named file.
+.PP
+If you have only one line of comment text to add, you can provide it on the
+command line with
+.BR \-comment .
+The comment text must be surrounded with quotes so that it is treated as a
+single argument.  Longer comments can be read from a text file.
+.PP
+If you give neither
+.B \-comment
+nor
+.BR \-cfile ,
+then
+.B wrjpgcom
+will read the comment text from standard input.  (In this case an input image
+file name MUST be supplied, so that the source JPEG file comes from somewhere
+else.)  You can enter multiple lines, up to 64KB worth.  Type an end-of-file
+indicator (usually control-D) to terminate the comment text entry.
+.PP
+.B wrjpgcom
+will not add a COM block if the provided comment string is empty.  Therefore
+\fB\-replace \-comment ""\fR can be used to delete all COM blocks from a file.
+.SH EXAMPLES
+.LP
+Add a short comment to in.jpg, producing out.jpg:
+.IP
+.B wrjpgcom \-c
+\fI"View of my back yard" in.jpg
+.B >
+.I out.jpg
+.PP
+Attach a long comment previously stored in comment.txt:
+.IP
+.B wrjpgcom
+.I in.jpg
+.B <
+.I comment.txt
+.B >
+.I out.jpg
+.PP
+or equivalently
+.IP
+.B wrjpgcom
+.B -cfile
+.I comment.txt
+.B <
+.I in.jpg
+.B >
+.I out.jpg
+.SH SEE ALSO
+.BR cjpeg (1),
+.BR djpeg (1),
+.BR jpegtran (1),
+.BR rdjpgcom (1)
+.SH AUTHOR
+Independent JPEG Group
diff --git a/wrjpgcom.c b/wrjpgcom.c
new file mode 100644
index 0000000..8c04b05
--- /dev/null
+++ b/wrjpgcom.c
@@ -0,0 +1,583 @@
+/*
+ * wrjpgcom.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a very simple stand-alone application that inserts
+ * user-supplied text as a COM (comment) marker in a JFIF file.
+ * This may be useful as an example of the minimum logic needed to parse
+ * JPEG markers.
+ */
+
+#define JPEG_CJPEG_DJPEG	/* to get the command-line config symbols */
+#include "jinclude.h"		/* get auto-config symbols, <stdio.h> */
+
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare malloc() */
+extern void * malloc ();
+#endif
+#include <ctype.h>		/* to declare isupper(), tolower() */
+#ifdef USE_SETMODE
+#include <fcntl.h>		/* to declare setmode()'s parameter macros */
+/* If you have setmode() but not <io.h>, just delete this line: */
+#include <io.h>			/* to declare setmode() */
+#endif
+
+#ifdef USE_CCOMMAND		/* command-line reader for Macintosh */
+#ifdef __MWERKS__
+#include <SIOUX.h>              /* Metrowerks needs this */
+#include <console.h>		/* ... and this */
+#endif
+#ifdef THINK_C
+#include <console.h>		/* Think declares it here */
+#endif
+#endif
+
+#ifdef DONT_USE_B_MODE		/* define mode parameters for fopen() */
+#define READ_BINARY	"r"
+#define WRITE_BINARY	"w"
+#else
+#ifdef VMS			/* VMS is very nonstandard */
+#define READ_BINARY	"rb", "ctx=stm"
+#define WRITE_BINARY	"wb", "ctx=stm"
+#else				/* standard ANSI-compliant case */
+#define READ_BINARY	"rb"
+#define WRITE_BINARY	"wb"
+#endif
+#endif
+
+#ifndef EXIT_FAILURE		/* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+#ifndef EXIT_SUCCESS
+#ifdef VMS
+#define EXIT_SUCCESS  1		/* VMS is very nonstandard */
+#else
+#define EXIT_SUCCESS  0
+#endif
+#endif
+
+/* Reduce this value if your malloc() can't allocate blocks up to 64K.
+ * On DOS, compiling in large model is usually a better solution.
+ */
+
+#ifndef MAX_COM_LENGTH
+#define MAX_COM_LENGTH 65000L	/* must be <= 65533 in any case */
+#endif
+
+
+/*
+ * These macros are used to read the input file and write the output file.
+ * To reuse this code in another application, you might need to change these.
+ */
+
+static FILE * infile;		/* input JPEG file */
+
+/* Return next input byte, or EOF if no more */
+#define NEXTBYTE()  getc(infile)
+
+static FILE * outfile;		/* output JPEG file */
+
+/* Emit an output byte */
+#define PUTBYTE(x)  putc((x), outfile)
+
+
+/* Error exit handler */
+#define ERREXIT(msg)  (fprintf(stderr, "%s\n", msg), exit(EXIT_FAILURE))
+
+
+/* Read one byte, testing for EOF */
+static int
+read_1_byte (void)
+{
+  int c;
+
+  c = NEXTBYTE();
+  if (c == EOF)
+    ERREXIT("Premature EOF in JPEG file");
+  return c;
+}
+
+/* Read 2 bytes, convert to unsigned int */
+/* All 2-byte quantities in JPEG markers are MSB first */
+static unsigned int
+read_2_bytes (void)
+{
+  int c1, c2;
+
+  c1 = NEXTBYTE();
+  if (c1 == EOF)
+    ERREXIT("Premature EOF in JPEG file");
+  c2 = NEXTBYTE();
+  if (c2 == EOF)
+    ERREXIT("Premature EOF in JPEG file");
+  return (((unsigned int) c1) << 8) + ((unsigned int) c2);
+}
+
+
+/* Routines to write data to output file */
+
+static void
+write_1_byte (int c)
+{
+  PUTBYTE(c);
+}
+
+static void
+write_2_bytes (unsigned int val)
+{
+  PUTBYTE((val >> 8) & 0xFF);
+  PUTBYTE(val & 0xFF);
+}
+
+static void
+write_marker (int marker)
+{
+  PUTBYTE(0xFF);
+  PUTBYTE(marker);
+}
+
+static void
+copy_rest_of_file (void)
+{
+  int c;
+
+  while ((c = NEXTBYTE()) != EOF)
+    PUTBYTE(c);
+}
+
+
+/*
+ * JPEG markers consist of one or more 0xFF bytes, followed by a marker
+ * code byte (which is not an FF).  Here are the marker codes of interest
+ * in this program.  (See jdmarker.c for a more complete list.)
+ */
+
+#define M_SOF0  0xC0		/* Start Of Frame N */
+#define M_SOF1  0xC1		/* N indicates which compression process */
+#define M_SOF2  0xC2		/* Only SOF0-SOF2 are now in common use */
+#define M_SOF3  0xC3
+#define M_SOF5  0xC5		/* NB: codes C4 and CC are NOT SOF markers */
+#define M_SOF6  0xC6
+#define M_SOF7  0xC7
+#define M_SOF9  0xC9
+#define M_SOF10 0xCA
+#define M_SOF11 0xCB
+#define M_SOF13 0xCD
+#define M_SOF14 0xCE
+#define M_SOF15 0xCF
+#define M_SOI   0xD8		/* Start Of Image (beginning of datastream) */
+#define M_EOI   0xD9		/* End Of Image (end of datastream) */
+#define M_SOS   0xDA		/* Start Of Scan (begins compressed data) */
+#define M_COM   0xFE		/* COMment */
+
+
+/*
+ * Find the next JPEG marker and return its marker code.
+ * We expect at least one FF byte, possibly more if the compressor used FFs
+ * to pad the file.  (Padding FFs will NOT be replicated in the output file.)
+ * There could also be non-FF garbage between markers.  The treatment of such
+ * garbage is unspecified; we choose to skip over it but emit a warning msg.
+ * NB: this routine must not be used after seeing SOS marker, since it will
+ * not deal correctly with FF/00 sequences in the compressed image data...
+ */
+
+static int
+next_marker (void)
+{
+  int c;
+  int discarded_bytes = 0;
+
+  /* Find 0xFF byte; count and skip any non-FFs. */
+  c = read_1_byte();
+  while (c != 0xFF) {
+    discarded_bytes++;
+    c = read_1_byte();
+  }
+  /* Get marker code byte, swallowing any duplicate FF bytes.  Extra FFs
+   * are legal as pad bytes, so don't count them in discarded_bytes.
+   */
+  do {
+    c = read_1_byte();
+  } while (c == 0xFF);
+
+  if (discarded_bytes != 0) {
+    fprintf(stderr, "Warning: garbage data found in JPEG file\n");
+  }
+
+  return c;
+}
+
+
+/*
+ * Read the initial marker, which should be SOI.
+ * For a JFIF file, the first two bytes of the file should be literally
+ * 0xFF M_SOI.  To be more general, we could use next_marker, but if the
+ * input file weren't actually JPEG at all, next_marker might read the whole
+ * file and then return a misleading error message...
+ */
+
+static int
+first_marker (void)
+{
+  int c1, c2;
+
+  c1 = NEXTBYTE();
+  c2 = NEXTBYTE();
+  if (c1 != 0xFF || c2 != M_SOI)
+    ERREXIT("Not a JPEG file");
+  return c2;
+}
+
+
+/*
+ * Most types of marker are followed by a variable-length parameter segment.
+ * This routine skips over the parameters for any marker we don't otherwise
+ * want to process.
+ * Note that we MUST skip the parameter segment explicitly in order not to
+ * be fooled by 0xFF bytes that might appear within the parameter segment;
+ * such bytes do NOT introduce new markers.
+ */
+
+static void
+copy_variable (void)
+/* Copy an unknown or uninteresting variable-length marker */
+{
+  unsigned int length;
+
+  /* Get the marker parameter length count */
+  length = read_2_bytes();
+  write_2_bytes(length);
+  /* Length includes itself, so must be at least 2 */
+  if (length < 2)
+    ERREXIT("Erroneous JPEG marker length");
+  length -= 2;
+  /* Skip over the remaining bytes */
+  while (length > 0) {
+    write_1_byte(read_1_byte());
+    length--;
+  }
+}
+
+static void
+skip_variable (void)
+/* Skip over an unknown or uninteresting variable-length marker */
+{
+  unsigned int length;
+
+  /* Get the marker parameter length count */
+  length = read_2_bytes();
+  /* Length includes itself, so must be at least 2 */
+  if (length < 2)
+    ERREXIT("Erroneous JPEG marker length");
+  length -= 2;
+  /* Skip over the remaining bytes */
+  while (length > 0) {
+    (void) read_1_byte();
+    length--;
+  }
+}
+
+
+/*
+ * Parse the marker stream until SOFn or EOI is seen;
+ * copy data to output, but discard COM markers unless keep_COM is true.
+ */
+
+static int
+scan_JPEG_header (int keep_COM)
+{
+  int marker;
+
+  /* Expect SOI at start of file */
+  if (first_marker() != M_SOI)
+    ERREXIT("Expected SOI marker first");
+  write_marker(M_SOI);
+
+  /* Scan miscellaneous markers until we reach SOFn. */
+  for (;;) {
+    marker = next_marker();
+    switch (marker) {
+      /* Note that marker codes 0xC4, 0xC8, 0xCC are not, and must not be,
+       * treated as SOFn.  C4 in particular is actually DHT.
+       */
+    case M_SOF0:		/* Baseline */
+    case M_SOF1:		/* Extended sequential, Huffman */
+    case M_SOF2:		/* Progressive, Huffman */
+    case M_SOF3:		/* Lossless, Huffman */
+    case M_SOF5:		/* Differential sequential, Huffman */
+    case M_SOF6:		/* Differential progressive, Huffman */
+    case M_SOF7:		/* Differential lossless, Huffman */
+    case M_SOF9:		/* Extended sequential, arithmetic */
+    case M_SOF10:		/* Progressive, arithmetic */
+    case M_SOF11:		/* Lossless, arithmetic */
+    case M_SOF13:		/* Differential sequential, arithmetic */
+    case M_SOF14:		/* Differential progressive, arithmetic */
+    case M_SOF15:		/* Differential lossless, arithmetic */
+      return marker;
+
+    case M_SOS:			/* should not see compressed data before SOF */
+      ERREXIT("SOS without prior SOFn");
+      break;
+
+    case M_EOI:			/* in case it's a tables-only JPEG stream */
+      return marker;
+
+    case M_COM:			/* Existing COM: conditionally discard */
+      if (keep_COM) {
+	write_marker(marker);
+	copy_variable();
+      } else {
+	skip_variable();
+      }
+      break;
+
+    default:			/* Anything else just gets copied */
+      write_marker(marker);
+      copy_variable();		/* we assume it has a parameter count... */
+      break;
+    }
+  } /* end loop */
+}
+
+
+/* Command line parsing code */
+
+static const char * progname;	/* program name for error messages */
+
+
+static void
+usage (void)
+/* complain about bad command line */
+{
+  fprintf(stderr, "wrjpgcom inserts a textual comment in a JPEG file.\n");
+  fprintf(stderr, "You can add to or replace any existing comment(s).\n");
+
+  fprintf(stderr, "Usage: %s [switches] ", progname);
+#ifdef TWO_FILE_COMMANDLINE
+  fprintf(stderr, "inputfile outputfile\n");
+#else
+  fprintf(stderr, "[inputfile]\n");
+#endif
+
+  fprintf(stderr, "Switches (names may be abbreviated):\n");
+  fprintf(stderr, "  -replace         Delete any existing comments\n");
+  fprintf(stderr, "  -comment \"text\"  Insert comment with given text\n");
+  fprintf(stderr, "  -cfile name      Read comment from named file\n");
+  fprintf(stderr, "Notice that you must put quotes around the comment text\n");
+  fprintf(stderr, "when you use -comment.\n");
+  fprintf(stderr, "If you do not give either -comment or -cfile on the command line,\n");
+  fprintf(stderr, "then the comment text is read from standard input.\n");
+  fprintf(stderr, "It can be multiple lines, up to %u characters total.\n",
+	  (unsigned int) MAX_COM_LENGTH);
+#ifndef TWO_FILE_COMMANDLINE
+  fprintf(stderr, "You must specify an input JPEG file name when supplying\n");
+  fprintf(stderr, "comment text from standard input.\n");
+#endif
+
+  exit(EXIT_FAILURE);
+}
+
+
+static int
+keymatch (char * arg, const char * keyword, int minchars)
+/* Case-insensitive matching of (possibly abbreviated) keyword switches. */
+/* keyword is the constant keyword (must be lower case already), */
+/* minchars is length of minimum legal abbreviation. */
+{
+  register int ca, ck;
+  register int nmatched = 0;
+
+  while ((ca = *arg++) != '\0') {
+    if ((ck = *keyword++) == '\0')
+      return 0;			/* arg longer than keyword, no good */
+    if (isupper(ca))		/* force arg to lcase (assume ck is already) */
+      ca = tolower(ca);
+    if (ca != ck)
+      return 0;			/* no good */
+    nmatched++;			/* count matched characters */
+  }
+  /* reached end of argument; fail if it's too short for unique abbrev */
+  if (nmatched < minchars)
+    return 0;
+  return 1;			/* A-OK */
+}
+
+
+/*
+ * The main program.
+ */
+
+int
+main (int argc, char **argv)
+{
+  int argn;
+  char * arg;
+  int keep_COM = 1;
+  char * comment_arg = NULL;
+  FILE * comment_file = NULL;
+  unsigned int comment_length = 0;
+  int marker;
+
+  /* On Mac, fetch a command line. */
+#ifdef USE_CCOMMAND
+  argc = ccommand(&argv);
+#endif
+
+  progname = argv[0];
+  if (progname == NULL || progname[0] == 0)
+    progname = "wrjpgcom";	/* in case C library doesn't provide it */
+
+  /* Parse switches, if any */
+  for (argn = 1; argn < argc; argn++) {
+    arg = argv[argn];
+    if (arg[0] != '-')
+      break;			/* not switch, must be file name */
+    arg++;			/* advance over '-' */
+    if (keymatch(arg, "replace", 1)) {
+      keep_COM = 0;
+    } else if (keymatch(arg, "cfile", 2)) {
+      if (++argn >= argc) usage();
+      if ((comment_file = fopen(argv[argn], "r")) == NULL) {
+	fprintf(stderr, "%s: can't open %s\n", progname, argv[argn]);
+	exit(EXIT_FAILURE);
+      }
+    } else if (keymatch(arg, "comment", 1)) {
+      if (++argn >= argc) usage();
+      comment_arg = argv[argn];
+      /* If the comment text starts with '"', then we are probably running
+       * under MS-DOG and must parse out the quoted string ourselves.  Sigh.
+       */
+      if (comment_arg[0] == '"') {
+	comment_arg = (char *) malloc((size_t) MAX_COM_LENGTH);
+	if (comment_arg == NULL)
+	  ERREXIT("Insufficient memory");
+	strcpy(comment_arg, argv[argn]+1);
+	for (;;) {
+	  comment_length = (unsigned int) strlen(comment_arg);
+	  if (comment_length > 0 && comment_arg[comment_length-1] == '"') {
+	    comment_arg[comment_length-1] = '\0'; /* zap terminating quote */
+	    break;
+	  }
+	  if (++argn >= argc)
+	    ERREXIT("Missing ending quote mark");
+	  strcat(comment_arg, " ");
+	  strcat(comment_arg, argv[argn]);
+	}
+      }
+      comment_length = (unsigned int) strlen(comment_arg);
+    } else
+      usage();
+  }
+
+  /* Cannot use both -comment and -cfile. */
+  if (comment_arg != NULL && comment_file != NULL)
+    usage();
+  /* If there is neither -comment nor -cfile, we will read the comment text
+   * from stdin; in this case there MUST be an input JPEG file name.
+   */
+  if (comment_arg == NULL && comment_file == NULL && argn >= argc)
+    usage();
+
+  /* Open the input file. */
+  if (argn < argc) {
+    if ((infile = fopen(argv[argn], READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open %s\n", progname, argv[argn]);
+      exit(EXIT_FAILURE);
+    }
+  } else {
+    /* default input file is stdin */
+#ifdef USE_SETMODE		/* need to hack file mode? */
+    setmode(fileno(stdin), O_BINARY);
+#endif
+#ifdef USE_FDOPEN		/* need to re-open in binary mode? */
+    if ((infile = fdopen(fileno(stdin), READ_BINARY)) == NULL) {
+      fprintf(stderr, "%s: can't open stdin\n", progname);
+      exit(EXIT_FAILURE);
+    }
+#else
+    infile = stdin;
+#endif
+  }
+
+  /* Open the output file. */
+#ifdef TWO_FILE_COMMANDLINE
+  /* Must have explicit output file name */
+  if (argn != argc-2) {
+    fprintf(stderr, "%s: must name one input and one output file\n",
+	    progname);
+    usage();
+  }
+  if ((outfile = fopen(argv[argn+1], WRITE_BINARY)) == NULL) {
+    fprintf(stderr, "%s: can't open %s\n", progname, argv[argn+1]);
+    exit(EXIT_FAILURE);
+  }
+#else
+  /* Unix style: expect zero or one file name */
+  if (argn < argc-1) {
+    fprintf(stderr, "%s: only one input file\n", progname);
+    usage();
+  }
+  /* default output file is stdout */
+#ifdef USE_SETMODE		/* need to hack file mode? */
+  setmode(fileno(stdout), O_BINARY);
+#endif
+#ifdef USE_FDOPEN		/* need to re-open in binary mode? */
+  if ((outfile = fdopen(fileno(stdout), WRITE_BINARY)) == NULL) {
+    fprintf(stderr, "%s: can't open stdout\n", progname);
+    exit(EXIT_FAILURE);
+  }
+#else
+  outfile = stdout;
+#endif
+#endif /* TWO_FILE_COMMANDLINE */
+
+  /* Collect comment text from comment_file or stdin, if necessary */
+  if (comment_arg == NULL) {
+    FILE * src_file;
+    int c;
+
+    comment_arg = (char *) malloc((size_t) MAX_COM_LENGTH);
+    if (comment_arg == NULL)
+      ERREXIT("Insufficient memory");
+    comment_length = 0;
+    src_file = (comment_file != NULL ? comment_file : stdin);
+    while ((c = getc(src_file)) != EOF) {
+      if (comment_length >= (unsigned int) MAX_COM_LENGTH) {
+	fprintf(stderr, "Comment text may not exceed %u bytes\n",
+		(unsigned int) MAX_COM_LENGTH);
+	exit(EXIT_FAILURE);
+      }
+      comment_arg[comment_length++] = (char) c;
+    }
+    if (comment_file != NULL)
+      fclose(comment_file);
+  }
+
+  /* Copy JPEG headers until SOFn marker;
+   * we will insert the new comment marker just before SOFn.
+   * This (a) causes the new comment to appear after, rather than before,
+   * existing comments; and (b) ensures that comments come after any JFIF
+   * or JFXX markers, as required by the JFIF specification.
+   */
+  marker = scan_JPEG_header(keep_COM);
+  /* Insert the new COM marker, but only if nonempty text has been supplied */
+  if (comment_length > 0) {
+    write_marker(M_COM);
+    write_2_bytes(comment_length + 2);
+    while (comment_length > 0) {
+      write_1_byte(*comment_arg++);
+      comment_length--;
+    }
+  }
+  /* Duplicate the remainder of the source file.
+   * Note that any COM markers occuring after SOF will not be touched.
+   */
+  write_marker(marker);
+  copy_rest_of_file();
+
+  /* All done. */
+  exit(EXIT_SUCCESS);
+  return 0;			/* suppress no-return-value warnings */
+}
diff --git a/wrppm.c b/wrppm.c
new file mode 100644
index 0000000..68e0c85
--- /dev/null
+++ b/wrppm.c
@@ -0,0 +1,269 @@
+/*
+ * wrppm.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Modified 2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write output images in PPM/PGM format.
+ * The extended 2-byte-per-sample raw PPM/PGM formats are supported.
+ * The PBMPLUS library is NOT required to compile this software
+ * (but it is highly useful as a set of PPM image manipulation programs).
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume output to
+ * an ordinary stdio stream.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef PPM_SUPPORTED
+
+
+/*
+ * For 12-bit JPEG data, we either downscale the values to 8 bits
+ * (to write standard byte-per-sample PPM/PGM files), or output
+ * nonstandard word-per-sample PPM/PGM files.  Downscaling is done
+ * if PPM_NORAWWORD is defined (this can be done in the Makefile
+ * or in jconfig.h).
+ * (When the core library supports data precision reduction, a cleaner
+ * implementation will be to ask for that instead.)
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define PUTPPMSAMPLE(ptr,v)  *ptr++ = (char) (v)
+#define BYTESPERSAMPLE 1
+#define PPM_MAXVAL 255
+#else
+#ifdef PPM_NORAWWORD
+#define PUTPPMSAMPLE(ptr,v)  *ptr++ = (char) ((v) >> (BITS_IN_JSAMPLE-8))
+#define BYTESPERSAMPLE 1
+#define PPM_MAXVAL 255
+#else
+/* The word-per-sample format always puts the MSB first. */
+#define PUTPPMSAMPLE(ptr,v)			\
+	{ register int val_ = v;		\
+	  *ptr++ = (char) ((val_ >> 8) & 0xFF);	\
+	  *ptr++ = (char) (val_ & 0xFF);	\
+	}
+#define BYTESPERSAMPLE 2
+#define PPM_MAXVAL ((1<<BITS_IN_JSAMPLE)-1)
+#endif
+#endif
+
+
+/*
+ * When JSAMPLE is the same size as char, we can just fwrite() the
+ * decompressed data to the PPM or PGM file.  On PCs, in order to make this
+ * work the output buffer must be allocated in near data space, because we are
+ * assuming small-data memory model wherein fwrite() can't reach far memory.
+ * If you need to process very wide images on a PC, you might have to compile
+ * in large-memory model, or else replace fwrite() with a putc() loop ---
+ * which will be much slower.
+ */
+
+
+/* Private version of data destination object */
+
+typedef struct {
+  struct djpeg_dest_struct pub;	/* public fields */
+
+  /* Usually these two pointers point to the same place: */
+  char *iobuffer;		/* fwrite's I/O buffer */
+  JSAMPROW pixrow;		/* decompressor output buffer */
+  size_t buffer_width;		/* width of I/O buffer */
+  JDIMENSION samples_per_row;	/* JSAMPLEs per output row */
+} ppm_dest_struct;
+
+typedef ppm_dest_struct * ppm_dest_ptr;
+
+
+/*
+ * Write some pixel data.
+ * In this module rows_supplied will always be 1.
+ *
+ * put_pixel_rows handles the "normal" 8-bit case where the decompressor
+ * output buffer is physically the same as the fwrite buffer.
+ */
+
+METHODDEF(void)
+put_pixel_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		JDIMENSION rows_supplied)
+{
+  ppm_dest_ptr dest = (ppm_dest_ptr) dinfo;
+
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+
+/*
+ * This code is used when we have to copy the data and apply a pixel
+ * format translation.  Typically this only happens in 12-bit mode.
+ */
+
+METHODDEF(void)
+copy_pixel_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		 JDIMENSION rows_supplied)
+{
+  ppm_dest_ptr dest = (ppm_dest_ptr) dinfo;
+  register char * bufferptr;
+  register JSAMPROW ptr;
+  register JDIMENSION col;
+
+  ptr = dest->pub.buffer[0];
+  bufferptr = dest->iobuffer;
+  for (col = dest->samples_per_row; col > 0; col--) {
+    PUTPPMSAMPLE(bufferptr, GETJSAMPLE(*ptr++));
+  }
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+
+/*
+ * Write some pixel data when color quantization is in effect.
+ * We have to demap the color index values to straight data.
+ */
+
+METHODDEF(void)
+put_demapped_rgb (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		  JDIMENSION rows_supplied)
+{
+  ppm_dest_ptr dest = (ppm_dest_ptr) dinfo;
+  register char * bufferptr;
+  register int pixval;
+  register JSAMPROW ptr;
+  register JSAMPROW color_map0 = cinfo->colormap[0];
+  register JSAMPROW color_map1 = cinfo->colormap[1];
+  register JSAMPROW color_map2 = cinfo->colormap[2];
+  register JDIMENSION col;
+
+  ptr = dest->pub.buffer[0];
+  bufferptr = dest->iobuffer;
+  for (col = cinfo->output_width; col > 0; col--) {
+    pixval = GETJSAMPLE(*ptr++);
+    PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map0[pixval]));
+    PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map1[pixval]));
+    PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map2[pixval]));
+  }
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+
+METHODDEF(void)
+put_demapped_gray (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		   JDIMENSION rows_supplied)
+{
+  ppm_dest_ptr dest = (ppm_dest_ptr) dinfo;
+  register char * bufferptr;
+  register JSAMPROW ptr;
+  register JSAMPROW color_map = cinfo->colormap[0];
+  register JDIMENSION col;
+
+  ptr = dest->pub.buffer[0];
+  bufferptr = dest->iobuffer;
+  for (col = cinfo->output_width; col > 0; col--) {
+    PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map[GETJSAMPLE(*ptr++)]));
+  }
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+
+/*
+ * Startup: write the file header.
+ */
+
+METHODDEF(void)
+start_output_ppm (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  ppm_dest_ptr dest = (ppm_dest_ptr) dinfo;
+
+  /* Emit file header */
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    /* emit header for raw PGM format */
+    fprintf(dest->pub.output_file, "P5\n%ld %ld\n%d\n",
+	    (long) cinfo->output_width, (long) cinfo->output_height,
+	    PPM_MAXVAL);
+    break;
+  case JCS_RGB:
+    /* emit header for raw PPM format */
+    fprintf(dest->pub.output_file, "P6\n%ld %ld\n%d\n",
+	    (long) cinfo->output_width, (long) cinfo->output_height,
+	    PPM_MAXVAL);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_PPM_COLORSPACE);
+  }
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_output_ppm (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  /* Make sure we wrote the output file OK */
+  fflush(dinfo->output_file);
+  if (ferror(dinfo->output_file))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * The module selection routine for PPM format output.
+ */
+
+GLOBAL(djpeg_dest_ptr)
+jinit_write_ppm (j_decompress_ptr cinfo)
+{
+  ppm_dest_ptr dest;
+
+  /* Create module interface object, fill in method pointers */
+  dest = (ppm_dest_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(ppm_dest_struct));
+  dest->pub.start_output = start_output_ppm;
+  dest->pub.finish_output = finish_output_ppm;
+
+  /* Calculate output image dimensions so we can allocate space */
+  jpeg_calc_output_dimensions(cinfo);
+
+  /* Create physical I/O buffer.  Note we make this near on a PC. */
+  dest->samples_per_row = cinfo->output_width * cinfo->out_color_components;
+  dest->buffer_width = dest->samples_per_row * (BYTESPERSAMPLE * SIZEOF(char));
+  dest->iobuffer = (char *) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, dest->buffer_width);
+
+  if (cinfo->quantize_colors || BITS_IN_JSAMPLE != 8 ||
+      SIZEOF(JSAMPLE) != SIZEOF(char)) {
+    /* When quantizing, we need an output buffer for colormap indexes
+     * that's separate from the physical I/O buffer.  We also need a
+     * separate buffer if pixel format translation must take place.
+     */
+    dest->pub.buffer = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       cinfo->output_width * cinfo->output_components, (JDIMENSION) 1);
+    dest->pub.buffer_height = 1;
+    if (! cinfo->quantize_colors)
+      dest->pub.put_pixel_rows = copy_pixel_rows;
+    else if (cinfo->out_color_space == JCS_GRAYSCALE)
+      dest->pub.put_pixel_rows = put_demapped_gray;
+    else
+      dest->pub.put_pixel_rows = put_demapped_rgb;
+  } else {
+    /* We will fwrite() directly from decompressor output buffer. */
+    /* Synthesize a JSAMPARRAY pointer structure */
+    /* Cast here implies near->far pointer conversion on PCs */
+    dest->pixrow = (JSAMPROW) dest->iobuffer;
+    dest->pub.buffer = & dest->pixrow;
+    dest->pub.buffer_height = 1;
+    dest->pub.put_pixel_rows = put_pixel_rows;
+  }
+
+  return (djpeg_dest_ptr) dest;
+}
+
+#endif /* PPM_SUPPORTED */
diff --git a/wrrle.c b/wrrle.c
new file mode 100644
index 0000000..a4e7337
--- /dev/null
+++ b/wrrle.c
@@ -0,0 +1,305 @@
+/*
+ * wrrle.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write output images in RLE format.
+ * The Utah Raster Toolkit library is required (version 3.1 or later).
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume output to
+ * an ordinary stdio stream.
+ *
+ * Based on code contributed by Mike Lijewski,
+ * with updates from Robert Hutchinson.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef RLE_SUPPORTED
+
+/* rle.h is provided by the Utah Raster Toolkit. */
+
+#include <rle.h>
+
+/*
+ * We assume that JSAMPLE has the same representation as rle_pixel,
+ * to wit, "unsigned char".  Hence we can't cope with 12- or 16-bit samples.
+ */
+
+#if BITS_IN_JSAMPLE != 8
+  Sorry, this code only copes with 8-bit JSAMPLEs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * Since RLE stores scanlines bottom-to-top, we have to invert the image
+ * from JPEG's top-to-bottom order.  To do this, we save the outgoing data
+ * in a virtual array during put_pixel_row calls, then actually emit the
+ * RLE file during finish_output.
+ */
+
+
+/*
+ * For now, if we emit an RLE color map then it is always 256 entries long,
+ * though not all of the entries need be used.
+ */
+
+#define CMAPBITS	8
+#define CMAPLENGTH	(1<<(CMAPBITS))
+
+typedef struct {
+  struct djpeg_dest_struct pub; /* public fields */
+
+  jvirt_sarray_ptr image;	/* virtual array to store the output image */
+  rle_map *colormap;	 	/* RLE-style color map, or NULL if none */
+  rle_pixel **rle_row;		/* To pass rows to rle_putrow() */
+
+} rle_dest_struct;
+
+typedef rle_dest_struct * rle_dest_ptr;
+
+/* Forward declarations */
+METHODDEF(void) rle_put_pixel_rows
+    JPP((j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+	 JDIMENSION rows_supplied));
+
+
+/*
+ * Write the file header.
+ *
+ * In this module it's easier to wait till finish_output to write anything.
+ */
+
+METHODDEF(void)
+start_output_rle (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  rle_dest_ptr dest = (rle_dest_ptr) dinfo;
+  size_t cmapsize;
+  int i, ci;
+#ifdef PROGRESS_REPORT
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+#endif
+
+  /*
+   * Make sure the image can be stored in RLE format.
+   *
+   * - RLE stores image dimensions as *signed* 16 bit integers.  JPEG
+   *   uses unsigned, so we have to check the width.
+   *
+   * - Colorspace is expected to be grayscale or RGB.
+   *
+   * - The number of channels (components) is expected to be 1 (grayscale/
+   *   pseudocolor) or 3 (truecolor/directcolor).
+   *   (could be 2 or 4 if using an alpha channel, but we aren't)
+   */
+
+  if (cinfo->output_width > 32767 || cinfo->output_height > 32767)
+    ERREXIT2(cinfo, JERR_RLE_DIMENSIONS, cinfo->output_width, 
+	     cinfo->output_height);
+
+  if (cinfo->out_color_space != JCS_GRAYSCALE &&
+      cinfo->out_color_space != JCS_RGB)
+    ERREXIT(cinfo, JERR_RLE_COLORSPACE);
+
+  if (cinfo->output_components != 1 && cinfo->output_components != 3)
+    ERREXIT1(cinfo, JERR_RLE_TOOMANYCHANNELS, cinfo->num_components);
+
+  /* Convert colormap, if any, to RLE format. */
+
+  dest->colormap = NULL;
+
+  if (cinfo->quantize_colors) {
+    /* Allocate storage for RLE-style cmap, zero any extra entries */
+    cmapsize = cinfo->out_color_components * CMAPLENGTH * SIZEOF(rle_map);
+    dest->colormap = (rle_map *) (*cinfo->mem->alloc_small)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE, cmapsize);
+    MEMZERO(dest->colormap, cmapsize);
+
+    /* Save away data in RLE format --- note 8-bit left shift! */
+    /* Shifting would need adjustment for JSAMPLEs wider than 8 bits. */
+    for (ci = 0; ci < cinfo->out_color_components; ci++) {
+      for (i = 0; i < cinfo->actual_number_of_colors; i++) {
+        dest->colormap[ci * CMAPLENGTH + i] =
+          GETJSAMPLE(cinfo->colormap[ci][i]) << 8;
+      }
+    }
+  }
+
+  /* Set the output buffer to the first row */
+  dest->pub.buffer = (*cinfo->mem->access_virt_sarray)
+    ((j_common_ptr) cinfo, dest->image, (JDIMENSION) 0, (JDIMENSION) 1, TRUE);
+  dest->pub.buffer_height = 1;
+
+  dest->pub.put_pixel_rows = rle_put_pixel_rows;
+
+#ifdef PROGRESS_REPORT
+  if (progress != NULL) {
+    progress->total_extra_passes++;  /* count file writing as separate pass */
+  }
+#endif
+}
+
+
+/*
+ * Write some pixel data.
+ *
+ * This routine just saves the data away in a virtual array.
+ */
+
+METHODDEF(void)
+rle_put_pixel_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		    JDIMENSION rows_supplied)
+{
+  rle_dest_ptr dest = (rle_dest_ptr) dinfo;
+
+  if (cinfo->output_scanline < cinfo->output_height) {
+    dest->pub.buffer = (*cinfo->mem->access_virt_sarray)
+      ((j_common_ptr) cinfo, dest->image,
+       cinfo->output_scanline, (JDIMENSION) 1, TRUE);
+  }
+}
+
+/*
+ * Finish up at the end of the file.
+ *
+ * Here is where we really output the RLE file.
+ */
+
+METHODDEF(void)
+finish_output_rle (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  rle_dest_ptr dest = (rle_dest_ptr) dinfo;
+  rle_hdr header;		/* Output file information */
+  rle_pixel **rle_row, *red, *green, *blue;
+  JSAMPROW output_row;
+  char cmapcomment[80];
+  int row, col;
+  int ci;
+#ifdef PROGRESS_REPORT
+  cd_progress_ptr progress = (cd_progress_ptr) cinfo->progress;
+#endif
+
+  /* Initialize the header info */
+  header = *rle_hdr_init(NULL);
+  header.rle_file = dest->pub.output_file;
+  header.xmin     = 0;
+  header.xmax     = cinfo->output_width  - 1;
+  header.ymin     = 0;
+  header.ymax     = cinfo->output_height - 1;
+  header.alpha    = 0;
+  header.ncolors  = cinfo->output_components;
+  for (ci = 0; ci < cinfo->output_components; ci++) {
+    RLE_SET_BIT(header, ci);
+  }
+  if (cinfo->quantize_colors) {
+    header.ncmap   = cinfo->out_color_components;
+    header.cmaplen = CMAPBITS;
+    header.cmap    = dest->colormap;
+    /* Add a comment to the output image with the true colormap length. */
+    sprintf(cmapcomment, "color_map_length=%d", cinfo->actual_number_of_colors);
+    rle_putcom(cmapcomment, &header);
+  }
+
+  /* Emit the RLE header and color map (if any) */
+  rle_put_setup(&header);
+
+  /* Now output the RLE data from our virtual array.
+   * We assume here that (a) rle_pixel is represented the same as JSAMPLE,
+   * and (b) we are not on a machine where FAR pointers differ from regular.
+   */
+
+#ifdef PROGRESS_REPORT
+  if (progress != NULL) {
+    progress->pub.pass_limit = cinfo->output_height;
+    progress->pub.pass_counter = 0;
+    (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+  }
+#endif
+
+  if (cinfo->output_components == 1) {
+    for (row = cinfo->output_height-1; row >= 0; row--) {
+      rle_row = (rle_pixel **) (*cinfo->mem->access_virt_sarray)
+        ((j_common_ptr) cinfo, dest->image,
+	 (JDIMENSION) row, (JDIMENSION) 1, FALSE);
+      rle_putrow(rle_row, (int) cinfo->output_width, &header);
+#ifdef PROGRESS_REPORT
+      if (progress != NULL) {
+        progress->pub.pass_counter++;
+        (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+      }
+#endif
+    }
+  } else {
+    for (row = cinfo->output_height-1; row >= 0; row--) {
+      rle_row = (rle_pixel **) dest->rle_row;
+      output_row = * (*cinfo->mem->access_virt_sarray)
+        ((j_common_ptr) cinfo, dest->image,
+	 (JDIMENSION) row, (JDIMENSION) 1, FALSE);
+      red = rle_row[0];
+      green = rle_row[1];
+      blue = rle_row[2];
+      for (col = cinfo->output_width; col > 0; col--) {
+        *red++ = GETJSAMPLE(*output_row++);
+        *green++ = GETJSAMPLE(*output_row++);
+        *blue++ = GETJSAMPLE(*output_row++);
+      }
+      rle_putrow(rle_row, (int) cinfo->output_width, &header);
+#ifdef PROGRESS_REPORT
+      if (progress != NULL) {
+        progress->pub.pass_counter++;
+        (*progress->pub.progress_monitor) ((j_common_ptr) cinfo);
+      }
+#endif
+    }
+  }
+
+#ifdef PROGRESS_REPORT
+  if (progress != NULL)
+    progress->completed_extra_passes++;
+#endif
+
+  /* Emit file trailer */
+  rle_puteof(&header);
+  fflush(dest->pub.output_file);
+  if (ferror(dest->pub.output_file))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * The module selection routine for RLE format output.
+ */
+
+GLOBAL(djpeg_dest_ptr)
+jinit_write_rle (j_decompress_ptr cinfo)
+{
+  rle_dest_ptr dest;
+
+  /* Create module interface object, fill in method pointers */
+  dest = (rle_dest_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+                                  SIZEOF(rle_dest_struct));
+  dest->pub.start_output = start_output_rle;
+  dest->pub.finish_output = finish_output_rle;
+
+  /* Calculate output image dimensions so we can allocate space */
+  jpeg_calc_output_dimensions(cinfo);
+
+  /* Allocate a work array for output to the RLE library. */
+  dest->rle_row = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     cinfo->output_width, (JDIMENSION) cinfo->output_components);
+
+  /* Allocate a virtual array to hold the image. */
+  dest->image = (*cinfo->mem->request_virt_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+     (JDIMENSION) (cinfo->output_width * cinfo->output_components),
+     cinfo->output_height, (JDIMENSION) 1);
+
+  return (djpeg_dest_ptr) dest;
+}
+
+#endif /* RLE_SUPPORTED */
diff --git a/wrtarga.c b/wrtarga.c
new file mode 100644
index 0000000..cf104d2
--- /dev/null
+++ b/wrtarga.c
@@ -0,0 +1,253 @@
+/*
+ * wrtarga.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write output images in Targa format.
+ *
+ * These routines may need modification for non-Unix environments or
+ * specialized applications.  As they stand, they assume output to
+ * an ordinary stdio stream.
+ *
+ * Based on code contributed by Lee Daniel Crocker.
+ */
+
+#include "cdjpeg.h"		/* Common decls for cjpeg/djpeg applications */
+
+#ifdef TARGA_SUPPORTED
+
+
+/*
+ * To support 12-bit JPEG data, we'd have to scale output down to 8 bits.
+ * This is not yet implemented.
+ */
+
+#if BITS_IN_JSAMPLE != 8
+  Sorry, this code only copes with 8-bit JSAMPLEs. /* deliberate syntax err */
+#endif
+
+/*
+ * The output buffer needs to be writable by fwrite().  On PCs, we must
+ * allocate the buffer in near data space, because we are assuming small-data
+ * memory model, wherein fwrite() can't reach far memory.  If you need to
+ * process very wide images on a PC, you might have to compile in large-memory
+ * model, or else replace fwrite() with a putc() loop --- which will be much
+ * slower.
+ */
+
+
+/* Private version of data destination object */
+
+typedef struct {
+  struct djpeg_dest_struct pub;	/* public fields */
+
+  char *iobuffer;		/* physical I/O buffer */
+  JDIMENSION buffer_width;	/* width of one row */
+} tga_dest_struct;
+
+typedef tga_dest_struct * tga_dest_ptr;
+
+
+LOCAL(void)
+write_header (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo, int num_colors)
+/* Create and write a Targa header */
+{
+  char targaheader[18];
+
+  /* Set unused fields of header to 0 */
+  MEMZERO(targaheader, SIZEOF(targaheader));
+
+  if (num_colors > 0) {
+    targaheader[1] = 1;		/* color map type 1 */
+    targaheader[5] = (char) (num_colors & 0xFF);
+    targaheader[6] = (char) (num_colors >> 8);
+    targaheader[7] = 24;	/* 24 bits per cmap entry */
+  }
+
+  targaheader[12] = (char) (cinfo->output_width & 0xFF);
+  targaheader[13] = (char) (cinfo->output_width >> 8);
+  targaheader[14] = (char) (cinfo->output_height & 0xFF);
+  targaheader[15] = (char) (cinfo->output_height >> 8);
+  targaheader[17] = 0x20;	/* Top-down, non-interlaced */
+
+  if (cinfo->out_color_space == JCS_GRAYSCALE) {
+    targaheader[2] = 3;		/* image type = uncompressed gray-scale */
+    targaheader[16] = 8;	/* bits per pixel */
+  } else {			/* must be RGB */
+    if (num_colors > 0) {
+      targaheader[2] = 1;	/* image type = colormapped RGB */
+      targaheader[16] = 8;
+    } else {
+      targaheader[2] = 2;	/* image type = uncompressed RGB */
+      targaheader[16] = 24;
+    }
+  }
+
+  if (JFWRITE(dinfo->output_file, targaheader, 18) != (size_t) 18)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * Write some pixel data.
+ * In this module rows_supplied will always be 1.
+ */
+
+METHODDEF(void)
+put_pixel_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		JDIMENSION rows_supplied)
+/* used for unquantized full-color output */
+{
+  tga_dest_ptr dest = (tga_dest_ptr) dinfo;
+  register JSAMPROW inptr;
+  register char * outptr;
+  register JDIMENSION col;
+
+  inptr = dest->pub.buffer[0];
+  outptr = dest->iobuffer;
+  for (col = cinfo->output_width; col > 0; col--) {
+    outptr[0] = (char) GETJSAMPLE(inptr[2]); /* RGB to BGR order */
+    outptr[1] = (char) GETJSAMPLE(inptr[1]);
+    outptr[2] = (char) GETJSAMPLE(inptr[0]);
+    inptr += 3, outptr += 3;
+  }
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+METHODDEF(void)
+put_gray_rows (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+	       JDIMENSION rows_supplied)
+/* used for grayscale OR quantized color output */
+{
+  tga_dest_ptr dest = (tga_dest_ptr) dinfo;
+  register JSAMPROW inptr;
+  register char * outptr;
+  register JDIMENSION col;
+
+  inptr = dest->pub.buffer[0];
+  outptr = dest->iobuffer;
+  for (col = cinfo->output_width; col > 0; col--) {
+    *outptr++ = (char) GETJSAMPLE(*inptr++);
+  }
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+
+/*
+ * Write some demapped pixel data when color quantization is in effect.
+ * For Targa, this is only applied to grayscale data.
+ */
+
+METHODDEF(void)
+put_demapped_gray (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+		   JDIMENSION rows_supplied)
+{
+  tga_dest_ptr dest = (tga_dest_ptr) dinfo;
+  register JSAMPROW inptr;
+  register char * outptr;
+  register JSAMPROW color_map0 = cinfo->colormap[0];
+  register JDIMENSION col;
+
+  inptr = dest->pub.buffer[0];
+  outptr = dest->iobuffer;
+  for (col = cinfo->output_width; col > 0; col--) {
+    *outptr++ = (char) GETJSAMPLE(color_map0[GETJSAMPLE(*inptr++)]);
+  }
+  (void) JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
+}
+
+
+/*
+ * Startup: write the file header.
+ */
+
+METHODDEF(void)
+start_output_tga (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  tga_dest_ptr dest = (tga_dest_ptr) dinfo;
+  int num_colors, i;
+  FILE *outfile;
+
+  if (cinfo->out_color_space == JCS_GRAYSCALE) {
+    /* Targa doesn't have a mapped grayscale format, so we will */
+    /* demap quantized gray output.  Never emit a colormap. */
+    write_header(cinfo, dinfo, 0);
+    if (cinfo->quantize_colors)
+      dest->pub.put_pixel_rows = put_demapped_gray;
+    else
+      dest->pub.put_pixel_rows = put_gray_rows;
+  } else if (cinfo->out_color_space == JCS_RGB) {
+    if (cinfo->quantize_colors) {
+      /* We only support 8-bit colormap indexes, so only 256 colors */
+      num_colors = cinfo->actual_number_of_colors;
+      if (num_colors > 256)
+	ERREXIT1(cinfo, JERR_TOO_MANY_COLORS, num_colors);
+      write_header(cinfo, dinfo, num_colors);
+      /* Write the colormap.  Note Targa uses BGR byte order */
+      outfile = dest->pub.output_file;
+      for (i = 0; i < num_colors; i++) {
+	putc(GETJSAMPLE(cinfo->colormap[2][i]), outfile);
+	putc(GETJSAMPLE(cinfo->colormap[1][i]), outfile);
+	putc(GETJSAMPLE(cinfo->colormap[0][i]), outfile);
+      }
+      dest->pub.put_pixel_rows = put_gray_rows;
+    } else {
+      write_header(cinfo, dinfo, 0);
+      dest->pub.put_pixel_rows = put_pixel_rows;
+    }
+  } else {
+    ERREXIT(cinfo, JERR_TGA_COLORSPACE);
+  }
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_output_tga (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
+{
+  /* Make sure we wrote the output file OK */
+  fflush(dinfo->output_file);
+  if (ferror(dinfo->output_file))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * The module selection routine for Targa format output.
+ */
+
+GLOBAL(djpeg_dest_ptr)
+jinit_write_targa (j_decompress_ptr cinfo)
+{
+  tga_dest_ptr dest;
+
+  /* Create module interface object, fill in method pointers */
+  dest = (tga_dest_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(tga_dest_struct));
+  dest->pub.start_output = start_output_tga;
+  dest->pub.finish_output = finish_output_tga;
+
+  /* Calculate output image dimensions so we can allocate space */
+  jpeg_calc_output_dimensions(cinfo);
+
+  /* Create I/O buffer.  Note we make this near on a PC. */
+  dest->buffer_width = cinfo->output_width * cinfo->output_components;
+  dest->iobuffer = (char *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(size_t) (dest->buffer_width * SIZEOF(char)));
+
+  /* Create decompressor output buffer. */
+  dest->pub.buffer = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, dest->buffer_width, (JDIMENSION) 1);
+  dest->pub.buffer_height = 1;
+
+  return (djpeg_dest_ptr) dest;
+}
+
+#endif /* TARGA_SUPPORTED */