1 /* 2 * Implement fast Fletcher4 with AVX2 instructions. (x86_64) 3 * 4 * Use the 256-bit AVX2 SIMD instructions and registers to compute 5 * Fletcher4 in four incremental 64-bit parallel accumulator streams, 6 * and then combine the streams to form the final four checksum words. 7 * 8 * Copyright (C) 2015 Intel Corporation. 9 * 10 * Authors: 11 * James Guilford <james.guilford@intel.com> 12 * Jinshan Xiong <jinshan.xiong@intel.com> 13 * 14 * This software is available to you under a choice of one of two 15 * licenses. You may choose to be licensed under the terms of the GNU 16 * General Public License (GPL) Version 2, available from the file 17 * COPYING in the main directory of this source tree, or the 18 * OpenIB.org BSD license below: 19 * 20 * Redistribution and use in source and binary forms, with or 21 * without modification, are permitted provided that the following 22 * conditions are met: 23 * 24 * - Redistributions of source code must retain the above 25 * copyright notice, this list of conditions and the following 26 * disclaimer. 27 * 28 * - Redistributions in binary form must reproduce the above 29 * copyright notice, this list of conditions and the following 30 * disclaimer in the documentation and/or other materials 31 * provided with the distribution. 32 * 33 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 34 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 35 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 36 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 37 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 38 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 39 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 40 * SOFTWARE. 41 */ 42 43 #if defined(HAVE_AVX) && defined(HAVE_AVX2) 44 45 #include <sys/spa_checksum.h> 46 #include <sys/string.h> 47 #include <sys/simd.h> 48 #include <zfs_fletcher.h> 49 50 static void 51 fletcher_4_avx2_init(fletcher_4_ctx_t *ctx) 52 { 53 memset(ctx->avx, 0, 4 * sizeof (zfs_fletcher_avx_t)); 54 } 55 56 static void 57 fletcher_4_avx2_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp) 58 { 59 uint64_t A, B, C, D; 60 61 A = ctx->avx[0].v[0] + ctx->avx[0].v[1] + 62 ctx->avx[0].v[2] + ctx->avx[0].v[3]; 63 B = 0 - ctx->avx[0].v[1] - 2 * ctx->avx[0].v[2] - 3 * ctx->avx[0].v[3] + 64 4 * ctx->avx[1].v[0] + 4 * ctx->avx[1].v[1] + 4 * ctx->avx[1].v[2] + 65 4 * ctx->avx[1].v[3]; 66 67 C = ctx->avx[0].v[2] + 3 * ctx->avx[0].v[3] - 6 * ctx->avx[1].v[0] - 68 10 * ctx->avx[1].v[1] - 14 * ctx->avx[1].v[2] - 69 18 * ctx->avx[1].v[3] + 16 * ctx->avx[2].v[0] + 70 16 * ctx->avx[2].v[1] + 16 * ctx->avx[2].v[2] + 71 16 * ctx->avx[2].v[3]; 72 73 D = 0 - ctx->avx[0].v[3] + 4 * ctx->avx[1].v[0] + 74 10 * ctx->avx[1].v[1] + 20 * ctx->avx[1].v[2] + 75 34 * ctx->avx[1].v[3] - 48 * ctx->avx[2].v[0] - 76 64 * ctx->avx[2].v[1] - 80 * ctx->avx[2].v[2] - 77 96 * ctx->avx[2].v[3] + 64 * ctx->avx[3].v[0] + 78 64 * ctx->avx[3].v[1] + 64 * ctx->avx[3].v[2] + 79 64 * ctx->avx[3].v[3]; 80 81 ZIO_SET_CHECKSUM(zcp, A, B, C, D); 82 } 83 84 #define FLETCHER_4_AVX2_RESTORE_CTX(ctx) \ 85 { \ 86 asm volatile("vmovdqu %0, %%ymm0" :: "m" ((ctx)->avx[0])); \ 87 asm volatile("vmovdqu %0, %%ymm1" :: "m" ((ctx)->avx[1])); \ 88 asm volatile("vmovdqu %0, %%ymm2" :: "m" ((ctx)->avx[2])); \ 89 asm volatile("vmovdqu %0, %%ymm3" :: "m" ((ctx)->avx[3])); \ 90 } 91 92 #define FLETCHER_4_AVX2_SAVE_CTX(ctx) \ 93 { \ 94 asm volatile("vmovdqu %%ymm0, %0" : "=m" ((ctx)->avx[0])); \ 95 asm volatile("vmovdqu %%ymm1, %0" : "=m" ((ctx)->avx[1])); \ 96 asm volatile("vmovdqu %%ymm2, %0" : "=m" ((ctx)->avx[2])); \ 97 asm volatile("vmovdqu %%ymm3, %0" : "=m" ((ctx)->avx[3])); \ 98 } 99 100 101 static void 102 fletcher_4_avx2_native(fletcher_4_ctx_t *ctx, const void *buf, uint64_t size) 103 { 104 const uint64_t *ip = buf; 105 const uint64_t *ipend = (uint64_t *)((uint8_t *)ip + size); 106 107 FLETCHER_4_AVX2_RESTORE_CTX(ctx); 108 109 do { 110 asm volatile("vpmovzxdq %0, %%ymm4"::"m" (*ip)); 111 asm volatile("vpaddq %ymm4, %ymm0, %ymm0"); 112 asm volatile("vpaddq %ymm0, %ymm1, %ymm1"); 113 asm volatile("vpaddq %ymm1, %ymm2, %ymm2"); 114 asm volatile("vpaddq %ymm2, %ymm3, %ymm3"); 115 } while ((ip += 2) < ipend); 116 117 FLETCHER_4_AVX2_SAVE_CTX(ctx); 118 asm volatile("vzeroupper"); 119 } 120 121 static void 122 fletcher_4_avx2_byteswap(fletcher_4_ctx_t *ctx, const void *buf, uint64_t size) 123 { 124 static const zfs_fletcher_avx_t mask = { 125 .v = { 0xFFFFFFFF00010203, 0xFFFFFFFF08090A0B, 126 0xFFFFFFFF00010203, 0xFFFFFFFF08090A0B } 127 }; 128 const uint64_t *ip = buf; 129 const uint64_t *ipend = (uint64_t *)((uint8_t *)ip + size); 130 131 FLETCHER_4_AVX2_RESTORE_CTX(ctx); 132 133 asm volatile("vmovdqu %0, %%ymm5" :: "m" (mask)); 134 135 do { 136 asm volatile("vpmovzxdq %0, %%ymm4"::"m" (*ip)); 137 asm volatile("vpshufb %ymm5, %ymm4, %ymm4"); 138 139 asm volatile("vpaddq %ymm4, %ymm0, %ymm0"); 140 asm volatile("vpaddq %ymm0, %ymm1, %ymm1"); 141 asm volatile("vpaddq %ymm1, %ymm2, %ymm2"); 142 asm volatile("vpaddq %ymm2, %ymm3, %ymm3"); 143 } while ((ip += 2) < ipend); 144 145 FLETCHER_4_AVX2_SAVE_CTX(ctx); 146 asm volatile("vzeroupper"); 147 } 148 149 static boolean_t fletcher_4_avx2_valid(void) 150 { 151 return (kfpu_allowed() && zfs_avx_available() && zfs_avx2_available()); 152 } 153 154 const fletcher_4_ops_t fletcher_4_avx2_ops = { 155 .init_native = fletcher_4_avx2_init, 156 .fini_native = fletcher_4_avx2_fini, 157 .compute_native = fletcher_4_avx2_native, 158 .init_byteswap = fletcher_4_avx2_init, 159 .fini_byteswap = fletcher_4_avx2_fini, 160 .compute_byteswap = fletcher_4_avx2_byteswap, 161 .valid = fletcher_4_avx2_valid, 162 .uses_fpu = B_TRUE, 163 .name = "avx2" 164 }; 165 166 #endif /* defined(HAVE_AVX) && defined(HAVE_AVX2) */ 167