OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

util: bump blake3 from 1.3.3 to 1.5.1, improve armv7 and aarch64 performance

Browse Source 
Steps for uprev:
 - copy files from BLAKE3/c src/util/blake3/
 - edit README
 - `for file in *.asm; do mv "$file" "${file%.asm}.masm"; done`
 - keep
  - blake3.h (no relevant changes), only change BLAKE3_VERSION_STRING
  - blake3_sse2_x86-64_unix.S (no changes)
  - blake3_avx512_x86-64_unix.S (no changes)
  - blake3_sse41_x86-64_unix.S (no changes)

Acked-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: David Heidelberg <david@ixit.cz>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29687>
This commit is contained in:
David Heidelberg
2024-06-11 21:20:07 -07:00
parent 9f1effb03b
commit a0c09eef93
14 changed files with 2762 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/util/blake3/README
View File

@@ -1,7 +1,7 @@
This folder contains a local copy of BLAKE3 cryptographic hash library, version 1.3.3.
This folder contains a local copy of BLAKE3 cryptographic hash library, version 1.5.1.
Except for changes listed in the "Changes" section, this is a verbatim copy from
https://github.com/BLAKE3-team/BLAKE3, tag 1.3.3.
https://github.com/BLAKE3-team/BLAKE3, tag 1.5.1.
Files will be periodically synchronized with the upstream, and any local changes should
be clearly documented below.
@@ -13,4 +13,4 @@ Changes:
- Add "static" to blake3_hash4_neon, to comply with -Werror=missing-prototypes.
- Add mesa_blake3_visibility.h and set symbol visibility to hidden for assembly sources.
- Add mesa_blake3_visibility.h and set symbol visibility to hidden for assembly sources.

21
src/util/blake3/blake3.c
View File

@@ -254,7 +254,7 @@ INLINE size_t compress_parents_parallel(const uint8_t *child_chaining_values,
// As a special case when the SIMD degree is 1, this function will still return
// at least 2 outputs. This guarantees that this function doesn't perform the
// root compression. (If it did, it would use the wrong flags, and also we
// wouldn't be able to implement exendable output.) Note that this function is
// wouldn't be able to implement extendable output.) Note that this function is
// not used when the whole input is only 1 chunk long; that's a different
// codepath.
//
@@ -341,21 +341,24 @@ INLINE void compress_subtree_to_parent_node(
size_t num_cvs = blake3_compress_subtree_wide(input, input_len, key,
chunk_counter, flags, cv_array);
assert(num_cvs <= MAX_SIMD_DEGREE_OR_2);
// If MAX_SIMD_DEGREE is greater than 2 and there's enough input,
// The following loop never executes when MAX_SIMD_DEGREE_OR_2 is 2, because
// as we just asserted, num_cvs will always be <=2 in that case. But GCC
// (particularly GCC 8.5) can't tell that it never executes, and if NDEBUG is
// set then it emits incorrect warnings here. We tried a few different
// hacks to silence these, but in the end our hacks just produced different
// warnings (see https://github.com/BLAKE3-team/BLAKE3/pull/380). Out of
// desperation, we ifdef out this entire loop when we know it's not needed.
#if MAX_SIMD_DEGREE_OR_2 > 2
// If MAX_SIMD_DEGREE_OR_2 is greater than 2 and there's enough input,
// compress_subtree_wide() returns more than 2 chaining values. Condense
// them into 2 by forming parent nodes repeatedly.
uint8_t out_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN / 2];
// The second half of this loop condition is always true, and we just
// asserted it above. But GCC can't tell that it's always true, and if NDEBUG
// is set on platforms where MAX_SIMD_DEGREE_OR_2 == 2, GCC emits spurious
// warnings here. GCC 8.5 is particularly sensitive, so if you're changing
// this code, test it against that version.
while (num_cvs > 2 && num_cvs <= MAX_SIMD_DEGREE_OR_2) {
while (num_cvs > 2) {
num_cvs =
compress_parents_parallel(cv_array, num_cvs, key, flags, out_array);
memcpy(cv_array, out_array, num_cvs * BLAKE3_OUT_LEN);
}
#endif
memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN);
}

2
src/util/blake3/blake3.h
View File

@@ -8,7 +8,7 @@
extern "C" {
#endif
#define BLAKE3_VERSION_STRING "1.3.3"
#define BLAKE3_VERSION_STRING "1.5.1"
#define BLAKE3_KEY_LEN 32
#define BLAKE3_OUT_LEN 32
#define BLAKE3_BLOCK_LEN 64

326
src/util/blake3/blake3_avx2.c Normal file
View File

@@ -0,0 +1,326 @@
#include "blake3_impl.h"
#include <immintrin.h>
#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
}

2
src/util/blake3/blake3_avx2_x86-64_windows_gnu.S
View File

@@ -1784,7 +1784,7 @@ blake3_hash_many_avx2:
vmovdqu xmmword ptr [rbx+0x10], xmm1
jmp 4b
.section .rodata
.section .rdata
.p2align 6
ADD0:
.long 0, 1, 2, 3, 4, 5, 6, 7

1220
src/util/blake3/blake3_avx512.c Normal file
View File

File diff suppressed because it is too large Load Diff

2
src/util/blake3/blake3_avx512_x86-64_windows_gnu.S
View File

@@ -2587,7 +2587,7 @@ blake3_compress_xof_avx512:
add rsp, 72
ret
.section .rodata
.section .rdata
.p2align 6
INDEX0:
.long 0, 1, 2, 3, 16, 17, 18, 19

41
src/util/blake3/blake3_dispatch.c
View File

@@ -6,6 +6,7 @@
#if defined(IS_X86)
#if defined(_MSC_VER)
#include <Windows.h>
#include <intrin.h>
#elif defined(__GNUC__)
#include <immintrin.h>
@@ -14,6 +15,32 @@
#endif
#endif
#if !defined(BLAKE3_ATOMICS)
#if defined(__has_include)
#if __has_include(<stdatomic.h>) && !defined(_MSC_VER)
#define BLAKE3_ATOMICS 1
#else
#define BLAKE3_ATOMICS 0
#endif /* __has_include(<stdatomic.h>) && !defined(_MSC_VER) */
#else
#define BLAKE3_ATOMICS 0
#endif /* defined(__has_include) */
#endif /* BLAKE3_ATOMICS */
#if BLAKE3_ATOMICS
#define ATOMIC_INT _Atomic int
#define ATOMIC_LOAD(x) x
#define ATOMIC_STORE(x, y) x = y
#elif defined(_MSC_VER)
#define ATOMIC_INT LONG
#define ATOMIC_LOAD(x) InterlockedOr(&x, 0)
#define ATOMIC_STORE(x, y) InterlockedExchange(&x, y)
#else
#define ATOMIC_INT int
#define ATOMIC_LOAD(x) x
#define ATOMIC_STORE(x, y) x = y
#endif
#define MAYBE_UNUSED(x) (void)((x))
#if defined(IS_X86)
@@ -76,7 +103,7 @@ enum cpu_feature {
#if !defined(BLAKE3_TESTING)
static /* Allow the variable to be controlled manually for testing */
#endif
enum cpu_feature g_cpu_features = UNDEFINED;
ATOMIC_INT g_cpu_features = UNDEFINED;
#if !defined(BLAKE3_TESTING)
static
@@ -84,14 +111,16 @@ static
enum cpu_feature
get_cpu_features(void) {
if (g_cpu_features != UNDEFINED) {
return g_cpu_features;
/* If TSAN detects a data race here, try compiling with -DBLAKE3_ATOMICS=1 */
enum cpu_feature features = ATOMIC_LOAD(g_cpu_features);
if (features != UNDEFINED) {
return features;
} else {
#if defined(IS_X86)
uint32_t regs[4] = {0};
uint32_t *eax = &regs[0], *ebx = &regs[1], *ecx = &regs[2], *edx = &regs[3];
(void)edx;
enum cpu_feature features = 0;
features = 0;
cpuid(regs, 0);
const int max_id = *eax;
cpuid(regs, 1);
@@ -101,7 +130,7 @@ static
if (*edx & (1UL << 26))
features |= SSE2;
#endif
if (*ecx & (1UL << 0))
if (*ecx & (1UL << 9))
features |= SSSE3;
if (*ecx & (1UL << 19))
features |= SSE41;
@@ -124,7 +153,7 @@ static
}
}
}
g_cpu_features = features;
ATOMIC_STORE(g_cpu_features, features);
return features;
#else
/* How to detect NEON? */

14
src/util/blake3/blake3_impl.h
View File

@@ -28,7 +28,7 @@ enum blake3_flags {
#define INLINE static inline __attribute__((always_inline))
#endif
#if (defined(__x86_64__) || defined(_M_X64)) && !defined(_M_ARM64EC)
#if defined(__x86_64__) || defined(_M_X64)
#define IS_X86
#define IS_X86_64
#endif
@@ -38,7 +38,7 @@ enum blake3_flags {
#define IS_X86_32
#endif
#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
#if defined(__aarch64__) || defined(_M_ARM64)
#define IS_AARCH64
#endif
@@ -51,7 +51,11 @@ enum blake3_flags {
#if !defined(BLAKE3_USE_NEON)
// If BLAKE3_USE_NEON not manually set, autodetect based on AArch64ness
#if defined(IS_AARCH64)
#define BLAKE3_USE_NEON 1
#if defined(__ARM_BIG_ENDIAN)
#define BLAKE3_USE_NEON 0
#else
#define BLAKE3_USE_NEON 1
#endif
#else
#define BLAKE3_USE_NEON 0
#endif
@@ -87,7 +91,7 @@ static const uint8_t MSG_SCHEDULE[7][16] = {
/* x is assumed to be nonzero. */
static unsigned int highest_one(uint64_t x) {
#if defined(__GNUC__) || defined(__clang__)
return 63 ^ __builtin_clzll(x);
return 63 ^ (unsigned int)__builtin_clzll(x);
#elif defined(_MSC_VER) && defined(IS_X86_64)
unsigned long index;
_BitScanReverse64(&index, x);
@@ -117,7 +121,7 @@ static unsigned int highest_one(uint64_t x) {
// Count the number of 1 bits.
INLINE unsigned int popcnt(uint64_t x) {
#if defined(__GNUC__) || defined(__clang__)
return __builtin_popcountll(x);
return (unsigned int)__builtin_popcountll(x);
#else
unsigned int count = 0;
while (x != 0) {

37
src/util/blake3/blake3_neon.c
View File

@@ -10,14 +10,12 @@
INLINE uint32x4_t loadu_128(const uint8_t src[16]) {
// vld1q_u32 has alignment requirements. Don't use it.
uint32x4_t x;
memcpy(&x, src, 16);
return x;
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.
memcpy(dest, &src, 16);
vst1q_u8(dest, vreinterpretq_u8_u32(src));
}
INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) {
@@ -36,19 +34,36 @@ INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
}
INLINE uint32x4_t rot16_128(uint32x4_t x) {
return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16));
// The straightfoward 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) {
return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12));
// 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) {
return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8));
// 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) {
return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7));
// 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
@@ -229,9 +244,9 @@ INLINE void load_counters4(uint64_t counter, bool increment_counter,
}
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) {
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]),

566
src/util/blake3/blake3_sse2.c Normal file
View File

@@ -0,0 +1,566 @@
#include "blake3_impl.h"
#include <immintrin.h>
#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];
}
}

2
src/util/blake3/blake3_sse2_x86-64_windows_gnu.S
View File

@@ -2301,7 +2301,7 @@ blake3_compress_xof_sse2:
ret
.section .rodata
.section .rdata
.p2align 6
BLAKE3_IV:
.long 0x6A09E667, 0xBB67AE85

560
src/util/blake3/blake3_sse41.c Normal file
View File

@@ -0,0 +1,560 @@
#include "blake3_impl.h"
#include <immintrin.h>
#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];
}
}

2
src/util/blake3/blake3_sse41_x86-64_windows_gnu.S
View File

@@ -2042,7 +2042,7 @@ blake3_compress_xof_sse41:
ret
.section .rodata
.section .rdata
.p2align 6
BLAKE3_IV:
.long 0x6A09E667, 0xBB67AE85