1// Copyright (c) 2017-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_AVX2
6
7#include <cstdint>
8#include <immintrin.h>
9
10#include <attributes.h>
11#include <crypto/common.h>
12
13namespace sha256d64_avx2 {
14namespace {
15
16__m256i inline K(uint32_t x) { return _mm256_set1_epi32(x); }
17
18__m256i inline Add(__m256i x, __m256i y) { return _mm256_add_epi32(x, y); }
19__m256i inline Add(__m256i x, __m256i y, __m256i z) { return Add(Add(x, y), z); }
20__m256i inline Add(__m256i x, __m256i y, __m256i z, __m256i w) { return Add(Add(x, y), Add(z, w)); }
21__m256i inline Add(__m256i x, __m256i y, __m256i z, __m256i w, __m256i v) { return Add(Add(x, y, z), Add(w, v)); }
22__m256i inline Inc(__m256i& x, __m256i y) { x = Add(x, y); return x; }
23__m256i inline Inc(__m256i& x, __m256i y, __m256i z) { x = Add(x, y, z); return x; }
24__m256i inline Inc(__m256i& x, __m256i y, __m256i z, __m256i w) { x = Add(x, y, z, w); return x; }
25__m256i inline Xor(__m256i x, __m256i y) { return _mm256_xor_si256(x, y); }
26__m256i inline Xor(__m256i x, __m256i y, __m256i z) { return Xor(Xor(x, y), z); }
27__m256i inline Or(__m256i x, __m256i y) { return _mm256_or_si256(x, y); }
28__m256i inline And(__m256i x, __m256i y) { return _mm256_and_si256(x, y); }
29__m256i inline ShR(__m256i x, int n) { return _mm256_srli_epi32(x, n); }
30__m256i inline ShL(__m256i x, int n) { return _mm256_slli_epi32(x, n); }
31
32__m256i inline Ch(__m256i x, __m256i y, __m256i z) { return Xor(z, And(x, Xor(y, z))); }
33__m256i inline Maj(__m256i x, __m256i y, __m256i z) { return Or(And(x, y), And(z, Or(x, y))); }
34__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))); }
35__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))); }
36__m256i inline sigma0(__m256i x) { return Xor(Or(ShR(x, 7), ShL(x, 25)), Or(ShR(x, 18), ShL(x, 14)), ShR(x, 3)); }
37__m256i inline sigma1(__m256i 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. */
40void ALWAYS_INLINE Round(__m256i a, __m256i b, __m256i c, __m256i& d, __m256i e, __m256i f, __m256i g, __m256i& h, __m256i k)
41{
42    __m256i t1 = Add(h, Sigma1(e), Ch(e, f, g), k);
43    __m256i t2 = Add(Sigma0(a), Maj(a, b, c));
44    d = Add(d, t1);
45    h = Add(t1, t2);
46}
47
48__m256i inline Read8(const unsigned char* chunk, int offset) {
49    __m256i ret = _mm256_set_epi32(
50        ReadLE32(chunk + 0 + offset),
51        ReadLE32(chunk + 64 + offset),
52        ReadLE32(chunk + 128 + offset),
53        ReadLE32(chunk + 192 + offset),
54        ReadLE32(chunk + 256 + offset),
55        ReadLE32(chunk + 320 + offset),
56        ReadLE32(chunk + 384 + offset),
57        ReadLE32(chunk + 448 + offset)
58    );
59    return _mm256_shuffle_epi8(ret, _mm256_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL, 0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
60}
61
62void inline Write8(unsigned char* out, int offset, __m256i v) {
63    v = _mm256_shuffle_epi8(v, _mm256_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL, 0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
64    WriteLE32(out + 0 + offset, _mm256_extract_epi32(v, 7));
65    WriteLE32(out + 32 + offset, _mm256_extract_epi32(v, 6));
66    WriteLE32(out + 64 + offset, _mm256_extract_epi32(v, 5));
67    WriteLE32(out + 96 + offset, _mm256_extract_epi32(v, 4));
68    WriteLE32(out + 128 + offset, _mm256_extract_epi32(v, 3));
69    WriteLE32(out + 160 + offset, _mm256_extract_epi32(v, 2));
70    WriteLE32(out + 192 + offset, _mm256_extract_epi32(v, 1));
71    WriteLE32(out + 224 + offset, _mm256_extract_epi32(v, 0));
72}
73
74}
75
76void Transform_8way(unsigned char* out, const unsigned char* in)
77{
78    // Transform 1
79    __m256i a = K(0x6a09e667ul);
80    __m256i b = K(0xbb67ae85ul);
81    __m256i c = K(0x3c6ef372ul);
82    __m256i d = K(0xa54ff53aul);
83    __m256i e = K(0x510e527ful);
84    __m256i f = K(0x9b05688cul);
85    __m256i g = K(0x1f83d9abul);
86    __m256i h = K(0x5be0cd19ul);
87
88    __m256i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
89
90    Round(a, b, c, d, e, f, g, h, Add(K(0x428a2f98ul), w0 = Read8(in, 0)));
91    Round(h, a, b, c, d, e, f, g, Add(K(0x71374491ul), w1 = Read8(in, 4)));
92    Round(g, h, a, b, c, d, e, f, Add(K(0xb5c0fbcful), w2 = Read8(in, 8)));
93    Round(f, g, h, a, b, c, d, e, Add(K(0xe9b5dba5ul), w3 = Read8(in, 12)));
94    Round(e, f, g, h, a, b, c, d, Add(K(0x3956c25bul), w4 = Read8(in, 16)));
95    Round(d, e, f, g, h, a, b, c, Add(K(0x59f111f1ul), w5 = Read8(in, 20)));
96    Round(c, d, e, f, g, h, a, b, Add(K(0x923f82a4ul), w6 = Read8(in, 24)));
97    Round(b, c, d, e, f, g, h, a, Add(K(0xab1c5ed5ul), w7 = Read8(in, 28)));
98    Round(a, b, c, d, e, f, g, h, Add(K(0xd807aa98ul), w8 = Read8(in, 32)));
99    Round(h, a, b, c, d, e, f, g, Add(K(0x12835b01ul), w9 = Read8(in, 36)));
100    Round(g, h, a, b, c, d, e, f, Add(K(0x243185beul), w10 = Read8(in, 40)));
101    Round(f, g, h, a, b, c, d, e, Add(K(0x550c7dc3ul), w11 = Read8(in, 44)));
102    Round(e, f, g, h, a, b, c, d, Add(K(0x72be5d74ul), w12 = Read8(in, 48)));
103    Round(d, e, f, g, h, a, b, c, Add(K(0x80deb1feul), w13 = Read8(in, 52)));
104    Round(c, d, e, f, g, h, a, b, Add(K(0x9bdc06a7ul), w14 = Read8(in, 56)));
105    Round(b, c, d, e, f, g, h, a, Add(K(0xc19bf174ul), w15 = Read8(in, 60)));
106    Round(a, b, c, d, e, f, g, h, Add(K(0xe49b69c1ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
107    Round(h, a, b, c, d, e, f, g, Add(K(0xefbe4786ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
108    Round(g, h, a, b, c, d, e, f, Add(K(0x0fc19dc6ul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
109    Round(f, g, h, a, b, c, d, e, Add(K(0x240ca1ccul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
110    Round(e, f, g, h, a, b, c, d, Add(K(0x2de92c6ful), Inc(w4, sigma1(w2), w13, sigma0(w5))));
111    Round(d, e, f, g, h, a, b, c, Add(K(0x4a7484aaul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
112    Round(c, d, e, f, g, h, a, b, Add(K(0x5cb0a9dcul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
113    Round(b, c, d, e, f, g, h, a, Add(K(0x76f988daul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
114    Round(a, b, c, d, e, f, g, h, Add(K(0x983e5152ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
115    Round(h, a, b, c, d, e, f, g, Add(K(0xa831c66dul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
116    Round(g, h, a, b, c, d, e, f, Add(K(0xb00327c8ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
117    Round(f, g, h, a, b, c, d, e, Add(K(0xbf597fc7ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
118    Round(e, f, g, h, a, b, c, d, Add(K(0xc6e00bf3ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
119    Round(d, e, f, g, h, a, b, c, Add(K(0xd5a79147ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
120    Round(c, d, e, f, g, h, a, b, Add(K(0x06ca6351ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
121    Round(b, c, d, e, f, g, h, a, Add(K(0x14292967ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
122    Round(a, b, c, d, e, f, g, h, Add(K(0x27b70a85ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
123    Round(h, a, b, c, d, e, f, g, Add(K(0x2e1b2138ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
124    Round(g, h, a, b, c, d, e, f, Add(K(0x4d2c6dfcul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
125    Round(f, g, h, a, b, c, d, e, Add(K(0x53380d13ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
126    Round(e, f, g, h, a, b, c, d, Add(K(0x650a7354ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
127    Round(d, e, f, g, h, a, b, c, Add(K(0x766a0abbul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
128    Round(c, d, e, f, g, h, a, b, Add(K(0x81c2c92eul), Inc(w6, sigma1(w4), w15, sigma0(w7))));
129    Round(b, c, d, e, f, g, h, a, Add(K(0x92722c85ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
130    Round(a, b, c, d, e, f, g, h, Add(K(0xa2bfe8a1ul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
131    Round(h, a, b, c, d, e, f, g, Add(K(0xa81a664bul), Inc(w9, sigma1(w7), w2, sigma0(w10))));
132    Round(g, h, a, b, c, d, e, f, Add(K(0xc24b8b70ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
133    Round(f, g, h, a, b, c, d, e, Add(K(0xc76c51a3ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
134    Round(e, f, g, h, a, b, c, d, Add(K(0xd192e819ul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
135    Round(d, e, f, g, h, a, b, c, Add(K(0xd6990624ul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
136    Round(c, d, e, f, g, h, a, b, Add(K(0xf40e3585ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
137    Round(b, c, d, e, f, g, h, a, Add(K(0x106aa070ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
138    Round(a, b, c, d, e, f, g, h, Add(K(0x19a4c116ul), Inc(w0, sigma1(w14), w9, sigma0(w1))));
139    Round(h, a, b, c, d, e, f, g, Add(K(0x1e376c08ul), Inc(w1, sigma1(w15), w10, sigma0(w2))));
140    Round(g, h, a, b, c, d, e, f, Add(K(0x2748774cul), Inc(w2, sigma1(w0), w11, sigma0(w3))));
141    Round(f, g, h, a, b, c, d, e, Add(K(0x34b0bcb5ul), Inc(w3, sigma1(w1), w12, sigma0(w4))));
142    Round(e, f, g, h, a, b, c, d, Add(K(0x391c0cb3ul), Inc(w4, sigma1(w2), w13, sigma0(w5))));
143    Round(d, e, f, g, h, a, b, c, Add(K(0x4ed8aa4aul), Inc(w5, sigma1(w3), w14, sigma0(w6))));
144    Round(c, d, e, f, g, h, a, b, Add(K(0x5b9cca4ful), Inc(w6, sigma1(w4), w15, sigma0(w7))));
145    Round(b, c, d, e, f, g, h, a, Add(K(0x682e6ff3ul), Inc(w7, sigma1(w5), w0, sigma0(w8))));
146    Round(a, b, c, d, e, f, g, h, Add(K(0x748f82eeul), Inc(w8, sigma1(w6), w1, sigma0(w9))));
147    Round(h, a, b, c, d, e, f, g, Add(K(0x78a5636ful), Inc(w9, sigma1(w7), w2, sigma0(w10))));
148    Round(g, h, a, b, c, d, e, f, Add(K(0x84c87814ul), Inc(w10, sigma1(w8), w3, sigma0(w11))));
149    Round(f, g, h, a, b, c, d, e, Add(K(0x8cc70208ul), Inc(w11, sigma1(w9), w4, sigma0(w12))));
150    Round(e, f, g, h, a, b, c, d, Add(K(0x90befffaul), Inc(w12, sigma1(w10), w5, sigma0(w13))));
151    Round(d, e, f, g, h, a, b, c, Add(K(0xa4506cebul), Inc(w13, sigma1(w11), w6, sigma0(w14))));
152    Round(c, d, e, f, g, h, a, b, Add(K(0xbef9a3f7ul), Inc(w14, sigma1(w12), w7, sigma0(w15))));
153    Round(b, c, d, e, f, g, h, a, Add(K(0xc67178f2ul), Inc(w15, sigma1(w13), w8, sigma0(w0))));
154
155    a = Add(a, K(0x6a09e667ul));
156    b = Add(b, K(0xbb67ae85ul));
157    c = Add(c, K(0x3c6ef372ul));
158    d = Add(d, K(0xa54ff53aul));
159    e = Add(e, K(0x510e527ful));
160    f = Add(f, K(0x9b05688cul));
161    g = Add(g, K(0x1f83d9abul));
162    h = Add(h, K(0x5be0cd19ul));
163
164    __m256i t0 = a, t1 = b, t2 = c, t3 = d, t4 = e, t5 = f, t6 = g, t7 = h;
165
166    // Transform 2
167    Round(a, b, c, d, e, f, g, h, K(0xc28a2f98ul));
168    Round(h, a, b, c, d, e, f, g, K(0x71374491ul));
169    Round(g, h, a, b, c, d, e, f, K(0xb5c0fbcful));
170    Round(f, g, h, a, b, c, d, e, K(0xe9b5dba5ul));
171    Round(e, f, g, h, a, b, c, d, K(0x3956c25bul));
172    Round(d, e, f, g, h, a, b, c, K(0x59f111f1ul));
173    Round(c, d, e, f, g, h, a, b, K(0x923f82a4ul));
174    Round(b, c, d, e, f, g, h, a, K(0xab1c5ed5ul));
175    Round(a, b, c, d, e, f, g, h, K(0xd807aa98ul));
176    Round(h, a, b, c, d, e, f, g, K(0x12835b01ul));
177    Round(g, h, a, b, c, d, e, f, K(0x243185beul));
178    Round(f, g, h, a, b, c, d, e, K(0x550c7dc3ul));
179    Round(e, f, g, h, a, b, c, d, K(0x72be5d74ul));
180    Round(d, e, f, g, h, a, b, c, K(0x80deb1feul));
181    Round(c, d, e, f, g, h, a, b, K(0x9bdc06a7ul));
182    Round(b, c, d, e, f, g, h, a, K(0xc19bf374ul));
183    Round(a, b, c, d, e, f, g, h, K(0x649b69c1ul));
184    Round(h, a, b, c, d, e, f, g, K(0xf0fe4786ul));
185    Round(g, h, a, b, c, d, e, f, K(0x0fe1edc6ul));
186    Round(f, g, h, a, b, c, d, e, K(0x240cf254ul));
187    Round(e, f, g, h, a, b, c, d, K(0x4fe9346ful));
188    Round(d, e, f, g, h, a, b, c, K(0x6cc984beul));
189    Round(c, d, e, f, g, h, a, b, K(0x61b9411eul));
190    Round(b, c, d, e, f, g, h, a, K(0x16f988faul));
191    Round(a, b, c, d, e, f, g, h, K(0xf2c65152ul));
192    Round(h, a, b, c, d, e, f, g, K(0xa88e5a6dul));
193    Round(g, h, a, b, c, d, e, f, K(0xb019fc65ul));
194    Round(f, g, h, a, b, c, d, e, K(0xb9d99ec7ul));
195    Round(e, f, g, h, a, b, c, d, K(0x9a1231c3ul));
196    Round(d, e, f, g, h, a, b, c, K(0xe70eeaa0ul));
197    Round(c, d, e, f, g, h, a, b, K(0xfdb1232bul));
198    Round(b, c, d, e, f, g, h, a, K(0xc7353eb0ul));
199    Round(a, b, c, d, e, f, g, h, K(0x3069bad5ul));
200    Round(h, a, b, c, d, e, f, g, K(0xcb976d5ful));
201    Round(g, h, a, b, c, d, e, f, K(0x5a0f118ful));
202    Round(f, g, h, a, b, c, d, e, K(0xdc1eeefdul));
203    Round(e, f, g, h, a, b, c, d, K(0x0a35b689ul));
204    Round(d, e, f, g, h, a, b, c, K(0xde0b7a04ul));
205    Round(c, d, e, f, g, h, a, b, K(0x58f4ca9dul));
206    Round(b, c, d, e, f, g, h, a, K(0xe15d5b16ul));
207    Round(a, b, c, d, e, f, g, h, K(0x007f3e86ul));
208    Round(h, a, b, c, d, e, f, g, K(0x37088980ul));
209    Round(g, h, a, b, c, d, e, f, K(0xa507ea32ul));
210    Round(f, g, h, a, b, c, d, e, K(0x6fab9537ul));
211    Round(e, f, g, h, a, b, c, d, K(0x17406110ul));
212    Round(d, e, f, g, h, a, b, c, K(0x0d8cd6f1ul));
213    Round(c, d, e, f, g, h, a, b, K(0xcdaa3b6dul));
214    Round(b, c, d, e, f, g, h, a, K(0xc0bbbe37ul));
215    Round(a, b, c, d, e, f, g, h, K(0x83613bdaul));
216    Round(h, a, b, c, d, e, f, g, K(0xdb48a363ul));
217    Round(g, h, a, b, c, d, e, f, K(0x0b02e931ul));
218    Round(f, g, h, a, b, c, d, e, K(0x6fd15ca7ul));
219    Round(e, f, g, h, a, b, c, d, K(0x521afacaul));
220    Round(d, e, f, g, h, a, b, c, K(0x31338431ul));
221    Round(c, d, e, f, g, h, a, b, K(0x6ed41a95ul));
222    Round(b, c, d, e, f, g, h, a, K(0x6d437890ul));
223    Round(a, b, c, d, e, f, g, h, K(0xc39c91f2ul));
224    Round(h, a, b, c, d, e, f, g, K(0x9eccabbdul));
225    Round(g, h, a, b, c, d, e, f, K(0xb5c9a0e6ul));
226    Round(f, g, h, a, b, c, d, e, K(0x532fb63cul));
227    Round(e, f, g, h, a, b, c, d, K(0xd2c741c6ul));
228    Round(d, e, f, g, h, a, b, c, K(0x07237ea3ul));
229    Round(c, d, e, f, g, h, a, b, K(0xa4954b68ul));
230    Round(b, c, d, e, f, g, h, a, K(0x4c191d76ul));
231
232    w0 = Add(t0, a);
233    w1 = Add(t1, b);
234    w2 = Add(t2, c);
235    w3 = Add(t3, d);
236    w4 = Add(t4, e);
237    w5 = Add(t5, f);
238    w6 = Add(t6, g);
239    w7 = Add(t7, h);
240
241    // Transform 3
242    a = K(0x6a09e667ul);
243    b = K(0xbb67ae85ul);
244    c = K(0x3c6ef372ul);
245    d = K(0xa54ff53aul);
246    e = K(0x510e527ful);
247    f = K(0x9b05688cul);
248    g = K(0x1f83d9abul);
249    h = K(0x5be0cd19ul);
250