| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* pg_crc32c_sse42.c |
| 4 |
* Compute CRC-32C checksum using Intel SSE 4.2 or AVX-512 instructions. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/port/pg_crc32c_sse42.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
#include "c.h" |
| 16 |
|
| 17 |
#include <nmmintrin.h> |
| 18 |
#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK |
| 19 |
#include <immintrin.h> |
| 20 |
#endif |
| 21 |
|
| 22 |
#include "port/pg_crc32c.h" |
| 23 |
|
| 24 |
pg_attribute_no_sanitize_alignment() |
| 25 |
pg_attribute_target("sse4.2") |
| 26 |
pg_crc32c |
| 27 |
pg_comp_crc32c_sse42(pg_crc32c crc, const void *data, size_t len) |
| 28 |
{ |
| 29 |
const unsigned char *p = data; |
| 30 |
const unsigned char *pend = p + len; |
| 31 |
|
| 32 |
/* |
| 33 |
* Process eight bytes of data at a time. |
| 34 |
* |
| 35 |
* NB: We do unaligned accesses here. The Intel architecture allows that, |
| 36 |
* and performance testing didn't show any performance gain from aligning |
| 37 |
* the begin address. |
| 38 |
*/ |
| 39 |
#ifdef __x86_64__ |
| 40 |
while (p + 8 <= pend) |
| 41 |
{ |
| 42 |
crc = (uint32) _mm_crc32_u64(crc, *((const uint64 *) p)); |
| 43 |
p += 8; |
| 44 |
} |
| 45 |
|
| 46 |
/* Process remaining full four bytes if any */ |
| 47 |
if (p + 4 <= pend) |
| 48 |
{ |
| 49 |
crc = _mm_crc32_u32(crc, *((const unsigned int *) p)); |
| 50 |
p += 4; |
| 51 |
} |
| 52 |
#else |
| 53 |
|
| 54 |
/* |
| 55 |
* Process four bytes at a time. (The eight byte instruction is not |
| 56 |
* available on the 32-bit x86 architecture). |
| 57 |
*/ |
| 58 |
while (p + 4 <= pend) |
| 59 |
{ |
| 60 |
crc = _mm_crc32_u32(crc, *((const unsigned int *) p)); |
| 61 |
p += 4; |
| 62 |
} |
| 63 |
#endif /* __x86_64__ */ |
| 64 |
|
| 65 |
/* Process any remaining bytes one at a time. */ |
| 66 |
while (p < pend) |
| 67 |
{ |
| 68 |
crc = _mm_crc32_u8(crc, *p); |
| 69 |
p++; |
| 70 |
} |
| 71 |
|
| 72 |
return crc; |
| 73 |
} |
| 74 |
|
| 75 |
#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK |
| 76 |
|
| 77 |
/* |
| 78 |
* Note: There is no copyright notice in the following generated code. |
| 79 |
* |
| 80 |
* We have modified the output to |
| 81 |
* - match our function declaration |
| 82 |
* - match whitespace to our project style |
| 83 |
* - add a threshold for the alignment stanza |
| 84 |
*/ |
| 85 |
|
| 86 |
/* Generated by https://github.com/corsix/fast-crc32/ using: */ |
| 87 |
/* ./generate -i avx512_vpclmulqdq -p crc32c -a v1e */ |
| 88 |
/* MIT licensed */ |
| 89 |
|
| 90 |
#define clmul_lo(a, b) (_mm512_clmulepi64_epi128((a), (b), 0)) |
| 91 |
#define clmul_hi(a, b) (_mm512_clmulepi64_epi128((a), (b), 17)) |
| 92 |
|
| 93 |
pg_attribute_target("vpclmulqdq,avx512vl") |
| 94 |
pg_crc32c |
| 95 |
pg_comp_crc32c_avx512(pg_crc32c crc, const void *data, size_t len) |
| 96 |
{ |
| 97 |
/* adjust names to match generated code */ |
| 98 |
pg_crc32c crc0 = crc; |
| 99 |
const char *buf = data; |
| 100 |
|
| 101 |
/* Align on cacheline boundary. The threshold is somewhat arbitrary. */ |
| 102 |
if (unlikely(len > 256)) |
| 103 |
{ |
| 104 |
for (; len && ((uintptr_t) buf & 7); --len) |
| 105 |
crc0 = _mm_crc32_u8(crc0, *buf++); |
| 106 |
while (((uintptr_t) buf & 56) && len >= 8) |
| 107 |
{ |
| 108 |
crc0 = _mm_crc32_u64(crc0, *(const uint64_t *) buf); |
| 109 |
buf += 8; |
| 110 |
len -= 8; |
| 111 |
} |
| 112 |
} |
| 113 |
|
| 114 |
if (len >= 64) |
| 115 |
{ |
| 116 |
const char *end = buf + len; |
| 117 |
const char *limit = buf + len - 64; |
| 118 |
__m128i z0; |
| 119 |
|
| 120 |
/* First vector chunk. */ |
| 121 |
__m512i x0 = _mm512_loadu_si512((const void *) buf), |
| 122 |
y0; |
| 123 |
__m512i k; |
| 124 |
|
| 125 |
k = _mm512_broadcast_i32x4(_mm_setr_epi32(0x740eef02, 0, 0x9e4addf8, 0)); |
| 126 |
x0 = _mm512_xor_si512(_mm512_zextsi128_si512(_mm_cvtsi32_si128(crc0)), x0); |
| 127 |
buf += 64; |
| 128 |
|
| 129 |
/* Main loop. */ |
| 130 |
while (buf <= limit) |
| 131 |
{ |
| 132 |
y0 = clmul_lo(x0, k), x0 = clmul_hi(x0, k); |
| 133 |
x0 = _mm512_ternarylogic_epi64(x0, y0, |
| 134 |
_mm512_loadu_si512((const void *) buf), |
| 135 |
0x96); |
| 136 |
buf += 64; |
| 137 |
} |
| 138 |
|
| 139 |
/* Reduce 512 bits to 128 bits. */ |
| 140 |
k = _mm512_setr_epi32(0x1c291d04, 0, 0xddc0152b, 0, |
| 141 |
0x3da6d0cb, 0, 0xba4fc28e, 0, |
| 142 |
0xf20c0dfe, 0, 0x493c7d27, 0, |
| 143 |
0, 0, 0, 0); |
| 144 |
y0 = clmul_lo(x0, k), k = clmul_hi(x0, k); |
| 145 |
y0 = _mm512_xor_si512(y0, k); |
| 146 |
z0 = _mm_ternarylogic_epi64(_mm512_castsi512_si128(y0), |
| 147 |
_mm512_extracti32x4_epi32(y0, 1), |
| 148 |
_mm512_extracti32x4_epi32(y0, 2), |
| 149 |
0x96); |
| 150 |
z0 = _mm_xor_si128(z0, _mm512_extracti32x4_epi32(x0, 3)); |
| 151 |
|
| 152 |
/* Reduce 128 bits to 32 bits, and multiply by x^32. */ |
| 153 |
crc0 = _mm_crc32_u64(0, _mm_extract_epi64(z0, 0)); |
| 154 |
crc0 = _mm_crc32_u64(crc0, _mm_extract_epi64(z0, 1)); |
| 155 |
len = end - buf; |
| 156 |
} |
| 157 |
|
| 158 |
return pg_comp_crc32c_sse42(crc0, buf, len); |
| 159 |
} |
| 160 |
|
| 161 |
#endif |