llvm/lib/Support/BLAKE3/blake3_neon.c

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