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