From 505194e3967f1de7d4a356aa759b8283f40167f3 Mon Sep 17 00:00:00 2001 From: Zhihao Zhang Date: Sat, 23 May 2026 08:23:22 +0000 Subject: [PATCH 1/5] vendor: import BLAKE3 1.8.4 SIMD C sources (not yet wired) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit These five files are copied verbatim from the upstream BLAKE3 1.8.4 distribution. They are not yet compiled or wired into CMake — that will happen in a subsequent commit. Co-Authored-By: Claude Opus 4.7 --- include/blake3/blake3_avx2.c | 326 ++++++++ include/blake3/blake3_avx512.c | 1388 ++++++++++++++++++++++++++++++++ include/blake3/blake3_neon.c | 367 +++++++++ include/blake3/blake3_sse2.c | 566 +++++++++++++ include/blake3/blake3_sse41.c | 560 +++++++++++++ 5 files changed, 3207 insertions(+) create mode 100644 include/blake3/blake3_avx2.c create mode 100644 include/blake3/blake3_avx512.c create mode 100644 include/blake3/blake3_neon.c create mode 100644 include/blake3/blake3_sse2.c create mode 100644 include/blake3/blake3_sse41.c diff --git a/include/blake3/blake3_avx2.c b/include/blake3/blake3_avx2.c new file mode 100644 index 0000000..381e7c4 --- /dev/null +++ b/include/blake3/blake3_avx2.c @@ -0,0 +1,326 @@ +#include "blake3_impl.h" + +#include + +#define DEGREE 8 + +INLINE __m256i loadu(const uint8_t src[32]) { + return _mm256_loadu_si256((const __m256i *)src); +} + +INLINE void storeu(__m256i src, uint8_t dest[16]) { + _mm256_storeu_si256((__m256i *)dest, src); +} + +INLINE __m256i addv(__m256i a, __m256i b) { return _mm256_add_epi32(a, b); } + +// Note that clang-format doesn't like the name "xor" for some reason. +INLINE __m256i xorv(__m256i a, __m256i b) { return _mm256_xor_si256(a, b); } + +INLINE __m256i set1(uint32_t x) { return _mm256_set1_epi32((int32_t)x); } + +INLINE __m256i rot16(__m256i x) { + return _mm256_shuffle_epi8( + x, _mm256_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2, + 13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2)); +} + +INLINE __m256i rot12(__m256i x) { + return _mm256_or_si256(_mm256_srli_epi32(x, 12), _mm256_slli_epi32(x, 32 - 12)); +} + +INLINE __m256i rot8(__m256i x) { + return _mm256_shuffle_epi8( + x, _mm256_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1, + 12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1)); +} + +INLINE __m256i rot7(__m256i x) { + return _mm256_or_si256(_mm256_srli_epi32(x, 7), _mm256_slli_epi32(x, 32 - 7)); +} + +INLINE void round_fn(__m256i v[16], __m256i m[16], size_t r) { + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[15] = rot16(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot12(v[4]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[15] = rot8(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot7(v[4]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot16(v[15]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[4] = rot12(v[4]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot8(v[15]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + v[4] = rot7(v[4]); +} + +INLINE void transpose_vecs(__m256i vecs[DEGREE]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11/44/55, and the high + // is 22/33/66/77. + __m256i ab_0145 = _mm256_unpacklo_epi32(vecs[0], vecs[1]); + __m256i ab_2367 = _mm256_unpackhi_epi32(vecs[0], vecs[1]); + __m256i cd_0145 = _mm256_unpacklo_epi32(vecs[2], vecs[3]); + __m256i cd_2367 = _mm256_unpackhi_epi32(vecs[2], vecs[3]); + __m256i ef_0145 = _mm256_unpacklo_epi32(vecs[4], vecs[5]); + __m256i ef_2367 = _mm256_unpackhi_epi32(vecs[4], vecs[5]); + __m256i gh_0145 = _mm256_unpacklo_epi32(vecs[6], vecs[7]); + __m256i gh_2367 = _mm256_unpackhi_epi32(vecs[6], vecs[7]); + + // Interleave 64-bit lanes. The low unpack is lanes 00/22 and the high is + // 11/33. + __m256i abcd_04 = _mm256_unpacklo_epi64(ab_0145, cd_0145); + __m256i abcd_15 = _mm256_unpackhi_epi64(ab_0145, cd_0145); + __m256i abcd_26 = _mm256_unpacklo_epi64(ab_2367, cd_2367); + __m256i abcd_37 = _mm256_unpackhi_epi64(ab_2367, cd_2367); + __m256i efgh_04 = _mm256_unpacklo_epi64(ef_0145, gh_0145); + __m256i efgh_15 = _mm256_unpackhi_epi64(ef_0145, gh_0145); + __m256i efgh_26 = _mm256_unpacklo_epi64(ef_2367, gh_2367); + __m256i efgh_37 = _mm256_unpackhi_epi64(ef_2367, gh_2367); + + // Interleave 128-bit lanes. + vecs[0] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x20); + vecs[1] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x20); + vecs[2] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x20); + vecs[3] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x20); + vecs[4] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x31); + vecs[5] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x31); + vecs[6] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x31); + vecs[7] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x31); +} + +INLINE void transpose_msg_vecs(const uint8_t *const *inputs, + size_t block_offset, __m256i out[16]) { + out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m256i)]); + out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m256i)]); + out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m256i)]); + out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m256i)]); + out[4] = loadu(&inputs[4][block_offset + 0 * sizeof(__m256i)]); + out[5] = loadu(&inputs[5][block_offset + 0 * sizeof(__m256i)]); + out[6] = loadu(&inputs[6][block_offset + 0 * sizeof(__m256i)]); + out[7] = loadu(&inputs[7][block_offset + 0 * sizeof(__m256i)]); + out[8] = loadu(&inputs[0][block_offset + 1 * sizeof(__m256i)]); + out[9] = loadu(&inputs[1][block_offset + 1 * sizeof(__m256i)]); + out[10] = loadu(&inputs[2][block_offset + 1 * sizeof(__m256i)]); + out[11] = loadu(&inputs[3][block_offset + 1 * sizeof(__m256i)]); + out[12] = loadu(&inputs[4][block_offset + 1 * sizeof(__m256i)]); + out[13] = loadu(&inputs[5][block_offset + 1 * sizeof(__m256i)]); + out[14] = loadu(&inputs[6][block_offset + 1 * sizeof(__m256i)]); + out[15] = loadu(&inputs[7][block_offset + 1 * sizeof(__m256i)]); + for (size_t i = 0; i < 8; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs(&out[0]); + transpose_vecs(&out[8]); +} + +INLINE void load_counters(uint64_t counter, bool increment_counter, + __m256i *out_lo, __m256i *out_hi) { + const __m256i mask = _mm256_set1_epi32(-(int32_t)increment_counter); + const __m256i add0 = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0); + const __m256i add1 = _mm256_and_si256(mask, add0); + __m256i l = _mm256_add_epi32(_mm256_set1_epi32((int32_t)counter), add1); + __m256i carry = _mm256_cmpgt_epi32(_mm256_xor_si256(add1, _mm256_set1_epi32(0x80000000)), + _mm256_xor_si256( l, _mm256_set1_epi32(0x80000000))); + __m256i h = _mm256_sub_epi32(_mm256_set1_epi32((int32_t)(counter >> 32)), carry); + *out_lo = l; + *out_hi = h; +} + +static +void blake3_hash8_avx2(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) { + __m256i h_vecs[8] = { + set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]), + set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]), + }; + __m256i counter_low_vec, counter_high_vec; + load_counters(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m256i block_len_vec = set1(BLAKE3_BLOCK_LEN); + __m256i block_flags_vec = set1(block_flags); + __m256i msg_vecs[16]; + transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m256i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn(v, msg_vecs, 0); + round_fn(v, msg_vecs, 1); + round_fn(v, msg_vecs, 2); + round_fn(v, msg_vecs, 3); + round_fn(v, msg_vecs, 4); + round_fn(v, msg_vecs, 5); + round_fn(v, msg_vecs, 6); + h_vecs[0] = xorv(v[0], v[8]); + h_vecs[1] = xorv(v[1], v[9]); + h_vecs[2] = xorv(v[2], v[10]); + h_vecs[3] = xorv(v[3], v[11]); + h_vecs[4] = xorv(v[4], v[12]); + h_vecs[5] = xorv(v[5], v[13]); + h_vecs[6] = xorv(v[6], v[14]); + h_vecs[7] = xorv(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs(h_vecs); + storeu(h_vecs[0], &out[0 * sizeof(__m256i)]); + storeu(h_vecs[1], &out[1 * sizeof(__m256i)]); + storeu(h_vecs[2], &out[2 * sizeof(__m256i)]); + storeu(h_vecs[3], &out[3 * sizeof(__m256i)]); + storeu(h_vecs[4], &out[4 * sizeof(__m256i)]); + storeu(h_vecs[5], &out[5 * sizeof(__m256i)]); + storeu(h_vecs[6], &out[6 * sizeof(__m256i)]); + storeu(h_vecs[7], &out[7 * sizeof(__m256i)]); +} + +#if !defined(BLAKE3_NO_SSE41) +void blake3_hash_many_sse41(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); +#else +void blake3_hash_many_portable(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); +#endif + +void blake3_hash_many_avx2(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) { + while (num_inputs >= DEGREE) { + blake3_hash8_avx2(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += DEGREE; + } + inputs += DEGREE; + num_inputs -= DEGREE; + out = &out[DEGREE * BLAKE3_OUT_LEN]; + } +#if !defined(BLAKE3_NO_SSE41) + blake3_hash_many_sse41(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, out); +#else + blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter, + increment_counter, flags, flags_start, flags_end, + out); +#endif +} diff --git a/include/blake3/blake3_avx512.c b/include/blake3/blake3_avx512.c new file mode 100644 index 0000000..f88a32d --- /dev/null +++ b/include/blake3/blake3_avx512.c @@ -0,0 +1,1388 @@ +#include "blake3_impl.h" + +#include + +#define _mm_shuffle_ps2(a, b, c) \ + (_mm_castps_si128( \ + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c)))) + +INLINE __m128i loadu_128(const uint8_t src[16]) { + return _mm_loadu_si128((void*)src); +} + +INLINE __m256i loadu_256(const uint8_t src[32]) { + return _mm256_loadu_si256((void*)src); +} + +INLINE __m512i loadu_512(const uint8_t src[64]) { + return _mm512_loadu_si512((void*)src); +} + +INLINE void storeu_128(__m128i src, uint8_t dest[16]) { + _mm_storeu_si128((void*)dest, src); +} + +INLINE void storeu_256(__m256i src, uint8_t dest[16]) { + _mm256_storeu_si256((void*)dest, src); +} + +INLINE void storeu_512(__m512i src, uint8_t dest[16]) { + _mm512_storeu_si512((void*)dest, src); +} + +INLINE __m128i add_128(__m128i a, __m128i b) { return _mm_add_epi32(a, b); } + +INLINE __m256i add_256(__m256i a, __m256i b) { return _mm256_add_epi32(a, b); } + +INLINE __m512i add_512(__m512i a, __m512i b) { return _mm512_add_epi32(a, b); } + +INLINE __m128i xor_128(__m128i a, __m128i b) { return _mm_xor_si128(a, b); } + +INLINE __m256i xor_256(__m256i a, __m256i b) { return _mm256_xor_si256(a, b); } + +INLINE __m512i xor_512(__m512i a, __m512i b) { return _mm512_xor_si512(a, b); } + +INLINE __m128i set1_128(uint32_t x) { return _mm_set1_epi32((int32_t)x); } + +INLINE __m256i set1_256(uint32_t x) { return _mm256_set1_epi32((int32_t)x); } + +INLINE __m512i set1_512(uint32_t x) { return _mm512_set1_epi32((int32_t)x); } + +INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d); +} + +INLINE __m128i rot16_128(__m128i x) { return _mm_ror_epi32(x, 16); } + +INLINE __m256i rot16_256(__m256i x) { return _mm256_ror_epi32(x, 16); } + +INLINE __m512i rot16_512(__m512i x) { return _mm512_ror_epi32(x, 16); } + +INLINE __m128i rot12_128(__m128i x) { return _mm_ror_epi32(x, 12); } + +INLINE __m256i rot12_256(__m256i x) { return _mm256_ror_epi32(x, 12); } + +INLINE __m512i rot12_512(__m512i x) { return _mm512_ror_epi32(x, 12); } + +INLINE __m128i rot8_128(__m128i x) { return _mm_ror_epi32(x, 8); } + +INLINE __m256i rot8_256(__m256i x) { return _mm256_ror_epi32(x, 8); } + +INLINE __m512i rot8_512(__m512i x) { return _mm512_ror_epi32(x, 8); } + +INLINE __m128i rot7_128(__m128i x) { return _mm_ror_epi32(x, 7); } + +INLINE __m256i rot7_256(__m256i x) { return _mm256_ror_epi32(x, 7); } + +INLINE __m512i rot7_512(__m512i x) { return _mm512_ror_epi32(x, 7); } + +/* + * ---------------------------------------------------------------------------- + * compress_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = add_128(add_128(*row0, m), *row1); + *row3 = xor_128(*row3, *row0); + *row3 = rot16_128(*row3); + *row2 = add_128(*row2, *row3); + *row1 = xor_128(*row1, *row2); + *row1 = rot12_128(*row1); +} + +INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = add_128(add_128(*row0, m), *row1); + *row3 = xor_128(*row3, *row0); + *row3 = rot8_128(*row3); + *row2 = add_128(*row2, *row3); + *row1 = xor_128(*row1, *row2); + *row1 = rot7_128(*row1); +} + +// Note the optimization here of leaving row1 as the unrotated row, rather than +// row0. All the message loads below are adjusted to compensate for this. See +// discussion at https://github.com/sneves/blake2-avx2/pull/4 +INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1)); +} + +INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3)); +} + +INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + rows[0] = loadu_128((uint8_t *)&cv[0]); + rows[1] = loadu_128((uint8_t *)&cv[4]); + rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); + rows[3] = set4(counter_low(counter), counter_high(counter), + (uint32_t)block_len, (uint32_t)flags); + + __m128i m0 = loadu_128(&block[sizeof(__m128i) * 0]); + __m128i m1 = loadu_128(&block[sizeof(__m128i) * 1]); + __m128i m2 = loadu_128(&block[sizeof(__m128i) * 2]); + __m128i m3 = loadu_128(&block[sizeof(__m128i) * 3]); + + __m128i t0, t1, t2, t3, tt; + + // Round 1. The first round permutes the message words from the original + // input order, into the groups that get mixed in parallel. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8 + t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9 + t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 2. This round and all following rounds apply a fixed permutation + // to the message words from the round before. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 3 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 4 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 5 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 6 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 7 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); +} + +void blake3_compress_xof_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu_128(xor_128(rows[0], rows[2]), &out[0]); + storeu_128(xor_128(rows[1], rows[3]), &out[16]); + storeu_128(xor_128(rows[2], loadu_128((uint8_t *)&cv[0])), &out[32]); + storeu_128(xor_128(rows[3], loadu_128((uint8_t *)&cv[4])), &out[48]); +} + +void blake3_compress_in_place_avx512(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu_128(xor_128(rows[0], rows[2]), (uint8_t *)&cv[0]); + storeu_128(xor_128(rows[1], rows[3]), (uint8_t *)&cv[4]); +} + +/* + * ---------------------------------------------------------------------------- + * hash4_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn4(__m128i v[16], __m128i m[16], size_t r) { + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[15] = rot16_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot12_128(v[4]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[15] = rot8_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot7_128(v[4]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot16_128(v[15]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[4] = rot12_128(v[4]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot8_128(v[15]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + v[4] = rot7_128(v[4]); +} + +INLINE void transpose_vecs_128(__m128i vecs[4]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is + // 22/33. Note that this doesn't split the vector into two lanes, as the + // AVX2 counterparts do. + __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + + // Interleave 64-bit lanes. + __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01); + __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01); + __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23); + __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23); + + vecs[0] = abcd_0; + vecs[1] = abcd_1; + vecs[2] = abcd_2; + vecs[3] = abcd_3; +} + +INLINE void transpose_msg_vecs4(const uint8_t *const *inputs, + size_t block_offset, __m128i out[16]) { + out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(__m128i)]); + out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(__m128i)]); + out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(__m128i)]); + out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(__m128i)]); + out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(__m128i)]); + out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(__m128i)]); + out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(__m128i)]); + out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(__m128i)]); + out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(__m128i)]); + out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(__m128i)]); + out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(__m128i)]); + out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(__m128i)]); + out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(__m128i)]); + out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(__m128i)]); + out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(__m128i)]); + out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(__m128i)]); + for (size_t i = 0; i < 4; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs_128(&out[0]); + transpose_vecs_128(&out[4]); + transpose_vecs_128(&out[8]); + transpose_vecs_128(&out[12]); +} + +INLINE void load_counters4(uint64_t counter, bool increment_counter, + __m128i *out_lo, __m128i *out_hi) { + uint64_t mask = (increment_counter ? ~0 : 0); + __m256i mask_vec = _mm256_set1_epi64x(mask); + __m256i deltas = _mm256_setr_epi64x(0, 1, 2, 3); + deltas = _mm256_and_si256(mask_vec, deltas); + __m256i counters = + _mm256_add_epi64(_mm256_set1_epi64x((int64_t)counter), deltas); + *out_lo = _mm256_cvtepi64_epi32(counters); + *out_hi = _mm256_cvtepi64_epi32(_mm256_srli_epi64(counters, 32)); +} + +static +void blake3_hash4_avx512(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) { + __m128i h_vecs[8] = { + set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]), + set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]), + }; + __m128i counter_low_vec, counter_high_vec; + load_counters4(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m128i block_len_vec = set1_128(BLAKE3_BLOCK_LEN); + __m128i block_flags_vec = set1_128(block_flags); + __m128i msg_vecs[16]; + transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn4(v, msg_vecs, 0); + round_fn4(v, msg_vecs, 1); + round_fn4(v, msg_vecs, 2); + round_fn4(v, msg_vecs, 3); + round_fn4(v, msg_vecs, 4); + round_fn4(v, msg_vecs, 5); + round_fn4(v, msg_vecs, 6); + h_vecs[0] = xor_128(v[0], v[8]); + h_vecs[1] = xor_128(v[1], v[9]); + h_vecs[2] = xor_128(v[2], v[10]); + h_vecs[3] = xor_128(v[3], v[11]); + h_vecs[4] = xor_128(v[4], v[12]); + h_vecs[5] = xor_128(v[5], v[13]); + h_vecs[6] = xor_128(v[6], v[14]); + h_vecs[7] = xor_128(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs_128(&h_vecs[0]); + transpose_vecs_128(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu_128(h_vecs[0], &out[0 * sizeof(__m128i)]); + storeu_128(h_vecs[4], &out[1 * sizeof(__m128i)]); + storeu_128(h_vecs[1], &out[2 * sizeof(__m128i)]); + storeu_128(h_vecs[5], &out[3 * sizeof(__m128i)]); + storeu_128(h_vecs[2], &out[4 * sizeof(__m128i)]); + storeu_128(h_vecs[6], &out[5 * sizeof(__m128i)]); + storeu_128(h_vecs[3], &out[6 * sizeof(__m128i)]); + storeu_128(h_vecs[7], &out[7 * sizeof(__m128i)]); +} + +static +void blake3_xof4_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[4 * 64]) { + __m128i h_vecs[8] = { + set1_128(cv[0]), set1_128(cv[1]), set1_128(cv[2]), set1_128(cv[3]), + set1_128(cv[4]), set1_128(cv[5]), set1_128(cv[6]), set1_128(cv[7]), + }; + uint32_t block_words[16]; + load_block_words(block, block_words); + __m128i msg_vecs[16]; + for (size_t i = 0; i < 16; i++) { + msg_vecs[i] = set1_128(block_words[i]); + } + __m128i counter_low_vec, counter_high_vec; + load_counters4(counter, true, &counter_low_vec, &counter_high_vec); + __m128i block_len_vec = set1_128(block_len); + __m128i block_flags_vec = set1_128(flags); + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn4(v, msg_vecs, 0); + round_fn4(v, msg_vecs, 1); + round_fn4(v, msg_vecs, 2); + round_fn4(v, msg_vecs, 3); + round_fn4(v, msg_vecs, 4); + round_fn4(v, msg_vecs, 5); + round_fn4(v, msg_vecs, 6); + for (size_t i = 0; i < 8; i++) { + v[i] = xor_128(v[i], v[i+8]); + v[i+8] = xor_128(v[i+8], h_vecs[i]); + } + transpose_vecs_128(&v[0]); + transpose_vecs_128(&v[4]); + transpose_vecs_128(&v[8]); + transpose_vecs_128(&v[12]); + for (size_t i = 0; i < 4; i++) { + storeu_128(v[i+ 0], &out[(4*i+0) * sizeof(__m128i)]); + storeu_128(v[i+ 4], &out[(4*i+1) * sizeof(__m128i)]); + storeu_128(v[i+ 8], &out[(4*i+2) * sizeof(__m128i)]); + storeu_128(v[i+12], &out[(4*i+3) * sizeof(__m128i)]); + } +} + +/* + * ---------------------------------------------------------------------------- + * hash8_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn8(__m256i v[16], __m256i m[16], size_t r) { + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_256(v[0], v[4]); + v[1] = add_256(v[1], v[5]); + v[2] = add_256(v[2], v[6]); + v[3] = add_256(v[3], v[7]); + v[12] = xor_256(v[12], v[0]); + v[13] = xor_256(v[13], v[1]); + v[14] = xor_256(v[14], v[2]); + v[15] = xor_256(v[15], v[3]); + v[12] = rot16_256(v[12]); + v[13] = rot16_256(v[13]); + v[14] = rot16_256(v[14]); + v[15] = rot16_256(v[15]); + v[8] = add_256(v[8], v[12]); + v[9] = add_256(v[9], v[13]); + v[10] = add_256(v[10], v[14]); + v[11] = add_256(v[11], v[15]); + v[4] = xor_256(v[4], v[8]); + v[5] = xor_256(v[5], v[9]); + v[6] = xor_256(v[6], v[10]); + v[7] = xor_256(v[7], v[11]); + v[4] = rot12_256(v[4]); + v[5] = rot12_256(v[5]); + v[6] = rot12_256(v[6]); + v[7] = rot12_256(v[7]); + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_256(v[0], v[4]); + v[1] = add_256(v[1], v[5]); + v[2] = add_256(v[2], v[6]); + v[3] = add_256(v[3], v[7]); + v[12] = xor_256(v[12], v[0]); + v[13] = xor_256(v[13], v[1]); + v[14] = xor_256(v[14], v[2]); + v[15] = xor_256(v[15], v[3]); + v[12] = rot8_256(v[12]); + v[13] = rot8_256(v[13]); + v[14] = rot8_256(v[14]); + v[15] = rot8_256(v[15]); + v[8] = add_256(v[8], v[12]); + v[9] = add_256(v[9], v[13]); + v[10] = add_256(v[10], v[14]); + v[11] = add_256(v[11], v[15]); + v[4] = xor_256(v[4], v[8]); + v[5] = xor_256(v[5], v[9]); + v[6] = xor_256(v[6], v[10]); + v[7] = xor_256(v[7], v[11]); + v[4] = rot7_256(v[4]); + v[5] = rot7_256(v[5]); + v[6] = rot7_256(v[6]); + v[7] = rot7_256(v[7]); + + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_256(v[0], v[5]); + v[1] = add_256(v[1], v[6]); + v[2] = add_256(v[2], v[7]); + v[3] = add_256(v[3], v[4]); + v[15] = xor_256(v[15], v[0]); + v[12] = xor_256(v[12], v[1]); + v[13] = xor_256(v[13], v[2]); + v[14] = xor_256(v[14], v[3]); + v[15] = rot16_256(v[15]); + v[12] = rot16_256(v[12]); + v[13] = rot16_256(v[13]); + v[14] = rot16_256(v[14]); + v[10] = add_256(v[10], v[15]); + v[11] = add_256(v[11], v[12]); + v[8] = add_256(v[8], v[13]); + v[9] = add_256(v[9], v[14]); + v[5] = xor_256(v[5], v[10]); + v[6] = xor_256(v[6], v[11]); + v[7] = xor_256(v[7], v[8]); + v[4] = xor_256(v[4], v[9]); + v[5] = rot12_256(v[5]); + v[6] = rot12_256(v[6]); + v[7] = rot12_256(v[7]); + v[4] = rot12_256(v[4]); + v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_256(v[0], v[5]); + v[1] = add_256(v[1], v[6]); + v[2] = add_256(v[2], v[7]); + v[3] = add_256(v[3], v[4]); + v[15] = xor_256(v[15], v[0]); + v[12] = xor_256(v[12], v[1]); + v[13] = xor_256(v[13], v[2]); + v[14] = xor_256(v[14], v[3]); + v[15] = rot8_256(v[15]); + v[12] = rot8_256(v[12]); + v[13] = rot8_256(v[13]); + v[14] = rot8_256(v[14]); + v[10] = add_256(v[10], v[15]); + v[11] = add_256(v[11], v[12]); + v[8] = add_256(v[8], v[13]); + v[9] = add_256(v[9], v[14]); + v[5] = xor_256(v[5], v[10]); + v[6] = xor_256(v[6], v[11]); + v[7] = xor_256(v[7], v[8]); + v[4] = xor_256(v[4], v[9]); + v[5] = rot7_256(v[5]); + v[6] = rot7_256(v[6]); + v[7] = rot7_256(v[7]); + v[4] = rot7_256(v[4]); +} + +INLINE void transpose_vecs_256(__m256i vecs[8]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11/44/55, and the high + // is 22/33/66/77. + __m256i ab_0145 = _mm256_unpacklo_epi32(vecs[0], vecs[1]); + __m256i ab_2367 = _mm256_unpackhi_epi32(vecs[0], vecs[1]); + __m256i cd_0145 = _mm256_unpacklo_epi32(vecs[2], vecs[3]); + __m256i cd_2367 = _mm256_unpackhi_epi32(vecs[2], vecs[3]); + __m256i ef_0145 = _mm256_unpacklo_epi32(vecs[4], vecs[5]); + __m256i ef_2367 = _mm256_unpackhi_epi32(vecs[4], vecs[5]); + __m256i gh_0145 = _mm256_unpacklo_epi32(vecs[6], vecs[7]); + __m256i gh_2367 = _mm256_unpackhi_epi32(vecs[6], vecs[7]); + + // Interleave 64-bit lanes. The low unpack is lanes 00/22 and the high is + // 11/33. + __m256i abcd_04 = _mm256_unpacklo_epi64(ab_0145, cd_0145); + __m256i abcd_15 = _mm256_unpackhi_epi64(ab_0145, cd_0145); + __m256i abcd_26 = _mm256_unpacklo_epi64(ab_2367, cd_2367); + __m256i abcd_37 = _mm256_unpackhi_epi64(ab_2367, cd_2367); + __m256i efgh_04 = _mm256_unpacklo_epi64(ef_0145, gh_0145); + __m256i efgh_15 = _mm256_unpackhi_epi64(ef_0145, gh_0145); + __m256i efgh_26 = _mm256_unpacklo_epi64(ef_2367, gh_2367); + __m256i efgh_37 = _mm256_unpackhi_epi64(ef_2367, gh_2367); + + // Interleave 128-bit lanes. + vecs[0] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x20); + vecs[1] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x20); + vecs[2] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x20); + vecs[3] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x20); + vecs[4] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x31); + vecs[5] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x31); + vecs[6] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x31); + vecs[7] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x31); +} + +INLINE void transpose_msg_vecs8(const uint8_t *const *inputs, + size_t block_offset, __m256i out[16]) { + out[0] = loadu_256(&inputs[0][block_offset + 0 * sizeof(__m256i)]); + out[1] = loadu_256(&inputs[1][block_offset + 0 * sizeof(__m256i)]); + out[2] = loadu_256(&inputs[2][block_offset + 0 * sizeof(__m256i)]); + out[3] = loadu_256(&inputs[3][block_offset + 0 * sizeof(__m256i)]); + out[4] = loadu_256(&inputs[4][block_offset + 0 * sizeof(__m256i)]); + out[5] = loadu_256(&inputs[5][block_offset + 0 * sizeof(__m256i)]); + out[6] = loadu_256(&inputs[6][block_offset + 0 * sizeof(__m256i)]); + out[7] = loadu_256(&inputs[7][block_offset + 0 * sizeof(__m256i)]); + out[8] = loadu_256(&inputs[0][block_offset + 1 * sizeof(__m256i)]); + out[9] = loadu_256(&inputs[1][block_offset + 1 * sizeof(__m256i)]); + out[10] = loadu_256(&inputs[2][block_offset + 1 * sizeof(__m256i)]); + out[11] = loadu_256(&inputs[3][block_offset + 1 * sizeof(__m256i)]); + out[12] = loadu_256(&inputs[4][block_offset + 1 * sizeof(__m256i)]); + out[13] = loadu_256(&inputs[5][block_offset + 1 * sizeof(__m256i)]); + out[14] = loadu_256(&inputs[6][block_offset + 1 * sizeof(__m256i)]); + out[15] = loadu_256(&inputs[7][block_offset + 1 * sizeof(__m256i)]); + for (size_t i = 0; i < 8; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs_256(&out[0]); + transpose_vecs_256(&out[8]); +} + +INLINE void load_counters8(uint64_t counter, bool increment_counter, + __m256i *out_lo, __m256i *out_hi) { + uint64_t mask = (increment_counter ? ~0 : 0); + __m512i mask_vec = _mm512_set1_epi64(mask); + __m512i deltas = _mm512_setr_epi64(0, 1, 2, 3, 4, 5, 6, 7); + deltas = _mm512_and_si512(mask_vec, deltas); + __m512i counters = + _mm512_add_epi64(_mm512_set1_epi64((int64_t)counter), deltas); + *out_lo = _mm512_cvtepi64_epi32(counters); + *out_hi = _mm512_cvtepi64_epi32(_mm512_srli_epi64(counters, 32)); +} + +static +void blake3_hash8_avx512(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) { + __m256i h_vecs[8] = { + set1_256(key[0]), set1_256(key[1]), set1_256(key[2]), set1_256(key[3]), + set1_256(key[4]), set1_256(key[5]), set1_256(key[6]), set1_256(key[7]), + }; + __m256i counter_low_vec, counter_high_vec; + load_counters8(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m256i block_len_vec = set1_256(BLAKE3_BLOCK_LEN); + __m256i block_flags_vec = set1_256(block_flags); + __m256i msg_vecs[16]; + transpose_msg_vecs8(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m256i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_256(IV[0]), set1_256(IV[1]), set1_256(IV[2]), set1_256(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn8(v, msg_vecs, 0); + round_fn8(v, msg_vecs, 1); + round_fn8(v, msg_vecs, 2); + round_fn8(v, msg_vecs, 3); + round_fn8(v, msg_vecs, 4); + round_fn8(v, msg_vecs, 5); + round_fn8(v, msg_vecs, 6); + h_vecs[0] = xor_256(v[0], v[8]); + h_vecs[1] = xor_256(v[1], v[9]); + h_vecs[2] = xor_256(v[2], v[10]); + h_vecs[3] = xor_256(v[3], v[11]); + h_vecs[4] = xor_256(v[4], v[12]); + h_vecs[5] = xor_256(v[5], v[13]); + h_vecs[6] = xor_256(v[6], v[14]); + h_vecs[7] = xor_256(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs_256(h_vecs); + storeu_256(h_vecs[0], &out[0 * sizeof(__m256i)]); + storeu_256(h_vecs[1], &out[1 * sizeof(__m256i)]); + storeu_256(h_vecs[2], &out[2 * sizeof(__m256i)]); + storeu_256(h_vecs[3], &out[3 * sizeof(__m256i)]); + storeu_256(h_vecs[4], &out[4 * sizeof(__m256i)]); + storeu_256(h_vecs[5], &out[5 * sizeof(__m256i)]); + storeu_256(h_vecs[6], &out[6 * sizeof(__m256i)]); + storeu_256(h_vecs[7], &out[7 * sizeof(__m256i)]); +} + +static +void blake3_xof8_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[8 * 64]) { + __m256i h_vecs[8] = { + set1_256(cv[0]), set1_256(cv[1]), set1_256(cv[2]), set1_256(cv[3]), + set1_256(cv[4]), set1_256(cv[5]), set1_256(cv[6]), set1_256(cv[7]), + }; + uint32_t block_words[16]; + load_block_words(block, block_words); + __m256i msg_vecs[16]; + for (size_t i = 0; i < 16; i++) { + msg_vecs[i] = set1_256(block_words[i]); + } + __m256i counter_low_vec, counter_high_vec; + load_counters8(counter, true, &counter_low_vec, &counter_high_vec); + __m256i block_len_vec = set1_256(block_len); + __m256i block_flags_vec = set1_256(flags); + __m256i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_256(IV[0]), set1_256(IV[1]), set1_256(IV[2]), set1_256(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn8(v, msg_vecs, 0); + round_fn8(v, msg_vecs, 1); + round_fn8(v, msg_vecs, 2); + round_fn8(v, msg_vecs, 3); + round_fn8(v, msg_vecs, 4); + round_fn8(v, msg_vecs, 5); + round_fn8(v, msg_vecs, 6); + for (size_t i = 0; i < 8; i++) { + v[i] = xor_256(v[i], v[i+8]); + v[i+8] = xor_256(v[i+8], h_vecs[i]); + } + transpose_vecs_256(&v[0]); + transpose_vecs_256(&v[8]); + for (size_t i = 0; i < 8; i++) { + storeu_256(v[i+0], &out[(2*i+0) * sizeof(__m256i)]); + storeu_256(v[i+8], &out[(2*i+1) * sizeof(__m256i)]); + } +} + +/* + * ---------------------------------------------------------------------------- + * hash16_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn16(__m512i v[16], __m512i m[16], size_t r) { + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_512(v[0], v[4]); + v[1] = add_512(v[1], v[5]); + v[2] = add_512(v[2], v[6]); + v[3] = add_512(v[3], v[7]); + v[12] = xor_512(v[12], v[0]); + v[13] = xor_512(v[13], v[1]); + v[14] = xor_512(v[14], v[2]); + v[15] = xor_512(v[15], v[3]); + v[12] = rot16_512(v[12]); + v[13] = rot16_512(v[13]); + v[14] = rot16_512(v[14]); + v[15] = rot16_512(v[15]); + v[8] = add_512(v[8], v[12]); + v[9] = add_512(v[9], v[13]); + v[10] = add_512(v[10], v[14]); + v[11] = add_512(v[11], v[15]); + v[4] = xor_512(v[4], v[8]); + v[5] = xor_512(v[5], v[9]); + v[6] = xor_512(v[6], v[10]); + v[7] = xor_512(v[7], v[11]); + v[4] = rot12_512(v[4]); + v[5] = rot12_512(v[5]); + v[6] = rot12_512(v[6]); + v[7] = rot12_512(v[7]); + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_512(v[0], v[4]); + v[1] = add_512(v[1], v[5]); + v[2] = add_512(v[2], v[6]); + v[3] = add_512(v[3], v[7]); + v[12] = xor_512(v[12], v[0]); + v[13] = xor_512(v[13], v[1]); + v[14] = xor_512(v[14], v[2]); + v[15] = xor_512(v[15], v[3]); + v[12] = rot8_512(v[12]); + v[13] = rot8_512(v[13]); + v[14] = rot8_512(v[14]); + v[15] = rot8_512(v[15]); + v[8] = add_512(v[8], v[12]); + v[9] = add_512(v[9], v[13]); + v[10] = add_512(v[10], v[14]); + v[11] = add_512(v[11], v[15]); + v[4] = xor_512(v[4], v[8]); + v[5] = xor_512(v[5], v[9]); + v[6] = xor_512(v[6], v[10]); + v[7] = xor_512(v[7], v[11]); + v[4] = rot7_512(v[4]); + v[5] = rot7_512(v[5]); + v[6] = rot7_512(v[6]); + v[7] = rot7_512(v[7]); + + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_512(v[0], v[5]); + v[1] = add_512(v[1], v[6]); + v[2] = add_512(v[2], v[7]); + v[3] = add_512(v[3], v[4]); + v[15] = xor_512(v[15], v[0]); + v[12] = xor_512(v[12], v[1]); + v[13] = xor_512(v[13], v[2]); + v[14] = xor_512(v[14], v[3]); + v[15] = rot16_512(v[15]); + v[12] = rot16_512(v[12]); + v[13] = rot16_512(v[13]); + v[14] = rot16_512(v[14]); + v[10] = add_512(v[10], v[15]); + v[11] = add_512(v[11], v[12]); + v[8] = add_512(v[8], v[13]); + v[9] = add_512(v[9], v[14]); + v[5] = xor_512(v[5], v[10]); + v[6] = xor_512(v[6], v[11]); + v[7] = xor_512(v[7], v[8]); + v[4] = xor_512(v[4], v[9]); + v[5] = rot12_512(v[5]); + v[6] = rot12_512(v[6]); + v[7] = rot12_512(v[7]); + v[4] = rot12_512(v[4]); + v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_512(v[0], v[5]); + v[1] = add_512(v[1], v[6]); + v[2] = add_512(v[2], v[7]); + v[3] = add_512(v[3], v[4]); + v[15] = xor_512(v[15], v[0]); + v[12] = xor_512(v[12], v[1]); + v[13] = xor_512(v[13], v[2]); + v[14] = xor_512(v[14], v[3]); + v[15] = rot8_512(v[15]); + v[12] = rot8_512(v[12]); + v[13] = rot8_512(v[13]); + v[14] = rot8_512(v[14]); + v[10] = add_512(v[10], v[15]); + v[11] = add_512(v[11], v[12]); + v[8] = add_512(v[8], v[13]); + v[9] = add_512(v[9], v[14]); + v[5] = xor_512(v[5], v[10]); + v[6] = xor_512(v[6], v[11]); + v[7] = xor_512(v[7], v[8]); + v[4] = xor_512(v[4], v[9]); + v[5] = rot7_512(v[5]); + v[6] = rot7_512(v[6]); + v[7] = rot7_512(v[7]); + v[4] = rot7_512(v[4]); +} + +// 0b10001000, or lanes a0/a2/b0/b2 in little-endian order +#define LO_IMM8 0x88 + +INLINE __m512i unpack_lo_128(__m512i a, __m512i b) { + return _mm512_shuffle_i32x4(a, b, LO_IMM8); +} + +// 0b11011101, or lanes a1/a3/b1/b3 in little-endian order +#define HI_IMM8 0xdd + +INLINE __m512i unpack_hi_128(__m512i a, __m512i b) { + return _mm512_shuffle_i32x4(a, b, HI_IMM8); +} + +INLINE void transpose_vecs_512(__m512i vecs[16]) { + // Interleave 32-bit lanes. The _0 unpack is lanes + // 0/0/1/1/4/4/5/5/8/8/9/9/12/12/13/13, and the _2 unpack is lanes + // 2/2/3/3/6/6/7/7/10/10/11/11/14/14/15/15. + __m512i ab_0 = _mm512_unpacklo_epi32(vecs[0], vecs[1]); + __m512i ab_2 = _mm512_unpackhi_epi32(vecs[0], vecs[1]); + __m512i cd_0 = _mm512_unpacklo_epi32(vecs[2], vecs[3]); + __m512i cd_2 = _mm512_unpackhi_epi32(vecs[2], vecs[3]); + __m512i ef_0 = _mm512_unpacklo_epi32(vecs[4], vecs[5]); + __m512i ef_2 = _mm512_unpackhi_epi32(vecs[4], vecs[5]); + __m512i gh_0 = _mm512_unpacklo_epi32(vecs[6], vecs[7]); + __m512i gh_2 = _mm512_unpackhi_epi32(vecs[6], vecs[7]); + __m512i ij_0 = _mm512_unpacklo_epi32(vecs[8], vecs[9]); + __m512i ij_2 = _mm512_unpackhi_epi32(vecs[8], vecs[9]); + __m512i kl_0 = _mm512_unpacklo_epi32(vecs[10], vecs[11]); + __m512i kl_2 = _mm512_unpackhi_epi32(vecs[10], vecs[11]); + __m512i mn_0 = _mm512_unpacklo_epi32(vecs[12], vecs[13]); + __m512i mn_2 = _mm512_unpackhi_epi32(vecs[12], vecs[13]); + __m512i op_0 = _mm512_unpacklo_epi32(vecs[14], vecs[15]); + __m512i op_2 = _mm512_unpackhi_epi32(vecs[14], vecs[15]); + + // Interleave 64-bit lanes. The _0 unpack is lanes + // 0/0/0/0/4/4/4/4/8/8/8/8/12/12/12/12, the _1 unpack is lanes + // 1/1/1/1/5/5/5/5/9/9/9/9/13/13/13/13, the _2 unpack is lanes + // 2/2/2/2/6/6/6/6/10/10/10/10/14/14/14/14, and the _3 unpack is lanes + // 3/3/3/3/7/7/7/7/11/11/11/11/15/15/15/15. + __m512i abcd_0 = _mm512_unpacklo_epi64(ab_0, cd_0); + __m512i abcd_1 = _mm512_unpackhi_epi64(ab_0, cd_0); + __m512i abcd_2 = _mm512_unpacklo_epi64(ab_2, cd_2); + __m512i abcd_3 = _mm512_unpackhi_epi64(ab_2, cd_2); + __m512i efgh_0 = _mm512_unpacklo_epi64(ef_0, gh_0); + __m512i efgh_1 = _mm512_unpackhi_epi64(ef_0, gh_0); + __m512i efgh_2 = _mm512_unpacklo_epi64(ef_2, gh_2); + __m512i efgh_3 = _mm512_unpackhi_epi64(ef_2, gh_2); + __m512i ijkl_0 = _mm512_unpacklo_epi64(ij_0, kl_0); + __m512i ijkl_1 = _mm512_unpackhi_epi64(ij_0, kl_0); + __m512i ijkl_2 = _mm512_unpacklo_epi64(ij_2, kl_2); + __m512i ijkl_3 = _mm512_unpackhi_epi64(ij_2, kl_2); + __m512i mnop_0 = _mm512_unpacklo_epi64(mn_0, op_0); + __m512i mnop_1 = _mm512_unpackhi_epi64(mn_0, op_0); + __m512i mnop_2 = _mm512_unpacklo_epi64(mn_2, op_2); + __m512i mnop_3 = _mm512_unpackhi_epi64(mn_2, op_2); + + // Interleave 128-bit lanes. The _0 unpack is + // 0/0/0/0/8/8/8/8/0/0/0/0/8/8/8/8, the _1 unpack is + // 1/1/1/1/9/9/9/9/1/1/1/1/9/9/9/9, and so on. + __m512i abcdefgh_0 = unpack_lo_128(abcd_0, efgh_0); + __m512i abcdefgh_1 = unpack_lo_128(abcd_1, efgh_1); + __m512i abcdefgh_2 = unpack_lo_128(abcd_2, efgh_2); + __m512i abcdefgh_3 = unpack_lo_128(abcd_3, efgh_3); + __m512i abcdefgh_4 = unpack_hi_128(abcd_0, efgh_0); + __m512i abcdefgh_5 = unpack_hi_128(abcd_1, efgh_1); + __m512i abcdefgh_6 = unpack_hi_128(abcd_2, efgh_2); + __m512i abcdefgh_7 = unpack_hi_128(abcd_3, efgh_3); + __m512i ijklmnop_0 = unpack_lo_128(ijkl_0, mnop_0); + __m512i ijklmnop_1 = unpack_lo_128(ijkl_1, mnop_1); + __m512i ijklmnop_2 = unpack_lo_128(ijkl_2, mnop_2); + __m512i ijklmnop_3 = unpack_lo_128(ijkl_3, mnop_3); + __m512i ijklmnop_4 = unpack_hi_128(ijkl_0, mnop_0); + __m512i ijklmnop_5 = unpack_hi_128(ijkl_1, mnop_1); + __m512i ijklmnop_6 = unpack_hi_128(ijkl_2, mnop_2); + __m512i ijklmnop_7 = unpack_hi_128(ijkl_3, mnop_3); + + // Interleave 128-bit lanes again for the final outputs. + vecs[0] = unpack_lo_128(abcdefgh_0, ijklmnop_0); + vecs[1] = unpack_lo_128(abcdefgh_1, ijklmnop_1); + vecs[2] = unpack_lo_128(abcdefgh_2, ijklmnop_2); + vecs[3] = unpack_lo_128(abcdefgh_3, ijklmnop_3); + vecs[4] = unpack_lo_128(abcdefgh_4, ijklmnop_4); + vecs[5] = unpack_lo_128(abcdefgh_5, ijklmnop_5); + vecs[6] = unpack_lo_128(abcdefgh_6, ijklmnop_6); + vecs[7] = unpack_lo_128(abcdefgh_7, ijklmnop_7); + vecs[8] = unpack_hi_128(abcdefgh_0, ijklmnop_0); + vecs[9] = unpack_hi_128(abcdefgh_1, ijklmnop_1); + vecs[10] = unpack_hi_128(abcdefgh_2, ijklmnop_2); + vecs[11] = unpack_hi_128(abcdefgh_3, ijklmnop_3); + vecs[12] = unpack_hi_128(abcdefgh_4, ijklmnop_4); + vecs[13] = unpack_hi_128(abcdefgh_5, ijklmnop_5); + vecs[14] = unpack_hi_128(abcdefgh_6, ijklmnop_6); + vecs[15] = unpack_hi_128(abcdefgh_7, ijklmnop_7); +} + +INLINE void transpose_msg_vecs16(const uint8_t *const *inputs, + size_t block_offset, __m512i out[16]) { + out[0] = loadu_512(&inputs[0][block_offset]); + out[1] = loadu_512(&inputs[1][block_offset]); + out[2] = loadu_512(&inputs[2][block_offset]); + out[3] = loadu_512(&inputs[3][block_offset]); + out[4] = loadu_512(&inputs[4][block_offset]); + out[5] = loadu_512(&inputs[5][block_offset]); + out[6] = loadu_512(&inputs[6][block_offset]); + out[7] = loadu_512(&inputs[7][block_offset]); + out[8] = loadu_512(&inputs[8][block_offset]); + out[9] = loadu_512(&inputs[9][block_offset]); + out[10] = loadu_512(&inputs[10][block_offset]); + out[11] = loadu_512(&inputs[11][block_offset]); + out[12] = loadu_512(&inputs[12][block_offset]); + out[13] = loadu_512(&inputs[13][block_offset]); + out[14] = loadu_512(&inputs[14][block_offset]); + out[15] = loadu_512(&inputs[15][block_offset]); + for (size_t i = 0; i < 16; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs_512(out); +} + +INLINE void load_counters16(uint64_t counter, bool increment_counter, + __m512i *out_lo, __m512i *out_hi) { + const __m512i mask = _mm512_set1_epi32(-(int32_t)increment_counter); + const __m512i deltas = _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); + const __m512i masked_deltas = _mm512_and_si512(deltas, mask); + const __m512i low_words = _mm512_add_epi32( + _mm512_set1_epi32((int32_t)counter), + masked_deltas); + // The carry bit is 1 if the high bit of the word was 1 before addition and is + // 0 after. + // NOTE: It would be a bit more natural to use _mm512_cmp_epu32_mask to + // compute the carry bits here, and originally we did, but that intrinsic is + // broken under GCC 5.4. See https://github.com/BLAKE3-team/BLAKE3/issues/271. + const __m512i carries = _mm512_srli_epi32( + _mm512_andnot_si512( + low_words, // 0 after (gets inverted by andnot) + _mm512_set1_epi32((int32_t)counter)), // and 1 before + 31); + const __m512i high_words = _mm512_add_epi32( + _mm512_set1_epi32((int32_t)(counter >> 32)), + carries); + *out_lo = low_words; + *out_hi = high_words; +} + +static +void blake3_hash16_avx512(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) { + __m512i h_vecs[8] = { + set1_512(key[0]), set1_512(key[1]), set1_512(key[2]), set1_512(key[3]), + set1_512(key[4]), set1_512(key[5]), set1_512(key[6]), set1_512(key[7]), + }; + __m512i counter_low_vec, counter_high_vec; + load_counters16(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m512i block_len_vec = set1_512(BLAKE3_BLOCK_LEN); + __m512i block_flags_vec = set1_512(block_flags); + __m512i msg_vecs[16]; + transpose_msg_vecs16(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m512i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_512(IV[0]), set1_512(IV[1]), set1_512(IV[2]), set1_512(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn16(v, msg_vecs, 0); + round_fn16(v, msg_vecs, 1); + round_fn16(v, msg_vecs, 2); + round_fn16(v, msg_vecs, 3); + round_fn16(v, msg_vecs, 4); + round_fn16(v, msg_vecs, 5); + round_fn16(v, msg_vecs, 6); + h_vecs[0] = xor_512(v[0], v[8]); + h_vecs[1] = xor_512(v[1], v[9]); + h_vecs[2] = xor_512(v[2], v[10]); + h_vecs[3] = xor_512(v[3], v[11]); + h_vecs[4] = xor_512(v[4], v[12]); + h_vecs[5] = xor_512(v[5], v[13]); + h_vecs[6] = xor_512(v[6], v[14]); + h_vecs[7] = xor_512(v[7], v[15]); + + block_flags = flags; + } + + // transpose_vecs_512 operates on a 16x16 matrix of words, but we only have 8 + // state vectors. Pad the matrix with zeros. After transposition, store the + // lower half of each vector. + __m512i padded[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_512(0), set1_512(0), set1_512(0), set1_512(0), + set1_512(0), set1_512(0), set1_512(0), set1_512(0), + }; + transpose_vecs_512(padded); + _mm256_mask_storeu_epi32(&out[0 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[0])); + _mm256_mask_storeu_epi32(&out[1 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[1])); + _mm256_mask_storeu_epi32(&out[2 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[2])); + _mm256_mask_storeu_epi32(&out[3 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[3])); + _mm256_mask_storeu_epi32(&out[4 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[4])); + _mm256_mask_storeu_epi32(&out[5 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[5])); + _mm256_mask_storeu_epi32(&out[6 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[6])); + _mm256_mask_storeu_epi32(&out[7 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[7])); + _mm256_mask_storeu_epi32(&out[8 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[8])); + _mm256_mask_storeu_epi32(&out[9 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[9])); + _mm256_mask_storeu_epi32(&out[10 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[10])); + _mm256_mask_storeu_epi32(&out[11 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[11])); + _mm256_mask_storeu_epi32(&out[12 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[12])); + _mm256_mask_storeu_epi32(&out[13 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[13])); + _mm256_mask_storeu_epi32(&out[14 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[14])); + _mm256_mask_storeu_epi32(&out[15 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[15])); +} + +static +void blake3_xof16_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t out[16 * 64]) { + __m512i h_vecs[8] = { + set1_512(cv[0]), set1_512(cv[1]), set1_512(cv[2]), set1_512(cv[3]), + set1_512(cv[4]), set1_512(cv[5]), set1_512(cv[6]), set1_512(cv[7]), + }; + uint32_t block_words[16]; + load_block_words(block, block_words); + __m512i msg_vecs[16]; + for (size_t i = 0; i < 16; i++) { + msg_vecs[i] = set1_512(block_words[i]); + } + __m512i counter_low_vec, counter_high_vec; + load_counters16(counter, true, &counter_low_vec, &counter_high_vec); + __m512i block_len_vec = set1_512(block_len); + __m512i block_flags_vec = set1_512(flags); + __m512i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_512(IV[0]), set1_512(IV[1]), set1_512(IV[2]), set1_512(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn16(v, msg_vecs, 0); + round_fn16(v, msg_vecs, 1); + round_fn16(v, msg_vecs, 2); + round_fn16(v, msg_vecs, 3); + round_fn16(v, msg_vecs, 4); + round_fn16(v, msg_vecs, 5); + round_fn16(v, msg_vecs, 6); + for (size_t i = 0; i < 8; i++) { + v[i] = xor_512(v[i], v[i+8]); + v[i+8] = xor_512(v[i+8], h_vecs[i]); + } + transpose_vecs_512(&v[0]); + for (size_t i = 0; i < 16; i++) { + storeu_512(v[i], &out[i * sizeof(__m512i)]); + } +} + +/* + * ---------------------------------------------------------------------------- + * hash_many_avx512 + * ---------------------------------------------------------------------------- + */ + +INLINE void hash_one_avx512(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]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_avx512(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_avx512(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) { + while (num_inputs >= 16) { + blake3_hash16_avx512(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 16; + } + inputs += 16; + num_inputs -= 16; + out = &out[16 * BLAKE3_OUT_LEN]; + } + while (num_inputs >= 8) { + blake3_hash8_avx512(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 8; + } + inputs += 8; + num_inputs -= 8; + out = &out[8 * BLAKE3_OUT_LEN]; + } + while (num_inputs >= 4) { + blake3_hash4_avx512(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 4; + } + inputs += 4; + num_inputs -= 4; + out = &out[4 * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_avx512(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} + +void blake3_xof_many_avx512(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags, + uint8_t* out, size_t outblocks) { + while (outblocks >= 16) { + blake3_xof16_avx512(cv, block, block_len, counter, flags, out); + counter += 16; + outblocks -= 16; + out += 16 * BLAKE3_BLOCK_LEN; + } + while (outblocks >= 8) { + blake3_xof8_avx512(cv, block, block_len, counter, flags, out); + counter += 8; + outblocks -= 8; + out += 8 * BLAKE3_BLOCK_LEN; + } + while (outblocks >= 4) { + blake3_xof4_avx512(cv, block, block_len, counter, flags, out); + counter += 4; + outblocks -= 4; + out += 4 * BLAKE3_BLOCK_LEN; + } + while (outblocks > 0) { + blake3_compress_xof_avx512(cv, block, block_len, counter, flags, out); + counter += 1; + outblocks -= 1; + out += BLAKE3_BLOCK_LEN; + } +} diff --git a/include/blake3/blake3_neon.c b/include/blake3/blake3_neon.c new file mode 100644 index 0000000..794ea80 --- /dev/null +++ b/include/blake3/blake3_neon.c @@ -0,0 +1,367 @@ +#include "blake3_impl.h" + +#include + +#ifdef __ARM_BIG_ENDIAN +#error "This implementation only supports little-endian ARM." +// It might be that all we need for big-endian support here is to get the loads +// and stores right, but step zero would be finding a way to test it in CI. +#endif + +INLINE uint32x4_t loadu_128(const uint8_t src[16]) { + // vld1q_u32 has alignment requirements. Don't use it. + return vreinterpretq_u32_u8(vld1q_u8(src)); +} + +INLINE void storeu_128(uint32x4_t src, uint8_t dest[16]) { + // vst1q_u32 has alignment requirements. Don't use it. + vst1q_u8(dest, vreinterpretq_u8_u32(src)); +} + +INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) { + return vaddq_u32(a, b); +} + +INLINE uint32x4_t xor_128(uint32x4_t a, uint32x4_t b) { + return veorq_u32(a, b); +} + +INLINE uint32x4_t set1_128(uint32_t x) { return vld1q_dup_u32(&x); } + +INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + uint32_t array[4] = {a, b, c, d}; + return vld1q_u32(array); +} + +INLINE uint32x4_t rot16_128(uint32x4_t x) { + // The straightforward implementation would be two shifts and an or, but that's + // slower on microarchitectures we've tested. See + // https://github.com/BLAKE3-team/BLAKE3/pull/319. + // return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16)); + return vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(x))); +} + +INLINE uint32x4_t rot12_128(uint32x4_t x) { + // See comment in rot16_128. + // return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12)); + return vsriq_n_u32(vshlq_n_u32(x, 32-12), x, 12); +} + +INLINE uint32x4_t rot8_128(uint32x4_t x) { + // See comment in rot16_128. + // return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8)); +#if defined(__clang__) + 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)); +#elif __GNUC__ * 10000 + __GNUC_MINOR__ * 100 >=40700 + static const uint8x16_t r8 = {1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12}; + return vreinterpretq_u32_u8(__builtin_shuffle(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), r8)); +#else + return vsriq_n_u32(vshlq_n_u32(x, 32-8), x, 8); +#endif +} + +INLINE uint32x4_t rot7_128(uint32x4_t x) { + // See comment in rot16_128. + // return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7)); + return vsriq_n_u32(vshlq_n_u32(x, 32-7), x, 7); +} + +// TODO: compress_neon + +// TODO: hash2_neon + +/* + * ---------------------------------------------------------------------------- + * hash4_neon + * ---------------------------------------------------------------------------- + */ + +INLINE void round_fn4(uint32x4_t v[16], uint32x4_t m[16], size_t r) { + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[15] = rot16_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot12_128(v[4]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = add_128(v[0], v[4]); + v[1] = add_128(v[1], v[5]); + v[2] = add_128(v[2], v[6]); + v[3] = add_128(v[3], v[7]); + v[12] = xor_128(v[12], v[0]); + v[13] = xor_128(v[13], v[1]); + v[14] = xor_128(v[14], v[2]); + v[15] = xor_128(v[15], v[3]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[15] = rot8_128(v[15]); + v[8] = add_128(v[8], v[12]); + v[9] = add_128(v[9], v[13]); + v[10] = add_128(v[10], v[14]); + v[11] = add_128(v[11], v[15]); + v[4] = xor_128(v[4], v[8]); + v[5] = xor_128(v[5], v[9]); + v[6] = xor_128(v[6], v[10]); + v[7] = xor_128(v[7], v[11]); + v[4] = rot7_128(v[4]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot16_128(v[15]); + v[12] = rot16_128(v[12]); + v[13] = rot16_128(v[13]); + v[14] = rot16_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot12_128(v[5]); + v[6] = rot12_128(v[6]); + v[7] = rot12_128(v[7]); + v[4] = rot12_128(v[4]); + v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = add_128(v[0], v[5]); + v[1] = add_128(v[1], v[6]); + v[2] = add_128(v[2], v[7]); + v[3] = add_128(v[3], v[4]); + v[15] = xor_128(v[15], v[0]); + v[12] = xor_128(v[12], v[1]); + v[13] = xor_128(v[13], v[2]); + v[14] = xor_128(v[14], v[3]); + v[15] = rot8_128(v[15]); + v[12] = rot8_128(v[12]); + v[13] = rot8_128(v[13]); + v[14] = rot8_128(v[14]); + v[10] = add_128(v[10], v[15]); + v[11] = add_128(v[11], v[12]); + v[8] = add_128(v[8], v[13]); + v[9] = add_128(v[9], v[14]); + v[5] = xor_128(v[5], v[10]); + v[6] = xor_128(v[6], v[11]); + v[7] = xor_128(v[7], v[8]); + v[4] = xor_128(v[4], v[9]); + v[5] = rot7_128(v[5]); + v[6] = rot7_128(v[6]); + v[7] = rot7_128(v[7]); + v[4] = rot7_128(v[4]); +} + +INLINE void transpose_vecs_128(uint32x4_t vecs[4]) { + // Individually transpose the four 2x2 sub-matrices in each corner. + uint32x4x2_t rows01 = vtrnq_u32(vecs[0], vecs[1]); + uint32x4x2_t rows23 = vtrnq_u32(vecs[2], vecs[3]); + + // Swap the top-right and bottom-left 2x2s (which just got transposed). + vecs[0] = + vcombine_u32(vget_low_u32(rows01.val[0]), vget_low_u32(rows23.val[0])); + vecs[1] = + vcombine_u32(vget_low_u32(rows01.val[1]), vget_low_u32(rows23.val[1])); + vecs[2] = + vcombine_u32(vget_high_u32(rows01.val[0]), vget_high_u32(rows23.val[0])); + vecs[3] = + vcombine_u32(vget_high_u32(rows01.val[1]), vget_high_u32(rows23.val[1])); +} + +INLINE void transpose_msg_vecs4(const uint8_t *const *inputs, + size_t block_offset, uint32x4_t out[16]) { + out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]); + out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]); + out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]); + out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]); + out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]); + out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]); + out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]); + out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]); + out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]); + out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(uint32x4_t)]); + out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(uint32x4_t)]); + out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(uint32x4_t)]); + out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(uint32x4_t)]); + out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(uint32x4_t)]); + out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(uint32x4_t)]); + out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(uint32x4_t)]); + transpose_vecs_128(&out[0]); + transpose_vecs_128(&out[4]); + transpose_vecs_128(&out[8]); + transpose_vecs_128(&out[12]); +} + +INLINE void load_counters4(uint64_t counter, bool increment_counter, + uint32x4_t *out_low, uint32x4_t *out_high) { + uint64_t mask = (increment_counter ? ~0 : 0); + *out_low = set4( + counter_low(counter + (mask & 0)), counter_low(counter + (mask & 1)), + counter_low(counter + (mask & 2)), counter_low(counter + (mask & 3))); + *out_high = set4( + counter_high(counter + (mask & 0)), counter_high(counter + (mask & 1)), + counter_high(counter + (mask & 2)), counter_high(counter + (mask & 3))); +} + +static void 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) { + uint32x4_t h_vecs[8] = { + set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]), + set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]), + }; + uint32x4_t counter_low_vec, counter_high_vec; + load_counters4(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + uint32x4_t block_len_vec = set1_128(BLAKE3_BLOCK_LEN); + uint32x4_t block_flags_vec = set1_128(block_flags); + uint32x4_t msg_vecs[16]; + transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + uint32x4_t v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1_128(IV[0]), set1_128(IV[1]), set1_128(IV[2]), set1_128(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn4(v, msg_vecs, 0); + round_fn4(v, msg_vecs, 1); + round_fn4(v, msg_vecs, 2); + round_fn4(v, msg_vecs, 3); + round_fn4(v, msg_vecs, 4); + round_fn4(v, msg_vecs, 5); + round_fn4(v, msg_vecs, 6); + h_vecs[0] = xor_128(v[0], v[8]); + h_vecs[1] = xor_128(v[1], v[9]); + h_vecs[2] = xor_128(v[2], v[10]); + h_vecs[3] = xor_128(v[3], v[11]); + h_vecs[4] = xor_128(v[4], v[12]); + h_vecs[5] = xor_128(v[5], v[13]); + h_vecs[6] = xor_128(v[6], v[14]); + h_vecs[7] = xor_128(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs_128(&h_vecs[0]); + transpose_vecs_128(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu_128(h_vecs[0], &out[0 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[4], &out[1 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[1], &out[2 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[5], &out[3 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[2], &out[4 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[6], &out[5 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[3], &out[6 * sizeof(uint32x4_t)]); + storeu_128(h_vecs[7], &out[7 * sizeof(uint32x4_t)]); +} + +/* + * ---------------------------------------------------------------------------- + * hash_many_neon + * ---------------------------------------------------------------------------- + */ + +void blake3_compress_in_place_portable(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags); + +INLINE void 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]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + // TODO: Implement compress_neon. However note that according to + // https://github.com/BLAKE2/BLAKE2/commit/7965d3e6e1b4193438b8d3a656787587d2579227, + // compress_neon might not be any faster than compress_portable. + blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void 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) { + while (num_inputs >= 4) { + blake3_hash4_neon(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += 4; + } + inputs += 4; + num_inputs -= 4; + out = &out[4 * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_neon(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} diff --git a/include/blake3/blake3_sse2.c b/include/blake3/blake3_sse2.c new file mode 100644 index 0000000..691e1c6 --- /dev/null +++ b/include/blake3/blake3_sse2.c @@ -0,0 +1,566 @@ +#include "blake3_impl.h" + +#include + +#define DEGREE 4 + +#define _mm_shuffle_ps2(a, b, c) \ + (_mm_castps_si128( \ + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c)))) + +INLINE __m128i loadu(const uint8_t src[16]) { + return _mm_loadu_si128((const __m128i *)src); +} + +INLINE void storeu(__m128i src, uint8_t dest[16]) { + _mm_storeu_si128((__m128i *)dest, src); +} + +INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); } + +// Note that clang-format doesn't like the name "xor" for some reason. +INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); } + +INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); } + +INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d); +} + +INLINE __m128i rot16(__m128i x) { + return _mm_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1); +} + +INLINE __m128i rot12(__m128i x) { + return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12)); +} + +INLINE __m128i rot8(__m128i x) { + return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8)); +} + +INLINE __m128i rot7(__m128i x) { + return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7)); +} + +INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot16(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot12(*row1); +} + +INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot8(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot7(*row1); +} + +// Note the optimization here of leaving row1 as the unrotated row, rather than +// row0. All the message loads below are adjusted to compensate for this. See +// discussion at https://github.com/sneves/blake2-avx2/pull/4 +INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1)); +} + +INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3)); +} + +INLINE __m128i blend_epi16(__m128i a, __m128i b, const int16_t imm8) { + const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01); + __m128i mask = _mm_set1_epi16(imm8); + mask = _mm_and_si128(mask, bits); + mask = _mm_cmpeq_epi16(mask, bits); + return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a)); +} + +INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + rows[0] = loadu((uint8_t *)&cv[0]); + rows[1] = loadu((uint8_t *)&cv[4]); + rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); + rows[3] = set4(counter_low(counter), counter_high(counter), + (uint32_t)block_len, (uint32_t)flags); + + __m128i m0 = loadu(&block[sizeof(__m128i) * 0]); + __m128i m1 = loadu(&block[sizeof(__m128i) * 1]); + __m128i m2 = loadu(&block[sizeof(__m128i) * 2]); + __m128i m3 = loadu(&block[sizeof(__m128i) * 3]); + + __m128i t0, t1, t2, t3, tt; + + // Round 1. The first round permutes the message words from the original + // input order, into the groups that get mixed in parallel. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8 + t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9 + t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 2. This round and all following rounds apply a fixed permutation + // to the message words from the round before. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 3 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 4 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 5 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 6 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 7 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); +} + +void blake3_compress_in_place_sse2(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]); + storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]); +} + +void blake3_compress_xof_sse2(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), &out[0]); + storeu(xorv(rows[1], rows[3]), &out[16]); + storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]); + storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]); +} + +INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) { + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[15] = rot16(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot12(v[4]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[15] = rot8(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot7(v[4]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot16(v[15]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[4] = rot12(v[4]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot8(v[15]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + v[4] = rot7(v[4]); +} + +INLINE void transpose_vecs(__m128i vecs[DEGREE]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is + // 22/33. Note that this doesn't split the vector into two lanes, as the + // AVX2 counterparts do. + __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + + // Interleave 64-bit lanes. + __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01); + __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01); + __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23); + __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23); + + vecs[0] = abcd_0; + vecs[1] = abcd_1; + vecs[2] = abcd_2; + vecs[3] = abcd_3; +} + +INLINE void transpose_msg_vecs(const uint8_t *const *inputs, + size_t block_offset, __m128i out[16]) { + out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]); + out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]); + out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]); + out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]); + out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]); + out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]); + out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]); + out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]); + out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]); + out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]); + out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]); + out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]); + out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]); + out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]); + out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]); + out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]); + for (size_t i = 0; i < 4; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs(&out[0]); + transpose_vecs(&out[4]); + transpose_vecs(&out[8]); + transpose_vecs(&out[12]); +} + +INLINE void load_counters(uint64_t counter, bool increment_counter, + __m128i *out_lo, __m128i *out_hi) { + const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter); + const __m128i add0 = _mm_set_epi32(3, 2, 1, 0); + const __m128i add1 = _mm_and_si128(mask, add0); + __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1); + __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), + _mm_xor_si128( l, _mm_set1_epi32(0x80000000))); + __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry); + *out_lo = l; + *out_hi = h; +} + +static +void blake3_hash4_sse2(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) { + __m128i h_vecs[8] = { + set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]), + set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]), + }; + __m128i counter_low_vec, counter_high_vec; + load_counters(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN); + __m128i block_flags_vec = set1(block_flags); + __m128i msg_vecs[16]; + transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn(v, msg_vecs, 0); + round_fn(v, msg_vecs, 1); + round_fn(v, msg_vecs, 2); + round_fn(v, msg_vecs, 3); + round_fn(v, msg_vecs, 4); + round_fn(v, msg_vecs, 5); + round_fn(v, msg_vecs, 6); + h_vecs[0] = xorv(v[0], v[8]); + h_vecs[1] = xorv(v[1], v[9]); + h_vecs[2] = xorv(v[2], v[10]); + h_vecs[3] = xorv(v[3], v[11]); + h_vecs[4] = xorv(v[4], v[12]); + h_vecs[5] = xorv(v[5], v[13]); + h_vecs[6] = xorv(v[6], v[14]); + h_vecs[7] = xorv(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs(&h_vecs[0]); + transpose_vecs(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu(h_vecs[0], &out[0 * sizeof(__m128i)]); + storeu(h_vecs[4], &out[1 * sizeof(__m128i)]); + storeu(h_vecs[1], &out[2 * sizeof(__m128i)]); + storeu(h_vecs[5], &out[3 * sizeof(__m128i)]); + storeu(h_vecs[2], &out[4 * sizeof(__m128i)]); + storeu(h_vecs[6], &out[5 * sizeof(__m128i)]); + storeu(h_vecs[3], &out[6 * sizeof(__m128i)]); + storeu(h_vecs[7], &out[7 * sizeof(__m128i)]); +} + +INLINE void hash_one_sse2(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]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_sse2(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_sse2(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) { + while (num_inputs >= DEGREE) { + blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += DEGREE; + } + inputs += DEGREE; + num_inputs -= DEGREE; + out = &out[DEGREE * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_sse2(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} diff --git a/include/blake3/blake3_sse41.c b/include/blake3/blake3_sse41.c new file mode 100644 index 0000000..4653a85 --- /dev/null +++ b/include/blake3/blake3_sse41.c @@ -0,0 +1,560 @@ +#include "blake3_impl.h" + +#include + +#define DEGREE 4 + +#define _mm_shuffle_ps2(a, b, c) \ + (_mm_castps_si128( \ + _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c)))) + +INLINE __m128i loadu(const uint8_t src[16]) { + return _mm_loadu_si128((const __m128i *)src); +} + +INLINE void storeu(__m128i src, uint8_t dest[16]) { + _mm_storeu_si128((__m128i *)dest, src); +} + +INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); } + +// Note that clang-format doesn't like the name "xor" for some reason. +INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); } + +INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); } + +INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) { + return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d); +} + +INLINE __m128i rot16(__m128i x) { + return _mm_shuffle_epi8( + x, _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2)); +} + +INLINE __m128i rot12(__m128i x) { + return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12)); +} + +INLINE __m128i rot8(__m128i x) { + return _mm_shuffle_epi8( + x, _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1)); +} + +INLINE __m128i rot7(__m128i x) { + return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7)); +} + +INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot16(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot12(*row1); +} + +INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3, + __m128i m) { + *row0 = addv(addv(*row0, m), *row1); + *row3 = xorv(*row3, *row0); + *row3 = rot8(*row3); + *row2 = addv(*row2, *row3); + *row1 = xorv(*row1, *row2); + *row1 = rot7(*row1); +} + +// Note the optimization here of leaving row1 as the unrotated row, rather than +// row0. All the message loads below are adjusted to compensate for this. See +// discussion at https://github.com/sneves/blake2-avx2/pull/4 +INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1)); +} + +INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) { + *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1)); + *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2)); + *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3)); +} + +INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, uint8_t flags) { + rows[0] = loadu((uint8_t *)&cv[0]); + rows[1] = loadu((uint8_t *)&cv[4]); + rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); + rows[3] = set4(counter_low(counter), counter_high(counter), + (uint32_t)block_len, (uint32_t)flags); + + __m128i m0 = loadu(&block[sizeof(__m128i) * 0]); + __m128i m1 = loadu(&block[sizeof(__m128i) * 1]); + __m128i m2 = loadu(&block[sizeof(__m128i) * 2]); + __m128i m3 = loadu(&block[sizeof(__m128i) * 3]); + + __m128i t0, t1, t2, t3, tt; + + // Round 1. The first round permutes the message words from the original + // input order, into the groups that get mixed in parallel. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8 + t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14 + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9 + t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15 + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 2. This round and all following rounds apply a fixed permutation + // to the message words from the round before. + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 3 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 4 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 5 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 6 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); + m0 = t0; + m1 = t1; + m2 = t2; + m3 = t3; + + // Round 7 + t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2)); + t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); + t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2)); + tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3)); + t1 = _mm_blend_epi16(tt, t1, 0xCC); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); + diagonalize(&rows[0], &rows[2], &rows[3]); + t2 = _mm_unpacklo_epi64(m3, m1); + tt = _mm_blend_epi16(t2, m2, 0xC0); + t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0)); + g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); + t3 = _mm_unpackhi_epi32(m1, m3); + tt = _mm_unpacklo_epi32(m2, t3); + t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2)); + g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); + undiagonalize(&rows[0], &rows[2], &rows[3]); +} + +void blake3_compress_in_place_sse41(uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]); + storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]); +} + +void blake3_compress_xof_sse41(const uint32_t cv[8], + const uint8_t block[BLAKE3_BLOCK_LEN], + uint8_t block_len, uint64_t counter, + uint8_t flags, uint8_t out[64]) { + __m128i rows[4]; + compress_pre(rows, cv, block, block_len, counter, flags); + storeu(xorv(rows[0], rows[2]), &out[0]); + storeu(xorv(rows[1], rows[3]), &out[16]); + storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]); + storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]); +} + +INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) { + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[15] = rot16(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot12(v[4]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); + v[0] = addv(v[0], v[4]); + v[1] = addv(v[1], v[5]); + v[2] = addv(v[2], v[6]); + v[3] = addv(v[3], v[7]); + v[12] = xorv(v[12], v[0]); + v[13] = xorv(v[13], v[1]); + v[14] = xorv(v[14], v[2]); + v[15] = xorv(v[15], v[3]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[15] = rot8(v[15]); + v[8] = addv(v[8], v[12]); + v[9] = addv(v[9], v[13]); + v[10] = addv(v[10], v[14]); + v[11] = addv(v[11], v[15]); + v[4] = xorv(v[4], v[8]); + v[5] = xorv(v[5], v[9]); + v[6] = xorv(v[6], v[10]); + v[7] = xorv(v[7], v[11]); + v[4] = rot7(v[4]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot16(v[15]); + v[12] = rot16(v[12]); + v[13] = rot16(v[13]); + v[14] = rot16(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot12(v[5]); + v[6] = rot12(v[6]); + v[7] = rot12(v[7]); + v[4] = rot12(v[4]); + v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); + v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); + v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); + v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); + v[0] = addv(v[0], v[5]); + v[1] = addv(v[1], v[6]); + v[2] = addv(v[2], v[7]); + v[3] = addv(v[3], v[4]); + v[15] = xorv(v[15], v[0]); + v[12] = xorv(v[12], v[1]); + v[13] = xorv(v[13], v[2]); + v[14] = xorv(v[14], v[3]); + v[15] = rot8(v[15]); + v[12] = rot8(v[12]); + v[13] = rot8(v[13]); + v[14] = rot8(v[14]); + v[10] = addv(v[10], v[15]); + v[11] = addv(v[11], v[12]); + v[8] = addv(v[8], v[13]); + v[9] = addv(v[9], v[14]); + v[5] = xorv(v[5], v[10]); + v[6] = xorv(v[6], v[11]); + v[7] = xorv(v[7], v[8]); + v[4] = xorv(v[4], v[9]); + v[5] = rot7(v[5]); + v[6] = rot7(v[6]); + v[7] = rot7(v[7]); + v[4] = rot7(v[4]); +} + +INLINE void transpose_vecs(__m128i vecs[DEGREE]) { + // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is + // 22/33. Note that this doesn't split the vector into two lanes, as the + // AVX2 counterparts do. + __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + + // Interleave 64-bit lanes. + __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01); + __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01); + __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23); + __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23); + + vecs[0] = abcd_0; + vecs[1] = abcd_1; + vecs[2] = abcd_2; + vecs[3] = abcd_3; +} + +INLINE void transpose_msg_vecs(const uint8_t *const *inputs, + size_t block_offset, __m128i out[16]) { + out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]); + out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]); + out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]); + out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]); + out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]); + out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]); + out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]); + out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]); + out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]); + out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]); + out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]); + out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]); + out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]); + out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]); + out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]); + out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]); + for (size_t i = 0; i < 4; ++i) { + _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0); + } + transpose_vecs(&out[0]); + transpose_vecs(&out[4]); + transpose_vecs(&out[8]); + transpose_vecs(&out[12]); +} + +INLINE void load_counters(uint64_t counter, bool increment_counter, + __m128i *out_lo, __m128i *out_hi) { + const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter); + const __m128i add0 = _mm_set_epi32(3, 2, 1, 0); + const __m128i add1 = _mm_and_si128(mask, add0); + __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1); + __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), + _mm_xor_si128( l, _mm_set1_epi32(0x80000000))); + __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry); + *out_lo = l; + *out_hi = h; +} + +static +void blake3_hash4_sse41(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) { + __m128i h_vecs[8] = { + set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]), + set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]), + }; + __m128i counter_low_vec, counter_high_vec; + load_counters(counter, increment_counter, &counter_low_vec, + &counter_high_vec); + uint8_t block_flags = flags | flags_start; + + for (size_t block = 0; block < blocks; block++) { + if (block + 1 == blocks) { + block_flags |= flags_end; + } + __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN); + __m128i block_flags_vec = set1(block_flags); + __m128i msg_vecs[16]; + transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs); + + __m128i v[16] = { + h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3], + h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7], + set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]), + counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec, + }; + round_fn(v, msg_vecs, 0); + round_fn(v, msg_vecs, 1); + round_fn(v, msg_vecs, 2); + round_fn(v, msg_vecs, 3); + round_fn(v, msg_vecs, 4); + round_fn(v, msg_vecs, 5); + round_fn(v, msg_vecs, 6); + h_vecs[0] = xorv(v[0], v[8]); + h_vecs[1] = xorv(v[1], v[9]); + h_vecs[2] = xorv(v[2], v[10]); + h_vecs[3] = xorv(v[3], v[11]); + h_vecs[4] = xorv(v[4], v[12]); + h_vecs[5] = xorv(v[5], v[13]); + h_vecs[6] = xorv(v[6], v[14]); + h_vecs[7] = xorv(v[7], v[15]); + + block_flags = flags; + } + + transpose_vecs(&h_vecs[0]); + transpose_vecs(&h_vecs[4]); + // The first four vecs now contain the first half of each output, and the + // second four vecs contain the second half of each output. + storeu(h_vecs[0], &out[0 * sizeof(__m128i)]); + storeu(h_vecs[4], &out[1 * sizeof(__m128i)]); + storeu(h_vecs[1], &out[2 * sizeof(__m128i)]); + storeu(h_vecs[5], &out[3 * sizeof(__m128i)]); + storeu(h_vecs[2], &out[4 * sizeof(__m128i)]); + storeu(h_vecs[6], &out[5 * sizeof(__m128i)]); + storeu(h_vecs[3], &out[6 * sizeof(__m128i)]); + storeu(h_vecs[7], &out[7 * sizeof(__m128i)]); +} + +INLINE void hash_one_sse41(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]) { + uint32_t cv[8]; + memcpy(cv, key, BLAKE3_KEY_LEN); + uint8_t block_flags = flags | flags_start; + while (blocks > 0) { + if (blocks == 1) { + block_flags |= flags_end; + } + blake3_compress_in_place_sse41(cv, input, BLAKE3_BLOCK_LEN, counter, + block_flags); + input = &input[BLAKE3_BLOCK_LEN]; + blocks -= 1; + block_flags = flags; + } + memcpy(out, cv, BLAKE3_OUT_LEN); +} + +void blake3_hash_many_sse41(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) { + while (num_inputs >= DEGREE) { + blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags, + flags_start, flags_end, out); + if (increment_counter) { + counter += DEGREE; + } + inputs += DEGREE; + num_inputs -= DEGREE; + out = &out[DEGREE * BLAKE3_OUT_LEN]; + } + while (num_inputs > 0) { + hash_one_sse41(inputs[0], blocks, key, counter, flags, flags_start, + flags_end, out); + if (increment_counter) { + counter += 1; + } + inputs += 1; + num_inputs -= 1; + out = &out[BLAKE3_OUT_LEN]; + } +} From f28990debe164fed185d26a1257747454c3aeb71 Mon Sep 17 00:00:00 2001 From: Zhihao Zhang Date: Sat, 23 May 2026 08:30:57 +0000 Subject: [PATCH 2/5] build: enable runtime-dispatched BLAKE3 SIMD in cas_core Adds AGENTVFS_BLAKE3_SIMD CMake option (default ON), wires the vendored SIMD C sources into cas_core with per-file -msse/-mavx flags scoped to those translation units only. Drops the BLAKE3_NO_AVX512 / BLAKE3_NO_AVX2 / BLAKE3_NO_SSE41 / BLAKE3_NO_SSE2 defines and the forced BLAKE3_USE_NEON=0. Adds cas_test_blake3_simd which verifies (a) BLAKE3 of the empty input matches the canonical spec vector and (b) the per-arch SIMD entry point is linked into cas_core. Gates the test target on AGENTVFS_BLAKE3_SIMD=ON so the rollback build is clean. Co-Authored-By: Claude Opus 4.7 --- CMakeLists.txt | 63 +++++++++++++++-- tests/cas/test_blake3_simd.cpp | 120 +++++++++++++++++++++++++++++++++ 2 files changed, 177 insertions(+), 6 deletions(-) create mode 100644 tests/cas/test_blake3_simd.cpp diff --git a/CMakeLists.txt b/CMakeLists.txt index 1b79dfe..15abc31 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -148,17 +148,63 @@ set(CAS_CORE_PORTABLE_SOURCES set(CAS_CORE_POSIX_SOURCES src/cas/control_socket.cpp ) + +option(AGENTVFS_BLAKE3_SIMD "Enable runtime-dispatched BLAKE3 SIMD paths" ON) + add_library(cas_core STATIC ${CAS_CORE_PORTABLE_SOURCES}) if(UNIX) target_sources(cas_core PRIVATE ${CAS_CORE_POSIX_SOURCES}) endif() target_include_directories(cas_core PUBLIC src/cas include include/blake3) -target_compile_definitions(cas_core PRIVATE - BLAKE3_NO_AVX512 BLAKE3_NO_AVX2 BLAKE3_NO_SSE41 BLAKE3_NO_SSE2 - # On aarch64 the vendored blake3 auto-enables NEON, but we only - # compile blake3_portable.c — link would fail with an unresolved - # blake3_hash_many_neon. Force the portable path on every arch. - BLAKE3_USE_NEON=0) + +if(AGENTVFS_BLAKE3_SIMD) + set(_blake3_amd64_names amd64 AMD64 x86_64) + set(_blake3_arm64_names aarch64 AArch64 arm64 ARM64 armv8 armv8a) + + if(CMAKE_SYSTEM_PROCESSOR IN_LIST _blake3_amd64_names) + target_sources(cas_core PRIVATE + include/blake3/blake3_sse2.c + include/blake3/blake3_sse41.c + include/blake3/blake3_avx2.c + include/blake3/blake3_avx512.c) + if(MSVC) + # MSVC has no dedicated /arch:SSE4.1 — /arch:AVX is the + # narrowest flag that turns on SSE4.1 codegen. + set_source_files_properties(include/blake3/blake3_sse2.c + PROPERTIES COMPILE_OPTIONS "/arch:SSE2") + set_source_files_properties(include/blake3/blake3_sse41.c + PROPERTIES COMPILE_OPTIONS "/arch:AVX") + set_source_files_properties(include/blake3/blake3_avx2.c + PROPERTIES COMPILE_OPTIONS "/arch:AVX2") + set_source_files_properties(include/blake3/blake3_avx512.c + PROPERTIES COMPILE_OPTIONS "/arch:AVX512") + else() + set_source_files_properties(include/blake3/blake3_sse2.c + PROPERTIES COMPILE_OPTIONS "-msse2") + set_source_files_properties(include/blake3/blake3_sse41.c + PROPERTIES COMPILE_OPTIONS "-msse4.1") + set_source_files_properties(include/blake3/blake3_avx2.c + PROPERTIES COMPILE_OPTIONS "-mavx2") + set_source_files_properties(include/blake3/blake3_avx512.c + PROPERTIES COMPILE_OPTIONS "-mavx512f;-mavx512vl") + endif() + elseif(CMAKE_SYSTEM_PROCESSOR IN_LIST _blake3_arm64_names) + target_sources(cas_core PRIVATE include/blake3/blake3_neon.c) + # 64-bit aarch64 has NEON unconditionally; no -mfpu flag needed. + # blake3_dispatch.c picks NEON automatically when BLAKE3_USE_NEON + # is unset (the upstream default). + else() + # Unknown arch — fall back to portable, same as + # AGENTVFS_BLAKE3_SIMD=OFF. + target_compile_definitions(cas_core PRIVATE + BLAKE3_NO_AVX512 BLAKE3_NO_AVX2 BLAKE3_NO_SSE41 BLAKE3_NO_SSE2 + BLAKE3_USE_NEON=0) + endif() +else() + target_compile_definitions(cas_core PRIVATE + BLAKE3_NO_AVX512 BLAKE3_NO_AVX2 BLAKE3_NO_SSE41 BLAKE3_NO_SSE2 + BLAKE3_USE_NEON=0) +endif() if(NOT MSVC) target_compile_options(cas_core PRIVATE -Wall -Wextra -Wpedantic) endif() @@ -523,6 +569,11 @@ target_include_directories(cas_test_telemetry_registry PRIVATE include src src/c target_compile_features(cas_test_telemetry_registry PRIVATE cxx_std_17) endif() +if(AGENTVFS_BLAKE3_SIMD) +add_executable(cas_test_blake3_simd tests/cas/test_blake3_simd.cpp) +target_link_libraries(cas_test_blake3_simd PRIVATE cas_core) +endif() + if(LINUX) add_executable(cas_test_ebpf_backend tests/cas/test_ebpf_backend.cpp diff --git a/tests/cas/test_blake3_simd.cpp b/tests/cas/test_blake3_simd.cpp new file mode 100644 index 0000000..bd52d88 --- /dev/null +++ b/tests/cas/test_blake3_simd.cpp @@ -0,0 +1,120 @@ +// Verifies that hardware-accelerated BLAKE3 is wired into cas_core. +// +// Two assertions: +// +// (1) Known-answer test: hashing the empty input yields the BLAKE3 +// canonical empty-string digest published in the upstream spec. +// This is correctness coverage that runs on every platform. +// +// (2) SIMD-presence test: takes the address of the per-arch SIMD entry +// point declared by blake3_impl.h. If the entry point's translation +// unit (blake3_avx2.c / blake3_neon.c / ...) is not compiled into +// cas_core, the link step fails with an unresolved-symbol error, +// which is exactly the regression we want to catch. + +#include +#include +#include +#include +#include +#include + +#include "blake3.h" +// Intentionally NOT including blake3_impl.h — its SIMD function +// declarations are gated by BLAKE3_NO_* defines that cas_core may or +// may not set. Declaring the SIMD entry point ourselves via extern "C" +// below keeps the test source decoupled from cas_core's build options. + +namespace { + +// Canonical BLAKE3 empty-input hash from the upstream spec / +// https://github.com/BLAKE3-team/BLAKE3/blob/master/test_vectors/test_vectors.json +// — input "", hash = +constexpr std::array kEmptyInputHash = { + 0xaf, 0x13, 0x49, 0xb9, 0xf5, 0xf9, 0xa1, 0xa6, + 0xa0, 0x40, 0x4d, 0xea, 0x36, 0xdc, 0xc9, 0x49, + 0x9b, 0xcb, 0x25, 0xc9, 0xad, 0xc1, 0x12, 0xb7, + 0xcc, 0x9a, 0x93, 0xca, 0xe4, 0x1f, 0x32, 0x62, +}; + +void test_empty_input_known_answer() { + blake3_hasher h; + blake3_hasher_init(&h); + std::array got{}; + blake3_hasher_finalize(&h, got.data(), got.size()); + if (got != kEmptyInputHash) { + std::fprintf(stderr, "empty-input BLAKE3 mismatch\n"); + std::fprintf(stderr, "expected: "); + for (auto b : kEmptyInputHash) std::fprintf(stderr, "%02x", b); + std::fprintf(stderr, "\n got: "); + for (auto b : got) std::fprintf(stderr, "%02x", b); + std::fprintf(stderr, "\n"); + std::abort(); + } +} + +// SIMD entry-point presence check. We take the address of the per-arch +// SIMD entry point and write it through a volatile pointer so the +// compiler cannot optimize the reference away. If the translation unit +// (blake3_avx2.c / blake3_neon.c / ...) is not compiled into cas_core, +// the link step fails with an unresolved-symbol error — exactly the +// regression we want to catch. +// +// Skipped on architectures other than x86_64 / aarch64 — we only enable +// SIMD on those (see CMakeLists.txt). +static volatile void* g_simd_entry_volatile = nullptr; + +#if (defined(__x86_64__) || defined(_M_X64)) && !defined(_M_ARM64EC) +extern "C" void blake3_hash_many_avx2( + 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); +// The volatile store forces the compiler to emit an actual reference to +// blake3_hash_many_avx2 that the linker must resolve. +static void* compute_simd_entry() { + void* p = reinterpret_cast(&blake3_hash_many_avx2); + g_simd_entry_volatile = p; // volatile store: not optimizable + return p; +} +static void* const simd_entry_point = compute_simd_entry(); +constexpr const char* simd_entry_name = "blake3_hash_many_avx2"; +constexpr bool simd_expected = true; +#elif defined(__aarch64__) || defined(_M_ARM64) +extern "C" void 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); +static void* compute_simd_entry() { + void* p = reinterpret_cast(&blake3_hash_many_neon); + g_simd_entry_volatile = p; // volatile store: not optimizable + return p; +} +static void* const simd_entry_point = compute_simd_entry(); +constexpr const char* simd_entry_name = "blake3_hash_many_neon"; +constexpr bool simd_expected = true; +#else +static void* const simd_entry_point = nullptr; +constexpr const char* simd_entry_name = "n/a"; +constexpr bool simd_expected = false; +#endif + +void test_simd_entry_point_present() { + if (simd_expected && simd_entry_point == nullptr) { + std::fprintf(stderr, "expected SIMD entry point %s but got null\n", + simd_entry_name); + std::abort(); + } + if (!simd_expected) { + std::fprintf(stderr, + "[skip] SIMD presence check (architecture has no SIMD path)\n"); + } +} + +} // namespace + +int main() { + test_empty_input_known_answer(); + test_simd_entry_point_present(); + std::puts("cas_test_blake3_simd: ok"); + return 0; +} From dc8ba3c531223de355b50e405e51546997ca4103 Mon Sep 17 00:00:00 2001 From: Zhihao Zhang Date: Sat, 23 May 2026 08:37:43 +0000 Subject: [PATCH 3/5] ci: run cas_test_blake3_simd on linux / macos / windows Co-Authored-By: Claude Opus 4.7 --- .github/workflows/ci.yml | 3 +++ 1 file changed, 3 insertions(+) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 94ccc77..07ba057 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -30,6 +30,7 @@ jobs: ./build/cas_test_fh_lifecycle /tmp /tmp/test.sock || true ./build/cas_test_branch_merge_daemon || true ./build/cas_test_branch_persistence || true + ./build/cas_test_blake3_simd ./build/cas_test_telemetry_event ./build/cas_test_telemetry_registry @@ -56,6 +57,7 @@ jobs: ./build/cas_test_branch_context ./build/cas_test_branch_merge ./build/cas_test_branch_merge_commit + ./build/cas_test_blake3_simd ./build/cas_test_telemetry_event ./build/cas_test_telemetry_registry @@ -83,6 +85,7 @@ jobs: .\build\Release\cas_test_write_buffer.exe .\build\Release\cas_test_branch_context.exe .\build\Release\cas_test_branch_merge.exe + .\build\Release\cas_test_blake3_simd.exe .\build\Release\cas_test_telemetry_event.exe windows-daemon: From 1859ef9e2be25b67eb0c88c4d83c22eb3604f697 Mon Sep 17 00:00:00 2001 From: Zhihao Zhang Date: Sat, 23 May 2026 08:39:28 +0000 Subject: [PATCH 4/5] docs: note MSVC version requirement for BLAKE3 SIMD path Co-Authored-By: Claude Opus 4.7 --- README.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index f85103b..6ee6c31 100644 --- a/README.md +++ b/README.md @@ -103,7 +103,9 @@ cmake --build build -j ``` ```powershell -# Windows — requires WinFsp 2.0+ from https://winfsp.dev +# Windows — requires WinFsp 2.0+ from https://winfsp.dev, +# and MSVC v141 (Visual Studio 2017 15.3) or newer for /arch:AVX512 +# codegen on the BLAKE3 SIMD path (current VS releases all qualify). cmake -B build -DAGENTVFS_EBPF=OFF -DAGENTVFS_WINFSP=ON cmake --build build --config Release -j .\build\Release\agentvfs.exe --source C:\some\dir --mountpoint Z: From e11a7785babb2215f08e60b4565c025f7ac78dac Mon Sep 17 00:00:00 2001 From: Zhihao Zhang Date: Sat, 23 May 2026 08:53:45 +0000 Subject: [PATCH 5/5] fix(blake3-simd): cover _M_ARM64EC and clarify MSVC /arch:SSE2 note Two review fixes against the BLAKE3 SIMD enablement: - Add `_M_ARM64EC` to the test's aarch64 branch in tests/cas/test_blake3_simd.cpp. blake3_impl.h treats ARM64EC as aarch64 and dispatches NEON at runtime, so the test should verify the NEON entry point on that target too. - Expand the MSVC comment in CMakeLists.txt to note that /arch:SSE2 is x86-only and is silently ignored on x64 (where SSE2 is the implicit baseline). Documents that the flag is harmless rather than load-bearing. Also fills in the canonical empty-input hash hex in the test source's KAT comment (was a stray "hash =" left over from an earlier edit). Co-Authored-By: Claude Opus 4.7 (1M context) --- CMakeLists.txt | 4 +++- tests/cas/test_blake3_simd.cpp | 4 ++-- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 15abc31..52fd04e 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -169,7 +169,9 @@ if(AGENTVFS_BLAKE3_SIMD) include/blake3/blake3_avx512.c) if(MSVC) # MSVC has no dedicated /arch:SSE4.1 — /arch:AVX is the - # narrowest flag that turns on SSE4.1 codegen. + # narrowest flag that turns on SSE4.1 codegen. /arch:SSE2 is + # x86-only; on x64 SSE2 is the implicit baseline and MSVC + # silently ignores the flag, which is harmless. set_source_files_properties(include/blake3/blake3_sse2.c PROPERTIES COMPILE_OPTIONS "/arch:SSE2") set_source_files_properties(include/blake3/blake3_sse41.c diff --git a/tests/cas/test_blake3_simd.cpp b/tests/cas/test_blake3_simd.cpp index bd52d88..4a3cbc1 100644 --- a/tests/cas/test_blake3_simd.cpp +++ b/tests/cas/test_blake3_simd.cpp @@ -29,7 +29,7 @@ namespace { // Canonical BLAKE3 empty-input hash from the upstream spec / // https://github.com/BLAKE3-team/BLAKE3/blob/master/test_vectors/test_vectors.json -// — input "", hash = +// — input "", hash = af1349b9f5f9a1a6a0404dea36dcc9499bcb25c9adc112b7cc9a93cae41f3262 constexpr std::array kEmptyInputHash = { 0xaf, 0x13, 0x49, 0xb9, 0xf5, 0xf9, 0xa1, 0xa6, 0xa0, 0x40, 0x4d, 0xea, 0x36, 0xdc, 0xc9, 0x49, @@ -79,7 +79,7 @@ static void* compute_simd_entry() { static void* const simd_entry_point = compute_simd_entry(); constexpr const char* simd_entry_name = "blake3_hash_many_avx2"; constexpr bool simd_expected = true; -#elif defined(__aarch64__) || defined(_M_ARM64) +#elif defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC) extern "C" void 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,