sha256_avx2.cpp raw
1 // Copyright (c) 2017-2019 The Limenka developers
2 // Distributed under the MIT software license, see the accompanying
3 // file COPYING or http://www.opensource.org/licenses/mit-license.php.
4
5 #ifdef ENABLE_AVX2
6
7 #include <stdint.h>
8 #include <immintrin.h>
9
10 #include <attributes.h>
11 #include <crypto/common.h>
12
13 #if defined(__clang__)
14 #pragma clang attribute push(__attribute__((__target__("avx,avx2"))), apply_to = function)
15 #elif defined(__GNUC__)
16 #pragma GCC target ("avx,avx2")
17 #endif
18
19 namespace sha256d64_avx2 {
20 namespace {
21
22 __m256i inline K(uint32_t x) { return _mm256_set1_epi32(x); }
23
24 __m256i inline Add(__m256i x, __m256i y) { return _mm256_add_epi32(x, y); }
25 __m256i inline Add(__m256i x, __m256i y, __m256i z) { return Add(Add(x, y), z); }
26 __m256i inline Add(__m256i x, __m256i y, __m256i z, __m256i w) { return Add(Add(x, y), Add(z, w)); }
27 __m256i inline Add(__m256i x, __m256i y, __m256i z, __m256i w, __m256i v) { return Add(Add(x, y, z), Add(w, v)); }
28 __m256i inline Inc(__m256i& x, __m256i y) { x = Add(x, y); return x; }
29 __m256i inline Inc(__m256i& x, __m256i y, __m256i z) { x = Add(x, y, z); return x; }
30 __m256i inline Inc(__m256i& x, __m256i y, __m256i z, __m256i w) { x = Add(x, y, z, w); return x; }
31 __m256i inline Xor(__m256i x, __m256i y) { return _mm256_xor_si256(x, y); }
32 __m256i inline Xor(__m256i x, __m256i y, __m256i z) { return Xor(Xor(x, y), z); }
33 __m256i inline Or(__m256i x, __m256i y) { return _mm256_or_si256(x, y); }
34 __m256i inline And(__m256i x, __m256i y) { return _mm256_and_si256(x, y); }
35 __m256i inline ShR(__m256i x, int n) { return _mm256_srli_epi32(x, n); }
36 __m256i inline ShL(__m256i x, int n) { return _mm256_slli_epi32(x, n); }
37
38 __m256i inline Ch(__m256i x, __m256i y, __m256i z) { return Xor(z, And(x, Xor(y, z))); }
39 __m256i inline Maj(__m256i x, __m256i y, __m256i z) { return Or(And(x, y), And(z, Or(x, y))); }
40 __m256i inline Sigma0(__m256i x) { return Xor(Or(ShR(x, 2), ShL(x, 30)), Or(ShR(x, 13), ShL(x, 19)), Or(ShR(x, 22), ShL(x, 10))); }
41 __m256i inline Sigma1(__m256i x) { return Xor(Or(ShR(x, 6), ShL(x, 26)), Or(ShR(x, 11), ShL(x, 21)), Or(ShR(x, 25), ShL(x, 7))); }
42 __m256i inline sigma0(__m256i x) { return Xor(Or(ShR(x, 7), ShL(x, 25)), Or(ShR(x, 18), ShL(x, 14)), ShR(x, 3)); }
43 __m256i inline sigma1(__m256i x) { return Xor(Or(ShR(x, 17), ShL(x, 15)), Or(ShR(x, 19), ShL(x, 13)), ShR(x, 10)); }
44
45 /** One round of SHA-256. */
46 void ALWAYS_INLINE Round(__m256i a, __m256i b, __m256i c, __m256i& d, __m256i e, __m256i f, __m256i g, __m256i& h, __m256i k)
47 {
48 __m256i t1 = Add(h, Sigma1(e), Ch(e, f, g), k);
49 __m256i t2 = Add(Sigma0(a), Maj(a, b, c));
50 d = Add(d, t1);
51 h = Add(t1, t2);
52 }
53
54 __m256i inline Read8(const unsigned char* chunk, int offset) {
55 __m256i ret = _mm256_set_epi32(
56 ReadLE32(chunk + 0 + offset),
57 ReadLE32(chunk + 64 + offset),
58 ReadLE32(chunk + 128 + offset),
59 ReadLE32(chunk + 192 + offset),
60 ReadLE32(chunk + 256 + offset),
61 ReadLE32(chunk + 320 + offset),
62 ReadLE32(chunk + 384 + offset),
63 ReadLE32(chunk + 448 + offset)
64 );
65 return _mm256_shuffle_epi8(ret, _mm256_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL, 0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
66 }
67
68 void inline Write8(unsigned char* out, int offset, __m256i v) {
69 v = _mm256_shuffle_epi8(v, _mm256_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL, 0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
70 WriteLE32(out + 0 + offset, _mm256_extract_epi32(v, 7));
71 WriteLE32(out + 32 + offset, _mm256_extract_epi32(v, 6));
72 WriteLE32(out + 64 + offset, _mm256_extract_epi32(v, 5));
73 WriteLE32(out + 96 + offset, _mm256_extract_epi32(v, 4));
74 WriteLE32(out + 128 + offset, _mm256_extract_epi32(v, 3));
75 WriteLE32(out + 160 + offset, _mm256_extract_epi32(v, 2));
76 WriteLE32(out + 192 + offset, _mm256_extract_epi32(v, 1));
77 WriteLE32(out + 224 + offset, _mm256_extract_epi32(v, 0));
78 }
79
80 }
81
82 void Transform_8way(unsigned char* out, const unsigned char* in)
83 {
84 // Transform 1
85 __m256i a = K(0x6a09e667ul);
86 __m256i b = K(0xbb67ae85ul);
87 __m256i c = K(0x3c6ef372ul);
88 __m256i d = K(0xa54ff53aul);
89 __m256i e = K(0x510e527ful);
90 __m256i f = K(0x9b05688cul);
91 __m256i g = K(0x1f83d9abul);
92 __m256i h = K(0x5be0cd19ul);
93
94 __m256i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
95
96 Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0 = Read8(in, 0)));
97 Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1 = Read8(in, 4)));
98 Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2 = Read8(in, 8)));
99 Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3 = Read8(in, 12)));
100 Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4 = Read8(in, 16)));
101 Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5 = Read8(in, 20)));
102 Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6 = Read8(in, 24)));
103 Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7 = Read8(in, 28)));
104 Round(a, b, c, d, e, f, g, h, Add(K(0xd807aa98ul), w8 = Read8(in, 32)));
105 Round(h, a, b, c, d, e, f, g, Add(K(0x12835b01ul), w9 = Read8(in, 36)));
106 Round(g, h, a, b, c, d, e, f, Add(K(0x243185beul), w10 = Read8(in, 40)));
107 Round(f, g, h, a, b, c, d, e, Add(K(0x550c7dc3ul), w11 = Read8(in, 44)));
108 Round(e, f, g, h, a, b, c, d, Add(K(0x72be5d74ul), w12 = Read8(in, 48)));
109 Round(d, e, f, g, h, a, b, c, Add(K(0x80deb1feul), w13 = Read8(in, 52)));
110 Round(c, d, e, f, g, h, a, b, Add(K(0x9bdc06a7ul), w14 = Read8(in, 56)));
111 Round(b, c, d, e, f, g, h, a, Add(K(0xc19bf174ul), w15 = Read8(in, 60)));
112 Round(a, b, c, d, e, f, g, h, Add(K(0xe49b69c1ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
113 Round(h, a, b, c, d, e, f, g, Add(K(0xefbe4786ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
114 Round(g, h, a, b, c, d, e, f, Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
115 Round(f, g, h, a, b, c, d, e, Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
116 Round(e, f, g, h, a, b, c, d, Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), w13, sigma0(w5))));
117 Round(d, e, f, g, h, a, b, c, Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
118 Round(c, d, e, f, g, h, a, b, Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
119 Round(b, c, d, e, f, g, h, a, Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
120 Round(a, b, c, d, e, f, g, h, Add(K(0x983e5152ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
121 Round(h, a, b, c, d, e, f, g, Add(K(0xa831c66dul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
122 Round(g, h, a, b, c, d, e, f, Add(K(0xb00327c8ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
123 Round(f, g, h, a, b, c, d, e, Add(K(0xbf597fc7ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
124 Round(e, f, g, h, a, b, c, d, Add(K(0xc6e00bf3ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
125 Round(d, e, f, g, h, a, b, c, Add(K(0xd5a79147ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
126 Round(c, d, e, f, g, h, a, b, Add(K(0x06ca6351ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
127 Round(b, c, d, e, f, g, h, a, Add(K(0x14292967ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
128 Round(a, b, c, d, e, f, g, h, Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
129 Round(h, a, b, c, d, e, f, g, Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
130 Round(g, h, a, b, c, d, e, f, Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
131 Round(f, g, h, a, b, c, d, e, Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
132 Round(e, f, g, h, a, b, c, d, Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
133 Round(d, e, f, g, h, a, b, c, Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
134 Round(c, d, e, f, g, h, a, b, Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
135 Round(b, c, d, e, f, g, h, a, Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
136 Round(a, b, c, d, e, f, g, h, Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
137 Round(h, a, b, c, d, e, f, g, Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
138 Round(g, h, a, b, c, d, e, f, Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
139 Round(f, g, h, a, b, c, d, e, Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
140 Round(e, f, g, h, a, b, c, d, Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
141 Round(d, e, f, g, h, a, b, c, Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
142 Round(c, d, e, f, g, h, a, b, Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
143 Round(b, c, d, e, f, g, h, a, Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
144 Round(a, b, c, d, e, f, g, h, Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
145 Round(h, a, b, c, d, e, f, g, Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
146 Round(g, h, a, b, c, d, e, f, Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
147 Round(f, g, h, a, b, c, d, e, Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
148 Round(e, f, g, h, a, b, c, d, Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
149 Round(d, e, f, g, h, a, b, c, Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
150 Round(c, d, e, f, g, h, a, b, Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
151 Round(b, c, d, e, f, g, h, a, Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
152 Round(a, b, c, d, e, f, g, h, Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
153 Round(h, a, b, c, d, e, f, g, Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
154 Round(g, h, a, b, c, d, e, f, Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
155 Round(f, g, h, a, b, c, d, e, Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
156 Round(e, f, g, h, a, b, c, d, Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
157 Round(d, e, f, g, h, a, b, c, Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
158 Round(c, d, e, f, g, h, a, b, Add(K(0xbef9a3f7ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
159 Round(b, c, d, e, f, g, h, a, Add(K(0xc67178f2ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
160
161 a = Add(a, K(0x6a09e667ul));
162 b = Add(b, K(0xbb67ae85ul));
163 c = Add(c, K(0x3c6ef372ul));
164 d = Add(d, K(0xa54ff53aul));
165 e = Add(e, K(0x510e527ful));
166 f = Add(f, K(0x9b05688cul));
167 g = Add(g, K(0x1f83d9abul));
168 h = Add(h, K(0x5be0cd19ul));
169
170 __m256i t0 = a, t1 = b, t2 = c, t3 = d, t4 = e, t5 = f, t6 = g, t7 = h;
171
172 // Transform 2
173 Round(a, b, c, d, e, f, g, h, K(0xc28a2f98ul));
174 Round(h, a, b, c, d, e, f, g, K(0x71374491ul));
175 Round(g, h, a, b, c, d, e, f, K(0xb5c0fbcful));
176 Round(f, g, h, a, b, c, d, e, K(0xe9b5dba5ul));
177 Round(e, f, g, h, a, b, c, d, K(0x3956c25bul));
178 Round(d, e, f, g, h, a, b, c, K(0x59f111f1ul));
179 Round(c, d, e, f, g, h, a, b, K(0x923f82a4ul));
180 Round(b, c, d, e, f, g, h, a, K(0xab1c5ed5ul));
181 Round(a, b, c, d, e, f, g, h, K(0xd807aa98ul));
182 Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
183 Round(g, h, a, b, c, d, e, f, K(0x243185beul));
184 Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
185 Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
186 Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
187 Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
188 Round(b, c, d, e, f, g, h, a, K(0xc19bf374ul));
189 Round(a, b, c, d, e, f, g, h, K(0x649b69c1ul));
190 Round(h, a, b, c, d, e, f, g, K(0xf0fe4786ul));
191 Round(g, h, a, b, c, d, e, f, K(0x0fe1edc6ul));
192 Round(f, g, h, a, b, c, d, e, K(0x240cf254ul));
193 Round(e, f, g, h, a, b, c, d, K(0x4fe9346ful));
194 Round(d, e, f, g, h, a, b, c, K(0x6cc984beul));
195 Round(c, d, e, f, g, h, a, b, K(0x61b9411eul));
196 Round(b, c, d, e, f, g, h, a, K(0x16f988faul));
197 Round(a, b, c, d, e, f, g, h, K(0xf2c65152ul));
198 Round(h, a, b, c, d, e, f, g, K(0xa88e5a6dul));
199 Round(g, h, a, b, c, d, e, f, K(0xb019fc65ul));
200 Round(f, g, h, a, b, c, d, e, K(0xb9d99ec7ul));
201 Round(e, f, g, h, a, b, c, d, K(0x9a1231c3ul));
202 Round(d, e, f, g, h, a, b, c, K(0xe70eeaa0ul));
203 Round(c, d, e, f, g, h, a, b, K(0xfdb1232bul));
204 Round(b, c, d, e, f, g, h, a, K(0xc7353eb0ul));
205 Round(a, b, c, d, e, f, g, h, K(0x3069bad5ul));
206 Round(h, a, b, c, d, e, f, g, K(0xcb976d5ful));
207 Round(g, h, a, b, c, d, e, f, K(0x5a0f118ful));
208 Round(f, g, h, a, b, c, d, e, K(0xdc1eeefdul));
209 Round(e, f, g, h, a, b, c, d, K(0x0a35b689ul));
210 Round(d, e, f, g, h, a, b, c, K(0xde0b7a04ul));
211 Round(c, d, e, f, g, h, a, b, K(0x58f4ca9dul));
212 Round(b, c, d, e, f, g, h, a, K(0xe15d5b16ul));
213 Round(a, b, c, d, e, f, g, h, K(0x007f3e86ul));
214 Round(h, a, b, c, d, e, f, g, K(0x37088980ul));
215 Round(g, h, a, b, c, d, e, f, K(0xa507ea32ul));
216 Round(f, g, h, a, b, c, d, e, K(0x6fab9537ul));
217 Round(e, f, g, h, a, b, c, d, K(0x17406110ul));
218 Round(d, e, f, g, h, a, b, c, K(0x0d8cd6f1ul));
219 Round(c, d, e, f, g, h, a, b, K(0xcdaa3b6dul));
220 Round(b, c, d, e, f, g, h, a, K(0xc0bbbe37ul));
221 Round(a, b, c, d, e, f, g, h, K(0x83613bdaul));
222 Round(h, a, b, c, d, e, f, g, K(0xdb48a363ul));
223 Round(g, h, a, b, c, d, e, f, K(0x0b02e931ul));
224 Round(f, g, h, a, b, c, d, e, K(0x6fd15ca7ul));
225 Round(e, f, g, h, a, b, c, d, K(0x521afacaul));
226 Round(d, e, f, g, h, a, b, c, K(0x31338431ul));
227 Round(c, d, e, f, g, h, a, b, K(0x6ed41a95ul));
228 Round(b, c, d, e, f, g, h, a, K(0x6d437890ul));
229 Round(a, b, c, d, e, f, g, h, K(0xc39c91f2ul));
230 Round(h, a, b, c, d, e, f, g, K(0x9eccabbdul));
231 Round(g, h, a, b, c, d, e, f, K(0xb5c9a0e6ul));
232 Round(f, g, h, a, b, c, d, e, K(0x532fb63cul));
233 Round(e, f, g, h, a, b, c, d, K(0xd2c741c6ul));
234 Round(d, e, f, g, h, a, b, c, K(0x07237ea3ul));
235 Round(c, d, e, f, g, h, a, b, K(0xa4954b68ul));
236 Round(b, c, d, e, f, g, h, a, K(0x4c191d76ul));
237
238 w0 = Add(t0, a);
239 w1 = Add(t1, b);
240 w2 = Add(t2, c);
241 w3 = Add(t3, d);
242 w4 = Add(t4, e);
243 w5 = Add(t5, f);
244 w6 = Add(t6, g);
245 w7 = Add(t7, h);
246
247 // Transform 3
248 a = K(0x6a09e667ul);
249 b = K(0xbb67ae85ul);
250 c = K(0x3c6ef372ul);
251 d = K(0xa54ff53aul);
252 e = K(0x510e527ful);
253 f = K(0x9b05688cul);
254 g = K(0x1f83d9abul);
255 h = K(0x5be0cd19ul);
256
257 Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0));
258 Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1));
259 Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2));
260 Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3));
261 Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4));
262 Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5));
263 Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6));
264 Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7));
265 Round(a, b, c, d, e, f, g, h, K(0x5807aa98ul));
266 Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
267 Round(g, h, a, b, c, d, e, f, K(0x243185beul));
268 Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
269 Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
270 Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
271 Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
272 Round(b, c, d, e, f, g, h, a, K(0xc19bf274ul));
273 Round(a, b, c, d, e, f, g, h, Add(K(0xe49b69c1ul), Inc(w0, sigma0(w1))));
274 Round(h, a, b, c, d, e, f, g, Add(K(0xefbe4786ul), Inc(w1, K(0xa00000ul), sigma0(w2))));
275 Round(g, h, a, b, c, d, e, f, Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), sigma0(w3))));
276 Round(f, g, h, a, b, c, d, e, Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), sigma0(w4))));
277 Round(e, f, g, h, a, b, c, d, Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), sigma0(w5))));
278 Round(d, e, f, g, h, a, b, c, Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), sigma0(w6))));
279 Round(c, d, e, f, g, h, a, b, Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), K(0x100ul), sigma0(w7))));
280 Round(b, c, d, e, f, g, h, a, Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, K(0x11002000ul))));
281 Round(a, b, c, d, e, f, g, h, Add(K(0x983e5152ul), w8 = Add(K(0x80000000ul), sigma1(w6), w1)));
282 Round(h, a, b, c, d, e, f, g, Add(K(0xa831c66dul), w9 = Add(sigma1(w7), w2)));
283 Round(g, h, a, b, c, d, e, f, Add(K(0xb00327c8ul), w10 = Add(sigma1(w8), w3)));
284 Round(f, g, h, a, b, c, d, e, Add(K(0xbf597fc7ul), w11 = Add(sigma1(w9), w4)));
285 Round(e, f, g, h, a, b, c, d, Add(K(0xc6e00bf3ul), w12 = Add(sigma1(w10), w5)));
286 Round(d, e, f, g, h, a, b, c, Add(K(0xd5a79147ul), w13 = Add(sigma1(w11), w6)));
287 Round(c, d, e, f, g, h, a, b, Add(K(0x06ca6351ul), w14 = Add(sigma1(w12), w7, K(0x400022ul))));
288 Round(b, c, d, e, f, g, h, a, Add(K(0x14292967ul), w15 = Add(K(0x100ul), sigma1(w13), w8, sigma0(w0))));
289 Round(a, b, c, d, e, f, g, h, Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
290 Round(h, a, b, c, d, e, f, g, Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
291 Round(g, h, a, b, c, d, e, f, Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
292 Round(f, g, h, a, b, c, d, e, Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
293 Round(e, f, g, h, a, b, c, d, Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
294 Round(d, e, f, g, h, a, b, c, Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
295 Round(c, d, e, f, g, h, a, b, Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
296 Round(b, c, d, e, f, g, h, a, Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
297 Round(a, b, c, d, e, f, g, h, Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
298 Round(h, a, b, c, d, e, f, g, Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
299 Round(g, h, a, b, c, d, e, f, Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
300 Round(f, g, h, a, b, c, d, e, Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
301 Round(e, f, g, h, a, b, c, d, Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
302 Round(d, e, f, g, h, a, b, c, Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
303 Round(c, d, e, f, g, h, a, b, Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
304 Round(b, c, d, e, f, g, h, a, Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
305 Round(a, b, c, d, e, f, g, h, Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
306 Round(h, a, b, c, d, e, f, g, Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
307 Round(g, h, a, b, c, d, e, f, Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
308 Round(f, g, h, a, b, c, d, e, Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
309 Round(e, f, g, h, a, b, c, d, Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
310 Round(d, e, f, g, h, a, b, c, Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
311 Round(c, d, e, f, g, h, a, b, Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
312 Round(b, c, d, e, f, g, h, a, Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
313 Round(a, b, c, d, e, f, g, h, Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
314 Round(h, a, b, c, d, e, f, g, Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
315 Round(g, h, a, b, c, d, e, f, Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
316 Round(f, g, h, a, b, c, d, e, Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
317 Round(e, f, g, h, a, b, c, d, Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
318 Round(d, e, f, g, h, a, b, c, Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
319 Round(c, d, e, f, g, h, a, b, Add(K(0xbef9a3f7ul), w14, sigma1(w12), w7, sigma0(w15)));
320 Round(b, c, d, e, f, g, h, a, Add(K(0xc67178f2ul), w15, sigma1(w13), w8, sigma0(w0)));
321
322 // Output
323 Write8(out, 0, Add(a, K(0x6a09e667ul)));
324 Write8(out, 4, Add(b, K(0xbb67ae85ul)));
325 Write8(out, 8, Add(c, K(0x3c6ef372ul)));
326 Write8(out, 12, Add(d, K(0xa54ff53aul)));
327 Write8(out, 16, Add(e, K(0x510e527ful)));
328 Write8(out, 20, Add(f, K(0x9b05688cul)));
329 Write8(out, 24, Add(g, K(0x1f83d9abul)));
330 Write8(out, 28, Add(h, K(0x5be0cd19ul)));
331 }
332
333 }
334
335 #if defined(__clang__)
336 #pragma clang attribute pop
337 #endif
338
339 #endif
340