chibipub

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

blake3_sse2.c (20960B)


      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_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1);
     32 }
     33 
     34 INLINE __m128i rot12(__m128i x) {
     35   return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
     36 }
     37 
     38 INLINE __m128i rot8(__m128i x) {
     39   return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8));
     40 }
     41 
     42 INLINE __m128i rot7(__m128i x) {
     43   return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
     44 }
     45 
     46 INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
     47                __m128i m) {
     48   *row0 = addv(addv(*row0, m), *row1);
     49   *row3 = xorv(*row3, *row0);
     50   *row3 = rot16(*row3);
     51   *row2 = addv(*row2, *row3);
     52   *row1 = xorv(*row1, *row2);
     53   *row1 = rot12(*row1);
     54 }
     55 
     56 INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
     57                __m128i m) {
     58   *row0 = addv(addv(*row0, m), *row1);
     59   *row3 = xorv(*row3, *row0);
     60   *row3 = rot8(*row3);
     61   *row2 = addv(*row2, *row3);
     62   *row1 = xorv(*row1, *row2);
     63   *row1 = rot7(*row1);
     64 }
     65 
     66 // Note the optimization here of leaving row1 as the unrotated row, rather than
     67 // row0. All the message loads below are adjusted to compensate for this. See
     68 // discussion at https://github.com/sneves/blake2-avx2/pull/4
     69 INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
     70   *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
     71   *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
     72   *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
     73 }
     74 
     75 INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
     76   *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
     77   *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
     78   *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
     79 }
     80 
     81 INLINE __m128i blend_epi16(__m128i a, __m128i b, const int imm8) {
     82   const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
     83   __m128i mask = _mm_set1_epi16(imm8);
     84   mask = _mm_and_si128(mask, bits);
     85   mask = _mm_cmpeq_epi16(mask, bits);
     86   return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
     87 }
     88 
     89 INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
     90                          const uint8_t block[BLAKE3_BLOCK_LEN],
     91                          uint8_t block_len, uint64_t counter, uint8_t flags) {
     92   rows[0] = loadu((uint8_t *)&cv[0]);
     93   rows[1] = loadu((uint8_t *)&cv[4]);
     94   rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
     95   rows[3] = set4(counter_low(counter), counter_high(counter),
     96                  (uint32_t)block_len, (uint32_t)flags);
     97 
     98   __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
     99   __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
    100   __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
    101   __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
    102 
    103   __m128i t0, t1, t2, t3, tt;
    104 
    105   // Round 1. The first round permutes the message words from the original
    106   // input order, into the groups that get mixed in parallel.
    107   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); //  6  4  2  0
    108   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    109   t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); //  7  5  3  1
    110   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    111   diagonalize(&rows[0], &rows[2], &rows[3]);
    112   t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10  8
    113   t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3));   // 12 10  8 14
    114   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    115   t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11  9
    116   t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3));   // 13 11  9 15
    117   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    118   undiagonalize(&rows[0], &rows[2], &rows[3]);
    119   m0 = t0;
    120   m1 = t1;
    121   m2 = t2;
    122   m3 = t3;
    123 
    124   // Round 2. This round and all following rounds apply a fixed permutation
    125   // to the message words from the round before.
    126   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    127   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    128   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    129   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    130   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    131   t1 = blend_epi16(tt, t1, 0xCC);
    132   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    133   diagonalize(&rows[0], &rows[2], &rows[3]);
    134   t2 = _mm_unpacklo_epi64(m3, m1);
    135   tt = blend_epi16(t2, m2, 0xC0);
    136   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    137   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    138   t3 = _mm_unpackhi_epi32(m1, m3);
    139   tt = _mm_unpacklo_epi32(m2, t3);
    140   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    141   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    142   undiagonalize(&rows[0], &rows[2], &rows[3]);
    143   m0 = t0;
    144   m1 = t1;
    145   m2 = t2;
    146   m3 = t3;
    147 
    148   // Round 3
    149   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    150   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    151   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    152   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    153   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    154   t1 = blend_epi16(tt, t1, 0xCC);
    155   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    156   diagonalize(&rows[0], &rows[2], &rows[3]);
    157   t2 = _mm_unpacklo_epi64(m3, m1);
    158   tt = blend_epi16(t2, m2, 0xC0);
    159   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    160   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    161   t3 = _mm_unpackhi_epi32(m1, m3);
    162   tt = _mm_unpacklo_epi32(m2, t3);
    163   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    164   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    165   undiagonalize(&rows[0], &rows[2], &rows[3]);
    166   m0 = t0;
    167   m1 = t1;
    168   m2 = t2;
    169   m3 = t3;
    170 
    171   // Round 4
    172   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    173   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    174   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    175   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    176   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    177   t1 = blend_epi16(tt, t1, 0xCC);
    178   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    179   diagonalize(&rows[0], &rows[2], &rows[3]);
    180   t2 = _mm_unpacklo_epi64(m3, m1);
    181   tt = blend_epi16(t2, m2, 0xC0);
    182   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    183   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    184   t3 = _mm_unpackhi_epi32(m1, m3);
    185   tt = _mm_unpacklo_epi32(m2, t3);
    186   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    187   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    188   undiagonalize(&rows[0], &rows[2], &rows[3]);
    189   m0 = t0;
    190   m1 = t1;
    191   m2 = t2;
    192   m3 = t3;
    193 
    194   // Round 5
    195   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    196   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    197   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    198   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    199   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    200   t1 = blend_epi16(tt, t1, 0xCC);
    201   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    202   diagonalize(&rows[0], &rows[2], &rows[3]);
    203   t2 = _mm_unpacklo_epi64(m3, m1);
    204   tt = blend_epi16(t2, m2, 0xC0);
    205   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    206   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    207   t3 = _mm_unpackhi_epi32(m1, m3);
    208   tt = _mm_unpacklo_epi32(m2, t3);
    209   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    210   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    211   undiagonalize(&rows[0], &rows[2], &rows[3]);
    212   m0 = t0;
    213   m1 = t1;
    214   m2 = t2;
    215   m3 = t3;
    216 
    217   // Round 6
    218   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    219   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    220   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    221   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    222   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    223   t1 = blend_epi16(tt, t1, 0xCC);
    224   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    225   diagonalize(&rows[0], &rows[2], &rows[3]);
    226   t2 = _mm_unpacklo_epi64(m3, m1);
    227   tt = blend_epi16(t2, m2, 0xC0);
    228   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    229   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    230   t3 = _mm_unpackhi_epi32(m1, m3);
    231   tt = _mm_unpacklo_epi32(m2, t3);
    232   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    233   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    234   undiagonalize(&rows[0], &rows[2], &rows[3]);
    235   m0 = t0;
    236   m1 = t1;
    237   m2 = t2;
    238   m3 = t3;
    239 
    240   // Round 7
    241   t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
    242   t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
    243   g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
    244   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
    245   tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
    246   t1 = blend_epi16(tt, t1, 0xCC);
    247   g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
    248   diagonalize(&rows[0], &rows[2], &rows[3]);
    249   t2 = _mm_unpacklo_epi64(m3, m1);
    250   tt = blend_epi16(t2, m2, 0xC0);
    251   t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
    252   g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
    253   t3 = _mm_unpackhi_epi32(m1, m3);
    254   tt = _mm_unpacklo_epi32(m2, t3);
    255   t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
    256   g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
    257   undiagonalize(&rows[0], &rows[2], &rows[3]);
    258 }
    259 
    260 void blake3_compress_in_place_sse2(uint32_t cv[8],
    261                                    const uint8_t block[BLAKE3_BLOCK_LEN],
    262                                    uint8_t block_len, uint64_t counter,
    263                                    uint8_t flags) {
    264   __m128i rows[4];
    265   compress_pre(rows, cv, block, block_len, counter, flags);
    266   storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
    267   storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
    268 }
    269 
    270 void blake3_compress_xof_sse2(const uint32_t cv[8],
    271                               const uint8_t block[BLAKE3_BLOCK_LEN],
    272                               uint8_t block_len, uint64_t counter,
    273                               uint8_t flags, uint8_t out[64]) {
    274   __m128i rows[4];
    275   compress_pre(rows, cv, block, block_len, counter, flags);
    276   storeu(xorv(rows[0], rows[2]), &out[0]);
    277   storeu(xorv(rows[1], rows[3]), &out[16]);
    278   storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
    279   storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
    280 }
    281 
    282 INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
    283   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
    284   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
    285   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
    286   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
    287   v[0] = addv(v[0], v[4]);
    288   v[1] = addv(v[1], v[5]);
    289   v[2] = addv(v[2], v[6]);
    290   v[3] = addv(v[3], v[7]);
    291   v[12] = xorv(v[12], v[0]);
    292   v[13] = xorv(v[13], v[1]);
    293   v[14] = xorv(v[14], v[2]);
    294   v[15] = xorv(v[15], v[3]);
    295   v[12] = rot16(v[12]);
    296   v[13] = rot16(v[13]);
    297   v[14] = rot16(v[14]);
    298   v[15] = rot16(v[15]);
    299   v[8] = addv(v[8], v[12]);
    300   v[9] = addv(v[9], v[13]);
    301   v[10] = addv(v[10], v[14]);
    302   v[11] = addv(v[11], v[15]);
    303   v[4] = xorv(v[4], v[8]);
    304   v[5] = xorv(v[5], v[9]);
    305   v[6] = xorv(v[6], v[10]);
    306   v[7] = xorv(v[7], v[11]);
    307   v[4] = rot12(v[4]);
    308   v[5] = rot12(v[5]);
    309   v[6] = rot12(v[6]);
    310   v[7] = rot12(v[7]);
    311   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
    312   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
    313   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
    314   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
    315   v[0] = addv(v[0], v[4]);
    316   v[1] = addv(v[1], v[5]);
    317   v[2] = addv(v[2], v[6]);
    318   v[3] = addv(v[3], v[7]);
    319   v[12] = xorv(v[12], v[0]);
    320   v[13] = xorv(v[13], v[1]);
    321   v[14] = xorv(v[14], v[2]);
    322   v[15] = xorv(v[15], v[3]);
    323   v[12] = rot8(v[12]);
    324   v[13] = rot8(v[13]);
    325   v[14] = rot8(v[14]);
    326   v[15] = rot8(v[15]);
    327   v[8] = addv(v[8], v[12]);
    328   v[9] = addv(v[9], v[13]);
    329   v[10] = addv(v[10], v[14]);
    330   v[11] = addv(v[11], v[15]);
    331   v[4] = xorv(v[4], v[8]);
    332   v[5] = xorv(v[5], v[9]);
    333   v[6] = xorv(v[6], v[10]);
    334   v[7] = xorv(v[7], v[11]);
    335   v[4] = rot7(v[4]);
    336   v[5] = rot7(v[5]);
    337   v[6] = rot7(v[6]);
    338   v[7] = rot7(v[7]);
    339 
    340   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
    341   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
    342   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
    343   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
    344   v[0] = addv(v[0], v[5]);
    345   v[1] = addv(v[1], v[6]);
    346   v[2] = addv(v[2], v[7]);
    347   v[3] = addv(v[3], v[4]);
    348   v[15] = xorv(v[15], v[0]);
    349   v[12] = xorv(v[12], v[1]);
    350   v[13] = xorv(v[13], v[2]);
    351   v[14] = xorv(v[14], v[3]);
    352   v[15] = rot16(v[15]);
    353   v[12] = rot16(v[12]);
    354   v[13] = rot16(v[13]);
    355   v[14] = rot16(v[14]);
    356   v[10] = addv(v[10], v[15]);
    357   v[11] = addv(v[11], v[12]);
    358   v[8] = addv(v[8], v[13]);
    359   v[9] = addv(v[9], v[14]);
    360   v[5] = xorv(v[5], v[10]);
    361   v[6] = xorv(v[6], v[11]);
    362   v[7] = xorv(v[7], v[8]);
    363   v[4] = xorv(v[4], v[9]);
    364   v[5] = rot12(v[5]);
    365   v[6] = rot12(v[6]);
    366   v[7] = rot12(v[7]);
    367   v[4] = rot12(v[4]);
    368   v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
    369   v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
    370   v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
    371   v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
    372   v[0] = addv(v[0], v[5]);
    373   v[1] = addv(v[1], v[6]);
    374   v[2] = addv(v[2], v[7]);
    375   v[3] = addv(v[3], v[4]);
    376   v[15] = xorv(v[15], v[0]);
    377   v[12] = xorv(v[12], v[1]);
    378   v[13] = xorv(v[13], v[2]);
    379   v[14] = xorv(v[14], v[3]);
    380   v[15] = rot8(v[15]);
    381   v[12] = rot8(v[12]);
    382   v[13] = rot8(v[13]);
    383   v[14] = rot8(v[14]);
    384   v[10] = addv(v[10], v[15]);
    385   v[11] = addv(v[11], v[12]);
    386   v[8] = addv(v[8], v[13]);
    387   v[9] = addv(v[9], v[14]);
    388   v[5] = xorv(v[5], v[10]);
    389   v[6] = xorv(v[6], v[11]);
    390   v[7] = xorv(v[7], v[8]);
    391   v[4] = xorv(v[4], v[9]);
    392   v[5] = rot7(v[5]);
    393   v[6] = rot7(v[6]);
    394   v[7] = rot7(v[7]);
    395   v[4] = rot7(v[4]);
    396 }
    397 
    398 INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
    399   // Interleave 32-bit lates. The low unpack is lanes 00/11 and the high is
    400   // 22/33. Note that this doesn't split the vector into two lanes, as the
    401   // AVX2 counterparts do.
    402   __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
    403   __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
    404   __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
    405   __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
    406 
    407   // Interleave 64-bit lanes.
    408   __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
    409   __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
    410   __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
    411   __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
    412 
    413   vecs[0] = abcd_0;
    414   vecs[1] = abcd_1;
    415   vecs[2] = abcd_2;
    416   vecs[3] = abcd_3;
    417 }
    418 
    419 INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
    420                                size_t block_offset, __m128i out[16]) {
    421   out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
    422   out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
    423   out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
    424   out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
    425   out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
    426   out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
    427   out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
    428   out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
    429   out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
    430   out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
    431   out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
    432   out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
    433   out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
    434   out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
    435   out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
    436   out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
    437   for (size_t i = 0; i < 4; ++i) {
    438     _mm_prefetch(&inputs[i][block_offset + 256], _MM_HINT_T0);
    439   }
    440   transpose_vecs(&out[0]);
    441   transpose_vecs(&out[4]);
    442   transpose_vecs(&out[8]);
    443   transpose_vecs(&out[12]);
    444 }
    445 
    446 INLINE void load_counters(uint64_t counter, bool increment_counter,
    447                           __m128i *out_lo, __m128i *out_hi) {
    448   const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
    449   const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
    450   const __m128i add1 = _mm_and_si128(mask, add0);
    451   __m128i l = _mm_add_epi32(_mm_set1_epi32(counter), add1);
    452   __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), 
    453                                   _mm_xor_si128(   l, _mm_set1_epi32(0x80000000)));
    454   __m128i h = _mm_sub_epi32(_mm_set1_epi32(counter >> 32), carry);
    455   *out_lo = l;
    456   *out_hi = h;
    457 }
    458 
    459 void blake3_hash4_sse2(const uint8_t *const *inputs, size_t blocks,
    460                        const uint32_t key[8], uint64_t counter,
    461                        bool increment_counter, uint8_t flags,
    462                        uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
    463   __m128i h_vecs[8] = {
    464       set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
    465       set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
    466   };
    467   __m128i counter_low_vec, counter_high_vec;
    468   load_counters(counter, increment_counter, &counter_low_vec,
    469                 &counter_high_vec);
    470   uint8_t block_flags = flags | flags_start;
    471 
    472   for (size_t block = 0; block < blocks; block++) {
    473     if (block + 1 == blocks) {
    474       block_flags |= flags_end;
    475     }
    476     __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
    477     __m128i block_flags_vec = set1(block_flags);
    478     __m128i msg_vecs[16];
    479     transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
    480 
    481     __m128i v[16] = {
    482         h_vecs[0],       h_vecs[1],        h_vecs[2],     h_vecs[3],
    483         h_vecs[4],       h_vecs[5],        h_vecs[6],     h_vecs[7],
    484         set1(IV[0]),     set1(IV[1]),      set1(IV[2]),   set1(IV[3]),
    485         counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
    486     };
    487     round_fn(v, msg_vecs, 0);
    488     round_fn(v, msg_vecs, 1);
    489     round_fn(v, msg_vecs, 2);
    490     round_fn(v, msg_vecs, 3);
    491     round_fn(v, msg_vecs, 4);
    492     round_fn(v, msg_vecs, 5);
    493     round_fn(v, msg_vecs, 6);
    494     h_vecs[0] = xorv(v[0], v[8]);
    495     h_vecs[1] = xorv(v[1], v[9]);
    496     h_vecs[2] = xorv(v[2], v[10]);
    497     h_vecs[3] = xorv(v[3], v[11]);
    498     h_vecs[4] = xorv(v[4], v[12]);
    499     h_vecs[5] = xorv(v[5], v[13]);
    500     h_vecs[6] = xorv(v[6], v[14]);
    501     h_vecs[7] = xorv(v[7], v[15]);
    502 
    503     block_flags = flags;
    504   }
    505 
    506   transpose_vecs(&h_vecs[0]);
    507   transpose_vecs(&h_vecs[4]);
    508   // The first four vecs now contain the first half of each output, and the
    509   // second four vecs contain the second half of each output.
    510   storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
    511   storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
    512   storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
    513   storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
    514   storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
    515   storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
    516   storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
    517   storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
    518 }
    519 
    520 INLINE void hash_one_sse2(const uint8_t *input, size_t blocks,
    521                           const uint32_t key[8], uint64_t counter,
    522                           uint8_t flags, uint8_t flags_start,
    523                           uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
    524   uint32_t cv[8];
    525   memcpy(cv, key, BLAKE3_KEY_LEN);
    526   uint8_t block_flags = flags | flags_start;
    527   while (blocks > 0) {
    528     if (blocks == 1) {
    529       block_flags |= flags_end;
    530     }
    531     blake3_compress_in_place_sse2(cv, input, BLAKE3_BLOCK_LEN, counter,
    532                                   block_flags);
    533     input = &input[BLAKE3_BLOCK_LEN];
    534     blocks -= 1;
    535     block_flags = flags;
    536   }
    537   memcpy(out, cv, BLAKE3_OUT_LEN);
    538 }
    539 
    540 void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
    541                            size_t blocks, const uint32_t key[8],
    542                            uint64_t counter, bool increment_counter,
    543                            uint8_t flags, uint8_t flags_start,
    544                            uint8_t flags_end, uint8_t *out) {
    545   while (num_inputs >= DEGREE) {
    546     blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags,
    547                       flags_start, flags_end, out);
    548     if (increment_counter) {
    549       counter += DEGREE;
    550     }
    551     inputs += DEGREE;
    552     num_inputs -= DEGREE;
    553     out = &out[DEGREE * BLAKE3_OUT_LEN];
    554   }
    555   while (num_inputs > 0) {
    556     hash_one_sse2(inputs[0], blocks, key, counter, flags, flags_start,
    557                   flags_end, out);
    558     if (increment_counter) {
    559       counter += 1;
    560     }
    561     inputs += 1;
    562     num_inputs -= 1;
    563     out = &out[BLAKE3_OUT_LEN];
    564   }
    565 }