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 0 : (_mm_castps_si128( \
13 0 : _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
14 :
15 0 : INLINE __m128i loadu(const uint8_t src[16]) {
16 0 : return _mm_loadu_si128((const __m128i *)src);
17 0 : }
18 :
19 0 : INLINE void storeu(__m128i src, uint8_t dest[16]) {
20 0 : _mm_storeu_si128((__m128i *)dest, src);
21 0 : }
22 :
23 0 : 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 0 : 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 0 : INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
31 0 : return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
32 0 : }
33 :
34 0 : INLINE __m128i rot16(__m128i x) {
35 0 : return _mm_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1);
36 0 : }
37 :
38 0 : INLINE __m128i rot12(__m128i x) {
39 0 : return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
40 0 : }
41 :
42 0 : INLINE __m128i rot8(__m128i x) {
43 0 : return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8));
44 0 : }
45 :
46 0 : INLINE __m128i rot7(__m128i x) {
47 0 : return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
48 0 : }
49 :
50 : INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
51 0 : __m128i m) {
52 0 : *row0 = addv(addv(*row0, m), *row1);
53 0 : *row3 = xorv(*row3, *row0);
54 0 : *row3 = rot16(*row3);
55 0 : *row2 = addv(*row2, *row3);
56 0 : *row1 = xorv(*row1, *row2);
57 0 : *row1 = rot12(*row1);
58 0 : }
59 :
60 : INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
61 0 : __m128i m) {
62 0 : *row0 = addv(addv(*row0, m), *row1);
63 0 : *row3 = xorv(*row3, *row0);
64 0 : *row3 = rot8(*row3);
65 0 : *row2 = addv(*row2, *row3);
66 0 : *row1 = xorv(*row1, *row2);
67 0 : *row1 = rot7(*row1);
68 0 : }
69 :
70 : // Note the optimization here of leaving row1 as the unrotated row, rather than
71 : // row0. All the message loads below are adjusted to compensate for this. See
72 : // discussion at https://github.com/sneves/blake2-avx2/pull/4
73 0 : INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
74 0 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
75 0 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
76 0 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
77 0 : }
78 :
79 0 : INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
80 0 : *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
81 0 : *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
82 0 : *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
83 0 : }
84 :
85 0 : INLINE __m128i blend_epi16(__m128i a, __m128i b, const int16_t imm8) {
86 0 : const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
87 0 : __m128i mask = _mm_set1_epi16(imm8);
88 0 : mask = _mm_and_si128(mask, bits);
89 0 : mask = _mm_cmpeq_epi16(mask, bits);
90 0 : return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
91 0 : }
92 :
93 : INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
94 : const uint8_t block[BLAKE3_BLOCK_LEN],
95 0 : uint8_t block_len, uint64_t counter, uint8_t flags) {
96 0 : rows[0] = loadu((uint8_t *)&cv[0]);
97 0 : rows[1] = loadu((uint8_t *)&cv[4]);
98 0 : rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
99 0 : rows[3] = set4(counter_low(counter), counter_high(counter),
100 0 : (uint32_t)block_len, (uint32_t)flags);
101 :
102 0 : __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
103 0 : __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
104 0 : __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
105 0 : __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
106 :
107 0 : __m128i t0, t1, t2, t3, tt;
108 :
109 : // Round 1. The first round permutes the message words from the original
110 : // input order, into the groups that get mixed in parallel.
111 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0
112 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
113 0 : t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1
114 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
115 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
116 0 : t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8
117 0 : t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14
118 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
119 0 : t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9
120 0 : t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15
121 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
122 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
123 0 : m0 = t0;
124 0 : m1 = t1;
125 0 : m2 = t2;
126 0 : m3 = t3;
127 :
128 : // Round 2. This round and all following rounds apply a fixed permutation
129 : // to the message words from the round before.
130 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
131 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
132 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
133 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
134 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
135 0 : t1 = blend_epi16(tt, t1, 0xCC);
136 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
137 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
138 0 : t2 = _mm_unpacklo_epi64(m3, m1);
139 0 : tt = blend_epi16(t2, m2, 0xC0);
140 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
141 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
142 0 : t3 = _mm_unpackhi_epi32(m1, m3);
143 0 : tt = _mm_unpacklo_epi32(m2, t3);
144 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
145 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
146 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
147 0 : m0 = t0;
148 0 : m1 = t1;
149 0 : m2 = t2;
150 0 : m3 = t3;
151 :
152 : // Round 3
153 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
154 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
155 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
156 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
157 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
158 0 : t1 = blend_epi16(tt, t1, 0xCC);
159 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
160 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
161 0 : t2 = _mm_unpacklo_epi64(m3, m1);
162 0 : tt = blend_epi16(t2, m2, 0xC0);
163 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
164 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
165 0 : t3 = _mm_unpackhi_epi32(m1, m3);
166 0 : tt = _mm_unpacklo_epi32(m2, t3);
167 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
168 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
169 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
170 0 : m0 = t0;
171 0 : m1 = t1;
172 0 : m2 = t2;
173 0 : m3 = t3;
174 :
175 : // Round 4
176 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
177 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
178 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
179 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
180 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
181 0 : t1 = blend_epi16(tt, t1, 0xCC);
182 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
183 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
184 0 : t2 = _mm_unpacklo_epi64(m3, m1);
185 0 : tt = blend_epi16(t2, m2, 0xC0);
186 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
187 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
188 0 : t3 = _mm_unpackhi_epi32(m1, m3);
189 0 : tt = _mm_unpacklo_epi32(m2, t3);
190 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
191 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
192 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
193 0 : m0 = t0;
194 0 : m1 = t1;
195 0 : m2 = t2;
196 0 : m3 = t3;
197 :
198 : // Round 5
199 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
200 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
201 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
202 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
203 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
204 0 : t1 = blend_epi16(tt, t1, 0xCC);
205 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
206 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
207 0 : t2 = _mm_unpacklo_epi64(m3, m1);
208 0 : tt = blend_epi16(t2, m2, 0xC0);
209 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
210 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
211 0 : t3 = _mm_unpackhi_epi32(m1, m3);
212 0 : tt = _mm_unpacklo_epi32(m2, t3);
213 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
214 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
215 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
216 0 : m0 = t0;
217 0 : m1 = t1;
218 0 : m2 = t2;
219 0 : m3 = t3;
220 :
221 : // Round 6
222 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
223 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
224 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
225 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
226 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
227 0 : t1 = blend_epi16(tt, t1, 0xCC);
228 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
229 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
230 0 : t2 = _mm_unpacklo_epi64(m3, m1);
231 0 : tt = blend_epi16(t2, m2, 0xC0);
232 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
233 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
234 0 : t3 = _mm_unpackhi_epi32(m1, m3);
235 0 : tt = _mm_unpacklo_epi32(m2, t3);
236 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
237 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
238 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
239 0 : m0 = t0;
240 0 : m1 = t1;
241 0 : m2 = t2;
242 0 : m3 = t3;
243 :
244 : // Round 7
245 0 : t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
246 0 : t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
247 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
248 0 : t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
249 0 : tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
250 0 : t1 = blend_epi16(tt, t1, 0xCC);
251 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
252 0 : diagonalize(&rows[0], &rows[2], &rows[3]);
253 0 : t2 = _mm_unpacklo_epi64(m3, m1);
254 0 : tt = blend_epi16(t2, m2, 0xC0);
255 0 : t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
256 0 : g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
257 0 : t3 = _mm_unpackhi_epi32(m1, m3);
258 0 : tt = _mm_unpacklo_epi32(m2, t3);
259 0 : t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
260 0 : g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
261 0 : undiagonalize(&rows[0], &rows[2], &rows[3]);
262 0 : }
263 :
264 : void blake3_compress_in_place_sse2(uint32_t cv[8],
265 : const uint8_t block[BLAKE3_BLOCK_LEN],
266 : uint8_t block_len, uint64_t counter,
267 0 : uint8_t flags) {
268 0 : __m128i rows[4];
269 0 : compress_pre(rows, cv, block, block_len, counter, flags);
270 0 : storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
271 0 : storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
272 0 : }
273 :
274 : void blake3_compress_xof_sse2(const uint32_t cv[8],
275 : const uint8_t block[BLAKE3_BLOCK_LEN],
276 : uint8_t block_len, uint64_t counter,
277 0 : uint8_t flags, uint8_t out[64]) {
278 0 : __m128i rows[4];
279 0 : compress_pre(rows, cv, block, block_len, counter, flags);
280 0 : storeu(xorv(rows[0], rows[2]), &out[0]);
281 0 : storeu(xorv(rows[1], rows[3]), &out[16]);
282 0 : storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
283 0 : storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
284 0 : }
285 :
286 0 : INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
287 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
288 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
289 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
290 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
291 0 : v[0] = addv(v[0], v[4]);
292 0 : v[1] = addv(v[1], v[5]);
293 0 : v[2] = addv(v[2], v[6]);
294 0 : v[3] = addv(v[3], v[7]);
295 0 : v[12] = xorv(v[12], v[0]);
296 0 : v[13] = xorv(v[13], v[1]);
297 0 : v[14] = xorv(v[14], v[2]);
298 0 : v[15] = xorv(v[15], v[3]);
299 0 : v[12] = rot16(v[12]);
300 0 : v[13] = rot16(v[13]);
301 0 : v[14] = rot16(v[14]);
302 0 : v[15] = rot16(v[15]);
303 0 : v[8] = addv(v[8], v[12]);
304 0 : v[9] = addv(v[9], v[13]);
305 0 : v[10] = addv(v[10], v[14]);
306 0 : v[11] = addv(v[11], v[15]);
307 0 : v[4] = xorv(v[4], v[8]);
308 0 : v[5] = xorv(v[5], v[9]);
309 0 : v[6] = xorv(v[6], v[10]);
310 0 : v[7] = xorv(v[7], v[11]);
311 0 : v[4] = rot12(v[4]);
312 0 : v[5] = rot12(v[5]);
313 0 : v[6] = rot12(v[6]);
314 0 : v[7] = rot12(v[7]);
315 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
316 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
317 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
318 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
319 0 : v[0] = addv(v[0], v[4]);
320 0 : v[1] = addv(v[1], v[5]);
321 0 : v[2] = addv(v[2], v[6]);
322 0 : v[3] = addv(v[3], v[7]);
323 0 : v[12] = xorv(v[12], v[0]);
324 0 : v[13] = xorv(v[13], v[1]);
325 0 : v[14] = xorv(v[14], v[2]);
326 0 : v[15] = xorv(v[15], v[3]);
327 0 : v[12] = rot8(v[12]);
328 0 : v[13] = rot8(v[13]);
329 0 : v[14] = rot8(v[14]);
330 0 : v[15] = rot8(v[15]);
331 0 : v[8] = addv(v[8], v[12]);
332 0 : v[9] = addv(v[9], v[13]);
333 0 : v[10] = addv(v[10], v[14]);
334 0 : v[11] = addv(v[11], v[15]);
335 0 : v[4] = xorv(v[4], v[8]);
336 0 : v[5] = xorv(v[5], v[9]);
337 0 : v[6] = xorv(v[6], v[10]);
338 0 : v[7] = xorv(v[7], v[11]);
339 0 : v[4] = rot7(v[4]);
340 0 : v[5] = rot7(v[5]);
341 0 : v[6] = rot7(v[6]);
342 0 : v[7] = rot7(v[7]);
343 :
344 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
345 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
346 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
347 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
348 0 : v[0] = addv(v[0], v[5]);
349 0 : v[1] = addv(v[1], v[6]);
350 0 : v[2] = addv(v[2], v[7]);
351 0 : v[3] = addv(v[3], v[4]);
352 0 : v[15] = xorv(v[15], v[0]);
353 0 : v[12] = xorv(v[12], v[1]);
354 0 : v[13] = xorv(v[13], v[2]);
355 0 : v[14] = xorv(v[14], v[3]);
356 0 : v[15] = rot16(v[15]);
357 0 : v[12] = rot16(v[12]);
358 0 : v[13] = rot16(v[13]);
359 0 : v[14] = rot16(v[14]);
360 0 : v[10] = addv(v[10], v[15]);
361 0 : v[11] = addv(v[11], v[12]);
362 0 : v[8] = addv(v[8], v[13]);
363 0 : v[9] = addv(v[9], v[14]);
364 0 : v[5] = xorv(v[5], v[10]);
365 0 : v[6] = xorv(v[6], v[11]);
366 0 : v[7] = xorv(v[7], v[8]);
367 0 : v[4] = xorv(v[4], v[9]);
368 0 : v[5] = rot12(v[5]);
369 0 : v[6] = rot12(v[6]);
370 0 : v[7] = rot12(v[7]);
371 0 : v[4] = rot12(v[4]);
372 0 : v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
373 0 : v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
374 0 : v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
375 0 : v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
376 0 : v[0] = addv(v[0], v[5]);
377 0 : v[1] = addv(v[1], v[6]);
378 0 : v[2] = addv(v[2], v[7]);
379 0 : v[3] = addv(v[3], v[4]);
380 0 : v[15] = xorv(v[15], v[0]);
381 0 : v[12] = xorv(v[12], v[1]);
382 0 : v[13] = xorv(v[13], v[2]);
383 0 : v[14] = xorv(v[14], v[3]);
384 0 : v[15] = rot8(v[15]);
385 0 : v[12] = rot8(v[12]);
386 0 : v[13] = rot8(v[13]);
387 0 : v[14] = rot8(v[14]);
388 0 : v[10] = addv(v[10], v[15]);
389 0 : v[11] = addv(v[11], v[12]);
390 0 : v[8] = addv(v[8], v[13]);
391 0 : v[9] = addv(v[9], v[14]);
392 0 : v[5] = xorv(v[5], v[10]);
393 0 : v[6] = xorv(v[6], v[11]);
394 0 : v[7] = xorv(v[7], v[8]);
395 0 : v[4] = xorv(v[4], v[9]);
396 0 : v[5] = rot7(v[5]);
397 0 : v[6] = rot7(v[6]);
398 0 : v[7] = rot7(v[7]);
399 0 : v[4] = rot7(v[4]);
400 0 : }
401 :
402 0 : INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
403 : // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is
404 : // 22/33. Note that this doesn't split the vector into two lanes, as the
405 : // AVX2 counterparts do.
406 0 : __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
407 0 : __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
408 0 : __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
409 0 : __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
410 :
411 : // Interleave 64-bit lanes.
412 0 : __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
413 0 : __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
414 0 : __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
415 0 : __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
416 :
417 0 : vecs[0] = abcd_0;
418 0 : vecs[1] = abcd_1;
419 0 : vecs[2] = abcd_2;
420 0 : vecs[3] = abcd_3;
421 0 : }
422 :
423 : INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
424 0 : size_t block_offset, __m128i out[16]) {
425 0 : out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
426 0 : out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
427 0 : out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
428 0 : out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
429 0 : out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
430 0 : out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
431 0 : out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
432 0 : out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
433 0 : out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
434 0 : out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
435 0 : out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
436 0 : out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
437 0 : out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
438 0 : out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
439 0 : out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
440 0 : out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
441 0 : for (size_t i = 0; i < 4; ++i) {
442 0 : _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
443 0 : }
444 0 : transpose_vecs(&out[0]);
445 0 : transpose_vecs(&out[4]);
446 0 : transpose_vecs(&out[8]);
447 0 : transpose_vecs(&out[12]);
448 0 : }
449 :
450 : INLINE void load_counters(uint64_t counter, bool increment_counter,
451 0 : __m128i *out_lo, __m128i *out_hi) {
452 0 : const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
453 0 : const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
454 0 : const __m128i add1 = _mm_and_si128(mask, add0);
455 0 : __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1);
456 0 : __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32((int)0x80000000)),
457 0 : _mm_xor_si128( l, _mm_set1_epi32((int)0x80000000)));
458 0 : __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry);
459 0 : *out_lo = l;
460 0 : *out_hi = h;
461 0 : }
462 :
463 : static
464 : void blake3_hash4_sse2(const uint8_t *const *inputs, size_t blocks,
465 : const uint32_t key[8], uint64_t counter,
466 : bool increment_counter, uint8_t flags,
467 0 : uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
468 0 : __m128i h_vecs[8] = {
469 0 : set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
470 0 : set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
471 0 : };
472 0 : __m128i counter_low_vec, counter_high_vec;
473 0 : load_counters(counter, increment_counter, &counter_low_vec,
474 0 : &counter_high_vec);
475 0 : uint8_t block_flags = flags | flags_start;
476 :
477 0 : for (size_t block = 0; block < blocks; block++) {
478 0 : if (block + 1 == blocks) {
479 0 : block_flags |= flags_end;
480 0 : }
481 0 : __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
482 0 : __m128i block_flags_vec = set1(block_flags);
483 0 : __m128i msg_vecs[16];
484 0 : transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
485 :
486 0 : __m128i v[16] = {
487 0 : h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
488 0 : h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
489 0 : set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]),
490 0 : counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
491 0 : };
492 0 : round_fn(v, msg_vecs, 0);
493 0 : round_fn(v, msg_vecs, 1);
494 0 : round_fn(v, msg_vecs, 2);
495 0 : round_fn(v, msg_vecs, 3);
496 0 : round_fn(v, msg_vecs, 4);
497 0 : round_fn(v, msg_vecs, 5);
498 0 : round_fn(v, msg_vecs, 6);
499 0 : h_vecs[0] = xorv(v[0], v[8]);
500 0 : h_vecs[1] = xorv(v[1], v[9]);
501 0 : h_vecs[2] = xorv(v[2], v[10]);
502 0 : h_vecs[3] = xorv(v[3], v[11]);
503 0 : h_vecs[4] = xorv(v[4], v[12]);
504 0 : h_vecs[5] = xorv(v[5], v[13]);
505 0 : h_vecs[6] = xorv(v[6], v[14]);
506 0 : h_vecs[7] = xorv(v[7], v[15]);
507 :
508 0 : block_flags = flags;
509 0 : }
510 :
511 0 : transpose_vecs(&h_vecs[0]);
512 0 : transpose_vecs(&h_vecs[4]);
513 : // The first four vecs now contain the first half of each output, and the
514 : // second four vecs contain the second half of each output.
515 0 : storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
516 0 : storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
517 0 : storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
518 0 : storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
519 0 : storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
520 0 : storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
521 0 : storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
522 0 : storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
523 0 : }
524 :
525 : INLINE void hash_one_sse2(const uint8_t *input, size_t blocks,
526 : const uint32_t key[8], uint64_t counter,
527 : uint8_t flags, uint8_t flags_start,
528 0 : uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
529 0 : uint32_t cv[8];
530 0 : memcpy(cv, key, BLAKE3_KEY_LEN);
531 0 : uint8_t block_flags = flags | flags_start;
532 0 : while (blocks > 0) {
533 0 : if (blocks == 1) {
534 0 : block_flags |= flags_end;
535 0 : }
536 0 : blake3_compress_in_place_sse2(cv, input, BLAKE3_BLOCK_LEN, counter,
537 0 : block_flags);
538 0 : input = &input[BLAKE3_BLOCK_LEN];
539 0 : blocks -= 1;
540 0 : block_flags = flags;
541 0 : }
542 0 : memcpy(out, cv, BLAKE3_OUT_LEN);
543 0 : }
544 :
545 : void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
546 : size_t blocks, const uint32_t key[8],
547 : uint64_t counter, bool increment_counter,
548 : uint8_t flags, uint8_t flags_start,
549 0 : uint8_t flags_end, uint8_t *out) {
550 0 : while (num_inputs >= DEGREE) {
551 0 : blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags,
552 0 : flags_start, flags_end, out);
553 0 : if (increment_counter) {
554 0 : counter += DEGREE;
555 0 : }
556 0 : inputs += DEGREE;
557 0 : num_inputs -= DEGREE;
558 0 : out = &out[DEGREE * BLAKE3_OUT_LEN];
559 0 : }
560 0 : while (num_inputs > 0) {
561 0 : hash_one_sse2(inputs[0], blocks, key, counter, flags, flags_start,
562 0 : flags_end, out);
563 0 : if (increment_counter) {
564 0 : counter += 1;
565 0 : }
566 0 : inputs += 1;
567 0 : num_inputs -= 1;
568 0 : out = &out[BLAKE3_OUT_LEN];
569 0 : }
570 0 : }
|
