Bob Pearson | fbedceb | 2012-03-23 15:02:22 -0700 | [diff] [blame] | 1 | A brief CRC tutorial. |
| 2 | |
| 3 | A CRC is a long-division remainder. You add the CRC to the message, |
| 4 | and the whole thing (message+CRC) is a multiple of the given |
| 5 | CRC polynomial. To check the CRC, you can either check that the |
| 6 | CRC matches the recomputed value, *or* you can check that the |
| 7 | remainder computed on the message+CRC is 0. This latter approach |
| 8 | is used by a lot of hardware implementations, and is why so many |
| 9 | protocols put the end-of-frame flag after the CRC. |
| 10 | |
| 11 | It's actually the same long division you learned in school, except that |
| 12 | - We're working in binary, so the digits are only 0 and 1, and |
| 13 | - When dividing polynomials, there are no carries. Rather than add and |
| 14 | subtract, we just xor. Thus, we tend to get a bit sloppy about |
| 15 | the difference between adding and subtracting. |
| 16 | |
| 17 | Like all division, the remainder is always smaller than the divisor. |
| 18 | To produce a 32-bit CRC, the divisor is actually a 33-bit CRC polynomial. |
| 19 | Since it's 33 bits long, bit 32 is always going to be set, so usually the |
| 20 | CRC is written in hex with the most significant bit omitted. (If you're |
| 21 | familiar with the IEEE 754 floating-point format, it's the same idea.) |
| 22 | |
| 23 | Note that a CRC is computed over a string of *bits*, so you have |
| 24 | to decide on the endianness of the bits within each byte. To get |
| 25 | the best error-detecting properties, this should correspond to the |
| 26 | order they're actually sent. For example, standard RS-232 serial is |
| 27 | little-endian; the most significant bit (sometimes used for parity) |
| 28 | is sent last. And when appending a CRC word to a message, you should |
| 29 | do it in the right order, matching the endianness. |
| 30 | |
| 31 | Just like with ordinary division, you proceed one digit (bit) at a time. |
| 32 | Each step of the division you take one more digit (bit) of the dividend |
| 33 | and append it to the current remainder. Then you figure out the |
| 34 | appropriate multiple of the divisor to subtract to being the remainder |
| 35 | back into range. In binary, this is easy - it has to be either 0 or 1, |
| 36 | and to make the XOR cancel, it's just a copy of bit 32 of the remainder. |
| 37 | |
| 38 | When computing a CRC, we don't care about the quotient, so we can |
| 39 | throw the quotient bit away, but subtract the appropriate multiple of |
| 40 | the polynomial from the remainder and we're back to where we started, |
| 41 | ready to process the next bit. |
| 42 | |
| 43 | A big-endian CRC written this way would be coded like: |
| 44 | for (i = 0; i < input_bits; i++) { |
| 45 | multiple = remainder & 0x80000000 ? CRCPOLY : 0; |
| 46 | remainder = (remainder << 1 | next_input_bit()) ^ multiple; |
| 47 | } |
| 48 | |
| 49 | Notice how, to get at bit 32 of the shifted remainder, we look |
| 50 | at bit 31 of the remainder *before* shifting it. |
| 51 | |
| 52 | But also notice how the next_input_bit() bits we're shifting into |
| 53 | the remainder don't actually affect any decision-making until |
| 54 | 32 bits later. Thus, the first 32 cycles of this are pretty boring. |
| 55 | Also, to add the CRC to a message, we need a 32-bit-long hole for it at |
| 56 | the end, so we have to add 32 extra cycles shifting in zeros at the |
| 57 | end of every message, |
| 58 | |
| 59 | These details lead to a standard trick: rearrange merging in the |
| 60 | next_input_bit() until the moment it's needed. Then the first 32 cycles |
| 61 | can be precomputed, and merging in the final 32 zero bits to make room |
| 62 | for the CRC can be skipped entirely. This changes the code to: |
| 63 | |
| 64 | for (i = 0; i < input_bits; i++) { |
| 65 | remainder ^= next_input_bit() << 31; |
| 66 | multiple = (remainder & 0x80000000) ? CRCPOLY : 0; |
| 67 | remainder = (remainder << 1) ^ multiple; |
| 68 | } |
| 69 | |
| 70 | With this optimization, the little-endian code is particularly simple: |
| 71 | for (i = 0; i < input_bits; i++) { |
| 72 | remainder ^= next_input_bit(); |
| 73 | multiple = (remainder & 1) ? CRCPOLY : 0; |
| 74 | remainder = (remainder >> 1) ^ multiple; |
| 75 | } |
| 76 | |
| 77 | The most significant coefficient of the remainder polynomial is stored |
| 78 | in the least significant bit of the binary "remainder" variable. |
| 79 | The other details of endianness have been hidden in CRCPOLY (which must |
| 80 | be bit-reversed) and next_input_bit(). |
| 81 | |
| 82 | As long as next_input_bit is returning the bits in a sensible order, we don't |
| 83 | *have* to wait until the last possible moment to merge in additional bits. |
| 84 | We can do it 8 bits at a time rather than 1 bit at a time: |
| 85 | for (i = 0; i < input_bytes; i++) { |
| 86 | remainder ^= next_input_byte() << 24; |
| 87 | for (j = 0; j < 8; j++) { |
| 88 | multiple = (remainder & 0x80000000) ? CRCPOLY : 0; |
| 89 | remainder = (remainder << 1) ^ multiple; |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | Or in little-endian: |
| 94 | for (i = 0; i < input_bytes; i++) { |
| 95 | remainder ^= next_input_byte(); |
| 96 | for (j = 0; j < 8; j++) { |
| 97 | multiple = (remainder & 1) ? CRCPOLY : 0; |
| 98 | remainder = (remainder >> 1) ^ multiple; |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | If the input is a multiple of 32 bits, you can even XOR in a 32-bit |
| 103 | word at a time and increase the inner loop count to 32. |
| 104 | |
| 105 | You can also mix and match the two loop styles, for example doing the |
| 106 | bulk of a message byte-at-a-time and adding bit-at-a-time processing |
| 107 | for any fractional bytes at the end. |
| 108 | |
| 109 | To reduce the number of conditional branches, software commonly uses |
| 110 | the byte-at-a-time table method, popularized by Dilip V. Sarwate, |
| 111 | "Computation of Cyclic Redundancy Checks via Table Look-Up", Comm. ACM |
| 112 | v.31 no.8 (August 1998) p. 1008-1013. |
| 113 | |
| 114 | Here, rather than just shifting one bit of the remainder to decide |
| 115 | in the correct multiple to subtract, we can shift a byte at a time. |
| 116 | This produces a 40-bit (rather than a 33-bit) intermediate remainder, |
| 117 | and the correct multiple of the polynomial to subtract is found using |
| 118 | a 256-entry lookup table indexed by the high 8 bits. |
| 119 | |
| 120 | (The table entries are simply the CRC-32 of the given one-byte messages.) |
| 121 | |
| 122 | When space is more constrained, smaller tables can be used, e.g. two |
| 123 | 4-bit shifts followed by a lookup in a 16-entry table. |
| 124 | |
| 125 | It is not practical to process much more than 8 bits at a time using this |
| 126 | technique, because tables larger than 256 entries use too much memory and, |
| 127 | more importantly, too much of the L1 cache. |
| 128 | |
| 129 | To get higher software performance, a "slicing" technique can be used. |
| 130 | See "High Octane CRC Generation with the Intel Slicing-by-8 Algorithm", |
| 131 | ftp://download.intel.com/technology/comms/perfnet/download/slicing-by-8.pdf |
| 132 | |
| 133 | This does not change the number of table lookups, but does increase |
| 134 | the parallelism. With the classic Sarwate algorithm, each table lookup |
| 135 | must be completed before the index of the next can be computed. |
| 136 | |
| 137 | A "slicing by 2" technique would shift the remainder 16 bits at a time, |
| 138 | producing a 48-bit intermediate remainder. Rather than doing a single |
| 139 | lookup in a 65536-entry table, the two high bytes are looked up in |
| 140 | two different 256-entry tables. Each contains the remainder required |
| 141 | to cancel out the corresponding byte. The tables are different because the |
| 142 | polynomials to cancel are different. One has non-zero coefficients from |
| 143 | x^32 to x^39, while the other goes from x^40 to x^47. |
| 144 | |
| 145 | Since modern processors can handle many parallel memory operations, this |
| 146 | takes barely longer than a single table look-up and thus performs almost |
| 147 | twice as fast as the basic Sarwate algorithm. |
| 148 | |
| 149 | This can be extended to "slicing by 4" using 4 256-entry tables. |
| 150 | Each step, 32 bits of data is fetched, XORed with the CRC, and the result |
| 151 | broken into bytes and looked up in the tables. Because the 32-bit shift |
| 152 | leaves the low-order bits of the intermediate remainder zero, the |
| 153 | final CRC is simply the XOR of the 4 table look-ups. |
| 154 | |
| 155 | But this still enforces sequential execution: a second group of table |
| 156 | look-ups cannot begin until the previous groups 4 table look-ups have all |
| 157 | been completed. Thus, the processor's load/store unit is sometimes idle. |
| 158 | |
| 159 | To make maximum use of the processor, "slicing by 8" performs 8 look-ups |
| 160 | in parallel. Each step, the 32-bit CRC is shifted 64 bits and XORed |
| 161 | with 64 bits of input data. What is important to note is that 4 of |
| 162 | those 8 bytes are simply copies of the input data; they do not depend |
| 163 | on the previous CRC at all. Thus, those 4 table look-ups may commence |
| 164 | immediately, without waiting for the previous loop iteration. |
| 165 | |
| 166 | By always having 4 loads in flight, a modern superscalar processor can |
| 167 | be kept busy and make full use of its L1 cache. |
| 168 | |
| 169 | Two more details about CRC implementation in the real world: |
| 170 | |
| 171 | Normally, appending zero bits to a message which is already a multiple |
| 172 | of a polynomial produces a larger multiple of that polynomial. Thus, |
| 173 | a basic CRC will not detect appended zero bits (or bytes). To enable |
| 174 | a CRC to detect this condition, it's common to invert the CRC before |
| 175 | appending it. This makes the remainder of the message+crc come out not |
| 176 | as zero, but some fixed non-zero value. (The CRC of the inversion |
| 177 | pattern, 0xffffffff.) |
| 178 | |
| 179 | The same problem applies to zero bits prepended to the message, and a |
| 180 | similar solution is used. Instead of starting the CRC computation with |
| 181 | a remainder of 0, an initial remainder of all ones is used. As long as |
| 182 | you start the same way on decoding, it doesn't make a difference. |