| 1 |
/* |
| 2 |
* Hardware-accelerated CRC-32 variants for Linux on z Systems |
| 3 |
* |
| 4 |
* Use the z/Architecture Vector Extension Facility to accelerate the |
| 5 |
* computing of bitreflected CRC-32 checksums. |
| 6 |
* |
| 7 |
* This CRC-32 implementation algorithm is bitreflected and processes |
| 8 |
* the least-significant bit first (Little-Endian). |
| 9 |
* |
| 10 |
* This code was originally written by Hendrik Brueckner |
| 11 |
* <[email protected]> for use in the Linux kernel and has been |
| 12 |
* relicensed under the zlib license. |
| 13 |
*/ |
| 14 |
#define Z_ONCE |
| 15 |
#include "../../zutil.h" |
| 16 |
#include "crc32_vx_hooks.h" |
| 17 |
|
| 18 |
#include <stdint.h> |
| 19 |
#include <stdio.h> |
| 20 |
#include <vecintrin.h> |
| 21 |
#include <sys/auxv.h> |
| 22 |
|
| 23 |
#ifdef __clang__ |
| 24 |
# if ((__clang_major__ == 18) || (__clang_major__ == 19 && (__clang_minor__ < 1 || (__clang_minor__ == 1 && __clang_patchlevel__ < 2)))) |
| 25 |
# error crc32_vx optimizations are broken due to compiler bug in Clang versions: 18.0.0 <= clang_version < 19.1.2. \ |
| 26 |
Either disable the zlib crc32_vx optimization, or switch to another compiler/compiler version. |
| 27 |
# endif |
| 28 |
#endif |
| 29 |
|
| 30 |
#define VX_MIN_LEN 64 |
| 31 |
#define VX_ALIGNMENT 16L |
| 32 |
#define VX_ALIGN_MASK (VX_ALIGNMENT - 1) |
| 33 |
|
| 34 |
typedef unsigned char uv16qi __attribute__((vector_size(16))); |
| 35 |
typedef unsigned int uv4si __attribute__((vector_size(16))); |
| 36 |
typedef unsigned long long uv2di __attribute__((vector_size(16))); |
| 37 |
|
| 38 |
local uint32_t crc32_le_vgfm_16(uint32_t crc, const unsigned char *buf, size_t len) { |
| 39 |
/* |
| 40 |
* The CRC-32 constant block contains reduction constants to fold and |
| 41 |
* process particular chunks of the input data stream in parallel. |
| 42 |
* |
| 43 |
* For the CRC-32 variants, the constants are precomputed according to |
| 44 |
* these definitions: |
| 45 |
* |
| 46 |
* R1 = [(x4*128+32 mod P'(x) << 32)]' << 1 |
| 47 |
* R2 = [(x4*128-32 mod P'(x) << 32)]' << 1 |
| 48 |
* R3 = [(x128+32 mod P'(x) << 32)]' << 1 |
| 49 |
* R4 = [(x128-32 mod P'(x) << 32)]' << 1 |
| 50 |
* R5 = [(x64 mod P'(x) << 32)]' << 1 |
| 51 |
* R6 = [(x32 mod P'(x) << 32)]' << 1 |
| 52 |
* |
| 53 |
* The bitreflected Barret reduction constant, u', is defined as |
| 54 |
* the bit reversal of floor(x**64 / P(x)). |
| 55 |
* |
| 56 |
* where P(x) is the polynomial in the normal domain and the P'(x) is the |
| 57 |
* polynomial in the reversed (bitreflected) domain. |
| 58 |
* |
| 59 |
* CRC-32 (IEEE 802.3 Ethernet, ...) polynomials: |
| 60 |
* |
| 61 |
* P(x) = 0x04C11DB7 |
| 62 |
* P'(x) = 0xEDB88320 |
| 63 |
*/ |
| 64 |
const uv16qi perm_le2be = {15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0}; /* BE->LE mask */ |
| 65 |
const uv2di r2r1 = {0x1C6E41596, 0x154442BD4}; /* R2, R1 */ |
| 66 |
const uv2di r4r3 = {0x0CCAA009E, 0x1751997D0}; /* R4, R3 */ |
| 67 |
const uv2di r5 = {0, 0x163CD6124}; /* R5 */ |
| 68 |
const uv2di ru_poly = {0, 0x1F7011641}; /* u' */ |
| 69 |
const uv2di crc_poly = {0, 0x1DB710641}; /* P'(x) << 1 */ |
| 70 |
|
| 71 |
/* |
| 72 |
* Load the initial CRC value. |
| 73 |
* |
| 74 |
* The CRC value is loaded into the rightmost word of the |
| 75 |
* vector register and is later XORed with the LSB portion |
| 76 |
* of the loaded input data. |
| 77 |
*/ |
| 78 |
uv2di v0 = {0, 0}; |
| 79 |
v0 = (uv2di)vec_insert(crc, (uv4si)v0, 3); |
| 80 |
|
| 81 |
/* Load a 64-byte data chunk and XOR with CRC */ |
| 82 |
uv2di v1 = vec_perm(((uv2di *)buf)[0], ((uv2di *)buf)[0], perm_le2be); |
| 83 |
uv2di v2 = vec_perm(((uv2di *)buf)[1], ((uv2di *)buf)[1], perm_le2be); |
| 84 |
uv2di v3 = vec_perm(((uv2di *)buf)[2], ((uv2di *)buf)[2], perm_le2be); |
| 85 |
uv2di v4 = vec_perm(((uv2di *)buf)[3], ((uv2di *)buf)[3], perm_le2be); |
| 86 |
|
| 87 |
v1 ^= v0; |
| 88 |
buf += 64; |
| 89 |
len -= 64; |
| 90 |
|
| 91 |
while (len >= 64) { |
| 92 |
/* Load the next 64-byte data chunk */ |
| 93 |
uv16qi part1 = vec_perm(((uv16qi *)buf)[0], ((uv16qi *)buf)[0], perm_le2be); |
| 94 |
uv16qi part2 = vec_perm(((uv16qi *)buf)[1], ((uv16qi *)buf)[1], perm_le2be); |
| 95 |
uv16qi part3 = vec_perm(((uv16qi *)buf)[2], ((uv16qi *)buf)[2], perm_le2be); |
| 96 |
uv16qi part4 = vec_perm(((uv16qi *)buf)[3], ((uv16qi *)buf)[3], perm_le2be); |
| 97 |
|
| 98 |
/* |
| 99 |
* Perform a GF(2) multiplication of the doublewords in V1 with |
| 100 |
* the R1 and R2 reduction constants in V0. The intermediate result |
| 101 |
* is then folded (accumulated) with the next data chunk in PART1 and |
| 102 |
* stored in V1. Repeat this step for the register contents |
| 103 |
* in V2, V3, and V4 respectively. |
| 104 |
*/ |
| 105 |
v1 = (uv2di)vec_gfmsum_accum_128(r2r1, v1, part1); |
| 106 |
v2 = (uv2di)vec_gfmsum_accum_128(r2r1, v2, part2); |
| 107 |
v3 = (uv2di)vec_gfmsum_accum_128(r2r1, v3, part3); |
| 108 |
v4 = (uv2di)vec_gfmsum_accum_128(r2r1, v4, part4); |
| 109 |
|
| 110 |
buf += 64; |
| 111 |
len -= 64; |
| 112 |
} |
| 113 |
|
| 114 |
/* |
| 115 |
* Fold V1 to V4 into a single 128-bit value in V1. Multiply V1 with R3 |
| 116 |
* and R4 and accumulating the next 128-bit chunk until a single 128-bit |
| 117 |
* value remains. |
| 118 |
*/ |
| 119 |
v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v2); |
| 120 |
v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v3); |
| 121 |
v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v4); |
| 122 |
|
| 123 |
while (len >= 16) { |
| 124 |
/* Load next data chunk */ |
| 125 |
v2 = vec_perm(*(uv2di *)buf, *(uv2di *)buf, perm_le2be); |
| 126 |
|
| 127 |
/* Fold next data chunk */ |
| 128 |
v1 = (uv2di)vec_gfmsum_accum_128(r4r3, v1, (uv16qi)v2); |
| 129 |
|
| 130 |
buf += 16; |
| 131 |
len -= 16; |
| 132 |
} |
| 133 |
|
| 134 |
/* |
| 135 |
* Set up a vector register for byte shifts. The shift value must |
| 136 |
* be loaded in bits 1-4 in byte element 7 of a vector register. |
| 137 |
* Shift by 8 bytes: 0x40 |
| 138 |
* Shift by 4 bytes: 0x20 |
| 139 |
*/ |
| 140 |
uv16qi v9 = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}; |
| 141 |
v9 = vec_insert((unsigned char)0x40, v9, 7); |
| 142 |
|
| 143 |
/* |
| 144 |
* Prepare V0 for the next GF(2) multiplication: shift V0 by 8 bytes |
| 145 |
* to move R4 into the rightmost doubleword and set the leftmost |
| 146 |
* doubleword to 0x1. |
| 147 |
*/ |
| 148 |
v0 = vec_srb(r4r3, (uv2di)v9); |
| 149 |
v0[0] = 1; |
| 150 |
|
| 151 |
/* |
| 152 |
* Compute GF(2) product of V1 and V0. The rightmost doubleword |
| 153 |
* of V1 is multiplied with R4. The leftmost doubleword of V1 is |
| 154 |
* multiplied by 0x1 and is then XORed with rightmost product. |
| 155 |
* Implicitly, the intermediate leftmost product becomes padded |
| 156 |
*/ |
| 157 |
v1 = (uv2di)vec_gfmsum_128(v0, v1); |
| 158 |
|
| 159 |
/* |
| 160 |
* Now do the final 32-bit fold by multiplying the rightmost word |
| 161 |
* in V1 with R5 and XOR the result with the remaining bits in V1. |
| 162 |
* |
| 163 |
* To achieve this by a single VGFMAG, right shift V1 by a word |
| 164 |
* and store the result in V2 which is then accumulated. Use the |
| 165 |
* vector unpack instruction to load the rightmost half of the |
| 166 |
* doubleword into the rightmost doubleword element of V1; the other |
| 167 |
* half is loaded in the leftmost doubleword. |
| 168 |
* The vector register with CONST_R5 contains the R5 constant in the |
| 169 |
* rightmost doubleword and the leftmost doubleword is zero to ignore |
| 170 |
* the leftmost product of V1. |
| 171 |
*/ |
| 172 |
v9 = vec_insert((unsigned char)0x20, v9, 7); |
| 173 |
v2 = vec_srb(v1, (uv2di)v9); |
| 174 |
v1 = vec_unpackl((uv4si)v1); /* Split rightmost doubleword */ |
| 175 |
v1 = (uv2di)vec_gfmsum_accum_128(r5, v1, (uv16qi)v2); |
| 176 |
|
| 177 |
/* |
| 178 |
* Apply a Barret reduction to compute the final 32-bit CRC value. |
| 179 |
* |
| 180 |
* The input values to the Barret reduction are the degree-63 polynomial |
| 181 |
* in V1 (R(x)), degree-32 generator polynomial, and the reduction |
| 182 |
* constant u. The Barret reduction result is the CRC value of R(x) mod |
| 183 |
* P(x). |
| 184 |
* |
| 185 |
* The Barret reduction algorithm is defined as: |
| 186 |
* |
| 187 |
* 1. T1(x) = floor( R(x) / x^32 ) GF2MUL u |
| 188 |
* 2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x) |
| 189 |
* 3. C(x) = R(x) XOR T2(x) mod x^32 |
| 190 |
* |
| 191 |
* Note: The leftmost doubleword of vector register containing |
| 192 |
* CONST_RU_POLY is zero and, thus, the intermediate GF(2) product |
| 193 |
* is zero and does not contribute to the final result. |
| 194 |
*/ |
| 195 |
|
| 196 |
/* T1(x) = floor( R(x) / x^32 ) GF2MUL u */ |
| 197 |
v2 = vec_unpackl((uv4si)v1); |
| 198 |
v2 = (uv2di)vec_gfmsum_128(ru_poly, v2); |
| 199 |
|
| 200 |
/* |
| 201 |
* Compute the GF(2) product of the CRC polynomial with T1(x) in |
| 202 |
* V2 and XOR the intermediate result, T2(x), with the value in V1. |
| 203 |
* The final result is stored in word element 2 of V2. |
| 204 |
*/ |
| 205 |
v2 = vec_unpackl((uv4si)v2); |
| 206 |
v2 = (uv2di)vec_gfmsum_accum_128(crc_poly, v2, (uv16qi)v1); |
| 207 |
|
| 208 |
return ((uv4si)v2)[2]; |
| 209 |
} |
| 210 |
|
| 211 |
|
| 212 |
local unsigned long s390_crc32_vx(unsigned long crc, const unsigned char FAR *buf, z_size_t len) |
| 213 |
{ |
| 214 |
uintptr_t prealign, aligned, remaining; |
| 215 |
|
| 216 |
if (buf == Z_NULL) return 0UL; |
| 217 |
|
| 218 |
if (len < VX_MIN_LEN + VX_ALIGN_MASK) |
| 219 |
return crc32_z(crc, buf, len); |
| 220 |
|
| 221 |
if ((uintptr_t)buf & VX_ALIGN_MASK) { |
| 222 |
prealign = VX_ALIGNMENT - ((uintptr_t)buf & VX_ALIGN_MASK); |
| 223 |
len -= prealign; |
| 224 |
crc = crc32_z(crc, buf, prealign); |
| 225 |
buf += prealign; |
| 226 |
} |
| 227 |
aligned = len & ~VX_ALIGN_MASK; |
| 228 |
remaining = len & VX_ALIGN_MASK; |
| 229 |
|
| 230 |
crc = crc32_le_vgfm_16(crc ^ 0xffffffff, buf, (size_t)aligned) ^ 0xffffffff; |
| 231 |
|
| 232 |
if (remaining) |
| 233 |
crc = crc32_z(crc, buf + aligned, remaining); |
| 234 |
|
| 235 |
return crc; |
| 236 |
} |
| 237 |
|
| 238 |
local z_once_t s390_crc32_made = Z_ONCE_INIT; |
| 239 |
local void s390_crc32_setup() { |
| 240 |
unsigned long hwcap = getauxval(AT_HWCAP); |
| 241 |
|
| 242 |
if (hwcap & HWCAP_S390_VX) |
| 243 |
crc32_z_hook = s390_crc32_vx; |
| 244 |
else |
| 245 |
crc32_z_hook = crc32_z; |
| 246 |
} |
| 247 |
|
| 248 |
local unsigned long s390_crc32_init(unsigned long crc, const unsigned char FAR *buf, z_size_t len) |
| 249 |
{ |
| 250 |
z_once(&s390_crc32_made,s390_crc32_setup); |
| 251 |
return crc32_z_hook(crc, buf, len); |
| 252 |
} |
| 253 |
|
| 254 |
ZLIB_INTERNAL unsigned long (*crc32_z_hook)(unsigned long crc, const unsigned char FAR *buf, z_size_t len) = s390_crc32_init; |