sha256blockAvx512_amd64.go raw
1 //go:build !noasm && !appengine && gc
2 // +build !noasm,!appengine,gc
3
4 /*
5 * Minio Cloud Storage, (C) 2017 Minio, Inc.
6 *
7 * Licensed under the Apache License, Version 2.0 (the "License");
8 * you may not use this file except in compliance with the License.
9 * You may obtain a copy of the License at
10 *
11 * http://www.apache.org/licenses/LICENSE-2.0
12 *
13 * Unless required by applicable law or agreed to in writing, software
14 * distributed under the License is distributed on an "AS IS" BASIS,
15 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 * See the License for the specific language governing permissions and
17 * limitations under the License.
18 */
19
20 package sha256
21
22 import (
23 "encoding/binary"
24 "errors"
25 "hash"
26 "sort"
27 "sync/atomic"
28 "time"
29 )
30
31 //go:noescape
32 func sha256X16Avx512(digests *[512]byte, scratch *[512]byte, table *[512]uint64, mask []uint64, inputs [16][]byte)
33
34 // Avx512ServerUID - Do not start at 0 but next multiple of 16 so as to be able to
35 // differentiate with default initialiation value of 0
36 const Avx512ServerUID = 16
37
38 var uidCounter uint64
39
40 // NewAvx512 - initialize sha256 Avx512 implementation.
41 func NewAvx512(a512srv *Avx512Server) hash.Hash {
42 uid := atomic.AddUint64(&uidCounter, 1)
43 return &Avx512Digest{uid: uid, a512srv: a512srv}
44 }
45
46 // Avx512Digest - Type for computing SHA256 using Avx512
47 type Avx512Digest struct {
48 uid uint64
49 a512srv *Avx512Server
50 x [chunk]byte
51 nx int
52 len uint64
53 final bool
54 result [Size]byte
55 }
56
57 // Size - Return size of checksum
58 func (d *Avx512Digest) Size() int { return Size }
59
60 // BlockSize - Return blocksize of checksum
61 func (d Avx512Digest) BlockSize() int { return BlockSize }
62
63 // Reset - reset sha digest to its initial values
64 func (d *Avx512Digest) Reset() {
65 d.a512srv.blocksCh <- blockInput{uid: d.uid, reset: true}
66 d.nx = 0
67 d.len = 0
68 d.final = false
69 }
70
71 // Write to digest
72 func (d *Avx512Digest) Write(p []byte) (nn int, err error) {
73
74 if d.final {
75 return 0, errors.New("Avx512Digest already finalized. Reset first before writing again")
76 }
77
78 nn = len(p)
79 d.len += uint64(nn)
80 if d.nx > 0 {
81 n := copy(d.x[d.nx:], p)
82 d.nx += n
83 if d.nx == chunk {
84 d.a512srv.blocksCh <- blockInput{uid: d.uid, msg: d.x[:]}
85 d.nx = 0
86 }
87 p = p[n:]
88 }
89 if len(p) >= chunk {
90 n := len(p) &^ (chunk - 1)
91 d.a512srv.blocksCh <- blockInput{uid: d.uid, msg: p[:n]}
92 p = p[n:]
93 }
94 if len(p) > 0 {
95 d.nx = copy(d.x[:], p)
96 }
97 return
98 }
99
100 // Sum - Return sha256 sum in bytes
101 func (d *Avx512Digest) Sum(in []byte) (result []byte) {
102
103 if d.final {
104 return append(in, d.result[:]...)
105 }
106
107 trail := make([]byte, 0, 128)
108 trail = append(trail, d.x[:d.nx]...)
109
110 len := d.len
111 // Padding. Add a 1 bit and 0 bits until 56 bytes mod 64.
112 var tmp [64]byte
113 tmp[0] = 0x80
114 if len%64 < 56 {
115 trail = append(trail, tmp[0:56-len%64]...)
116 } else {
117 trail = append(trail, tmp[0:64+56-len%64]...)
118 }
119 d.nx = 0
120
121 // Length in bits.
122 len <<= 3
123 for i := uint(0); i < 8; i++ {
124 tmp[i] = byte(len >> (56 - 8*i))
125 }
126 trail = append(trail, tmp[0:8]...)
127
128 sumCh := make(chan [Size]byte)
129 d.a512srv.blocksCh <- blockInput{uid: d.uid, msg: trail, final: true, sumCh: sumCh}
130 d.result = <-sumCh
131 d.final = true
132 return append(in, d.result[:]...)
133 }
134
135 var table = [512]uint64{
136 0x428a2f98428a2f98, 0x428a2f98428a2f98, 0x428a2f98428a2f98, 0x428a2f98428a2f98,
137 0x428a2f98428a2f98, 0x428a2f98428a2f98, 0x428a2f98428a2f98, 0x428a2f98428a2f98,
138 0x7137449171374491, 0x7137449171374491, 0x7137449171374491, 0x7137449171374491,
139 0x7137449171374491, 0x7137449171374491, 0x7137449171374491, 0x7137449171374491,
140 0xb5c0fbcfb5c0fbcf, 0xb5c0fbcfb5c0fbcf, 0xb5c0fbcfb5c0fbcf, 0xb5c0fbcfb5c0fbcf,
141 0xb5c0fbcfb5c0fbcf, 0xb5c0fbcfb5c0fbcf, 0xb5c0fbcfb5c0fbcf, 0xb5c0fbcfb5c0fbcf,
142 0xe9b5dba5e9b5dba5, 0xe9b5dba5e9b5dba5, 0xe9b5dba5e9b5dba5, 0xe9b5dba5e9b5dba5,
143 0xe9b5dba5e9b5dba5, 0xe9b5dba5e9b5dba5, 0xe9b5dba5e9b5dba5, 0xe9b5dba5e9b5dba5,
144 0x3956c25b3956c25b, 0x3956c25b3956c25b, 0x3956c25b3956c25b, 0x3956c25b3956c25b,
145 0x3956c25b3956c25b, 0x3956c25b3956c25b, 0x3956c25b3956c25b, 0x3956c25b3956c25b,
146 0x59f111f159f111f1, 0x59f111f159f111f1, 0x59f111f159f111f1, 0x59f111f159f111f1,
147 0x59f111f159f111f1, 0x59f111f159f111f1, 0x59f111f159f111f1, 0x59f111f159f111f1,
148 0x923f82a4923f82a4, 0x923f82a4923f82a4, 0x923f82a4923f82a4, 0x923f82a4923f82a4,
149 0x923f82a4923f82a4, 0x923f82a4923f82a4, 0x923f82a4923f82a4, 0x923f82a4923f82a4,
150 0xab1c5ed5ab1c5ed5, 0xab1c5ed5ab1c5ed5, 0xab1c5ed5ab1c5ed5, 0xab1c5ed5ab1c5ed5,
151 0xab1c5ed5ab1c5ed5, 0xab1c5ed5ab1c5ed5, 0xab1c5ed5ab1c5ed5, 0xab1c5ed5ab1c5ed5,
152 0xd807aa98d807aa98, 0xd807aa98d807aa98, 0xd807aa98d807aa98, 0xd807aa98d807aa98,
153 0xd807aa98d807aa98, 0xd807aa98d807aa98, 0xd807aa98d807aa98, 0xd807aa98d807aa98,
154 0x12835b0112835b01, 0x12835b0112835b01, 0x12835b0112835b01, 0x12835b0112835b01,
155 0x12835b0112835b01, 0x12835b0112835b01, 0x12835b0112835b01, 0x12835b0112835b01,
156 0x243185be243185be, 0x243185be243185be, 0x243185be243185be, 0x243185be243185be,
157 0x243185be243185be, 0x243185be243185be, 0x243185be243185be, 0x243185be243185be,
158 0x550c7dc3550c7dc3, 0x550c7dc3550c7dc3, 0x550c7dc3550c7dc3, 0x550c7dc3550c7dc3,
159 0x550c7dc3550c7dc3, 0x550c7dc3550c7dc3, 0x550c7dc3550c7dc3, 0x550c7dc3550c7dc3,
160 0x72be5d7472be5d74, 0x72be5d7472be5d74, 0x72be5d7472be5d74, 0x72be5d7472be5d74,
161 0x72be5d7472be5d74, 0x72be5d7472be5d74, 0x72be5d7472be5d74, 0x72be5d7472be5d74,
162 0x80deb1fe80deb1fe, 0x80deb1fe80deb1fe, 0x80deb1fe80deb1fe, 0x80deb1fe80deb1fe,
163 0x80deb1fe80deb1fe, 0x80deb1fe80deb1fe, 0x80deb1fe80deb1fe, 0x80deb1fe80deb1fe,
164 0x9bdc06a79bdc06a7, 0x9bdc06a79bdc06a7, 0x9bdc06a79bdc06a7, 0x9bdc06a79bdc06a7,
165 0x9bdc06a79bdc06a7, 0x9bdc06a79bdc06a7, 0x9bdc06a79bdc06a7, 0x9bdc06a79bdc06a7,
166 0xc19bf174c19bf174, 0xc19bf174c19bf174, 0xc19bf174c19bf174, 0xc19bf174c19bf174,
167 0xc19bf174c19bf174, 0xc19bf174c19bf174, 0xc19bf174c19bf174, 0xc19bf174c19bf174,
168 0xe49b69c1e49b69c1, 0xe49b69c1e49b69c1, 0xe49b69c1e49b69c1, 0xe49b69c1e49b69c1,
169 0xe49b69c1e49b69c1, 0xe49b69c1e49b69c1, 0xe49b69c1e49b69c1, 0xe49b69c1e49b69c1,
170 0xefbe4786efbe4786, 0xefbe4786efbe4786, 0xefbe4786efbe4786, 0xefbe4786efbe4786,
171 0xefbe4786efbe4786, 0xefbe4786efbe4786, 0xefbe4786efbe4786, 0xefbe4786efbe4786,
172 0x0fc19dc60fc19dc6, 0x0fc19dc60fc19dc6, 0x0fc19dc60fc19dc6, 0x0fc19dc60fc19dc6,
173 0x0fc19dc60fc19dc6, 0x0fc19dc60fc19dc6, 0x0fc19dc60fc19dc6, 0x0fc19dc60fc19dc6,
174 0x240ca1cc240ca1cc, 0x240ca1cc240ca1cc, 0x240ca1cc240ca1cc, 0x240ca1cc240ca1cc,
175 0x240ca1cc240ca1cc, 0x240ca1cc240ca1cc, 0x240ca1cc240ca1cc, 0x240ca1cc240ca1cc,
176 0x2de92c6f2de92c6f, 0x2de92c6f2de92c6f, 0x2de92c6f2de92c6f, 0x2de92c6f2de92c6f,
177 0x2de92c6f2de92c6f, 0x2de92c6f2de92c6f, 0x2de92c6f2de92c6f, 0x2de92c6f2de92c6f,
178 0x4a7484aa4a7484aa, 0x4a7484aa4a7484aa, 0x4a7484aa4a7484aa, 0x4a7484aa4a7484aa,
179 0x4a7484aa4a7484aa, 0x4a7484aa4a7484aa, 0x4a7484aa4a7484aa, 0x4a7484aa4a7484aa,
180 0x5cb0a9dc5cb0a9dc, 0x5cb0a9dc5cb0a9dc, 0x5cb0a9dc5cb0a9dc, 0x5cb0a9dc5cb0a9dc,
181 0x5cb0a9dc5cb0a9dc, 0x5cb0a9dc5cb0a9dc, 0x5cb0a9dc5cb0a9dc, 0x5cb0a9dc5cb0a9dc,
182 0x76f988da76f988da, 0x76f988da76f988da, 0x76f988da76f988da, 0x76f988da76f988da,
183 0x76f988da76f988da, 0x76f988da76f988da, 0x76f988da76f988da, 0x76f988da76f988da,
184 0x983e5152983e5152, 0x983e5152983e5152, 0x983e5152983e5152, 0x983e5152983e5152,
185 0x983e5152983e5152, 0x983e5152983e5152, 0x983e5152983e5152, 0x983e5152983e5152,
186 0xa831c66da831c66d, 0xa831c66da831c66d, 0xa831c66da831c66d, 0xa831c66da831c66d,
187 0xa831c66da831c66d, 0xa831c66da831c66d, 0xa831c66da831c66d, 0xa831c66da831c66d,
188 0xb00327c8b00327c8, 0xb00327c8b00327c8, 0xb00327c8b00327c8, 0xb00327c8b00327c8,
189 0xb00327c8b00327c8, 0xb00327c8b00327c8, 0xb00327c8b00327c8, 0xb00327c8b00327c8,
190 0xbf597fc7bf597fc7, 0xbf597fc7bf597fc7, 0xbf597fc7bf597fc7, 0xbf597fc7bf597fc7,
191 0xbf597fc7bf597fc7, 0xbf597fc7bf597fc7, 0xbf597fc7bf597fc7, 0xbf597fc7bf597fc7,
192 0xc6e00bf3c6e00bf3, 0xc6e00bf3c6e00bf3, 0xc6e00bf3c6e00bf3, 0xc6e00bf3c6e00bf3,
193 0xc6e00bf3c6e00bf3, 0xc6e00bf3c6e00bf3, 0xc6e00bf3c6e00bf3, 0xc6e00bf3c6e00bf3,
194 0xd5a79147d5a79147, 0xd5a79147d5a79147, 0xd5a79147d5a79147, 0xd5a79147d5a79147,
195 0xd5a79147d5a79147, 0xd5a79147d5a79147, 0xd5a79147d5a79147, 0xd5a79147d5a79147,
196 0x06ca635106ca6351, 0x06ca635106ca6351, 0x06ca635106ca6351, 0x06ca635106ca6351,
197 0x06ca635106ca6351, 0x06ca635106ca6351, 0x06ca635106ca6351, 0x06ca635106ca6351,
198 0x1429296714292967, 0x1429296714292967, 0x1429296714292967, 0x1429296714292967,
199 0x1429296714292967, 0x1429296714292967, 0x1429296714292967, 0x1429296714292967,
200 0x27b70a8527b70a85, 0x27b70a8527b70a85, 0x27b70a8527b70a85, 0x27b70a8527b70a85,
201 0x27b70a8527b70a85, 0x27b70a8527b70a85, 0x27b70a8527b70a85, 0x27b70a8527b70a85,
202 0x2e1b21382e1b2138, 0x2e1b21382e1b2138, 0x2e1b21382e1b2138, 0x2e1b21382e1b2138,
203 0x2e1b21382e1b2138, 0x2e1b21382e1b2138, 0x2e1b21382e1b2138, 0x2e1b21382e1b2138,
204 0x4d2c6dfc4d2c6dfc, 0x4d2c6dfc4d2c6dfc, 0x4d2c6dfc4d2c6dfc, 0x4d2c6dfc4d2c6dfc,
205 0x4d2c6dfc4d2c6dfc, 0x4d2c6dfc4d2c6dfc, 0x4d2c6dfc4d2c6dfc, 0x4d2c6dfc4d2c6dfc,
206 0x53380d1353380d13, 0x53380d1353380d13, 0x53380d1353380d13, 0x53380d1353380d13,
207 0x53380d1353380d13, 0x53380d1353380d13, 0x53380d1353380d13, 0x53380d1353380d13,
208 0x650a7354650a7354, 0x650a7354650a7354, 0x650a7354650a7354, 0x650a7354650a7354,
209 0x650a7354650a7354, 0x650a7354650a7354, 0x650a7354650a7354, 0x650a7354650a7354,
210 0x766a0abb766a0abb, 0x766a0abb766a0abb, 0x766a0abb766a0abb, 0x766a0abb766a0abb,
211 0x766a0abb766a0abb, 0x766a0abb766a0abb, 0x766a0abb766a0abb, 0x766a0abb766a0abb,
212 0x81c2c92e81c2c92e, 0x81c2c92e81c2c92e, 0x81c2c92e81c2c92e, 0x81c2c92e81c2c92e,
213 0x81c2c92e81c2c92e, 0x81c2c92e81c2c92e, 0x81c2c92e81c2c92e, 0x81c2c92e81c2c92e,
214 0x92722c8592722c85, 0x92722c8592722c85, 0x92722c8592722c85, 0x92722c8592722c85,
215 0x92722c8592722c85, 0x92722c8592722c85, 0x92722c8592722c85, 0x92722c8592722c85,
216 0xa2bfe8a1a2bfe8a1, 0xa2bfe8a1a2bfe8a1, 0xa2bfe8a1a2bfe8a1, 0xa2bfe8a1a2bfe8a1,
217 0xa2bfe8a1a2bfe8a1, 0xa2bfe8a1a2bfe8a1, 0xa2bfe8a1a2bfe8a1, 0xa2bfe8a1a2bfe8a1,
218 0xa81a664ba81a664b, 0xa81a664ba81a664b, 0xa81a664ba81a664b, 0xa81a664ba81a664b,
219 0xa81a664ba81a664b, 0xa81a664ba81a664b, 0xa81a664ba81a664b, 0xa81a664ba81a664b,
220 0xc24b8b70c24b8b70, 0xc24b8b70c24b8b70, 0xc24b8b70c24b8b70, 0xc24b8b70c24b8b70,
221 0xc24b8b70c24b8b70, 0xc24b8b70c24b8b70, 0xc24b8b70c24b8b70, 0xc24b8b70c24b8b70,
222 0xc76c51a3c76c51a3, 0xc76c51a3c76c51a3, 0xc76c51a3c76c51a3, 0xc76c51a3c76c51a3,
223 0xc76c51a3c76c51a3, 0xc76c51a3c76c51a3, 0xc76c51a3c76c51a3, 0xc76c51a3c76c51a3,
224 0xd192e819d192e819, 0xd192e819d192e819, 0xd192e819d192e819, 0xd192e819d192e819,
225 0xd192e819d192e819, 0xd192e819d192e819, 0xd192e819d192e819, 0xd192e819d192e819,
226 0xd6990624d6990624, 0xd6990624d6990624, 0xd6990624d6990624, 0xd6990624d6990624,
227 0xd6990624d6990624, 0xd6990624d6990624, 0xd6990624d6990624, 0xd6990624d6990624,
228 0xf40e3585f40e3585, 0xf40e3585f40e3585, 0xf40e3585f40e3585, 0xf40e3585f40e3585,
229 0xf40e3585f40e3585, 0xf40e3585f40e3585, 0xf40e3585f40e3585, 0xf40e3585f40e3585,
230 0x106aa070106aa070, 0x106aa070106aa070, 0x106aa070106aa070, 0x106aa070106aa070,
231 0x106aa070106aa070, 0x106aa070106aa070, 0x106aa070106aa070, 0x106aa070106aa070,
232 0x19a4c11619a4c116, 0x19a4c11619a4c116, 0x19a4c11619a4c116, 0x19a4c11619a4c116,
233 0x19a4c11619a4c116, 0x19a4c11619a4c116, 0x19a4c11619a4c116, 0x19a4c11619a4c116,
234 0x1e376c081e376c08, 0x1e376c081e376c08, 0x1e376c081e376c08, 0x1e376c081e376c08,
235 0x1e376c081e376c08, 0x1e376c081e376c08, 0x1e376c081e376c08, 0x1e376c081e376c08,
236 0x2748774c2748774c, 0x2748774c2748774c, 0x2748774c2748774c, 0x2748774c2748774c,
237 0x2748774c2748774c, 0x2748774c2748774c, 0x2748774c2748774c, 0x2748774c2748774c,
238 0x34b0bcb534b0bcb5, 0x34b0bcb534b0bcb5, 0x34b0bcb534b0bcb5, 0x34b0bcb534b0bcb5,
239 0x34b0bcb534b0bcb5, 0x34b0bcb534b0bcb5, 0x34b0bcb534b0bcb5, 0x34b0bcb534b0bcb5,
240 0x391c0cb3391c0cb3, 0x391c0cb3391c0cb3, 0x391c0cb3391c0cb3, 0x391c0cb3391c0cb3,
241 0x391c0cb3391c0cb3, 0x391c0cb3391c0cb3, 0x391c0cb3391c0cb3, 0x391c0cb3391c0cb3,
242 0x4ed8aa4a4ed8aa4a, 0x4ed8aa4a4ed8aa4a, 0x4ed8aa4a4ed8aa4a, 0x4ed8aa4a4ed8aa4a,
243 0x4ed8aa4a4ed8aa4a, 0x4ed8aa4a4ed8aa4a, 0x4ed8aa4a4ed8aa4a, 0x4ed8aa4a4ed8aa4a,
244 0x5b9cca4f5b9cca4f, 0x5b9cca4f5b9cca4f, 0x5b9cca4f5b9cca4f, 0x5b9cca4f5b9cca4f,
245 0x5b9cca4f5b9cca4f, 0x5b9cca4f5b9cca4f, 0x5b9cca4f5b9cca4f, 0x5b9cca4f5b9cca4f,
246 0x682e6ff3682e6ff3, 0x682e6ff3682e6ff3, 0x682e6ff3682e6ff3, 0x682e6ff3682e6ff3,
247 0x682e6ff3682e6ff3, 0x682e6ff3682e6ff3, 0x682e6ff3682e6ff3, 0x682e6ff3682e6ff3,
248 0x748f82ee748f82ee, 0x748f82ee748f82ee, 0x748f82ee748f82ee, 0x748f82ee748f82ee,
249 0x748f82ee748f82ee, 0x748f82ee748f82ee, 0x748f82ee748f82ee, 0x748f82ee748f82ee,
250 0x78a5636f78a5636f, 0x78a5636f78a5636f, 0x78a5636f78a5636f, 0x78a5636f78a5636f,
251 0x78a5636f78a5636f, 0x78a5636f78a5636f, 0x78a5636f78a5636f, 0x78a5636f78a5636f,
252 0x84c8781484c87814, 0x84c8781484c87814, 0x84c8781484c87814, 0x84c8781484c87814,
253 0x84c8781484c87814, 0x84c8781484c87814, 0x84c8781484c87814, 0x84c8781484c87814,
254 0x8cc702088cc70208, 0x8cc702088cc70208, 0x8cc702088cc70208, 0x8cc702088cc70208,
255 0x8cc702088cc70208, 0x8cc702088cc70208, 0x8cc702088cc70208, 0x8cc702088cc70208,
256 0x90befffa90befffa, 0x90befffa90befffa, 0x90befffa90befffa, 0x90befffa90befffa,
257 0x90befffa90befffa, 0x90befffa90befffa, 0x90befffa90befffa, 0x90befffa90befffa,
258 0xa4506ceba4506ceb, 0xa4506ceba4506ceb, 0xa4506ceba4506ceb, 0xa4506ceba4506ceb,
259 0xa4506ceba4506ceb, 0xa4506ceba4506ceb, 0xa4506ceba4506ceb, 0xa4506ceba4506ceb,
260 0xbef9a3f7bef9a3f7, 0xbef9a3f7bef9a3f7, 0xbef9a3f7bef9a3f7, 0xbef9a3f7bef9a3f7,
261 0xbef9a3f7bef9a3f7, 0xbef9a3f7bef9a3f7, 0xbef9a3f7bef9a3f7, 0xbef9a3f7bef9a3f7,
262 0xc67178f2c67178f2, 0xc67178f2c67178f2, 0xc67178f2c67178f2, 0xc67178f2c67178f2,
263 0xc67178f2c67178f2, 0xc67178f2c67178f2, 0xc67178f2c67178f2, 0xc67178f2c67178f2}
264
265 // Interface function to assembly ode
266 func blockAvx512(digests *[512]byte, input [16][]byte, mask []uint64) [16][Size]byte {
267
268 scratch := [512]byte{}
269 sha256X16Avx512(digests, &scratch, &table, mask, input)
270
271 output := [16][Size]byte{}
272 for i := 0; i < 16; i++ {
273 output[i] = getDigest(i, digests[:])
274 }
275
276 return output
277 }
278
279 func getDigest(index int, state []byte) (sum [Size]byte) {
280 for j := 0; j < 16; j += 2 {
281 for i := index*4 + j*Size; i < index*4+(j+1)*Size; i += Size {
282 binary.BigEndian.PutUint32(sum[j*2:], binary.LittleEndian.Uint32(state[i:i+4]))
283 }
284 }
285 return
286 }
287
288 // Message to send across input channel
289 type blockInput struct {
290 uid uint64
291 msg []byte
292 reset bool
293 final bool
294 sumCh chan [Size]byte
295 }
296
297 // Avx512Server - Type to implement 16x parallel handling of SHA256 invocations
298 type Avx512Server struct {
299 blocksCh chan blockInput // Input channel
300 totalIn int // Total number of inputs waiting to be processed
301 lanes [16]Avx512LaneInfo // Array with info per lane (out of 16)
302 digests map[uint64][Size]byte // Map of uids to (interim) digest results
303 }
304
305 // Avx512LaneInfo - Info for each lane
306 type Avx512LaneInfo struct {
307 uid uint64 // unique identification for this SHA processing
308 block []byte // input block to be processed
309 outputCh chan [Size]byte // channel for output result
310 }
311
312 // NewAvx512Server - Create new object for parallel processing handling
313 func NewAvx512Server() *Avx512Server {
314 a512srv := &Avx512Server{}
315 a512srv.digests = make(map[uint64][Size]byte)
316 a512srv.blocksCh = make(chan blockInput)
317
318 // Start a single thread for reading from the input channel
319 go a512srv.Process()
320 return a512srv
321 }
322
323 // Process - Sole handler for reading from the input channel
324 func (a512srv *Avx512Server) Process() {
325 for {
326 select {
327 case block := <-a512srv.blocksCh:
328 if block.reset {
329 a512srv.reset(block.uid)
330 continue
331 }
332 index := block.uid & 0xf
333 // fmt.Println("Adding message:", block.uid, index)
334
335 if a512srv.lanes[index].block != nil { // If slot is already filled, process all inputs
336 //fmt.Println("Invoking Blocks()")
337 a512srv.blocks()
338 }
339 a512srv.totalIn++
340 a512srv.lanes[index] = Avx512LaneInfo{uid: block.uid, block: block.msg}
341 if block.final {
342 a512srv.lanes[index].outputCh = block.sumCh
343 }
344 if a512srv.totalIn == len(a512srv.lanes) {
345 // fmt.Println("Invoking Blocks() while FULL: ")
346 a512srv.blocks()
347 }
348
349 // TODO: test with larger timeout
350 case <-time.After(1 * time.Microsecond):
351 for _, lane := range a512srv.lanes {
352 if lane.block != nil { // check if there is any input to process
353 // fmt.Println("Invoking Blocks() on TIMEOUT: ")
354 a512srv.blocks()
355 break // we are done
356 }
357 }
358 }
359 }
360 }
361
362 // Do a reset for this calculation
363 func (a512srv *Avx512Server) reset(uid uint64) {
364
365 // Check if there is a message still waiting to be processed (and remove if so)
366 for i, lane := range a512srv.lanes {
367 if lane.uid == uid {
368 if lane.block != nil {
369 a512srv.lanes[i] = Avx512LaneInfo{} // clear message
370 a512srv.totalIn--
371 }
372 }
373 }
374
375 // Delete entry from hash map
376 delete(a512srv.digests, uid)
377 }
378
379 // Invoke assembly and send results back
380 func (a512srv *Avx512Server) blocks() {
381
382 inputs := [16][]byte{}
383 for i := range inputs {
384 inputs[i] = a512srv.lanes[i].block
385 }
386
387 mask := expandMask(genMask(inputs))
388 outputs := blockAvx512(a512srv.getDigests(), inputs, mask)
389
390 a512srv.totalIn = 0
391 for i := 0; i < len(outputs); i++ {
392 uid, outputCh := a512srv.lanes[i].uid, a512srv.lanes[i].outputCh
393 a512srv.digests[uid] = outputs[i]
394 a512srv.lanes[i] = Avx512LaneInfo{}
395
396 if outputCh != nil {
397 // Send back result
398 outputCh <- outputs[i]
399 delete(a512srv.digests, uid) // Delete entry from hashmap
400 }
401 }
402 }
403
404 func (a512srv *Avx512Server) Write(uid uint64, p []byte) (nn int, err error) {
405 a512srv.blocksCh <- blockInput{uid: uid, msg: p}
406 return len(p), nil
407 }
408
409 // Sum - return sha256 sum in bytes for a given sum id.
410 func (a512srv *Avx512Server) Sum(uid uint64, p []byte) [32]byte {
411 sumCh := make(chan [32]byte)
412 a512srv.blocksCh <- blockInput{uid: uid, msg: p, final: true, sumCh: sumCh}
413 return <-sumCh
414 }
415
416 func (a512srv *Avx512Server) getDigests() *[512]byte {
417 digests := [512]byte{}
418 for i, lane := range a512srv.lanes {
419 a, ok := a512srv.digests[lane.uid]
420 if ok {
421 binary.BigEndian.PutUint32(digests[(i+0*16)*4:], binary.LittleEndian.Uint32(a[0:4]))
422 binary.BigEndian.PutUint32(digests[(i+1*16)*4:], binary.LittleEndian.Uint32(a[4:8]))
423 binary.BigEndian.PutUint32(digests[(i+2*16)*4:], binary.LittleEndian.Uint32(a[8:12]))
424 binary.BigEndian.PutUint32(digests[(i+3*16)*4:], binary.LittleEndian.Uint32(a[12:16]))
425 binary.BigEndian.PutUint32(digests[(i+4*16)*4:], binary.LittleEndian.Uint32(a[16:20]))
426 binary.BigEndian.PutUint32(digests[(i+5*16)*4:], binary.LittleEndian.Uint32(a[20:24]))
427 binary.BigEndian.PutUint32(digests[(i+6*16)*4:], binary.LittleEndian.Uint32(a[24:28]))
428 binary.BigEndian.PutUint32(digests[(i+7*16)*4:], binary.LittleEndian.Uint32(a[28:32]))
429 } else {
430 binary.LittleEndian.PutUint32(digests[(i+0*16)*4:], init0)
431 binary.LittleEndian.PutUint32(digests[(i+1*16)*4:], init1)
432 binary.LittleEndian.PutUint32(digests[(i+2*16)*4:], init2)
433 binary.LittleEndian.PutUint32(digests[(i+3*16)*4:], init3)
434 binary.LittleEndian.PutUint32(digests[(i+4*16)*4:], init4)
435 binary.LittleEndian.PutUint32(digests[(i+5*16)*4:], init5)
436 binary.LittleEndian.PutUint32(digests[(i+6*16)*4:], init6)
437 binary.LittleEndian.PutUint32(digests[(i+7*16)*4:], init7)
438 }
439 }
440 return &digests
441 }
442
443 // Helper struct for sorting blocks based on length
444 type lane struct {
445 len uint
446 pos uint
447 }
448
449 type lanes []lane
450
451 func (lns lanes) Len() int { return len(lns) }
452 func (lns lanes) Swap(i, j int) { lns[i], lns[j] = lns[j], lns[i] }
453 func (lns lanes) Less(i, j int) bool { return lns[i].len < lns[j].len }
454
455 // Helper struct for
456 type maskRounds struct {
457 mask uint64
458 rounds uint64
459 }
460
461 func genMask(input [16][]byte) [16]maskRounds {
462
463 // Sort on blocks length small to large
464 var sorted [16]lane
465 for c, inpt := range input {
466 sorted[c] = lane{uint(len(inpt)), uint(c)}
467 }
468 sort.Sort(lanes(sorted[:]))
469
470 // Create mask array including 'rounds' between masks
471 m, round, index := uint64(0xffff), uint64(0), 0
472 var mr [16]maskRounds
473 for _, s := range sorted {
474 if s.len > 0 {
475 if uint64(s.len)>>6 > round {
476 mr[index] = maskRounds{m, (uint64(s.len) >> 6) - round}
477 index++
478 }
479 round = uint64(s.len) >> 6
480 }
481 m = m & ^(1 << uint(s.pos))
482 }
483
484 return mr
485 }
486
487 // TODO: remove function
488 func expandMask(mr [16]maskRounds) []uint64 {
489 size := uint64(0)
490 for _, r := range mr {
491 size += r.rounds
492 }
493 result, index := make([]uint64, size), 0
494 for _, r := range mr {
495 for j := uint64(0); j < r.rounds; j++ {
496 result[index] = r.mask
497 index++
498 }
499 }
500 return result
501 }
502