dotya.ml/ mkproof
2
0
Fork 0
mirror of https://git.sr.ht/~sircmpwn/mkproof synced 2024-06-09 05:06:06 +02:00
Code Releases Activity

BLAMKA with AVX2 optimizations

This commit is contained in:
luc-lynx 2017-04-25 00:49:11 +03:00
parent 30c78a1e5e
commit 2e67e5e985
3 changed files with 204 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

150
src/blake2/blamka-round-opt.h
View File

@ -29,6 +29,7 @@
#include <x86intrin.h>
#endif
#if !defined(__AVX2__)
#if !defined(__XOP__)
#if defined(__SSSE3__)
#define r16 \
@ -176,5 +177,152 @@ static BLAKE2_INLINE __m128i fBlaMka(__m128i x, __m128i y) {
\
UNDIAGONALIZE(A0, B0, C0, D0, A1, B1, C1, D1); \
} while ((void)0, 0)
#else /* __AVX2__ */
#endif
#include <immintrin.h>
#define rotr32(x) _mm256_permutevar8x32_epi32(x, _mm256_setr_epi32(1, 0, 3, 2, 5, 4, 7, 6))
#define rotr24(x) _mm256_shuffle_epi8(x, _mm256_setr_epi8(3, 4, 5, 6, 7, 0, 1, 2, 11, 12, 13, 14, 15, 8, 9, 10, 3, 4, 5, 6, 7, 0, 1, 2, 11, 12, 13, 14, 15, 8, 9, 10))
#define rotr16(x) _mm256_shuffle_epi8(x, _mm256_setr_epi8(2, 3, 4, 5, 6, 7, 0, 1, 10, 11, 12, 13, 14, 15, 8, 9, 2, 3, 4, 5, 6, 7, 0, 1, 10, 11, 12, 13, 14, 15, 8, 9))
#define rotr63(x) _mm256_xor_si256(_mm256_srli_epi64((x), 63), _mm256_add_epi64((x), (x)))
#define G1_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
do { \
__m256i ml = _mm256_mul_epu32(A0, B0); \
ml = _mm256_add_epi64(ml, ml); \
A0 = _mm256_add_epi64(A0, _mm256_add_epi64(B0, ml)); \
D0 = _mm256_xor_si256(D0, A0); \
D0 = rotr32(D0); \
\
ml = _mm256_mul_epu32(C0, D0); \
ml = _mm256_add_epi64(ml, ml); \
C0 = _mm256_add_epi64(C0, _mm256_add_epi64(D0, ml)); \
\
B0 = _mm256_xor_si256(B0, C0); \
B0 = rotr24(B0); \
\
ml = _mm256_mul_epu32(A1, B1); \
ml = _mm256_add_epi64(ml, ml); \
A1 = _mm256_add_epi64(A1, _mm256_add_epi64(B1, ml)); \
D1 = _mm256_xor_si256(D1, A1); \
D1 = rotr32(D1); \
\
ml = _mm256_mul_epu32(C1, D1); \
ml = _mm256_add_epi64(ml, ml); \
C1 = _mm256_add_epi64(C1, _mm256_add_epi64(D1, ml)); \
\
B1 = _mm256_xor_si256(B1, C1); \
B1 = rotr24(B1); \
} while((void)0, 0);
#define G2_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
do { \
__m256i ml = _mm256_mul_epu32(A0, B0); \
ml = _mm256_add_epi64(ml, ml); \
A0 = _mm256_add_epi64(A0, _mm256_add_epi64(B0, ml)); \
D0 = _mm256_xor_si256(D0, A0); \
D0 = rotr16(D0); \
\
ml = _mm256_mul_epu32(C0, D0); \
ml = _mm256_add_epi64(ml, ml); \
C0 = _mm256_add_epi64(C0, _mm256_add_epi64(D0, ml)); \
B0 = _mm256_xor_si256(B0, C0); \
B0 = rotr63(B0); \
\
ml = _mm256_mul_epu32(A1, B1); \
ml = _mm256_add_epi64(ml, ml); \
A1 = _mm256_add_epi64(A1, _mm256_add_epi64(B1, ml)); \
D1 = _mm256_xor_si256(D1, A1); \
D1 = rotr16(D1); \
\
ml = _mm256_mul_epu32(C1, D1); \
ml = _mm256_add_epi64(ml, ml); \
C1 = _mm256_add_epi64(C1, _mm256_add_epi64(D1, ml)); \
B1 = _mm256_xor_si256(B1, C1); \
B1 = rotr63(B1); \
} while((void)0, 0);
#define DIAGONALIZE_1(A0, B0, C0, D0, A1, B1, C1, D1) \
do { \
B0 = _mm256_permute4x64_epi64(B0, _MM_SHUFFLE(0, 3, 2, 1)); \
C0 = _mm256_permute4x64_epi64(C0, _MM_SHUFFLE(1, 0, 3, 2)); \
D0 = _mm256_permute4x64_epi64(D0, _MM_SHUFFLE(2, 1, 0, 3)); \
\
B1 = _mm256_permute4x64_epi64(B1, _MM_SHUFFLE(0, 3, 2, 1)); \
C1 = _mm256_permute4x64_epi64(C1, _MM_SHUFFLE(1, 0, 3, 2)); \
D1 = _mm256_permute4x64_epi64(D1, _MM_SHUFFLE(2, 1, 0, 3)); \
} while((void)0, 0);
#define DIAGONALIZE_2(A0, A1, B0, B1, C0, C1, D0, D1) \
do { \
__m256i tmp1 = _mm256_blend_epi32(B0, B1, 0xCC); \
__m256i tmp2 = _mm256_blend_epi32(B0, B1, 0x33); \
B1 = _mm256_permute4x64_epi64(tmp1, _MM_SHUFFLE(2,3,0,1)); \
B0 = _mm256_permute4x64_epi64(tmp2, _MM_SHUFFLE(2,3,0,1)); \
\
tmp1 = C0; \
C0 = C1; \
C1 = tmp1; \
\
tmp1 = _mm256_blend_epi32(D0, D1, 0xCC); \
tmp2 = _mm256_blend_epi32(D0, D1, 0x33); \
D0 = _mm256_permute4x64_epi64(tmp1, _MM_SHUFFLE(2,3,0,1)); \
D1 = _mm256_permute4x64_epi64(tmp2, _MM_SHUFFLE(2,3,0,1)); \
} while(0);
#define UNDIAGONALIZE_1(A0, B0, C0, D0, A1, B1, C1, D1) \
do { \
B0 = _mm256_permute4x64_epi64(B0, _MM_SHUFFLE(2, 1, 0, 3)); \
C0 = _mm256_permute4x64_epi64(C0, _MM_SHUFFLE(1, 0, 3, 2)); \
D0 = _mm256_permute4x64_epi64(D0, _MM_SHUFFLE(0, 3, 2, 1)); \
\
B1 = _mm256_permute4x64_epi64(B1, _MM_SHUFFLE(2, 1, 0, 3)); \
C1 = _mm256_permute4x64_epi64(C1, _MM_SHUFFLE(1, 0, 3, 2)); \
D1 = _mm256_permute4x64_epi64(D1, _MM_SHUFFLE(0, 3, 2, 1)); \
} while((void)0, 0);
#define UNDIAGONALIZE_2(A0, A1, B0, B1, C0, C1, D0, D1) \
do { \
__m256i tmp1 = _mm256_blend_epi32(B0, B1, 0xCC); \
__m256i tmp2 = _mm256_blend_epi32(B0, B1, 0x33); \
B0 = _mm256_permute4x64_epi64(tmp1, _MM_SHUFFLE(2,3,0,1)); \
B1 = _mm256_permute4x64_epi64(tmp2, _MM_SHUFFLE(2,3,0,1)); \
\
tmp1 = C0; \
C0 = C1; \
C1 = tmp1; \
\
tmp1 = _mm256_blend_epi32(D0, D1, 0x33); \
tmp2 = _mm256_blend_epi32(D0, D1, 0xCC); \
D0 = _mm256_permute4x64_epi64(tmp1, _MM_SHUFFLE(2,3,0,1)); \
D1 = _mm256_permute4x64_epi64(tmp2, _MM_SHUFFLE(2,3,0,1)); \
} while((void)0, 0);
#define BLAKE2_ROUND_1(A0, A1, B0, B1, C0, C1, D0, D1) \
do{ \
G1_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
G2_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
\
DIAGONALIZE_1(A0, B0, C0, D0, A1, B1, C1, D1) \
\
G1_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
G2_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
\
UNDIAGONALIZE_1(A0, B0, C0, D0, A1, B1, C1, D1) \
} while((void)0, 0);
#define BLAKE2_ROUND_2(A0, A1, B0, B1, C0, C1, D0, D1) \
do{ \
G1_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
G2_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
\
DIAGONALIZE_2(A0, A1, B0, B1, C0, C1, D0, D1) \
\
G1_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
G2_AVX2(A0, A1, B0, B1, C0, C1, D0, D1) \
\
UNDIAGONALIZE_2(A0, A1, B0, B1, C0, C1, D0, D1) \
} while((void)0, 0);
#endif /* __AVX2__ */
#endif /* BLAKE_ROUND_MKA_OPT_H */

9
src/core.h
View File

@ -20,6 +20,14 @@
#include "argon2.h"
#if defined(_MSC_VER)
#define ALIGN(n) __declspec(align(16))
#elif defined(__GNUC__) || defined(__clang)
#define ALIGN(x) __attribute__((__aligned__(x)))
#else
#define ALIGN(x)
#endif
#define CONST_CAST(x) (x)(uintptr_t)
/**********************Argon2 internal constants*******************************/
@ -29,6 +37,7 @@ enum argon2_core_constants {
ARGON2_BLOCK_SIZE = 1024,
ARGON2_QWORDS_IN_BLOCK = ARGON2_BLOCK_SIZE / 8,
ARGON2_OWORDS_IN_BLOCK = ARGON2_BLOCK_SIZE / 16,
ARGON2_HWORDS_IN_BLOCK = ARGON2_BLOCK_SIZE / 32,
/* Number of pseudo-random values generated by one call to Blake in Argon2i
to

46
src/opt.c
View File

@ -34,6 +34,42 @@
* @param with_xor Whether to XOR into the new block (1) or just overwrite (0)
* @pre all block pointers must be valid
*/
#if defined(__AVX2__)
static void fill_block(__m256i *state, const block *ref_block,
block *next_block, int with_xor) {
__m256i block_XY[ARGON2_HWORDS_IN_BLOCK];
unsigned int i;
if (with_xor) {
for (i = 0; i < ARGON2_HWORDS_IN_BLOCK; i++) {
state[i] = _mm256_xor_si256(
state[i], _mm256_loadu_si256((const __m256i *)ref_block->v + i));
block_XY[i] = _mm256_xor_si256(
state[i], _mm256_loadu_si256((const __m256i *)next_block->v + i));
}
} else {
for (i = 0; i < ARGON2_HWORDS_IN_BLOCK; i++) {
block_XY[i] = state[i] = _mm256_xor_si256(
state[i], _mm256_loadu_si256((const __m256i *)ref_block->v + i));
}
}
for (i = 0; i < 4; ++i) {
BLAKE2_ROUND_1(state[8 * i + 0], state[8 * i + 4], state[8 * i + 1], state[8 * i + 5],
state[8 * i + 2], state[8 * i + 6], state[8 * i + 3], state[8 * i + 7]);
}
for (i = 0; i < 4; ++i) {
BLAKE2_ROUND_2(state[ 0 + i], state[ 4 + i], state[ 8 + i], state[12 + i],
state[16 + i], state[20 + i], state[24 + i], state[28 + i]);
}
for (i = 0; i < ARGON2_HWORDS_IN_BLOCK; i++) {
state[i] = _mm256_xor_si256(state[i], block_XY[i]);
_mm256_storeu_si256((__m256i *)next_block->v + i, state[i]);
}
}
#else
static void fill_block(__m128i *state, const block *ref_block,
block *next_block, int with_xor) {
__m128i block_XY[ARGON2_OWORDS_IN_BLOCK];
@ -70,11 +106,17 @@ static void fill_block(__m128i *state, const block *ref_block,
_mm_storeu_si128((__m128i *)next_block->v + i, state[i]);
}
}
#endif
static void next_addresses(block *address_block, block *input_block) {
/*Temporary zero-initialized blocks*/
#if defined(__AVX2__)
__m256i zero_block[ARGON2_HWORDS_IN_BLOCK];
__m256i zero2_block[ARGON2_HWORDS_IN_BLOCK];
#else
__m128i zero_block[ARGON2_OWORDS_IN_BLOCK];
__m128i zero2_block[ARGON2_OWORDS_IN_BLOCK];
#endif
memset(zero_block, 0, sizeof(zero_block));
memset(zero2_block, 0, sizeof(zero2_block));
@ -96,7 +138,11 @@ void fill_segment(const argon2_instance_t *instance,
uint64_t pseudo_rand, ref_index, ref_lane;
uint32_t prev_offset, curr_offset;
uint32_t starting_index, i;
#if defined(__AVX2__)
__m256i state[32];
#else
__m128i state[64];
#endif
int data_independent_addressing;
if (instance == NULL) {