James Kuszmaul | 6439136 | 2021-01-17 11:32:00 -0800 | [diff] [blame^] | 1 | /* |
| 2 | * crc32c.c -- compute CRC-32C using the Intel crc32 instruction |
| 3 | * Copyright (C) 2013 Mark Adler |
| 4 | * Version 1.1 1 Aug 2013 Mark Adler |
| 5 | * |
| 6 | * Use hardware CRC instruction on Intel SSE 4.2 processors. This computes a |
| 7 | * CRC-32C, *not* the CRC-32 used by Ethernet and zip, gzip, etc. A software |
| 8 | * version is provided as a fall-back, as well as for speed comparisons. |
| 9 | * |
| 10 | * (Modified for the use in RAWRTC. Software fallback has been stripped.) |
| 11 | * |
| 12 | * |
| 13 | * |
| 14 | * This software is provided 'as-is', without any express or implied |
| 15 | * warranty. In no event will the author be held liable for any damages |
| 16 | * arising from the use of this software. |
| 17 | * |
| 18 | * Permission is granted to anyone to use this software for any purpose, |
| 19 | * including commercial applications, and to alter it and redistribute it |
| 20 | * freely, subject to the following restrictions: |
| 21 | * |
| 22 | * 1. The origin of this software must not be misrepresented; you must not |
| 23 | * claim that you wrote the original software. If you use this software |
| 24 | * in a product, an acknowledgment in the product documentation would be |
| 25 | * appreciated but is not required. |
| 26 | * 2. Altered source versions must be plainly marked as such, and must not be |
| 27 | * misrepresented as being the original software. |
| 28 | * 3. This notice may not be removed or altered from any source distribution. |
| 29 | * |
| 30 | * Mark Adler |
| 31 | * madler@alumni.caltech.edu |
| 32 | */ |
| 33 | #include "sse42.h" |
| 34 | #include <re.h> |
| 35 | |
| 36 | /* CRC-32C (iSCSI) polynomial in reversed bit order. */ |
| 37 | #define POLY 0x82f63b78 |
| 38 | |
| 39 | /* Multiply a matrix times a vector over the Galois field of two elements, |
| 40 | GF(2). Each element is a bit in an unsigned integer. mat must have at |
| 41 | least as many entries as the power of two for most significant one bit in |
| 42 | vec. */ |
| 43 | static inline uint32_t gf2_matrix_times(uint32_t* mat, uint32_t vec) { |
| 44 | uint32_t sum; |
| 45 | |
| 46 | sum = 0; |
| 47 | while (vec) { |
| 48 | if (vec & 1) { |
| 49 | sum ^= *mat; |
| 50 | } |
| 51 | vec >>= 1; |
| 52 | mat++; |
| 53 | } |
| 54 | return sum; |
| 55 | } |
| 56 | |
| 57 | /* Multiply a matrix by itself over GF(2). Both mat and square must have 32 |
| 58 | rows. */ |
| 59 | static inline void gf2_matrix_square(uint32_t* square, uint32_t* mat) { |
| 60 | int n; |
| 61 | |
| 62 | for (n = 0; n < 32; n++) { |
| 63 | square[n] = gf2_matrix_times(mat, mat[n]); |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | /* Construct an operator to apply len zeros to a crc. len must be a power of |
| 68 | two. If len is not a power of two, then the result is the same as for the |
| 69 | largest power of two less than len. The result for len == 0 is the same as |
| 70 | for len == 1. A version of this routine could be easily written for any |
| 71 | len, but that is not needed for this application. */ |
| 72 | static void crc32c_zeros_op(uint32_t* even, size_t len) { |
| 73 | int n; |
| 74 | uint32_t row; |
| 75 | uint32_t odd[32]; /* odd-power-of-two zeros operator */ |
| 76 | |
| 77 | /* put operator for one zero bit in odd */ |
| 78 | odd[0] = POLY; /* CRC-32C polynomial */ |
| 79 | row = 1; |
| 80 | for (n = 1; n < 32; n++) { |
| 81 | odd[n] = row; |
| 82 | row <<= 1; |
| 83 | } |
| 84 | |
| 85 | /* put operator for two zero bits in even */ |
| 86 | gf2_matrix_square(even, odd); |
| 87 | |
| 88 | /* put operator for four zero bits in odd */ |
| 89 | gf2_matrix_square(odd, even); |
| 90 | |
| 91 | /* first square will put the operator for one zero byte (eight zero bits), |
| 92 | in even -- next square puts operator for two zero bytes in odd, and so |
| 93 | on, until len has been rotated down to zero */ |
| 94 | do { |
| 95 | gf2_matrix_square(even, odd); |
| 96 | len >>= 1; |
| 97 | if (len == 0) { |
| 98 | return; |
| 99 | } |
| 100 | gf2_matrix_square(odd, even); |
| 101 | len >>= 1; |
| 102 | } while (len); |
| 103 | |
| 104 | /* answer ended up in odd -- copy to even */ |
| 105 | for (n = 0; n < 32; n++) { |
| 106 | even[n] = odd[n]; |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | /* Take a length and build four lookup tables for applying the zeros operator |
| 111 | for that length, byte-by-byte on the operand. */ |
| 112 | static void crc32c_zeros(uint32_t zeros[][256], size_t len) { |
| 113 | uint32_t n; |
| 114 | uint32_t op[32]; |
| 115 | |
| 116 | crc32c_zeros_op(op, len); |
| 117 | for (n = 0; n < 256; n++) { |
| 118 | zeros[0][n] = gf2_matrix_times(op, n); |
| 119 | zeros[1][n] = gf2_matrix_times(op, n << 8); |
| 120 | zeros[2][n] = gf2_matrix_times(op, n << 16); |
| 121 | zeros[3][n] = gf2_matrix_times(op, n << 24); |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | /* Apply the zeros operator table to crc. */ |
| 126 | static inline uint32_t crc32c_shift(uint32_t zeros[][256], uint32_t crc) { |
| 127 | // clang-format off |
| 128 | return zeros[0][crc & 0xff] ^ zeros[1][(crc >> 8) & 0xff] ^ |
| 129 | zeros[2][(crc >> 16) & 0xff] ^ zeros[3][crc >> 24]; |
| 130 | // clang-format on |
| 131 | } |
| 132 | |
| 133 | /* Block sizes for three-way parallel crc computation. LONG and SHORT must |
| 134 | both be powers of two. The associated string constants must be set |
| 135 | accordingly, for use in constructing the assembler instructions. */ |
| 136 | #define LONG 8192 |
| 137 | #define LONGx1 "8192" |
| 138 | #define LONGx2 "16384" |
| 139 | #define SHORT 256 |
| 140 | #define SHORTx1 "256" |
| 141 | #define SHORTx2 "512" |
| 142 | |
| 143 | /* Tables for hardware crc that shift a crc by LONG and SHORT zeros. */ |
| 144 | static uint32_t crc32c_long[4][256]; |
| 145 | static uint32_t crc32c_short[4][256]; |
| 146 | |
| 147 | /* Compute CRC-32C using the Intel hardware instruction. */ |
| 148 | static uint32_t crc32c_hw(uint32_t crc, const void* buf, size_t len) { |
| 149 | const unsigned char* next = buf; |
| 150 | const unsigned char* end; |
| 151 | uint64_t crc0, crc1, crc2; /* need to be 64 bits for crc32q */ |
| 152 | |
| 153 | /* pre-process the crc */ |
| 154 | crc0 = crc ^ 0xffffffff; |
| 155 | |
| 156 | /* compute the crc for up to seven leading bytes to bring the data pointer |
| 157 | to an eight-byte boundary */ |
| 158 | while (len && ((uintptr_t) next & 7) != 0) { |
| 159 | __asm__("crc32b\t" |
| 160 | "(%1), %0" |
| 161 | : "=r"(crc0) |
| 162 | : "r"(next), "0"(crc0)); |
| 163 | next++; |
| 164 | len--; |
| 165 | } |
| 166 | |
| 167 | /* compute the crc on sets of LONG*3 bytes, executing three independent crc |
| 168 | instructions, each on LONG bytes -- this is optimized for the Nehalem, |
| 169 | Westmere, Sandy Bridge, and Ivy Bridge architectures, which have a |
| 170 | throughput of one crc per cycle, but a latency of three cycles */ |
| 171 | while (len >= LONG * 3) { |
| 172 | crc1 = 0; |
| 173 | crc2 = 0; |
| 174 | end = next + LONG; |
| 175 | do { |
| 176 | __asm__("crc32q\t" |
| 177 | "(%3), %0\n\t" |
| 178 | "crc32q\t" LONGx1 "(%3), %1\n\t" |
| 179 | "crc32q\t" LONGx2 "(%3), %2" |
| 180 | : "=r"(crc0), "=r"(crc1), "=r"(crc2) |
| 181 | : "r"(next), "0"(crc0), "1"(crc1), "2"(crc2)); |
| 182 | next += 8; |
| 183 | } while (next < end); |
| 184 | crc0 = crc32c_shift(crc32c_long, (uint32_t) crc0) ^ crc1; |
| 185 | crc0 = crc32c_shift(crc32c_long, (uint32_t) crc0) ^ crc2; |
| 186 | next += LONG * 2; |
| 187 | len -= LONG * 3; |
| 188 | } |
| 189 | |
| 190 | /* do the same thing, but now on SHORT*3 blocks for the remaining data less |
| 191 | than a LONG*3 block */ |
| 192 | while (len >= SHORT * 3) { |
| 193 | crc1 = 0; |
| 194 | crc2 = 0; |
| 195 | end = next + SHORT; |
| 196 | do { |
| 197 | __asm__("crc32q\t" |
| 198 | "(%3), %0\n\t" |
| 199 | "crc32q\t" SHORTx1 "(%3), %1\n\t" |
| 200 | "crc32q\t" SHORTx2 "(%3), %2" |
| 201 | : "=r"(crc0), "=r"(crc1), "=r"(crc2) |
| 202 | : "r"(next), "0"(crc0), "1"(crc1), "2"(crc2)); |
| 203 | next += 8; |
| 204 | } while (next < end); |
| 205 | crc0 = crc32c_shift(crc32c_short, (uint32_t) crc0) ^ crc1; |
| 206 | crc0 = crc32c_shift(crc32c_short, (uint32_t) crc0) ^ crc2; |
| 207 | next += SHORT * 2; |
| 208 | len -= SHORT * 3; |
| 209 | } |
| 210 | |
| 211 | /* compute the crc on the remaining eight-byte units less than a SHORT*3 |
| 212 | block */ |
| 213 | end = next + (len - (len & 7)); |
| 214 | while (next < end) { |
| 215 | __asm__("crc32q\t" |
| 216 | "(%1), %0" |
| 217 | : "=r"(crc0) |
| 218 | : "r"(next), "0"(crc0)); |
| 219 | next += 8; |
| 220 | } |
| 221 | len &= 7; |
| 222 | |
| 223 | /* compute the crc for up to seven trailing bytes */ |
| 224 | while (len) { |
| 225 | __asm__("crc32b\t" |
| 226 | "(%1), %0" |
| 227 | : "=r"(crc0) |
| 228 | : "r"(next), "0"(crc0)); |
| 229 | next++; |
| 230 | len--; |
| 231 | } |
| 232 | |
| 233 | /* return a post-processed crc */ |
| 234 | return (uint32_t) crc0 ^ 0xffffffff; |
| 235 | } |
| 236 | |
| 237 | /* |
| 238 | * Probe for SSE 4.2 support. |
| 239 | */ |
| 240 | bool rawrtc_crc32c_sse42_supported(void) { |
| 241 | uint32_t eax, ecx; |
| 242 | eax = 1; |
| 243 | __asm__("cpuid" : "=c"(ecx) : "a"(eax) : "%ebx", "%edx"); |
| 244 | return ((ecx >> 20) & 1) != 0; |
| 245 | } |
| 246 | |
| 247 | /* |
| 248 | * Initialize tables for shifting CRCs. |
| 249 | */ |
| 250 | void rawrtc_crc32c_init_sse42(void) { |
| 251 | crc32c_zeros(crc32c_long, LONG); |
| 252 | crc32c_zeros(crc32c_short, SHORT); |
| 253 | } |
| 254 | |
| 255 | /* |
| 256 | * Compute CRC-32C using hardware instructions. |
| 257 | */ |
| 258 | uint32_t rawrtc_crc32c_sse42(void const* buffer, size_t length) { |
| 259 | return crc32c_hw(0x00000000, buffer, length); |
| 260 | } |