1#include "blake3_impl.h"
2
3#if BLAKE3_USE_NEON
4
5#include <arm_neon.h>
6
7#ifdef __ARM_BIG_ENDIAN
8#error "This implementation only supports little-endian ARM."
9// It might be that all we need for big-endian support here is to get the loads
10// and stores right, but step zero would be finding a way to test it in CI.
11#endif
12
13INLINE uint32x4_t loadu_128(const uint8_t src[16]) {
14 // vld1q_u32 has alignment requirements. Don't use it.
15 return vreinterpretq_u32_u8(vld1q_u8(src));
16}
17
18INLINE void storeu_128(uint32x4_t src, uint8_t dest[16]) {
19 // vst1q_u32 has alignment requirements. Don't use it.
20 vst1q_u8(dest, vreinterpretq_u8_u32(src));
21}
22
23INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) {
24 return vaddq_u32(a, b);
25}
26
27INLINE uint32x4_t xor_128(uint32x4_t a, uint32x4_t b) {
28 return veorq_u32(a, b);
29}
30
31INLINE uint32x4_t set1_128(uint32_t x) { return vld1q_dup_u32(&x); }
32
33INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
34 uint32_t array[4] = {a, b, c, d};
35 return vld1q_u32(array);
36}
37
38INLINE uint32x4_t rot16_128(uint32x4_t x) {
39 // The straightforward implementation would be two shifts and an or, but that's
40 // slower on microarchitectures we've tested. See
41 // https://github.com/BLAKE3-team/BLAKE3/pull/319.
42 // return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16));
43 return vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(x)));
44}
45
46INLINE uint32x4_t rot12_128(uint32x4_t x) {
47 // See comment in rot16_128.
48 // return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12));
49 return vsriq_n_u32(vshlq_n_u32(x, 32-12), x, 12);
50}
51
52INLINE uint32x4_t rot8_128(uint32x4_t x) {
53 // See comment in rot16_128.
54 // return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8));
55#if defined(__clang__)
56 return vreinterpretq_u32_u8(__builtin_shufflevector(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), 1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12));
57#elif defined(__GNUC__)
58 static const uint8x16_t r8 = {1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12};
59 return vreinterpretq_u32_u8(__builtin_shuffle(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), r8));
60#else
61 return vsriq_n_u32(vshlq_n_u32(x, 32-8), x, 8);
62#endif
63}
64
65INLINE uint32x4_t rot7_128(uint32x4_t x) {
66 // See comment in rot16_128.
67 // return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7));
68 return vsriq_n_u32(vshlq_n_u32(x, 32-7), x, 7);
69}
70
71// TODO: compress_neon
72
73// TODO: hash2_neon
74
75/*
76 * ----------------------------------------------------------------------------
77 * hash4_neon
78 * ----------------------------------------------------------------------------
79 */
80
81INLINE void round_fn4(uint32x4_t v[16], uint32x4_t m[16], size_t r) {
82 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
83 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
84 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
85 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
86 v[0] = add_128(v[0], v[4]);
87 v[1] = add_128(v[1], v[5]);
88 v[2] = add_128(v[2], v[6]);
89 v[3] = add_128(v[3], v[7]);
90 v[12] = xor_128(v[12], v[0]);
91 v[13] = xor_128(v[13], v[1]);
92 v[14] = xor_128(v[14], v[2]);
93 v[15] = xor_128(v[15], v[3]);
94 v[12] = rot16_128(v[12]);
95 v[13] = rot16_128(v[13]);
96 v[14] = rot16_128(v[14]);
97 v[15] = rot16_128(v[15]);
98 v[8] = add_128(v[8], v[12]);
99 v[9] = add_128(v[9], v[13]);
100 v[10] = add_128(v[10], v[14]);
101 v[11] = add_128(v[11], v[15]);
102 v[4] = xor_128(v[4], v[8]);
103 v[5] = xor_128(v[5], v[9]);
104 v[6] = xor_128(v[6], v[10]);
105 v[7] = xor_128(v[7], v[11]);
106 v[4] = rot12_128(v[4]);
107 v[5] = rot12_128(v[5]);
108 v[6] = rot12_128(v[6]);
109 v[7] = rot12_128(v[7]);
110 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
111 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
112 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
113 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
114 v[0] = add_128(v[0], v[4]);
115 v[1] = add_128(v[1], v[5]);
116 v[2] = add_128(v[2], v[6]);
117 v[3] = add_128(v[3], v[7]);
118 v[12] = xor_128(v[12], v[0]);
119 v[13] = xor_128(v[13], v[1]);
120 v[14] = xor_128(v[14], v[2]);
121 v[15] = xor_128(v[15], v[3]);
122 v[12] = rot8_128(v[12]);
123 v[13] = rot8_128(v[13]);
124 v[14] = rot8_128(v[14]);
125 v[15] = rot8_128(v[15]);
126 v[8] = add_128(v[8], v[12]);
127 v[9] = add_128(v[9], v[13]);
128 v[10] = add_128(v[10], v[14]);
129 v[11] = add_128(v[11], v[15]);
130 v[4] = xor_128(v[4], v[8]);
131 v[5] = xor_128(v[5], v[9]);
132 v[6] = xor_128(v[6], v[10]);
133 v[7] = xor_128(v[7], v[11]);
134 v[4] = rot7_128(v[4]);
135 v[5] = rot7_128(v[5]);
136 v[6] = rot7_128(v[6]);
137 v[7] = rot7_128(v[7]);
138
139 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
140 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
141 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
142 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
143 v[0] = add_128(v[0], v[5]);
144 v[1] = add_128(v[1], v[6]);
145 v[2] = add_128(v[2], v[7]);
146 v[3] = add_128(v[3], v[4]);
147 v[15] = xor_128(v[15], v[0]);
148 v[12] = xor_128(v[12], v[1]);
149 v[13] = xor_128(v[13], v[2]);
150 v[14] = xor_128(v[14], v[3]);
151 v[15] = rot16_128(v[15]);
152 v[12] = rot16_128(v[12]);
153 v[13] = rot16_128(v[13]);
154 v[14] = rot16_128(v[14]);
155 v[10] = add_128(v[10], v[15]);
156 v[11] = add_128(v[11], v[12]);
157 v[8] = add_128(v[8], v[13]);
158 v[9] = add_128(v[9], v[14]);
159 v[5] = xor_128(v[5], v[10]);
160 v[6] = xor_128(v[6], v[11]);
161 v[7] = xor_128(v[7], v[8]);
162 v[4] = xor_128(v[4], v[9]);
163 v[5] = rot12_128(v[5]);
164 v[6] = rot12_128(v[6]);
165 v[7] = rot12_128(v[7]);
166 v[4] = rot12_128(v[4]);
167 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
168 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
169 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
170 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
171 v[0] = add_128(v[0], v[5]);
172 v[1] = add_128(v[1], v[6]);
173 v[2] = add_128(v[2], v[7]);
174 v[3] = add_128(v[3], v[4]);
175 v[15] = xor_128(v[15], v[0]);
176 v[12] = xor_128(v[12], v[1]);
177 v[13] = xor_128(v[13], v[2]);
178 v[14] = xor_128(v[14], v[3]);
179 v[15] = rot8_128(v[15]);
180 v[12] = rot8_128(v[12]);
181 v[13] = rot8_128(v[13]);
182 v[14] = rot8_128(v[14]);
183 v[10] = add_128(v[10], v[15]);
184 v[11] = add_128(v[11], v[12]);
185 v[8] = add_128(v[8], v[13]);
186 v[9] = add_128(v[9], v[14]);
187 v[5] = xor_128(v[5], v[10]);
188 v[6] = xor_128(v[6], v[11]);
189 v[7] = xor_128(v[7], v[8]);
190 v[4] = xor_128(v[4], v[9]);
191 v[5] = rot7_128(v[5]);
192 v[6] = rot7_128(v[6]);
193 v[7] = rot7_128(v[7]);
194 v[4] = rot7_128(v[4]);
195}
196
197INLINE void transpose_vecs_128(uint32x4_t vecs[4]) {
198 // Individually transpose the four 2x2 sub-matrices in each corner.
199 uint32x4x2_t rows01 = vtrnq_u32(vecs[0], vecs[1]);
200 uint32x4x2_t rows23 = vtrnq_u32(vecs[2], vecs[3]);
201
202 // Swap the top-right and bottom-left 2x2s (which just got transposed).
203 vecs[0] =
204 vcombine_u32(vget_low_u32(rows01.val[0]), vget_low_u32(rows23.val[0]));
205 vecs[1] =
206 vcombine_u32(vget_low_u32(rows01.val[1]), vget_low_u32(rows23.val[1]));
207 vecs[2] =
208 vcombine_u32(vget_high_u32(rows01.val[0]), vget_high_u32(rows23.val[0]));
209 vecs[3] =
210 vcombine_u32(vget_high_u32(rows01.val[1]), vget_high_u32(rows23.val[1]));
211}
212
213INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
214 size_t block_offset, uint32x4_t out[16]) {
215 out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]);
216 out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]);
217 out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]);
218 out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]);
219 out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]);
220 out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]);
221 out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]);
222 out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]);
223 out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]);
224 out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(uint32x4_t)]);
225 out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(uint32x4_t)]);
226 out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(uint32x4_t)]);
227 out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(uint32x4_t)]);
228 out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(uint32x4_t)]);
229 out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(uint32x4_t)]);
230 out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(uint32x4_t)]);
231 transpose_vecs_128(&out[0]);
232 transpose_vecs_128(&out[4]);
233 transpose_vecs_128(&out[8]);
234 transpose_vecs_128(&out[12]);
235}
236
237INLINE void load_counters4(uint64_t counter, bool increment_counter,
238 uint32x4_t *out_low, uint32x4_t *out_high) {
239 uint64_t mask = (increment_counter ? ~0 : 0);
240 *out_low = set4(
241 counter_low(counter + (mask & 0)), counter_low(counter + (mask & 1)),
242 counter_low(counter + (mask & 2)), counter_low(counter + (mask & 3)));
243 *out_high = set4(
244 counter_high(counter + (mask & 0)), counter_high(counter + (mask & 1)),
245 counter_high(counter + (mask & 2)), counter_high(counter + (mask & 3)));
246}
247
248static void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks,
249 const uint32_t key[8], uint64_t counter,
250 bool increment_counter, uint8_t flags,
251 uint8_t flags_start, uint8_t flags_end,
252 uint8_t *out) {
253 uint32x4_t h_vecs[8] = {
254 set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]),
255 set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]),
256 };
257 uint32x4_t counter_low_vec, counter_high_vec;
258 load_counters4(counter, increment_counter, &counter_low_vec,
259 &counter_high_vec);
260 uint8_t block_flags = flags | flags_start;
261
262 for (size_t block = 0; block < blocks; block++) {
263 if (block + 1 == blocks) {
264 block_flags |= flags_end;
265 }
266 uint32x4_t block_len_vec = set1_128(BLAKE3_BLOCK_LEN);
267 uint32x4_t block_flags_vec = set1_128(block_flags);
268 uint32x4_t msg_vecs[16];
269 transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
270
271 uint32x4_t v[16] = {
272 h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
273 h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
274 set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]),
275 counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
276 };
277 round_fn4(v, msg_vecs, 0);
278 round_fn4(v, msg_vecs, 1);
279 round_fn4(v, msg_vecs, 2);
280 round_fn4(v, msg_vecs, 3);
281 round_fn4(v, msg_vecs, 4);
282 round_fn4(v, msg_vecs, 5);
283 round_fn4(v, msg_vecs, 6);
284 h_vecs[0] = xor_128(v[0], v[8]);
285 h_vecs[1] = xor_128(v[1], v[9]);
286 h_vecs[2] = xor_128(v[2], v[10]);
287 h_vecs[3] = xor_128(v[3], v[11]);
288 h_vecs[4] = xor_128(v[4], v[12]);
289 h_vecs[5] = xor_128(v[5], v[13]);
290 h_vecs[6] = xor_128(v[6], v[14]);
291 h_vecs[7] = xor_128(v[7], v[15]);
292
293 block_flags = flags;
294 }
295
296 transpose_vecs_128(&h_vecs[0]);
297 transpose_vecs_128(&h_vecs[4]);
298 // The first four vecs now contain the first half of each output, and the
299 // second four vecs contain the second half of each output.
300 storeu_128(h_vecs[0], &out[0 * sizeof(uint32x4_t)]);
301 storeu_128(h_vecs[4], &out[1 * sizeof(uint32x4_t)]);
302 storeu_128(h_vecs[1], &out[2 * sizeof(uint32x4_t)]);
303 storeu_128(h_vecs[5], &out[3 * sizeof(uint32x4_t)]);
304 storeu_128(h_vecs[2], &out[4 * sizeof(uint32x4_t)]);
305 storeu_128(h_vecs[6], &out[5 * sizeof(uint32x4_t)]);
306 storeu_128(h_vecs[3], &out[6 * sizeof(uint32x4_t)]);
307 storeu_128(h_vecs[7], &out[7 * sizeof(uint32x4_t)]);
308}
309
310/*
311 * ----------------------------------------------------------------------------
312 * hash_many_neon
313 * ----------------------------------------------------------------------------
314 */
315
316void blake3_compress_in_place_portable(uint32_t cv[8],
317 const uint8_t block[BLAKE3_BLOCK_LEN],
318 uint8_t block_len, uint64_t counter,
319 uint8_t flags);
320
321INLINE void hash_one_neon(const uint8_t *input, size_t blocks,
322 const uint32_t key[8], uint64_t counter,
323 uint8_t flags, uint8_t flags_start, uint8_t flags_end,
324 uint8_t out[BLAKE3_OUT_LEN]) {
325 uint32_t cv[8];
326 memcpy(cv, key, BLAKE3_KEY_LEN);
327 uint8_t block_flags = flags | flags_start;
328 while (blocks > 0) {
329 if (blocks == 1) {
330 block_flags |= flags_end;
331 }
332 // TODO: Implement compress_neon. However note that according to
333 // https://github.com/BLAKE2/BLAKE2/commit/7965d3e6e1b4193438b8d3a656787587d2579227,
334 // compress_neon might not be any faster than compress_portable.
335 blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter,
336 block_flags);
337 input = &input[BLAKE3_BLOCK_LEN];
338 blocks -= 1;
339 block_flags = flags;
340 }
341 memcpy(out, cv, BLAKE3_OUT_LEN);
342}
343
344void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
345 size_t blocks, const uint32_t key[8],
346 uint64_t counter, bool increment_counter,
347 uint8_t flags, uint8_t flags_start,
348 uint8_t flags_end, uint8_t *out) {
349 while (num_inputs >= 4) {
350 blake3_hash4_neon(inputs, blocks, key, counter, increment_counter, flags,
351 flags_start, flags_end, out);
352 if (increment_counter) {
353 counter += 4;
354 }
355 inputs += 4;
356 num_inputs -= 4;
357 out = &out[4 * BLAKE3_OUT_LEN];
358 }
359 while (num_inputs > 0) {
360 hash_one_neon(inputs[0], blocks, key, counter, flags, flags_start,
361 flags_end, out);
362 if (increment_counter) {
363 counter += 1;
364 }
365 inputs += 1;
366 num_inputs -= 1;
367 out = &out[BLAKE3_OUT_LEN];
368 }
369}
370
371#endif // BLAKE3_USE_NEON
372