Line data Source code
1 :
2 : // Source originally from https://github.com/BLAKE3-team/BLAKE3
3 : // From commit: 2dd4e57f68d85f3983b1880b66250fc7bdf0b7c8
4 :
5 : #include "blake3_impl.h"
6 :
7 : #include <immintrin.h>
8 :
9 : #pragma GCC diagnostic ignored "-Wsign-conversion"
10 :
11 0 : #define DEGREE 4
12 :
13 : #define _mm_shuffle_ps2(a, b, c) \
14 0 : (_mm_castps_si128( \
15 0 : _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
16 :
17 0 : INLINE __m128i loadu(const uint8_t src[16]) {
18 0 : return _mm_loadu_si128((const __m128i *)src);
19 0 : }
20 :
21 0 : INLINE void storeu(__m128i src, uint8_t dest[16]) {
22 0 : _mm_storeu_si128((__m128i *)dest, src);
23 0 : }
24 :
25 0 : INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
26 :
27 : // Note that clang-format doesn't like the name "xor" for some reason.
28 0 : INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
29 :
30 0 : INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
31 :
32 0 : INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
33 0 : return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
34 0 : }
35 :
36 0 : INLINE __m128i rot16(__m128i x) {
37 0 : return _mm_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1);
38 0 : }
39 :
40 0 : INLINE __m128i rot12(__m128i x) {
41 0 : return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
42 0 : }
43 :
44 0 : INLINE __m128i rot8(__m128i x) {
45 0 : return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8));
46 0 : }
47 :
48 0 : INLINE __m128i rot7(__m128i x) {
49 0 : return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
50 0 : }
51 :
52 : INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
53 0 : __m128i m) {
54 0 : *row0 = addv(addv(*row0, m), *row1);
55 0 : *row3 = xorv(*row3, *row0);
56 0 : *row3 = rot16(*row3);
57 0 : *row2 = addv(*row2, *row3);
58 0 : *row1 = xorv(*row1, *row2);
59 0 : *row1 = rot12(*row1);
60 0 : }
61 :
62 : INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
63 0 : __m128i m) {
64 0 : *row0 = addv(addv(*row0, m), *row1);
65 0 : *row3 = xorv(*row3, *row0);
66 0 : *row3 = rot8(*row3);
67 0 : *row2 = addv(*row2, *row3);
68 0 : *row1 = xorv(*row1, *row2);
69 0 : *row1 = rot7(*row1);
70 0 : }
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 0 : INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
76 0 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
77 0 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
78 0 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
79 0 : }
80 :
81 0 : INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
82 0 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
83 0 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
84 0 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
85 0 : }
86 :
87 0 : INLINE __m128i blend_epi16(__m128i a, __m128i b, const int16_t imm8) {
88 0 : const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
89 0 : __m128i mask = _mm_set1_epi16(imm8);
90 0 : mask = _mm_and_si128(mask, bits);
91 0 : mask = _mm_cmpeq_epi16(mask, bits);
92 0 : return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
93 0 : }
94 :
95 : INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
96 : const uint8_t block[BLAKE3_BLOCK_LEN],
97 0 : uint8_t block_len, uint64_t counter, uint8_t flags) {
98 0 : rows[0] = loadu((uint8_t *)&cv[0]);
99 0 : rows[1] = loadu((uint8_t *)&cv[4]);
100 0 : rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
101 0 : rows[3] = set4(counter_low(counter), counter_high(counter),
102 0 : (uint32_t)block_len, (uint32_t)flags);
103 :
104 0 : __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
105 0 : __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
106 0 : __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
107 0 : __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
108 :
109 0 : __m128i t0, t1, t2, t3, tt;
110 :
111 : // Round 1. The first round permutes the message words from the original
112 : // input order, into the groups that get mixed in parallel.
113 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0
114 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
115 0 : t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1
116 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
117 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
118 0 : t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8
119 0 : t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14
120 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
121 0 : t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9
122 0 : t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15
123 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
124 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
125 0 : m0 = t0;
126 0 : m1 = t1;
127 0 : m2 = t2;
128 0 : m3 = t3;
129 :
130 : // Round 2. This round and all following rounds apply a fixed permutation
131 : // to the message words from the round before.
132 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
133 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
134 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
135 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
136 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
137 0 : t1 = blend_epi16(tt, t1, 0xCC);
138 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
139 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
140 0 : t2 = _mm_unpacklo_epi64(m3, m1);
141 0 : tt = blend_epi16(t2, m2, 0xC0);
142 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
143 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
144 0 : t3 = _mm_unpackhi_epi32(m1, m3);
145 0 : tt = _mm_unpacklo_epi32(m2, t3);
146 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
147 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
148 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
149 0 : m0 = t0;
150 0 : m1 = t1;
151 0 : m2 = t2;
152 0 : m3 = t3;
153 :
154 : // Round 3
155 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
156 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
157 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
158 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
159 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
160 0 : t1 = blend_epi16(tt, t1, 0xCC);
161 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
162 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
163 0 : t2 = _mm_unpacklo_epi64(m3, m1);
164 0 : tt = blend_epi16(t2, m2, 0xC0);
165 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
166 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
167 0 : t3 = _mm_unpackhi_epi32(m1, m3);
168 0 : tt = _mm_unpacklo_epi32(m2, t3);
169 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
170 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
171 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
172 0 : m0 = t0;
173 0 : m1 = t1;
174 0 : m2 = t2;
175 0 : m3 = t3;
176 :
177 : // Round 4
178 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
179 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
180 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
181 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
182 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
183 0 : t1 = blend_epi16(tt, t1, 0xCC);
184 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
185 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
186 0 : t2 = _mm_unpacklo_epi64(m3, m1);
187 0 : tt = blend_epi16(t2, m2, 0xC0);
188 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
189 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
190 0 : t3 = _mm_unpackhi_epi32(m1, m3);
191 0 : tt = _mm_unpacklo_epi32(m2, t3);
192 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
193 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
194 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
195 0 : m0 = t0;
196 0 : m1 = t1;
197 0 : m2 = t2;
198 0 : m3 = t3;
199 :
200 : // Round 5
201 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
202 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
203 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
204 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
205 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
206 0 : t1 = blend_epi16(tt, t1, 0xCC);
207 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
208 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
209 0 : t2 = _mm_unpacklo_epi64(m3, m1);
210 0 : tt = blend_epi16(t2, m2, 0xC0);
211 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
212 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
213 0 : t3 = _mm_unpackhi_epi32(m1, m3);
214 0 : tt = _mm_unpacklo_epi32(m2, t3);
215 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
216 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
217 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
218 0 : m0 = t0;
219 0 : m1 = t1;
220 0 : m2 = t2;
221 0 : m3 = t3;
222 :
223 : // Round 6
224 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
225 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
226 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
227 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
228 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
229 0 : t1 = blend_epi16(tt, t1, 0xCC);
230 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
231 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
232 0 : t2 = _mm_unpacklo_epi64(m3, m1);
233 0 : tt = blend_epi16(t2, m2, 0xC0);
234 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
235 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
236 0 : t3 = _mm_unpackhi_epi32(m1, m3);
237 0 : tt = _mm_unpacklo_epi32(m2, t3);
238 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
239 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
240 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
241 0 : m0 = t0;
242 0 : m1 = t1;
243 0 : m2 = t2;
244 0 : m3 = t3;
245 :
246 : // Round 7
247 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
248 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
249 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
250 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
251 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
252 0 : t1 = blend_epi16(tt, t1, 0xCC);
253 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
254 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
255 0 : t2 = _mm_unpacklo_epi64(m3, m1);
256 0 : tt = blend_epi16(t2, m2, 0xC0);
257 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
258 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
259 0 : t3 = _mm_unpackhi_epi32(m1, m3);
260 0 : tt = _mm_unpacklo_epi32(m2, t3);
261 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
262 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
263 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
264 0 : }
265 :
266 : void blake3_compress_in_place_sse2(uint32_t cv[8],
267 : const uint8_t block[BLAKE3_BLOCK_LEN],
268 : uint8_t block_len, uint64_t counter,
269 0 : uint8_t flags) {
270 0 : __m128i rows[4];
271 0 : compress_pre(rows, cv, block, block_len, counter, flags);
272 0 : storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
273 0 : storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
274 0 : }
275 :
276 : void blake3_compress_xof_sse2(const uint32_t cv[8],
277 : const uint8_t block[BLAKE3_BLOCK_LEN],
278 : uint8_t block_len, uint64_t counter,
279 0 : uint8_t flags, uint8_t out[64]) {
280 0 : __m128i rows[4];
281 0 : compress_pre(rows, cv, block, block_len, counter, flags);
282 0 : storeu(xorv(rows[0], rows[2]), &out[0]);
283 0 : storeu(xorv(rows[1], rows[3]), &out[16]);
284 0 : storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
285 0 : storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
286 0 : }
287 :
288 0 : INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
289 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
290 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
291 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
292 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
293 0 : v[0] = addv(v[0], v[4]);
294 0 : v[1] = addv(v[1], v[5]);
295 0 : v[2] = addv(v[2], v[6]);
296 0 : v[3] = addv(v[3], v[7]);
297 0 : v[12] = xorv(v[12], v[0]);
298 0 : v[13] = xorv(v[13], v[1]);
299 0 : v[14] = xorv(v[14], v[2]);
300 0 : v[15] = xorv(v[15], v[3]);
301 0 : v[12] = rot16(v[12]);
302 0 : v[13] = rot16(v[13]);
303 0 : v[14] = rot16(v[14]);
304 0 : v[15] = rot16(v[15]);
305 0 : v[8] = addv(v[8], v[12]);
306 0 : v[9] = addv(v[9], v[13]);
307 0 : v[10] = addv(v[10], v[14]);
308 0 : v[11] = addv(v[11], v[15]);
309 0 : v[4] = xorv(v[4], v[8]);
310 0 : v[5] = xorv(v[5], v[9]);
311 0 : v[6] = xorv(v[6], v[10]);
312 0 : v[7] = xorv(v[7], v[11]);
313 0 : v[4] = rot12(v[4]);
314 0 : v[5] = rot12(v[5]);
315 0 : v[6] = rot12(v[6]);
316 0 : v[7] = rot12(v[7]);
317 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
318 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
319 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
320 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
321 0 : v[0] = addv(v[0], v[4]);
322 0 : v[1] = addv(v[1], v[5]);
323 0 : v[2] = addv(v[2], v[6]);
324 0 : v[3] = addv(v[3], v[7]);
325 0 : v[12] = xorv(v[12], v[0]);
326 0 : v[13] = xorv(v[13], v[1]);
327 0 : v[14] = xorv(v[14], v[2]);
328 0 : v[15] = xorv(v[15], v[3]);
329 0 : v[12] = rot8(v[12]);
330 0 : v[13] = rot8(v[13]);
331 0 : v[14] = rot8(v[14]);
332 0 : v[15] = rot8(v[15]);
333 0 : v[8] = addv(v[8], v[12]);
334 0 : v[9] = addv(v[9], v[13]);
335 0 : v[10] = addv(v[10], v[14]);
336 0 : v[11] = addv(v[11], v[15]);
337 0 : v[4] = xorv(v[4], v[8]);
338 0 : v[5] = xorv(v[5], v[9]);
339 0 : v[6] = xorv(v[6], v[10]);
340 0 : v[7] = xorv(v[7], v[11]);
341 0 : v[4] = rot7(v[4]);
342 0 : v[5] = rot7(v[5]);
343 0 : v[6] = rot7(v[6]);
344 0 : v[7] = rot7(v[7]);
345 :
346 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
347 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
348 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
349 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
350 0 : v[0] = addv(v[0], v[5]);
351 0 : v[1] = addv(v[1], v[6]);
352 0 : v[2] = addv(v[2], v[7]);
353 0 : v[3] = addv(v[3], v[4]);
354 0 : v[15] = xorv(v[15], v[0]);
355 0 : v[12] = xorv(v[12], v[1]);
356 0 : v[13] = xorv(v[13], v[2]);
357 0 : v[14] = xorv(v[14], v[3]);
358 0 : v[15] = rot16(v[15]);
359 0 : v[12] = rot16(v[12]);
360 0 : v[13] = rot16(v[13]);
361 0 : v[14] = rot16(v[14]);
362 0 : v[10] = addv(v[10], v[15]);
363 0 : v[11] = addv(v[11], v[12]);
364 0 : v[8] = addv(v[8], v[13]);
365 0 : v[9] = addv(v[9], v[14]);
366 0 : v[5] = xorv(v[5], v[10]);
367 0 : v[6] = xorv(v[6], v[11]);
368 0 : v[7] = xorv(v[7], v[8]);
369 0 : v[4] = xorv(v[4], v[9]);
370 0 : v[5] = rot12(v[5]);
371 0 : v[6] = rot12(v[6]);
372 0 : v[7] = rot12(v[7]);
373 0 : v[4] = rot12(v[4]);
374 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
375 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
376 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
377 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
378 0 : v[0] = addv(v[0], v[5]);
379 0 : v[1] = addv(v[1], v[6]);
380 0 : v[2] = addv(v[2], v[7]);
381 0 : v[3] = addv(v[3], v[4]);
382 0 : v[15] = xorv(v[15], v[0]);
383 0 : v[12] = xorv(v[12], v[1]);
384 0 : v[13] = xorv(v[13], v[2]);
385 0 : v[14] = xorv(v[14], v[3]);
386 0 : v[15] = rot8(v[15]);
387 0 : v[12] = rot8(v[12]);
388 0 : v[13] = rot8(v[13]);
389 0 : v[14] = rot8(v[14]);
390 0 : v[10] = addv(v[10], v[15]);
391 0 : v[11] = addv(v[11], v[12]);
392 0 : v[8] = addv(v[8], v[13]);
393 0 : v[9] = addv(v[9], v[14]);
394 0 : v[5] = xorv(v[5], v[10]);
395 0 : v[6] = xorv(v[6], v[11]);
396 0 : v[7] = xorv(v[7], v[8]);
397 0 : v[4] = xorv(v[4], v[9]);
398 0 : v[5] = rot7(v[5]);
399 0 : v[6] = rot7(v[6]);
400 0 : v[7] = rot7(v[7]);
401 0 : v[4] = rot7(v[4]);
402 0 : }
403 :
404 0 : INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
405 : // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is
406 : // 22/33. Note that this doesn't split the vector into two lanes, as the
407 : // AVX2 counterparts do.
408 0 : __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
409 0 : __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
410 0 : __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
411 0 : __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
412 :
413 : // Interleave 64-bit lanes.
414 0 : __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
415 0 : __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
416 0 : __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
417 0 : __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
418 :
419 0 : vecs[0] = abcd_0;
420 0 : vecs[1] = abcd_1;
421 0 : vecs[2] = abcd_2;
422 0 : vecs[3] = abcd_3;
423 0 : }
424 :
425 : INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
426 0 : size_t block_offset, __m128i out[16]) {
427 0 : out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
428 0 : out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
429 0 : out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
430 0 : out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
431 0 : out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
432 0 : out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
433 0 : out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
434 0 : out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
435 0 : out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
436 0 : out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
437 0 : out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
438 0 : out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
439 0 : out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
440 0 : out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
441 0 : out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
442 0 : out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
443 0 : for (size_t i = 0; i < 4; ++i) {
444 0 : _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
445 0 : }
446 0 : transpose_vecs(&out[0]);
447 0 : transpose_vecs(&out[4]);
448 0 : transpose_vecs(&out[8]);
449 0 : transpose_vecs(&out[12]);
450 0 : }
451 :
452 : INLINE void load_counters(uint64_t counter, bool increment_counter,
453 0 : __m128i *out_lo, __m128i *out_hi) {
454 0 : const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
455 0 : const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
456 0 : const __m128i add1 = _mm_and_si128(mask, add0);
457 0 : __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1);
458 0 : __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)),
459 0 : _mm_xor_si128( l, _mm_set1_epi32(0x80000000)));
460 0 : __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry);
461 0 : *out_lo = l;
462 0 : *out_hi = h;
463 0 : }
464 :
465 : static
466 : void blake3_hash4_sse2(const uint8_t *const *inputs, size_t blocks,
467 : const uint32_t key[8], uint64_t counter,
468 : bool increment_counter, uint8_t flags,
469 0 : uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
470 0 : __m128i h_vecs[8] = {
471 0 : set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
472 0 : set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
473 0 : };
474 0 : __m128i counter_low_vec, counter_high_vec;
475 0 : load_counters(counter, increment_counter, &counter_low_vec,
476 0 : &counter_high_vec);
477 0 : uint8_t block_flags = flags | flags_start;
478 :
479 0 : for (size_t block = 0; block < blocks; block++) {
480 0 : if (block + 1 == blocks) {
481 0 : block_flags |= flags_end;
482 0 : }
483 0 : __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
484 0 : __m128i block_flags_vec = set1(block_flags);
485 0 : __m128i msg_vecs[16];
486 0 : transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
487 :
488 0 : __m128i v[16] = {
489 0 : h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
490 0 : h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
491 0 : set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]),
492 0 : counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
493 0 : };
494 0 : round_fn(v, msg_vecs, 0);
495 0 : round_fn(v, msg_vecs, 1);
496 0 : round_fn(v, msg_vecs, 2);
497 0 : round_fn(v, msg_vecs, 3);
498 0 : round_fn(v, msg_vecs, 4);
499 0 : round_fn(v, msg_vecs, 5);
500 0 : round_fn(v, msg_vecs, 6);
501 0 : h_vecs[0] = xorv(v[0], v[8]);
502 0 : h_vecs[1] = xorv(v[1], v[9]);
503 0 : h_vecs[2] = xorv(v[2], v[10]);
504 0 : h_vecs[3] = xorv(v[3], v[11]);
505 0 : h_vecs[4] = xorv(v[4], v[12]);
506 0 : h_vecs[5] = xorv(v[5], v[13]);
507 0 : h_vecs[6] = xorv(v[6], v[14]);
508 0 : h_vecs[7] = xorv(v[7], v[15]);
509 :
510 0 : block_flags = flags;
511 0 : }
512 :
513 0 : transpose_vecs(&h_vecs[0]);
514 0 : transpose_vecs(&h_vecs[4]);
515 : // The first four vecs now contain the first half of each output, and the
516 : // second four vecs contain the second half of each output.
517 0 : storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
518 0 : storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
519 0 : storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
520 0 : storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
521 0 : storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
522 0 : storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
523 0 : storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
524 0 : storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
525 0 : }
526 :
527 : INLINE void hash_one_sse2(const uint8_t *input, size_t blocks,
528 : const uint32_t key[8], uint64_t counter,
529 : uint8_t flags, uint8_t flags_start,
530 0 : uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
531 0 : uint32_t cv[8];
532 0 : memcpy(cv, key, BLAKE3_KEY_LEN);
533 0 : uint8_t block_flags = flags | flags_start;
534 0 : while (blocks > 0) {
535 0 : if (blocks == 1) {
536 0 : block_flags |= flags_end;
537 0 : }
538 0 : blake3_compress_in_place_sse2(cv, input, BLAKE3_BLOCK_LEN, counter,
539 0 : block_flags);
540 0 : input = &input[BLAKE3_BLOCK_LEN];
541 0 : blocks -= 1;
542 0 : block_flags = flags;
543 0 : }
544 0 : memcpy(out, cv, BLAKE3_OUT_LEN);
545 0 : }
546 :
547 : void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
548 : size_t blocks, const uint32_t key[8],
549 : uint64_t counter, bool increment_counter,
550 : uint8_t flags, uint8_t flags_start,
551 0 : uint8_t flags_end, uint8_t *out) {
552 0 : while (num_inputs >= DEGREE) {
553 0 : blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags,
554 0 : flags_start, flags_end, out);
555 0 : if (increment_counter) {
556 0 : counter += DEGREE;
557 0 : }
558 0 : inputs += DEGREE;
559 0 : num_inputs -= DEGREE;
560 0 : out = &out[DEGREE * BLAKE3_OUT_LEN];
561 0 : }
562 0 : while (num_inputs > 0) {
563 0 : hash_one_sse2(inputs[0], blocks, key, counter, flags, flags_start,
564 0 : flags_end, out);
565 0 : if (increment_counter) {
566 0 : counter += 1;
567 0 : }
568 0 : inputs += 1;
569 0 : num_inputs -= 1;
570 0 : out = &out[BLAKE3_OUT_LEN];
571 0 : }
572 0 : }
|
