1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182

A brief CRC tutorial.
A CRC is a longdivision remainder. You add the CRC to the message,
and the whole thing (message+CRC) is a multiple of the given
CRC polynomial. To check the CRC, you can either check that the
CRC matches the recomputed value, *or* you can check that the
remainder computed on the message+CRC is 0. This latter approach
is used by a lot of hardware implementations, and is why so many
protocols put the endofframe flag after the CRC.
It's actually the same long division you learned in school, except that
 We're working in binary, so the digits are only 0 and 1, and
 When dividing polynomials, there are no carries. Rather than add and
subtract, we just xor. Thus, we tend to get a bit sloppy about
the difference between adding and subtracting.
Like all division, the remainder is always smaller than the divisor.
To produce a 32bit CRC, the divisor is actually a 33bit CRC polynomial.
Since it's 33 bits long, bit 32 is always going to be set, so usually the
CRC is written in hex with the most significant bit omitted. (If you're
familiar with the IEEE 754 floatingpoint format, it's the same idea.)
Note that a CRC is computed over a string of *bits*, so you have
to decide on the endianness of the bits within each byte. To get
the best errordetecting properties, this should correspond to the
order they're actually sent. For example, standard RS232 serial is
littleendian; the most significant bit (sometimes used for parity)
is sent last. And when appending a CRC word to a message, you should
do it in the right order, matching the endianness.
Just like with ordinary division, you proceed one digit (bit) at a time.
Each step of the division you take one more digit (bit) of the dividend
and append it to the current remainder. Then you figure out the
appropriate multiple of the divisor to subtract to being the remainder
back into range. In binary, this is easy  it has to be either 0 or 1,
and to make the XOR cancel, it's just a copy of bit 32 of the remainder.
When computing a CRC, we don't care about the quotient, so we can
throw the quotient bit away, but subtract the appropriate multiple of
the polynomial from the remainder and we're back to where we started,
ready to process the next bit.
A bigendian CRC written this way would be coded like:
for (i = 0; i < input_bits; i++) {
multiple = remainder & 0x80000000 ? CRCPOLY : 0;
remainder = (remainder << 1  next_input_bit()) ^ multiple;
}
Notice how, to get at bit 32 of the shifted remainder, we look
at bit 31 of the remainder *before* shifting it.
But also notice how the next_input_bit() bits we're shifting into
the remainder don't actually affect any decisionmaking until
32 bits later. Thus, the first 32 cycles of this are pretty boring.
Also, to add the CRC to a message, we need a 32bitlong hole for it at
the end, so we have to add 32 extra cycles shifting in zeros at the
end of every message,
These details lead to a standard trick: rearrange merging in the
next_input_bit() until the moment it's needed. Then the first 32 cycles
can be precomputed, and merging in the final 32 zero bits to make room
for the CRC can be skipped entirely. This changes the code to:
for (i = 0; i < input_bits; i++) {
remainder ^= next_input_bit() << 31;
multiple = (remainder & 0x80000000) ? CRCPOLY : 0;
remainder = (remainder << 1) ^ multiple;
}
With this optimization, the littleendian code is particularly simple:
for (i = 0; i < input_bits; i++) {
remainder ^= next_input_bit();
multiple = (remainder & 1) ? CRCPOLY : 0;
remainder = (remainder >> 1) ^ multiple;
}
The most significant coefficient of the remainder polynomial is stored
in the least significant bit of the binary "remainder" variable.
The other details of endianness have been hidden in CRCPOLY (which must
be bitreversed) and next_input_bit().
As long as next_input_bit is returning the bits in a sensible order, we don't
*have* to wait until the last possible moment to merge in additional bits.
We can do it 8 bits at a time rather than 1 bit at a time:
for (i = 0; i < input_bytes; i++) {
remainder ^= next_input_byte() << 24;
for (j = 0; j < 8; j++) {
multiple = (remainder & 0x80000000) ? CRCPOLY : 0;
remainder = (remainder << 1) ^ multiple;
}
}
Or in littleendian:
for (i = 0; i < input_bytes; i++) {
remainder ^= next_input_byte();
for (j = 0; j < 8; j++) {
multiple = (remainder & 1) ? CRCPOLY : 0;
remainder = (remainder >> 1) ^ multiple;
}
}
If the input is a multiple of 32 bits, you can even XOR in a 32bit
word at a time and increase the inner loop count to 32.
You can also mix and match the two loop styles, for example doing the
bulk of a message byteatatime and adding bitatatime processing
for any fractional bytes at the end.
To reduce the number of conditional branches, software commonly uses
the byteatatime table method, popularized by Dilip V. Sarwate,
"Computation of Cyclic Redundancy Checks via Table LookUp", Comm. ACM
v.31 no.8 (August 1998) p. 10081013.
Here, rather than just shifting one bit of the remainder to decide
in the correct multiple to subtract, we can shift a byte at a time.
This produces a 40bit (rather than a 33bit) intermediate remainder,
and the correct multiple of the polynomial to subtract is found using
a 256entry lookup table indexed by the high 8 bits.
(The table entries are simply the CRC32 of the given onebyte messages.)
When space is more constrained, smaller tables can be used, e.g. two
4bit shifts followed by a lookup in a 16entry table.
It is not practical to process much more than 8 bits at a time using this
technique, because tables larger than 256 entries use too much memory and,
more importantly, too much of the L1 cache.
To get higher software performance, a "slicing" technique can be used.
See "High Octane CRC Generation with the Intel Slicingby8 Algorithm",
ftp://download.intel.com/technology/comms/perfnet/download/slicingby8.pdf
This does not change the number of table lookups, but does increase
the parallelism. With the classic Sarwate algorithm, each table lookup
must be completed before the index of the next can be computed.
A "slicing by 2" technique would shift the remainder 16 bits at a time,
producing a 48bit intermediate remainder. Rather than doing a single
lookup in a 65536entry table, the two high bytes are looked up in
two different 256entry tables. Each contains the remainder required
to cancel out the corresponding byte. The tables are different because the
polynomials to cancel are different. One has nonzero coefficients from
x^32 to x^39, while the other goes from x^40 to x^47.
Since modern processors can handle many parallel memory operations, this
takes barely longer than a single table lookup and thus performs almost
twice as fast as the basic Sarwate algorithm.
This can be extended to "slicing by 4" using 4 256entry tables.
Each step, 32 bits of data is fetched, XORed with the CRC, and the result
broken into bytes and looked up in the tables. Because the 32bit shift
leaves the loworder bits of the intermediate remainder zero, the
final CRC is simply the XOR of the 4 table lookups.
But this still enforces sequential execution: a second group of table
lookups cannot begin until the previous groups 4 table lookups have all
been completed. Thus, the processor's load/store unit is sometimes idle.
To make maximum use of the processor, "slicing by 8" performs 8 lookups
in parallel. Each step, the 32bit CRC is shifted 64 bits and XORed
with 64 bits of input data. What is important to note is that 4 of
those 8 bytes are simply copies of the input data; they do not depend
on the previous CRC at all. Thus, those 4 table lookups may commence
immediately, without waiting for the previous loop iteration.
By always having 4 loads in flight, a modern superscalar processor can
be kept busy and make full use of its L1 cache.
Two more details about CRC implementation in the real world:
Normally, appending zero bits to a message which is already a multiple
of a polynomial produces a larger multiple of that polynomial. Thus,
a basic CRC will not detect appended zero bits (or bytes). To enable
a CRC to detect this condition, it's common to invert the CRC before
appending it. This makes the remainder of the message+crc come out not
as zero, but some fixed nonzero value. (The CRC of the inversion
pattern, 0xffffffff.)
The same problem applies to zero bits prepended to the message, and a
similar solution is used. Instead of starting the CRC computation with
a remainder of 0, an initial remainder of all ones is used. As long as
you start the same way on decoding, it doesn't make a difference.
