1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
|
//+build !noasm,!appengine,gc
// Copyright (c) 2020 MinIO Inc. All rights reserved.
// Use of this source code is governed by a license that can be
// found in the LICENSE file.
package md5simd
import (
"fmt"
"math"
"unsafe"
"github.com/klauspost/cpuid/v2"
)
var hasAVX512 bool
func init() {
// VANDNPD requires AVX512DQ. Technically it could be VPTERNLOGQ which is AVX512F.
hasAVX512 = cpuid.CPU.Supports(cpuid.AVX512F, cpuid.AVX512DQ)
}
//go:noescape
func block8(state *uint32, base uintptr, bufs *int32, cache *byte, n int)
//go:noescape
func block16(state *uint32, base uintptr, ptrs *int32, mask uint64, n int)
// 8-way 4x uint32 digests in 4 ymm registers
// (ymm0, ymm1, ymm2, ymm3)
type digest8 struct {
v0, v1, v2, v3 [8]uint32
}
// Stack cache for 8x64 byte md5.BlockSize bytes.
// Must be 32-byte aligned, so allocate 512+32 and
// align upwards at runtime.
type cache8 [512 + 32]byte
// MD5 magic numbers for one lane of hashing; inflated
// 8x below at init time.
var md5consts = [64]uint32{
0xd76aa478, 0xe8c7b756, 0x242070db, 0xc1bdceee,
0xf57c0faf, 0x4787c62a, 0xa8304613, 0xfd469501,
0x698098d8, 0x8b44f7af, 0xffff5bb1, 0x895cd7be,
0x6b901122, 0xfd987193, 0xa679438e, 0x49b40821,
0xf61e2562, 0xc040b340, 0x265e5a51, 0xe9b6c7aa,
0xd62f105d, 0x02441453, 0xd8a1e681, 0xe7d3fbc8,
0x21e1cde6, 0xc33707d6, 0xf4d50d87, 0x455a14ed,
0xa9e3e905, 0xfcefa3f8, 0x676f02d9, 0x8d2a4c8a,
0xfffa3942, 0x8771f681, 0x6d9d6122, 0xfde5380c,
0xa4beea44, 0x4bdecfa9, 0xf6bb4b60, 0xbebfbc70,
0x289b7ec6, 0xeaa127fa, 0xd4ef3085, 0x04881d05,
0xd9d4d039, 0xe6db99e5, 0x1fa27cf8, 0xc4ac5665,
0xf4292244, 0x432aff97, 0xab9423a7, 0xfc93a039,
0x655b59c3, 0x8f0ccc92, 0xffeff47d, 0x85845dd1,
0x6fa87e4f, 0xfe2ce6e0, 0xa3014314, 0x4e0811a1,
0xf7537e82, 0xbd3af235, 0x2ad7d2bb, 0xeb86d391,
}
// inflate the consts 8-way for 8x md5 (256 bit ymm registers)
var avx256md5consts = func(c []uint32) []uint32 {
inf := make([]uint32, 8*len(c))
for i := range c {
for j := 0; j < 8; j++ {
inf[(i*8)+j] = c[i]
}
}
return inf
}(md5consts[:])
// 16-way 4x uint32 digests in 4 zmm registers
type digest16 struct {
v0, v1, v2, v3 [16]uint32
}
// inflate the consts 16-way for 16x md5 (512 bit zmm registers)
var avx512md5consts = func(c []uint32) []uint32 {
inf := make([]uint32, 16*len(c))
for i := range c {
for j := 0; j < 16; j++ {
inf[(i*16)+j] = c[i]
}
}
return inf
}(md5consts[:])
// Interface function to assembly code
func (s *md5Server) blockMd5_x16(d *digest16, input [16][]byte, half bool) {
if hasAVX512 {
blockMd5_avx512(d, input, s.allBufs, &s.maskRounds16)
return
}
// Preparing data using copy is slower since copies aren't inlined.
// Calculate on this goroutine
if half {
for i := range s.i8[0][:] {
s.i8[0][i] = input[i]
}
for i := range s.d8a.v0[:] {
s.d8a.v0[i], s.d8a.v1[i], s.d8a.v2[i], s.d8a.v3[i] = d.v0[i], d.v1[i], d.v2[i], d.v3[i]
}
blockMd5_avx2(&s.d8a, s.i8[0], s.allBufs, &s.maskRounds8a)
for i := range s.d8a.v0[:] {
d.v0[i], d.v1[i], d.v2[i], d.v3[i] = s.d8a.v0[i], s.d8a.v1[i], s.d8a.v2[i], s.d8a.v3[i]
}
return
}
for i := range s.i8[0][:] {
s.i8[0][i], s.i8[1][i] = input[i], input[8+i]
}
for i := range s.d8a.v0[:] {
j := (i + 8) & 15
s.d8a.v0[i], s.d8a.v1[i], s.d8a.v2[i], s.d8a.v3[i] = d.v0[i], d.v1[i], d.v2[i], d.v3[i]
s.d8b.v0[i], s.d8b.v1[i], s.d8b.v2[i], s.d8b.v3[i] = d.v0[j], d.v1[j], d.v2[j], d.v3[j]
}
// Benchmarks appears to be slightly faster when spinning up 2 goroutines instead
// of using the current for one of the blocks.
s.wg.Add(2)
go func() { blockMd5_avx2(&s.d8a, s.i8[0], s.allBufs, &s.maskRounds8a); s.wg.Done() }()
go func() { blockMd5_avx2(&s.d8b, s.i8[1], s.allBufs, &s.maskRounds8b); s.wg.Done() }()
s.wg.Wait()
for i := range s.d8a.v0[:] {
d.v0[i], d.v1[i], d.v2[i], d.v3[i] = s.d8a.v0[i], s.d8a.v1[i], s.d8a.v2[i], s.d8a.v3[i]
}
for i := range s.d8b.v0[:] {
j := (i + 8) & 15
d.v0[j], d.v1[j], d.v2[j], d.v3[j] = s.d8b.v0[i], s.d8b.v1[i], s.d8b.v2[i], s.d8b.v3[i]
}
}
// Interface function to AVX512 assembly code
func blockMd5_avx512(s *digest16, input [16][]byte, base []byte, maskRounds *[16]maskRounds) {
baseMin := uint64(uintptr(unsafe.Pointer(&(base[0]))))
ptrs := [16]int32{}
for i := range ptrs {
if len(input[i]) > 0 {
if len(input[i]) > internalBlockSize {
panic(fmt.Sprintf("Sanity check fails for lane %d: maximum input length cannot exceed internalBlockSize", i))
}
off := uint64(uintptr(unsafe.Pointer(&(input[i][0])))) - baseMin
if off > math.MaxUint32 {
panic(fmt.Sprintf("invalid buffer sent with offset %x", off))
}
ptrs[i] = int32(off)
}
}
sdup := *s // create copy of initial states to receive intermediate updates
rounds := generateMaskAndRounds16(input, maskRounds)
for r := 0; r < rounds; r++ {
m := maskRounds[r]
block16(&sdup.v0[0], uintptr(baseMin), &ptrs[0], m.mask, int(64*m.rounds))
for j := 0; j < len(ptrs); j++ {
ptrs[j] += int32(64 * m.rounds) // update pointers for next round
if m.mask&(1<<j) != 0 { // update digest if still masked as active
(*s).v0[j], (*s).v1[j], (*s).v2[j], (*s).v3[j] = sdup.v0[j], sdup.v1[j], sdup.v2[j], sdup.v3[j]
}
}
}
}
// Interface function to AVX2 assembly code
func blockMd5_avx2(s *digest8, input [8][]byte, base []byte, maskRounds *[8]maskRounds) {
baseMin := uint64(uintptr(unsafe.Pointer(&(base[0])))) - 4
ptrs := [8]int32{}
for i := range ptrs {
if len(input[i]) > 0 {
if len(input[i]) > internalBlockSize {
panic(fmt.Sprintf("Sanity check fails for lane %d: maximum input length cannot exceed internalBlockSize", i))
}
off := uint64(uintptr(unsafe.Pointer(&(input[i][0])))) - baseMin
if off > math.MaxUint32 {
panic(fmt.Sprintf("invalid buffer sent with offset %x", off))
}
ptrs[i] = int32(off)
}
}
sdup := *s // create copy of initial states to receive intermediate updates
rounds := generateMaskAndRounds8(input, maskRounds)
for r := 0; r < rounds; r++ {
m := maskRounds[r]
var cache cache8 // stack storage for block8 tmp state
block8(&sdup.v0[0], uintptr(baseMin), &ptrs[0], &cache[0], int(64*m.rounds))
for j := 0; j < len(ptrs); j++ {
ptrs[j] += int32(64 * m.rounds) // update pointers for next round
if m.mask&(1<<j) != 0 { // update digest if still masked as active
(*s).v0[j], (*s).v1[j], (*s).v2[j], (*s).v3[j] = sdup.v0[j], sdup.v1[j], sdup.v2[j], sdup.v3[j]
}
}
}
}
|
