sha256_sse41.cpp raw
1 // Copyright (c) 2018-present The Bitcoin Core 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_SSE41
6
7 #include <cstdint>
8 #include <immintrin.h>
9
10 #include <attributes.h>
11 #include <crypto/common.h>
12
13 namespace sha256d64_sse41 {
14 namespace {
15
16 __m128i inline K(uint32_t x) { return _mm_set1_epi32(x); }
17
18 __m128i inline Add(__m128i x, __m128i y) { return _mm_add_epi32(x, y); }
19 __m128i inline Add(__m128i x, __m128i y, __m128i z) { return Add(Add(x, y), z); }
20 __m128i inline Add(__m128i x, __m128i y, __m128i z, __m128i w) { return Add(Add(x, y), Add(z, w)); }
21 __m128i inline Add(__m128i x, __m128i y, __m128i z, __m128i w, __m128i v) { return Add(Add(x, y, z), Add(w, v)); }
22 __m128i inline Inc(__m128i& x, __m128i y) { x = Add(x, y); return x; }
23 __m128i inline Inc(__m128i& x, __m128i y, __m128i z) { x = Add(x, y, z); return x; }
24 __m128i inline Inc(__m128i& x, __m128i y, __m128i z, __m128i w) { x = Add(x, y, z, w); return x; }
25 __m128i inline Xor(__m128i x, __m128i y) { return _mm_xor_si128(x, y); }
26 __m128i inline Xor(__m128i x, __m128i y, __m128i z) { return Xor(Xor(x, y), z); }
27 __m128i inline Or(__m128i x, __m128i y) { return _mm_or_si128(x, y); }
28 __m128i inline And(__m128i x, __m128i y) { return _mm_and_si128(x, y); }
29 __m128i inline ShR(__m128i x, int n) { return _mm_srli_epi32(x, n); }
30 __m128i inline ShL(__m128i x, int n) { return _mm_slli_epi32(x, n); }
31
32 __m128i inline Ch(__m128i x, __m128i y, __m128i z) { return Xor(z, And(x, Xor(y, z))); }
33 __m128i inline Maj(__m128i x, __m128i y, __m128i z) { return Or(And(x, y), And(z, Or(x, y))); }
34 __m128i inline Sigma0(__m128i x) { return Xor(Or(ShR(x, 2), ShL(x, 30)), Or(ShR(x, 13), ShL(x, 19)), Or(ShR(x, 22), ShL(x, 10))); }
35 __m128i inline Sigma1(__m128i x) { return Xor(Or(ShR(x, 6), ShL(x, 26)), Or(ShR(x, 11), ShL(x, 21)), Or(ShR(x, 25), ShL(x, 7))); }
36 __m128i inline sigma0(__m128i x) { return Xor(Or(ShR(x, 7), ShL(x, 25)), Or(ShR(x, 18), ShL(x, 14)), ShR(x, 3)); }
37 __m128i inline sigma1(__m128i x) { return Xor(Or(ShR(x, 17), ShL(x, 15)), Or(ShR(x, 19), ShL(x, 13)), ShR(x, 10)); }
38
39 /** One round of SHA-256. */
40 void ALWAYS_INLINE Round(__m128i a, __m128i b, __m128i c, __m128i& d, __m128i e, __m128i f, __m128i g, __m128i& h, __m128i k)
41 {
42 __m128i t1 = Add(h, Sigma1(e), Ch(e, f, g), k);
43 __m128i t2 = Add(Sigma0(a), Maj(a, b, c));
44 d = Add(d, t1);
45 h = Add(t1, t2);
46 }
47
48 __m128i inline Read4(const unsigned char* chunk, int offset) {
49 __m128i ret = _mm_set_epi32(
50 ReadLE32(chunk + 0 + offset),
51 ReadLE32(chunk + 64 + offset),
52 ReadLE32(chunk + 128 + offset),
53 ReadLE32(chunk + 192 + offset)
54 );
55 return _mm_shuffle_epi8(ret, _mm_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
56 }
57
58 void inline Write4(unsigned char* out, int offset, __m128i v) {
59 v = _mm_shuffle_epi8(v, _mm_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
60 WriteLE32(out + 0 + offset, _mm_extract_epi32(v, 3));
61 WriteLE32(out + 32 + offset, _mm_extract_epi32(v, 2));
62 WriteLE32(out + 64 + offset, _mm_extract_epi32(v, 1));
63 WriteLE32(out + 96 + offset, _mm_extract_epi32(v, 0));
64 }
65
66 }
67
68 void Transform_4way(unsigned char* out, const unsigned char* in)
69 {
70 // Transform 1
71 __m128i a = K(0x6a09e667ul);
72 __m128i b = K(0xbb67ae85ul);
73 __m128i c = K(0x3c6ef372ul);
74 __m128i d = K(0xa54ff53aul);
75 __m128i e = K(0x510e527ful);
76 __m128i f = K(0x9b05688cul);
77 __m128i g = K(0x1f83d9abul);
78 __m128i h = K(0x5be0cd19ul);
79
80 __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
81
82 Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0 = Read4(in, 0)));
83 Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1 = Read4(in, 4)));
84 Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2 = Read4(in, 8)));
85 Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3 = Read4(in, 12)));
86 Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4 = Read4(in, 16)));
87 Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5 = Read4(in, 20)));
88 Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6 = Read4(in, 24)));
89 Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7 = Read4(in, 28)));
90 Round(a, b, c, d, e, f, g, h, Add(K(0xd807aa98ul), w8 = Read4(in, 32)));
91 Round(h, a, b, c, d, e, f, g, Add(K(0x12835b01ul), w9 = Read4(in, 36)));
92 Round(g, h, a, b, c, d, e, f, Add(K(0x243185beul), w10 = Read4(in, 40)));
93 Round(f, g, h, a, b, c, d, e, Add(K(0x550c7dc3ul), w11 = Read4(in, 44)));
94 Round(e, f, g, h, a, b, c, d, Add(K(0x72be5d74ul), w12 = Read4(in, 48)));
95 Round(d, e, f, g, h, a, b, c, Add(K(0x80deb1feul), w13 = Read4(in, 52)));
96 Round(c, d, e, f, g, h, a, b, Add(K(0x9bdc06a7ul), w14 = Read4(in, 56)));
97 Round(b, c, d, e, f, g, h, a, Add(K(0xc19bf174ul), w15 = Read4(in, 60)));
98 Round(a, b, c, d, e, f, g, h, Add(K(0xe49b69c1ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
99 Round(h, a, b, c, d, e, f, g, Add(K(0xefbe4786ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
100 Round(g, h, a, b, c, d, e, f, Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
101 Round(f, g, h, a, b, c, d, e, Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
102 Round(e, f, g, h, a, b, c, d, Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), w13, sigma0(w5))));
103 Round(d, e, f, g, h, a, b, c, Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
104 Round(c, d, e, f, g, h, a, b, Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
105 Round(b, c, d, e, f, g, h, a, Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
106 Round(a, b, c, d, e, f, g, h, Add(K(0x983e5152ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
107 Round(h, a, b, c, d, e, f, g, Add(K(0xa831c66dul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
108 Round(g, h, a, b, c, d, e, f, Add(K(0xb00327c8ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
109 Round(f, g, h, a, b, c, d, e, Add(K(0xbf597fc7ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
110 Round(e, f, g, h, a, b, c, d, Add(K(0xc6e00bf3ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
111 Round(d, e, f, g, h, a, b, c, Add(K(0xd5a79147ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
112 Round(c, d, e, f, g, h, a, b, Add(K(0x06ca6351ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
113 Round(b, c, d, e, f, g, h, a, Add(K(0x14292967ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
114 Round(a, b, c, d, e, f, g, h, Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
115 Round(h, a, b, c, d, e, f, g, Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
116 Round(g, h, a, b, c, d, e, f, Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
117 Round(f, g, h, a, b, c, d, e, Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
118 Round(e, f, g, h, a, b, c, d, Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
119 Round(d, e, f, g, h, a, b, c, Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
120 Round(c, d, e, f, g, h, a, b, Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
121 Round(b, c, d, e, f, g, h, a, Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
122 Round(a, b, c, d, e, f, g, h, Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
123 Round(h, a, b, c, d, e, f, g, Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
124 Round(g, h, a, b, c, d, e, f, Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
125 Round(f, g, h, a, b, c, d, e, Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
126 Round(e, f, g, h, a, b, c, d, Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
127 Round(d, e, f, g, h, a, b, c, Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
128 Round(c, d, e, f, g, h, a, b, Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
129 Round(b, c, d, e, f, g, h, a, Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
130 Round(a, b, c, d, e, f, g, h, Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
131 Round(h, a, b, c, d, e, f, g, Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
132 Round(g, h, a, b, c, d, e, f, Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
133 Round(f, g, h, a, b, c, d, e, Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
134 Round(e, f, g, h, a, b, c, d, Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
135 Round(d, e, f, g, h, a, b, c, Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
136 Round(c, d, e, f, g, h, a, b, Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
137 Round(b, c, d, e, f, g, h, a, Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
138 Round(a, b, c, d, e, f, g, h, Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
139 Round(h, a, b, c, d, e, f, g, Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
140 Round(g, h, a, b, c, d, e, f, Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
141 Round(f, g, h, a, b, c, d, e, Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
142 Round(e, f, g, h, a, b, c, d, Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
143 Round(d, e, f, g, h, a, b, c, Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
144 Round(c, d, e, f, g, h, a, b, Add(K(0xbef9a3f7ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
145 Round(b, c, d, e, f, g, h, a, Add(K(0xc67178f2ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
146
147 a = Add(a, K(0x6a09e667ul));
148 b = Add(b, K(0xbb67ae85ul));
149 c = Add(c, K(0x3c6ef372ul));
150 d = Add(d, K(0xa54ff53aul));
151 e = Add(e, K(0x510e527ful));
152 f = Add(f, K(0x9b05688cul));
153 g = Add(g, K(0x1f83d9abul));
154 h = Add(h, K(0x5be0cd19ul));
155
156 __m128i t0 = a, t1 = b, t2 = c, t3 = d, t4 = e, t5 = f, t6 = g, t7 = h;
157
158 // Transform 2
159 Round(a, b, c, d, e, f, g, h, K(0xc28a2f98ul));
160 Round(h, a, b, c, d, e, f, g, K(0x71374491ul));
161 Round(g, h, a, b, c, d, e, f, K(0xb5c0fbcful));
162 Round(f, g, h, a, b, c, d, e, K(0xe9b5dba5ul));
163 Round(e, f, g, h, a, b, c, d, K(0x3956c25bul));
164 Round(d, e, f, g, h, a, b, c, K(0x59f111f1ul));
165 Round(c, d, e, f, g, h, a, b, K(0x923f82a4ul));
166 Round(b, c, d, e, f, g, h, a, K(0xab1c5ed5ul));
167 Round(a, b, c, d, e, f, g, h, K(0xd807aa98ul));
168 Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
169 Round(g, h, a, b, c, d, e, f, K(0x243185beul));
170 Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
171 Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
172 Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
173 Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
174 Round(b, c, d, e, f, g, h, a, K(0xc19bf374ul));
175 Round(a, b, c, d, e, f, g, h, K(0x649b69c1ul));
176 Round(h, a, b, c, d, e, f, g, K(0xf0fe4786ul));
177 Round(g, h, a, b, c, d, e, f, K(0x0fe1edc6ul));
178 Round(f, g, h, a, b, c, d, e, K(0x240cf254ul));
179 Round(e, f, g, h, a, b, c, d, K(0x4fe9346ful));
180 Round(d, e, f, g, h, a, b, c, K(0x6cc984beul));
181 Round(c, d, e, f, g, h, a, b, K(0x61b9411eul));
182 Round(b, c, d, e, f, g, h, a, K(0x16f988faul));
183 Round(a, b, c, d, e, f, g, h, K(0xf2c65152ul));
184 Round(h, a, b, c, d, e, f, g, K(0xa88e5a6dul));
185 Round(g, h, a, b, c, d, e, f, K(0xb019fc65ul));
186 Round(f, g, h, a, b, c, d, e, K(0xb9d99ec7ul));
187 Round(e, f, g, h, a, b, c, d, K(0x9a1231c3ul));
188 Round(d, e, f, g, h, a, b, c, K(0xe70eeaa0ul));
189 Round(c, d, e, f, g, h, a, b, K(0xfdb1232bul));
190 Round(b, c, d, e, f, g, h, a, K(0xc7353eb0ul));
191 Round(a, b, c, d, e, f, g, h, K(0x3069bad5ul));
192 Round(h, a, b, c, d, e, f, g, K(0xcb976d5ful));
193 Round(g, h, a, b, c, d, e, f, K(0x5a0f118ful));
194 Round(f, g, h, a, b, c, d, e, K(0xdc1eeefdul));
195 Round(e, f, g, h, a, b, c, d, K(0x0a35b689ul));
196 Round(d, e, f, g, h, a, b, c, K(0xde0b7a04ul));
197 Round(c, d, e, f, g, h, a, b, K(0x58f4ca9dul));
198 Round(b, c, d, e, f, g, h, a, K(0xe15d5b16ul));
199 Round(a, b, c, d, e, f, g, h, K(0x007f3e86ul));
200 Round(h, a, b, c, d, e, f, g, K(0x37088980ul));
201 Round(g, h, a, b, c, d, e, f, K(0xa507ea32ul));
202 Round(f, g, h, a, b, c, d, e, K(0x6fab9537ul));
203 Round(e, f, g, h, a, b, c, d, K(0x17406110ul));
204 Round(d, e, f, g, h, a, b, c, K(0x0d8cd6f1ul));
205 Round(c, d, e, f, g, h, a, b, K(0xcdaa3b6dul));
206 Round(b, c, d, e, f, g, h, a, K(0xc0bbbe37ul));
207 Round(a, b, c, d, e, f, g, h, K(0x83613bdaul));
208 Round(h, a, b, c, d, e, f, g, K(0xdb48a363ul));
209 Round(g, h, a, b, c, d, e, f, K(0x0b02e931ul));
210 Round(f, g, h, a, b, c, d, e, K(0x6fd15ca7ul));
211 Round(e, f, g, h, a, b, c, d, K(0x521afacaul));
212 Round(d, e, f, g, h, a, b, c, K(0x31338431ul));
213 Round(c, d, e, f, g, h, a, b, K(0x6ed41a95ul));
214 Round(b, c, d, e, f, g, h, a, K(0x6d437890ul));
215 Round(a, b, c, d, e, f, g, h, K(0xc39c91f2ul));
216 Round(h, a, b, c, d, e, f, g, K(0x9eccabbdul));
217 Round(g, h, a, b, c, d, e, f, K(0xb5c9a0e6ul));
218 Round(f, g, h, a, b, c, d, e, K(0x532fb63cul));
219 Round(e, f, g, h, a, b, c, d, K(0xd2c741c6ul));
220 Round(d, e, f, g, h, a, b, c, K(0x07237ea3ul));
221 Round(c, d, e, f, g, h, a, b, K(0xa4954b68ul));
222 Round(b, c, d, e, f, g, h, a, K(0x4c191d76ul));
223
224 w0 = Add(t0, a);
225 w1 = Add(t1, b);
226 w2 = Add(t2, c);
227 w3 = Add(t3, d);
228 w4 = Add(t4, e);
229 w5 = Add(t5, f);
230 w6 = Add(t6, g);
231 w7 = Add(t7, h);
232
233 // Transform 3
234 a = K(0x6a09e667ul);
235 b = K(0xbb67ae85ul);
236 c = K(0x3c6ef372ul);
237 d = K(0xa54ff53aul);
238 e = K(0x510e527ful);
239 f = K(0x9b05688cul);
240 g = K(0x1f83d9abul);
241 h = K(0x5be0cd19ul);
242
243 Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0));
244 Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1));
245 Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2));
246 Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3));
247 Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4));
248 Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5));
249 Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6));
250 Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7));
251 Round(a, b, c, d, e, f, g, h, K(0x5807aa98ul));
252 Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
253 Round(g, h, a, b, c, d, e, f, K(0x243185beul));
254 Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
255 Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
256 Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
257 Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
258 Round(b, c, d, e, f, g, h, a, K(0xc19bf274ul));
259 Round(a, b, c, d, e, f, g, h, Add(K(0xe49b69c1ul), Inc(w0, sigma0(w1))));
260 Round(h, a, b, c, d, e, f, g, Add(K(0xefbe4786ul), Inc(w1, K(0xa00000ul), sigma0(w2))));
261 Round(g, h, a, b, c, d, e, f, Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), sigma0(w3))));
262 Round(f, g, h, a, b, c, d, e, Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), sigma0(w4))));
263 Round(e, f, g, h, a, b, c, d, Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), sigma0(w5))));
264 Round(d, e, f, g, h, a, b, c, Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), sigma0(w6))));
265 Round(c, d, e, f, g, h, a, b, Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), K(0x100ul), sigma0(w7))));
266 Round(b, c, d, e, f, g, h, a, Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, K(0x11002000ul))));
267 Round(a, b, c, d, e, f, g, h, Add(K(0x983e5152ul), w8 = Add(K(0x80000000ul), sigma1(w6), w1)));
268 Round(h, a, b, c, d, e, f, g, Add(K(0xa831c66dul), w9 = Add(sigma1(w7), w2)));
269 Round(g, h, a, b, c, d, e, f, Add(K(0xb00327c8ul), w10 = Add(sigma1(w8), w3)));
270 Round(f, g, h, a, b, c, d, e, Add(K(0xbf597fc7ul), w11 = Add(sigma1(w9), w4)));
271 Round(e, f, g, h, a, b, c, d, Add(K(0xc6e00bf3ul), w12 = Add(sigma1(w10), w5)));
272 Round(d, e, f, g, h, a, b, c, Add(K(0xd5a79147ul), w13 = Add(sigma1(w11), w6)));
273 Round(c, d, e, f, g, h, a, b, Add(K(0x06ca6351ul), w14 = Add(sigma1(w12), w7, K(0x400022ul))));
274 Round(b, c, d, e, f, g, h, a, Add(K(0x14292967ul), w15 = Add(K(0x100ul), sigma1(w13), w8, sigma0(w0))));
275 Round(a, b, c, d, e, f, g, h, Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
276 Round(h, a, b, c, d, e, f, g, Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
277 Round(g, h, a, b, c, d, e, f, Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
278 Round(f, g, h, a, b, c, d, e, Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
279 Round(e, f, g, h, a, b, c, d, Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
280 Round(d, e, f, g, h, a, b, c, Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
281 Round(c, d, e, f, g, h, a, b, Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
282 Round(b, c, d, e, f, g, h, a, Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
283 Round(a, b, c, d, e, f, g, h, Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
284 Round(h, a, b, c, d, e, f, g, Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
285 Round(g, h, a, b, c, d, e, f, Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
286 Round(f, g, h, a, b, c, d, e, Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
287 Round(e, f, g, h, a, b, c, d, Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
288 Round(d, e, f, g, h, a, b, c, Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
289 Round(c, d, e, f, g, h, a, b, Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
290 Round(b, c, d, e, f, g, h, a, Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
291 Round(a, b, c, d, e, f, g, h, Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
292 Round(h, a, b, c, d, e, f, g, Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
293 Round(g, h, a, b, c, d, e, f, Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
294 Round(f, g, h, a, b, c, d, e, Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
295 Round(e, f, g, h, a, b, c, d, Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
296 Round(d, e, f, g, h, a, b, c, Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
297 Round(c, d, e, f, g, h, a, b, Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
298 Round(b, c, d, e, f, g, h, a, Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
299 Round(a, b, c, d, e, f, g, h, Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
300 Round(h, a, b, c, d, e, f, g, Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
301 Round(g, h, a, b, c, d, e, f, Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
302 Round(f, g, h, a, b, c, d, e, Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
303 Round(e, f, g, h, a, b, c, d, Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
304 Round(d, e, f, g, h, a, b, c, Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
305 Round(c, d, e, f, g, h, a, b, Add(K(0xbef9a3f7ul), w14, sigma1(w12), w7, sigma0(w15)));
306 Round(b, c, d, e, f, g, h, a, Add(K(0xc67178f2ul), w15, sigma1(w13), w8, sigma0(w0)));
307
308 // Output
309 Write4(out, 0, Add(a, K(0x6a09e667ul)));
310 Write4(out, 4, Add(b, K(0xbb67ae85ul)));
311 Write4(out, 8, Add(c, K(0x3c6ef372ul)));
312 Write4(out, 12, Add(d, K(0xa54ff53aul)));
313 Write4(out, 16, Add(e, K(0x510e527ful)));
314 Write4(out, 20, Add(f, K(0x9b05688cul)));
315 Write4(out, 24, Add(g, K(0x1f83d9abul)));
316 Write4(out, 28, Add(h, K(0x5be0cd19ul)));
317 }
318
319 }
320
321 #endif
322