Line data Source code
1 :
2 : // Source originally from https://github.com/BLAKE3-team/BLAKE3
3 : // From commit: c0ea395cf91d242f078c23d5f8d87eb9dd5f7b78
4 :
5 : #include "blake3_impl.h"
6 :
7 : #include <immintrin.h>
8 :
9 3305846 : #define DEGREE 4
10 :
11 : #define _mm_shuffle_ps2(a, b, c) \
12 830314720 : (_mm_castps_si128( \
13 830314720 : _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
14 :
15 381264692 : INLINE __m128i loadu(const uint8_t src[16]) {
16 381264692 : return _mm_loadu_si128((const __m128i *)src);
17 381264692 : }
18 :
19 139796828 : INLINE void storeu(__m128i src, uint8_t dest[16]) {
20 139796828 : _mm_storeu_si128((__m128i *)dest, src);
21 139796828 : }
22 :
23 5115292728 : INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
24 :
25 : // Note that clang-format doesn't like the name "xor" for some reason.
26 5270975396 : INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
27 :
28 15620012 : INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
29 :
30 103789340 : INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
31 103789340 : return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
32 103789340 : }
33 :
34 852548788 : INLINE __m128i rot16(__m128i x) {
35 852548788 : return _mm_shuffle_epi8(
36 852548788 : x, _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2));
37 852548788 : }
38 :
39 852548788 : INLINE __m128i rot12(__m128i x) {
40 852548788 : return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
41 852548788 : }
42 :
43 852548788 : INLINE __m128i rot8(__m128i x) {
44 852548788 : return _mm_shuffle_epi8(
45 852548788 : x, _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1));
46 852548788 : }
47 :
48 852548788 : INLINE __m128i rot7(__m128i x) {
49 852548788 : return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
50 852548788 : }
51 :
52 : INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
53 726525380 : __m128i m) {
54 726525380 : *row0 = addv(addv(*row0, m), *row1);
55 726525380 : *row3 = xorv(*row3, *row0);
56 726525380 : *row3 = rot16(*row3);
57 726525380 : *row2 = addv(*row2, *row3);
58 726525380 : *row1 = xorv(*row1, *row2);
59 726525380 : *row1 = rot12(*row1);
60 726525380 : }
61 :
62 : INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
63 726525380 : __m128i m) {
64 726525380 : *row0 = addv(addv(*row0, m), *row1);
65 726525380 : *row3 = xorv(*row3, *row0);
66 726525380 : *row3 = rot8(*row3);
67 726525380 : *row2 = addv(*row2, *row3);
68 726525380 : *row1 = xorv(*row1, *row2);
69 726525380 : *row1 = rot7(*row1);
70 726525380 : }
71 :
72 : // Note the optimization here of leaving row1 as the unrotated row, rather than
73 : // row0. All the message loads below are adjusted to compensate for this. See
74 : // discussion at https://github.com/sneves/blake2-avx2/pull/4
75 363262690 : INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
76 363262690 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
77 363262690 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
78 363262690 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
79 363262690 : }
80 :
81 363262690 : INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
82 363262690 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
83 363262690 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
84 363262690 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
85 363262690 : }
86 :
87 : INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
88 : const uint8_t block[BLAKE3_BLOCK_LEN],
89 51894670 : uint8_t block_len, uint64_t counter, uint8_t flags) {
90 51894670 : rows[0] = loadu((uint8_t *)&cv[0]);
91 51894670 : rows[1] = loadu((uint8_t *)&cv[4]);
92 51894670 : rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
93 51894670 : rows[3] = set4(counter_low(counter), counter_high(counter),
94 51894670 : (uint32_t)block_len, (uint32_t)flags);
95 :
96 51894670 : __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
97 51894670 : __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
98 51894670 : __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
99 51894670 : __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
100 :
101 51894670 : __m128i t0, t1, t2, t3, tt;
102 :
103 : // Round 1. The first round permutes the message words from the original
104 : // input order, into the groups that get mixed in parallel.
105 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0
106 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
107 51894670 : t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1
108 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
109 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
110 51894670 : t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8
111 51894670 : t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14
112 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
113 51894670 : t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9
114 51894670 : t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15
115 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
116 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
117 51894670 : m0 = t0;
118 51894670 : m1 = t1;
119 51894670 : m2 = t2;
120 51894670 : m3 = t3;
121 :
122 : // Round 2. This round and all following rounds apply a fixed permutation
123 : // to the message words from the round before.
124 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
125 51894670 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
126 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
127 51894670 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
128 51894670 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
129 51894670 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
130 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
131 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
132 51894670 : t2 = _mm_unpacklo_epi64(m3, m1);
133 51894670 : tt = _mm_blend_epi16(t2, m2, 0xC0);
134 51894670 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
135 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
136 51894670 : t3 = _mm_unpackhi_epi32(m1, m3);
137 51894670 : tt = _mm_unpacklo_epi32(m2, t3);
138 51894670 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
139 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
140 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
141 51894670 : m0 = t0;
142 51894670 : m1 = t1;
143 51894670 : m2 = t2;
144 51894670 : m3 = t3;
145 :
146 : // Round 3
147 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
148 51894670 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
149 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
150 51894670 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
151 51894670 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
152 51894670 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
153 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
154 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
155 51894670 : t2 = _mm_unpacklo_epi64(m3, m1);
156 51894670 : tt = _mm_blend_epi16(t2, m2, 0xC0);
157 51894670 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
158 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
159 51894670 : t3 = _mm_unpackhi_epi32(m1, m3);
160 51894670 : tt = _mm_unpacklo_epi32(m2, t3);
161 51894670 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
162 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
163 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
164 51894670 : m0 = t0;
165 51894670 : m1 = t1;
166 51894670 : m2 = t2;
167 51894670 : m3 = t3;
168 :
169 : // Round 4
170 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
171 51894670 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
172 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
173 51894670 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
174 51894670 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
175 51894670 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
176 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
177 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
178 51894670 : t2 = _mm_unpacklo_epi64(m3, m1);
179 51894670 : tt = _mm_blend_epi16(t2, m2, 0xC0);
180 51894670 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
181 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
182 51894670 : t3 = _mm_unpackhi_epi32(m1, m3);
183 51894670 : tt = _mm_unpacklo_epi32(m2, t3);
184 51894670 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
185 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
186 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
187 51894670 : m0 = t0;
188 51894670 : m1 = t1;
189 51894670 : m2 = t2;
190 51894670 : m3 = t3;
191 :
192 : // Round 5
193 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
194 51894670 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
195 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
196 51894670 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
197 51894670 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
198 51894670 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
199 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
200 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
201 51894670 : t2 = _mm_unpacklo_epi64(m3, m1);
202 51894670 : tt = _mm_blend_epi16(t2, m2, 0xC0);
203 51894670 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
204 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
205 51894670 : t3 = _mm_unpackhi_epi32(m1, m3);
206 51894670 : tt = _mm_unpacklo_epi32(m2, t3);
207 51894670 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
208 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
209 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
210 51894670 : m0 = t0;
211 51894670 : m1 = t1;
212 51894670 : m2 = t2;
213 51894670 : m3 = t3;
214 :
215 : // Round 6
216 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
217 51894670 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
218 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
219 51894670 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
220 51894670 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
221 51894670 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
222 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
223 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
224 51894670 : t2 = _mm_unpacklo_epi64(m3, m1);
225 51894670 : tt = _mm_blend_epi16(t2, m2, 0xC0);
226 51894670 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
227 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
228 51894670 : t3 = _mm_unpackhi_epi32(m1, m3);
229 51894670 : tt = _mm_unpacklo_epi32(m2, t3);
230 51894670 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
231 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
232 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
233 51894670 : m0 = t0;
234 51894670 : m1 = t1;
235 51894670 : m2 = t2;
236 51894670 : m3 = t3;
237 :
238 : // Round 7
239 51894670 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
240 51894670 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
241 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
242 51894670 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
243 51894670 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
244 51894670 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
245 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
246 51894670 : diagonalize(&rows[0], &rows[2], &rows[3]);
247 51894670 : t2 = _mm_unpacklo_epi64(m3, m1);
248 51894670 : tt = _mm_blend_epi16(t2, m2, 0xC0);
249 51894670 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
250 51894670 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
251 51894670 : t3 = _mm_unpackhi_epi32(m1, m3);
252 51894670 : tt = _mm_unpacklo_epi32(m2, t3);
253 51894670 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
254 51894670 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
255 51894670 : undiagonalize(&rows[0], &rows[2], &rows[3]);
256 51894670 : }
257 :
258 : void fd_blake3_compress_in_place_sse41(uint32_t cv[8],
259 : const uint8_t block[BLAKE3_BLOCK_LEN],
260 : uint8_t block_len, uint64_t counter,
261 34949678 : uint8_t flags) {
262 34949678 : __m128i rows[4];
263 34949678 : compress_pre(rows, cv, block, block_len, counter, flags);
264 34949678 : storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
265 34949678 : storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
266 34949678 : }
267 :
268 : void fd_blake3_compress_xof_sse41(const uint32_t cv[8],
269 : const uint8_t block[BLAKE3_BLOCK_LEN],
270 : uint8_t block_len, uint64_t counter,
271 16944992 : uint8_t flags, uint8_t out[64]) {
272 16944992 : __m128i rows[4];
273 16944992 : compress_pre(rows, cv, block, block_len, counter, flags);
274 16944992 : storeu(xorv(rows[0], rows[2]), &out[0]);
275 16944992 : storeu(xorv(rows[1], rows[3]), &out[16]);
276 16944992 : storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
277 16944992 : storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
278 16944992 : }
279 :
280 15752926 : INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
281 15752926 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
282 15752926 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
283 15752926 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
284 15752926 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
285 15752926 : v[0] = addv(v[0], v[4]);
286 15752926 : v[1] = addv(v[1], v[5]);
287 15752926 : v[2] = addv(v[2], v[6]);
288 15752926 : v[3] = addv(v[3], v[7]);
289 15752926 : v[12] = xorv(v[12], v[0]);
290 15752926 : v[13] = xorv(v[13], v[1]);
291 15752926 : v[14] = xorv(v[14], v[2]);
292 15752926 : v[15] = xorv(v[15], v[3]);
293 15752926 : v[12] = rot16(v[12]);
294 15752926 : v[13] = rot16(v[13]);
295 15752926 : v[14] = rot16(v[14]);
296 15752926 : v[15] = rot16(v[15]);
297 15752926 : v[8] = addv(v[8], v[12]);
298 15752926 : v[9] = addv(v[9], v[13]);
299 15752926 : v[10] = addv(v[10], v[14]);
300 15752926 : v[11] = addv(v[11], v[15]);
301 15752926 : v[4] = xorv(v[4], v[8]);
302 15752926 : v[5] = xorv(v[5], v[9]);
303 15752926 : v[6] = xorv(v[6], v[10]);
304 15752926 : v[7] = xorv(v[7], v[11]);
305 15752926 : v[4] = rot12(v[4]);
306 15752926 : v[5] = rot12(v[5]);
307 15752926 : v[6] = rot12(v[6]);
308 15752926 : v[7] = rot12(v[7]);
309 15752926 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
310 15752926 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
311 15752926 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
312 15752926 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
313 15752926 : v[0] = addv(v[0], v[4]);
314 15752926 : v[1] = addv(v[1], v[5]);
315 15752926 : v[2] = addv(v[2], v[6]);
316 15752926 : v[3] = addv(v[3], v[7]);
317 15752926 : v[12] = xorv(v[12], v[0]);
318 15752926 : v[13] = xorv(v[13], v[1]);
319 15752926 : v[14] = xorv(v[14], v[2]);
320 15752926 : v[15] = xorv(v[15], v[3]);
321 15752926 : v[12] = rot8(v[12]);
322 15752926 : v[13] = rot8(v[13]);
323 15752926 : v[14] = rot8(v[14]);
324 15752926 : v[15] = rot8(v[15]);
325 15752926 : v[8] = addv(v[8], v[12]);
326 15752926 : v[9] = addv(v[9], v[13]);
327 15752926 : v[10] = addv(v[10], v[14]);
328 15752926 : v[11] = addv(v[11], v[15]);
329 15752926 : v[4] = xorv(v[4], v[8]);
330 15752926 : v[5] = xorv(v[5], v[9]);
331 15752926 : v[6] = xorv(v[6], v[10]);
332 15752926 : v[7] = xorv(v[7], v[11]);
333 15752926 : v[4] = rot7(v[4]);
334 15752926 : v[5] = rot7(v[5]);
335 15752926 : v[6] = rot7(v[6]);
336 15752926 : v[7] = rot7(v[7]);
337 :
338 15752926 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
339 15752926 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
340 15752926 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
341 15752926 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
342 15752926 : v[0] = addv(v[0], v[5]);
343 15752926 : v[1] = addv(v[1], v[6]);
344 15752926 : v[2] = addv(v[2], v[7]);
345 15752926 : v[3] = addv(v[3], v[4]);
346 15752926 : v[15] = xorv(v[15], v[0]);
347 15752926 : v[12] = xorv(v[12], v[1]);
348 15752926 : v[13] = xorv(v[13], v[2]);
349 15752926 : v[14] = xorv(v[14], v[3]);
350 15752926 : v[15] = rot16(v[15]);
351 15752926 : v[12] = rot16(v[12]);
352 15752926 : v[13] = rot16(v[13]);
353 15752926 : v[14] = rot16(v[14]);
354 15752926 : v[10] = addv(v[10], v[15]);
355 15752926 : v[11] = addv(v[11], v[12]);
356 15752926 : v[8] = addv(v[8], v[13]);
357 15752926 : v[9] = addv(v[9], v[14]);
358 15752926 : v[5] = xorv(v[5], v[10]);
359 15752926 : v[6] = xorv(v[6], v[11]);
360 15752926 : v[7] = xorv(v[7], v[8]);
361 15752926 : v[4] = xorv(v[4], v[9]);
362 15752926 : v[5] = rot12(v[5]);
363 15752926 : v[6] = rot12(v[6]);
364 15752926 : v[7] = rot12(v[7]);
365 15752926 : v[4] = rot12(v[4]);
366 15752926 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
367 15752926 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
368 15752926 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
369 15752926 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
370 15752926 : v[0] = addv(v[0], v[5]);
371 15752926 : v[1] = addv(v[1], v[6]);
372 15752926 : v[2] = addv(v[2], v[7]);
373 15752926 : v[3] = addv(v[3], v[4]);
374 15752926 : v[15] = xorv(v[15], v[0]);
375 15752926 : v[12] = xorv(v[12], v[1]);
376 15752926 : v[13] = xorv(v[13], v[2]);
377 15752926 : v[14] = xorv(v[14], v[3]);
378 15752926 : v[15] = rot8(v[15]);
379 15752926 : v[12] = rot8(v[12]);
380 15752926 : v[13] = rot8(v[13]);
381 15752926 : v[14] = rot8(v[14]);
382 15752926 : v[10] = addv(v[10], v[15]);
383 15752926 : v[11] = addv(v[11], v[12]);
384 15752926 : v[8] = addv(v[8], v[13]);
385 15752926 : v[9] = addv(v[9], v[14]);
386 15752926 : v[5] = xorv(v[5], v[10]);
387 15752926 : v[6] = xorv(v[6], v[11]);
388 15752926 : v[7] = xorv(v[7], v[8]);
389 15752926 : v[4] = xorv(v[4], v[9]);
390 15752926 : v[5] = rot7(v[5]);
391 15752926 : v[6] = rot7(v[6]);
392 15752926 : v[7] = rot7(v[7]);
393 15752926 : v[4] = rot7(v[4]);
394 15752926 : }
395 :
396 9531048 : INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
397 : // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is
398 : // 22/33. Note that this doesn't split the vector into two lanes, as the
399 : // AVX2 counterparts do.
400 9531048 : __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
401 9531048 : __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
402 9531048 : __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
403 9531048 : __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
404 :
405 : // Interleave 64-bit lanes.
406 9531048 : __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
407 9531048 : __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
408 9531048 : __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
409 9531048 : __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
410 :
411 9531048 : vecs[0] = abcd_0;
412 9531048 : vecs[1] = abcd_1;
413 9531048 : vecs[2] = abcd_2;
414 9531048 : vecs[3] = abcd_3;
415 9531048 : }
416 :
417 : INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
418 2250418 : size_t block_offset, __m128i out[16]) {
419 2250418 : out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
420 2250418 : out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
421 2250418 : out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
422 2250418 : out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
423 2250418 : out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
424 2250418 : out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
425 2250418 : out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
426 2250418 : out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
427 2250418 : out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
428 2250418 : out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
429 2250418 : out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
430 2250418 : out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
431 2250418 : out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
432 2250418 : out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
433 2250418 : out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
434 2250418 : out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
435 11252090 : for (size_t i = 0; i < 4; ++i) {
436 9001672 : _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
437 9001672 : }
438 2250418 : transpose_vecs(&out[0]);
439 2250418 : transpose_vecs(&out[4]);
440 2250418 : transpose_vecs(&out[8]);
441 2250418 : transpose_vecs(&out[12]);
442 2250418 : }
443 :
444 : INLINE void load_counters(uint64_t counter, bool increment_counter,
445 264688 : __m128i *out_lo, __m128i *out_hi) {
446 264688 : const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
447 264688 : const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
448 264688 : const __m128i add1 = _mm_and_si128(mask, add0);
449 264688 : __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1);
450 264688 : __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32((int)0x80000000)),
451 264688 : _mm_xor_si128( l, _mm_set1_epi32((int)0x80000000)));
452 264688 : __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry);
453 264688 : *out_lo = l;
454 264688 : *out_hi = h;
455 264688 : }
456 :
457 : static
458 : void fd_blake3_hash4_sse41(const uint8_t *const *inputs, size_t blocks,
459 : const uint32_t key[8], uint64_t counter,
460 : bool increment_counter, uint8_t flags,
461 264688 : uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
462 264688 : __m128i h_vecs[8] = {
463 264688 : set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
464 264688 : set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
465 264688 : };
466 264688 : __m128i counter_low_vec, counter_high_vec;
467 264688 : load_counters(counter, increment_counter, &counter_low_vec,
468 264688 : &counter_high_vec);
469 264688 : uint8_t block_flags = flags | flags_start;
470 :
471 2515106 : for (size_t block = 0; block < blocks; block++) {
472 2250418 : if (block + 1 == blocks) {
473 264688 : block_flags |= flags_end;
474 264688 : }
475 2250418 : __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
476 2250418 : __m128i block_flags_vec = set1(block_flags);
477 2250418 : __m128i msg_vecs[16];
478 2250418 : transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
479 :
480 2250418 : __m128i v[16] = {
481 2250418 : h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
482 2250418 : h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
483 2250418 : set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]),
484 2250418 : counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
485 2250418 : };
486 2250418 : round_fn(v, msg_vecs, 0);
487 2250418 : round_fn(v, msg_vecs, 1);
488 2250418 : round_fn(v, msg_vecs, 2);
489 2250418 : round_fn(v, msg_vecs, 3);
490 2250418 : round_fn(v, msg_vecs, 4);
491 2250418 : round_fn(v, msg_vecs, 5);
492 2250418 : round_fn(v, msg_vecs, 6);
493 2250418 : h_vecs[0] = xorv(v[0], v[8]);
494 2250418 : h_vecs[1] = xorv(v[1], v[9]);
495 2250418 : h_vecs[2] = xorv(v[2], v[10]);
496 2250418 : h_vecs[3] = xorv(v[3], v[11]);
497 2250418 : h_vecs[4] = xorv(v[4], v[12]);
498 2250418 : h_vecs[5] = xorv(v[5], v[13]);
499 2250418 : h_vecs[6] = xorv(v[6], v[14]);
500 2250418 : h_vecs[7] = xorv(v[7], v[15]);
501 :
502 2250418 : block_flags = flags;
503 2250418 : }
504 :
505 264688 : transpose_vecs(&h_vecs[0]);
506 264688 : transpose_vecs(&h_vecs[4]);
507 : // The first four vecs now contain the first half of each output, and the
508 : // second four vecs contain the second half of each output.
509 264688 : storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
510 264688 : storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
511 264688 : storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
512 264688 : storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
513 264688 : storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
514 264688 : storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
515 264688 : storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
516 264688 : storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
517 264688 : }
518 :
519 : INLINE void hash_one_sse41(const uint8_t *input, size_t blocks,
520 : const uint32_t key[8], uint64_t counter,
521 : uint8_t flags, uint8_t flags_start,
522 1058924 : uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
523 1058924 : uint32_t cv[8];
524 1058924 : memcpy(cv, key, BLAKE3_KEY_LEN);
525 1058924 : uint8_t block_flags = flags | flags_start;
526 10061068 : while (blocks > 0) {
527 9002144 : if (blocks == 1) {
528 1058924 : block_flags |= flags_end;
529 1058924 : }
530 9002144 : fd_blake3_compress_in_place_sse41(cv, input, BLAKE3_BLOCK_LEN, counter,
531 9002144 : block_flags);
532 9002144 : input = &input[BLAKE3_BLOCK_LEN];
533 9002144 : blocks -= 1;
534 9002144 : block_flags = flags;
535 9002144 : }
536 1058924 : memcpy(out, cv, BLAKE3_OUT_LEN);
537 1058924 : }
538 :
539 : void fd_blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs,
540 : size_t blocks, const uint32_t key[8],
541 : uint64_t counter, bool increment_counter,
542 : uint8_t flags, uint8_t flags_start,
543 2114712 : uint8_t flags_end, uint8_t *out) {
544 2379400 : while (num_inputs >= DEGREE) {
545 264688 : fd_blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags,
546 264688 : flags_start, flags_end, out);
547 264688 : if (increment_counter) {
548 132382 : counter += DEGREE;
549 132382 : }
550 264688 : inputs += DEGREE;
551 264688 : num_inputs -= DEGREE;
552 264688 : out = &out[DEGREE * BLAKE3_OUT_LEN];
553 264688 : }
554 3173636 : while (num_inputs > 0) {
555 1058924 : hash_one_sse41(inputs[0], blocks, key, counter, flags, flags_start,
556 1058924 : flags_end, out);
557 1058924 : if (increment_counter) {
558 529548 : counter += 1;
559 529548 : }
560 1058924 : inputs += 1;
561 1058924 : num_inputs -= 1;
562 1058924 : out = &out[BLAKE3_OUT_LEN];
563 1058924 : }
564 2114712 : }
|
