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 0 : #define DEGREE 4
10 :
11 : #define _mm_shuffle_ps2(a, b, c) \
12 160 : (_mm_castps_si128( \
13 160 : _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
14 :
15 80 : INLINE __m128i loadu(const uint8_t src[16]) {
16 80 : return _mm_loadu_si128((const __m128i *)src);
17 80 : }
18 :
19 40 : INLINE void storeu(__m128i src, uint8_t dest[16]) {
20 40 : _mm_storeu_si128((__m128i *)dest, src);
21 40 : }
22 :
23 840 : 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.
26 880 : INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
27 :
28 0 : INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
29 :
30 20 : INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
31 20 : return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
32 20 : }
33 :
34 140 : INLINE __m128i rot16(__m128i x) {
35 140 : return _mm_shuffle_epi8(
36 140 : x, _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2));
37 140 : }
38 :
39 140 : INLINE __m128i rot12(__m128i x) {
40 140 : return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
41 140 : }
42 :
43 140 : INLINE __m128i rot8(__m128i x) {
44 140 : return _mm_shuffle_epi8(
45 140 : x, _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1));
46 140 : }
47 :
48 140 : INLINE __m128i rot7(__m128i x) {
49 140 : return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
50 140 : }
51 :
52 : INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
53 140 : __m128i m) {
54 140 : *row0 = addv(addv(*row0, m), *row1);
55 140 : *row3 = xorv(*row3, *row0);
56 140 : *row3 = rot16(*row3);
57 140 : *row2 = addv(*row2, *row3);
58 140 : *row1 = xorv(*row1, *row2);
59 140 : *row1 = rot12(*row1);
60 140 : }
61 :
62 : INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
63 140 : __m128i m) {
64 140 : *row0 = addv(addv(*row0, m), *row1);
65 140 : *row3 = xorv(*row3, *row0);
66 140 : *row3 = rot8(*row3);
67 140 : *row2 = addv(*row2, *row3);
68 140 : *row1 = xorv(*row1, *row2);
69 140 : *row1 = rot7(*row1);
70 140 : }
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 70 : INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
76 70 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
77 70 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
78 70 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
79 70 : }
80 :
81 70 : INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
82 70 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
83 70 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
84 70 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
85 70 : }
86 :
87 : INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
88 : const uint8_t block[BLAKE3_BLOCK_LEN],
89 10 : uint8_t block_len, uint64_t counter, uint8_t flags) {
90 10 : rows[0] = loadu((uint8_t *)&cv[0]);
91 10 : rows[1] = loadu((uint8_t *)&cv[4]);
92 10 : rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
93 10 : rows[3] = set4(counter_low(counter), counter_high(counter),
94 10 : (uint32_t)block_len, (uint32_t)flags);
95 :
96 10 : __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
97 10 : __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
98 10 : __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
99 10 : __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
100 :
101 10 : __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 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0
106 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
107 10 : t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1
108 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
109 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
110 10 : t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8
111 10 : t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14
112 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
113 10 : t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9
114 10 : t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15
115 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
116 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
117 10 : m0 = t0;
118 10 : m1 = t1;
119 10 : m2 = t2;
120 10 : 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 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
125 10 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
126 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
127 10 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
128 10 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
129 10 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
130 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
131 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
132 10 : t2 = _mm_unpacklo_epi64(m3, m1);
133 10 : tt = _mm_blend_epi16(t2, m2, 0xC0);
134 10 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
135 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
136 10 : t3 = _mm_unpackhi_epi32(m1, m3);
137 10 : tt = _mm_unpacklo_epi32(m2, t3);
138 10 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
139 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
140 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
141 10 : m0 = t0;
142 10 : m1 = t1;
143 10 : m2 = t2;
144 10 : m3 = t3;
145 :
146 : // Round 3
147 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
148 10 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
149 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
150 10 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
151 10 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
152 10 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
153 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
154 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
155 10 : t2 = _mm_unpacklo_epi64(m3, m1);
156 10 : tt = _mm_blend_epi16(t2, m2, 0xC0);
157 10 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
158 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
159 10 : t3 = _mm_unpackhi_epi32(m1, m3);
160 10 : tt = _mm_unpacklo_epi32(m2, t3);
161 10 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
162 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
163 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
164 10 : m0 = t0;
165 10 : m1 = t1;
166 10 : m2 = t2;
167 10 : m3 = t3;
168 :
169 : // Round 4
170 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
171 10 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
172 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
173 10 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
174 10 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
175 10 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
176 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
177 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
178 10 : t2 = _mm_unpacklo_epi64(m3, m1);
179 10 : tt = _mm_blend_epi16(t2, m2, 0xC0);
180 10 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
181 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
182 10 : t3 = _mm_unpackhi_epi32(m1, m3);
183 10 : tt = _mm_unpacklo_epi32(m2, t3);
184 10 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
185 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
186 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
187 10 : m0 = t0;
188 10 : m1 = t1;
189 10 : m2 = t2;
190 10 : m3 = t3;
191 :
192 : // Round 5
193 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
194 10 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
195 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
196 10 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
197 10 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
198 10 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
199 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
200 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
201 10 : t2 = _mm_unpacklo_epi64(m3, m1);
202 10 : tt = _mm_blend_epi16(t2, m2, 0xC0);
203 10 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
204 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
205 10 : t3 = _mm_unpackhi_epi32(m1, m3);
206 10 : tt = _mm_unpacklo_epi32(m2, t3);
207 10 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
208 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
209 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
210 10 : m0 = t0;
211 10 : m1 = t1;
212 10 : m2 = t2;
213 10 : m3 = t3;
214 :
215 : // Round 6
216 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
217 10 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
218 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
219 10 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
220 10 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
221 10 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
222 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
223 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
224 10 : t2 = _mm_unpacklo_epi64(m3, m1);
225 10 : tt = _mm_blend_epi16(t2, m2, 0xC0);
226 10 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
227 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
228 10 : t3 = _mm_unpackhi_epi32(m1, m3);
229 10 : tt = _mm_unpacklo_epi32(m2, t3);
230 10 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
231 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
232 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
233 10 : m0 = t0;
234 10 : m1 = t1;
235 10 : m2 = t2;
236 10 : m3 = t3;
237 :
238 : // Round 7
239 10 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
240 10 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
241 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
242 10 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
243 10 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
244 10 : t1 = _mm_blend_epi16(tt, t1, 0xCC);
245 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
246 10 : diagonalize(&rows[0], &rows[2], &rows[3]);
247 10 : t2 = _mm_unpacklo_epi64(m3, m1);
248 10 : tt = _mm_blend_epi16(t2, m2, 0xC0);
249 10 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
250 10 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
251 10 : t3 = _mm_unpackhi_epi32(m1, m3);
252 10 : tt = _mm_unpacklo_epi32(m2, t3);
253 10 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
254 10 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
255 10 : undiagonalize(&rows[0], &rows[2], &rows[3]);
256 10 : }
257 :
258 : void 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 0 : uint8_t flags) {
262 0 : __m128i rows[4];
263 0 : compress_pre(rows, cv, block, block_len, counter, flags);
264 0 : storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
265 0 : storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
266 0 : }
267 :
268 : void 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 10 : uint8_t flags, uint8_t out[64]) {
272 10 : __m128i rows[4];
273 10 : compress_pre(rows, cv, block, block_len, counter, flags);
274 10 : storeu(xorv(rows[0], rows[2]), &out[0]);
275 10 : storeu(xorv(rows[1], rows[3]), &out[16]);
276 10 : storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
277 10 : storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
278 10 : }
279 :
280 0 : INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
281 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
282 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
283 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
284 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
285 0 : v[0] = addv(v[0], v[4]);
286 0 : v[1] = addv(v[1], v[5]);
287 0 : v[2] = addv(v[2], v[6]);
288 0 : v[3] = addv(v[3], v[7]);
289 0 : v[12] = xorv(v[12], v[0]);
290 0 : v[13] = xorv(v[13], v[1]);
291 0 : v[14] = xorv(v[14], v[2]);
292 0 : v[15] = xorv(v[15], v[3]);
293 0 : v[12] = rot16(v[12]);
294 0 : v[13] = rot16(v[13]);
295 0 : v[14] = rot16(v[14]);
296 0 : v[15] = rot16(v[15]);
297 0 : v[8] = addv(v[8], v[12]);
298 0 : v[9] = addv(v[9], v[13]);
299 0 : v[10] = addv(v[10], v[14]);
300 0 : v[11] = addv(v[11], v[15]);
301 0 : v[4] = xorv(v[4], v[8]);
302 0 : v[5] = xorv(v[5], v[9]);
303 0 : v[6] = xorv(v[6], v[10]);
304 0 : v[7] = xorv(v[7], v[11]);
305 0 : v[4] = rot12(v[4]);
306 0 : v[5] = rot12(v[5]);
307 0 : v[6] = rot12(v[6]);
308 0 : v[7] = rot12(v[7]);
309 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
310 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
311 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
312 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
313 0 : v[0] = addv(v[0], v[4]);
314 0 : v[1] = addv(v[1], v[5]);
315 0 : v[2] = addv(v[2], v[6]);
316 0 : v[3] = addv(v[3], v[7]);
317 0 : v[12] = xorv(v[12], v[0]);
318 0 : v[13] = xorv(v[13], v[1]);
319 0 : v[14] = xorv(v[14], v[2]);
320 0 : v[15] = xorv(v[15], v[3]);
321 0 : v[12] = rot8(v[12]);
322 0 : v[13] = rot8(v[13]);
323 0 : v[14] = rot8(v[14]);
324 0 : v[15] = rot8(v[15]);
325 0 : v[8] = addv(v[8], v[12]);
326 0 : v[9] = addv(v[9], v[13]);
327 0 : v[10] = addv(v[10], v[14]);
328 0 : v[11] = addv(v[11], v[15]);
329 0 : v[4] = xorv(v[4], v[8]);
330 0 : v[5] = xorv(v[5], v[9]);
331 0 : v[6] = xorv(v[6], v[10]);
332 0 : v[7] = xorv(v[7], v[11]);
333 0 : v[4] = rot7(v[4]);
334 0 : v[5] = rot7(v[5]);
335 0 : v[6] = rot7(v[6]);
336 0 : v[7] = rot7(v[7]);
337 :
338 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
339 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
340 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
341 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
342 0 : v[0] = addv(v[0], v[5]);
343 0 : v[1] = addv(v[1], v[6]);
344 0 : v[2] = addv(v[2], v[7]);
345 0 : v[3] = addv(v[3], v[4]);
346 0 : v[15] = xorv(v[15], v[0]);
347 0 : v[12] = xorv(v[12], v[1]);
348 0 : v[13] = xorv(v[13], v[2]);
349 0 : v[14] = xorv(v[14], v[3]);
350 0 : v[15] = rot16(v[15]);
351 0 : v[12] = rot16(v[12]);
352 0 : v[13] = rot16(v[13]);
353 0 : v[14] = rot16(v[14]);
354 0 : v[10] = addv(v[10], v[15]);
355 0 : v[11] = addv(v[11], v[12]);
356 0 : v[8] = addv(v[8], v[13]);
357 0 : v[9] = addv(v[9], v[14]);
358 0 : v[5] = xorv(v[5], v[10]);
359 0 : v[6] = xorv(v[6], v[11]);
360 0 : v[7] = xorv(v[7], v[8]);
361 0 : v[4] = xorv(v[4], v[9]);
362 0 : v[5] = rot12(v[5]);
363 0 : v[6] = rot12(v[6]);
364 0 : v[7] = rot12(v[7]);
365 0 : v[4] = rot12(v[4]);
366 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
367 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
368 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
369 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
370 0 : v[0] = addv(v[0], v[5]);
371 0 : v[1] = addv(v[1], v[6]);
372 0 : v[2] = addv(v[2], v[7]);
373 0 : v[3] = addv(v[3], v[4]);
374 0 : v[15] = xorv(v[15], v[0]);
375 0 : v[12] = xorv(v[12], v[1]);
376 0 : v[13] = xorv(v[13], v[2]);
377 0 : v[14] = xorv(v[14], v[3]);
378 0 : v[15] = rot8(v[15]);
379 0 : v[12] = rot8(v[12]);
380 0 : v[13] = rot8(v[13]);
381 0 : v[14] = rot8(v[14]);
382 0 : v[10] = addv(v[10], v[15]);
383 0 : v[11] = addv(v[11], v[12]);
384 0 : v[8] = addv(v[8], v[13]);
385 0 : v[9] = addv(v[9], v[14]);
386 0 : v[5] = xorv(v[5], v[10]);
387 0 : v[6] = xorv(v[6], v[11]);
388 0 : v[7] = xorv(v[7], v[8]);
389 0 : v[4] = xorv(v[4], v[9]);
390 0 : v[5] = rot7(v[5]);
391 0 : v[6] = rot7(v[6]);
392 0 : v[7] = rot7(v[7]);
393 0 : v[4] = rot7(v[4]);
394 0 : }
395 :
396 0 : 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 0 : __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
401 0 : __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
402 0 : __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
403 0 : __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
404 :
405 : // Interleave 64-bit lanes.
406 0 : __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
407 0 : __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
408 0 : __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
409 0 : __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
410 :
411 0 : vecs[0] = abcd_0;
412 0 : vecs[1] = abcd_1;
413 0 : vecs[2] = abcd_2;
414 0 : vecs[3] = abcd_3;
415 0 : }
416 :
417 : INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
418 0 : size_t block_offset, __m128i out[16]) {
419 0 : out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
420 0 : out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
421 0 : out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
422 0 : out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
423 0 : out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
424 0 : out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
425 0 : out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
426 0 : out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
427 0 : out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
428 0 : out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
429 0 : out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
430 0 : out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
431 0 : out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
432 0 : out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
433 0 : out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
434 0 : out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
435 0 : for (size_t i = 0; i < 4; ++i) {
436 0 : _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
437 0 : }
438 0 : transpose_vecs(&out[0]);
439 0 : transpose_vecs(&out[4]);
440 0 : transpose_vecs(&out[8]);
441 0 : transpose_vecs(&out[12]);
442 0 : }
443 :
444 : INLINE void load_counters(uint64_t counter, bool increment_counter,
445 0 : __m128i *out_lo, __m128i *out_hi) {
446 0 : const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
447 0 : const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
448 0 : const __m128i add1 = _mm_and_si128(mask, add0);
449 0 : __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1);
450 0 : __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32((int)0x80000000)),
451 0 : _mm_xor_si128( l, _mm_set1_epi32((int)0x80000000)));
452 0 : __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry);
453 0 : *out_lo = l;
454 0 : *out_hi = h;
455 0 : }
456 :
457 : static
458 : void 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 0 : uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
462 0 : __m128i h_vecs[8] = {
463 0 : set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
464 0 : set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
465 0 : };
466 0 : __m128i counter_low_vec, counter_high_vec;
467 0 : load_counters(counter, increment_counter, &counter_low_vec,
468 0 : &counter_high_vec);
469 0 : uint8_t block_flags = flags | flags_start;
470 :
471 0 : for (size_t block = 0; block < blocks; block++) {
472 0 : if (block + 1 == blocks) {
473 0 : block_flags |= flags_end;
474 0 : }
475 0 : __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
476 0 : __m128i block_flags_vec = set1(block_flags);
477 0 : __m128i msg_vecs[16];
478 0 : transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
479 :
480 0 : __m128i v[16] = {
481 0 : h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
482 0 : h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
483 0 : set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]),
484 0 : counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
485 0 : };
486 0 : round_fn(v, msg_vecs, 0);
487 0 : round_fn(v, msg_vecs, 1);
488 0 : round_fn(v, msg_vecs, 2);
489 0 : round_fn(v, msg_vecs, 3);
490 0 : round_fn(v, msg_vecs, 4);
491 0 : round_fn(v, msg_vecs, 5);
492 0 : round_fn(v, msg_vecs, 6);
493 0 : h_vecs[0] = xorv(v[0], v[8]);
494 0 : h_vecs[1] = xorv(v[1], v[9]);
495 0 : h_vecs[2] = xorv(v[2], v[10]);
496 0 : h_vecs[3] = xorv(v[3], v[11]);
497 0 : h_vecs[4] = xorv(v[4], v[12]);
498 0 : h_vecs[5] = xorv(v[5], v[13]);
499 0 : h_vecs[6] = xorv(v[6], v[14]);
500 0 : h_vecs[7] = xorv(v[7], v[15]);
501 :
502 0 : block_flags = flags;
503 0 : }
504 :
505 0 : transpose_vecs(&h_vecs[0]);
506 0 : 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 0 : storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
510 0 : storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
511 0 : storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
512 0 : storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
513 0 : storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
514 0 : storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
515 0 : storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
516 0 : storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
517 0 : }
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 0 : uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
523 0 : uint32_t cv[8];
524 0 : memcpy(cv, key, BLAKE3_KEY_LEN);
525 0 : uint8_t block_flags = flags | flags_start;
526 0 : while (blocks > 0) {
527 0 : if (blocks == 1) {
528 0 : block_flags |= flags_end;
529 0 : }
530 0 : blake3_compress_in_place_sse41(cv, input, BLAKE3_BLOCK_LEN, counter,
531 0 : block_flags);
532 0 : input = &input[BLAKE3_BLOCK_LEN];
533 0 : blocks -= 1;
534 0 : block_flags = flags;
535 0 : }
536 0 : memcpy(out, cv, BLAKE3_OUT_LEN);
537 0 : }
538 :
539 : void 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 0 : uint8_t flags_end, uint8_t *out) {
544 0 : while (num_inputs >= DEGREE) {
545 0 : blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags,
546 0 : flags_start, flags_end, out);
547 0 : if (increment_counter) {
548 0 : counter += DEGREE;
549 0 : }
550 0 : inputs += DEGREE;
551 0 : num_inputs -= DEGREE;
552 0 : out = &out[DEGREE * BLAKE3_OUT_LEN];
553 0 : }
554 0 : while (num_inputs > 0) {
555 0 : hash_one_sse41(inputs[0], blocks, key, counter, flags, flags_start,
556 0 : flags_end, out);
557 0 : if (increment_counter) {
558 0 : counter += 1;
559 0 : }
560 0 : inputs += 1;
561 0 : num_inputs -= 1;
562 0 : out = &out[BLAKE3_OUT_LEN];
563 0 : }
564 0 : }
|
