1 #include "blake3_impl.h"
2
3 #include <arm_neon.h>
4
5 #ifdef __ARM_BIG_ENDIAN
6 #error "This implementation only supports little-endian ARM."
7 // It might be that all we need for big-endian support here is to get the loads
8 // and stores right, but step zero would be finding a way to test it in CI.
9 #endif
10
loadu_128(const uint8_t src[16])11 INLINE uint32x4_t loadu_128(const uint8_t src[16]) {
12 // vld1q_u32 has alignment requirements. Don't use it.
13 return vreinterpretq_u32_u8(vld1q_u8(src));
14 }
15
storeu_128(uint32x4_t src,uint8_t dest[16])16 INLINE void storeu_128(uint32x4_t src, uint8_t dest[16]) {
17 // vst1q_u32 has alignment requirements. Don't use it.
18 vst1q_u8(dest, vreinterpretq_u8_u32(src));
19 }
20
add_128(uint32x4_t a,uint32x4_t b)21 INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) {
22 return vaddq_u32(a, b);
23 }
24
xor_128(uint32x4_t a,uint32x4_t b)25 INLINE uint32x4_t xor_128(uint32x4_t a, uint32x4_t b) {
26 return veorq_u32(a, b);
27 }
28
set1_128(uint32_t x)29 INLINE uint32x4_t set1_128(uint32_t x) { return vld1q_dup_u32(&x); }
30
set4(uint32_t a,uint32_t b,uint32_t c,uint32_t d)31 INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
32 uint32_t array[4] = {a, b, c, d};
33 return vld1q_u32(array);
34 }
35
rot16_128(uint32x4_t x)36 INLINE uint32x4_t rot16_128(uint32x4_t x) {
37 // The straightfoward implementation would be two shifts and an or, but that's
38 // slower on microarchitectures we've tested. See
39 // https://github.com/BLAKE3-team/BLAKE3/pull/319.
40 // return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16));
41 return vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(x)));
42 }
43
rot12_128(uint32x4_t x)44 INLINE uint32x4_t rot12_128(uint32x4_t x) {
45 // See comment in rot16_128.
46 // return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12));
47 return vsriq_n_u32(vshlq_n_u32(x, 32-12), x, 12);
48 }
49
rot8_128(uint32x4_t x)50 INLINE uint32x4_t rot8_128(uint32x4_t x) {
51 // See comment in rot16_128.
52 // return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8));
53 #if defined(__clang__)
54 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));
55 #elif __GNUC__ * 10000 + __GNUC_MINOR__ * 100 >=40700
56 static const uint8x16_t r8 = {1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12};
57 return vreinterpretq_u32_u8(__builtin_shuffle(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), r8));
58 #else
59 return vsriq_n_u32(vshlq_n_u32(x, 32-8), x, 8);
60 #endif
61 }
62
rot7_128(uint32x4_t x)63 INLINE uint32x4_t rot7_128(uint32x4_t x) {
64 // See comment in rot16_128.
65 // return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7));
66 return vsriq_n_u32(vshlq_n_u32(x, 32-7), x, 7);
67 }
68
69 // TODO: compress_neon
70
71 // TODO: hash2_neon
72
73 /*
74 * ----------------------------------------------------------------------------
75 * hash4_neon
76 * ----------------------------------------------------------------------------
77 */
78
round_fn4(uint32x4_t v[16],uint32x4_t m[16],size_t r)79 INLINE void round_fn4(uint32x4_t v[16], uint32x4_t m[16], size_t r) {
80 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
81 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
82 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
83 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
84 v[0] = add_128(v[0], v[4]);
85 v[1] = add_128(v[1], v[5]);
86 v[2] = add_128(v[2], v[6]);
87 v[3] = add_128(v[3], v[7]);
88 v[12] = xor_128(v[12], v[0]);
89 v[13] = xor_128(v[13], v[1]);
90 v[14] = xor_128(v[14], v[2]);
91 v[15] = xor_128(v[15], v[3]);
92 v[12] = rot16_128(v[12]);
93 v[13] = rot16_128(v[13]);
94 v[14] = rot16_128(v[14]);
95 v[15] = rot16_128(v[15]);
96 v[8] = add_128(v[8], v[12]);
97 v[9] = add_128(v[9], v[13]);
98 v[10] = add_128(v[10], v[14]);
99 v[11] = add_128(v[11], v[15]);
100 v[4] = xor_128(v[4], v[8]);
101 v[5] = xor_128(v[5], v[9]);
102 v[6] = xor_128(v[6], v[10]);
103 v[7] = xor_128(v[7], v[11]);
104 v[4] = rot12_128(v[4]);
105 v[5] = rot12_128(v[5]);
106 v[6] = rot12_128(v[6]);
107 v[7] = rot12_128(v[7]);
108 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
109 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
110 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
111 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
112 v[0] = add_128(v[0], v[4]);
113 v[1] = add_128(v[1], v[5]);
114 v[2] = add_128(v[2], v[6]);
115 v[3] = add_128(v[3], v[7]);
116 v[12] = xor_128(v[12], v[0]);
117 v[13] = xor_128(v[13], v[1]);
118 v[14] = xor_128(v[14], v[2]);
119 v[15] = xor_128(v[15], v[3]);
120 v[12] = rot8_128(v[12]);
121 v[13] = rot8_128(v[13]);
122 v[14] = rot8_128(v[14]);
123 v[15] = rot8_128(v[15]);
124 v[8] = add_128(v[8], v[12]);
125 v[9] = add_128(v[9], v[13]);
126 v[10] = add_128(v[10], v[14]);
127 v[11] = add_128(v[11], v[15]);
128 v[4] = xor_128(v[4], v[8]);
129 v[5] = xor_128(v[5], v[9]);
130 v[6] = xor_128(v[6], v[10]);
131 v[7] = xor_128(v[7], v[11]);
132 v[4] = rot7_128(v[4]);
133 v[5] = rot7_128(v[5]);
134 v[6] = rot7_128(v[6]);
135 v[7] = rot7_128(v[7]);
136
137 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
138 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
139 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
140 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
141 v[0] = add_128(v[0], v[5]);
142 v[1] = add_128(v[1], v[6]);
143 v[2] = add_128(v[2], v[7]);
144 v[3] = add_128(v[3], v[4]);
145 v[15] = xor_128(v[15], v[0]);
146 v[12] = xor_128(v[12], v[1]);
147 v[13] = xor_128(v[13], v[2]);
148 v[14] = xor_128(v[14], v[3]);
149 v[15] = rot16_128(v[15]);
150 v[12] = rot16_128(v[12]);
151 v[13] = rot16_128(v[13]);
152 v[14] = rot16_128(v[14]);
153 v[10] = add_128(v[10], v[15]);
154 v[11] = add_128(v[11], v[12]);
155 v[8] = add_128(v[8], v[13]);
156 v[9] = add_128(v[9], v[14]);
157 v[5] = xor_128(v[5], v[10]);
158 v[6] = xor_128(v[6], v[11]);
159 v[7] = xor_128(v[7], v[8]);
160 v[4] = xor_128(v[4], v[9]);
161 v[5] = rot12_128(v[5]);
162 v[6] = rot12_128(v[6]);
163 v[7] = rot12_128(v[7]);
164 v[4] = rot12_128(v[4]);
165 v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
166 v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
167 v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
168 v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
169 v[0] = add_128(v[0], v[5]);
170 v[1] = add_128(v[1], v[6]);
171 v[2] = add_128(v[2], v[7]);
172 v[3] = add_128(v[3], v[4]);
173 v[15] = xor_128(v[15], v[0]);
174 v[12] = xor_128(v[12], v[1]);
175 v[13] = xor_128(v[13], v[2]);
176 v[14] = xor_128(v[14], v[3]);
177 v[15] = rot8_128(v[15]);
178 v[12] = rot8_128(v[12]);
179 v[13] = rot8_128(v[13]);
180 v[14] = rot8_128(v[14]);
181 v[10] = add_128(v[10], v[15]);
182 v[11] = add_128(v[11], v[12]);
183 v[8] = add_128(v[8], v[13]);
184 v[9] = add_128(v[9], v[14]);
185 v[5] = xor_128(v[5], v[10]);
186 v[6] = xor_128(v[6], v[11]);
187 v[7] = xor_128(v[7], v[8]);
188 v[4] = xor_128(v[4], v[9]);
189 v[5] = rot7_128(v[5]);
190 v[6] = rot7_128(v[6]);
191 v[7] = rot7_128(v[7]);
192 v[4] = rot7_128(v[4]);
193 }
194
transpose_vecs_128(uint32x4_t vecs[4])195 INLINE void transpose_vecs_128(uint32x4_t vecs[4]) {
196 // Individually transpose the four 2x2 sub-matrices in each corner.
197 uint32x4x2_t rows01 = vtrnq_u32(vecs[0], vecs[1]);
198 uint32x4x2_t rows23 = vtrnq_u32(vecs[2], vecs[3]);
199
200 // Swap the top-right and bottom-left 2x2s (which just got transposed).
201 vecs[0] =
202 vcombine_u32(vget_low_u32(rows01.val[0]), vget_low_u32(rows23.val[0]));
203 vecs[1] =
204 vcombine_u32(vget_low_u32(rows01.val[1]), vget_low_u32(rows23.val[1]));
205 vecs[2] =
206 vcombine_u32(vget_high_u32(rows01.val[0]), vget_high_u32(rows23.val[0]));
207 vecs[3] =
208 vcombine_u32(vget_high_u32(rows01.val[1]), vget_high_u32(rows23.val[1]));
209 }
210
transpose_msg_vecs4(const uint8_t * const * inputs,size_t block_offset,uint32x4_t out[16])211 INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
212 size_t block_offset, uint32x4_t out[16]) {
213 out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]);
214 out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]);
215 out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]);
216 out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]);
217 out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]);
218 out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]);
219 out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]);
220 out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]);
221 out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]);
222 out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(uint32x4_t)]);
223 out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(uint32x4_t)]);
224 out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(uint32x4_t)]);
225 out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(uint32x4_t)]);
226 out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(uint32x4_t)]);
227 out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(uint32x4_t)]);
228 out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(uint32x4_t)]);
229 transpose_vecs_128(&out[0]);
230 transpose_vecs_128(&out[4]);
231 transpose_vecs_128(&out[8]);
232 transpose_vecs_128(&out[12]);
233 }
234
load_counters4(uint64_t counter,bool increment_counter,uint32x4_t * out_low,uint32x4_t * out_high)235 INLINE void load_counters4(uint64_t counter, bool increment_counter,
236 uint32x4_t *out_low, uint32x4_t *out_high) {
237 uint64_t mask = (increment_counter ? ~0 : 0);
238 *out_low = set4(
239 counter_low(counter + (mask & 0)), counter_low(counter + (mask & 1)),
240 counter_low(counter + (mask & 2)), counter_low(counter + (mask & 3)));
241 *out_high = set4(
242 counter_high(counter + (mask & 0)), counter_high(counter + (mask & 1)),
243 counter_high(counter + (mask & 2)), counter_high(counter + (mask & 3)));
244 }
245
blake3_hash4_neon(const uint8_t * const * inputs,size_t blocks,const uint32_t key[8],uint64_t counter,bool increment_counter,uint8_t flags,uint8_t flags_start,uint8_t flags_end,uint8_t * out)246 static void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks,
247 const uint32_t key[8], uint64_t counter,
248 bool increment_counter, uint8_t flags,
249 uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
250 uint32x4_t h_vecs[8] = {
251 set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]),
252 set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]),
253 };
254 uint32x4_t counter_low_vec, counter_high_vec;
255 load_counters4(counter, increment_counter, &counter_low_vec,
256 &counter_high_vec);
257 uint8_t block_flags = flags | flags_start;
258
259 for (size_t block = 0; block < blocks; block++) {
260 if (block + 1 == blocks) {
261 block_flags |= flags_end;
262 }
263 uint32x4_t block_len_vec = set1_128(BLAKE3_BLOCK_LEN);
264 uint32x4_t block_flags_vec = set1_128(block_flags);
265 uint32x4_t msg_vecs[16];
266 transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
267
268 uint32x4_t v[16] = {
269 h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
270 h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
271 set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]),
272 counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
273 };
274 round_fn4(v, msg_vecs, 0);
275 round_fn4(v, msg_vecs, 1);
276 round_fn4(v, msg_vecs, 2);
277 round_fn4(v, msg_vecs, 3);
278 round_fn4(v, msg_vecs, 4);
279 round_fn4(v, msg_vecs, 5);
280 round_fn4(v, msg_vecs, 6);
281 h_vecs[0] = xor_128(v[0], v[8]);
282 h_vecs[1] = xor_128(v[1], v[9]);
283 h_vecs[2] = xor_128(v[2], v[10]);
284 h_vecs[3] = xor_128(v[3], v[11]);
285 h_vecs[4] = xor_128(v[4], v[12]);
286 h_vecs[5] = xor_128(v[5], v[13]);
287 h_vecs[6] = xor_128(v[6], v[14]);
288 h_vecs[7] = xor_128(v[7], v[15]);
289
290 block_flags = flags;
291 }
292
293 transpose_vecs_128(&h_vecs[0]);
294 transpose_vecs_128(&h_vecs[4]);
295 // The first four vecs now contain the first half of each output, and the
296 // second four vecs contain the second half of each output.
297 storeu_128(h_vecs[0], &out[0 * sizeof(uint32x4_t)]);
298 storeu_128(h_vecs[4], &out[1 * sizeof(uint32x4_t)]);
299 storeu_128(h_vecs[1], &out[2 * sizeof(uint32x4_t)]);
300 storeu_128(h_vecs[5], &out[3 * sizeof(uint32x4_t)]);
301 storeu_128(h_vecs[2], &out[4 * sizeof(uint32x4_t)]);
302 storeu_128(h_vecs[6], &out[5 * sizeof(uint32x4_t)]);
303 storeu_128(h_vecs[3], &out[6 * sizeof(uint32x4_t)]);
304 storeu_128(h_vecs[7], &out[7 * sizeof(uint32x4_t)]);
305 }
306
307 /*
308 * ----------------------------------------------------------------------------
309 * hash_many_neon
310 * ----------------------------------------------------------------------------
311 */
312
313 void blake3_compress_in_place_portable(uint32_t cv[8],
314 const uint8_t block[BLAKE3_BLOCK_LEN],
315 uint8_t block_len, uint64_t counter,
316 uint8_t flags);
317
hash_one_neon(const uint8_t * input,size_t blocks,const uint32_t key[8],uint64_t counter,uint8_t flags,uint8_t flags_start,uint8_t flags_end,uint8_t out[BLAKE3_OUT_LEN])318 INLINE void hash_one_neon(const uint8_t *input, size_t blocks,
319 const uint32_t key[8], uint64_t counter,
320 uint8_t flags, uint8_t flags_start, uint8_t flags_end,
321 uint8_t out[BLAKE3_OUT_LEN]) {
322 uint32_t cv[8];
323 memcpy(cv, key, BLAKE3_KEY_LEN);
324 uint8_t block_flags = flags | flags_start;
325 while (blocks > 0) {
326 if (blocks == 1) {
327 block_flags |= flags_end;
328 }
329 // TODO: Implement compress_neon. However note that according to
330 // https://github.com/BLAKE2/BLAKE2/commit/7965d3e6e1b4193438b8d3a656787587d2579227,
331 // compress_neon might not be any faster than compress_portable.
332 blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter,
333 block_flags);
334 input = &input[BLAKE3_BLOCK_LEN];
335 blocks -= 1;
336 block_flags = flags;
337 }
338 memcpy(out, cv, BLAKE3_OUT_LEN);
339 }
340
blake3_hash_many_neon(const uint8_t * const * inputs,size_t num_inputs,size_t blocks,const uint32_t key[8],uint64_t counter,bool increment_counter,uint8_t flags,uint8_t flags_start,uint8_t flags_end,uint8_t * out)341 void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
342 size_t blocks, const uint32_t key[8],
343 uint64_t counter, bool increment_counter,
344 uint8_t flags, uint8_t flags_start,
345 uint8_t flags_end, uint8_t *out) {
346 while (num_inputs >= 4) {
347 blake3_hash4_neon(inputs, blocks, key, counter, increment_counter, flags,
348 flags_start, flags_end, out);
349 if (increment_counter) {
350 counter += 4;
351 }
352 inputs += 4;
353 num_inputs -= 4;
354 out = &out[4 * BLAKE3_OUT_LEN];
355 }
356 while (num_inputs > 0) {
357 hash_one_neon(inputs[0], blocks, key, counter, flags, flags_start,
358 flags_end, out);
359 if (increment_counter) {
360 counter += 1;
361 }
362 inputs += 1;
363 num_inputs -= 1;
364 out = &out[BLAKE3_OUT_LEN];
365 }
366 }
367