chibipub

experimental activitypub node in C
git clone git://jb55.com/chibipub
Log | Files | Refs | README | LICENSE

blake3_sse41.c (20772B)


      1 #include "blake3_impl.h"
      2 
      3 #include <immintrin.h>
      4 
      5 #define DEGREE 4
      6 
      7 #define _mm_shuffle_ps2(a, b, c)                                               \
      8   (_mm_castps_si128(                                                           \
      9       _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
     10 
     11 INLINE __m128i loadu(const uint8_t src[16]) {
     12   return _mm_loadu_si128((const __m128i *)src);
     13 }
     14 
     15 INLINE void storeu(__m128i src, uint8_t dest[16]) {
     16   _mm_storeu_si128((__m128i *)dest, src);
     17 }
     18 
     19 INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
     20 
     21 // Note that clang-format doesn't like the name "xor" for some reason.
     22 INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
     23 
     24 INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
     25 
     26 INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
     27   return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
     28 }
     29 
     30 INLINE __m128i rot16(__m128i x) {
     31   return _mm_shuffle_epi8(
     32       x, _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2));
     33 }
     34 
     35 INLINE __m128i rot12(__m128i x) {
     36   return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
     37 }
     38 
     39 INLINE __m128i rot8(__m128i x) {
     40   return _mm_shuffle_epi8(
     41       x, _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1));
     42 }
     43 
     44 INLINE __m128i rot7(__m128i x) {
     45   return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
     46 }
     47 
     48 INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
     49                __m128i m) {
     50   *row0 = addv(addv(*row0, m), *row1);
     51   *row3 = xorv(*row3, *row0);
     52   *row3 = rot16(*row3);
     53   *row2 = addv(*row2, *row3);
     54   *row1 = xorv(*row1, *row2);
     55   *row1 = rot12(*row1);
     56 }
     57 
     58 INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
     59                __m128i m) {
     60   *row0 = addv(addv(*row0, m), *row1);
     61   *row3 = xorv(*row3, *row0);
     62   *row3 = rot8(*row3);
     63   *row2 = addv(*row2, *row3);
     64   *row1 = xorv(*row1, *row2);
     65   *row1 = rot7(*row1);
     66 }
     67 
     68 // Note the optimization here of leaving row1 as the unrotated row, rather than
     69 // row0. All the message loads below are adjusted to compensate for this. See
     70 // discussion at https://github.com/sneves/blake2-avx2/pull/4
     71 INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
     72   *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
     73   *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
     74   *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
     75 }
     76 
     77 INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
     78   *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
     79   *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
     80   *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
     81 }
     82 
     83 INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
     84                          const uint8_t block[BLAKE3_BLOCK_LEN],
     85                          uint8_t block_len, uint64_t counter, uint8_t flags) {
     86   rows[0] = loadu((uint8_t *)&cv[0]);
     87   rows[1] = loadu((uint8_t *)&cv[4]);
     88   rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
     89   rows[3] = set4(counter_low(counter), counter_high(counter),
     90                  (uint32_t)block_len, (uint32_t)flags);
     91 
     92   __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
     93   __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
     94   __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
     95   __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
     96 
     97   __m128i t0, t1, t2, t3, tt;
     98 
     99   // Round 1. The first round permutes the message words from the original
    100   // input order, into the groups that get mixed in parallel.
    101   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); //  6  4  2  0
    102   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    103   t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); //  7  5  3  1
    104   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    105   diagonalize(&rows[0], &rows[2], &rows[3]);
    106   t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10  8
    107   t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3));   // 12 10  8 14
    108   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    109   t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11  9
    110   t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3));   // 13 11  9 15
    111   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    112   undiagonalize(&rows[0], &rows[2], &rows[3]);
    113   m0 = t0;
    114   m1 = t1;
    115   m2 = t2;
    116   m3 = t3;
    117 
    118   // Round 2. This round and all following rounds apply a fixed permutation
    119   // to the message words from the round before.
    120   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    121   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    122   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    123   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    124   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    125   t1 = _mm_blend_epi16(tt, t1, 0xCC);
    126   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    127   diagonalize(&rows[0], &rows[2], &rows[3]);
    128   t2 = _mm_unpacklo_epi64(m3, m1);
    129   tt = _mm_blend_epi16(t2, m2, 0xC0);
    130   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    131   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    132   t3 = _mm_unpackhi_epi32(m1, m3);
    133   tt = _mm_unpacklo_epi32(m2, t3);
    134   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    135   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    136   undiagonalize(&rows[0], &rows[2], &rows[3]);
    137   m0 = t0;
    138   m1 = t1;
    139   m2 = t2;
    140   m3 = t3;
    141 
    142   // Round 3
    143   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    144   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    145   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    146   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    147   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    148   t1 = _mm_blend_epi16(tt, t1, 0xCC);
    149   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    150   diagonalize(&rows[0], &rows[2], &rows[3]);
    151   t2 = _mm_unpacklo_epi64(m3, m1);
    152   tt = _mm_blend_epi16(t2, m2, 0xC0);
    153   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    154   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    155   t3 = _mm_unpackhi_epi32(m1, m3);
    156   tt = _mm_unpacklo_epi32(m2, t3);
    157   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    158   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    159   undiagonalize(&rows[0], &rows[2], &rows[3]);
    160   m0 = t0;
    161   m1 = t1;
    162   m2 = t2;
    163   m3 = t3;
    164 
    165   // Round 4
    166   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    167   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    168   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    169   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    170   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    171   t1 = _mm_blend_epi16(tt, t1, 0xCC);
    172   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    173   diagonalize(&rows[0], &rows[2], &rows[3]);
    174   t2 = _mm_unpacklo_epi64(m3, m1);
    175   tt = _mm_blend_epi16(t2, m2, 0xC0);
    176   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    177   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    178   t3 = _mm_unpackhi_epi32(m1, m3);
    179   tt = _mm_unpacklo_epi32(m2, t3);
    180   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    181   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    182   undiagonalize(&rows[0], &rows[2], &rows[3]);
    183   m0 = t0;
    184   m1 = t1;
    185   m2 = t2;
    186   m3 = t3;
    187 
    188   // Round 5
    189   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    190   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    191   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    192   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    193   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    194   t1 = _mm_blend_epi16(tt, t1, 0xCC);
    195   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    196   diagonalize(&rows[0], &rows[2], &rows[3]);
    197   t2 = _mm_unpacklo_epi64(m3, m1);
    198   tt = _mm_blend_epi16(t2, m2, 0xC0);
    199   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    200   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    201   t3 = _mm_unpackhi_epi32(m1, m3);
    202   tt = _mm_unpacklo_epi32(m2, t3);
    203   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    204   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    205   undiagonalize(&rows[0], &rows[2], &rows[3]);
    206   m0 = t0;
    207   m1 = t1;
    208   m2 = t2;
    209   m3 = t3;
    210 
    211   // Round 6
    212   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    213   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    214   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    215   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    216   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    217   t1 = _mm_blend_epi16(tt, t1, 0xCC);
    218   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    219   diagonalize(&rows[0], &rows[2], &rows[3]);
    220   t2 = _mm_unpacklo_epi64(m3, m1);
    221   tt = _mm_blend_epi16(t2, m2, 0xC0);
    222   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    223   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    224   t3 = _mm_unpackhi_epi32(m1, m3);
    225   tt = _mm_unpacklo_epi32(m2, t3);
    226   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    227   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    228   undiagonalize(&rows[0], &rows[2], &rows[3]);
    229   m0 = t0;
    230   m1 = t1;
    231   m2 = t2;
    232   m3 = t3;
    233 
    234   // Round 7
    235   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    236   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    237   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    238   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    239   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    240   t1 = _mm_blend_epi16(tt, t1, 0xCC);
    241   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    242   diagonalize(&rows[0], &rows[2], &rows[3]);
    243   t2 = _mm_unpacklo_epi64(m3, m1);
    244   tt = _mm_blend_epi16(t2, m2, 0xC0);
    245   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    246   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    247   t3 = _mm_unpackhi_epi32(m1, m3);
    248   tt = _mm_unpacklo_epi32(m2, t3);
    249   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    250   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    251   undiagonalize(&rows[0], &rows[2], &rows[3]);
    252 }
    253 
    254 void blake3_compress_in_place_sse41(uint32_t cv[8],
    255                                     const uint8_t block[BLAKE3_BLOCK_LEN],
    256                                     uint8_t block_len, uint64_t counter,
    257                                     uint8_t flags) {
    258   __m128i rows[4];
    259   compress_pre(rows, cv, block, block_len, counter, flags);
    260   storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
    261   storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
    262 }
    263 
    264 void blake3_compress_xof_sse41(const uint32_t cv[8],
    265                                const uint8_t block[BLAKE3_BLOCK_LEN],
    266                                uint8_t block_len, uint64_t counter,
    267                                uint8_t flags, uint8_t out[64]) {
    268   __m128i rows[4];
    269   compress_pre(rows, cv, block, block_len, counter, flags);
    270   storeu(xorv(rows[0], rows[2]), &out[0]);
    271   storeu(xorv(rows[1], rows[3]), &out[16]);
    272   storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
    273   storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
    274 }
    275 
    276 INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
    277   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
    278   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
    279   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
    280   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
    281   v[0] = addv(v[0], v[4]);
    282   v[1] = addv(v[1], v[5]);
    283   v[2] = addv(v[2], v[6]);
    284   v[3] = addv(v[3], v[7]);
    285   v[12] = xorv(v[12], v[0]);
    286   v[13] = xorv(v[13], v[1]);
    287   v[14] = xorv(v[14], v[2]);
    288   v[15] = xorv(v[15], v[3]);
    289   v[12] = rot16(v[12]);
    290   v[13] = rot16(v[13]);
    291   v[14] = rot16(v[14]);
    292   v[15] = rot16(v[15]);
    293   v[8] = addv(v[8], v[12]);
    294   v[9] = addv(v[9], v[13]);
    295   v[10] = addv(v[10], v[14]);
    296   v[11] = addv(v[11], v[15]);
    297   v[4] = xorv(v[4], v[8]);
    298   v[5] = xorv(v[5], v[9]);
    299   v[6] = xorv(v[6], v[10]);
    300   v[7] = xorv(v[7], v[11]);
    301   v[4] = rot12(v[4]);
    302   v[5] = rot12(v[5]);
    303   v[6] = rot12(v[6]);
    304   v[7] = rot12(v[7]);
    305   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
    306   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
    307   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
    308   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
    309   v[0] = addv(v[0], v[4]);
    310   v[1] = addv(v[1], v[5]);
    311   v[2] = addv(v[2], v[6]);
    312   v[3] = addv(v[3], v[7]);
    313   v[12] = xorv(v[12], v[0]);
    314   v[13] = xorv(v[13], v[1]);
    315   v[14] = xorv(v[14], v[2]);
    316   v[15] = xorv(v[15], v[3]);
    317   v[12] = rot8(v[12]);
    318   v[13] = rot8(v[13]);
    319   v[14] = rot8(v[14]);
    320   v[15] = rot8(v[15]);
    321   v[8] = addv(v[8], v[12]);
    322   v[9] = addv(v[9], v[13]);
    323   v[10] = addv(v[10], v[14]);
    324   v[11] = addv(v[11], v[15]);
    325   v[4] = xorv(v[4], v[8]);
    326   v[5] = xorv(v[5], v[9]);
    327   v[6] = xorv(v[6], v[10]);
    328   v[7] = xorv(v[7], v[11]);
    329   v[4] = rot7(v[4]);
    330   v[5] = rot7(v[5]);
    331   v[6] = rot7(v[6]);
    332   v[7] = rot7(v[7]);
    333 
    334   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
    335   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
    336   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
    337   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
    338   v[0] = addv(v[0], v[5]);
    339   v[1] = addv(v[1], v[6]);
    340   v[2] = addv(v[2], v[7]);
    341   v[3] = addv(v[3], v[4]);
    342   v[15] = xorv(v[15], v[0]);
    343   v[12] = xorv(v[12], v[1]);
    344   v[13] = xorv(v[13], v[2]);
    345   v[14] = xorv(v[14], v[3]);
    346   v[15] = rot16(v[15]);
    347   v[12] = rot16(v[12]);
    348   v[13] = rot16(v[13]);
    349   v[14] = rot16(v[14]);
    350   v[10] = addv(v[10], v[15]);
    351   v[11] = addv(v[11], v[12]);
    352   v[8] = addv(v[8], v[13]);
    353   v[9] = addv(v[9], v[14]);
    354   v[5] = xorv(v[5], v[10]);
    355   v[6] = xorv(v[6], v[11]);
    356   v[7] = xorv(v[7], v[8]);
    357   v[4] = xorv(v[4], v[9]);
    358   v[5] = rot12(v[5]);
    359   v[6] = rot12(v[6]);
    360   v[7] = rot12(v[7]);
    361   v[4] = rot12(v[4]);
    362   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
    363   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
    364   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
    365   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
    366   v[0] = addv(v[0], v[5]);
    367   v[1] = addv(v[1], v[6]);
    368   v[2] = addv(v[2], v[7]);
    369   v[3] = addv(v[3], v[4]);
    370   v[15] = xorv(v[15], v[0]);
    371   v[12] = xorv(v[12], v[1]);
    372   v[13] = xorv(v[13], v[2]);
    373   v[14] = xorv(v[14], v[3]);
    374   v[15] = rot8(v[15]);
    375   v[12] = rot8(v[12]);
    376   v[13] = rot8(v[13]);
    377   v[14] = rot8(v[14]);
    378   v[10] = addv(v[10], v[15]);
    379   v[11] = addv(v[11], v[12]);
    380   v[8] = addv(v[8], v[13]);
    381   v[9] = addv(v[9], v[14]);
    382   v[5] = xorv(v[5], v[10]);
    383   v[6] = xorv(v[6], v[11]);
    384   v[7] = xorv(v[7], v[8]);
    385   v[4] = xorv(v[4], v[9]);
    386   v[5] = rot7(v[5]);
    387   v[6] = rot7(v[6]);
    388   v[7] = rot7(v[7]);
    389   v[4] = rot7(v[4]);
    390 }
    391 
    392 INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
    393   // Interleave 32-bit lates. The low unpack is lanes 00/11 and the high is
    394   // 22/33. Note that this doesn't split the vector into two lanes, as the
    395   // AVX2 counterparts do.
    396   __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
    397   __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
    398   __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
    399   __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
    400 
    401   // Interleave 64-bit lanes.
    402   __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
    403   __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
    404   __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
    405   __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
    406 
    407   vecs[0] = abcd_0;
    408   vecs[1] = abcd_1;
    409   vecs[2] = abcd_2;
    410   vecs[3] = abcd_3;
    411 }
    412 
    413 INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
    414                                size_t block_offset, __m128i out[16]) {
    415   out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
    416   out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
    417   out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
    418   out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
    419   out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
    420   out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
    421   out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
    422   out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
    423   out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
    424   out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
    425   out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
    426   out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
    427   out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
    428   out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
    429   out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
    430   out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
    431   for (size_t i = 0; i < 4; ++i) {
    432     _mm_prefetch(&inputs[i][block_offset + 256], _MM_HINT_T0);
    433   }
    434   transpose_vecs(&out[0]);
    435   transpose_vecs(&out[4]);
    436   transpose_vecs(&out[8]);
    437   transpose_vecs(&out[12]);
    438 }
    439 
    440 INLINE void load_counters(uint64_t counter, bool increment_counter,
    441                           __m128i *out_lo, __m128i *out_hi) {
    442   const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
    443   const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
    444   const __m128i add1 = _mm_and_si128(mask, add0);
    445   __m128i l = _mm_add_epi32(_mm_set1_epi32(counter), add1);
    446   __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), 
    447                                   _mm_xor_si128(   l, _mm_set1_epi32(0x80000000)));
    448   __m128i h = _mm_sub_epi32(_mm_set1_epi32(counter >> 32), carry);
    449   *out_lo = l;
    450   *out_hi = h;
    451 }
    452 
    453 void blake3_hash4_sse41(const uint8_t *const *inputs, size_t blocks,
    454                         const uint32_t key[8], uint64_t counter,
    455                         bool increment_counter, uint8_t flags,
    456                         uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
    457   __m128i h_vecs[8] = {
    458       set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
    459       set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
    460   };
    461   __m128i counter_low_vec, counter_high_vec;
    462   load_counters(counter, increment_counter, &counter_low_vec,
    463                 &counter_high_vec);
    464   uint8_t block_flags = flags | flags_start;
    465 
    466   for (size_t block = 0; block < blocks; block++) {
    467     if (block + 1 == blocks) {
    468       block_flags |= flags_end;
    469     }
    470     __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
    471     __m128i block_flags_vec = set1(block_flags);
    472     __m128i msg_vecs[16];
    473     transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
    474 
    475     __m128i v[16] = {
    476         h_vecs[0],       h_vecs[1],        h_vecs[2],     h_vecs[3],
    477         h_vecs[4],       h_vecs[5],        h_vecs[6],     h_vecs[7],
    478         set1(IV[0]),     set1(IV[1]),      set1(IV[2]),   set1(IV[3]),
    479         counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
    480     };
    481     round_fn(v, msg_vecs, 0);
    482     round_fn(v, msg_vecs, 1);
    483     round_fn(v, msg_vecs, 2);
    484     round_fn(v, msg_vecs, 3);
    485     round_fn(v, msg_vecs, 4);
    486     round_fn(v, msg_vecs, 5);
    487     round_fn(v, msg_vecs, 6);
    488     h_vecs[0] = xorv(v[0], v[8]);
    489     h_vecs[1] = xorv(v[1], v[9]);
    490     h_vecs[2] = xorv(v[2], v[10]);
    491     h_vecs[3] = xorv(v[3], v[11]);
    492     h_vecs[4] = xorv(v[4], v[12]);
    493     h_vecs[5] = xorv(v[5], v[13]);
    494     h_vecs[6] = xorv(v[6], v[14]);
    495     h_vecs[7] = xorv(v[7], v[15]);
    496 
    497     block_flags = flags;
    498   }
    499 
    500   transpose_vecs(&h_vecs[0]);
    501   transpose_vecs(&h_vecs[4]);
    502   // The first four vecs now contain the first half of each output, and the
    503   // second four vecs contain the second half of each output.
    504   storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
    505   storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
    506   storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
    507   storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
    508   storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
    509   storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
    510   storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
    511   storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
    512 }
    513 
    514 INLINE void hash_one_sse41(const uint8_t *input, size_t blocks,
    515                            const uint32_t key[8], uint64_t counter,
    516                            uint8_t flags, uint8_t flags_start,
    517                            uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
    518   uint32_t cv[8];
    519   memcpy(cv, key, BLAKE3_KEY_LEN);
    520   uint8_t block_flags = flags | flags_start;
    521   while (blocks > 0) {
    522     if (blocks == 1) {
    523       block_flags |= flags_end;
    524     }
    525     blake3_compress_in_place_sse41(cv, input, BLAKE3_BLOCK_LEN, counter,
    526                                    block_flags);
    527     input = &input[BLAKE3_BLOCK_LEN];
    528     blocks -= 1;
    529     block_flags = flags;
    530   }
    531   memcpy(out, cv, BLAKE3_OUT_LEN);
    532 }
    533 
    534 void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs,
    535                             size_t blocks, const uint32_t key[8],
    536                             uint64_t counter, bool increment_counter,
    537                             uint8_t flags, uint8_t flags_start,
    538                             uint8_t flags_end, uint8_t *out) {
    539   while (num_inputs >= DEGREE) {
    540     blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags,
    541                        flags_start, flags_end, out);
    542     if (increment_counter) {
    543       counter += DEGREE;
    544     }
    545     inputs += DEGREE;
    546     num_inputs -= DEGREE;
    547     out = &out[DEGREE * BLAKE3_OUT_LEN];
    548   }
    549   while (num_inputs > 0) {
    550     hash_one_sse41(inputs[0], blocks, key, counter, flags, flags_start,
    551                    flags_end, out);
    552     if (increment_counter) {
    553       counter += 1;
    554     }
    555     inputs += 1;
    556     num_inputs -= 1;
    557     out = &out[BLAKE3_OUT_LEN];
    558   }
    559 }