diff options
Diffstat (limited to 'simd')
73 files changed, 30610 insertions, 0 deletions
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt new file mode 100755 index 0000000..fc2ce99 --- /dev/null +++ b/simd/CMakeLists.txt @@ -0,0 +1,68 @@ +if(NOT DEFINED NASM) + set(NASM nasm CACHE PATH "Path to NASM/YASM executable") +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() + +# 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 ${NAFLAGS} -g) +endif() + +if(SIMD_X86_64) + set(SIMD_BASENAMES jfsseflt-64 jccolss2-64 jdcolss2-64 jcgrass2-64 + jcsamss2-64 jdsamss2-64 jdmerss2-64 jcqnts2i-64 jfss2fst-64 jfss2int-64 + jiss2red-64 jiss2int-64 jiss2fst-64 jcqnts2f-64 jiss2flt-64) + message(STATUS "Building x86_64 SIMD extensions") +else() + set(SIMD_BASENAMES jsimdcpu jccolmmx jcgrammx jdcolmmx jcsammmx jdsammmx + jdmermmx jcqntmmx jfmmxfst jfmmxint jimmxred jimmxint jimmxfst jcqnt3dn + jf3dnflt ji3dnflt jcqntsse jfsseflt jisseflt jccolss2 jcgrass2 jdcolss2 + jcsamss2 jdsamss2 jdmerss2 jcqnts2i jfss2fst jfss2int jiss2red jiss2int + jiss2fst jcqnts2f jiss2flt) + message(STATUS "Building i386 SIMD extensions") +endif() + +if(MSVC_IDE) + set(OBJDIR "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_CFG_INTDIR}") +else() + set(OBJDIR ${CMAKE_CURRENT_BINARY_DIR}) +endif() + +file(GLOB INC_FILES *.inc) + +foreach(file ${SIMD_BASENAMES}) + set(DEPFILE "") + set(SIMD_SRC ${CMAKE_CURRENT_SOURCE_DIR}/${file}.asm) + if(${file} MATCHES col) + set(DEPFILE ${file}) + string(REGEX REPLACE "col" "clr" DEPFILE ${DEPFILE}) + set(DEPFILE ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE}.asm) + endif() + if(${file} MATCHES mer) + set(DEPFILE ${file}) + string(REGEX REPLACE "mer" "mrg" DEPFILE ${DEPFILE}) + set(DEPFILE ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE}.asm) + endif() + if(${file} MATCHES gra) + set(DEPFILE ${file}) + string(REGEX REPLACE "gra" "gry" DEPFILE ${DEPFILE}) + set(DEPFILE ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE}.asm) + endif() + set(SIMD_OBJ ${OBJDIR}/${file}.obj) + add_custom_command(OUTPUT ${SIMD_OBJ} + DEPENDS ${SIMD_SRC} ${DEPFILE} ${INC_FILES} + 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 ${SIMD_OBJS}) diff --git a/simd/Makefile.am b/simd/Makefile.am new file mode 100644 index 0000000..a12ff6e --- /dev/null +++ b/simd/Makefile.am @@ -0,0 +1,67 @@ +noinst_LTLIBRARIES = libsimd.la + +BUILT_SOURCES = jsimdcfg.inc + +EXTRA_DIST = nasm_lt.sh jcclrmmx.asm jcclrss2.asm jdclrmmx.asm jdclrss2.asm \ + jdmrgmmx.asm jdmrgss2.asm jcclrss2-64.asm jdclrss2-64.asm \ + jdmrgss2-64.asm jcgryss2-64.asm jcgrymmx.asm jcgryss2.asm CMakeLists.txt + +if SIMD_X86_64 + +libsimd_la_SOURCES = jsimd_x86_64.c \ + jsimd.h jsimdcfg.inc.h \ + jsimdext.inc jcolsamp.inc jdct.inc \ + jfsseflt-64.asm jcgrass2-64.asm \ + jccolss2-64.asm jdcolss2-64.asm \ + jcsamss2-64.asm jdsamss2-64.asm jdmerss2-64.asm \ + jcqnts2i-64.asm jfss2fst-64.asm jfss2int-64.asm \ + jiss2red-64.asm jiss2int-64.asm jiss2fst-64.asm \ + jcqnts2f-64.asm jiss2flt-64.asm + +jccolss2-64.lo: jcclrss2-64.asm +jdcolss2-64.lo: jdclrss2-64.asm +jcgrass2-64.lo: jcgryss2-64.asm +jdmerss2-64.lo: jdmrgss2-64.asm +endif + +if SIMD_I386 + +libsimd_la_SOURCES = jsimd_i386.c \ + jsimd.h jsimdcfg.inc.h \ + jsimdext.inc jcolsamp.inc jdct.inc \ + jsimdcpu.asm \ + jccolmmx.asm jdcolmmx.asm jcgrammx.asm \ + jcsammmx.asm jdsammmx.asm jdmermmx.asm \ + jcqntmmx.asm jfmmxfst.asm jfmmxint.asm \ + jimmxred.asm jimmxint.asm jimmxfst.asm \ + jcqnt3dn.asm jf3dnflt.asm ji3dnflt.asm \ + jcqntsse.asm jfsseflt.asm jisseflt.asm \ + jccolss2.asm jdcolss2.asm jcgrass2.asm \ + jcsamss2.asm jdsamss2.asm jdmerss2.asm \ + jcqnts2i.asm jfss2fst.asm jfss2int.asm \ + jiss2red.asm jiss2int.asm jiss2fst.asm \ + jcqnts2f.asm jiss2flt.asm + +jccolmmx.lo: jcclrmmx.asm +jcgrammx.lo: jcgrymmx.asm +jccolss2.lo: jcclrss2.asm +jcgrass2.lo: jcgryss2.asm +jdcolmmx.lo: jdclrmmx.asm +jdcolss2.lo: jdclrss2.asm +jdmermmx.lo: jdmrgmmx.asm +jdmerss2.lo: jdmrgss2.asm +endif + +if SIMD_ARM + +libsimd_la_SOURCES = jsimd_arm.c jsimd_arm_neon.S + +endif + +AM_CPPFLAGS = -I$(top_srcdir) + +.asm.lo: + $(LIBTOOL) --mode=compile --tag NASM $(srcdir)/nasm_lt.sh $(NASM) $(NAFLAGS) -I$(srcdir) -I. $< -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/Makefile.in b/simd/Makefile.in new file mode 100644 index 0000000..d40e96c --- /dev/null +++ b/simd/Makefile.in @@ -0,0 +1,590 @@ +# Makefile.in generated by automake 1.9.2 from Makefile.am. +# @configure_input@ + +# Copyright (C) 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, +# 2003, 2004 Free Software Foundation, Inc. +# This Makefile.in is free software; the Free Software Foundation +# gives unlimited permission to copy and/or distribute it, +# with or without modifications, as long as this notice is preserved. + +# This program is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY, to the extent permitted by law; without +# even the implied warranty of MERCHANTABILITY or FITNESS FOR A +# PARTICULAR PURPOSE. + +@SET_MAKE@ + +SOURCES = $(libsimd_la_SOURCES) + +srcdir = @srcdir@ +top_srcdir = @top_srcdir@ +VPATH = @srcdir@ +pkgdatadir = $(datadir)/@PACKAGE@ +pkglibdir = $(libdir)/@PACKAGE@ +pkgincludedir = $(includedir)/@PACKAGE@ +top_builddir = .. +am__cd = CDPATH="$${ZSH_VERSION+.}$(PATH_SEPARATOR)" && cd +INSTALL = @INSTALL@ +install_sh_DATA = $(install_sh) -c -m 644 +install_sh_PROGRAM = $(install_sh) -c +install_sh_SCRIPT = $(install_sh) -c +INSTALL_HEADER = $(INSTALL_DATA) +transform = $(program_transform_name) +NORMAL_INSTALL = : +PRE_INSTALL = : +POST_INSTALL = : +NORMAL_UNINSTALL = : +PRE_UNINSTALL = : +POST_UNINSTALL = : +build_triplet = @build@ +host_triplet = @host@ +subdir = simd +DIST_COMMON = $(srcdir)/Makefile.am $(srcdir)/Makefile.in +ACLOCAL_M4 = $(top_srcdir)/aclocal.m4 +am__aclocal_m4_deps = $(top_srcdir)/acinclude.m4 \ + $(top_srcdir)/configure.ac +am__configure_deps = $(am__aclocal_m4_deps) $(CONFIGURE_DEPENDENCIES) \ + $(ACLOCAL_M4) +mkinstalldirs = $(install_sh) -d +CONFIG_HEADER = $(top_builddir)/config.h $(top_builddir)/jconfig.h +CONFIG_CLEAN_FILES = +LTLIBRARIES = $(noinst_LTLIBRARIES) +libsimd_la_LIBADD = +am__libsimd_la_SOURCES_DIST = jsimd_arm.c jsimd_arm_neon.S \ + jsimd_i386.c jsimd.h jsimdcfg.inc.h jsimdext.inc jcolsamp.inc \ + jdct.inc jsimdcpu.asm jccolmmx.asm jdcolmmx.asm jcgrammx.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 jcgrass2.asm jcsamss2.asm jdsamss2.asm \ + jdmerss2.asm jcqnts2i.asm jfss2fst.asm jfss2int.asm \ + jiss2red.asm jiss2int.asm jiss2fst.asm jcqnts2f.asm \ + jiss2flt.asm jsimd_x86_64.c jfsseflt-64.asm jcgrass2-64.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 +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@am_libsimd_la_OBJECTS = jsimd_x86_64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jfsseflt-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jcgrass2-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jccolss2-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jdcolss2-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jcsamss2-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jdsamss2-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jdmerss2-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jcqnts2i-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jfss2fst-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jfss2int-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jiss2red-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jiss2int-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jiss2fst-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jcqnts2f-64.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_FALSE@@SIMD_X86_64_TRUE@ jiss2flt-64.lo +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@am_libsimd_la_OBJECTS = jsimd_i386.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jsimdcpu.lo jccolmmx.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jdcolmmx.lo jcgrammx.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jcsammmx.lo jdsammmx.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jdmermmx.lo jcqntmmx.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jfmmxfst.lo jfmmxint.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jimmxred.lo jimmxint.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jimmxfst.lo jcqnt3dn.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jf3dnflt.lo ji3dnflt.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jcqntsse.lo jfsseflt.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jisseflt.lo jccolss2.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jdcolss2.lo jcgrass2.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jcsamss2.lo jdsamss2.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jdmerss2.lo jcqnts2i.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jfss2fst.lo jfss2int.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jiss2red.lo jiss2int.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jiss2fst.lo jcqnts2f.lo \ +@SIMD_ARM_FALSE@@SIMD_I386_TRUE@ jiss2flt.lo +@SIMD_ARM_TRUE@am_libsimd_la_OBJECTS = jsimd_arm.lo jsimd_arm_neon.lo +libsimd_la_OBJECTS = $(am_libsimd_la_OBJECTS) +DEFAULT_INCLUDES = -I. -I$(srcdir) -I$(top_builddir) -I$(top_builddir) +depcomp = $(SHELL) $(top_srcdir)/depcomp +am__depfiles_maybe = depfiles +CCASCOMPILE = $(CCAS) $(AM_CCASFLAGS) $(CCASFLAGS) +LTCCASCOMPILE = $(LIBTOOL) --mode=compile $(CCAS) $(AM_CCASFLAGS) \ + $(CCASFLAGS) +COMPILE = $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) \ + $(CPPFLAGS) $(AM_CFLAGS) $(CFLAGS) +LTCOMPILE = $(LIBTOOL) --mode=compile --tag=CC $(CC) $(DEFS) \ + $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $(CPPFLAGS) \ + $(AM_CFLAGS) $(CFLAGS) +CCLD = $(CC) +LINK = $(LIBTOOL) --mode=link --tag=CC $(CCLD) $(AM_CFLAGS) $(CFLAGS) \ + $(AM_LDFLAGS) $(LDFLAGS) -o $@ +SOURCES = $(libsimd_la_SOURCES) +DIST_SOURCES = $(am__libsimd_la_SOURCES_DIST) +ETAGS = etags +CTAGS = ctags +DISTFILES = $(DIST_COMMON) $(DIST_SOURCES) $(TEXINFOS) $(EXTRA_DIST) +ACLOCAL = @ACLOCAL@ +AMDEP_FALSE = @AMDEP_FALSE@ +AMDEP_TRUE = @AMDEP_TRUE@ +AMTAR = @AMTAR@ +AR = @AR@ +AUTOCONF = @AUTOCONF@ +AUTOHEADER = @AUTOHEADER@ +AUTOMAKE = @AUTOMAKE@ +AWK = @AWK@ +BUILD = @BUILD@ +CC = @CC@ +CCAS = @CCAS@ +CCASFLAGS = @CCASFLAGS@ +CCDEPMODE = @CCDEPMODE@ +CFLAGS = @CFLAGS@ +CPP = @CPP@ +CPPFLAGS = @CPPFLAGS@ +CXX = @CXX@ +CXXCPP = @CXXCPP@ +CXXDEPMODE = @CXXDEPMODE@ +CXXFLAGS = @CXXFLAGS@ +CYGPATH_W = @CYGPATH_W@ +DEBARCH = @DEBARCH@ +DEFS = @DEFS@ +DEPDIR = @DEPDIR@ +ECHO = @ECHO@ +ECHO_C = @ECHO_C@ +ECHO_N = @ECHO_N@ +ECHO_T = @ECHO_T@ +EGREP = @EGREP@ +EXEEXT = @EXEEXT@ +F77 = @F77@ +FFLAGS = @FFLAGS@ +INSTALL_DATA = @INSTALL_DATA@ +INSTALL_PROGRAM = @INSTALL_PROGRAM@ +INSTALL_SCRIPT = @INSTALL_SCRIPT@ +INSTALL_STRIP_PROGRAM = @INSTALL_STRIP_PROGRAM@ +JAR = @JAR@ +JAVA = @JAVA@ +JAVAC = @JAVAC@ +JAVACFLAGS = @JAVACFLAGS@ +JAVA_RPM_CONTENTS_1 = @JAVA_RPM_CONTENTS_1@ +JAVA_RPM_CONTENTS_2 = @JAVA_RPM_CONTENTS_2@ +JNI_CFLAGS = @JNI_CFLAGS@ +JPEG_LIB_VERSION = @JPEG_LIB_VERSION@ +JPEG_LIB_VERSION_DECIMAL = @JPEG_LIB_VERSION_DECIMAL@ +LDFLAGS = @LDFLAGS@ +LIBOBJS = @LIBOBJS@ +LIBS = @LIBS@ +LIBTOOL = @LIBTOOL@ +LIBTOOL_CURRENT = @LIBTOOL_CURRENT@ +LN_S = @LN_S@ +LTLIBOBJS = @LTLIBOBJS@ +MAKEINFO = @MAKEINFO@ +MEM_SRCDST_FUNCTIONS = @MEM_SRCDST_FUNCTIONS@ +NAFLAGS = @NAFLAGS@ +NASM = @NASM@ +OBJEXT = @OBJEXT@ +PACKAGE = @PACKAGE@ +PACKAGE_BUGREPORT = @PACKAGE_BUGREPORT@ +PACKAGE_NAME = @PACKAGE_NAME@ +PACKAGE_STRING = @PACKAGE_STRING@ +PACKAGE_TARNAME = @PACKAGE_TARNAME@ +PACKAGE_VERSION = @PACKAGE_VERSION@ +PATH_SEPARATOR = @PATH_SEPARATOR@ +PKGNAME = @PKGNAME@ +RANLIB = @RANLIB@ +RPMARCH = @RPMARCH@ +RPM_CONFIG_ARGS = @RPM_CONFIG_ARGS@ +SET_MAKE = @SET_MAKE@ +SHELL = @SHELL@ +SIMD_ARM_FALSE = @SIMD_ARM_FALSE@ +SIMD_ARM_TRUE = @SIMD_ARM_TRUE@ +SIMD_I386_FALSE = @SIMD_I386_FALSE@ +SIMD_I386_TRUE = @SIMD_I386_TRUE@ +SIMD_X86_64_FALSE = @SIMD_X86_64_FALSE@ +SIMD_X86_64_TRUE = @SIMD_X86_64_TRUE@ +SO_AGE = @SO_AGE@ +SO_MAJOR_VERSION = @SO_MAJOR_VERSION@ +SO_MINOR_VERSION = @SO_MINOR_VERSION@ +STRIP = @STRIP@ +VERSION = @VERSION@ +VERSION_SCRIPT_FALSE = @VERSION_SCRIPT_FALSE@ +VERSION_SCRIPT_FLAG = @VERSION_SCRIPT_FLAG@ +VERSION_SCRIPT_TRUE = @VERSION_SCRIPT_TRUE@ +WITH_ARITH_DEC_FALSE = @WITH_ARITH_DEC_FALSE@ +WITH_ARITH_DEC_TRUE = @WITH_ARITH_DEC_TRUE@ +WITH_ARITH_ENC_FALSE = @WITH_ARITH_ENC_FALSE@ +WITH_ARITH_ENC_TRUE = @WITH_ARITH_ENC_TRUE@ +WITH_ARITH_FALSE = @WITH_ARITH_FALSE@ +WITH_ARITH_TRUE = @WITH_ARITH_TRUE@ +WITH_JAVA = @WITH_JAVA@ +WITH_JAVA_FALSE = @WITH_JAVA_FALSE@ +WITH_JAVA_TRUE = @WITH_JAVA_TRUE@ +WITH_SIMD_FALSE = @WITH_SIMD_FALSE@ +WITH_SIMD_TRUE = @WITH_SIMD_TRUE@ +WITH_SSE_FLOAT_DCT_FALSE = @WITH_SSE_FLOAT_DCT_FALSE@ +WITH_SSE_FLOAT_DCT_TRUE = @WITH_SSE_FLOAT_DCT_TRUE@ +WITH_TURBOJPEG_FALSE = @WITH_TURBOJPEG_FALSE@ +WITH_TURBOJPEG_TRUE = @WITH_TURBOJPEG_TRUE@ +X86_64_FALSE = @X86_64_FALSE@ +X86_64_TRUE = @X86_64_TRUE@ +ac_ct_AR = @ac_ct_AR@ +ac_ct_CC = @ac_ct_CC@ +ac_ct_CXX = @ac_ct_CXX@ +ac_ct_F77 = @ac_ct_F77@ +ac_ct_RANLIB = @ac_ct_RANLIB@ +ac_ct_STRIP = @ac_ct_STRIP@ +am__fastdepCC_FALSE = @am__fastdepCC_FALSE@ +am__fastdepCC_TRUE = @am__fastdepCC_TRUE@ +am__fastdepCXX_FALSE = @am__fastdepCXX_FALSE@ +am__fastdepCXX_TRUE = @am__fastdepCXX_TRUE@ +am__include = @am__include@ +am__leading_dot = @am__leading_dot@ +am__quote = @am__quote@ +am__tar = @am__tar@ +am__untar = @am__untar@ +bindir = @bindir@ +build = @build@ +build_alias = @build_alias@ +build_cpu = @build_cpu@ +build_os = @build_os@ +build_vendor = @build_vendor@ +datadir = @datadir@ +exec_prefix = @exec_prefix@ +host = @host@ +host_alias = @host_alias@ +host_cpu = @host_cpu@ +host_os = @host_os@ +host_vendor = @host_vendor@ +includedir = @includedir@ +infodir = @infodir@ +install_sh = @install_sh@ +libdir = @libdir@ +libexecdir = @libexecdir@ +localstatedir = @localstatedir@ +mandir = @mandir@ +mkdir_p = @mkdir_p@ +oldincludedir = @oldincludedir@ +prefix = @prefix@ +program_transform_name = @program_transform_name@ +sbindir = @sbindir@ +sharedstatedir = @sharedstatedir@ +sysconfdir = @sysconfdir@ +target_alias = @target_alias@ +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 jcgryss2-64.asm jcgrymmx.asm jcgryss2.asm CMakeLists.txt + +@SIMD_ARM_TRUE@libsimd_la_SOURCES = jsimd_arm.c jsimd_arm_neon.S +@SIMD_I386_TRUE@libsimd_la_SOURCES = jsimd_i386.c \ +@SIMD_I386_TRUE@ jsimd.h jsimdcfg.inc.h \ +@SIMD_I386_TRUE@ jsimdext.inc jcolsamp.inc jdct.inc \ +@SIMD_I386_TRUE@ jsimdcpu.asm \ +@SIMD_I386_TRUE@ jccolmmx.asm jdcolmmx.asm jcgrammx.asm \ +@SIMD_I386_TRUE@ jcsammmx.asm jdsammmx.asm jdmermmx.asm \ +@SIMD_I386_TRUE@ jcqntmmx.asm jfmmxfst.asm jfmmxint.asm \ +@SIMD_I386_TRUE@ jimmxred.asm jimmxint.asm jimmxfst.asm \ +@SIMD_I386_TRUE@ jcqnt3dn.asm jf3dnflt.asm ji3dnflt.asm \ +@SIMD_I386_TRUE@ jcqntsse.asm jfsseflt.asm jisseflt.asm \ +@SIMD_I386_TRUE@ jccolss2.asm jdcolss2.asm jcgrass2.asm \ +@SIMD_I386_TRUE@ jcsamss2.asm jdsamss2.asm jdmerss2.asm \ +@SIMD_I386_TRUE@ jcqnts2i.asm jfss2fst.asm jfss2int.asm \ +@SIMD_I386_TRUE@ jiss2red.asm jiss2int.asm jiss2fst.asm \ +@SIMD_I386_TRUE@ jcqnts2f.asm jiss2flt.asm + +@SIMD_X86_64_TRUE@libsimd_la_SOURCES = jsimd_x86_64.c \ +@SIMD_X86_64_TRUE@ jsimd.h jsimdcfg.inc.h \ +@SIMD_X86_64_TRUE@ jsimdext.inc jcolsamp.inc jdct.inc \ +@SIMD_X86_64_TRUE@ jfsseflt-64.asm jcgrass2-64.asm \ +@SIMD_X86_64_TRUE@ jccolss2-64.asm jdcolss2-64.asm \ +@SIMD_X86_64_TRUE@ jcsamss2-64.asm jdsamss2-64.asm jdmerss2-64.asm \ +@SIMD_X86_64_TRUE@ jcqnts2i-64.asm jfss2fst-64.asm jfss2int-64.asm \ +@SIMD_X86_64_TRUE@ jiss2red-64.asm jiss2int-64.asm jiss2fst-64.asm \ +@SIMD_X86_64_TRUE@ jcqnts2f-64.asm jiss2flt-64.asm + +AM_CPPFLAGS = -I$(top_srcdir) +all: $(BUILT_SOURCES) + $(MAKE) $(AM_MAKEFLAGS) all-am + +.SUFFIXES: +.SUFFIXES: .S .asm .c .lo .o .obj +$(srcdir)/Makefile.in: $(srcdir)/Makefile.am $(am__configure_deps) + @for dep in $?; do \ + case '$(am__configure_deps)' in \ + *$$dep*) \ + cd $(top_builddir) && $(MAKE) $(AM_MAKEFLAGS) am--refresh \ + && exit 0; \ + exit 1;; \ + esac; \ + done; \ + echo ' cd $(top_srcdir) && $(AUTOMAKE) --foreign simd/Makefile'; \ + cd $(top_srcdir) && \ + $(AUTOMAKE) --foreign simd/Makefile +.PRECIOUS: Makefile +Makefile: $(srcdir)/Makefile.in $(top_builddir)/config.status + @case '$?' in \ + *config.status*) \ + cd $(top_builddir) && $(MAKE) $(AM_MAKEFLAGS) am--refresh;; \ + *) \ + echo ' cd $(top_builddir) && $(SHELL) ./config.status $(subdir)/$@ $(am__depfiles_maybe)'; \ + cd $(top_builddir) && $(SHELL) ./config.status $(subdir)/$@ $(am__depfiles_maybe);; \ + esac; + +$(top_builddir)/config.status: $(top_srcdir)/configure $(CONFIG_STATUS_DEPENDENCIES) + cd $(top_builddir) && $(MAKE) $(AM_MAKEFLAGS) am--refresh + +$(top_srcdir)/configure: $(am__configure_deps) + cd $(top_builddir) && $(MAKE) $(AM_MAKEFLAGS) am--refresh +$(ACLOCAL_M4): $(am__aclocal_m4_deps) + cd $(top_builddir) && $(MAKE) $(AM_MAKEFLAGS) am--refresh + +clean-noinstLTLIBRARIES: + -test -z "$(noinst_LTLIBRARIES)" || rm -f $(noinst_LTLIBRARIES) + @list='$(noinst_LTLIBRARIES)'; for p in $$list; do \ + dir="`echo $$p | sed -e 's|/[^/]*$$||'`"; \ + test "$$dir" != "$$p" || dir=.; \ + echo "rm -f \"$${dir}/so_locations\""; \ + rm -f "$${dir}/so_locations"; \ + done +libsimd.la: $(libsimd_la_OBJECTS) $(libsimd_la_DEPENDENCIES) + $(LINK) $(libsimd_la_LDFLAGS) $(libsimd_la_OBJECTS) $(libsimd_la_LIBADD) $(LIBS) + +mostlyclean-compile: + -rm -f *.$(OBJEXT) + +distclean-compile: + -rm -f *.tab.c + +@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/jsimd_arm.Plo@am__quote@ +@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/jsimd_i386.Plo@am__quote@ +@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/jsimd_x86_64.Plo@am__quote@ + +.S.o: + $(CCASCOMPILE) -c $< + +.S.obj: + $(CCASCOMPILE) -c `$(CYGPATH_W) '$<'` + +.S.lo: + $(LTCCASCOMPILE) -c -o $@ $< + +.c.o: +@am__fastdepCC_TRUE@ if $(COMPILE) -MT $@ -MD -MP -MF "$(DEPDIR)/$*.Tpo" -c -o $@ $<; \ +@am__fastdepCC_TRUE@ then mv -f "$(DEPDIR)/$*.Tpo" "$(DEPDIR)/$*.Po"; else rm -f "$(DEPDIR)/$*.Tpo"; exit 1; fi +@AMDEP_TRUE@@am__fastdepCC_FALSE@ source='$<' object='$@' libtool=no @AMDEPBACKSLASH@ +@AMDEP_TRUE@@am__fastdepCC_FALSE@ DEPDIR=$(DEPDIR) $(CCDEPMODE) $(depcomp) @AMDEPBACKSLASH@ +@am__fastdepCC_FALSE@ $(COMPILE) -c $< + +.c.obj: +@am__fastdepCC_TRUE@ if $(COMPILE) -MT $@ -MD -MP -MF "$(DEPDIR)/$*.Tpo" -c -o $@ `$(CYGPATH_W) '$<'`; \ +@am__fastdepCC_TRUE@ then mv -f "$(DEPDIR)/$*.Tpo" "$(DEPDIR)/$*.Po"; else rm -f "$(DEPDIR)/$*.Tpo"; exit 1; fi +@AMDEP_TRUE@@am__fastdepCC_FALSE@ source='$<' object='$@' libtool=no @AMDEPBACKSLASH@ +@AMDEP_TRUE@@am__fastdepCC_FALSE@ DEPDIR=$(DEPDIR) $(CCDEPMODE) $(depcomp) @AMDEPBACKSLASH@ +@am__fastdepCC_FALSE@ $(COMPILE) -c `$(CYGPATH_W) '$<'` + +.c.lo: +@am__fastdepCC_TRUE@ if $(LTCOMPILE) -MT $@ -MD -MP -MF "$(DEPDIR)/$*.Tpo" -c -o $@ $<; \ +@am__fastdepCC_TRUE@ then mv -f "$(DEPDIR)/$*.Tpo" "$(DEPDIR)/$*.Plo"; else rm -f "$(DEPDIR)/$*.Tpo"; exit 1; fi +@AMDEP_TRUE@@am__fastdepCC_FALSE@ source='$<' object='$@' libtool=yes @AMDEPBACKSLASH@ +@AMDEP_TRUE@@am__fastdepCC_FALSE@ DEPDIR=$(DEPDIR) $(CCDEPMODE) $(depcomp) @AMDEPBACKSLASH@ +@am__fastdepCC_FALSE@ $(LTCOMPILE) -c -o $@ $< + +mostlyclean-libtool: + -rm -f *.lo + +clean-libtool: + -rm -rf .libs _libs + +distclean-libtool: + -rm -f libtool +uninstall-info-am: + +ID: $(HEADERS) $(SOURCES) $(LISP) $(TAGS_FILES) + list='$(SOURCES) $(HEADERS) $(LISP) $(TAGS_FILES)'; \ + unique=`for i in $$list; do \ + if test -f "$$i"; then echo $$i; else echo $(srcdir)/$$i; fi; \ + done | \ + $(AWK) ' { files[$$0] = 1; } \ + END { for (i in files) print i; }'`; \ + mkid -fID $$unique +tags: TAGS + +TAGS: $(HEADERS) $(SOURCES) $(TAGS_DEPENDENCIES) \ + $(TAGS_FILES) $(LISP) + tags=; \ + here=`pwd`; \ + list='$(SOURCES) $(HEADERS) $(LISP) $(TAGS_FILES)'; \ + unique=`for i in $$list; do \ + if test -f "$$i"; then echo $$i; else echo $(srcdir)/$$i; fi; \ + done | \ + $(AWK) ' { files[$$0] = 1; } \ + END { for (i in files) print i; }'`; \ + if test -z "$(ETAGS_ARGS)$$tags$$unique"; then :; else \ + test -n "$$unique" || unique=$$empty_fix; \ + $(ETAGS) $(ETAGSFLAGS) $(AM_ETAGSFLAGS) $(ETAGS_ARGS) \ + $$tags $$unique; \ + fi +ctags: CTAGS +CTAGS: $(HEADERS) $(SOURCES) $(TAGS_DEPENDENCIES) \ + $(TAGS_FILES) $(LISP) + tags=; \ + here=`pwd`; \ + list='$(SOURCES) $(HEADERS) $(LISP) $(TAGS_FILES)'; \ + unique=`for i in $$list; do \ + if test -f "$$i"; then echo $$i; else echo $(srcdir)/$$i; fi; \ + done | \ + $(AWK) ' { files[$$0] = 1; } \ + END { for (i in files) print i; }'`; \ + test -z "$(CTAGS_ARGS)$$tags$$unique" \ + || $(CTAGS) $(CTAGSFLAGS) $(AM_CTAGSFLAGS) $(CTAGS_ARGS) \ + $$tags $$unique + +GTAGS: + here=`$(am__cd) $(top_builddir) && pwd` \ + && cd $(top_srcdir) \ + && gtags -i $(GTAGS_ARGS) $$here + +distclean-tags: + -rm -f TAGS ID GTAGS GRTAGS GSYMS GPATH tags + +distdir: $(DISTFILES) + @srcdirstrip=`echo "$(srcdir)" | sed 's|.|.|g'`; \ + topsrcdirstrip=`echo "$(top_srcdir)" | sed 's|.|.|g'`; \ + list='$(DISTFILES)'; for file in $$list; do \ + case $$file in \ + $(srcdir)/*) file=`echo "$$file" | sed "s|^$$srcdirstrip/||"`;; \ + $(top_srcdir)/*) file=`echo "$$file" | sed "s|^$$topsrcdirstrip/|$(top_builddir)/|"`;; \ + esac; \ + if test -f $$file || test -d $$file; then d=.; else d=$(srcdir); fi; \ + dir=`echo "$$file" | sed -e 's,/[^/]*$$,,'`; \ + if test "$$dir" != "$$file" && test "$$dir" != "."; then \ + dir="/$$dir"; \ + $(mkdir_p) "$(distdir)$$dir"; \ + else \ + dir=''; \ + fi; \ + if test -d $$d/$$file; then \ + if test -d $(srcdir)/$$file && test $$d != $(srcdir); then \ + cp -pR $(srcdir)/$$file $(distdir)$$dir || exit 1; \ + fi; \ + cp -pR $$d/$$file $(distdir)$$dir || exit 1; \ + else \ + test -f $(distdir)/$$file \ + || cp -p $$d/$$file $(distdir)/$$file \ + || exit 1; \ + fi; \ + done +check-am: all-am +check: $(BUILT_SOURCES) + $(MAKE) $(AM_MAKEFLAGS) check-am +all-am: Makefile $(LTLIBRARIES) +installdirs: +install: $(BUILT_SOURCES) + $(MAKE) $(AM_MAKEFLAGS) install-am +install-exec: install-exec-am +install-data: install-data-am +uninstall: uninstall-am + +install-am: all-am + @$(MAKE) $(AM_MAKEFLAGS) install-exec-am install-data-am + +installcheck: installcheck-am +install-strip: + $(MAKE) $(AM_MAKEFLAGS) INSTALL_PROGRAM="$(INSTALL_STRIP_PROGRAM)" \ + install_sh_PROGRAM="$(INSTALL_STRIP_PROGRAM)" INSTALL_STRIP_FLAG=-s \ + `test -z '$(STRIP)' || \ + echo "INSTALL_PROGRAM_ENV=STRIPPROG='$(STRIP)'"` install +mostlyclean-generic: + +clean-generic: + +distclean-generic: + -test -z "$(CONFIG_CLEAN_FILES)" || rm -f $(CONFIG_CLEAN_FILES) + +maintainer-clean-generic: + @echo "This command is intended for maintainers to use" + @echo "it deletes files that may require special tools to rebuild." + -test -z "$(BUILT_SOURCES)" || rm -f $(BUILT_SOURCES) +clean: clean-am + +clean-am: clean-generic clean-libtool clean-noinstLTLIBRARIES \ + mostlyclean-am + +distclean: distclean-am + -rm -rf ./$(DEPDIR) + -rm -f Makefile +distclean-am: clean-am distclean-compile distclean-generic \ + distclean-libtool distclean-tags + +dvi: dvi-am + +dvi-am: + +html: html-am + +info: info-am + +info-am: + +install-data-am: + +install-exec-am: + +install-info: install-info-am + +install-man: + +installcheck-am: + +maintainer-clean: maintainer-clean-am + -rm -rf ./$(DEPDIR) + -rm -f Makefile +maintainer-clean-am: distclean-am maintainer-clean-generic + +mostlyclean: mostlyclean-am + +mostlyclean-am: mostlyclean-compile mostlyclean-generic \ + mostlyclean-libtool + +pdf: pdf-am + +pdf-am: + +ps: ps-am + +ps-am: + +uninstall-am: uninstall-info-am + +.PHONY: CTAGS GTAGS all all-am check check-am clean clean-generic \ + clean-libtool clean-noinstLTLIBRARIES ctags distclean \ + distclean-compile distclean-generic distclean-libtool \ + distclean-tags distdir dvi dvi-am html html-am info info-am \ + install install-am install-data install-data-am install-exec \ + install-exec-am install-info install-info-am install-man \ + install-strip installcheck installcheck-am installdirs \ + maintainer-clean maintainer-clean-generic mostlyclean \ + mostlyclean-compile mostlyclean-generic mostlyclean-libtool \ + pdf pdf-am ps ps-am tags uninstall uninstall-am \ + uninstall-info-am + + +@SIMD_X86_64_TRUE@jccolss2-64.lo: jcclrss2-64.asm +@SIMD_X86_64_TRUE@jdcolss2-64.lo: jdclrss2-64.asm +@SIMD_X86_64_TRUE@jcgrass2-64.lo: jcgryss2-64.asm +@SIMD_X86_64_TRUE@jdmerss2-64.lo: jdmrgss2-64.asm + +@SIMD_I386_TRUE@jccolmmx.lo: jcclrmmx.asm +@SIMD_I386_TRUE@jcgrammx.lo: jcgrymmx.asm +@SIMD_I386_TRUE@jccolss2.lo: jcclrss2.asm +@SIMD_I386_TRUE@jcgrass2.lo: jcgryss2.asm +@SIMD_I386_TRUE@jdcolmmx.lo: jdclrmmx.asm +@SIMD_I386_TRUE@jdcolss2.lo: jdclrss2.asm +@SIMD_I386_TRUE@jdmermmx.lo: jdmrgmmx.asm +@SIMD_I386_TRUE@jdmerss2.lo: jdmrgss2.asm + +.asm.lo: + $(LIBTOOL) --mode=compile --tag NASM $(srcdir)/nasm_lt.sh $(NASM) $(NAFLAGS) -I$(srcdir) -I. $< -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' > $@ +# Tell versions [3.59,3.63) of GNU make to not export all variables. +# Otherwise a system limit (for SysV at least) may be exceeded. +.NOEXPORT: diff --git a/simd/jcclrmmx.asm b/simd/jcclrmmx.asm new file mode 100644 index 0000000..e095253 --- /dev/null +++ b/simd/jcclrmmx.asm @@ -0,0 +1,477 @@ +; +; 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" + +; -------------------------------------------------------------------------- +; +; 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..f5d6bed --- /dev/null +++ b/simd/jcclrss2-64.asm @@ -0,0 +1,485 @@ +; +; 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" + +; -------------------------------------------------------------------------- +; +; 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..517b705 --- /dev/null +++ b/simd/jcclrss2.asm @@ -0,0 +1,503 @@ +; +; 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" + +; -------------------------------------------------------------------------- +; +; 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..9650e47 --- /dev/null +++ b/simd/jccolmmx.asm @@ -0,0 +1,123 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jcclrmmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..ae60148 --- /dev/null +++ b/simd/jccolss2-64.asm @@ -0,0 +1,120 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 64 + +%include "jcclrss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..ac001d1 --- /dev/null +++ b/simd/jccolss2.asm @@ -0,0 +1,120 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jcclrss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%define jsimd_rgb_ycc_convert_sse2 jsimd_extxrgb_ycc_convert_sse2 +%include "jcclrss2.asm" diff --git a/simd/jcgrammx.asm b/simd/jcgrammx.asm new file mode 100644 index 0000000..b8b8dd3 --- /dev/null +++ b/simd/jcgrammx.asm @@ -0,0 +1,116 @@ +; +; jcgrammx.asm - grayscale colorspace conversion (MMX) +; +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB +; Copyright 2011 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_114 equ 7471 ; FIX(0.11400) +F_0_250 equ 16384 ; FIX(0.25000) +F_0_299 equ 19595 ; FIX(0.29900) +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_gray_convert_mmx) + +EXTN(jconst_rgb_gray_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 +PD_ONEHALF times 2 dd (1 << (SCALEBITS-1)) + + alignz 16 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jcgrymmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%define jsimd_rgb_gray_convert_mmx jsimd_extrgb_gray_convert_mmx +%include "jcgrymmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%define jsimd_rgb_gray_convert_mmx jsimd_extrgbx_gray_convert_mmx +%include "jcgrymmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%define jsimd_rgb_gray_convert_mmx jsimd_extbgr_gray_convert_mmx +%include "jcgrymmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%define jsimd_rgb_gray_convert_mmx jsimd_extbgrx_gray_convert_mmx +%include "jcgrymmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%define jsimd_rgb_gray_convert_mmx jsimd_extxbgr_gray_convert_mmx +%include "jcgrymmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%define jsimd_rgb_gray_convert_mmx jsimd_extxrgb_gray_convert_mmx +%include "jcgrymmx.asm" diff --git a/simd/jcgrass2-64.asm b/simd/jcgrass2-64.asm new file mode 100644 index 0000000..ba28cc3 --- /dev/null +++ b/simd/jcgrass2-64.asm @@ -0,0 +1,113 @@ +; +; jcgrass2-64.asm - grayscale colorspace conversion (64-bit SSE2) +; +; x86 SIMD extension for IJG JPEG library +; Copyright (C) 1999-2006, MIYASAKA Masaru. +; Copyright (C) 2011, 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_114 equ 7471 ; FIX(0.11400) +F_0_250 equ 16384 ; FIX(0.25000) +F_0_299 equ 19595 ; FIX(0.29900) +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_gray_convert_sse2) + +EXTN(jconst_rgb_gray_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 +PD_ONEHALF times 4 dd (1 << (SCALEBITS-1)) + + alignz 16 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 64 + +%include "jcgryss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extrgb_gray_convert_sse2 +%include "jcgryss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extrgbx_gray_convert_sse2 +%include "jcgryss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extbgr_gray_convert_sse2 +%include "jcgryss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extbgrx_gray_convert_sse2 +%include "jcgryss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extxbgr_gray_convert_sse2 +%include "jcgryss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extxrgb_gray_convert_sse2 +%include "jcgryss2-64.asm" diff --git a/simd/jcgrass2.asm b/simd/jcgrass2.asm new file mode 100644 index 0000000..998968e --- /dev/null +++ b/simd/jcgrass2.asm @@ -0,0 +1,113 @@ +; +; jcgrass2.asm - grayscale colorspace conversion (SSE2) +; +; x86 SIMD extension for IJG JPEG library +; Copyright (C) 1999-2006, MIYASAKA Masaru. +; Copyright (C) 2011, 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_114 equ 7471 ; FIX(0.11400) +F_0_250 equ 16384 ; FIX(0.25000) +F_0_299 equ 19595 ; FIX(0.29900) +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_gray_convert_sse2) + +EXTN(jconst_rgb_gray_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 +PD_ONEHALF times 4 dd (1 << (SCALEBITS-1)) + + alignz 16 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jcgryss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extrgb_gray_convert_sse2 +%include "jcgryss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extrgbx_gray_convert_sse2 +%include "jcgryss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extbgr_gray_convert_sse2 +%include "jcgryss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extbgrx_gray_convert_sse2 +%include "jcgryss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extxbgr_gray_convert_sse2 +%include "jcgryss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%define jsimd_rgb_gray_convert_sse2 jsimd_extxrgb_gray_convert_sse2 +%include "jcgryss2.asm" diff --git a/simd/jcgrymmx.asm b/simd/jcgrymmx.asm new file mode 100644 index 0000000..bbeea09 --- /dev/null +++ b/simd/jcgrymmx.asm @@ -0,0 +1,357 @@ +; +; jcgrymmx.asm - grayscale colorspace conversion (MMX) +; +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB +; Copyright 2011 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" + +; -------------------------------------------------------------------------- +; +; Convert some rows of samples to the output colorspace. +; +; GLOBAL(void) +; jsimd_rgb_gray_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 2 +%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr + + align 16 + global EXTN(jsimd_rgb_gray_convert_mmx) + +EXTN(jsimd_rgb_gray_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] + lea edi, [edi+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 edi + push esi + push ecx ; col + + mov esi, JSAMPROW [esi] ; inptr + mov edi, JSAMPROW [edi] ; outptr0 + 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_gray_cnv +.column_ld16: + test cl, 2*SIZEOF_MMWORD + mov ecx, SIZEOF_MMWORD + jz short .rgb_gray_cnv + movq mmF,mmA + movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] + movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] + jmp short .rgb_gray_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_gray_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_gray_cnv + movq mmD,mmA + movq mmC,mmF + movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] + movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] + jmp short .rgb_gray_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_gray_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 + ; + ; (This implementation) + ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G + + movq mm6,mm1 + punpcklwd mm1,mm3 + punpckhwd mm6,mm3 + 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) + + movq mm7, mm6 ; mm7=ROH*FIX(0.299)+GOH*FIX(0.337) + + movq mm6,mm0 + punpcklwd mm0,mm2 + punpckhwd mm6,mm2 + 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) + + movq MMWORD [wk(0)], mm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) + movq MMWORD [wk(1)], mm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) + + movq mm0, mm5 ; mm0=BO + movq mm6, mm4 ; mm6=BE + + movq mm4,mm0 + punpcklwd mm0,mm3 + punpckhwd mm4,mm3 + 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) + + movq mm3,[GOTOFF(eax,PD_ONEHALF)] ; mm3=[PD_ONEHALF] + + paddd mm0, mm1 + paddd mm4, mm7 + paddd mm0,mm3 + paddd mm4,mm3 + psrld mm0,SCALEBITS ; mm0=YOL + psrld mm4,SCALEBITS ; mm4=YOH + packssdw mm0,mm4 ; mm0=YO + + movq mm4,mm6 + punpcklwd mm6,mm2 + punpckhwd mm4,mm2 + 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) + + movq mm2,[GOTOFF(eax,PD_ONEHALF)] ; mm2=[PD_ONEHALF] + + paddd mm6, MMWORD [wk(0)] + paddd mm4, MMWORD [wk(1)] + 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 + + sub ecx, byte SIZEOF_MMWORD + add esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; inptr + add edi, byte SIZEOF_MMWORD ; outptr0 + cmp ecx, byte SIZEOF_MMWORD + jae near .columnloop + test ecx,ecx + jnz near .column_ld1 + + pop ecx ; col + pop esi + pop edi + poppic eax + + add esi, byte SIZEOF_JSAMPROW ; input_buf + add edi, 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/jcgryss2-64.asm b/simd/jcgryss2-64.asm new file mode 100644 index 0000000..23ae8af --- /dev/null +++ b/simd/jcgryss2-64.asm @@ -0,0 +1,364 @@ +; +; jcgryss2-64.asm - grayscale colorspace conversion (64-bit SSE2) +; +; x86 SIMD extension for IJG JPEG library +; Copyright (C) 1999-2006, MIYASAKA Masaru. +; Copyright (C) 2011, 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" + +; -------------------------------------------------------------------------- +; +; Convert some rows of samples to the output colorspace. +; +; GLOBAL(void) +; jsimd_rgb_gray_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 2 + + align 16 + + global EXTN(jsimd_rgb_gray_convert_sse2) + +EXTN(jsimd_rgb_gray_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] + lea rdi, [rdi+rcx*SIZEOF_JSAMPROW] + + pop rcx + + mov rsi, r11 + mov eax, r14d + test rax,rax + jle near .return +.rowloop: + push rdi + push rsi + push rcx ; col + + mov rsi, JSAMPROW [rsi] ; inptr + mov rdi, JSAMPROW [rdi] ; outptr0 + + 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_gray_cnv +.column_ld32: + test cl, 2*SIZEOF_XMMWORD + mov rcx, SIZEOF_XMMWORD + jz short .rgb_gray_cnv + movdqa xmmB,xmmA + movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] + movdqu xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD] + jmp short .rgb_gray_cnv + +.columnloop: + movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] + movdqu xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD] + movdqu xmmB, XMMWORD [rsi+2*SIZEOF_XMMWORD] + +.rgb_gray_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_gray_cnv + movdqa xmmF,xmmA + movdqa xmmH,xmmE + movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] + movdqu xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD] + jmp short .rgb_gray_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_gray_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 + ; + ; (This implementation) + ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G + + movdqa xmm6,xmm1 + punpcklwd xmm1,xmm3 + punpckhwd xmm6,xmm3 + 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) + + movdqa xmm7, xmm6 ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337) + + movdqa xmm6,xmm0 + punpcklwd xmm0,xmm2 + punpckhwd xmm6,xmm2 + 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) + + movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) + movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) + + movdqa xmm0, xmm5 ; xmm0=BO + movdqa xmm6, xmm4 ; xmm6=BE + + movdqa xmm4,xmm0 + punpcklwd xmm0,xmm3 + punpckhwd xmm4,xmm3 + 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) + + movdqa xmm3,[rel PD_ONEHALF] ; xmm3=[PD_ONEHALF] + + paddd xmm0, xmm1 + paddd xmm4, xmm7 + paddd xmm0,xmm3 + paddd xmm4,xmm3 + psrld xmm0,SCALEBITS ; xmm0=YOL + psrld xmm4,SCALEBITS ; xmm4=YOH + packssdw xmm0,xmm4 ; xmm0=YO + + movdqa xmm4,xmm6 + punpcklwd xmm6,xmm2 + punpckhwd xmm4,xmm2 + 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) + + movdqa xmm2,[rel PD_ONEHALF] ; xmm2=[PD_ONEHALF] + + paddd xmm6, XMMWORD [wk(0)] + paddd xmm4, XMMWORD [wk(1)] + 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 + + sub rcx, byte SIZEOF_XMMWORD + add rsi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; inptr + add rdi, byte SIZEOF_XMMWORD ; outptr0 + cmp rcx, byte SIZEOF_XMMWORD + jae near .columnloop + test rcx,rcx + jnz near .column_ld1 + + pop rcx ; col + pop rsi + pop rdi + + add rsi, byte SIZEOF_JSAMPROW ; input_buf + add rdi, 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/jcgryss2.asm b/simd/jcgryss2.asm new file mode 100644 index 0000000..c294287 --- /dev/null +++ b/simd/jcgryss2.asm @@ -0,0 +1,383 @@ +; +; jcgryss2.asm - grayscale colorspace conversion (SSE2) +; +; x86 SIMD extension for IJG JPEG library +; Copyright (C) 1999-2006, MIYASAKA Masaru. +; Copyright (C) 2011, 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" + +; -------------------------------------------------------------------------- +; +; Convert some rows of samples to the output colorspace. +; +; GLOBAL(void) +; jsimd_rgb_gray_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 2 +%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr + + align 16 + + global EXTN(jsimd_rgb_gray_convert_sse2) + +EXTN(jsimd_rgb_gray_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] + lea edi, [edi+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 edi + push esi + push ecx ; col + + mov esi, JSAMPROW [esi] ; inptr + mov edi, JSAMPROW [edi] ; outptr0 + 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_gray_cnv +.column_ld32: + test cl, 2*SIZEOF_XMMWORD + mov ecx, SIZEOF_XMMWORD + jz short .rgb_gray_cnv + movdqa xmmB,xmmA + movdqu xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD] + movdqu xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD] + jmp short .rgb_gray_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_gray_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_gray_cnv + movdqa xmmF,xmmA + movdqa xmmH,xmmE + movdqu xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD] + movdqu xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD] + jmp short .rgb_gray_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_gray_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 + ; + ; (This implementation) + ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G + + movdqa xmm6,xmm1 + punpcklwd xmm1,xmm3 + punpckhwd xmm6,xmm3 + 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) + + movdqa xmm7, xmm6 ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337) + + movdqa xmm6,xmm0 + punpcklwd xmm0,xmm2 + punpckhwd xmm6,xmm2 + 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) + + movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) + movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) + + movdqa xmm0, xmm5 ; xmm0=BO + movdqa xmm6, xmm4 ; xmm6=BE + + movdqa xmm4,xmm0 + punpcklwd xmm0,xmm3 + punpckhwd xmm4,xmm3 + 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) + + movdqa xmm3,[GOTOFF(eax,PD_ONEHALF)] ; xmm3=[PD_ONEHALF] + + paddd xmm0, xmm1 + paddd xmm4, xmm7 + paddd xmm0,xmm3 + paddd xmm4,xmm3 + psrld xmm0,SCALEBITS ; xmm0=YOL + psrld xmm4,SCALEBITS ; xmm4=YOH + packssdw xmm0,xmm4 ; xmm0=YO + + movdqa xmm4,xmm6 + punpcklwd xmm6,xmm2 + punpckhwd xmm4,xmm2 + 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) + + movdqa xmm2,[GOTOFF(eax,PD_ONEHALF)] ; xmm2=[PD_ONEHALF] + + paddd xmm6, XMMWORD [wk(0)] + paddd xmm4, XMMWORD [wk(1)] + 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 + + sub ecx, byte SIZEOF_XMMWORD + add esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; inptr + add edi, byte SIZEOF_XMMWORD ; outptr0 + cmp ecx, byte SIZEOF_XMMWORD + jae near .columnloop + test ecx,ecx + jnz near .column_ld1 + + pop ecx ; col + pop esi + pop edi + poppic eax + + add esi, byte SIZEOF_JSAMPROW ; input_buf + add edi, 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/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..1c255e8 --- /dev/null +++ b/simd/jdclrmmx.asm @@ -0,0 +1,405 @@ +; +; 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" + +; -------------------------------------------------------------------------- +; +; 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..7d17c52 --- /dev/null +++ b/simd/jdclrss2-64.asm @@ -0,0 +1,441 @@ +; +; jdclrss2-64.asm - colorspace conversion (64-bit SSE2) +; +; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB +; Copyright 2009, 2012 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" + +; -------------------------------------------------------------------------- +; +; 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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF +.out0: + add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE + cmp rcx, byte 2*SIZEOF_XMMWORD + jb short .column_st16 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + add rdi, byte 2*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 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + add rdi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub rcx, byte SIZEOF_XMMWORD +.column_st15: + ; Store the lower 8 bytes of xmmA to the output when it has enough + ; space. + cmp rcx, byte SIZEOF_MMWORD + jb short .column_st7 + movq XMM_MMWORD [rdi], xmmA + add rdi, byte SIZEOF_MMWORD + sub rcx, byte SIZEOF_MMWORD + psrldq xmmA, SIZEOF_MMWORD +.column_st7: + ; Store the lower 4 bytes of xmmA to the output when it has enough + ; space. + cmp rcx, byte SIZEOF_DWORD + jb short .column_st3 + movd XMM_DWORD [rdi], xmmA + add rdi, byte SIZEOF_DWORD + sub rcx, byte SIZEOF_DWORD + psrldq xmmA, SIZEOF_DWORD +.column_st3: + ; Store the lower 2 bytes of rax to the output when it has enough + ; space. + movd eax, xmmA + cmp rcx, byte SIZEOF_WORD + jb short .column_st1 + mov WORD [rdi], ax + add rdi, byte SIZEOF_WORD + sub rcx, byte SIZEOF_WORD + shr rax, 16 +.column_st1: + ; Store the lower 1 byte of rax to the output when it has enough + ; space. + test rcx, rcx + jz short .nextrow + mov BYTE [rdi], al + +%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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC + movdqu XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH +.out0: + add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + cmp rcx, byte SIZEOF_XMMWORD/2 + jb short .column_st16 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + add rdi, byte 2*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 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + add rdi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub rcx, byte SIZEOF_XMMWORD/4 +.column_st15: + ; Store two pixels (8 bytes) of xmmA to the output when it has enough + ; space. + cmp rcx, byte SIZEOF_XMMWORD/8 + jb short .column_st7 + movq MMWORD [rdi], xmmA + add rdi, byte SIZEOF_XMMWORD/8*4 + sub rcx, byte SIZEOF_XMMWORD/8 + psrldq xmmA, SIZEOF_XMMWORD/8*4 +.column_st7: + ; Store one pixel (4 bytes) of xmmA to the output when it has enough + ; space. + test rcx, rcx + jz short .nextrow + movd XMM_DWORD [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..97754cb --- /dev/null +++ b/simd/jdclrss2.asm @@ -0,0 +1,460 @@ +; +; jdclrss2.asm - colorspace conversion (SSE2) +; +; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB +; Copyright 2012 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" + +; -------------------------------------------------------------------------- +; +; 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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF +.out0: + add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE + cmp ecx, byte 2*SIZEOF_XMMWORD + jb short .column_st16 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + add edi, byte 2*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 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + add edi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub ecx, byte SIZEOF_XMMWORD +.column_st15: + ; Store the lower 8 bytes of xmmA to the output when it has enough + ; space. + cmp ecx, byte SIZEOF_MMWORD + jb short .column_st7 + movq XMM_MMWORD [edi], xmmA + add edi, byte SIZEOF_MMWORD + sub ecx, byte SIZEOF_MMWORD + psrldq xmmA, SIZEOF_MMWORD +.column_st7: + ; Store the lower 4 bytes of xmmA to the output when it has enough + ; space. + cmp ecx, byte SIZEOF_DWORD + jb short .column_st3 + movd XMM_DWORD [edi], xmmA + add edi, byte SIZEOF_DWORD + sub ecx, byte SIZEOF_DWORD + psrldq xmmA, SIZEOF_DWORD +.column_st3: + ; Store the lower 2 bytes of eax to the output when it has enough + ; space. + movd eax, xmmA + cmp ecx, byte SIZEOF_WORD + jb short .column_st1 + mov WORD [edi], ax + add edi, byte SIZEOF_WORD + sub ecx, byte SIZEOF_WORD + shr eax, 16 +.column_st1: + ; Store the lower 1 byte of eax to the output when it has enough + ; space. + test ecx, ecx + jz short .nextrow + mov BYTE [edi], al + +%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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC + movdqu XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH +.out0: + add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + cmp ecx, byte SIZEOF_XMMWORD/2 + jb short .column_st16 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + add edi, byte 2*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 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + add edi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub ecx, byte SIZEOF_XMMWORD/4 +.column_st15: + ; Store two pixels (8 bytes) of xmmA to the output when it has enough + ; space. + cmp ecx, byte SIZEOF_XMMWORD/8 + jb short .column_st7 + movq XMM_MMWORD [edi], xmmA + add edi, byte SIZEOF_XMMWORD/8*4 + sub ecx, byte SIZEOF_XMMWORD/8 + psrldq xmmA, SIZEOF_XMMWORD/8*4 +.column_st7: + ; Store one pixel (4 bytes) of xmmA to the output when it has enough + ; space. + test ecx, ecx + jz short .nextrow + movd XMM_DWORD [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..5e4e47d --- /dev/null +++ b/simd/jdcolmmx.asm @@ -0,0 +1,120 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jdclrmmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..01b3dce --- /dev/null +++ b/simd/jdcolss2-64.asm @@ -0,0 +1,120 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 64 + +%include "jdclrss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..1912d92 --- /dev/null +++ b/simd/jdcolss2.asm @@ -0,0 +1,120 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jdclrss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..7b86c74 --- /dev/null +++ b/simd/jdmermmx.asm @@ -0,0 +1,126 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jdmrgmmx.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..a184ea6 --- /dev/null +++ b/simd/jdmerss2-64.asm @@ -0,0 +1,126 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 64 + +%include "jdmrgss2-64.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..e536c80 --- /dev/null +++ b/simd/jdmerss2.asm @@ -0,0 +1,126 @@ +; +; 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 + +; -------------------------------------------------------------------------- + SECTION SEG_TEXT + BITS 32 + +%include "jdmrgss2.asm" + +%undef RGB_RED +%undef RGB_GREEN +%undef RGB_BLUE +%undef RGB_PIXELSIZE +%define RGB_RED EXT_RGB_RED +%define RGB_GREEN EXT_RGB_GREEN +%define RGB_BLUE EXT_RGB_BLUE +%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE +%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 EXT_RGBX_RED +%define RGB_GREEN EXT_RGBX_GREEN +%define RGB_BLUE EXT_RGBX_BLUE +%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE +%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 EXT_BGR_RED +%define RGB_GREEN EXT_BGR_GREEN +%define RGB_BLUE EXT_BGR_BLUE +%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE +%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 EXT_BGRX_RED +%define RGB_GREEN EXT_BGRX_GREEN +%define RGB_BLUE EXT_BGRX_BLUE +%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE +%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 EXT_XBGR_RED +%define RGB_GREEN EXT_XBGR_GREEN +%define RGB_BLUE EXT_XBGR_BLUE +%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE +%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 EXT_XRGB_RED +%define RGB_GREEN EXT_XRGB_GREEN +%define RGB_BLUE EXT_XRGB_BLUE +%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE +%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..d0800a7 --- /dev/null +++ b/simd/jdmrgmmx.asm @@ -0,0 +1,464 @@ +; +; 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" + +; -------------------------------------------------------------------------- +; +; 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..ffbf6b2 --- /dev/null +++ b/simd/jdmrgss2-64.asm @@ -0,0 +1,538 @@ +; +; jdmrgss2-64.asm - merged upsampling/color conversion (64-bit SSE2) +; +; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB +; Copyright 2009, 2012 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" + +; -------------------------------------------------------------------------- +; +; 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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF +.out0: + add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE + cmp rcx, byte 2*SIZEOF_XMMWORD + jb short .column_st16 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + add rdi, byte 2*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 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + add rdi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub rcx, byte SIZEOF_XMMWORD +.column_st15: + ; Store the lower 8 bytes of xmmA to the output when it has enough + ; space. + cmp rcx, byte SIZEOF_MMWORD + jb short .column_st7 + movq XMM_MMWORD [rdi], xmmA + add rdi, byte SIZEOF_MMWORD + sub rcx, byte SIZEOF_MMWORD + psrldq xmmA, SIZEOF_MMWORD +.column_st7: + ; Store the lower 4 bytes of xmmA to the output when it has enough + ; space. + cmp rcx, byte SIZEOF_DWORD + jb short .column_st3 + movd XMM_DWORD [rdi], xmmA + add rdi, byte SIZEOF_DWORD + sub rcx, byte SIZEOF_DWORD + psrldq xmmA, SIZEOF_DWORD +.column_st3: + ; Store the lower 2 bytes of rax to the output when it has enough + ; space. + movd eax, xmmA + cmp rcx, byte SIZEOF_WORD + jb short .column_st1 + mov WORD [rdi], ax + add rdi, byte SIZEOF_WORD + sub rcx, byte SIZEOF_WORD + shr rax, 16 +.column_st1: + ; Store the lower 1 byte of rax to the output when it has enough + ; space. + test rcx, rcx + jz short .endcolumn + mov BYTE [rdi], al + +%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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC + movdqu XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH +.out0: + add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + cmp rcx, byte SIZEOF_XMMWORD/2 + jb short .column_st16 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD + add rdi, byte 2*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 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA + add rdi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub rcx, byte SIZEOF_XMMWORD/4 +.column_st15: + ; Store two pixels (8 bytes) of xmmA to the output when it has enough + ; space. + cmp rcx, byte SIZEOF_XMMWORD/8 + jb short .column_st7 + movq XMM_MMWORD [rdi], xmmA + add rdi, byte SIZEOF_XMMWORD/8*4 + sub rcx, byte SIZEOF_XMMWORD/8 + psrldq xmmA, SIZEOF_XMMWORD/8*4 +.column_st7: + ; Store one pixel (4 bytes) of xmmA to the output when it has enough + ; space. + test rcx, rcx + jz short .endcolumn + movd XMM_DWORD [rdi], 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..6494340 --- /dev/null +++ b/simd/jdmrgss2.asm @@ -0,0 +1,519 @@ +; +; jdmrgss2.asm - merged upsampling/color conversion (SSE2) +; +; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB +; Copyright 2012 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" + +; -------------------------------------------------------------------------- +; +; 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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF +.out0: + add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE + cmp ecx, byte 2*SIZEOF_XMMWORD + jb short .column_st16 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + add edi, byte 2*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 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + add edi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub ecx, byte SIZEOF_XMMWORD +.column_st15: + ; Store the lower 8 bytes of xmmA to the output when it has enough + ; space. + cmp ecx, byte SIZEOF_MMWORD + jb short .column_st7 + movq XMM_MMWORD [edi], xmmA + add edi, byte SIZEOF_MMWORD + sub ecx, byte SIZEOF_MMWORD + psrldq xmmA, SIZEOF_MMWORD +.column_st7: + ; Store the lower 4 bytes of xmmA to the output when it has enough + ; space. + cmp ecx, byte SIZEOF_DWORD + jb short .column_st3 + movd XMM_DWORD [edi], xmmA + add edi, byte SIZEOF_DWORD + sub ecx, byte SIZEOF_DWORD + psrldq xmmA, SIZEOF_DWORD +.column_st3: + ; Store the lower 2 bytes of eax to the output when it has enough + ; space. + movd eax, xmmA + cmp ecx, byte SIZEOF_WORD + jb short .column_st1 + mov WORD [edi], ax + add edi, byte SIZEOF_WORD + sub ecx, byte SIZEOF_WORD + shr eax, 16 +.column_st1: + ; Store the lower 1 byte of eax to the output when it has enough + ; space. + test ecx, ecx + jz short .endcolumn + mov BYTE [edi], al + +%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 + jmp short .out0 +.out1: ; --(unaligned)----------------- + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + movdqu XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC + movdqu XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH +.out0: + add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr + 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: + cmp ecx, byte SIZEOF_XMMWORD/2 + jb short .column_st16 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + movdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD + add edi, byte 2*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 + movdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA + add edi, byte SIZEOF_XMMWORD ; outptr + movdqa xmmA,xmmD + sub ecx, byte SIZEOF_XMMWORD/4 +.column_st15: + ; Store two pixels (8 bytes) of xmmA to the output when it has enough + ; space. + cmp ecx, byte SIZEOF_XMMWORD/8 + jb short .column_st7 + movq XMM_MMWORD [edi], xmmA + add edi, byte SIZEOF_XMMWORD/8*4 + sub ecx, byte SIZEOF_XMMWORD/8 + psrldq xmmA, SIZEOF_XMMWORD/8*4 +.column_st7: + ; Store one pixel (4 bytes) of xmmA to the output when it has enough + ; space. + test ecx, ecx + jz short .endcolumn + movd XMM_DWORD [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..3d4751f --- /dev/null +++ b/simd/jsimd.h @@ -0,0 +1,670 @@ +/* + * simd/jsimd.h + * + * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB + * Copyright 2011 D. R. Commander + * + * Based on the 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 + * + */ + +/* 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 +#define JSIMD_ARM_NEON 0x10 + +/* 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_rgb_gray_convert_mmx jSRGBGRYM +#define jsimd_extrgb_gray_convert_mmx jSEXTRGBGRYM +#define jsimd_extrgbx_gray_convert_mmx jSEXTRGBXGRYM +#define jsimd_extbgr_gray_convert_mmx jSEXTBGRGRYM +#define jsimd_extbgrx_gray_convert_mmx jSEXTBGRXGRYM +#define jsimd_extxbgr_gray_convert_mmx jSEXTXBGRGRYM +#define jsimd_extxrgb_gray_convert_mmx jSEXTXRGBGRYM +#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_rgb_gray_convert_sse2 jSCRGBGRYS2 +#define jsimd_rgb_gray_convert_sse2 jSRGBGRYS2 +#define jsimd_extrgb_gray_convert_sse2 jSEXTRGBGRYS2 +#define jsimd_extrgbx_gray_convert_sse2 jSEXTRGBXGRYS2 +#define jsimd_extbgr_gray_convert_sse2 jSEXTBGRGRYS2 +#define jsimd_extbgrx_gray_convert_sse2 jSEXTBGRXGRYS2 +#define jsimd_extxbgr_gray_convert_sse2 jSEXTXBGRGRYS2 +#define jsimd_extxrgb_gray_convert_sse2 jSEXTXRGBGRYS2 +#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_rgb_gray_convert_mmx + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extrgb_gray_convert_mmx + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extrgbx_gray_convert_mmx + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extbgr_gray_convert_mmx + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extbgrx_gray_convert_mmx + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extxbgr_gray_convert_mmx + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extxrgb_gray_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_rgb_gray_convert_sse2[]; +EXTERN(void) jsimd_rgb_gray_convert_sse2 + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extrgb_gray_convert_sse2 + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extrgbx_gray_convert_sse2 + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extbgr_gray_convert_sse2 + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extbgrx_gray_convert_sse2 + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extxbgr_gray_convert_sse2 + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extxrgb_gray_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)); + +EXTERN(void) jsimd_rgb_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extrgb_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extrgbx_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extbgr_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extbgrx_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extxbgr_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); +EXTERN(void) jsimd_extxrgb_ycc_convert_neon + JPP((JDIMENSION img_width, + JSAMPARRAY input_buf, JSAMPIMAGE output_buf, + JDIMENSION output_row, int num_rows)); + +EXTERN(void) jsimd_ycc_rgb_convert_neon + JPP((JDIMENSION out_width, + JSAMPIMAGE input_buf, JDIMENSION input_row, + JSAMPARRAY output_buf, int num_rows)); +EXTERN(void) jsimd_ycc_extrgb_convert_neon + JPP((JDIMENSION out_width, + JSAMPIMAGE input_buf, JDIMENSION input_row, + JSAMPARRAY output_buf, int num_rows)); +EXTERN(void) jsimd_ycc_extrgbx_convert_neon + JPP((JDIMENSION out_width, + JSAMPIMAGE input_buf, JDIMENSION input_row, + JSAMPARRAY output_buf, int num_rows)); +EXTERN(void) jsimd_ycc_extbgr_convert_neon + JPP((JDIMENSION out_width, + JSAMPIMAGE input_buf, JDIMENSION input_row, + JSAMPARRAY output_buf, int num_rows)); +EXTERN(void) jsimd_ycc_extbgrx_convert_neon + JPP((JDIMENSION out_width, + JSAMPIMAGE input_buf, JDIMENSION input_row, + JSAMPARRAY output_buf, int num_rows)); +EXTERN(void) jsimd_ycc_extxbgr_convert_neon + JPP((JDIMENSION out_width, + JSAMPIMAGE input_buf, JDIMENSION input_row, + JSAMPARRAY output_buf, int num_rows)); +EXTERN(void) jsimd_ycc_extxrgb_convert_neon + 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)); + +EXTERN(void) jsimd_h2v1_fancy_upsample_neon + JPP((int max_v_samp_factor, JDIMENSION downsampled_width, + JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)); + +/* 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_neon 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_ifast_neon 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_neon 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)); + +EXTERN(void) jsimd_idct_2x2_neon JPP((void * dct_table, + JCOEFPTR coef_block, + JSAMPARRAY output_buf, + JDIMENSION output_col)); +EXTERN(void) jsimd_idct_4x4_neon 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_islow_neon JPP((void * dct_table, + JCOEFPTR coef_block, + JSAMPARRAY output_buf, + JDIMENSION output_col)); +EXTERN(void) jsimd_idct_ifast_neon 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_arm.c b/simd/jsimd_arm.c new file mode 100644 index 0000000..cae84df --- /dev/null +++ b/simd/jsimd_arm.c @@ -0,0 +1,682 @@ +/* + * jsimd_arm.c + * + * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB + * Copyright 2009-2011 D. R. Commander + * + * Based on the 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 contains the interface between the "normal" portions + * of the library and the SIMD implementations when running on + * ARM architecture. + * + * Based on the stubs from 'jsimd_none.c' + */ + +#define JPEG_INTERNALS +#include "../jinclude.h" +#include "../jpeglib.h" +#include "../jsimd.h" +#include "../jdct.h" +#include "../jsimddct.h" +#include "jsimd.h" + +#include <stdio.h> +#include <string.h> +#include <ctype.h> + +static unsigned int simd_support = ~0; + +#if defined(__linux__) || defined(ANDROID) || defined(__ANDROID__) + +#define SOMEWHAT_SANE_PROC_CPUINFO_SIZE_LIMIT (1024 * 1024) + +LOCAL(int) +check_feature (char *buffer, char *feature) +{ + char *p; + if (*feature == 0) + return 0; + if (strncmp(buffer, "Features", 8) != 0) + return 0; + buffer += 8; + while (isspace(*buffer)) + buffer++; + + /* Check if 'feature' is present in the buffer as a separate word */ + while ((p = strstr(buffer, feature))) { + if (p > buffer && !isspace(*(p - 1))) { + buffer++; + continue; + } + p += strlen(feature); + if (*p != 0 && !isspace(*p)) { + buffer++; + continue; + } + return 1; + } + return 0; +} + +LOCAL(int) +parse_proc_cpuinfo (int bufsize) +{ + char *buffer = (char *)malloc(bufsize); + FILE *fd; + simd_support = 0; + + if (!buffer) + return 0; + + fd = fopen("/proc/cpuinfo", "r"); + if (fd) { + while (fgets(buffer, bufsize, fd)) { + if (!strchr(buffer, '\n') && !feof(fd)) { + /* "impossible" happened - insufficient size of the buffer! */ + fclose(fd); + free(buffer); + return 0; + } + if (check_feature(buffer, "neon")) + simd_support |= JSIMD_ARM_NEON; + } + fclose(fd); + } + free(buffer); + return 1; +} + +#endif + +/* + * 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 !defined(__ARM_NEON__) && defined(__linux__) || defined(ANDROID) || defined(__ANDROID__) + int bufsize = 1024; /* an initial guess for the line buffer size limit */ +#endif + + if (simd_support != ~0U) + return; + + simd_support = 0; + +#if defined(__ARM_NEON__) + simd_support |= JSIMD_ARM_NEON; +#elif defined(__linux__) || defined(ANDROID) || defined(__ANDROID__) + /* We still have a chance to use NEON regardless of globally used + * -mcpu/-mfpu options passed to gcc by performing runtime detection via + * /proc/cpuinfo parsing on linux/android */ + while (!parse_proc_cpuinfo(bufsize)) { + bufsize *= 2; + if (bufsize > SOMEWHAT_SANE_PROC_CPUINFO_SIZE_LIMIT) + break; + } +#endif + + /* Force different settings through environment variables */ + env = getenv("JSIMD_FORCE_ARM_NEON"); + if ((env != NULL) && (strcmp(env, "1") == 0)) + simd_support &= JSIMD_ARM_NEON; + env = getenv("JSIMD_FORCE_NO_SIMD"); + if ((env != NULL) && (strcmp(env, "1") == 0)) + simd_support = 0; +} + +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_ARM_NEON) + return 1; + + return 0; +} + +GLOBAL(int) +jsimd_can_rgb_gray (void) +{ + init_simd(); + + 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_ARM_NEON) + 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 (*neonfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); + + switch(cinfo->in_color_space) + { + case JCS_EXT_RGB: + neonfct=jsimd_extrgb_ycc_convert_neon; + break; + case JCS_EXT_RGBX: + case JCS_EXT_RGBA: + neonfct=jsimd_extrgbx_ycc_convert_neon; + break; + case JCS_EXT_BGR: + neonfct=jsimd_extbgr_ycc_convert_neon; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + neonfct=jsimd_extbgrx_ycc_convert_neon; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + neonfct=jsimd_extxbgr_ycc_convert_neon; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + neonfct=jsimd_extxrgb_ycc_convert_neon; + break; + default: + neonfct=jsimd_extrgb_ycc_convert_neon; + break; + } + + if (simd_support & JSIMD_ARM_NEON) + neonfct(cinfo->image_width, input_buf, + output_buf, output_row, num_rows); +} + +GLOBAL(void) +jsimd_rgb_gray_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) +{ + void (*neonfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); + + switch(cinfo->out_color_space) + { + case JCS_EXT_RGB: + neonfct=jsimd_ycc_extrgb_convert_neon; + break; + case JCS_EXT_RGBX: + case JCS_EXT_RGBA: + neonfct=jsimd_ycc_extrgbx_convert_neon; + break; + case JCS_EXT_BGR: + neonfct=jsimd_ycc_extbgr_convert_neon; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + neonfct=jsimd_ycc_extbgrx_convert_neon; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + neonfct=jsimd_ycc_extxbgr_convert_neon; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + neonfct=jsimd_ycc_extxrgb_convert_neon; + break; + default: + neonfct=jsimd_ycc_extrgb_convert_neon; + break; + } + + if (simd_support & JSIMD_ARM_NEON) + neonfct(cinfo->output_width, input_buf, + input_row, output_buf, num_rows); +} + +GLOBAL(int) +jsimd_can_h2v2_downsample (void) +{ + init_simd(); + + return 0; +} + +GLOBAL(int) +jsimd_can_h2v1_downsample (void) +{ + init_simd(); + + 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) +{ + init_simd(); + + return 0; +} + +GLOBAL(int) +jsimd_can_h2v1_upsample (void) +{ + init_simd(); + + 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) +{ + init_simd(); + + 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_ARM_NEON) + 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) +{ +} + +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_ARM_NEON) + jsimd_h2v1_fancy_upsample_neon(cinfo->max_v_samp_factor, + compptr->downsampled_width, input_data, output_data_ptr); +} + +GLOBAL(int) +jsimd_can_h2v2_merged_upsample (void) +{ + init_simd(); + + return 0; +} + +GLOBAL(int) +jsimd_can_h2v1_merged_upsample (void) +{ + init_simd(); + + 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) +{ + 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_ARM_NEON) + return 1; + + return 0; +} + +GLOBAL(int) +jsimd_can_convsamp_float (void) +{ + init_simd(); + + return 0; +} + +GLOBAL(void) +jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, + DCTELEM * workspace) +{ + if (simd_support & JSIMD_ARM_NEON) + jsimd_convsamp_neon(sample_data, start_col, workspace); +} + +GLOBAL(void) +jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, + FAST_FLOAT * workspace) +{ +} + +GLOBAL(int) +jsimd_can_fdct_islow (void) +{ + init_simd(); + + 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_ARM_NEON) + return 1; + + return 0; +} + +GLOBAL(int) +jsimd_can_fdct_float (void) +{ + init_simd(); + + return 0; +} + +GLOBAL(void) +jsimd_fdct_islow (DCTELEM * data) +{ +} + +GLOBAL(void) +jsimd_fdct_ifast (DCTELEM * data) +{ + if (simd_support & JSIMD_ARM_NEON) + jsimd_fdct_ifast_neon(data); +} + +GLOBAL(void) +jsimd_fdct_float (FAST_FLOAT * 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_ARM_NEON) + return 1; + + return 0; +} + +GLOBAL(int) +jsimd_can_quantize_float (void) +{ + init_simd(); + + return 0; +} + +GLOBAL(void) +jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors, + DCTELEM * workspace) +{ + if (simd_support & JSIMD_ARM_NEON) + jsimd_quantize_neon(coef_block, divisors, workspace); +} + +GLOBAL(void) +jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors, + FAST_FLOAT * 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_ARM_NEON)) + 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_ARM_NEON)) + 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_ARM_NEON)) + jsimd_idct_2x2_neon(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_ARM_NEON)) + jsimd_idct_4x4_neon(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_ARM_NEON) + 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_ARM_NEON)) + return 1; + + return 0; +} + +GLOBAL(int) +jsimd_can_idct_float (void) +{ + init_simd(); + + 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_ARM_NEON)) + jsimd_idct_islow_neon(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_ARM_NEON)) + jsimd_idct_ifast_neon(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) +{ +} + diff --git a/simd/jsimd_arm_neon.S b/simd/jsimd_arm_neon.S new file mode 100644 index 0000000..9962b8a --- /dev/null +++ b/simd/jsimd_arm_neon.S @@ -0,0 +1,2397 @@ +/* + * ARM NEON optimizations for libjpeg-turbo + * + * Copyright (C) 2009-2011 Nokia Corporation and/or its subsidiary(-ies). + * All rights reserved. + * Author: Siarhei Siamashka <siarhei.siamashka@nokia.com> + * + * 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. + */ + +#if defined(__linux__) && defined(__ELF__) +.section .note.GNU-stack,"",%progbits /* mark stack as non-executable */ +#endif + +.text +.fpu neon +.arch armv7a +.object_arch armv4 +.arm + + +#define RESPECT_STRICT_ALIGNMENT 1 + +/*****************************************************************************/ + +/* Supplementary macro for setting function attributes */ +.macro asm_function fname +#ifdef __APPLE__ + .func _\fname + .globl _\fname +_\fname: +#else + .func \fname + .global \fname +#ifdef __ELF__ + .hidden \fname + .type \fname, %function +#endif +\fname: +#endif +.endm + +/* Transpose a block of 4x4 coefficients in four 64-bit registers */ +.macro transpose_4x4 x0, x1, x2, x3 + vtrn.16 \x0, \x1 + vtrn.16 \x2, \x3 + vtrn.32 \x0, \x2 + vtrn.32 \x1, \x3 +.endm + +#define CENTERJSAMPLE 128 + +/*****************************************************************************/ + +/* + * Perform dequantization and inverse DCT on one block of coefficients. + * + * GLOBAL(void) + * jsimd_idct_islow_neon (void * dct_table, JCOEFPTR coef_block, + * JSAMPARRAY output_buf, JDIMENSION output_col) + */ + +#define FIX_0_298631336 (2446) +#define FIX_0_390180644 (3196) +#define FIX_0_541196100 (4433) +#define FIX_0_765366865 (6270) +#define FIX_0_899976223 (7373) +#define FIX_1_175875602 (9633) +#define FIX_1_501321110 (12299) +#define FIX_1_847759065 (15137) +#define FIX_1_961570560 (16069) +#define FIX_2_053119869 (16819) +#define FIX_2_562915447 (20995) +#define FIX_3_072711026 (25172) + +#define FIX_1_175875602_MINUS_1_961570560 (FIX_1_175875602 - FIX_1_961570560) +#define FIX_1_175875602_MINUS_0_390180644 (FIX_1_175875602 - FIX_0_390180644) +#define FIX_0_541196100_MINUS_1_847759065 (FIX_0_541196100 - FIX_1_847759065) +#define FIX_3_072711026_MINUS_2_562915447 (FIX_3_072711026 - FIX_2_562915447) +#define FIX_0_298631336_MINUS_0_899976223 (FIX_0_298631336 - FIX_0_899976223) +#define FIX_1_501321110_MINUS_0_899976223 (FIX_1_501321110 - FIX_0_899976223) +#define FIX_2_053119869_MINUS_2_562915447 (FIX_2_053119869 - FIX_2_562915447) +#define FIX_0_541196100_PLUS_0_765366865 (FIX_0_541196100 + FIX_0_765366865) + +/* + * Reference SIMD-friendly 1-D ISLOW iDCT C implementation. + * Uses some ideas from the comments in 'simd/jiss2int-64.asm' + */ +#define REF_1D_IDCT(xrow0, xrow1, xrow2, xrow3, xrow4, xrow5, xrow6, xrow7) \ +{ \ + DCTELEM row0, row1, row2, row3, row4, row5, row6, row7; \ + INT32 q1, q2, q3, q4, q5, q6, q7; \ + INT32 tmp11_plus_tmp2, tmp11_minus_tmp2; \ + \ + /* 1-D iDCT input data */ \ + row0 = xrow0; \ + row1 = xrow1; \ + row2 = xrow2; \ + row3 = xrow3; \ + row4 = xrow4; \ + row5 = xrow5; \ + row6 = xrow6; \ + row7 = xrow7; \ + \ + q5 = row7 + row3; \ + q4 = row5 + row1; \ + q6 = MULTIPLY(q5, FIX_1_175875602_MINUS_1_961570560) + \ + MULTIPLY(q4, FIX_1_175875602); \ + q7 = MULTIPLY(q5, FIX_1_175875602) + \ + MULTIPLY(q4, FIX_1_175875602_MINUS_0_390180644); \ + q2 = MULTIPLY(row2, FIX_0_541196100) + \ + MULTIPLY(row6, FIX_0_541196100_MINUS_1_847759065); \ + q4 = q6; \ + q3 = ((INT32) row0 - (INT32) row4) << 13; \ + q6 += MULTIPLY(row5, -FIX_2_562915447) + \ + MULTIPLY(row3, FIX_3_072711026_MINUS_2_562915447); \ + /* now we can use q1 (reloadable constants have been used up) */ \ + q1 = q3 + q2; \ + q4 += MULTIPLY(row7, FIX_0_298631336_MINUS_0_899976223) + \ + MULTIPLY(row1, -FIX_0_899976223); \ + q5 = q7; \ + q1 = q1 + q6; \ + q7 += MULTIPLY(row7, -FIX_0_899976223) + \ + MULTIPLY(row1, FIX_1_501321110_MINUS_0_899976223); \ + \ + /* (tmp11 + tmp2) has been calculated (out_row1 before descale) */ \ + tmp11_plus_tmp2 = q1; \ + row1 = 0; \ + \ + q1 = q1 - q6; \ + q5 += MULTIPLY(row5, FIX_2_053119869_MINUS_2_562915447) + \ + MULTIPLY(row3, -FIX_2_562915447); \ + q1 = q1 - q6; \ + q6 = MULTIPLY(row2, FIX_0_541196100_PLUS_0_765366865) + \ + MULTIPLY(row6, FIX_0_541196100); \ + q3 = q3 - q2; \ + \ + /* (tmp11 - tmp2) has been calculated (out_row6 before descale) */ \ + tmp11_minus_tmp2 = q1; \ + \ + q1 = ((INT32) row0 + (INT32) row4) << 13; \ + q2 = q1 + q6; \ + q1 = q1 - q6; \ + \ + /* pick up the results */ \ + tmp0 = q4; \ + tmp1 = q5; \ + tmp2 = (tmp11_plus_tmp2 - tmp11_minus_tmp2) / 2; \ + tmp3 = q7; \ + tmp10 = q2; \ + tmp11 = (tmp11_plus_tmp2 + tmp11_minus_tmp2) / 2; \ + tmp12 = q3; \ + tmp13 = q1; \ +} + +#define XFIX_0_899976223 d0[0] +#define XFIX_0_541196100 d0[1] +#define XFIX_2_562915447 d0[2] +#define XFIX_0_298631336_MINUS_0_899976223 d0[3] +#define XFIX_1_501321110_MINUS_0_899976223 d1[0] +#define XFIX_2_053119869_MINUS_2_562915447 d1[1] +#define XFIX_0_541196100_PLUS_0_765366865 d1[2] +#define XFIX_1_175875602 d1[3] +#define XFIX_1_175875602_MINUS_0_390180644 d2[0] +#define XFIX_0_541196100_MINUS_1_847759065 d2[1] +#define XFIX_3_072711026_MINUS_2_562915447 d2[2] +#define XFIX_1_175875602_MINUS_1_961570560 d2[3] + +.balign 16 +jsimd_idct_islow_neon_consts: + .short FIX_0_899976223 /* d0[0] */ + .short FIX_0_541196100 /* d0[1] */ + .short FIX_2_562915447 /* d0[2] */ + .short FIX_0_298631336_MINUS_0_899976223 /* d0[3] */ + .short FIX_1_501321110_MINUS_0_899976223 /* d1[0] */ + .short FIX_2_053119869_MINUS_2_562915447 /* d1[1] */ + .short FIX_0_541196100_PLUS_0_765366865 /* d1[2] */ + .short FIX_1_175875602 /* d1[3] */ + /* reloadable constants */ + .short FIX_1_175875602_MINUS_0_390180644 /* d2[0] */ + .short FIX_0_541196100_MINUS_1_847759065 /* d2[1] */ + .short FIX_3_072711026_MINUS_2_562915447 /* d2[2] */ + .short FIX_1_175875602_MINUS_1_961570560 /* d2[3] */ + +asm_function jsimd_idct_islow_neon + + DCT_TABLE .req r0 + COEF_BLOCK .req r1 + OUTPUT_BUF .req r2 + OUTPUT_COL .req r3 + TMP1 .req r0 + TMP2 .req r1 + TMP3 .req r2 + TMP4 .req ip + + ROW0L .req d16 + ROW0R .req d17 + ROW1L .req d18 + ROW1R .req d19 + ROW2L .req d20 + ROW2R .req d21 + ROW3L .req d22 + ROW3R .req d23 + ROW4L .req d24 + ROW4R .req d25 + ROW5L .req d26 + ROW5R .req d27 + ROW6L .req d28 + ROW6R .req d29 + ROW7L .req d30 + ROW7R .req d31 + + /* Load and dequantize coefficients into NEON registers + * with the following allocation: + * 0 1 2 3 | 4 5 6 7 + * ---------+-------- + * 0 | d16 | d17 ( q8 ) + * 1 | d18 | d19 ( q9 ) + * 2 | d20 | d21 ( q10 ) + * 3 | d22 | d23 ( q11 ) + * 4 | d24 | d25 ( q12 ) + * 5 | d26 | d27 ( q13 ) + * 6 | d28 | d29 ( q14 ) + * 7 | d30 | d31 ( q15 ) + */ + adr ip, jsimd_idct_islow_neon_consts + vld1.16 {d16, d17, d18, d19}, [COEF_BLOCK, :128]! + vld1.16 {d0, d1, d2, d3}, [DCT_TABLE, :128]! + vld1.16 {d20, d21, d22, d23}, [COEF_BLOCK, :128]! + vmul.s16 q8, q8, q0 + vld1.16 {d4, d5, d6, d7}, [DCT_TABLE, :128]! + vmul.s16 q9, q9, q1 + vld1.16 {d24, d25, d26, d27}, [COEF_BLOCK, :128]! + vmul.s16 q10, q10, q2 + vld1.16 {d0, d1, d2, d3}, [DCT_TABLE, :128]! + vmul.s16 q11, q11, q3 + vld1.16 {d28, d29, d30, d31}, [COEF_BLOCK, :128] + vmul.s16 q12, q12, q0 + vld1.16 {d4, d5, d6, d7}, [DCT_TABLE, :128]! + vmul.s16 q14, q14, q2 + vmul.s16 q13, q13, q1 + vld1.16 {d0, d1, d2, d3}, [ip, :128] /* load constants */ + add ip, ip, #16 + vmul.s16 q15, q15, q3 + vpush {d8-d15} /* save NEON registers */ + /* 1-D IDCT, pass 1, left 4x8 half */ + vadd.s16 d4, ROW7L, ROW3L + vadd.s16 d5, ROW5L, ROW1L + vmull.s16 q6, d4, XFIX_1_175875602_MINUS_1_961570560 + vmlal.s16 q6, d5, XFIX_1_175875602 + vmull.s16 q7, d4, XFIX_1_175875602 + /* Check for the zero coefficients in the right 4x8 half */ + push {r4, r5} + vmlal.s16 q7, d5, XFIX_1_175875602_MINUS_0_390180644 + vsubl.s16 q3, ROW0L, ROW4L + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 1 * 8))] + vmull.s16 q2, ROW2L, XFIX_0_541196100 + vmlal.s16 q2, ROW6L, XFIX_0_541196100_MINUS_1_847759065 + orr r0, r4, r5 + vmov q4, q6 + vmlsl.s16 q6, ROW5L, XFIX_2_562915447 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 2 * 8))] + vmlal.s16 q6, ROW3L, XFIX_3_072711026_MINUS_2_562915447 + vshl.s32 q3, q3, #13 + orr r0, r0, r4 + vmlsl.s16 q4, ROW1L, XFIX_0_899976223 + orr r0, r0, r5 + vadd.s32 q1, q3, q2 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 3 * 8))] + vmov q5, q7 + vadd.s32 q1, q1, q6 + orr r0, r0, r4 + vmlsl.s16 q7, ROW7L, XFIX_0_899976223 + orr r0, r0, r5 + vmlal.s16 q7, ROW1L, XFIX_1_501321110_MINUS_0_899976223 + vrshrn.s32 ROW1L, q1, #11 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 4 * 8))] + vsub.s32 q1, q1, q6 + vmlal.s16 q5, ROW5L, XFIX_2_053119869_MINUS_2_562915447 + orr r0, r0, r4 + vmlsl.s16 q5, ROW3L, XFIX_2_562915447 + orr r0, r0, r5 + vsub.s32 q1, q1, q6 + vmull.s16 q6, ROW2L, XFIX_0_541196100_PLUS_0_765366865 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 5 * 8))] + vmlal.s16 q6, ROW6L, XFIX_0_541196100 + vsub.s32 q3, q3, q2 + orr r0, r0, r4 + vrshrn.s32 ROW6L, q1, #11 + orr r0, r0, r5 + vadd.s32 q1, q3, q5 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 6 * 8))] + vsub.s32 q3, q3, q5 + vaddl.s16 q5, ROW0L, ROW4L + orr r0, r0, r4 + vrshrn.s32 ROW2L, q1, #11 + orr r0, r0, r5 + vrshrn.s32 ROW5L, q3, #11 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 7 * 8))] + vshl.s32 q5, q5, #13 + vmlal.s16 q4, ROW7L, XFIX_0_298631336_MINUS_0_899976223 + orr r0, r0, r4 + vadd.s32 q2, q5, q6 + orrs r0, r0, r5 + vsub.s32 q1, q5, q6 + vadd.s32 q6, q2, q7 + ldrd r4, [COEF_BLOCK, #(-96 + 2 * (4 + 0 * 8))] + vsub.s32 q2, q2, q7 + vadd.s32 q5, q1, q4 + orr r0, r4, r5 + vsub.s32 q3, q1, q4 + pop {r4, r5} + vrshrn.s32 ROW7L, q2, #11 + vrshrn.s32 ROW3L, q5, #11 + vrshrn.s32 ROW0L, q6, #11 + vrshrn.s32 ROW4L, q3, #11 + + beq 3f /* Go to do some special handling for the sparse right 4x8 half */ + + /* 1-D IDCT, pass 1, right 4x8 half */ + vld1.s16 {d2}, [ip, :64] /* reload constants */ + vadd.s16 d10, ROW7R, ROW3R + vadd.s16 d8, ROW5R, ROW1R + /* Transpose left 4x8 half */ + vtrn.16 ROW6L, ROW7L + vmull.s16 q6, d10, XFIX_1_175875602_MINUS_1_961570560 + vmlal.s16 q6, d8, XFIX_1_175875602 + vtrn.16 ROW2L, ROW3L + vmull.s16 q7, d10, XFIX_1_175875602 + vmlal.s16 q7, d8, XFIX_1_175875602_MINUS_0_390180644 + vtrn.16 ROW0L, ROW1L + vsubl.s16 q3, ROW0R, ROW4R + vmull.s16 q2, ROW2R, XFIX_0_541196100 + vmlal.s16 q2, ROW6R, XFIX_0_541196100_MINUS_1_847759065 + vtrn.16 ROW4L, ROW5L + vmov q4, q6 + vmlsl.s16 q6, ROW5R, XFIX_2_562915447 + vmlal.s16 q6, ROW3R, XFIX_3_072711026_MINUS_2_562915447 + vtrn.32 ROW1L, ROW3L + vshl.s32 q3, q3, #13 + vmlsl.s16 q4, ROW1R, XFIX_0_899976223 + vtrn.32 ROW4L, ROW6L + vadd.s32 q1, q3, q2 + vmov q5, q7 + vadd.s32 q1, q1, q6 + vtrn.32 ROW0L, ROW2L + vmlsl.s16 q7, ROW7R, XFIX_0_899976223 + vmlal.s16 q7, ROW1R, XFIX_1_501321110_MINUS_0_899976223 + vrshrn.s32 ROW1R, q1, #11 + vtrn.32 ROW5L, ROW7L + vsub.s32 q1, q1, q6 + vmlal.s16 q5, ROW5R, XFIX_2_053119869_MINUS_2_562915447 + vmlsl.s16 q5, ROW3R, XFIX_2_562915447 + vsub.s32 q1, q1, q6 + vmull.s16 q6, ROW2R, XFIX_0_541196100_PLUS_0_765366865 + vmlal.s16 q6, ROW6R, XFIX_0_541196100 + vsub.s32 q3, q3, q2 + vrshrn.s32 ROW6R, q1, #11 + vadd.s32 q1, q3, q5 + vsub.s32 q3, q3, q5 + vaddl.s16 q5, ROW0R, ROW4R + vrshrn.s32 ROW2R, q1, #11 + vrshrn.s32 ROW5R, q3, #11 + vshl.s32 q5, q5, #13 + vmlal.s16 q4, ROW7R, XFIX_0_298631336_MINUS_0_899976223 + vadd.s32 q2, q5, q6 + vsub.s32 q1, q5, q6 + vadd.s32 q6, q2, q7 + vsub.s32 q2, q2, q7 + vadd.s32 q5, q1, q4 + vsub.s32 q3, q1, q4 + vrshrn.s32 ROW7R, q2, #11 + vrshrn.s32 ROW3R, q5, #11 + vrshrn.s32 ROW0R, q6, #11 + vrshrn.s32 ROW4R, q3, #11 + /* Transpose right 4x8 half */ + vtrn.16 ROW6R, ROW7R + vtrn.16 ROW2R, ROW3R + vtrn.16 ROW0R, ROW1R + vtrn.16 ROW4R, ROW5R + vtrn.32 ROW1R, ROW3R + vtrn.32 ROW4R, ROW6R + vtrn.32 ROW0R, ROW2R + vtrn.32 ROW5R, ROW7R + +1: /* 1-D IDCT, pass 2 (normal variant), left 4x8 half */ + vld1.s16 {d2}, [ip, :64] /* reload constants */ + vmull.s16 q6, ROW1R, XFIX_1_175875602 /* ROW5L <-> ROW1R */ + vmlal.s16 q6, ROW1L, XFIX_1_175875602 + vmlal.s16 q6, ROW3R, XFIX_1_175875602_MINUS_1_961570560 /* ROW7L <-> ROW3R */ + vmlal.s16 q6, ROW3L, XFIX_1_175875602_MINUS_1_961570560 + vmull.s16 q7, ROW3R, XFIX_1_175875602 /* ROW7L <-> ROW3R */ + vmlal.s16 q7, ROW3L, XFIX_1_175875602 + vmlal.s16 q7, ROW1R, XFIX_1_175875602_MINUS_0_390180644 /* ROW5L <-> ROW1R */ + vmlal.s16 q7, ROW1L, XFIX_1_175875602_MINUS_0_390180644 + vsubl.s16 q3, ROW0L, ROW0R /* ROW4L <-> ROW0R */ + vmull.s16 q2, ROW2L, XFIX_0_541196100 + vmlal.s16 q2, ROW2R, XFIX_0_541196100_MINUS_1_847759065 /* ROW6L <-> ROW2R */ + vmov q4, q6 + vmlsl.s16 q6, ROW1R, XFIX_2_562915447 /* ROW5L <-> ROW1R */ + vmlal.s16 q6, ROW3L, XFIX_3_072711026_MINUS_2_562915447 + vshl.s32 q3, q3, #13 + vmlsl.s16 q4, ROW1L, XFIX_0_899976223 + vadd.s32 q1, q3, q2 + vmov q5, q7 + vadd.s32 q1, q1, q6 + vmlsl.s16 q7, ROW3R, XFIX_0_899976223 /* ROW7L <-> ROW3R */ + vmlal.s16 q7, ROW1L, XFIX_1_501321110_MINUS_0_899976223 + vshrn.s32 ROW1L, q1, #16 + vsub.s32 q1, q1, q6 + vmlal.s16 q5, ROW1R, XFIX_2_053119869_MINUS_2_562915447 /* ROW5L <-> ROW1R */ + vmlsl.s16 q5, ROW3L, XFIX_2_562915447 + vsub.s32 q1, q1, q6 + vmull.s16 q6, ROW2L, XFIX_0_541196100_PLUS_0_765366865 + vmlal.s16 q6, ROW2R, XFIX_0_541196100 /* ROW6L <-> ROW2R */ + vsub.s32 q3, q3, q2 + vshrn.s32 ROW2R, q1, #16 /* ROW6L <-> ROW2R */ + vadd.s32 q1, q3, q5 + vsub.s32 q3, q3, q5 + vaddl.s16 q5, ROW0L, ROW0R /* ROW4L <-> ROW0R */ + vshrn.s32 ROW2L, q1, #16 + vshrn.s32 ROW1R, q3, #16 /* ROW5L <-> ROW1R */ + vshl.s32 q5, q5, #13 + vmlal.s16 q4, ROW3R, XFIX_0_298631336_MINUS_0_899976223 /* ROW7L <-> ROW3R */ + vadd.s32 q2, q5, q6 + vsub.s32 q1, q5, q6 + vadd.s32 q6, q2, q7 + vsub.s32 q2, q2, q7 + vadd.s32 q5, q1, q4 + vsub.s32 q3, q1, q4 + vshrn.s32 ROW3R, q2, #16 /* ROW7L <-> ROW3R */ + vshrn.s32 ROW3L, q5, #16 + vshrn.s32 ROW0L, q6, #16 + vshrn.s32 ROW0R, q3, #16 /* ROW4L <-> ROW0R */ + /* 1-D IDCT, pass 2, right 4x8 half */ + vld1.s16 {d2}, [ip, :64] /* reload constants */ + vmull.s16 q6, ROW5R, XFIX_1_175875602 + vmlal.s16 q6, ROW5L, XFIX_1_175875602 /* ROW5L <-> ROW1R */ + vmlal.s16 q6, ROW7R, XFIX_1_175875602_MINUS_1_961570560 + vmlal.s16 q6, ROW7L, XFIX_1_175875602_MINUS_1_961570560 /* ROW7L <-> ROW3R */ + vmull.s16 q7, ROW7R, XFIX_1_175875602 + vmlal.s16 q7, ROW7L, XFIX_1_175875602 /* ROW7L <-> ROW3R */ + vmlal.s16 q7, ROW5R, XFIX_1_175875602_MINUS_0_390180644 + vmlal.s16 q7, ROW5L, XFIX_1_175875602_MINUS_0_390180644 /* ROW5L <-> ROW1R */ + vsubl.s16 q3, ROW4L, ROW4R /* ROW4L <-> ROW0R */ + vmull.s16 q2, ROW6L, XFIX_0_541196100 /* ROW6L <-> ROW2R */ + vmlal.s16 q2, ROW6R, XFIX_0_541196100_MINUS_1_847759065 + vmov q4, q6 + vmlsl.s16 q6, ROW5R, XFIX_2_562915447 + vmlal.s16 q6, ROW7L, XFIX_3_072711026_MINUS_2_562915447 /* ROW7L <-> ROW3R */ + vshl.s32 q3, q3, #13 + vmlsl.s16 q4, ROW5L, XFIX_0_899976223 /* ROW5L <-> ROW1R */ + vadd.s32 q1, q3, q2 + vmov q5, q7 + vadd.s32 q1, q1, q6 + vmlsl.s16 q7, ROW7R, XFIX_0_899976223 + vmlal.s16 q7, ROW5L, XFIX_1_501321110_MINUS_0_899976223 /* ROW5L <-> ROW1R */ + vshrn.s32 ROW5L, q1, #16 /* ROW5L <-> ROW1R */ + vsub.s32 q1, q1, q6 + vmlal.s16 q5, ROW5R, XFIX_2_053119869_MINUS_2_562915447 + vmlsl.s16 q5, ROW7L, XFIX_2_562915447 /* ROW7L <-> ROW3R */ + vsub.s32 q1, q1, q6 + vmull.s16 q6, ROW6L, XFIX_0_541196100_PLUS_0_765366865 /* ROW6L <-> ROW2R */ + vmlal.s16 q6, ROW6R, XFIX_0_541196100 + vsub.s32 q3, q3, q2 + vshrn.s32 ROW6R, q1, #16 + vadd.s32 q1, q3, q5 + vsub.s32 q3, q3, q5 + vaddl.s16 q5, ROW4L, ROW4R /* ROW4L <-> ROW0R */ + vshrn.s32 ROW6L, q1, #16 /* ROW6L <-> ROW2R */ + vshrn.s32 ROW5R, q3, #16 + vshl.s32 q5, q5, #13 + vmlal.s16 q4, ROW7R, XFIX_0_298631336_MINUS_0_899976223 + vadd.s32 q2, q5, q6 + vsub.s32 q1, q5, q6 + vadd.s32 q6, q2, q7 + vsub.s32 q2, q2, q7 + vadd.s32 q5, q1, q4 + vsub.s32 q3, q1, q4 + vshrn.s32 ROW7R, q2, #16 + vshrn.s32 ROW7L, q5, #16 /* ROW7L <-> ROW3R */ + vshrn.s32 ROW4L, q6, #16 /* ROW4L <-> ROW0R */ + vshrn.s32 ROW4R, q3, #16 + +2: /* Descale to 8-bit and range limit */ + vqrshrn.s16 d16, q8, #2 + vqrshrn.s16 d17, q9, #2 + vqrshrn.s16 d18, q10, #2 + vqrshrn.s16 d19, q11, #2 + vpop {d8-d15} /* restore NEON registers */ + vqrshrn.s16 d20, q12, #2 + /* Transpose the final 8-bit samples and do signed->unsigned conversion */ + vtrn.16 q8, q9 + vqrshrn.s16 d21, q13, #2 + vqrshrn.s16 d22, q14, #2 + vmov.u8 q0, #(CENTERJSAMPLE) + vqrshrn.s16 d23, q15, #2 + vtrn.8 d16, d17 + vtrn.8 d18, d19 + vadd.u8 q8, q8, q0 + vadd.u8 q9, q9, q0 + vtrn.16 q10, q11 + /* Store results to the output buffer */ + ldmia OUTPUT_BUF!, {TMP1, TMP2} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + vst1.8 {d16}, [TMP1] + vtrn.8 d20, d21 + vst1.8 {d17}, [TMP2] + ldmia OUTPUT_BUF!, {TMP1, TMP2} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + vst1.8 {d18}, [TMP1] + vadd.u8 q10, q10, q0 + vst1.8 {d19}, [TMP2] + ldmia OUTPUT_BUF, {TMP1, TMP2, TMP3, TMP4} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + add TMP3, TMP3, OUTPUT_COL + add TMP4, TMP4, OUTPUT_COL + vtrn.8 d22, d23 + vst1.8 {d20}, [TMP1] + vadd.u8 q11, q11, q0 + vst1.8 {d21}, [TMP2] + vst1.8 {d22}, [TMP3] + vst1.8 {d23}, [TMP4] + bx lr + +3: /* Left 4x8 half is done, right 4x8 half contains mostly zeros */ + + /* Transpose left 4x8 half */ + vtrn.16 ROW6L, ROW7L + vtrn.16 ROW2L, ROW3L + vtrn.16 ROW0L, ROW1L + vtrn.16 ROW4L, ROW5L + vshl.s16 ROW0R, ROW0R, #2 /* PASS1_BITS */ + vtrn.32 ROW1L, ROW3L + vtrn.32 ROW4L, ROW6L + vtrn.32 ROW0L, ROW2L + vtrn.32 ROW5L, ROW7L + + cmp r0, #0 + beq 4f /* Right 4x8 half has all zeros, go to 'sparse' second pass */ + + /* Only row 0 is non-zero for the right 4x8 half */ + vdup.s16 ROW1R, ROW0R[1] + vdup.s16 ROW2R, ROW0R[2] + vdup.s16 ROW3R, ROW0R[3] + vdup.s16 ROW4R, ROW0R[0] + vdup.s16 ROW5R, ROW0R[1] + vdup.s16 ROW6R, ROW0R[2] + vdup.s16 ROW7R, ROW0R[3] + vdup.s16 ROW0R, ROW0R[0] + b 1b /* Go to 'normal' second pass */ + +4: /* 1-D IDCT, pass 2 (sparse variant with zero rows 4-7), left 4x8 half */ + vld1.s16 {d2}, [ip, :64] /* reload constants */ + vmull.s16 q6, ROW1L, XFIX_1_175875602 + vmlal.s16 q6, ROW3L, XFIX_1_175875602_MINUS_1_961570560 + vmull.s16 q7, ROW3L, XFIX_1_175875602 + vmlal.s16 q7, ROW1L, XFIX_1_175875602_MINUS_0_390180644 + vmull.s16 q2, ROW2L, XFIX_0_541196100 + vshll.s16 q3, ROW0L, #13 + vmov q4, q6 + vmlal.s16 q6, ROW3L, XFIX_3_072711026_MINUS_2_562915447 + vmlsl.s16 q4, ROW1L, XFIX_0_899976223 + vadd.s32 q1, q3, q2 + vmov q5, q7 + vmlal.s16 q7, ROW1L, XFIX_1_501321110_MINUS_0_899976223 + vadd.s32 q1, q1, q6 + vadd.s32 q6, q6, q6 + vmlsl.s16 q5, ROW3L, XFIX_2_562915447 + vshrn.s32 ROW1L, q1, #16 + vsub.s32 q1, q1, q6 + vmull.s16 q6, ROW2L, XFIX_0_541196100_PLUS_0_765366865 + vsub.s32 q3, q3, q2 + vshrn.s32 ROW2R, q1, #16 /* ROW6L <-> ROW2R */ + vadd.s32 q1, q3, q5 + vsub.s32 q3, q3, q5 + vshll.s16 q5, ROW0L, #13 + vshrn.s32 ROW2L, q1, #16 + vshrn.s32 ROW1R, q3, #16 /* ROW5L <-> ROW1R */ + vadd.s32 q2, q5, q6 + vsub.s32 q1, q5, q6 + vadd.s32 q6, q2, q7 + vsub.s32 q2, q2, q7 + vadd.s32 q5, q1, q4 + vsub.s32 q3, q1, q4 + vshrn.s32 ROW3R, q2, #16 /* ROW7L <-> ROW3R */ + vshrn.s32 ROW3L, q5, #16 + vshrn.s32 ROW0L, q6, #16 + vshrn.s32 ROW0R, q3, #16 /* ROW4L <-> ROW0R */ + /* 1-D IDCT, pass 2 (sparse variant with zero rows 4-7), right 4x8 half */ + vld1.s16 {d2}, [ip, :64] /* reload constants */ + vmull.s16 q6, ROW5L, XFIX_1_175875602 + vmlal.s16 q6, ROW7L, XFIX_1_175875602_MINUS_1_961570560 + vmull.s16 q7, ROW7L, XFIX_1_175875602 + vmlal.s16 q7, ROW5L, XFIX_1_175875602_MINUS_0_390180644 + vmull.s16 q2, ROW6L, XFIX_0_541196100 + vshll.s16 q3, ROW4L, #13 + vmov q4, q6 + vmlal.s16 q6, ROW7L, XFIX_3_072711026_MINUS_2_562915447 + vmlsl.s16 q4, ROW5L, XFIX_0_899976223 + vadd.s32 q1, q3, q2 + vmov q5, q7 + vmlal.s16 q7, ROW5L, XFIX_1_501321110_MINUS_0_899976223 + vadd.s32 q1, q1, q6 + vadd.s32 q6, q6, q6 + vmlsl.s16 q5, ROW7L, XFIX_2_562915447 + vshrn.s32 ROW5L, q1, #16 /* ROW5L <-> ROW1R */ + vsub.s32 q1, q1, q6 + vmull.s16 q6, ROW6L, XFIX_0_541196100_PLUS_0_765366865 + vsub.s32 q3, q3, q2 + vshrn.s32 ROW6R, q1, #16 + vadd.s32 q1, q3, q5 + vsub.s32 q3, q3, q5 + vshll.s16 q5, ROW4L, #13 + vshrn.s32 ROW6L, q1, #16 /* ROW6L <-> ROW2R */ + vshrn.s32 ROW5R, q3, #16 + vadd.s32 q2, q5, q6 + vsub.s32 q1, q5, q6 + vadd.s32 q6, q2, q7 + vsub.s32 q2, q2, q7 + vadd.s32 q5, q1, q4 + vsub.s32 q3, q1, q4 + vshrn.s32 ROW7R, q2, #16 + vshrn.s32 ROW7L, q5, #16 /* ROW7L <-> ROW3R */ + vshrn.s32 ROW4L, q6, #16 /* ROW4L <-> ROW0R */ + vshrn.s32 ROW4R, q3, #16 + b 2b /* Go to epilogue */ + + .unreq DCT_TABLE + .unreq COEF_BLOCK + .unreq OUTPUT_BUF + .unreq OUTPUT_COL + .unreq TMP1 + .unreq TMP2 + .unreq TMP3 + .unreq TMP4 + + .unreq ROW0L + .unreq ROW0R + .unreq ROW1L + .unreq ROW1R + .unreq ROW2L + .unreq ROW2R + .unreq ROW3L + .unreq ROW3R + .unreq ROW4L + .unreq ROW4R + .unreq ROW5L + .unreq ROW5R + .unreq ROW6L + .unreq ROW6R + .unreq ROW7L + .unreq ROW7R +.endfunc + +/*****************************************************************************/ + +/* + * jsimd_idct_ifast_neon + * + * This function contains a fast, not so accurate integer implementation of + * the inverse DCT (Discrete Cosine Transform). It uses the same calculations + * and produces exactly the same output as IJG's original 'jpeg_idct_ifast' + * function from jidctfst.c + * + * Normally 1-D AAN DCT needs 5 multiplications and 29 additions. + * But in ARM NEON case some extra additions are required because VQDMULH + * instruction can't handle the constants larger than 1. So the expressions + * like "x * 1.082392200" have to be converted to "x * 0.082392200 + x", + * which introduces an extra addition. Overall, there are 6 extra additions + * per 1-D IDCT pass, totalling to 5 VQDMULH and 35 VADD/VSUB instructions. + */ + +#define XFIX_1_082392200 d0[0] +#define XFIX_1_414213562 d0[1] +#define XFIX_1_847759065 d0[2] +#define XFIX_2_613125930 d0[3] + +.balign 16 +jsimd_idct_ifast_neon_consts: + .short (277 * 128 - 256 * 128) /* XFIX_1_082392200 */ + .short (362 * 128 - 256 * 128) /* XFIX_1_414213562 */ + .short (473 * 128 - 256 * 128) /* XFIX_1_847759065 */ + .short (669 * 128 - 512 * 128) /* XFIX_2_613125930 */ + +asm_function jsimd_idct_ifast_neon + + DCT_TABLE .req r0 + COEF_BLOCK .req r1 + OUTPUT_BUF .req r2 + OUTPUT_COL .req r3 + TMP1 .req r0 + TMP2 .req r1 + TMP3 .req r2 + TMP4 .req ip + + /* Load and dequantize coefficients into NEON registers + * with the following allocation: + * 0 1 2 3 | 4 5 6 7 + * ---------+-------- + * 0 | d16 | d17 ( q8 ) + * 1 | d18 | d19 ( q9 ) + * 2 | d20 | d21 ( q10 ) + * 3 | d22 | d23 ( q11 ) + * 4 | d24 | d25 ( q12 ) + * 5 | d26 | d27 ( q13 ) + * 6 | d28 | d29 ( q14 ) + * 7 | d30 | d31 ( q15 ) + */ + adr ip, jsimd_idct_ifast_neon_consts + vld1.16 {d16, d17, d18, d19}, [COEF_BLOCK, :128]! + vld1.16 {d0, d1, d2, d3}, [DCT_TABLE, :128]! + vld1.16 {d20, d21, d22, d23}, [COEF_BLOCK, :128]! + vmul.s16 q8, q8, q0 + vld1.16 {d4, d5, d6, d7}, [DCT_TABLE, :128]! + vmul.s16 q9, q9, q1 + vld1.16 {d24, d25, d26, d27}, [COEF_BLOCK, :128]! + vmul.s16 q10, q10, q2 + vld1.16 {d0, d1, d2, d3}, [DCT_TABLE, :128]! + vmul.s16 q11, q11, q3 + vld1.16 {d28, d29, d30, d31}, [COEF_BLOCK, :128] + vmul.s16 q12, q12, q0 + vld1.16 {d4, d5, d6, d7}, [DCT_TABLE, :128]! + vmul.s16 q14, q14, q2 + vmul.s16 q13, q13, q1 + vld1.16 {d0}, [ip, :64] /* load constants */ + vmul.s16 q15, q15, q3 + vpush {d8-d13} /* save NEON registers */ + /* 1-D IDCT, pass 1 */ + vsub.s16 q2, q10, q14 + vadd.s16 q14, q10, q14 + vsub.s16 q1, q11, q13 + vadd.s16 q13, q11, q13 + vsub.s16 q5, q9, q15 + vadd.s16 q15, q9, q15 + vqdmulh.s16 q4, q2, XFIX_1_414213562 + vqdmulh.s16 q6, q1, XFIX_2_613125930 + vadd.s16 q3, q1, q1 + vsub.s16 q1, q5, q1 + vadd.s16 q10, q2, q4 + vqdmulh.s16 q4, q1, XFIX_1_847759065 + vsub.s16 q2, q15, q13 + vadd.s16 q3, q3, q6 + vqdmulh.s16 q6, q2, XFIX_1_414213562 + vadd.s16 q1, q1, q4 + vqdmulh.s16 q4, q5, XFIX_1_082392200 + vsub.s16 q10, q10, q14 + vadd.s16 q2, q2, q6 + vsub.s16 q6, q8, q12 + vadd.s16 q12, q8, q12 + vadd.s16 q9, q5, q4 + vadd.s16 q5, q6, q10 + vsub.s16 q10, q6, q10 + vadd.s16 q6, q15, q13 + vadd.s16 q8, q12, q14 + vsub.s16 q3, q6, q3 + vsub.s16 q12, q12, q14 + vsub.s16 q3, q3, q1 + vsub.s16 q1, q9, q1 + vadd.s16 q2, q3, q2 + vsub.s16 q15, q8, q6 + vadd.s16 q1, q1, q2 + vadd.s16 q8, q8, q6 + vadd.s16 q14, q5, q3 + vsub.s16 q9, q5, q3 + vsub.s16 q13, q10, q2 + vadd.s16 q10, q10, q2 + /* Transpose */ + vtrn.16 q8, q9 + vsub.s16 q11, q12, q1 + vtrn.16 q14, q15 + vadd.s16 q12, q12, q1 + vtrn.16 q10, q11 + vtrn.16 q12, q13 + vtrn.32 q9, q11 + vtrn.32 q12, q14 + vtrn.32 q8, q10 + vtrn.32 q13, q15 + vswp d28, d21 + vswp d26, d19 + /* 1-D IDCT, pass 2 */ + vsub.s16 q2, q10, q14 + vswp d30, d23 + vadd.s16 q14, q10, q14 + vswp d24, d17 + vsub.s16 q1, q11, q13 + vadd.s16 q13, q11, q13 + vsub.s16 q5, q9, q15 + vadd.s16 q15, q9, q15 + vqdmulh.s16 q4, q2, XFIX_1_414213562 + vqdmulh.s16 q6, q1, XFIX_2_613125930 + vadd.s16 q3, q1, q1 + vsub.s16 q1, q5, q1 + vadd.s16 q10, q2, q4 + vqdmulh.s16 q4, q1, XFIX_1_847759065 + vsub.s16 q2, q15, q13 + vadd.s16 q3, q3, q6 + vqdmulh.s16 q6, q2, XFIX_1_414213562 + vadd.s16 q1, q1, q4 + vqdmulh.s16 q4, q5, XFIX_1_082392200 + vsub.s16 q10, q10, q14 + vadd.s16 q2, q2, q6 + vsub.s16 q6, q8, q12 + vadd.s16 q12, q8, q12 + vadd.s16 q9, q5, q4 + vadd.s16 q5, q6, q10 + vsub.s16 q10, q6, q10 + vadd.s16 q6, q15, q13 + vadd.s16 q8, q12, q14 + vsub.s16 q3, q6, q3 + vsub.s16 q12, q12, q14 + vsub.s16 q3, q3, q1 + vsub.s16 q1, q9, q1 + vadd.s16 q2, q3, q2 + vsub.s16 q15, q8, q6 + vadd.s16 q1, q1, q2 + vadd.s16 q8, q8, q6 + vadd.s16 q14, q5, q3 + vsub.s16 q9, q5, q3 + vsub.s16 q13, q10, q2 + vpop {d8-d13} /* restore NEON registers */ + vadd.s16 q10, q10, q2 + vsub.s16 q11, q12, q1 + vadd.s16 q12, q12, q1 + /* Descale to 8-bit and range limit */ + vmov.u8 q0, #0x80 + vqshrn.s16 d16, q8, #5 + vqshrn.s16 d17, q9, #5 + vqshrn.s16 d18, q10, #5 + vqshrn.s16 d19, q11, #5 + vqshrn.s16 d20, q12, #5 + vqshrn.s16 d21, q13, #5 + vqshrn.s16 d22, q14, #5 + vqshrn.s16 d23, q15, #5 + vadd.u8 q8, q8, q0 + vadd.u8 q9, q9, q0 + vadd.u8 q10, q10, q0 + vadd.u8 q11, q11, q0 + /* Transpose the final 8-bit samples */ + vtrn.16 q8, q9 + vtrn.16 q10, q11 + vtrn.32 q8, q10 + vtrn.32 q9, q11 + vtrn.8 d16, d17 + vtrn.8 d18, d19 + /* Store results to the output buffer */ + ldmia OUTPUT_BUF!, {TMP1, TMP2} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + vst1.8 {d16}, [TMP1] + vst1.8 {d17}, [TMP2] + ldmia OUTPUT_BUF!, {TMP1, TMP2} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + vst1.8 {d18}, [TMP1] + vtrn.8 d20, d21 + vst1.8 {d19}, [TMP2] + ldmia OUTPUT_BUF, {TMP1, TMP2, TMP3, TMP4} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + add TMP3, TMP3, OUTPUT_COL + add TMP4, TMP4, OUTPUT_COL + vst1.8 {d20}, [TMP1] + vtrn.8 d22, d23 + vst1.8 {d21}, [TMP2] + vst1.8 {d22}, [TMP3] + vst1.8 {d23}, [TMP4] + bx lr + + .unreq DCT_TABLE + .unreq COEF_BLOCK + .unreq OUTPUT_BUF + .unreq OUTPUT_COL + .unreq TMP1 + .unreq TMP2 + .unreq TMP3 + .unreq TMP4 +.endfunc + +/*****************************************************************************/ + +/* + * jsimd_idct_4x4_neon + * + * This function contains inverse-DCT code for getting reduced-size + * 4x4 pixels output from an 8x8 DCT block. It uses the same calculations + * and produces exactly the same output as IJG's original 'jpeg_idct_4x4' + * function from jpeg-6b (jidctred.c). + * + * NOTE: jpeg-8 has an improved implementation of 4x4 inverse-DCT, which + * requires much less arithmetic operations and hence should be faster. + * The primary purpose of this particular NEON optimized function is + * bit exact compatibility with jpeg-6b. + * + * TODO: a bit better instructions scheduling can be achieved by expanding + * idct_helper/transpose_4x4 macros and reordering instructions, + * but readability will suffer somewhat. + */ + +#define CONST_BITS 13 + +#define FIX_0_211164243 (1730) /* FIX(0.211164243) */ +#define FIX_0_509795579 (4176) /* FIX(0.509795579) */ +#define FIX_0_601344887 (4926) /* FIX(0.601344887) */ +#define FIX_0_720959822 (5906) /* FIX(0.720959822) */ +#define FIX_0_765366865 (6270) /* FIX(0.765366865) */ +#define FIX_0_850430095 (6967) /* FIX(0.850430095) */ +#define FIX_0_899976223 (7373) /* FIX(0.899976223) */ +#define FIX_1_061594337 (8697) /* FIX(1.061594337) */ +#define FIX_1_272758580 (10426) /* FIX(1.272758580) */ +#define FIX_1_451774981 (11893) /* FIX(1.451774981) */ +#define FIX_1_847759065 (15137) /* FIX(1.847759065) */ +#define FIX_2_172734803 (17799) /* FIX(2.172734803) */ +#define FIX_2_562915447 (20995) /* FIX(2.562915447) */ +#define FIX_3_624509785 (29692) /* FIX(3.624509785) */ + +.balign 16 +jsimd_idct_4x4_neon_consts: + .short FIX_1_847759065 /* d0[0] */ + .short -FIX_0_765366865 /* d0[1] */ + .short -FIX_0_211164243 /* d0[2] */ + .short FIX_1_451774981 /* d0[3] */ + .short -FIX_2_172734803 /* d1[0] */ + .short FIX_1_061594337 /* d1[1] */ + .short -FIX_0_509795579 /* d1[2] */ + .short -FIX_0_601344887 /* d1[3] */ + .short FIX_0_899976223 /* d2[0] */ + .short FIX_2_562915447 /* d2[1] */ + .short 1 << (CONST_BITS+1) /* d2[2] */ + .short 0 /* d2[3] */ + +.macro idct_helper x4, x6, x8, x10, x12, x14, x16, shift, y26, y27, y28, y29 + vmull.s16 q14, \x4, d2[2] + vmlal.s16 q14, \x8, d0[0] + vmlal.s16 q14, \x14, d0[1] + + vmull.s16 q13, \x16, d1[2] + vmlal.s16 q13, \x12, d1[3] + vmlal.s16 q13, \x10, d2[0] + vmlal.s16 q13, \x6, d2[1] + + vmull.s16 q15, \x4, d2[2] + vmlsl.s16 q15, \x8, d0[0] + vmlsl.s16 q15, \x14, d0[1] + + vmull.s16 q12, \x16, d0[2] + vmlal.s16 q12, \x12, d0[3] + vmlal.s16 q12, \x10, d1[0] + vmlal.s16 q12, \x6, d1[1] + + vadd.s32 q10, q14, q13 + vsub.s32 q14, q14, q13 + +.if \shift > 16 + vrshr.s32 q10, q10, #\shift + vrshr.s32 q14, q14, #\shift + vmovn.s32 \y26, q10 + vmovn.s32 \y29, q14 +.else + vrshrn.s32 \y26, q10, #\shift + vrshrn.s32 \y29, q14, #\shift +.endif + + vadd.s32 q10, q15, q12 + vsub.s32 q15, q15, q12 + +.if \shift > 16 + vrshr.s32 q10, q10, #\shift + vrshr.s32 q15, q15, #\shift + vmovn.s32 \y27, q10 + vmovn.s32 \y28, q15 +.else + vrshrn.s32 \y27, q10, #\shift + vrshrn.s32 \y28, q15, #\shift +.endif + +.endm + +asm_function jsimd_idct_4x4_neon + + DCT_TABLE .req r0 + COEF_BLOCK .req r1 + OUTPUT_BUF .req r2 + OUTPUT_COL .req r3 + TMP1 .req r0 + TMP2 .req r1 + TMP3 .req r2 + TMP4 .req ip + + vpush {d8-d15} + + /* Load constants (d3 is just used for padding) */ + adr TMP4, jsimd_idct_4x4_neon_consts + vld1.16 {d0, d1, d2, d3}, [TMP4, :128] + + /* Load all COEF_BLOCK into NEON registers with the following allocation: + * 0 1 2 3 | 4 5 6 7 + * ---------+-------- + * 0 | d4 | d5 + * 1 | d6 | d7 + * 2 | d8 | d9 + * 3 | d10 | d11 + * 4 | - | - + * 5 | d12 | d13 + * 6 | d14 | d15 + * 7 | d16 | d17 + */ + vld1.16 {d4, d5, d6, d7}, [COEF_BLOCK, :128]! + vld1.16 {d8, d9, d10, d11}, [COEF_BLOCK, :128]! + add COEF_BLOCK, COEF_BLOCK, #16 + vld1.16 {d12, d13, d14, d15}, [COEF_BLOCK, :128]! + vld1.16 {d16, d17}, [COEF_BLOCK, :128]! + /* dequantize */ + vld1.16 {d18, d19, d20, d21}, [DCT_TABLE, :128]! + vmul.s16 q2, q2, q9 + vld1.16 {d22, d23, d24, d25}, [DCT_TABLE, :128]! + vmul.s16 q3, q3, q10 + vmul.s16 q4, q4, q11 + add DCT_TABLE, DCT_TABLE, #16 + vld1.16 {d26, d27, d28, d29}, [DCT_TABLE, :128]! + vmul.s16 q5, q5, q12 + vmul.s16 q6, q6, q13 + vld1.16 {d30, d31}, [DCT_TABLE, :128]! + vmul.s16 q7, q7, q14 + vmul.s16 q8, q8, q15 + + /* Pass 1 */ + idct_helper d4, d6, d8, d10, d12, d14, d16, 12, d4, d6, d8, d10 + transpose_4x4 d4, d6, d8, d10 + idct_helper d5, d7, d9, d11, d13, d15, d17, 12, d5, d7, d9, d11 + transpose_4x4 d5, d7, d9, d11 + + /* Pass 2 */ + idct_helper d4, d6, d8, d10, d7, d9, d11, 19, d26, d27, d28, d29 + transpose_4x4 d26, d27, d28, d29 + + /* Range limit */ + vmov.u16 q15, #0x80 + vadd.s16 q13, q13, q15 + vadd.s16 q14, q14, q15 + vqmovun.s16 d26, q13 + vqmovun.s16 d27, q14 + + /* Store results to the output buffer */ + ldmia OUTPUT_BUF, {TMP1, TMP2, TMP3, TMP4} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + add TMP3, TMP3, OUTPUT_COL + add TMP4, TMP4, OUTPUT_COL + +#if defined(__ARMEL__) && !RESPECT_STRICT_ALIGNMENT + /* We can use much less instructions on little endian systems if the + * OS kernel is not configured to trap unaligned memory accesses + */ + vst1.32 {d26[0]}, [TMP1]! + vst1.32 {d27[0]}, [TMP3]! + vst1.32 {d26[1]}, [TMP2]! + vst1.32 {d27[1]}, [TMP4]! +#else + vst1.8 {d26[0]}, [TMP1]! + vst1.8 {d27[0]}, [TMP3]! + vst1.8 {d26[1]}, [TMP1]! + vst1.8 {d27[1]}, [TMP3]! + vst1.8 {d26[2]}, [TMP1]! + vst1.8 {d27[2]}, [TMP3]! + vst1.8 {d26[3]}, [TMP1]! + vst1.8 {d27[3]}, [TMP3]! + + vst1.8 {d26[4]}, [TMP2]! + vst1.8 {d27[4]}, [TMP4]! + vst1.8 {d26[5]}, [TMP2]! + vst1.8 {d27[5]}, [TMP4]! + vst1.8 {d26[6]}, [TMP2]! + vst1.8 {d27[6]}, [TMP4]! + vst1.8 {d26[7]}, [TMP2]! + vst1.8 {d27[7]}, [TMP4]! +#endif + + vpop {d8-d15} + bx lr + + .unreq DCT_TABLE + .unreq COEF_BLOCK + .unreq OUTPUT_BUF + .unreq OUTPUT_COL + .unreq TMP1 + .unreq TMP2 + .unreq TMP3 + .unreq TMP4 +.endfunc + +.purgem idct_helper + +/*****************************************************************************/ + +/* + * jsimd_idct_2x2_neon + * + * This function contains inverse-DCT code for getting reduced-size + * 2x2 pixels output from an 8x8 DCT block. It uses the same calculations + * and produces exactly the same output as IJG's original 'jpeg_idct_2x2' + * function from jpeg-6b (jidctred.c). + * + * NOTE: jpeg-8 has an improved implementation of 2x2 inverse-DCT, which + * requires much less arithmetic operations and hence should be faster. + * The primary purpose of this particular NEON optimized function is + * bit exact compatibility with jpeg-6b. + */ + +.balign 8 +jsimd_idct_2x2_neon_consts: + .short -FIX_0_720959822 /* d0[0] */ + .short FIX_0_850430095 /* d0[1] */ + .short -FIX_1_272758580 /* d0[2] */ + .short FIX_3_624509785 /* d0[3] */ + +.macro idct_helper x4, x6, x10, x12, x16, shift, y26, y27 + vshll.s16 q14, \x4, #15 + vmull.s16 q13, \x6, d0[3] + vmlal.s16 q13, \x10, d0[2] + vmlal.s16 q13, \x12, d0[1] + vmlal.s16 q13, \x16, d0[0] + + vadd.s32 q10, q14, q13 + vsub.s32 q14, q14, q13 + +.if \shift > 16 + vrshr.s32 q10, q10, #\shift + vrshr.s32 q14, q14, #\shift + vmovn.s32 \y26, q10 + vmovn.s32 \y27, q14 +.else + vrshrn.s32 \y26, q10, #\shift + vrshrn.s32 \y27, q14, #\shift +.endif + +.endm + +asm_function jsimd_idct_2x2_neon + + DCT_TABLE .req r0 + COEF_BLOCK .req r1 + OUTPUT_BUF .req r2 + OUTPUT_COL .req r3 + TMP1 .req r0 + TMP2 .req ip + + vpush {d8-d15} + + /* Load constants */ + adr TMP2, jsimd_idct_2x2_neon_consts + vld1.16 {d0}, [TMP2, :64] + + /* Load all COEF_BLOCK into NEON registers with the following allocation: + * 0 1 2 3 | 4 5 6 7 + * ---------+-------- + * 0 | d4 | d5 + * 1 | d6 | d7 + * 2 | - | - + * 3 | d10 | d11 + * 4 | - | - + * 5 | d12 | d13 + * 6 | - | - + * 7 | d16 | d17 + */ + vld1.16 {d4, d5, d6, d7}, [COEF_BLOCK, :128]! + add COEF_BLOCK, COEF_BLOCK, #16 + vld1.16 {d10, d11}, [COEF_BLOCK, :128]! + add COEF_BLOCK, COEF_BLOCK, #16 + vld1.16 {d12, d13}, [COEF_BLOCK, :128]! + add COEF_BLOCK, COEF_BLOCK, #16 + vld1.16 {d16, d17}, [COEF_BLOCK, :128]! + /* Dequantize */ + vld1.16 {d18, d19, d20, d21}, [DCT_TABLE, :128]! + vmul.s16 q2, q2, q9 + vmul.s16 q3, q3, q10 + add DCT_TABLE, DCT_TABLE, #16 + vld1.16 {d24, d25}, [DCT_TABLE, :128]! + vmul.s16 q5, q5, q12 + add DCT_TABLE, DCT_TABLE, #16 + vld1.16 {d26, d27}, [DCT_TABLE, :128]! + vmul.s16 q6, q6, q13 + add DCT_TABLE, DCT_TABLE, #16 + vld1.16 {d30, d31}, [DCT_TABLE, :128]! + vmul.s16 q8, q8, q15 + + /* Pass 1 */ +#if 0 + idct_helper d4, d6, d10, d12, d16, 13, d4, d6 + transpose_4x4 d4, d6, d8, d10 + idct_helper d5, d7, d11, d13, d17, 13, d5, d7 + transpose_4x4 d5, d7, d9, d11 +#else + vmull.s16 q13, d6, d0[3] + vmlal.s16 q13, d10, d0[2] + vmlal.s16 q13, d12, d0[1] + vmlal.s16 q13, d16, d0[0] + vmull.s16 q12, d7, d0[3] + vmlal.s16 q12, d11, d0[2] + vmlal.s16 q12, d13, d0[1] + vmlal.s16 q12, d17, d0[0] + vshll.s16 q14, d4, #15 + vshll.s16 q15, d5, #15 + vadd.s32 q10, q14, q13 + vsub.s32 q14, q14, q13 + vrshrn.s32 d4, q10, #13 + vrshrn.s32 d6, q14, #13 + vadd.s32 q10, q15, q12 + vsub.s32 q14, q15, q12 + vrshrn.s32 d5, q10, #13 + vrshrn.s32 d7, q14, #13 + vtrn.16 q2, q3 + vtrn.32 q3, q5 +#endif + + /* Pass 2 */ + idct_helper d4, d6, d10, d7, d11, 20, d26, d27 + + /* Range limit */ + vmov.u16 q15, #0x80 + vadd.s16 q13, q13, q15 + vqmovun.s16 d26, q13 + vqmovun.s16 d27, q13 + + /* Store results to the output buffer */ + ldmia OUTPUT_BUF, {TMP1, TMP2} + add TMP1, TMP1, OUTPUT_COL + add TMP2, TMP2, OUTPUT_COL + + vst1.8 {d26[0]}, [TMP1]! + vst1.8 {d27[4]}, [TMP1]! + vst1.8 {d26[1]}, [TMP2]! + vst1.8 {d27[5]}, [TMP2]! + + vpop {d8-d15} + bx lr + + .unreq DCT_TABLE + .unreq COEF_BLOCK + .unreq OUTPUT_BUF + .unreq OUTPUT_COL + .unreq TMP1 + .unreq TMP2 +.endfunc + +.purgem idct_helper + +/*****************************************************************************/ + +/* + * jsimd_ycc_extrgb_convert_neon + * jsimd_ycc_extbgr_convert_neon + * jsimd_ycc_extrgbx_convert_neon + * jsimd_ycc_extbgrx_convert_neon + * jsimd_ycc_extxbgr_convert_neon + * jsimd_ycc_extxrgb_convert_neon + * + * Colorspace conversion YCbCr -> RGB + */ + + +.macro do_load size + .if \size == 8 + vld1.8 {d4}, [U, :64]! + vld1.8 {d5}, [V, :64]! + vld1.8 {d0}, [Y, :64]! + pld [U, #64] + pld [V, #64] + pld [Y, #64] + .elseif \size == 4 + vld1.8 {d4[0]}, [U]! + vld1.8 {d4[1]}, [U]! + vld1.8 {d4[2]}, [U]! + vld1.8 {d4[3]}, [U]! + vld1.8 {d5[0]}, [V]! + vld1.8 {d5[1]}, [V]! + vld1.8 {d5[2]}, [V]! + vld1.8 {d5[3]}, [V]! + vld1.8 {d0[0]}, [Y]! + vld1.8 {d0[1]}, [Y]! + vld1.8 {d0[2]}, [Y]! + vld1.8 {d0[3]}, [Y]! + .elseif \size == 2 + vld1.8 {d4[4]}, [U]! + vld1.8 {d4[5]}, [U]! + vld1.8 {d5[4]}, [V]! + vld1.8 {d5[5]}, [V]! + vld1.8 {d0[4]}, [Y]! + vld1.8 {d0[5]}, [Y]! + .elseif \size == 1 + vld1.8 {d4[6]}, [U]! + vld1.8 {d5[6]}, [V]! + vld1.8 {d0[6]}, [Y]! + .else + .error unsupported macroblock size + .endif +.endm + +.macro do_store bpp, size + .if \bpp == 24 + .if \size == 8 + vst3.8 {d10, d11, d12}, [RGB]! + .elseif \size == 4 + vst3.8 {d10[0], d11[0], d12[0]}, [RGB]! + vst3.8 {d10[1], d11[1], d12[1]}, [RGB]! + vst3.8 {d10[2], d11[2], d12[2]}, [RGB]! + vst3.8 {d10[3], d11[3], d12[3]}, [RGB]! + .elseif \size == 2 + vst3.8 {d10[4], d11[4], d12[4]}, [RGB]! + vst3.8 {d10[5], d11[5], d12[5]}, [RGB]! + .elseif \size == 1 + vst3.8 {d10[6], d11[6], d12[6]}, [RGB]! + .else + .error unsupported macroblock size + .endif + .elseif \bpp == 32 + .if \size == 8 + vst4.8 {d10, d11, d12, d13}, [RGB]! + .elseif \size == 4 + vst4.8 {d10[0], d11[0], d12[0], d13[0]}, [RGB]! + vst4.8 {d10[1], d11[1], d12[1], d13[1]}, [RGB]! + vst4.8 {d10[2], d11[2], d12[2], d13[2]}, [RGB]! + vst4.8 {d10[3], d11[3], d12[3], d13[3]}, [RGB]! + .elseif \size == 2 + vst4.8 {d10[4], d11[4], d12[4], d13[4]}, [RGB]! + vst4.8 {d10[5], d11[5], d12[5], d13[5]}, [RGB]! + .elseif \size == 1 + vst4.8 {d10[6], d11[6], d12[6], d13[6]}, [RGB]! + .else + .error unsupported macroblock size + .endif + .else + .error unsupported bpp + .endif +.endm + +.macro generate_jsimd_ycc_rgb_convert_neon colorid, bpp, r_offs, g_offs, b_offs + +/* + * 2 stage pipelined YCbCr->RGB conversion + */ + +.macro do_yuv_to_rgb_stage1 + vaddw.u8 q3, q1, d4 /* q3 = u - 128 */ + vaddw.u8 q4, q1, d5 /* q2 = v - 128 */ + vmull.s16 q10, d6, d1[1] /* multiply by -11277 */ + vmlal.s16 q10, d8, d1[2] /* multiply by -23401 */ + vmull.s16 q11, d7, d1[1] /* multiply by -11277 */ + vmlal.s16 q11, d9, d1[2] /* multiply by -23401 */ + vmull.s16 q12, d8, d1[0] /* multiply by 22971 */ + vmull.s16 q13, d9, d1[0] /* multiply by 22971 */ + vmull.s16 q14, d6, d1[3] /* multiply by 29033 */ + vmull.s16 q15, d7, d1[3] /* multiply by 29033 */ +.endm + +.macro do_yuv_to_rgb_stage2 + vrshrn.s32 d20, q10, #15 + vrshrn.s32 d21, q11, #15 + vrshrn.s32 d24, q12, #14 + vrshrn.s32 d25, q13, #14 + vrshrn.s32 d28, q14, #14 + vrshrn.s32 d29, q15, #14 + vaddw.u8 q10, q10, d0 + vaddw.u8 q12, q12, d0 + vaddw.u8 q14, q14, d0 + vqmovun.s16 d1\g_offs, q10 + vqmovun.s16 d1\r_offs, q12 + vqmovun.s16 d1\b_offs, q14 +.endm + +.macro do_yuv_to_rgb_stage2_store_load_stage1 + vld1.8 {d4}, [U, :64]! + vrshrn.s32 d20, q10, #15 + vrshrn.s32 d21, q11, #15 + vrshrn.s32 d24, q12, #14 + vrshrn.s32 d25, q13, #14 + vrshrn.s32 d28, q14, #14 + vld1.8 {d5}, [V, :64]! + vrshrn.s32 d29, q15, #14 + vaddw.u8 q10, q10, d0 + vaddw.u8 q12, q12, d0 + vaddw.u8 q14, q14, d0 + vqmovun.s16 d1\g_offs, q10 + vld1.8 {d0}, [Y, :64]! + vqmovun.s16 d1\r_offs, q12 + pld [U, #64] + pld [V, #64] + pld [Y, #64] + vqmovun.s16 d1\b_offs, q14 + vaddw.u8 q3, q1, d4 /* q3 = u - 128 */ + vaddw.u8 q4, q1, d5 /* q2 = v - 128 */ + do_store \bpp, 8 + vmull.s16 q10, d6, d1[1] /* multiply by -11277 */ + vmlal.s16 q10, d8, d1[2] /* multiply by -23401 */ + vmull.s16 q11, d7, d1[1] /* multiply by -11277 */ + vmlal.s16 q11, d9, d1[2] /* multiply by -23401 */ + vmull.s16 q12, d8, d1[0] /* multiply by 22971 */ + vmull.s16 q13, d9, d1[0] /* multiply by 22971 */ + vmull.s16 q14, d6, d1[3] /* multiply by 29033 */ + vmull.s16 q15, d7, d1[3] /* multiply by 29033 */ +.endm + +.macro do_yuv_to_rgb + do_yuv_to_rgb_stage1 + do_yuv_to_rgb_stage2 +.endm + +/* Apple gas crashes on adrl, work around that by using adr. + * But this requires a copy of these constants for each function. + */ + +.balign 16 +jsimd_ycc_\colorid\()_neon_consts: + .short 0, 0, 0, 0 + .short 22971, -11277, -23401, 29033 + .short -128, -128, -128, -128 + .short -128, -128, -128, -128 + +asm_function jsimd_ycc_\colorid\()_convert_neon + OUTPUT_WIDTH .req r0 + INPUT_BUF .req r1 + INPUT_ROW .req r2 + OUTPUT_BUF .req r3 + NUM_ROWS .req r4 + + INPUT_BUF0 .req r5 + INPUT_BUF1 .req r6 + INPUT_BUF2 .req INPUT_BUF + + RGB .req r7 + Y .req r8 + U .req r9 + V .req r10 + N .req ip + + /* Load constants to d1, d2, d3 (d0 is just used for padding) */ + adr ip, jsimd_ycc_\colorid\()_neon_consts + vld1.16 {d0, d1, d2, d3}, [ip, :128] + + /* Save ARM registers and handle input arguments */ + push {r4, r5, r6, r7, r8, r9, r10, lr} + ldr NUM_ROWS, [sp, #(4 * 8)] + ldr INPUT_BUF0, [INPUT_BUF] + ldr INPUT_BUF1, [INPUT_BUF, #4] + ldr INPUT_BUF2, [INPUT_BUF, #8] + .unreq INPUT_BUF + + /* Save NEON registers */ + vpush {d8-d15} + + /* Initially set d10, d11, d12, d13 to 0xFF */ + vmov.u8 q5, #255 + vmov.u8 q6, #255 + + /* Outer loop over scanlines */ + cmp NUM_ROWS, #1 + blt 9f +0: + ldr Y, [INPUT_BUF0, INPUT_ROW, lsl #2] + ldr U, [INPUT_BUF1, INPUT_ROW, lsl #2] + mov N, OUTPUT_WIDTH + ldr V, [INPUT_BUF2, INPUT_ROW, lsl #2] + add INPUT_ROW, INPUT_ROW, #1 + ldr RGB, [OUTPUT_BUF], #4 + + /* Inner loop over pixels */ + subs N, N, #8 + blt 3f + do_load 8 + do_yuv_to_rgb_stage1 + subs N, N, #8 + blt 2f +1: + do_yuv_to_rgb_stage2_store_load_stage1 + subs N, N, #8 + bge 1b +2: + do_yuv_to_rgb_stage2 + do_store \bpp, 8 + tst N, #7 + beq 8f +3: + tst N, #4 + beq 3f + do_load 4 +3: + tst N, #2 + beq 4f + do_load 2 +4: + tst N, #1 + beq 5f + do_load 1 +5: + do_yuv_to_rgb + tst N, #4 + beq 6f + do_store \bpp, 4 +6: + tst N, #2 + beq 7f + do_store \bpp, 2 +7: + tst N, #1 + beq 8f + do_store \bpp, 1 +8: + subs NUM_ROWS, NUM_ROWS, #1 + bgt 0b +9: + /* Restore all registers and return */ + vpop {d8-d15} + pop {r4, r5, r6, r7, r8, r9, r10, pc} + + .unreq OUTPUT_WIDTH + .unreq INPUT_ROW + .unreq OUTPUT_BUF + .unreq NUM_ROWS + .unreq INPUT_BUF0 + .unreq INPUT_BUF1 + .unreq INPUT_BUF2 + .unreq RGB + .unreq Y + .unreq U + .unreq V + .unreq N +.endfunc + +.purgem do_yuv_to_rgb +.purgem do_yuv_to_rgb_stage1 +.purgem do_yuv_to_rgb_stage2 +.purgem do_yuv_to_rgb_stage2_store_load_stage1 + +.endm + +/*--------------------------------- id ----- bpp R G B */ +generate_jsimd_ycc_rgb_convert_neon extrgb, 24, 0, 1, 2 +generate_jsimd_ycc_rgb_convert_neon extbgr, 24, 2, 1, 0 +generate_jsimd_ycc_rgb_convert_neon extrgbx, 32, 0, 1, 2 +generate_jsimd_ycc_rgb_convert_neon extbgrx, 32, 2, 1, 0 +generate_jsimd_ycc_rgb_convert_neon extxbgr, 32, 3, 2, 1 +generate_jsimd_ycc_rgb_convert_neon extxrgb, 32, 1, 2, 3 + +.purgem do_load +.purgem do_store + +/*****************************************************************************/ + +/* + * jsimd_extrgb_ycc_convert_neon + * jsimd_extbgr_ycc_convert_neon + * jsimd_extrgbx_ycc_convert_neon + * jsimd_extbgrx_ycc_convert_neon + * jsimd_extxbgr_ycc_convert_neon + * jsimd_extxrgb_ycc_convert_neon + * + * Colorspace conversion RGB -> YCbCr + */ + +.macro do_store size + .if \size == 8 + vst1.8 {d20}, [Y]! + vst1.8 {d21}, [U]! + vst1.8 {d22}, [V]! + .elseif \size == 4 + vst1.8 {d20[0]}, [Y]! + vst1.8 {d20[1]}, [Y]! + vst1.8 {d20[2]}, [Y]! + vst1.8 {d20[3]}, [Y]! + vst1.8 {d21[0]}, [U]! + vst1.8 {d21[1]}, [U]! + vst1.8 {d21[2]}, [U]! + vst1.8 {d21[3]}, [U]! + vst1.8 {d22[0]}, [V]! + vst1.8 {d22[1]}, [V]! + vst1.8 {d22[2]}, [V]! + vst1.8 {d22[3]}, [V]! + .elseif \size == 2 + vst1.8 {d20[4]}, [Y]! + vst1.8 {d20[5]}, [Y]! + vst1.8 {d21[4]}, [U]! + vst1.8 {d21[5]}, [U]! + vst1.8 {d22[4]}, [V]! + vst1.8 {d22[5]}, [V]! + .elseif \size == 1 + vst1.8 {d20[6]}, [Y]! + vst1.8 {d21[6]}, [U]! + vst1.8 {d22[6]}, [V]! + .else + .error unsupported macroblock size + .endif +.endm + +.macro do_load bpp, size + .if \bpp == 24 + .if \size == 8 + vld3.8 {d10, d11, d12}, [RGB]! + pld [RGB, #128] + .elseif \size == 4 + vld3.8 {d10[0], d11[0], d12[0]}, [RGB]! + vld3.8 {d10[1], d11[1], d12[1]}, [RGB]! + vld3.8 {d10[2], d11[2], d12[2]}, [RGB]! + vld3.8 {d10[3], d11[3], d12[3]}, [RGB]! + .elseif \size == 2 + vld3.8 {d10[4], d11[4], d12[4]}, [RGB]! + vld3.8 {d10[5], d11[5], d12[5]}, [RGB]! + .elseif \size == 1 + vld3.8 {d10[6], d11[6], d12[6]}, [RGB]! + .else + .error unsupported macroblock size + .endif + .elseif \bpp == 32 + .if \size == 8 + vld4.8 {d10, d11, d12, d13}, [RGB]! + pld [RGB, #128] + .elseif \size == 4 + vld4.8 {d10[0], d11[0], d12[0], d13[0]}, [RGB]! + vld4.8 {d10[1], d11[1], d12[1], d13[1]}, [RGB]! + vld4.8 {d10[2], d11[2], d12[2], d13[2]}, [RGB]! + vld4.8 {d10[3], d11[3], d12[3], d13[3]}, [RGB]! + .elseif \size == 2 + vld4.8 {d10[4], d11[4], d12[4], d13[4]}, [RGB]! + vld4.8 {d10[5], d11[5], d12[5], d13[5]}, [RGB]! + .elseif \size == 1 + vld4.8 {d10[6], d11[6], d12[6], d13[6]}, [RGB]! + .else + .error unsupported macroblock size + .endif + .else + .error unsupported bpp + .endif +.endm + +.macro generate_jsimd_rgb_ycc_convert_neon colorid, bpp, r_offs, g_offs, b_offs + +/* + * 2 stage pipelined RGB->YCbCr conversion + */ + +.macro do_rgb_to_yuv_stage1 + vmovl.u8 q2, d1\r_offs /* r = { d4, d5 } */ + vmovl.u8 q3, d1\g_offs /* g = { d6, d7 } */ + vmovl.u8 q4, d1\b_offs /* b = { d8, d9 } */ + vmull.u16 q7, d4, d0[0] + vmlal.u16 q7, d6, d0[1] + vmlal.u16 q7, d8, d0[2] + vmull.u16 q8, d5, d0[0] + vmlal.u16 q8, d7, d0[1] + vmlal.u16 q8, d9, d0[2] + vrev64.32 q9, q1 + vrev64.32 q13, q1 + vmlsl.u16 q9, d4, d0[3] + vmlsl.u16 q9, d6, d1[0] + vmlal.u16 q9, d8, d1[1] + vmlsl.u16 q13, d5, d0[3] + vmlsl.u16 q13, d7, d1[0] + vmlal.u16 q13, d9, d1[1] + vrev64.32 q14, q1 + vrev64.32 q15, q1 + vmlal.u16 q14, d4, d1[1] + vmlsl.u16 q14, d6, d1[2] + vmlsl.u16 q14, d8, d1[3] + vmlal.u16 q15, d5, d1[1] + vmlsl.u16 q15, d7, d1[2] + vmlsl.u16 q15, d9, d1[3] +.endm + +.macro do_rgb_to_yuv_stage2 + vrshrn.u32 d20, q7, #16 + vrshrn.u32 d21, q8, #16 + vshrn.u32 d22, q9, #16 + vshrn.u32 d23, q13, #16 + vshrn.u32 d24, q14, #16 + vshrn.u32 d25, q15, #16 + vmovn.u16 d20, q10 /* d20 = y */ + vmovn.u16 d21, q11 /* d21 = u */ + vmovn.u16 d22, q12 /* d22 = v */ +.endm + +.macro do_rgb_to_yuv + do_rgb_to_yuv_stage1 + do_rgb_to_yuv_stage2 +.endm + +.macro do_rgb_to_yuv_stage2_store_load_stage1 + vrshrn.u32 d20, q7, #16 + vrshrn.u32 d21, q8, #16 + vshrn.u32 d22, q9, #16 + vrev64.32 q9, q1 + vshrn.u32 d23, q13, #16 + vrev64.32 q13, q1 + vshrn.u32 d24, q14, #16 + vshrn.u32 d25, q15, #16 + do_load \bpp, 8 + vmovn.u16 d20, q10 /* d20 = y */ + vmovl.u8 q2, d1\r_offs /* r = { d4, d5 } */ + vmovn.u16 d21, q11 /* d21 = u */ + vmovl.u8 q3, d1\g_offs /* g = { d6, d7 } */ + vmovn.u16 d22, q12 /* d22 = v */ + vmovl.u8 q4, d1\b_offs /* b = { d8, d9 } */ + vmull.u16 q7, d4, d0[0] + vmlal.u16 q7, d6, d0[1] + vmlal.u16 q7, d8, d0[2] + vst1.8 {d20}, [Y]! + vmull.u16 q8, d5, d0[0] + vmlal.u16 q8, d7, d0[1] + vmlal.u16 q8, d9, d0[2] + vmlsl.u16 q9, d4, d0[3] + vmlsl.u16 q9, d6, d1[0] + vmlal.u16 q9, d8, d1[1] + vst1.8 {d21}, [U]! + vmlsl.u16 q13, d5, d0[3] + vmlsl.u16 q13, d7, d1[0] + vmlal.u16 q13, d9, d1[1] + vrev64.32 q14, q1 + vrev64.32 q15, q1 + vmlal.u16 q14, d4, d1[1] + vmlsl.u16 q14, d6, d1[2] + vmlsl.u16 q14, d8, d1[3] + vst1.8 {d22}, [V]! + vmlal.u16 q15, d5, d1[1] + vmlsl.u16 q15, d7, d1[2] + vmlsl.u16 q15, d9, d1[3] +.endm + +.balign 16 +jsimd_\colorid\()_ycc_neon_consts: + .short 19595, 38470, 7471, 11059 + .short 21709, 32768, 27439, 5329 + .short 32767, 128, 32767, 128 + .short 32767, 128, 32767, 128 + +asm_function jsimd_\colorid\()_ycc_convert_neon + OUTPUT_WIDTH .req r0 + INPUT_BUF .req r1 + OUTPUT_BUF .req r2 + OUTPUT_ROW .req r3 + NUM_ROWS .req r4 + + OUTPUT_BUF0 .req r5 + OUTPUT_BUF1 .req r6 + OUTPUT_BUF2 .req OUTPUT_BUF + + RGB .req r7 + Y .req r8 + U .req r9 + V .req r10 + N .req ip + + /* Load constants to d0, d1, d2, d3 */ + adr ip, jsimd_\colorid\()_ycc_neon_consts + vld1.16 {d0, d1, d2, d3}, [ip, :128] + + /* Save ARM registers and handle input arguments */ + push {r4, r5, r6, r7, r8, r9, r10, lr} + ldr NUM_ROWS, [sp, #(4 * 8)] + ldr OUTPUT_BUF0, [OUTPUT_BUF] + ldr OUTPUT_BUF1, [OUTPUT_BUF, #4] + ldr OUTPUT_BUF2, [OUTPUT_BUF, #8] + .unreq OUTPUT_BUF + + /* Save NEON registers */ + vpush {d8-d15} + + /* Outer loop over scanlines */ + cmp NUM_ROWS, #1 + blt 9f +0: + ldr Y, [OUTPUT_BUF0, OUTPUT_ROW, lsl #2] + ldr U, [OUTPUT_BUF1, OUTPUT_ROW, lsl #2] + mov N, OUTPUT_WIDTH + ldr V, [OUTPUT_BUF2, OUTPUT_ROW, lsl #2] + add OUTPUT_ROW, OUTPUT_ROW, #1 + ldr RGB, [INPUT_BUF], #4 + + /* Inner loop over pixels */ + subs N, N, #8 + blt 3f + do_load \bpp, 8 + do_rgb_to_yuv_stage1 + subs N, N, #8 + blt 2f +1: + do_rgb_to_yuv_stage2_store_load_stage1 + subs N, N, #8 + bge 1b +2: + do_rgb_to_yuv_stage2 + do_store 8 + tst N, #7 + beq 8f +3: + tst N, #4 + beq 3f + do_load \bpp, 4 +3: + tst N, #2 + beq 4f + do_load \bpp, 2 +4: + tst N, #1 + beq 5f + do_load \bpp, 1 +5: + do_rgb_to_yuv + tst N, #4 + beq 6f + do_store 4 +6: + tst N, #2 + beq 7f + do_store 2 +7: + tst N, #1 + beq 8f + do_store 1 +8: + subs NUM_ROWS, NUM_ROWS, #1 + bgt 0b +9: + /* Restore all registers and return */ + vpop {d8-d15} + pop {r4, r5, r6, r7, r8, r9, r10, pc} + + .unreq OUTPUT_WIDTH + .unreq OUTPUT_ROW + .unreq INPUT_BUF + .unreq NUM_ROWS + .unreq OUTPUT_BUF0 + .unreq OUTPUT_BUF1 + .unreq OUTPUT_BUF2 + .unreq RGB + .unreq Y + .unreq U + .unreq V + .unreq N +.endfunc + +.purgem do_rgb_to_yuv +.purgem do_rgb_to_yuv_stage1 +.purgem do_rgb_to_yuv_stage2 +.purgem do_rgb_to_yuv_stage2_store_load_stage1 + +.endm + +/*--------------------------------- id ----- bpp R G B */ +generate_jsimd_rgb_ycc_convert_neon extrgb, 24, 0, 1, 2 +generate_jsimd_rgb_ycc_convert_neon extbgr, 24, 2, 1, 0 +generate_jsimd_rgb_ycc_convert_neon extrgbx, 32, 0, 1, 2 +generate_jsimd_rgb_ycc_convert_neon extbgrx, 32, 2, 1, 0 +generate_jsimd_rgb_ycc_convert_neon extxbgr, 32, 3, 2, 1 +generate_jsimd_rgb_ycc_convert_neon extxrgb, 32, 1, 2, 3 + +.purgem do_load +.purgem do_store + +/*****************************************************************************/ + +/* + * Load data into workspace, applying unsigned->signed conversion + * + * TODO: can be combined with 'jsimd_fdct_ifast_neon' to get + * rid of VST1.16 instructions + */ + +asm_function jsimd_convsamp_neon + SAMPLE_DATA .req r0 + START_COL .req r1 + WORKSPACE .req r2 + TMP1 .req r3 + TMP2 .req r4 + TMP3 .req r5 + TMP4 .req ip + + push {r4, r5} + vmov.u8 d0, #128 + + ldmia SAMPLE_DATA!, {TMP1, TMP2, TMP3, TMP4} + add TMP1, TMP1, START_COL + add TMP2, TMP2, START_COL + add TMP3, TMP3, START_COL + add TMP4, TMP4, START_COL + vld1.8 {d16}, [TMP1] + vsubl.u8 q8, d16, d0 + vld1.8 {d18}, [TMP2] + vsubl.u8 q9, d18, d0 + vld1.8 {d20}, [TMP3] + vsubl.u8 q10, d20, d0 + vld1.8 {d22}, [TMP4] + ldmia SAMPLE_DATA!, {TMP1, TMP2, TMP3, TMP4} + vsubl.u8 q11, d22, d0 + vst1.16 {d16, d17, d18, d19}, [WORKSPACE, :128]! + add TMP1, TMP1, START_COL + add TMP2, TMP2, START_COL + vst1.16 {d20, d21, d22, d23}, [WORKSPACE, :128]! + add TMP3, TMP3, START_COL + add TMP4, TMP4, START_COL + vld1.8 {d24}, [TMP1] + vsubl.u8 q12, d24, d0 + vld1.8 {d26}, [TMP2] + vsubl.u8 q13, d26, d0 + vld1.8 {d28}, [TMP3] + vsubl.u8 q14, d28, d0 + vld1.8 {d30}, [TMP4] + vsubl.u8 q15, d30, d0 + vst1.16 {d24, d25, d26, d27}, [WORKSPACE, :128]! + vst1.16 {d28, d29, d30, d31}, [WORKSPACE, :128]! + pop {r4, r5} + bx lr + + .unreq SAMPLE_DATA + .unreq START_COL + .unreq WORKSPACE + .unreq TMP1 + .unreq TMP2 + .unreq TMP3 + .unreq TMP4 +.endfunc + +/*****************************************************************************/ + +/* + * jsimd_fdct_ifast_neon + * + * This function contains a fast, not so accurate integer implementation of + * the forward DCT (Discrete Cosine Transform). It uses the same calculations + * and produces exactly the same output as IJG's original 'jpeg_fdct_ifast' + * function from jfdctfst.c + * + * TODO: can be combined with 'jsimd_convsamp_neon' to get + * rid of a bunch of VLD1.16 instructions + */ + +#define XFIX_0_382683433 d0[0] +#define XFIX_0_541196100 d0[1] +#define XFIX_0_707106781 d0[2] +#define XFIX_1_306562965 d0[3] + +.balign 16 +jsimd_fdct_ifast_neon_consts: + .short (98 * 128) /* XFIX_0_382683433 */ + .short (139 * 128) /* XFIX_0_541196100 */ + .short (181 * 128) /* XFIX_0_707106781 */ + .short (334 * 128 - 256 * 128) /* XFIX_1_306562965 */ + +asm_function jsimd_fdct_ifast_neon + + DATA .req r0 + TMP .req ip + + vpush {d8-d15} + + /* Load constants */ + adr TMP, jsimd_fdct_ifast_neon_consts + vld1.16 {d0}, [TMP, :64] + + /* Load all DATA into NEON registers with the following allocation: + * 0 1 2 3 | 4 5 6 7 + * ---------+-------- + * 0 | d16 | d17 | q8 + * 1 | d18 | d19 | q9 + * 2 | d20 | d21 | q10 + * 3 | d22 | d23 | q11 + * 4 | d24 | d25 | q12 + * 5 | d26 | d27 | q13 + * 6 | d28 | d29 | q14 + * 7 | d30 | d31 | q15 + */ + + vld1.16 {d16, d17, d18, d19}, [DATA, :128]! + vld1.16 {d20, d21, d22, d23}, [DATA, :128]! + vld1.16 {d24, d25, d26, d27}, [DATA, :128]! + vld1.16 {d28, d29, d30, d31}, [DATA, :128] + sub DATA, DATA, #(128 - 32) + + mov TMP, #2 +1: + /* Transpose */ + vtrn.16 q12, q13 + vtrn.16 q10, q11 + vtrn.16 q8, q9 + vtrn.16 q14, q15 + vtrn.32 q9, q11 + vtrn.32 q13, q15 + vtrn.32 q8, q10 + vtrn.32 q12, q14 + vswp d30, d23 + vswp d24, d17 + vswp d26, d19 + /* 1-D FDCT */ + vadd.s16 q2, q11, q12 + vswp d28, d21 + vsub.s16 q12, q11, q12 + vsub.s16 q6, q10, q13 + vadd.s16 q10, q10, q13 + vsub.s16 q7, q9, q14 + vadd.s16 q9, q9, q14 + vsub.s16 q1, q8, q15 + vadd.s16 q8, q8, q15 + vsub.s16 q4, q9, q10 + vsub.s16 q5, q8, q2 + vadd.s16 q3, q9, q10 + vadd.s16 q4, q4, q5 + vadd.s16 q2, q8, q2 + vqdmulh.s16 q4, q4, XFIX_0_707106781 + vadd.s16 q11, q12, q6 + vadd.s16 q8, q2, q3 + vsub.s16 q12, q2, q3 + vadd.s16 q3, q6, q7 + vadd.s16 q7, q7, q1 + vqdmulh.s16 q3, q3, XFIX_0_707106781 + vsub.s16 q6, q11, q7 + vadd.s16 q10, q5, q4 + vqdmulh.s16 q6, q6, XFIX_0_382683433 + vsub.s16 q14, q5, q4 + vqdmulh.s16 q11, q11, XFIX_0_541196100 + vqdmulh.s16 q5, q7, XFIX_1_306562965 + vadd.s16 q4, q1, q3 + vsub.s16 q3, q1, q3 + vadd.s16 q7, q7, q6 + vadd.s16 q11, q11, q6 + vadd.s16 q7, q7, q5 + vadd.s16 q13, q3, q11 + vsub.s16 q11, q3, q11 + vadd.s16 q9, q4, q7 + vsub.s16 q15, q4, q7 + subs TMP, TMP, #1 + bne 1b + + /* store results */ + vst1.16 {d16, d17, d18, d19}, [DATA, :128]! + vst1.16 {d20, d21, d22, d23}, [DATA, :128]! + vst1.16 {d24, d25, d26, d27}, [DATA, :128]! + vst1.16 {d28, d29, d30, d31}, [DATA, :128] + + vpop {d8-d15} + bx lr + + .unreq DATA + .unreq TMP +.endfunc + +/*****************************************************************************/ + +/* + * GLOBAL(void) + * jsimd_quantize_neon (JCOEFPTR coef_block, DCTELEM * divisors, + * DCTELEM * workspace); + * + * Note: the code uses 2 stage pipelining in order to improve instructions + * scheduling and eliminate stalls (this provides ~15% better + * performance for this function on both ARM Cortex-A8 and + * ARM Cortex-A9 when compared to the non-pipelined variant). + * The instructions which belong to the second stage use different + * indentation for better readiability. + */ +asm_function jsimd_quantize_neon + + COEF_BLOCK .req r0 + DIVISORS .req r1 + WORKSPACE .req r2 + + RECIPROCAL .req DIVISORS + CORRECTION .req r3 + SHIFT .req ip + LOOP_COUNT .req r4 + + vld1.16 {d0, d1, d2, d3}, [WORKSPACE, :128]! + vabs.s16 q12, q0 + add CORRECTION, DIVISORS, #(64 * 2) + add SHIFT, DIVISORS, #(64 * 6) + vld1.16 {d20, d21, d22, d23}, [CORRECTION, :128]! + vabs.s16 q13, q1 + vld1.16 {d16, d17, d18, d19}, [RECIPROCAL, :128]! + vadd.u16 q12, q12, q10 /* add correction */ + vadd.u16 q13, q13, q11 + vmull.u16 q10, d24, d16 /* multiply by reciprocal */ + vmull.u16 q11, d25, d17 + vmull.u16 q8, d26, d18 + vmull.u16 q9, d27, d19 + vld1.16 {d24, d25, d26, d27}, [SHIFT, :128]! + vshrn.u32 d20, q10, #16 + vshrn.u32 d21, q11, #16 + vshrn.u32 d22, q8, #16 + vshrn.u32 d23, q9, #16 + vneg.s16 q12, q12 + vneg.s16 q13, q13 + vshr.s16 q2, q0, #15 /* extract sign */ + vshr.s16 q3, q1, #15 + vshl.u16 q14, q10, q12 /* shift */ + vshl.u16 q15, q11, q13 + + push {r4, r5} + mov LOOP_COUNT, #3 +1: + vld1.16 {d0, d1, d2, d3}, [WORKSPACE, :128]! + veor.u16 q14, q14, q2 /* restore sign */ + vabs.s16 q12, q0 + vld1.16 {d20, d21, d22, d23}, [CORRECTION, :128]! + vabs.s16 q13, q1 + veor.u16 q15, q15, q3 + vld1.16 {d16, d17, d18, d19}, [RECIPROCAL, :128]! + vadd.u16 q12, q12, q10 /* add correction */ + vadd.u16 q13, q13, q11 + vmull.u16 q10, d24, d16 /* multiply by reciprocal */ + vmull.u16 q11, d25, d17 + vmull.u16 q8, d26, d18 + vmull.u16 q9, d27, d19 + vsub.u16 q14, q14, q2 + vld1.16 {d24, d25, d26, d27}, [SHIFT, :128]! + vsub.u16 q15, q15, q3 + vshrn.u32 d20, q10, #16 + vshrn.u32 d21, q11, #16 + vst1.16 {d28, d29, d30, d31}, [COEF_BLOCK, :128]! + vshrn.u32 d22, q8, #16 + vshrn.u32 d23, q9, #16 + vneg.s16 q12, q12 + vneg.s16 q13, q13 + vshr.s16 q2, q0, #15 /* extract sign */ + vshr.s16 q3, q1, #15 + vshl.u16 q14, q10, q12 /* shift */ + vshl.u16 q15, q11, q13 + subs LOOP_COUNT, LOOP_COUNT, #1 + bne 1b + pop {r4, r5} + + veor.u16 q14, q14, q2 /* restore sign */ + veor.u16 q15, q15, q3 + vsub.u16 q14, q14, q2 + vsub.u16 q15, q15, q3 + vst1.16 {d28, d29, d30, d31}, [COEF_BLOCK, :128]! + + bx lr /* return */ + + .unreq COEF_BLOCK + .unreq DIVISORS + .unreq WORKSPACE + .unreq RECIPROCAL + .unreq CORRECTION + .unreq SHIFT + .unreq LOOP_COUNT +.endfunc + +/*****************************************************************************/ + +/* + * GLOBAL(void) + * jsimd_h2v1_fancy_upsample_neon (int max_v_samp_factor, + * JDIMENSION downsampled_width, + * JSAMPARRAY input_data, + * JSAMPARRAY * output_data_ptr); + * + * Note: the use of unaligned writes is the main remaining bottleneck in + * this code, which can be potentially solved to get up to tens + * of percents performance improvement on Cortex-A8/Cortex-A9. + */ + +/* + * Upsample 16 source pixels to 32 destination pixels. The new 16 source + * pixels are loaded to q0. The previous 16 source pixels are in q1. The + * shifted-by-one source pixels are constructed in q2 by using q0 and q1. + * Register d28 is used for multiplication by 3. Register q15 is used + * for adding +1 bias. + */ +.macro upsample16 OUTPTR, INPTR + vld1.8 {q0}, [\INPTR]! + vmovl.u8 q8, d0 + vext.8 q2, q1, q0, #15 + vmovl.u8 q9, d1 + vaddw.u8 q10, q15, d4 + vaddw.u8 q11, q15, d5 + vmlal.u8 q8, d4, d28 + vmlal.u8 q9, d5, d28 + vmlal.u8 q10, d0, d28 + vmlal.u8 q11, d1, d28 + vmov q1, q0 /* backup source pixels to q1 */ + vrshrn.u16 d6, q8, #2 + vrshrn.u16 d7, q9, #2 + vshrn.u16 d8, q10, #2 + vshrn.u16 d9, q11, #2 + vst2.8 {d6, d7, d8, d9}, [\OUTPTR]! +.endm + +/* + * Upsample 32 source pixels to 64 destination pixels. Compared to 'usample16' + * macro, the roles of q0 and q1 registers are reversed for even and odd + * groups of 16 pixels, that's why "vmov q1, q0" instructions are not needed. + * Also this unrolling allows to reorder loads and stores to compensate + * multiplication latency and reduce stalls. + */ +.macro upsample32 OUTPTR, INPTR + /* even 16 pixels group */ + vld1.8 {q0}, [\INPTR]! + vmovl.u8 q8, d0 + vext.8 q2, q1, q0, #15 + vmovl.u8 q9, d1 + vaddw.u8 q10, q15, d4 + vaddw.u8 q11, q15, d5 + vmlal.u8 q8, d4, d28 + vmlal.u8 q9, d5, d28 + vmlal.u8 q10, d0, d28 + vmlal.u8 q11, d1, d28 + /* odd 16 pixels group */ + vld1.8 {q1}, [\INPTR]! + vrshrn.u16 d6, q8, #2 + vrshrn.u16 d7, q9, #2 + vshrn.u16 d8, q10, #2 + vshrn.u16 d9, q11, #2 + vmovl.u8 q8, d2 + vext.8 q2, q0, q1, #15 + vmovl.u8 q9, d3 + vaddw.u8 q10, q15, d4 + vaddw.u8 q11, q15, d5 + vmlal.u8 q8, d4, d28 + vmlal.u8 q9, d5, d28 + vmlal.u8 q10, d2, d28 + vmlal.u8 q11, d3, d28 + vst2.8 {d6, d7, d8, d9}, [\OUTPTR]! + vrshrn.u16 d6, q8, #2 + vrshrn.u16 d7, q9, #2 + vshrn.u16 d8, q10, #2 + vshrn.u16 d9, q11, #2 + vst2.8 {d6, d7, d8, d9}, [\OUTPTR]! +.endm + +/* + * Upsample a row of WIDTH pixels from INPTR to OUTPTR. + */ +.macro upsample_row OUTPTR, INPTR, WIDTH, TMP1 + /* special case for the first and last pixels */ + sub \WIDTH, \WIDTH, #1 + add \OUTPTR, \OUTPTR, #1 + ldrb \TMP1, [\INPTR, \WIDTH] + strb \TMP1, [\OUTPTR, \WIDTH, asl #1] + ldrb \TMP1, [\INPTR], #1 + strb \TMP1, [\OUTPTR, #-1] + vmov.8 d3[7], \TMP1 + + subs \WIDTH, \WIDTH, #32 + blt 5f +0: /* process 32 pixels per iteration */ + upsample32 \OUTPTR, \INPTR + subs \WIDTH, \WIDTH, #32 + bge 0b +5: + adds \WIDTH, \WIDTH, #16 + blt 1f +0: /* process 16 pixels if needed */ + upsample16 \OUTPTR, \INPTR + subs \WIDTH, \WIDTH, #16 +1: + adds \WIDTH, \WIDTH, #16 + beq 9f + + /* load the remaining 1-15 pixels */ + add \INPTR, \INPTR, \WIDTH + tst \WIDTH, #1 + beq 2f + sub \INPTR, \INPTR, #1 + vld1.8 {d0[0]}, [\INPTR] +2: + tst \WIDTH, #2 + beq 2f + vext.8 d0, d0, d0, #6 + sub \INPTR, \INPTR, #1 + vld1.8 {d0[1]}, [\INPTR] + sub \INPTR, \INPTR, #1 + vld1.8 {d0[0]}, [\INPTR] +2: + tst \WIDTH, #4 + beq 2f + vrev64.32 d0, d0 + sub \INPTR, \INPTR, #1 + vld1.8 {d0[3]}, [\INPTR] + sub \INPTR, \INPTR, #1 + vld1.8 {d0[2]}, [\INPTR] + sub \INPTR, \INPTR, #1 + vld1.8 {d0[1]}, [\INPTR] + sub \INPTR, \INPTR, #1 + vld1.8 {d0[0]}, [\INPTR] +2: + tst \WIDTH, #8 + beq 2f + vmov d1, d0 + sub \INPTR, \INPTR, #8 + vld1.8 {d0}, [\INPTR] +2: /* upsample the remaining pixels */ + vmovl.u8 q8, d0 + vext.8 q2, q1, q0, #15 + vmovl.u8 q9, d1 + vaddw.u8 q10, q15, d4 + vaddw.u8 q11, q15, d5 + vmlal.u8 q8, d4, d28 + vmlal.u8 q9, d5, d28 + vmlal.u8 q10, d0, d28 + vmlal.u8 q11, d1, d28 + vrshrn.u16 d10, q8, #2 + vrshrn.u16 d12, q9, #2 + vshrn.u16 d11, q10, #2 + vshrn.u16 d13, q11, #2 + vzip.8 d10, d11 + vzip.8 d12, d13 + /* store the remaining pixels */ + tst \WIDTH, #8 + beq 2f + vst1.8 {d10, d11}, [\OUTPTR]! + vmov q5, q6 +2: + tst \WIDTH, #4 + beq 2f + vst1.8 {d10}, [\OUTPTR]! + vmov d10, d11 +2: + tst \WIDTH, #2 + beq 2f + vst1.8 {d10[0]}, [\OUTPTR]! + vst1.8 {d10[1]}, [\OUTPTR]! + vst1.8 {d10[2]}, [\OUTPTR]! + vst1.8 {d10[3]}, [\OUTPTR]! + vext.8 d10, d10, d10, #4 +2: + tst \WIDTH, #1 + beq 2f + vst1.8 {d10[0]}, [\OUTPTR]! + vst1.8 {d10[1]}, [\OUTPTR]! +2: +9: +.endm + +asm_function jsimd_h2v1_fancy_upsample_neon + + MAX_V_SAMP_FACTOR .req r0 + DOWNSAMPLED_WIDTH .req r1 + INPUT_DATA .req r2 + OUTPUT_DATA_PTR .req r3 + OUTPUT_DATA .req OUTPUT_DATA_PTR + + OUTPTR .req r4 + INPTR .req r5 + WIDTH .req ip + TMP .req lr + + push {r4, r5, r6, lr} + vpush {d8-d15} + + ldr OUTPUT_DATA, [OUTPUT_DATA_PTR] + cmp MAX_V_SAMP_FACTOR, #0 + ble 99f + + /* initialize constants */ + vmov.u8 d28, #3 + vmov.u16 q15, #1 +11: + ldr INPTR, [INPUT_DATA], #4 + ldr OUTPTR, [OUTPUT_DATA], #4 + mov WIDTH, DOWNSAMPLED_WIDTH + upsample_row OUTPTR, INPTR, WIDTH, TMP + subs MAX_V_SAMP_FACTOR, MAX_V_SAMP_FACTOR, #1 + bgt 11b + +99: + vpop {d8-d15} + pop {r4, r5, r6, pc} + + .unreq MAX_V_SAMP_FACTOR + .unreq DOWNSAMPLED_WIDTH + .unreq INPUT_DATA + .unreq OUTPUT_DATA_PTR + .unreq OUTPUT_DATA + + .unreq OUTPTR + .unreq INPTR + .unreq WIDTH + .unreq TMP + +.endfunc + +.purgem upsample16 +.purgem upsample32 +.purgem upsample_row diff --git a/simd/jsimd_i386.c b/simd/jsimd_i386.c new file mode 100644 index 0000000..e96f5b8 --- /dev/null +++ b/simd/jsimd_i386.c @@ -0,0 +1,1048 @@ +/* + * jsimd_i386.c + * + * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB + * Copyright 2009-2011 D. R. Commander + * + * Based on the 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 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 != ~0U) + 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_rgb_gray (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_gray_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: + case JCS_EXT_RGBA: + 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: + case JCS_EXT_BGRA: + sse2fct=jsimd_extbgrx_ycc_convert_sse2; + mmxfct=jsimd_extbgrx_ycc_convert_mmx; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_extxbgr_ycc_convert_sse2; + mmxfct=jsimd_extxbgr_ycc_convert_mmx; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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_rgb_gray_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_gray_convert_sse2; + mmxfct=jsimd_extrgb_gray_convert_mmx; + break; + case JCS_EXT_RGBX: + case JCS_EXT_RGBA: + sse2fct=jsimd_extrgbx_gray_convert_sse2; + mmxfct=jsimd_extrgbx_gray_convert_mmx; + break; + case JCS_EXT_BGR: + sse2fct=jsimd_extbgr_gray_convert_sse2; + mmxfct=jsimd_extbgr_gray_convert_mmx; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + sse2fct=jsimd_extbgrx_gray_convert_sse2; + mmxfct=jsimd_extbgrx_gray_convert_mmx; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_extxbgr_gray_convert_sse2; + mmxfct=jsimd_extxbgr_gray_convert_mmx; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + sse2fct=jsimd_extxrgb_gray_convert_sse2; + mmxfct=jsimd_extxrgb_gray_convert_mmx; + break; + default: + sse2fct=jsimd_rgb_gray_convert_sse2; + mmxfct=jsimd_rgb_gray_convert_mmx; + break; + } + + if ((simd_support & JSIMD_SSE2) && + IS_ALIGNED_SSE(jconst_rgb_gray_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: + case JCS_EXT_RGBA: + 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: + case JCS_EXT_BGRA: + sse2fct=jsimd_ycc_extbgrx_convert_sse2; + mmxfct=jsimd_ycc_extbgrx_convert_mmx; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_ycc_extxbgr_convert_sse2; + mmxfct=jsimd_ycc_extxbgr_convert_mmx; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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: + case JCS_EXT_RGBA: + 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: + case JCS_EXT_BGRA: + sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2; + mmxfct=jsimd_h2v2_extbgrx_merged_upsample_mmx; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2; + mmxfct=jsimd_h2v2_extxbgr_merged_upsample_mmx; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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: + case JCS_EXT_RGBA: + 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: + case JCS_EXT_BGRA: + sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2; + mmxfct=jsimd_h2v1_extbgrx_merged_upsample_mmx; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2; + mmxfct=jsimd_h2v1_extxbgr_merged_upsample_mmx; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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..8d17db3 --- /dev/null +++ b/simd/jsimd_x86_64.c @@ -0,0 +1,753 @@ +/* + * jsimd_x86_64.c + * + * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB + * Copyright 2009-2011 D. R. Commander + * + * Based on the 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 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_rgb_gray (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_gray_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: + case JCS_EXT_RGBA: + sse2fct=jsimd_extrgbx_ycc_convert_sse2; + break; + case JCS_EXT_BGR: + sse2fct=jsimd_extbgr_ycc_convert_sse2; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + sse2fct=jsimd_extbgrx_ycc_convert_sse2; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_extxbgr_ycc_convert_sse2; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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_rgb_gray_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_gray_convert_sse2; + break; + case JCS_EXT_RGBX: + case JCS_EXT_RGBA: + sse2fct=jsimd_extrgbx_gray_convert_sse2; + break; + case JCS_EXT_BGR: + sse2fct=jsimd_extbgr_gray_convert_sse2; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + sse2fct=jsimd_extbgrx_gray_convert_sse2; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_extxbgr_gray_convert_sse2; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + sse2fct=jsimd_extxrgb_gray_convert_sse2; + break; + default: + sse2fct=jsimd_rgb_gray_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: + case JCS_EXT_RGBA: + sse2fct=jsimd_ycc_extrgbx_convert_sse2; + break; + case JCS_EXT_BGR: + sse2fct=jsimd_ycc_extbgr_convert_sse2; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + sse2fct=jsimd_ycc_extbgrx_convert_sse2; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_ycc_extxbgr_convert_sse2; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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: + case JCS_EXT_RGBA: + sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2; + break; + case JCS_EXT_BGR: + sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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: + case JCS_EXT_RGBA: + sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2; + break; + case JCS_EXT_BGR: + sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2; + break; + case JCS_EXT_BGRX: + case JCS_EXT_BGRA: + sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2; + break; + case JCS_EXT_XBGR: + case JCS_EXT_ABGR: + sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2; + break; + case JCS_EXT_XRGB: + case JCS_EXT_ARGB: + 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..583b7e3 --- /dev/null +++ b/simd/jsimdcfg.inc.h @@ -0,0 +1,196 @@ +// 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" + +; +; -- jpeglib.h +; + +%define _cpp_protection_DCTSIZE DCTSIZE +%define _cpp_protection_DCTSIZE2 DCTSIZE2 + +; +; -- jmorecfg.h +; + +%define _cpp_protection_RGB_RED RGB_RED +%define _cpp_protection_RGB_GREEN RGB_GREEN +%define _cpp_protection_RGB_BLUE RGB_BLUE +%define _cpp_protection_RGB_PIXELSIZE RGB_PIXELSIZE + +%define _cpp_protection_EXT_RGB_RED EXT_RGB_RED +%define _cpp_protection_EXT_RGB_GREEN EXT_RGB_GREEN +%define _cpp_protection_EXT_RGB_BLUE EXT_RGB_BLUE +%define _cpp_protection_EXT_RGB_PIXELSIZE EXT_RGB_PIXELSIZE + +%define _cpp_protection_EXT_RGBX_RED EXT_RGBX_RED +%define _cpp_protection_EXT_RGBX_GREEN EXT_RGBX_GREEN +%define _cpp_protection_EXT_RGBX_BLUE EXT_RGBX_BLUE +%define _cpp_protection_EXT_RGBX_PIXELSIZE EXT_RGBX_PIXELSIZE + +%define _cpp_protection_EXT_BGR_RED EXT_BGR_RED +%define _cpp_protection_EXT_BGR_GREEN EXT_BGR_GREEN +%define _cpp_protection_EXT_BGR_BLUE EXT_BGR_BLUE +%define _cpp_protection_EXT_BGR_PIXELSIZE EXT_BGR_PIXELSIZE + +%define _cpp_protection_EXT_BGRX_RED EXT_BGRX_RED +%define _cpp_protection_EXT_BGRX_GREEN EXT_BGRX_GREEN +%define _cpp_protection_EXT_BGRX_BLUE EXT_BGRX_BLUE +%define _cpp_protection_EXT_BGRX_PIXELSIZE EXT_BGRX_PIXELSIZE + +%define _cpp_protection_EXT_XBGR_RED EXT_XBGR_RED +%define _cpp_protection_EXT_XBGR_GREEN EXT_XBGR_GREEN +%define _cpp_protection_EXT_XBGR_BLUE EXT_XBGR_BLUE +%define _cpp_protection_EXT_XBGR_PIXELSIZE EXT_XBGR_PIXELSIZE + +%define _cpp_protection_EXT_XRGB_RED EXT_XRGB_RED +%define _cpp_protection_EXT_XRGB_GREEN EXT_XRGB_GREEN +%define _cpp_protection_EXT_XRGB_BLUE EXT_XRGB_BLUE +%define _cpp_protection_EXT_XRGB_PIXELSIZE EXT_XRGB_PIXELSIZE + +%define RGBX_FILLER_0XFF 1 + +; 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 _cpp_protection_CENTERJSAMPLE 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 +; + +%define _cpp_protection_JSIMD_NONE JSIMD_NONE +%define _cpp_protection_JSIMD_MMX JSIMD_MMX +%define _cpp_protection_JSIMD_3DNOW JSIMD_3DNOW +%define _cpp_protection_JSIMD_SSE JSIMD_SSE +%define _cpp_protection_JSIMD_SSE2 JSIMD_SSE2 + +; Short forms of external names for systems with brain-damaged linkers. +; +#ifdef NEED_SHORT_EXTERNAL_NAMES +%define _cpp_protection_jpeg_simd_cpu_support jpeg_simd_cpu_support +%define _cpp_protection_jsimd_rgb_ycc_convert_mmx jsimd_rgb_ycc_convert_mmx +%define _cpp_protection_jsimd_ycc_rgb_convert_mmx jsimd_ycc_rgb_convert_mmx +%define _cpp_protection_jconst_rgb_ycc_convert_sse2 jconst_rgb_ycc_convert_sse2 +%define _cpp_protection_jsimd_rgb_ycc_convert_sse2 jsimd_rgb_ycc_convert_sse2 +%define _cpp_protection_jconst_ycc_rgb_convert_sse2 jconst_ycc_rgb_convert_sse2 +%define _cpp_protection_jsimd_ycc_rgb_convert_sse2 jsimd_ycc_rgb_convert_sse2 +%define _cpp_protection_jsimd_h2v2_downsample_mmx jsimd_h2v2_downsample_mmx +%define _cpp_protection_jsimd_h2v1_downsample_mmx jsimd_h2v1_downsample_mmx +%define _cpp_protection_jsimd_h2v2_downsample_sse2 jsimd_h2v2_downsample_sse2 +%define _cpp_protection_jsimd_h2v1_downsample_sse2 jsimd_h2v1_downsample_sse2 +%define _cpp_protection_jsimd_h2v2_upsample_mmx jsimd_h2v2_upsample_mmx +%define _cpp_protection_jsimd_h2v1_upsample_mmx jsimd_h2v1_upsample_mmx +%define _cpp_protection_jsimd_h2v1_fancy_upsample_mmx jsimd_h2v1_fancy_upsample_mmx +%define _cpp_protection_jsimd_h2v2_fancy_upsample_mmx jsimd_h2v2_fancy_upsample_mmx +%define _cpp_protection_jsimd_h2v1_merged_upsample_mmx jsimd_h2v1_merged_upsample_mmx +%define _cpp_protection_jsimd_h2v2_merged_upsample_mmx jsimd_h2v2_merged_upsample_mmx +%define _cpp_protection_jsimd_h2v2_upsample_sse2 jsimd_h2v2_upsample_sse2 +%define _cpp_protection_jsimd_h2v1_upsample_sse2 jsimd_h2v1_upsample_sse2 +%define _cpp_protection_jconst_fancy_upsample_sse2 jconst_fancy_upsample_sse2 +%define _cpp_protection_jsimd_h2v1_fancy_upsample_sse2 jsimd_h2v1_fancy_upsample_sse2 +%define _cpp_protection_jsimd_h2v2_fancy_upsample_sse2 jsimd_h2v2_fancy_upsample_sse2 +%define _cpp_protection_jconst_merged_upsample_sse2 jconst_merged_upsample_sse2 +%define _cpp_protection_jsimd_h2v1_merged_upsample_sse2 jsimd_h2v1_merged_upsample_sse2 +%define _cpp_protection_jsimd_h2v2_merged_upsample_sse2 jsimd_h2v2_merged_upsample_sse2 +%define _cpp_protection_jsimd_convsamp_mmx jsimd_convsamp_mmx +%define _cpp_protection_jsimd_convsamp_sse2 jsimd_convsamp_sse2 +%define _cpp_protection_jsimd_convsamp_float_3dnow jsimd_convsamp_float_3dnow +%define _cpp_protection_jsimd_convsamp_float_sse jsimd_convsamp_float_sse +%define _cpp_protection_jsimd_convsamp_float_sse2 jsimd_convsamp_float_sse2 +%define _cpp_protection_jsimd_fdct_islow_mmx jsimd_fdct_islow_mmx +%define _cpp_protection_jsimd_fdct_ifast_mmx jsimd_fdct_ifast_mmx +%define _cpp_protection_jconst_fdct_islow_sse2 jconst_fdct_islow_sse2 +%define _cpp_protection_jsimd_fdct_islow_sse2 jsimd_fdct_islow_sse2 +%define _cpp_protection_jconst_fdct_ifast_sse2 jconst_fdct_ifast_sse2 +%define _cpp_protection_jsimd_fdct_ifast_sse2 jsimd_fdct_ifast_sse2 +%define _cpp_protection_jsimd_fdct_float_3dnow jsimd_fdct_float_3dnow +%define _cpp_protection_jconst_fdct_float_sse jconst_fdct_float_sse +%define _cpp_protection_jsimd_fdct_float_sse jsimd_fdct_float_sse +%define _cpp_protection_jsimd_quantize_mmx jsimd_quantize_mmx +%define _cpp_protection_jsimd_quantize_sse2 jsimd_quantize_sse2 +%define _cpp_protection_jsimd_quantize_float_3dnow jsimd_quantize_float_3dnow +%define _cpp_protection_jsimd_quantize_float_sse jsimd_quantize_float_sse +%define _cpp_protection_jsimd_quantize_float_sse2 jsimd_quantize_float_sse2 +%define _cpp_protection_jsimd_idct_2x2_mmx jsimd_idct_2x2_mmx +%define _cpp_protection_jsimd_idct_4x4_mmx jsimd_idct_4x4_mmx +%define _cpp_protection_jconst_idct_red_sse2 jconst_idct_red_sse2 +%define _cpp_protection_jsimd_idct_2x2_sse2 jsimd_idct_2x2_sse2 +%define _cpp_protection_jsimd_idct_4x4_sse2 jsimd_idct_4x4_sse2 +%define _cpp_protection_jsimd_idct_islow_mmx jsimd_idct_islow_mmx +%define _cpp_protection_jsimd_idct_ifast_mmx jsimd_idct_ifast_mmx +%define _cpp_protection_jconst_idct_islow_sse2 jconst_idct_islow_sse2 +%define _cpp_protection_jsimd_idct_islow_sse2 jsimd_idct_islow_sse2 +%define _cpp_protection_jconst_idct_ifast_sse2 jconst_idct_ifast_sse2 +%define _cpp_protection_jsimd_idct_ifast_sse2 jsimd_idct_ifast_sse2 +%define _cpp_protection_jsimd_idct_float_3dnow jsimd_idct_float_3dnow +%define _cpp_protection_jconst_idct_float_sse jconst_idct_float_sse +%define _cpp_protection_jsimd_idct_float_sse jsimd_idct_float_sse +%define _cpp_protection_jconst_idct_float_sse2 jconst_idct_float_sse2 +%define _cpp_protection_jsimd_idct_float_sse2 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..253b897 --- /dev/null +++ b/simd/jsimdext.inc @@ -0,0 +1,376 @@ +; +; 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 -- +; +%ifdef __YASM_VER__ +%define SEG_TEXT .text align=16 +%define SEG_CONST .rdata align=16 +%else +%define SEG_TEXT .text align=16 public use32 class=CODE +%define SEG_CONST .rdata align=16 public use32 class=CONST +%endif + +%elifdef WIN64 ; ----(nasm -fwin64 -DWIN64 ...)-------- +; * Microsoft Visual C++ + +; -- segment definition -- +; +%ifdef __YASM_VER__ +%define SEG_TEXT .text align=16 +%define SEG_CONST .rdata align=16 +%else +%define SEG_TEXT .text align=16 public use64 class=CODE +%define SEG_CONST .rdata align=16 public use64 class=CONST +%endif +%define EXTN(name) name ; foo() -> foo + +%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 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 + movaps XMMWORD [rsp], xmm6 + sub rsp, SIZEOF_XMMWORD + movaps XMMWORD [rsp], xmm7 +%endmacro + +%imacro uncollect_args 0 + movaps xmm7, XMMWORD [rsp] + add rsp, SIZEOF_XMMWORD + movaps xmm6, XMMWORD [rsp] + add rsp, SIZEOF_XMMWORD + pop rdi + pop rsi + pop r15 + pop r14 + pop r13 + pop r12 +%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 |