Bitcoin ABC  0.22.12
P2P Digital Currency
sha256_sse41.cpp
Go to the documentation of this file.
1 #ifdef ENABLE_SSE41
2 
3 #include <cstdint>
4 #include <immintrin.h>
5 
6 #include <crypto/common.h>
7 
8 namespace sha256d64_sse41 {
9 namespace {
10 
11  __m128i inline K(uint32_t x) { return _mm_set1_epi32(x); }
12 
13  __m128i inline Add(__m128i x, __m128i y) { return _mm_add_epi32(x, y); }
14  __m128i inline Add(__m128i x, __m128i y, __m128i z) {
15  return Add(Add(x, y), z);
16  }
17  __m128i inline Add(__m128i x, __m128i y, __m128i z, __m128i w) {
18  return Add(Add(x, y), Add(z, w));
19  }
20  __m128i inline Add(__m128i x, __m128i y, __m128i z, __m128i w, __m128i v) {
21  return Add(Add(x, y, z), Add(w, v));
22  }
23  __m128i inline Inc(__m128i &x, __m128i y) {
24  x = Add(x, y);
25  return x;
26  }
27  __m128i inline Inc(__m128i &x, __m128i y, __m128i z) {
28  x = Add(x, y, z);
29  return x;
30  }
31  __m128i inline Inc(__m128i &x, __m128i y, __m128i z, __m128i w) {
32  x = Add(x, y, z, w);
33  return x;
34  }
35  __m128i inline Xor(__m128i x, __m128i y) { return _mm_xor_si128(x, y); }
36  __m128i inline Xor(__m128i x, __m128i y, __m128i z) {
37  return Xor(Xor(x, y), z);
38  }
39  __m128i inline Or(__m128i x, __m128i y) { return _mm_or_si128(x, y); }
40  __m128i inline And(__m128i x, __m128i y) { return _mm_and_si128(x, y); }
41  __m128i inline ShR(__m128i x, int n) { return _mm_srli_epi32(x, n); }
42  __m128i inline ShL(__m128i x, int n) { return _mm_slli_epi32(x, n); }
43 
44  __m128i inline Ch(__m128i x, __m128i y, __m128i z) {
45  return Xor(z, And(x, Xor(y, z)));
46  }
47  __m128i inline Maj(__m128i x, __m128i y, __m128i z) {
48  return Or(And(x, y), And(z, Or(x, y)));
49  }
50  __m128i inline Sigma0(__m128i x) {
51  return Xor(Or(ShR(x, 2), ShL(x, 30)), Or(ShR(x, 13), ShL(x, 19)),
52  Or(ShR(x, 22), ShL(x, 10)));
53  }
54  __m128i inline Sigma1(__m128i x) {
55  return Xor(Or(ShR(x, 6), ShL(x, 26)), Or(ShR(x, 11), ShL(x, 21)),
56  Or(ShR(x, 25), ShL(x, 7)));
57  }
58  __m128i inline sigma0(__m128i x) {
59  return Xor(Or(ShR(x, 7), ShL(x, 25)), Or(ShR(x, 18), ShL(x, 14)),
60  ShR(x, 3));
61  }
62  __m128i inline sigma1(__m128i x) {
63  return Xor(Or(ShR(x, 17), ShL(x, 15)), Or(ShR(x, 19), ShL(x, 13)),
64  ShR(x, 10));
65  }
66 
68  inline void __attribute__((always_inline))
69  Round(__m128i a, __m128i b, __m128i c, __m128i &d, __m128i e, __m128i f,
70  __m128i g, __m128i &h, __m128i k) {
71  __m128i t1 = Add(h, Sigma1(e), Ch(e, f, g), k);
72  __m128i t2 = Add(Sigma0(a), Maj(a, b, c));
73  d = Add(d, t1);
74  h = Add(t1, t2);
75  }
76 
77  __m128i inline Read4(const uint8_t *chunk, int offset) {
78  __m128i ret = _mm_set_epi32(
79  ReadLE32(chunk + 0 + offset), ReadLE32(chunk + 64 + offset),
80  ReadLE32(chunk + 128 + offset), ReadLE32(chunk + 192 + offset));
81  return _mm_shuffle_epi8(ret, _mm_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL,
82  0x04050607UL, 0x00010203UL));
83  }
84 
85  inline void Write4(uint8_t *out, int offset, __m128i v) {
86  v = _mm_shuffle_epi8(v, _mm_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL,
87  0x04050607UL, 0x00010203UL));
88  WriteLE32(out + 0 + offset, _mm_extract_epi32(v, 3));
89  WriteLE32(out + 32 + offset, _mm_extract_epi32(v, 2));
90  WriteLE32(out + 64 + offset, _mm_extract_epi32(v, 1));
91  WriteLE32(out + 96 + offset, _mm_extract_epi32(v, 0));
92  }
93 } // namespace
94 
95 void Transform_4way(uint8_t *out, const uint8_t *in) {
96  // Transform 1
97  __m128i a = K(0x6a09e667ul);
98  __m128i b = K(0xbb67ae85ul);
99  __m128i c = K(0x3c6ef372ul);
100  __m128i d = K(0xa54ff53aul);
101  __m128i e = K(0x510e527ful);
102  __m128i f = K(0x9b05688cul);
103  __m128i g = K(0x1f83d9abul);
104  __m128i h = K(0x5be0cd19ul);
105 
106  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14,
107  w15;
108 
109  Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0 = Read4(in, 0)));
110  Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1 = Read4(in, 4)));
111  Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2 = Read4(in, 8)));
112  Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3 = Read4(in, 12)));
113  Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4 = Read4(in, 16)));
114  Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5 = Read4(in, 20)));
115  Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6 = Read4(in, 24)));
116  Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7 = Read4(in, 28)));
117  Round(a, b, c, d, e, f, g, h, Add(K(0xd807aa98ul), w8 = Read4(in, 32)));
118  Round(h, a, b, c, d, e, f, g, Add(K(0x12835b01ul), w9 = Read4(in, 36)));
119  Round(g, h, a, b, c, d, e, f, Add(K(0x243185beul), w10 = Read4(in, 40)));
120  Round(f, g, h, a, b, c, d, e, Add(K(0x550c7dc3ul), w11 = Read4(in, 44)));
121  Round(e, f, g, h, a, b, c, d, Add(K(0x72be5d74ul), w12 = Read4(in, 48)));
122  Round(d, e, f, g, h, a, b, c, Add(K(0x80deb1feul), w13 = Read4(in, 52)));
123  Round(c, d, e, f, g, h, a, b, Add(K(0x9bdc06a7ul), w14 = Read4(in, 56)));
124  Round(b, c, d, e, f, g, h, a, Add(K(0xc19bf174ul), w15 = Read4(in, 60)));
125  Round(a, b, c, d, e, f, g, h,
126  Add(K(0xe49b69c1ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
127  Round(h, a, b, c, d, e, f, g,
128  Add(K(0xefbe4786ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
129  Round(g, h, a, b, c, d, e, f,
130  Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
131  Round(f, g, h, a, b, c, d, e,
132  Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
133  Round(e, f, g, h, a, b, c, d,
134  Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), w13, sigma0(w5))));
135  Round(d, e, f, g, h, a, b, c,
136  Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
137  Round(c, d, e, f, g, h, a, b,
138  Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
139  Round(b, c, d, e, f, g, h, a,
140  Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
141  Round(a, b, c, d, e, f, g, h,
142  Add(K(0x983e5152ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
143  Round(h, a, b, c, d, e, f, g,
144  Add(K(0xa831c66dul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
145  Round(g, h, a, b, c, d, e, f,
146  Add(K(0xb00327c8ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
147  Round(f, g, h, a, b, c, d, e,
148  Add(K(0xbf597fc7ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
149  Round(e, f, g, h, a, b, c, d,
150  Add(K(0xc6e00bf3ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
151  Round(d, e, f, g, h, a, b, c,
152  Add(K(0xd5a79147ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
153  Round(c, d, e, f, g, h, a, b,
154  Add(K(0x06ca6351ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
155  Round(b, c, d, e, f, g, h, a,
156  Add(K(0x14292967ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
157  Round(a, b, c, d, e, f, g, h,
158  Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
159  Round(h, a, b, c, d, e, f, g,
160  Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
161  Round(g, h, a, b, c, d, e, f,
162  Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
163  Round(f, g, h, a, b, c, d, e,
164  Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
165  Round(e, f, g, h, a, b, c, d,
166  Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
167  Round(d, e, f, g, h, a, b, c,
168  Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
169  Round(c, d, e, f, g, h, a, b,
170  Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
171  Round(b, c, d, e, f, g, h, a,
172  Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
173  Round(a, b, c, d, e, f, g, h,
174  Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
175  Round(h, a, b, c, d, e, f, g,
176  Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
177  Round(g, h, a, b, c, d, e, f,
178  Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
179  Round(f, g, h, a, b, c, d, e,
180  Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
181  Round(e, f, g, h, a, b, c, d,
182  Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
183  Round(d, e, f, g, h, a, b, c,
184  Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
185  Round(c, d, e, f, g, h, a, b,
186  Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
187  Round(b, c, d, e, f, g, h, a,
188  Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
189  Round(a, b, c, d, e, f, g, h,
190  Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
191  Round(h, a, b, c, d, e, f, g,
192  Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
193  Round(g, h, a, b, c, d, e, f,
194  Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
195  Round(f, g, h, a, b, c, d, e,
196  Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
197  Round(e, f, g, h, a, b, c, d,
198  Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
199  Round(d, e, f, g, h, a, b, c,
200  Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
201  Round(c, d, e, f, g, h, a, b,
202  Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
203  Round(b, c, d, e, f, g, h, a,
204  Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
205  Round(a, b, c, d, e, f, g, h,
206  Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
207  Round(h, a, b, c, d, e, f, g,
208  Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
209  Round(g, h, a, b, c, d, e, f,
210  Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
211  Round(f, g, h, a, b, c, d, e,
212  Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
213  Round(e, f, g, h, a, b, c, d,
214  Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
215  Round(d, e, f, g, h, a, b, c,
216  Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
217  Round(c, d, e, f, g, h, a, b,
218  Add(K(0xbef9a3f7ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
219  Round(b, c, d, e, f, g, h, a,
220  Add(K(0xc67178f2ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
221 
222  a = Add(a, K(0x6a09e667ul));
223  b = Add(b, K(0xbb67ae85ul));
224  c = Add(c, K(0x3c6ef372ul));
225  d = Add(d, K(0xa54ff53aul));
226  e = Add(e, K(0x510e527ful));
227  f = Add(f, K(0x9b05688cul));
228  g = Add(g, K(0x1f83d9abul));
229  h = Add(h, K(0x5be0cd19ul));
230 
231  __m128i t0 = a, t1 = b, t2 = c, t3 = d, t4 = e, t5 = f, t6 = g, t7 = h;
232 
233  // Transform 2
234  Round(a, b, c, d, e, f, g, h, K(0xc28a2f98ul));
235  Round(h, a, b, c, d, e, f, g, K(0x71374491ul));
236  Round(g, h, a, b, c, d, e, f, K(0xb5c0fbcful));
237  Round(f, g, h, a, b, c, d, e, K(0xe9b5dba5ul));
238  Round(e, f, g, h, a, b, c, d, K(0x3956c25bul));
239  Round(d, e, f, g, h, a, b, c, K(0x59f111f1ul));
240  Round(c, d, e, f, g, h, a, b, K(0x923f82a4ul));
241  Round(b, c, d, e, f, g, h, a, K(0xab1c5ed5ul));
242  Round(a, b, c, d, e, f, g, h, K(0xd807aa98ul));
243  Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
244  Round(g, h, a, b, c, d, e, f, K(0x243185beul));
245  Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
246  Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
247  Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
248  Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
249  Round(b, c, d, e, f, g, h, a, K(0xc19bf374ul));
250  Round(a, b, c, d, e, f, g, h, K(0x649b69c1ul));
251  Round(h, a, b, c, d, e, f, g, K(0xf0fe4786ul));
252  Round(g, h, a, b, c, d, e, f, K(0x0fe1edc6ul));
253  Round(f, g, h, a, b, c, d, e, K(0x240cf254ul));
254  Round(e, f, g, h, a, b, c, d, K(0x4fe9346ful));
255  Round(d, e, f, g, h, a, b, c, K(0x6cc984beul));
256  Round(c, d, e, f, g, h, a, b, K(0x61b9411eul));
257  Round(b, c, d, e, f, g, h, a, K(0x16f988faul));
258  Round(a, b, c, d, e, f, g, h, K(0xf2c65152ul));
259  Round(h, a, b, c, d, e, f, g, K(0xa88e5a6dul));
260  Round(g, h, a, b, c, d, e, f, K(0xb019fc65ul));
261  Round(f, g, h, a, b, c, d, e, K(0xb9d99ec7ul));
262  Round(e, f, g, h, a, b, c, d, K(0x9a1231c3ul));
263  Round(d, e, f, g, h, a, b, c, K(0xe70eeaa0ul));
264  Round(c, d, e, f, g, h, a, b, K(0xfdb1232bul));
265  Round(b, c, d, e, f, g, h, a, K(0xc7353eb0ul));
266  Round(a, b, c, d, e, f, g, h, K(0x3069bad5ul));
267  Round(h, a, b, c, d, e, f, g, K(0xcb976d5ful));
268  Round(g, h, a, b, c, d, e, f, K(0x5a0f118ful));
269  Round(f, g, h, a, b, c, d, e, K(0xdc1eeefdul));
270  Round(e, f, g, h, a, b, c, d, K(0x0a35b689ul));
271  Round(d, e, f, g, h, a, b, c, K(0xde0b7a04ul));
272  Round(c, d, e, f, g, h, a, b, K(0x58f4ca9dul));
273  Round(b, c, d, e, f, g, h, a, K(0xe15d5b16ul));
274  Round(a, b, c, d, e, f, g, h, K(0x007f3e86ul));
275  Round(h, a, b, c, d, e, f, g, K(0x37088980ul));
276  Round(g, h, a, b, c, d, e, f, K(0xa507ea32ul));
277  Round(f, g, h, a, b, c, d, e, K(0x6fab9537ul));
278  Round(e, f, g, h, a, b, c, d, K(0x17406110ul));
279  Round(d, e, f, g, h, a, b, c, K(0x0d8cd6f1ul));
280  Round(c, d, e, f, g, h, a, b, K(0xcdaa3b6dul));
281  Round(b, c, d, e, f, g, h, a, K(0xc0bbbe37ul));
282  Round(a, b, c, d, e, f, g, h, K(0x83613bdaul));
283  Round(h, a, b, c, d, e, f, g, K(0xdb48a363ul));
284  Round(g, h, a, b, c, d, e, f, K(0x0b02e931ul));
285  Round(f, g, h, a, b, c, d, e, K(0x6fd15ca7ul));
286  Round(e, f, g, h, a, b, c, d, K(0x521afacaul));
287  Round(d, e, f, g, h, a, b, c, K(0x31338431ul));
288  Round(c, d, e, f, g, h, a, b, K(0x6ed41a95ul));
289  Round(b, c, d, e, f, g, h, a, K(0x6d437890ul));
290  Round(a, b, c, d, e, f, g, h, K(0xc39c91f2ul));
291  Round(h, a, b, c, d, e, f, g, K(0x9eccabbdul));
292  Round(g, h, a, b, c, d, e, f, K(0xb5c9a0e6ul));
293  Round(f, g, h, a, b, c, d, e, K(0x532fb63cul));
294  Round(e, f, g, h, a, b, c, d, K(0xd2c741c6ul));
295  Round(d, e, f, g, h, a, b, c, K(0x07237ea3ul));
296  Round(c, d, e, f, g, h, a, b, K(0xa4954b68ul));
297  Round(b, c, d, e, f, g, h, a, K(0x4c191d76ul));
298 
299  w0 = Add(t0, a);
300  w1 = Add(t1, b);
301  w2 = Add(t2, c);
302  w3 = Add(t3, d);
303  w4 = Add(t4, e);
304  w5 = Add(t5, f);
305  w6 = Add(t6, g);
306  w7 = Add(t7, h);
307 
308  // Transform 3
309  a = K(0x6a09e667ul);
310  b = K(0xbb67ae85ul);
311  c = K(0x3c6ef372ul);
312  d = K(0xa54ff53aul);
313  e = K(0x510e527ful);
314  f = K(0x9b05688cul);
315  g = K(0x1f83d9abul);
316  h = K(0x5be0cd19ul);
317 
318  Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0));
319  Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1));
320  Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2));
321  Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3));
322  Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4));
323  Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5));
324  Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6));
325  Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7));
326  Round(a, b, c, d, e, f, g, h, K(0x5807aa98ul));
327  Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
328  Round(g, h, a, b, c, d, e, f, K(0x243185beul));
329  Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
330  Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
331  Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
332  Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
333  Round(b, c, d, e, f, g, h, a, K(0xc19bf274ul));
334  Round(a, b, c, d, e, f, g, h, Add(K(0xe49b69c1ul), Inc(w0, sigma0(w1))));
335  Round(h, a, b, c, d, e, f, g,
336  Add(K(0xefbe4786ul), Inc(w1, K(0xa00000ul), sigma0(w2))));
337  Round(g, h, a, b, c, d, e, f,
338  Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), sigma0(w3))));
339  Round(f, g, h, a, b, c, d, e,
340  Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), sigma0(w4))));
341  Round(e, f, g, h, a, b, c, d,
342  Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), sigma0(w5))));
343  Round(d, e, f, g, h, a, b, c,
344  Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), sigma0(w6))));
345  Round(c, d, e, f, g, h, a, b,
346  Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), K(0x100ul), sigma0(w7))));
347  Round(b, c, d, e, f, g, h, a,
348  Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, K(0x11002000ul))));
349  Round(a, b, c, d, e, f, g, h,
350  Add(K(0x983e5152ul), w8 = Add(K(0x80000000ul), sigma1(w6), w1)));
351  Round(h, a, b, c, d, e, f, g,
352  Add(K(0xa831c66dul), w9 = Add(sigma1(w7), w2)));
353  Round(g, h, a, b, c, d, e, f,
354  Add(K(0xb00327c8ul), w10 = Add(sigma1(w8), w3)));
355  Round(f, g, h, a, b, c, d, e,
356  Add(K(0xbf597fc7ul), w11 = Add(sigma1(w9), w4)));
357  Round(e, f, g, h, a, b, c, d,
358  Add(K(0xc6e00bf3ul), w12 = Add(sigma1(w10), w5)));
359  Round(d, e, f, g, h, a, b, c,
360  Add(K(0xd5a79147ul), w13 = Add(sigma1(w11), w6)));
361  Round(c, d, e, f, g, h, a, b,
362  Add(K(0x06ca6351ul), w14 = Add(sigma1(w12), w7, K(0x400022ul))));
363  Round(b, c, d, e, f, g, h, a,
364  Add(K(0x14292967ul),
365  w15 = Add(K(0x100ul), sigma1(w13), w8, sigma0(w0))));
366  Round(a, b, c, d, e, f, g, h,
367  Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
368  Round(h, a, b, c, d, e, f, g,
369  Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
370  Round(g, h, a, b, c, d, e, f,
371  Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
372  Round(f, g, h, a, b, c, d, e,
373  Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
374  Round(e, f, g, h, a, b, c, d,
375  Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
376  Round(d, e, f, g, h, a, b, c,
377  Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
378  Round(c, d, e, f, g, h, a, b,
379  Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
380  Round(b, c, d, e, f, g, h, a,
381  Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
382  Round(a, b, c, d, e, f, g, h,
383  Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
384  Round(h, a, b, c, d, e, f, g,
385  Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
386  Round(g, h, a, b, c, d, e, f,
387  Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
388  Round(f, g, h, a, b, c, d, e,
389  Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
390  Round(e, f, g, h, a, b, c, d,
391  Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
392  Round(d, e, f, g, h, a, b, c,
393  Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
394  Round(c, d, e, f, g, h, a, b,
395  Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
396  Round(b, c, d, e, f, g, h, a,
397  Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
398  Round(a, b, c, d, e, f, g, h,
399  Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
400  Round(h, a, b, c, d, e, f, g,
401  Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
402  Round(g, h, a, b, c, d, e, f,
403  Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
404  Round(f, g, h, a, b, c, d, e,
405  Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
406  Round(e, f, g, h, a, b, c, d,
407  Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
408  Round(d, e, f, g, h, a, b, c,
409  Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
410  Round(c, d, e, f, g, h, a, b,
411  Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
412  Round(b, c, d, e, f, g, h, a,
413  Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
414  Round(a, b, c, d, e, f, g, h,
415  Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
416  Round(h, a, b, c, d, e, f, g,
417  Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
418  Round(g, h, a, b, c, d, e, f,
419  Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
420  Round(f, g, h, a, b, c, d, e,
421  Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
422  Round(e, f, g, h, a, b, c, d,
423  Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
424  Round(d, e, f, g, h, a, b, c,
425  Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
426  Round(c, d, e, f, g, h, a, b,
427  Add(K(0xbef9a3f7ul), w14, sigma1(w12), w7, sigma0(w15)));
428  Round(b, c, d, e, f, g, h, a,
429  Add(K(0xc67178f2ul), w15, sigma1(w13), w8, sigma0(w0)));
430 
431  // Output
432  Write4(out, 0, Add(a, K(0x6a09e667ul)));
433  Write4(out, 4, Add(b, K(0xbb67ae85ul)));
434  Write4(out, 8, Add(c, K(0x3c6ef372ul)));
435  Write4(out, 12, Add(d, K(0xa54ff53aul)));
436  Write4(out, 16, Add(e, K(0x510e527ful)));
437  Write4(out, 20, Add(f, K(0x9b05688cul)));
438  Write4(out, 24, Add(g, K(0x1f83d9abul)));
439  Write4(out, 28, Add(h, K(0x5be0cd19ul)));
440 }
441 } // namespace sha256d64_sse41
442 
443 #endif
#define Round(a, b, c, d, e, f, g, h, k, w)
Definition: hash_impl.h:24
static void WriteLE32(uint8_t *ptr, uint32_t x)
Definition: common.h:40
#define Maj(x, y, z)
Definition: hash_impl.h:18
#define sigma1(x)
Definition: hash_impl.h:22
#define sigma0(x)
Definition: hash_impl.h:21
void Transform_4way(uint8_t *out, const uint8_t *in)
#define Sigma0(x)
Definition: hash_impl.h:19
static uint32_t ReadLE32(const uint8_t *ptr)
Definition: common.h:23
#define Sigma1(x)
Definition: hash_impl.h:20
#define Ch(x, y, z)
Definition: hash_impl.h:17