How to implement CRC32 taking advantage of Intel specific instructions?

c crc32

RubenLaguna · Jul 2, 2015 · Viewed 11.2k times · Source

So I have a design which incorporates CRC32C checksums to ensure data hasn't been damaged. I decided to use CRC32C because I can have both a software version and a hardware-accelerated version if the computer the software runs on supports SSE 4.2

I'm going by Intel's developer manual (vol 2A), which seems to provide the algorithm behind the crc32 instruction. However, I'm having little luck. Intel's developer guide says the following:

BIT_REFLECT32: DEST[31-0] = SRC[0-31]
MOD2: Remainder from Polynomial division modulus 2

TEMP1[31-0] <- BIT_REFLECT(SRC[31-0])
TEMP2[31-0] <- BIT_REFLECT(DEST[31-0])
TEMP3[63-0] <- TEMP1[31-0] << 32
TEMP4[63-0] <- TEMP2[31-0] << 32
TEMP5[63-0] <- TEMP3[63-0] XOR TEMP4[63-0]
TEMP6[31-0] <- TEMP5[63-0] MOD2 0x11EDC6F41
DEST[31-0]  <- BIT_REFLECT(TEMP6[31-0])

Now, as far as I can tell, I've done everything up to the line starting TEMP6 correctly, but I think I may be either misunderstanding the polynomial division, or implementing it incorrectly. If my understanding is correct, 1 / 1 mod 2 = 1, 0 / 1 mod 2 = 0, and both divides-by-zero are undefined.

What I don't understand is how binary division with 64-bit and 33-bit operands will work. If SRC is 0x00000000, and DEST is 0xFFFFFFFF, TEMP5[63-32] will be all set bits, while TEMP5[31-0] will be all unset bits.

If I was to use the bits from TEMP5 as the numerator, there would be 30 divisions by zero as the polynomial 11EDC6F41 is only 33 bits long (and so converting it to a 64-bit unsigned integer leaves the top 30 bits unset), and so the denominator is unset for 30 bits.

However, if I was to use the polynomial as the numerator, the bottom 32 bits of TEMP5 are unset, resulting in divides by zero there, and the top 30 bits of the result would be zero, as the top 30 bits of the numerator would be zero, as 0 / 1 mod 2 = 0.

Am I misunderstanding how this works? Just plain missing something? Or has Intel left out some crucial step in their documentation?

The reason I went to Intel's developer guide for what appeared to be the algorithm they used is because they used a 33-bit polynomial, and I wanted to make outputs identical, which didn't happen when I used the 32-bit polynomial 1EDC6F41 (show below).

uint32_t poly = 0x1EDC6F41, sres, crcTable[256], data = 0x00000000;

for (n = 0; n < 256; n++) {
    sres = n;
    for (k = 0; k < 8; k++)
        sres = (sres & 1) == 1 ? poly ^ (sres >> 1) : (sres >> 1);
    crcTable[n] = sres;
}
sres = 0xFFFFFFFF;

for (n = 0; n < 4; n++) {
    sres = crcTable[(sres ^ data) & 0xFF] ^ (sres >> 8);
}

The above code produces 4138093821 as an output, and the crc32 opcode produces 2346497208 using the input 0x00000000.

Sorry if this is badly written or incomprehensible in places, it is rather late for me.

How to implement CRC32 taking advantage of Intel specific instructions?

Answer

Related questions