sha256_x86_shani.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 // Based on https://github.com/noloader/SHA-Intrinsics/blob/master/sha256-x86.c,
6 // Written and placed in public domain by Jeffrey Walton.
7 // Based on code from Intel, and by Sean Gulley for the miTLS project.
8
9 #if defined(ENABLE_SSE41) && defined(ENABLE_X86_SHANI)
10
11 #include <cstddef>
12 #include <cstdint>
13 #include <immintrin.h>
14
15 #include <attributes.h>
16
17 namespace {
18
19 alignas(__m128i) const uint8_t MASK[16] = {0x03, 0x02, 0x01, 0x00, 0x07, 0x06, 0x05, 0x04, 0x0b, 0x0a, 0x09, 0x08, 0x0f, 0x0e, 0x0d, 0x0c};
20 alignas(__m128i) const uint8_t INIT0[16] = {0x8c, 0x68, 0x05, 0x9b, 0x7f, 0x52, 0x0e, 0x51, 0x85, 0xae, 0x67, 0xbb, 0x67, 0xe6, 0x09, 0x6a};
21 alignas(__m128i) const uint8_t INIT1[16] = {0x19, 0xcd, 0xe0, 0x5b, 0xab, 0xd9, 0x83, 0x1f, 0x3a, 0xf5, 0x4f, 0xa5, 0x72, 0xf3, 0x6e, 0x3c};
22
23 void ALWAYS_INLINE QuadRound(__m128i& state0, __m128i& state1, uint64_t k1, uint64_t k0)
24 {
25 const __m128i msg = _mm_set_epi64x(k1, k0);
26 state1 = _mm_sha256rnds2_epu32(state1, state0, msg);
27 state0 = _mm_sha256rnds2_epu32(state0, state1, _mm_shuffle_epi32(msg, 0x0e));
28 }
29
30 void ALWAYS_INLINE QuadRound(__m128i& state0, __m128i& state1, __m128i m, uint64_t k1, uint64_t k0)
31 {
32 const __m128i msg = _mm_add_epi32(m, _mm_set_epi64x(k1, k0));
33 state1 = _mm_sha256rnds2_epu32(state1, state0, msg);
34 state0 = _mm_sha256rnds2_epu32(state0, state1, _mm_shuffle_epi32(msg, 0x0e));
35 }
36
37 void ALWAYS_INLINE ShiftMessageA(__m128i& m0, __m128i m1)
38 {
39 m0 = _mm_sha256msg1_epu32(m0, m1);
40 }
41
42 void ALWAYS_INLINE ShiftMessageC(__m128i& m0, __m128i m1, __m128i& m2)
43 {
44 m2 = _mm_sha256msg2_epu32(_mm_add_epi32(m2, _mm_alignr_epi8(m1, m0, 4)), m1);
45 }
46
47 void ALWAYS_INLINE ShiftMessageB(__m128i& m0, __m128i m1, __m128i& m2)
48 {
49 ShiftMessageC(m0, m1, m2);
50 ShiftMessageA(m0, m1);
51 }
52
53 void ALWAYS_INLINE Shuffle(__m128i& s0, __m128i& s1)
54 {
55 const __m128i t1 = _mm_shuffle_epi32(s0, 0xB1);
56 const __m128i t2 = _mm_shuffle_epi32(s1, 0x1B);
57 s0 = _mm_alignr_epi8(t1, t2, 0x08);
58 s1 = _mm_blend_epi16(t2, t1, 0xF0);
59 }
60
61 void ALWAYS_INLINE Unshuffle(__m128i& s0, __m128i& s1)
62 {
63 const __m128i t1 = _mm_shuffle_epi32(s0, 0x1B);
64 const __m128i t2 = _mm_shuffle_epi32(s1, 0xB1);
65 s0 = _mm_blend_epi16(t1, t2, 0xF0);
66 s1 = _mm_alignr_epi8(t2, t1, 0x08);
67 }
68
69 __m128i ALWAYS_INLINE Load(const unsigned char* in)
70 {
71 return _mm_shuffle_epi8(_mm_loadu_si128((const __m128i*)in), _mm_load_si128((const __m128i*)MASK));
72 }
73
74 void ALWAYS_INLINE Save(unsigned char* out, __m128i s)
75 {
76 _mm_storeu_si128((__m128i*)out, _mm_shuffle_epi8(s, _mm_load_si128((const __m128i*)MASK)));
77 }
78 }
79
80 namespace sha256_x86_shani {
81 void Transform(uint32_t* s, const unsigned char* chunk, size_t blocks)
82 {
83 __m128i m0, m1, m2, m3, s0, s1, so0, so1;
84
85 /* Load state */
86 s0 = _mm_loadu_si128((const __m128i*)s);
87 s1 = _mm_loadu_si128((const __m128i*)(s + 4));
88 Shuffle(s0, s1);
89
90 while (blocks--) {
91 /* Remember old state */
92 so0 = s0;
93 so1 = s1;
94
95 /* Load data and transform */
96 m0 = Load(chunk);
97 QuadRound(s0, s1, m0, 0xe9b5dba5b5c0fbcfull, 0x71374491428a2f98ull);
98 m1 = Load(chunk + 16);
99 QuadRound(s0, s1, m1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
100 ShiftMessageA(m0, m1);
101 m2 = Load(chunk + 32);
102 QuadRound(s0, s1, m2, 0x550c7dc3243185beull, 0x12835b01d807aa98ull);
103 ShiftMessageA(m1, m2);
104 m3 = Load(chunk + 48);
105 QuadRound(s0, s1, m3, 0xc19bf1749bdc06a7ull, 0x80deb1fe72be5d74ull);
106 ShiftMessageB(m2, m3, m0);
107 QuadRound(s0, s1, m0, 0x240ca1cc0fc19dc6ull, 0xefbe4786E49b69c1ull);
108 ShiftMessageB(m3, m0, m1);
109 QuadRound(s0, s1, m1, 0x76f988da5cb0a9dcull, 0x4a7484aa2de92c6full);
110 ShiftMessageB(m0, m1, m2);
111 QuadRound(s0, s1, m2, 0xbf597fc7b00327c8ull, 0xa831c66d983e5152ull);
112 ShiftMessageB(m1, m2, m3);
113 QuadRound(s0, s1, m3, 0x1429296706ca6351ull, 0xd5a79147c6e00bf3ull);
114 ShiftMessageB(m2, m3, m0);
115 QuadRound(s0, s1, m0, 0x53380d134d2c6dfcull, 0x2e1b213827b70a85ull);
116 ShiftMessageB(m3, m0, m1);
117 QuadRound(s0, s1, m1, 0x92722c8581c2c92eull, 0x766a0abb650a7354ull);
118 ShiftMessageB(m0, m1, m2);
119 QuadRound(s0, s1, m2, 0xc76c51A3c24b8b70ull, 0xa81a664ba2bfe8a1ull);
120 ShiftMessageB(m1, m2, m3);
121 QuadRound(s0, s1, m3, 0x106aa070f40e3585ull, 0xd6990624d192e819ull);
122 ShiftMessageB(m2, m3, m0);
123 QuadRound(s0, s1, m0, 0x34b0bcb52748774cull, 0x1e376c0819a4c116ull);
124 ShiftMessageB(m3, m0, m1);
125 QuadRound(s0, s1, m1, 0x682e6ff35b9cca4full, 0x4ed8aa4a391c0cb3ull);
126 ShiftMessageC(m0, m1, m2);
127 QuadRound(s0, s1, m2, 0x8cc7020884c87814ull, 0x78a5636f748f82eeull);
128 ShiftMessageC(m1, m2, m3);
129 QuadRound(s0, s1, m3, 0xc67178f2bef9A3f7ull, 0xa4506ceb90befffaull);
130
131 /* Combine with old state */
132 s0 = _mm_add_epi32(s0, so0);
133 s1 = _mm_add_epi32(s1, so1);
134
135 /* Advance */
136 chunk += 64;
137 }
138
139 Unshuffle(s0, s1);
140 _mm_storeu_si128((__m128i*)s, s0);
141 _mm_storeu_si128((__m128i*)(s + 4), s1);
142 }
143 }
144
145 namespace sha256d64_x86_shani {
146
147 void Transform_2way(unsigned char* out, const unsigned char* in)
148 {
149 __m128i am0, am1, am2, am3, as0, as1, aso0, aso1;
150 __m128i bm0, bm1, bm2, bm3, bs0, bs1, bso0, bso1;
151
152 /* Transform 1 */
153 bs0 = as0 = _mm_load_si128((const __m128i*)INIT0);
154 bs1 = as1 = _mm_load_si128((const __m128i*)INIT1);
155 am0 = Load(in);
156 bm0 = Load(in + 64);
157 QuadRound(as0, as1, am0, 0xe9b5dba5b5c0fbcfull, 0x71374491428a2f98ull);
158 QuadRound(bs0, bs1, bm0, 0xe9b5dba5b5c0fbcfull, 0x71374491428a2f98ull);
159 am1 = Load(in + 16);
160 bm1 = Load(in + 80);
161 QuadRound(as0, as1, am1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
162 QuadRound(bs0, bs1, bm1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
163 ShiftMessageA(am0, am1);
164 ShiftMessageA(bm0, bm1);
165 am2 = Load(in + 32);
166 bm2 = Load(in + 96);
167 QuadRound(as0, as1, am2, 0x550c7dc3243185beull, 0x12835b01d807aa98ull);
168 QuadRound(bs0, bs1, bm2, 0x550c7dc3243185beull, 0x12835b01d807aa98ull);
169 ShiftMessageA(am1, am2);
170 ShiftMessageA(bm1, bm2);
171 am3 = Load(in + 48);
172 bm3 = Load(in + 112);
173 QuadRound(as0, as1, am3, 0xc19bf1749bdc06a7ull, 0x80deb1fe72be5d74ull);
174 QuadRound(bs0, bs1, bm3, 0xc19bf1749bdc06a7ull, 0x80deb1fe72be5d74ull);
175 ShiftMessageB(am2, am3, am0);
176 ShiftMessageB(bm2, bm3, bm0);
177 QuadRound(as0, as1, am0, 0x240ca1cc0fc19dc6ull, 0xefbe4786E49b69c1ull);
178 QuadRound(bs0, bs1, bm0, 0x240ca1cc0fc19dc6ull, 0xefbe4786E49b69c1ull);
179 ShiftMessageB(am3, am0, am1);
180 ShiftMessageB(bm3, bm0, bm1);
181 QuadRound(as0, as1, am1, 0x76f988da5cb0a9dcull, 0x4a7484aa2de92c6full);
182 QuadRound(bs0, bs1, bm1, 0x76f988da5cb0a9dcull, 0x4a7484aa2de92c6full);
183 ShiftMessageB(am0, am1, am2);
184 ShiftMessageB(bm0, bm1, bm2);
185 QuadRound(as0, as1, am2, 0xbf597fc7b00327c8ull, 0xa831c66d983e5152ull);
186 QuadRound(bs0, bs1, bm2, 0xbf597fc7b00327c8ull, 0xa831c66d983e5152ull);
187 ShiftMessageB(am1, am2, am3);
188 ShiftMessageB(bm1, bm2, bm3);
189 QuadRound(as0, as1, am3, 0x1429296706ca6351ull, 0xd5a79147c6e00bf3ull);
190 QuadRound(bs0, bs1, bm3, 0x1429296706ca6351ull, 0xd5a79147c6e00bf3ull);
191 ShiftMessageB(am2, am3, am0);
192 ShiftMessageB(bm2, bm3, bm0);
193 QuadRound(as0, as1, am0, 0x53380d134d2c6dfcull, 0x2e1b213827b70a85ull);
194 QuadRound(bs0, bs1, bm0, 0x53380d134d2c6dfcull, 0x2e1b213827b70a85ull);
195 ShiftMessageB(am3, am0, am1);
196 ShiftMessageB(bm3, bm0, bm1);
197 QuadRound(as0, as1, am1, 0x92722c8581c2c92eull, 0x766a0abb650a7354ull);
198 QuadRound(bs0, bs1, bm1, 0x92722c8581c2c92eull, 0x766a0abb650a7354ull);
199 ShiftMessageB(am0, am1, am2);
200 ShiftMessageB(bm0, bm1, bm2);
201 QuadRound(as0, as1, am2, 0xc76c51A3c24b8b70ull, 0xa81a664ba2bfe8a1ull);
202 QuadRound(bs0, bs1, bm2, 0xc76c51A3c24b8b70ull, 0xa81a664ba2bfe8a1ull);
203 ShiftMessageB(am1, am2, am3);
204 ShiftMessageB(bm1, bm2, bm3);
205 QuadRound(as0, as1, am3, 0x106aa070f40e3585ull, 0xd6990624d192e819ull);
206 QuadRound(bs0, bs1, bm3, 0x106aa070f40e3585ull, 0xd6990624d192e819ull);
207 ShiftMessageB(am2, am3, am0);
208 ShiftMessageB(bm2, bm3, bm0);
209 QuadRound(as0, as1, am0, 0x34b0bcb52748774cull, 0x1e376c0819a4c116ull);
210 QuadRound(bs0, bs1, bm0, 0x34b0bcb52748774cull, 0x1e376c0819a4c116ull);
211 ShiftMessageB(am3, am0, am1);
212 ShiftMessageB(bm3, bm0, bm1);
213 QuadRound(as0, as1, am1, 0x682e6ff35b9cca4full, 0x4ed8aa4a391c0cb3ull);
214 QuadRound(bs0, bs1, bm1, 0x682e6ff35b9cca4full, 0x4ed8aa4a391c0cb3ull);
215 ShiftMessageC(am0, am1, am2);
216 ShiftMessageC(bm0, bm1, bm2);
217 QuadRound(as0, as1, am2, 0x8cc7020884c87814ull, 0x78a5636f748f82eeull);
218 QuadRound(bs0, bs1, bm2, 0x8cc7020884c87814ull, 0x78a5636f748f82eeull);
219 ShiftMessageC(am1, am2, am3);
220 ShiftMessageC(bm1, bm2, bm3);
221 QuadRound(as0, as1, am3, 0xc67178f2bef9A3f7ull, 0xa4506ceb90befffaull);
222 QuadRound(bs0, bs1, bm3, 0xc67178f2bef9A3f7ull, 0xa4506ceb90befffaull);
223 as0 = _mm_add_epi32(as0, _mm_load_si128((const __m128i*)INIT0));
224 bs0 = _mm_add_epi32(bs0, _mm_load_si128((const __m128i*)INIT0));
225 as1 = _mm_add_epi32(as1, _mm_load_si128((const __m128i*)INIT1));
226 bs1 = _mm_add_epi32(bs1, _mm_load_si128((const __m128i*)INIT1));
227
228 /* Transform 2 */
229 aso0 = as0;
230 bso0 = bs0;
231 aso1 = as1;
232 bso1 = bs1;
233 QuadRound(as0, as1, 0xe9b5dba5b5c0fbcfull, 0x71374491c28a2f98ull);
234 QuadRound(bs0, bs1, 0xe9b5dba5b5c0fbcfull, 0x71374491c28a2f98ull);
235 QuadRound(as0, as1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
236 QuadRound(bs0, bs1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
237 QuadRound(as0, as1, 0x550c7dc3243185beull, 0x12835b01d807aa98ull);
238 QuadRound(bs0, bs1, 0x550c7dc3243185beull, 0x12835b01d807aa98ull);
239 QuadRound(as0, as1, 0xc19bf3749bdc06a7ull, 0x80deb1fe72be5d74ull);
240 QuadRound(bs0, bs1, 0xc19bf3749bdc06a7ull, 0x80deb1fe72be5d74ull);
241 QuadRound(as0, as1, 0x240cf2540fe1edc6ull, 0xf0fe4786649b69c1ull);
242 QuadRound(bs0, bs1, 0x240cf2540fe1edc6ull, 0xf0fe4786649b69c1ull);
243 QuadRound(as0, as1, 0x16f988fa61b9411eull, 0x6cc984be4fe9346full);
244 QuadRound(bs0, bs1, 0x16f988fa61b9411eull, 0x6cc984be4fe9346full);
245 QuadRound(as0, as1, 0xb9d99ec7b019fc65ull, 0xa88e5a6df2c65152ull);
246 QuadRound(bs0, bs1, 0xb9d99ec7b019fc65ull, 0xa88e5a6df2c65152ull);
247 QuadRound(as0, as1, 0xc7353eb0fdb1232bull, 0xe70eeaa09a1231c3ull);
248 QuadRound(bs0, bs1, 0xc7353eb0fdb1232bull, 0xe70eeaa09a1231c3ull);
249 QuadRound(as0, as1, 0xdc1eeefd5a0f118full, 0xcb976d5f3069bad5ull);
250 QuadRound(bs0, bs1, 0xdc1eeefd5a0f118full, 0xcb976d5f3069bad5ull);
251 QuadRound(as0, as1, 0xe15d5b1658f4ca9dull, 0xde0b7a040a35b689ull);
252 QuadRound(bs0, bs1, 0xe15d5b1658f4ca9dull, 0xde0b7a040a35b689ull);
253 QuadRound(as0, as1, 0x6fab9537a507ea32ull, 0x37088980007f3e86ull);
254 QuadRound(bs0, bs1, 0x6fab9537a507ea32ull, 0x37088980007f3e86ull);
255 QuadRound(as0, as1, 0xc0bbbe37cdaa3b6dull, 0x0d8cd6f117406110ull);
256 QuadRound(bs0, bs1, 0xc0bbbe37cdaa3b6dull, 0x0d8cd6f117406110ull);
257 QuadRound(as0, as1, 0x6fd15ca70b02e931ull, 0xdb48a36383613bdaull);
258 QuadRound(bs0, bs1, 0x6fd15ca70b02e931ull, 0xdb48a36383613bdaull);
259 QuadRound(as0, as1, 0x6d4378906ed41a95ull, 0x31338431521afacaull);
260 QuadRound(bs0, bs1, 0x6d4378906ed41a95ull, 0x31338431521afacaull);
261 QuadRound(as0, as1, 0x532fb63cb5c9a0e6ull, 0x9eccabbdc39c91f2ull);
262 QuadRound(bs0, bs1, 0x532fb63cb5c9a0e6ull, 0x9eccabbdc39c91f2ull);
263 QuadRound(as0, as1, 0x4c191d76a4954b68ull, 0x07237ea3d2c741c6ull);
264 QuadRound(bs0, bs1, 0x4c191d76a4954b68ull, 0x07237ea3d2c741c6ull);
265 as0 = _mm_add_epi32(as0, aso0);
266 bs0 = _mm_add_epi32(bs0, bso0);
267 as1 = _mm_add_epi32(as1, aso1);
268 bs1 = _mm_add_epi32(bs1, bso1);
269
270 /* Extract hash */
271 Unshuffle(as0, as1);
272 Unshuffle(bs0, bs1);
273 am0 = as0;
274 bm0 = bs0;
275 am1 = as1;
276 bm1 = bs1;
277
278 /* Transform 3 */
279 bs0 = as0 = _mm_load_si128((const __m128i*)INIT0);
280 bs1 = as1 = _mm_load_si128((const __m128i*)INIT1);
281 QuadRound(as0, as1, am0, 0xe9b5dba5B5c0fbcfull, 0x71374491428a2f98ull);
282 QuadRound(bs0, bs1, bm0, 0xe9b5dba5B5c0fbcfull, 0x71374491428a2f98ull);
283 QuadRound(as0, as1, am1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
284 QuadRound(bs0, bs1, bm1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
285 ShiftMessageA(am0, am1);
286 ShiftMessageA(bm0, bm1);
287 bm2 = am2 = _mm_set_epi64x(0x0ull, 0x80000000ull);
288 QuadRound(as0, as1, 0x550c7dc3243185beull, 0x12835b015807aa98ull);
289 QuadRound(bs0, bs1, 0x550c7dc3243185beull, 0x12835b015807aa98ull);
290 ShiftMessageA(am1, am2);
291 ShiftMessageA(bm1, bm2);
292 bm3 = am3 = _mm_set_epi64x(0x10000000000ull, 0x0ull);
293 QuadRound(as0, as1, 0xc19bf2749bdc06a7ull, 0x80deb1fe72be5d74ull);
294 QuadRound(bs0, bs1, 0xc19bf2749bdc06a7ull, 0x80deb1fe72be5d74ull);
295 ShiftMessageB(am2, am3, am0);
296 ShiftMessageB(bm2, bm3, bm0);
297 QuadRound(as0, as1, am0, 0x240ca1cc0fc19dc6ull, 0xefbe4786e49b69c1ull);
298 QuadRound(bs0, bs1, bm0, 0x240ca1cc0fc19dc6ull, 0xefbe4786e49b69c1ull);
299 ShiftMessageB(am3, am0, am1);
300 ShiftMessageB(bm3, bm0, bm1);
301 QuadRound(as0, as1, am1, 0x76f988da5cb0a9dcull, 0x4a7484aa2de92c6full);
302 QuadRound(bs0, bs1, bm1, 0x76f988da5cb0a9dcull, 0x4a7484aa2de92c6full);
303 ShiftMessageB(am0, am1, am2);
304 ShiftMessageB(bm0, bm1, bm2);
305 QuadRound(as0, as1, am2, 0xbf597fc7b00327c8ull, 0xa831c66d983e5152ull);
306 QuadRound(bs0, bs1, bm2, 0xbf597fc7b00327c8ull, 0xa831c66d983e5152ull);
307 ShiftMessageB(am1, am2, am3);
308 ShiftMessageB(bm1, bm2, bm3);
309 QuadRound(as0, as1, am3, 0x1429296706ca6351ull, 0xd5a79147c6e00bf3ull);
310 QuadRound(bs0, bs1, bm3, 0x1429296706ca6351ull, 0xd5a79147c6e00bf3ull);
311 ShiftMessageB(am2, am3, am0);
312 ShiftMessageB(bm2, bm3, bm0);
313 QuadRound(as0, as1, am0, 0x53380d134d2c6dfcull, 0x2e1b213827b70a85ull);
314 QuadRound(bs0, bs1, bm0, 0x53380d134d2c6dfcull, 0x2e1b213827b70a85ull);
315 ShiftMessageB(am3, am0, am1);
316 ShiftMessageB(bm3, bm0, bm1);
317 QuadRound(as0, as1, am1, 0x92722c8581c2c92eull, 0x766a0abb650a7354ull);
318 QuadRound(bs0, bs1, bm1, 0x92722c8581c2c92eull, 0x766a0abb650a7354ull);
319 ShiftMessageB(am0, am1, am2);
320 ShiftMessageB(bm0, bm1, bm2);
321 QuadRound(as0, as1, am2, 0xc76c51a3c24b8b70ull, 0xa81a664ba2bfe8A1ull);
322 QuadRound(bs0, bs1, bm2, 0xc76c51a3c24b8b70ull, 0xa81a664ba2bfe8A1ull);
323 ShiftMessageB(am1, am2, am3);
324 ShiftMessageB(bm1, bm2, bm3);
325 QuadRound(as0, as1, am3, 0x106aa070f40e3585ull, 0xd6990624d192e819ull);
326 QuadRound(bs0, bs1, bm3, 0x106aa070f40e3585ull, 0xd6990624d192e819ull);
327 ShiftMessageB(am2, am3, am0);
328 ShiftMessageB(bm2, bm3, bm0);
329 QuadRound(as0, as1, am0, 0x34b0bcb52748774cull, 0x1e376c0819a4c116ull);
330 QuadRound(bs0, bs1, bm0, 0x34b0bcb52748774cull, 0x1e376c0819a4c116ull);
331 ShiftMessageB(am3, am0, am1);
332 ShiftMessageB(bm3, bm0, bm1);
333 QuadRound(as0, as1, am1, 0x682e6ff35b9cca4full, 0x4ed8aa4a391c0cb3ull);
334 QuadRound(bs0, bs1, bm1, 0x682e6ff35b9cca4full, 0x4ed8aa4a391c0cb3ull);
335 ShiftMessageC(am0, am1, am2);
336 ShiftMessageC(bm0, bm1, bm2);
337 QuadRound(as0, as1, am2, 0x8cc7020884c87814ull, 0x78a5636f748f82eeull);
338 QuadRound(bs0, bs1, bm2, 0x8cc7020884c87814ull, 0x78a5636f748f82eeull);
339 ShiftMessageC(am1, am2, am3);
340 ShiftMessageC(bm1, bm2, bm3);
341 QuadRound(as0, as1, am3, 0xc67178f2bef9a3f7ull, 0xa4506ceb90befffaull);
342 QuadRound(bs0, bs1, bm3, 0xc67178f2bef9a3f7ull, 0xa4506ceb90befffaull);
343 as0 = _mm_add_epi32(as0, _mm_load_si128((const __m128i*)INIT0));
344 bs0 = _mm_add_epi32(bs0, _mm_load_si128((const __m128i*)INIT0));
345 as1 = _mm_add_epi32(as1, _mm_load_si128((const __m128i*)INIT1));
346 bs1 = _mm_add_epi32(bs1, _mm_load_si128((const __m128i*)INIT1));
347
348 /* Extract hash into out */
349 Unshuffle(as0, as1);
350 Unshuffle(bs0, bs1);
351 Save(out, as0);
352 Save(out + 16, as1);
353 Save(out + 32, bs0);
354 Save(out + 48, bs1);
355 }
356
357 }
358
359 #endif
360