2 * Implement fast Fletcher4 using superscalar pipelines.
4 * Use regular C code to compute
5 * Fletcher4 in two incremental 64-bit parallel accumulator streams,
6 * and then combine the streams to form the final four checksum words.
7 * This implementation is a derivative of the AVX SIMD implementation by
8 * James Guilford and Jinshan Xiong from Intel (see zfs_fletcher_intel.c).
10 * Copyright (C) 2016 Romain Dolbeau.
13 * Romain Dolbeau <romain.dolbeau@atos.net>
15 * This software is available to you under a choice of one of two
16 * licenses. You may choose to be licensed under the terms of the GNU
17 * General Public License (GPL) Version 2, available from the file
18 * COPYING in the main directory of this source tree, or the
19 * OpenIB.org BSD license below:
21 * Redistribution and use in source and binary forms, with or
22 * without modification, are permitted provided that the following
25 * - Redistributions of source code must retain the above
26 * copyright notice, this list of conditions and the following
29 * - Redistributions in binary form must reproduce the above
30 * copyright notice, this list of conditions and the following
31 * disclaimer in the documentation and/or other materials
32 * provided with the distribution.
34 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
35 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
36 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
37 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
38 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
39 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
40 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
44 #include <sys/byteorder.h>
45 #include <sys/spa_checksum.h>
46 #include <sys/strings.h>
47 #include <zfs_fletcher.h>
50 fletcher_4_superscalar_init(fletcher_4_ctx_t *ctx)
52 bzero(ctx->superscalar, 4 * sizeof (zfs_fletcher_superscalar_t));
56 fletcher_4_superscalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp)
59 A = ctx->superscalar[0].v[0] + ctx->superscalar[0].v[1];
60 B = 2 * ctx->superscalar[1].v[0] + 2 * ctx->superscalar[1].v[1] -
61 ctx->superscalar[0].v[1];
62 C = 4 * ctx->superscalar[2].v[0] - ctx->superscalar[1].v[0] +
63 4 * ctx->superscalar[2].v[1] - 3 * ctx->superscalar[1].v[1];
64 D = 8 * ctx->superscalar[3].v[0] - 4 * ctx->superscalar[2].v[0] +
65 8 * ctx->superscalar[3].v[1] - 8 * ctx->superscalar[2].v[1] +
66 ctx->superscalar[1].v[1];
67 ZIO_SET_CHECKSUM(zcp, A, B, C, D);
71 fletcher_4_superscalar_native(fletcher_4_ctx_t *ctx,
72 const void *buf, uint64_t size)
74 const uint32_t *ip = buf;
75 const uint32_t *ipend = ip + (size / sizeof (uint32_t));
77 uint64_t a2, b2, c2, d2;
79 a = ctx->superscalar[0].v[0];
80 b = ctx->superscalar[1].v[0];
81 c = ctx->superscalar[2].v[0];
82 d = ctx->superscalar[3].v[0];
83 a2 = ctx->superscalar[0].v[1];
84 b2 = ctx->superscalar[1].v[1];
85 c2 = ctx->superscalar[2].v[1];
86 d2 = ctx->superscalar[3].v[1];
88 for (; ip < ipend; ip += 2) {
99 ctx->superscalar[0].v[0] = a;
100 ctx->superscalar[1].v[0] = b;
101 ctx->superscalar[2].v[0] = c;
102 ctx->superscalar[3].v[0] = d;
103 ctx->superscalar[0].v[1] = a2;
104 ctx->superscalar[1].v[1] = b2;
105 ctx->superscalar[2].v[1] = c2;
106 ctx->superscalar[3].v[1] = d2;
110 fletcher_4_superscalar_byteswap(fletcher_4_ctx_t *ctx,
111 const void *buf, uint64_t size)
113 const uint32_t *ip = buf;
114 const uint32_t *ipend = ip + (size / sizeof (uint32_t));
116 uint64_t a2, b2, c2, d2;
118 a = ctx->superscalar[0].v[0];
119 b = ctx->superscalar[1].v[0];
120 c = ctx->superscalar[2].v[0];
121 d = ctx->superscalar[3].v[0];
122 a2 = ctx->superscalar[0].v[1];
123 b2 = ctx->superscalar[1].v[1];
124 c2 = ctx->superscalar[2].v[1];
125 d2 = ctx->superscalar[3].v[1];
127 for (; ip < ipend; ip += 2) {
128 a += BSWAP_32(ip[0]);
129 a2 += BSWAP_32(ip[1]);
138 ctx->superscalar[0].v[0] = a;
139 ctx->superscalar[1].v[0] = b;
140 ctx->superscalar[2].v[0] = c;
141 ctx->superscalar[3].v[0] = d;
142 ctx->superscalar[0].v[1] = a2;
143 ctx->superscalar[1].v[1] = b2;
144 ctx->superscalar[2].v[1] = c2;
145 ctx->superscalar[3].v[1] = d2;
148 static boolean_t fletcher_4_superscalar_valid(void)
153 const fletcher_4_ops_t fletcher_4_superscalar_ops = {
154 .init_native = fletcher_4_superscalar_init,
155 .compute_native = fletcher_4_superscalar_native,
156 .fini_native = fletcher_4_superscalar_fini,
157 .init_byteswap = fletcher_4_superscalar_init,
158 .compute_byteswap = fletcher_4_superscalar_byteswap,
159 .fini_byteswap = fletcher_4_superscalar_fini,
160 .valid = fletcher_4_superscalar_valid,
161 .name = "superscalar"