Ecosystem/xmrig
Ecosystem/xmrig
1
0
Fork 0
mirror of https://github.com/xmrig/xmrig.git synced 2024-12-22 11:39:33 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

AstroBWT: add AVX2 Salsa20 implementation

Browse source 
+4.5% speedup on Ryzen 5 5600X
This commit is contained in:
SChernykh 2021-08-29 10:35:43 +02:00
parent df4532d9a1
commit 3dc192f63e
7 changed files with 1556 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
cmake/astrobwt.cmake
View file

@ -23,6 +23,7 @@ if (WITH_ASTROBWT)
else()
if (CMAKE_SIZEOF_VOID_P EQUAL 8)
add_definitions(/DASTROBWT_AVX2)
list(APPEND SOURCES_CRYPTO src/crypto/astrobwt/xmm6int/salsa20_xmm6int-avx2.c)
if (CMAKE_C_COMPILER_ID MATCHES MSVC)
enable_language(ASM_MASM)
list(APPEND SOURCES_CRYPTO src/crypto/astrobwt/sha3_256_avx2.asm)

31
src/crypto/astrobwt/AstroBWT.cpp
View file

@ -70,7 +70,17 @@ static void Salsa20_XORKeyStream(const void* key, void* output, size_t size)
{
const uint64_t iv = 0;
ZeroTier::Salsa20 s(key, &iv);
s.XORKeyStream(output, size);
s.XORKeyStream(output, static_cast<uint32_t>(size));
memset(static_cast<uint8_t*>(output) - 16, 0, 16);
memset(static_cast<uint8_t*>(output) + size, 0, 16);
}
extern "C" int salsa20_stream_avx2(void* c, uint64_t clen, const void* iv, const void* key);
static void Salsa20_XORKeyStream_AVX256(const void* key, void* output, size_t size)
{
const uint64_t iv = 0;
salsa20_stream_avx2(output, size, &iv, key);
memset(static_cast<uint8_t*>(output) - 16, 0, 16);
memset(static_cast<uint8_t*>(output) + size, 0, 16);
}
@ -167,13 +177,16 @@ bool xmrig::astrobwt::astrobwt_dero(const void* input_data, uint32_t input_size,
uint8_t* stage2_result = (uint8_t*)(tmp_indices);
#ifdef ASTROBWT_AVX2
if (hasAVX2 && avx2)
if (hasAVX2 && avx2) {
SHA3_256_AVX2_ASM(input_data, input_size, key);
Salsa20_XORKeyStream_AVX256(key, stage1_output, STAGE1_SIZE);
}
else
#endif
{
sha3_HashBuffer(256, SHA3_FLAGS_NONE, input_data, input_size, key, sizeof(key));
Salsa20_XORKeyStream(key, stage1_output, STAGE1_SIZE);
Salsa20_XORKeyStream(key, stage1_output, STAGE1_SIZE);
}
sort_indices(STAGE1_SIZE + 1, stage1_output, indices, tmp_indices);
@ -196,7 +209,15 @@ bool xmrig::astrobwt::astrobwt_dero(const void* input_data, uint32_t input_size,
return false;
}
Salsa20_XORKeyStream(key, stage2_output, stage2_size);
#ifdef ASTROBWT_AVX2
if (hasAVX2 && avx2) {
Salsa20_XORKeyStream_AVX256(key, stage2_output, stage2_size);
}
else
#endif
{
Salsa20_XORKeyStream(key, stage2_output, stage2_size);
}
sort_indices(stage2_size + 1, stage2_output, indices, tmp_indices);

105
src/crypto/astrobwt/xmm6int/salsa20_xmm6int-avx2.c Normal file
View file

@ -0,0 +1,105 @@
/*
* ISC License
*
* Copyright (c) 2013-2021
* Frank Denis <j at pureftpd dot org>
*
* Permission to use, copy, modify, and/or distribute this software for any
* purpose with or without fee is hereby granted, provided that the above
* copyright notice and this permission notice appear in all copies.
*
* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
*/
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#ifdef __GNUC__
#pragma GCC target("sse2")
#pragma GCC target("ssse3")
#pragma GCC target("sse4.1")
#pragma GCC target("avx2")
#endif
#include <emmintrin.h>
#include <immintrin.h>
#include <smmintrin.h>
#include <tmmintrin.h>
#define ROUNDS 20
typedef struct salsa_ctx {
uint32_t input[16];
} salsa_ctx;
static const int TR[16] = {
0, 5, 10, 15, 12, 1, 6, 11, 8, 13, 2, 7, 4, 9, 14, 3
};
#define LOAD32_LE(p) *((uint32_t*)(p))
#define STORE32_LE(dst, src) memcpy((dst), &(src), sizeof(uint32_t))
static void
salsa_keysetup(salsa_ctx *ctx, const uint8_t *k)
{
ctx->input[TR[1]] = LOAD32_LE(k + 0);
ctx->input[TR[2]] = LOAD32_LE(k + 4);
ctx->input[TR[3]] = LOAD32_LE(k + 8);
ctx->input[TR[4]] = LOAD32_LE(k + 12);
ctx->input[TR[11]] = LOAD32_LE(k + 16);
ctx->input[TR[12]] = LOAD32_LE(k + 20);
ctx->input[TR[13]] = LOAD32_LE(k + 24);
ctx->input[TR[14]] = LOAD32_LE(k + 28);
ctx->input[TR[0]] = 0x61707865;
ctx->input[TR[5]] = 0x3320646e;
ctx->input[TR[10]] = 0x79622d32;
ctx->input[TR[15]] = 0x6b206574;
}
static void
salsa_ivsetup(salsa_ctx *ctx, const uint8_t *iv, const uint8_t *counter)
{
ctx->input[TR[6]] = LOAD32_LE(iv + 0);
ctx->input[TR[7]] = LOAD32_LE(iv + 4);
ctx->input[TR[8]] = counter == NULL ? 0 : LOAD32_LE(counter + 0);
ctx->input[TR[9]] = counter == NULL ? 0 : LOAD32_LE(counter + 4);
}
static void
salsa20_encrypt_bytes(salsa_ctx *ctx, const uint8_t *m, uint8_t *c,
unsigned long long bytes)
{
uint32_t * const x = &ctx->input[0];
if (!bytes) {
return; /* LCOV_EXCL_LINE */
}
#include "u8.h"
#include "u4.h"
#include "u1.h"
#include "u0.h"
}
int salsa20_stream_avx2(void* c, uint64_t clen, const void* iv, const void* key)
{
struct salsa_ctx ctx;
if (!clen) {
return 0;
}
salsa_keysetup(&ctx, (const uint8_t*)key);
salsa_ivsetup(&ctx, (const uint8_t*)iv, NULL);
memset(c, 0, clen);
salsa20_encrypt_bytes(&ctx, (const uint8_t*)c, (uint8_t*)c, clen);
return 0;
}

193
src/crypto/astrobwt/xmm6int/u0.h Normal file
View file

@ -0,0 +1,193 @@
if (bytes > 0) {
__m128i diag0 = _mm_loadu_si128((const __m128i *) (x + 0));
__m128i diag1 = _mm_loadu_si128((const __m128i *) (x + 4));
__m128i diag2 = _mm_loadu_si128((const __m128i *) (x + 8));
__m128i diag3 = _mm_loadu_si128((const __m128i *) (x + 12));
__m128i a0, a1, a2, a3, a4, a5, a6, a7;
__m128i b0, b1, b2, b3, b4, b5, b6, b7;
uint8_t partialblock[64];
unsigned int i;
a0 = diag1;
for (i = 0; i < ROUNDS; i += 4) {
a0 = _mm_add_epi32(a0, diag0);
a1 = diag0;
b0 = a0;
a0 = _mm_slli_epi32(a0, 7);
b0 = _mm_srli_epi32(b0, 25);
diag3 = _mm_xor_si128(diag3, a0);
diag3 = _mm_xor_si128(diag3, b0);
a1 = _mm_add_epi32(a1, diag3);
a2 = diag3;
b1 = a1;
a1 = _mm_slli_epi32(a1, 9);
b1 = _mm_srli_epi32(b1, 23);
diag2 = _mm_xor_si128(diag2, a1);
diag3 = _mm_shuffle_epi32(diag3, 0x93);
diag2 = _mm_xor_si128(diag2, b1);
a2 = _mm_add_epi32(a2, diag2);
a3 = diag2;
b2 = a2;
a2 = _mm_slli_epi32(a2, 13);
b2 = _mm_srli_epi32(b2, 19);
diag1 = _mm_xor_si128(diag1, a2);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag1 = _mm_xor_si128(diag1, b2);
a3 = _mm_add_epi32(a3, diag1);
a4 = diag3;
b3 = a3;
a3 = _mm_slli_epi32(a3, 18);
b3 = _mm_srli_epi32(b3, 14);
diag0 = _mm_xor_si128(diag0, a3);
diag1 = _mm_shuffle_epi32(diag1, 0x39);
diag0 = _mm_xor_si128(diag0, b3);
a4 = _mm_add_epi32(a4, diag0);
a5 = diag0;
b4 = a4;
a4 = _mm_slli_epi32(a4, 7);
b4 = _mm_srli_epi32(b4, 25);
diag1 = _mm_xor_si128(diag1, a4);
diag1 = _mm_xor_si128(diag1, b4);
a5 = _mm_add_epi32(a5, diag1);
a6 = diag1;
b5 = a5;
a5 = _mm_slli_epi32(a5, 9);
b5 = _mm_srli_epi32(b5, 23);
diag2 = _mm_xor_si128(diag2, a5);
diag1 = _mm_shuffle_epi32(diag1, 0x93);
diag2 = _mm_xor_si128(diag2, b5);
a6 = _mm_add_epi32(a6, diag2);
a7 = diag2;
b6 = a6;
a6 = _mm_slli_epi32(a6, 13);
b6 = _mm_srli_epi32(b6, 19);
diag3 = _mm_xor_si128(diag3, a6);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag3 = _mm_xor_si128(diag3, b6);
a7 = _mm_add_epi32(a7, diag3);
a0 = diag1;
b7 = a7;
a7 = _mm_slli_epi32(a7, 18);
b7 = _mm_srli_epi32(b7, 14);
diag0 = _mm_xor_si128(diag0, a7);
diag3 = _mm_shuffle_epi32(diag3, 0x39);
diag0 = _mm_xor_si128(diag0, b7);
a0 = _mm_add_epi32(a0, diag0);
a1 = diag0;
b0 = a0;
a0 = _mm_slli_epi32(a0, 7);
b0 = _mm_srli_epi32(b0, 25);
diag3 = _mm_xor_si128(diag3, a0);
diag3 = _mm_xor_si128(diag3, b0);
a1 = _mm_add_epi32(a1, diag3);
a2 = diag3;
b1 = a1;
a1 = _mm_slli_epi32(a1, 9);
b1 = _mm_srli_epi32(b1, 23);
diag2 = _mm_xor_si128(diag2, a1);
diag3 = _mm_shuffle_epi32(diag3, 0x93);
diag2 = _mm_xor_si128(diag2, b1);
a2 = _mm_add_epi32(a2, diag2);
a3 = diag2;
b2 = a2;
a2 = _mm_slli_epi32(a2, 13);
b2 = _mm_srli_epi32(b2, 19);
diag1 = _mm_xor_si128(diag1, a2);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag1 = _mm_xor_si128(diag1, b2);
a3 = _mm_add_epi32(a3, diag1);
a4 = diag3;
b3 = a3;
a3 = _mm_slli_epi32(a3, 18);
b3 = _mm_srli_epi32(b3, 14);
diag0 = _mm_xor_si128(diag0, a3);
diag1 = _mm_shuffle_epi32(diag1, 0x39);
diag0 = _mm_xor_si128(diag0, b3);
a4 = _mm_add_epi32(a4, diag0);
a5 = diag0;
b4 = a4;
a4 = _mm_slli_epi32(a4, 7);
b4 = _mm_srli_epi32(b4, 25);
diag1 = _mm_xor_si128(diag1, a4);
diag1 = _mm_xor_si128(diag1, b4);
a5 = _mm_add_epi32(a5, diag1);
a6 = diag1;
b5 = a5;
a5 = _mm_slli_epi32(a5, 9);
b5 = _mm_srli_epi32(b5, 23);
diag2 = _mm_xor_si128(diag2, a5);
diag1 = _mm_shuffle_epi32(diag1, 0x93);
diag2 = _mm_xor_si128(diag2, b5);
a6 = _mm_add_epi32(a6, diag2);
a7 = diag2;
b6 = a6;
a6 = _mm_slli_epi32(a6, 13);
b6 = _mm_srli_epi32(b6, 19);
diag3 = _mm_xor_si128(diag3, a6);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag3 = _mm_xor_si128(diag3, b6);
a7 = _mm_add_epi32(a7, diag3);
a0 = diag1;
b7 = a7;
a7 = _mm_slli_epi32(a7, 18);
b7 = _mm_srli_epi32(b7, 14);
diag0 = _mm_xor_si128(diag0, a7);
diag3 = _mm_shuffle_epi32(diag3, 0x39);
diag0 = _mm_xor_si128(diag0, b7);
}
diag0 = _mm_add_epi32(diag0, _mm_loadu_si128((const __m128i *) (x + 0)));
diag1 = _mm_add_epi32(diag1, _mm_loadu_si128((const __m128i *) (x + 4)));
diag2 = _mm_add_epi32(diag2, _mm_loadu_si128((const __m128i *) (x + 8)));
diag3 = _mm_add_epi32(diag3, _mm_loadu_si128((const __m128i *) (x + 12)));
#define ONEQUAD_SHUFFLE(A, B, C, D) \
do { \
uint32_t in##A = _mm_cvtsi128_si32(diag0); \
uint32_t in##B = _mm_cvtsi128_si32(diag1); \
uint32_t in##C = _mm_cvtsi128_si32(diag2); \
uint32_t in##D = _mm_cvtsi128_si32(diag3); \
diag0 = _mm_shuffle_epi32(diag0, 0x39); \
diag1 = _mm_shuffle_epi32(diag1, 0x39); \
diag2 = _mm_shuffle_epi32(diag2, 0x39); \
diag3 = _mm_shuffle_epi32(diag3, 0x39); \
*(uint32_t *) (partialblock + (A * 4)) = in##A; \
*(uint32_t *) (partialblock + (B * 4)) = in##B; \
*(uint32_t *) (partialblock + (C * 4)) = in##C; \
*(uint32_t *) (partialblock + (D * 4)) = in##D; \
} while (0)
#define ONEQUAD(A, B, C, D) ONEQUAD_SHUFFLE(A, B, C, D)
ONEQUAD(0, 12, 8, 4);
ONEQUAD(5, 1, 13, 9);
ONEQUAD(10, 6, 2, 14);
ONEQUAD(15, 11, 7, 3);
#undef ONEQUAD
#undef ONEQUAD_SHUFFLE
for (i = 0; i < bytes; i++) {
c[i] = m[i] ^ partialblock[i];
}
}

207
src/crypto/astrobwt/xmm6int/u1.h Normal file
View file

@ -0,0 +1,207 @@
while (bytes >= 64) {
__m128i diag0 = _mm_loadu_si128((const __m128i *) (x + 0));
__m128i diag1 = _mm_loadu_si128((const __m128i *) (x + 4));
__m128i diag2 = _mm_loadu_si128((const __m128i *) (x + 8));
__m128i diag3 = _mm_loadu_si128((const __m128i *) (x + 12));
__m128i a0, a1, a2, a3, a4, a5, a6, a7;
__m128i b0, b1, b2, b3, b4, b5, b6, b7;
uint32_t in8;
uint32_t in9;
int i;
a0 = diag1;
for (i = 0; i < ROUNDS; i += 4) {
a0 = _mm_add_epi32(a0, diag0);
a1 = diag0;
b0 = a0;
a0 = _mm_slli_epi32(a0, 7);
b0 = _mm_srli_epi32(b0, 25);
diag3 = _mm_xor_si128(diag3, a0);
diag3 = _mm_xor_si128(diag3, b0);
a1 = _mm_add_epi32(a1, diag3);
a2 = diag3;
b1 = a1;
a1 = _mm_slli_epi32(a1, 9);
b1 = _mm_srli_epi32(b1, 23);
diag2 = _mm_xor_si128(diag2, a1);
diag3 = _mm_shuffle_epi32(diag3, 0x93);
diag2 = _mm_xor_si128(diag2, b1);
a2 = _mm_add_epi32(a2, diag2);
a3 = diag2;
b2 = a2;
a2 = _mm_slli_epi32(a2, 13);
b2 = _mm_srli_epi32(b2, 19);
diag1 = _mm_xor_si128(diag1, a2);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag1 = _mm_xor_si128(diag1, b2);
a3 = _mm_add_epi32(a3, diag1);
a4 = diag3;
b3 = a3;
a3 = _mm_slli_epi32(a3, 18);
b3 = _mm_srli_epi32(b3, 14);
diag0 = _mm_xor_si128(diag0, a3);
diag1 = _mm_shuffle_epi32(diag1, 0x39);
diag0 = _mm_xor_si128(diag0, b3);
a4 = _mm_add_epi32(a4, diag0);
a5 = diag0;
b4 = a4;
a4 = _mm_slli_epi32(a4, 7);
b4 = _mm_srli_epi32(b4, 25);
diag1 = _mm_xor_si128(diag1, a4);
diag1 = _mm_xor_si128(diag1, b4);
a5 = _mm_add_epi32(a5, diag1);
a6 = diag1;
b5 = a5;
a5 = _mm_slli_epi32(a5, 9);
b5 = _mm_srli_epi32(b5, 23);
diag2 = _mm_xor_si128(diag2, a5);
diag1 = _mm_shuffle_epi32(diag1, 0x93);
diag2 = _mm_xor_si128(diag2, b5);
a6 = _mm_add_epi32(a6, diag2);
a7 = diag2;
b6 = a6;
a6 = _mm_slli_epi32(a6, 13);
b6 = _mm_srli_epi32(b6, 19);
diag3 = _mm_xor_si128(diag3, a6);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag3 = _mm_xor_si128(diag3, b6);
a7 = _mm_add_epi32(a7, diag3);
a0 = diag1;
b7 = a7;
a7 = _mm_slli_epi32(a7, 18);
b7 = _mm_srli_epi32(b7, 14);
diag0 = _mm_xor_si128(diag0, a7);
diag3 = _mm_shuffle_epi32(diag3, 0x39);
diag0 = _mm_xor_si128(diag0, b7);
a0 = _mm_add_epi32(a0, diag0);
a1 = diag0;
b0 = a0;
a0 = _mm_slli_epi32(a0, 7);
b0 = _mm_srli_epi32(b0, 25);
diag3 = _mm_xor_si128(diag3, a0);
diag3 = _mm_xor_si128(diag3, b0);
a1 = _mm_add_epi32(a1, diag3);
a2 = diag3;
b1 = a1;
a1 = _mm_slli_epi32(a1, 9);
b1 = _mm_srli_epi32(b1, 23);
diag2 = _mm_xor_si128(diag2, a1);
diag3 = _mm_shuffle_epi32(diag3, 0x93);
diag2 = _mm_xor_si128(diag2, b1);
a2 = _mm_add_epi32(a2, diag2);
a3 = diag2;
b2 = a2;
a2 = _mm_slli_epi32(a2, 13);
b2 = _mm_srli_epi32(b2, 19);
diag1 = _mm_xor_si128(diag1, a2);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag1 = _mm_xor_si128(diag1, b2);
a3 = _mm_add_epi32(a3, diag1);
a4 = diag3;
b3 = a3;
a3 = _mm_slli_epi32(a3, 18);
b3 = _mm_srli_epi32(b3, 14);
diag0 = _mm_xor_si128(diag0, a3);
diag1 = _mm_shuffle_epi32(diag1, 0x39);
diag0 = _mm_xor_si128(diag0, b3);
a4 = _mm_add_epi32(a4, diag0);
a5 = diag0;
b4 = a4;
a4 = _mm_slli_epi32(a4, 7);
b4 = _mm_srli_epi32(b4, 25);
diag1 = _mm_xor_si128(diag1, a4);
diag1 = _mm_xor_si128(diag1, b4);
a5 = _mm_add_epi32(a5, diag1);
a6 = diag1;
b5 = a5;
a5 = _mm_slli_epi32(a5, 9);
b5 = _mm_srli_epi32(b5, 23);
diag2 = _mm_xor_si128(diag2, a5);
diag1 = _mm_shuffle_epi32(diag1, 0x93);
diag2 = _mm_xor_si128(diag2, b5);
a6 = _mm_add_epi32(a6, diag2);
a7 = diag2;
b6 = a6;
a6 = _mm_slli_epi32(a6, 13);
b6 = _mm_srli_epi32(b6, 19);
diag3 = _mm_xor_si128(diag3, a6);
diag2 = _mm_shuffle_epi32(diag2, 0x4e);
diag3 = _mm_xor_si128(diag3, b6);
a7 = _mm_add_epi32(a7, diag3);
a0 = diag1;
b7 = a7;
a7 = _mm_slli_epi32(a7, 18);
b7 = _mm_srli_epi32(b7, 14);
diag0 = _mm_xor_si128(diag0, a7);
diag3 = _mm_shuffle_epi32(diag3, 0x39);
diag0 = _mm_xor_si128(diag0, b7);
}
diag0 = _mm_add_epi32(diag0, _mm_loadu_si128((const __m128i *) (x + 0)));
diag1 = _mm_add_epi32(diag1, _mm_loadu_si128((const __m128i *) (x + 4)));
diag2 = _mm_add_epi32(diag2, _mm_loadu_si128((const __m128i *) (x + 8)));
diag3 = _mm_add_epi32(diag3, _mm_loadu_si128((const __m128i *) (x + 12)));
#define ONEQUAD_SHUFFLE(A, B, C, D) \
do { \
uint32_t in##A = _mm_cvtsi128_si32(diag0); \
uint32_t in##B = _mm_cvtsi128_si32(diag1); \
uint32_t in##C = _mm_cvtsi128_si32(diag2); \
uint32_t in##D = _mm_cvtsi128_si32(diag3); \
diag0 = _mm_shuffle_epi32(diag0, 0x39); \
diag1 = _mm_shuffle_epi32(diag1, 0x39); \
diag2 = _mm_shuffle_epi32(diag2, 0x39); \
diag3 = _mm_shuffle_epi32(diag3, 0x39); \
in##A ^= *(const uint32_t *) (m + (A * 4)); \
in##B ^= *(const uint32_t *) (m + (B * 4)); \
in##C ^= *(const uint32_t *) (m + (C * 4)); \
in##D ^= *(const uint32_t *) (m + (D * 4)); \
*(uint32_t *) (c + (A * 4)) = in##A; \
*(uint32_t *) (c + (B * 4)) = in##B; \
*(uint32_t *) (c + (C * 4)) = in##C; \
*(uint32_t *) (c + (D * 4)) = in##D; \
} while (0)
#define ONEQUAD(A, B, C, D) ONEQUAD_SHUFFLE(A, B, C, D)
ONEQUAD(0, 12, 8, 4);
ONEQUAD(5, 1, 13, 9);
ONEQUAD(10, 6, 2, 14);
ONEQUAD(15, 11, 7, 3);
#undef ONEQUAD
#undef ONEQUAD_SHUFFLE
in8 = x[8];
in9 = x[13];
in8++;
if (in8 == 0) {
in9++;
}
x[8] = in8;
x[13] = in9;
c += 64;
m += 64;
bytes -= 64;
}

547
src/crypto/astrobwt/xmm6int/u4.h Normal file
View file

@ -0,0 +1,547 @@
if (bytes >= 256) {
__m128i y0, y1, y2, y3, y4, y5, y6, y7, y8, y9, y10, y11, y12, y13, y14,
y15;
__m128i z0, z1, z2, z3, z4, z5, z6, z7, z8, z9, z10, z11, z12, z13, z14,
z15;
__m128i orig0, orig1, orig2, orig3, orig4, orig5, orig6, orig7, orig8,
orig9, orig10, orig11, orig12, orig13, orig14, orig15;
uint32_t in8;
uint32_t in9;
int i;
/* element broadcast immediate for _mm_shuffle_epi32 are in order:
0x00, 0x55, 0xaa, 0xff */
z0 = _mm_loadu_si128((const __m128i *) (x + 0));
z5 = _mm_shuffle_epi32(z0, 0x55);
z10 = _mm_shuffle_epi32(z0, 0xaa);
z15 = _mm_shuffle_epi32(z0, 0xff);
z0 = _mm_shuffle_epi32(z0, 0x00);
z1 = _mm_loadu_si128((const __m128i *) (x + 4));
z6 = _mm_shuffle_epi32(z1, 0xaa);
z11 = _mm_shuffle_epi32(z1, 0xff);
z12 = _mm_shuffle_epi32(z1, 0x00);
z1 = _mm_shuffle_epi32(z1, 0x55);
z2 = _mm_loadu_si128((const __m128i *) (x + 8));
z7 = _mm_shuffle_epi32(z2, 0xff);
z13 = _mm_shuffle_epi32(z2, 0x55);
z2 = _mm_shuffle_epi32(z2, 0xaa);
/* no z8 -> first half of the nonce, will fill later */
z3 = _mm_loadu_si128((const __m128i *) (x + 12));
z4 = _mm_shuffle_epi32(z3, 0x00);
z14 = _mm_shuffle_epi32(z3, 0xaa);
z3 = _mm_shuffle_epi32(z3, 0xff);
/* no z9 -> second half of the nonce, will fill later */
orig0 = z0;
orig1 = z1;
orig2 = z2;
orig3 = z3;
orig4 = z4;
orig5 = z5;
orig6 = z6;
orig7 = z7;
orig10 = z10;
orig11 = z11;
orig12 = z12;
orig13 = z13;
orig14 = z14;
orig15 = z15;
while (bytes >= 256) {
/* vector implementation for z8 and z9 */
/* not sure if it helps for only 4 blocks */
const __m128i addv8 = _mm_set_epi64x(1, 0);
const __m128i addv9 = _mm_set_epi64x(3, 2);
__m128i t8, t9;
uint64_t in89;
in8 = x[8];
in9 = x[13];
in89 = ((uint64_t) in8) | (((uint64_t) in9) << 32);
t8 = _mm_set1_epi64x(in89);
t9 = _mm_set1_epi64x(in89);
z8 = _mm_add_epi64(addv8, t8);
z9 = _mm_add_epi64(addv9, t9);
t8 = _mm_unpacklo_epi32(z8, z9);
t9 = _mm_unpackhi_epi32(z8, z9);
z8 = _mm_unpacklo_epi32(t8, t9);
z9 = _mm_unpackhi_epi32(t8, t9);
orig8 = z8;
orig9 = z9;
in89 += 4;
x[8] = in89 & 0xFFFFFFFF;
x[13] = (in89 >> 32) & 0xFFFFFFFF;
z5 = orig5;
z10 = orig10;
z15 = orig15;
z14 = orig14;
z3 = orig3;
z6 = orig6;
z11 = orig11;
z1 = orig1;
z7 = orig7;
z13 = orig13;
z2 = orig2;
z9 = orig9;
z0 = orig0;
z12 = orig12;
z4 = orig4;
z8 = orig8;
for (i = 0; i < ROUNDS; i += 2) {
/* the inner loop is a direct translation (regexp search/replace)
* from the amd64-xmm6 ASM */
__m128i r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13,
r14, r15;
y4 = z12;
y4 = _mm_add_epi32(y4, z0);
r4 = y4;
y4 = _mm_slli_epi32(y4, 7);
z4 = _mm_xor_si128(z4, y4);
r4 = _mm_srli_epi32(r4, 25);
z4 = _mm_xor_si128(z4, r4);
y9 = z1;
y9 = _mm_add_epi32(y9, z5);
r9 = y9;
y9 = _mm_slli_epi32(y9, 7);
z9 = _mm_xor_si128(z9, y9);
r9 = _mm_srli_epi32(r9, 25);
z9 = _mm_xor_si128(z9, r9);
y8 = z0;
y8 = _mm_add_epi32(y8, z4);
r8 = y8;
y8 = _mm_slli_epi32(y8, 9);
z8 = _mm_xor_si128(z8, y8);
r8 = _mm_srli_epi32(r8, 23);
z8 = _mm_xor_si128(z8, r8);
y13 = z5;
y13 = _mm_add_epi32(y13, z9);
r13 = y13;
y13 = _mm_slli_epi32(y13, 9);
z13 = _mm_xor_si128(z13, y13);
r13 = _mm_srli_epi32(r13, 23);
z13 = _mm_xor_si128(z13, r13);
y12 = z4;
y12 = _mm_add_epi32(y12, z8);
r12 = y12;
y12 = _mm_slli_epi32(y12, 13);
z12 = _mm_xor_si128(z12, y12);
r12 = _mm_srli_epi32(r12, 19);
z12 = _mm_xor_si128(z12, r12);
y1 = z9;
y1 = _mm_add_epi32(y1, z13);
r1 = y1;
y1 = _mm_slli_epi32(y1, 13);
z1 = _mm_xor_si128(z1, y1);
r1 = _mm_srli_epi32(r1, 19);
z1 = _mm_xor_si128(z1, r1);
y0 = z8;
y0 = _mm_add_epi32(y0, z12);
r0 = y0;
y0 = _mm_slli_epi32(y0, 18);
z0 = _mm_xor_si128(z0, y0);
r0 = _mm_srli_epi32(r0, 14);
z0 = _mm_xor_si128(z0, r0);
y5 = z13;
y5 = _mm_add_epi32(y5, z1);
r5 = y5;
y5 = _mm_slli_epi32(y5, 18);
z5 = _mm_xor_si128(z5, y5);
r5 = _mm_srli_epi32(r5, 14);
z5 = _mm_xor_si128(z5, r5);
y14 = z6;
y14 = _mm_add_epi32(y14, z10);
r14 = y14;
y14 = _mm_slli_epi32(y14, 7);
z14 = _mm_xor_si128(z14, y14);
r14 = _mm_srli_epi32(r14, 25);
z14 = _mm_xor_si128(z14, r14);
y3 = z11;
y3 = _mm_add_epi32(y3, z15);
r3 = y3;
y3 = _mm_slli_epi32(y3, 7);
z3 = _mm_xor_si128(z3, y3);
r3 = _mm_srli_epi32(r3, 25);
z3 = _mm_xor_si128(z3, r3);
y2 = z10;
y2 = _mm_add_epi32(y2, z14);
r2 = y2;
y2 = _mm_slli_epi32(y2, 9);
z2 = _mm_xor_si128(z2, y2);
r2 = _mm_srli_epi32(r2, 23);
z2 = _mm_xor_si128(z2, r2);
y7 = z15;
y7 = _mm_add_epi32(y7, z3);
r7 = y7;
y7 = _mm_slli_epi32(y7, 9);
z7 = _mm_xor_si128(z7, y7);
r7 = _mm_srli_epi32(r7, 23);
z7 = _mm_xor_si128(z7, r7);
y6 = z14;
y6 = _mm_add_epi32(y6, z2);
r6 = y6;
y6 = _mm_slli_epi32(y6, 13);
z6 = _mm_xor_si128(z6, y6);
r6 = _mm_srli_epi32(r6, 19);
z6 = _mm_xor_si128(z6, r6);
y11 = z3;
y11 = _mm_add_epi32(y11, z7);
r11 = y11;
y11 = _mm_slli_epi32(y11, 13);
z11 = _mm_xor_si128(z11, y11);
r11 = _mm_srli_epi32(r11, 19);
z11 = _mm_xor_si128(z11, r11);
y10 = z2;
y10 = _mm_add_epi32(y10, z6);
r10 = y10;
y10 = _mm_slli_epi32(y10, 18);
z10 = _mm_xor_si128(z10, y10);
r10 = _mm_srli_epi32(r10, 14);
z10 = _mm_xor_si128(z10, r10);
y1 = z3;
y1 = _mm_add_epi32(y1, z0);
r1 = y1;
y1 = _mm_slli_epi32(y1, 7);
z1 = _mm_xor_si128(z1, y1);
r1 = _mm_srli_epi32(r1, 25);
z1 = _mm_xor_si128(z1, r1);
y15 = z7;
y15 = _mm_add_epi32(y15, z11);
r15 = y15;
y15 = _mm_slli_epi32(y15, 18);
z15 = _mm_xor_si128(z15, y15);
r15 = _mm_srli_epi32(r15, 14);
z15 = _mm_xor_si128(z15, r15);
y6 = z4;
y6 = _mm_add_epi32(y6, z5);
r6 = y6;
y6 = _mm_slli_epi32(y6, 7);
z6 = _mm_xor_si128(z6, y6);
r6 = _mm_srli_epi32(r6, 25);
z6 = _mm_xor_si128(z6, r6);
y2 = z0;
y2 = _mm_add_epi32(y2, z1);
r2 = y2;
y2 = _mm_slli_epi32(y2, 9);
z2 = _mm_xor_si128(z2, y2);
r2 = _mm_srli_epi32(r2, 23);
z2 = _mm_xor_si128(z2, r2);
y7 = z5;
y7 = _mm_add_epi32(y7, z6);
r7 = y7;
y7 = _mm_slli_epi32(y7, 9);
z7 = _mm_xor_si128(z7, y7);
r7 = _mm_srli_epi32(r7, 23);
z7 = _mm_xor_si128(z7, r7);
y3 = z1;
y3 = _mm_add_epi32(y3, z2);
r3 = y3;
y3 = _mm_slli_epi32(y3, 13);
z3 = _mm_xor_si128(z3, y3);
r3 = _mm_srli_epi32(r3, 19);
z3 = _mm_xor_si128(z3, r3);
y4 = z6;
y4 = _mm_add_epi32(y4, z7);
r4 = y4;
y4 = _mm_slli_epi32(y4, 13);
z4 = _mm_xor_si128(z4, y4);
r4 = _mm_srli_epi32(r4, 19);
z4 = _mm_xor_si128(z4, r4);
y0 = z2;
y0 = _mm_add_epi32(y0, z3);
r0 = y0;
y0 = _mm_slli_epi32(y0, 18);
z0 = _mm_xor_si128(z0, y0);
r0 = _mm_srli_epi32(r0, 14);
z0 = _mm_xor_si128(z0, r0);
y5 = z7;
y5 = _mm_add_epi32(y5, z4);
r5 = y5;
y5 = _mm_slli_epi32(y5, 18);
z5 = _mm_xor_si128(z5, y5);
r5 = _mm_srli_epi32(r5, 14);
z5 = _mm_xor_si128(z5, r5);
y11 = z9;
y11 = _mm_add_epi32(y11, z10);
r11 = y11;
y11 = _mm_slli_epi32(y11, 7);
z11 = _mm_xor_si128(z11, y11);
r11 = _mm_srli_epi32(r11, 25);
z11 = _mm_xor_si128(z11, r11);
y12 = z14;
y12 = _mm_add_epi32(y12, z15);
r12 = y12;
y12 = _mm_slli_epi32(y12, 7);
z12 = _mm_xor_si128(z12, y12);
r12 = _mm_srli_epi32(r12, 25);
z12 = _mm_xor_si128(z12, r12);
y8 = z10;
y8 = _mm_add_epi32(y8, z11);
r8 = y8;
y8 = _mm_slli_epi32(y8, 9);
z8 = _mm_xor_si128(z8, y8);
r8 = _mm_srli_epi32(r8, 23);
z8 = _mm_xor_si128(z8, r8);
y13 = z15;
y13 = _mm_add_epi32(y13, z12);
r13 = y13;
y13 = _mm_slli_epi32(y13, 9);
z13 = _mm_xor_si128(z13, y13);
r13 = _mm_srli_epi32(r13, 23);
z13 = _mm_xor_si128(z13, r13);
y9 = z11;
y9 = _mm_add_epi32(y9, z8);
r9 = y9;
y9 = _mm_slli_epi32(y9, 13);
z9 = _mm_xor_si128(z9, y9);
r9 = _mm_srli_epi32(r9, 19);
z9 = _mm_xor_si128(z9, r9);
y14 = z12;
y14 = _mm_add_epi32(y14, z13);
r14 = y14;
y14 = _mm_slli_epi32(y14, 13);
z14 = _mm_xor_si128(z14, y14);
r14 = _mm_srli_epi32(r14, 19);
z14 = _mm_xor_si128(z14, r14);
y10 = z8;
y10 = _mm_add_epi32(y10, z9);
r10 = y10;
y10 = _mm_slli_epi32(y10, 18);
z10 = _mm_xor_si128(z10, y10);
r10 = _mm_srli_epi32(r10, 14);
z10 = _mm_xor_si128(z10, r10);
y15 = z13;
y15 = _mm_add_epi32(y15, z14);
r15 = y15;
y15 = _mm_slli_epi32(y15, 18);
z15 = _mm_xor_si128(z15, y15);
r15 = _mm_srli_epi32(r15, 14);
z15 = _mm_xor_si128(z15, r15);
}
/* store data ; this macro replicates the original amd64-xmm6 code */
#define ONEQUAD_SHUFFLE(A, B, C, D) \
z##A = _mm_add_epi32(z##A, orig##A); \
z##B = _mm_add_epi32(z##B, orig##B); \
z##C = _mm_add_epi32(z##C, orig##C); \
z##D = _mm_add_epi32(z##D, orig##D); \
in##A = _mm_cvtsi128_si32(z##A); \
in##B = _mm_cvtsi128_si32(z##B); \
in##C = _mm_cvtsi128_si32(z##C); \
in##D = _mm_cvtsi128_si32(z##D); \
z##A = _mm_shuffle_epi32(z##A, 0x39); \
z##B = _mm_shuffle_epi32(z##B, 0x39); \
z##C = _mm_shuffle_epi32(z##C, 0x39); \
z##D = _mm_shuffle_epi32(z##D, 0x39); \
\
in##A ^= *(uint32_t *) (m + 0); \
in##B ^= *(uint32_t *) (m + 4); \
in##C ^= *(uint32_t *) (m + 8); \
in##D ^= *(uint32_t *) (m + 12); \
\
*(uint32_t *) (c + 0) = in##A; \
*(uint32_t *) (c + 4) = in##B; \
*(uint32_t *) (c + 8) = in##C; \
*(uint32_t *) (c + 12) = in##D; \
\
in##A = _mm_cvtsi128_si32(z##A); \
in##B = _mm_cvtsi128_si32(z##B); \
in##C = _mm_cvtsi128_si32(z##C); \
in##D = _mm_cvtsi128_si32(z##D); \
z##A = _mm_shuffle_epi32(z##A, 0x39); \
z##B = _mm_shuffle_epi32(z##B, 0x39); \
z##C = _mm_shuffle_epi32(z##C, 0x39); \
z##D = _mm_shuffle_epi32(z##D, 0x39); \
\
in##A ^= *(uint32_t *) (m + 64); \
in##B ^= *(uint32_t *) (m + 68); \
in##C ^= *(uint32_t *) (m + 72); \
in##D ^= *(uint32_t *) (m + 76); \
*(uint32_t *) (c + 64) = in##A; \
*(uint32_t *) (c + 68) = in##B; \
*(uint32_t *) (c + 72) = in##C; \
*(uint32_t *) (c + 76) = in##D; \
\
in##A = _mm_cvtsi128_si32(z##A); \
in##B = _mm_cvtsi128_si32(z##B); \
in##C = _mm_cvtsi128_si32(z##C); \
in##D = _mm_cvtsi128_si32(z##D); \
z##A = _mm_shuffle_epi32(z##A, 0x39); \
z##B = _mm_shuffle_epi32(z##B, 0x39); \
z##C = _mm_shuffle_epi32(z##C, 0x39); \
z##D = _mm_shuffle_epi32(z##D, 0x39); \
\
in##A ^= *(uint32_t *) (m + 128); \
in##B ^= *(uint32_t *) (m + 132); \
in##C ^= *(uint32_t *) (m + 136); \
in##D ^= *(uint32_t *) (m + 140); \
*(uint32_t *) (c + 128) = in##A; \
*(uint32_t *) (c + 132) = in##B; \
*(uint32_t *) (c + 136) = in##C; \
*(uint32_t *) (c + 140) = in##D; \
\
in##A = _mm_cvtsi128_si32(z##A); \
in##B = _mm_cvtsi128_si32(z##B); \
in##C = _mm_cvtsi128_si32(z##C); \
in##D = _mm_cvtsi128_si32(z##D); \
\
in##A ^= *(uint32_t *) (m + 192); \
in##B ^= *(uint32_t *) (m + 196); \
in##C ^= *(uint32_t *) (m + 200); \
in##D ^= *(uint32_t *) (m + 204); \
*(uint32_t *) (c + 192) = in##A; \
*(uint32_t *) (c + 196) = in##B; \
*(uint32_t *) (c + 200) = in##C; \
*(uint32_t *) (c + 204) = in##D
/* store data ; this macro replaces shuffle+mov by a direct extract; not much
* difference */
#define ONEQUAD_EXTRACT(A, B, C, D) \
z##A = _mm_add_epi32(z##A, orig##A); \
z##B = _mm_add_epi32(z##B, orig##B); \
z##C = _mm_add_epi32(z##C, orig##C); \
z##D = _mm_add_epi32(z##D, orig##D); \
in##A = _mm_cvtsi128_si32(z##A); \
in##B = _mm_cvtsi128_si32(z##B); \
in##C = _mm_cvtsi128_si32(z##C); \
in##D = _mm_cvtsi128_si32(z##D); \
in##A ^= *(uint32_t *) (m + 0); \
in##B ^= *(uint32_t *) (m + 4); \
in##C ^= *(uint32_t *) (m + 8); \
in##D ^= *(uint32_t *) (m + 12); \
*(uint32_t *) (c + 0) = in##A; \
*(uint32_t *) (c + 4) = in##B; \
*(uint32_t *) (c + 8) = in##C; \
*(uint32_t *) (c + 12) = in##D; \
\
in##A = _mm_extract_epi32(z##A, 1); \
in##B = _mm_extract_epi32(z##B, 1); \
in##C = _mm_extract_epi32(z##C, 1); \
in##D = _mm_extract_epi32(z##D, 1); \
\
in##A ^= *(uint32_t *) (m + 64); \
in##B ^= *(uint32_t *) (m + 68); \
in##C ^= *(uint32_t *) (m + 72); \
in##D ^= *(uint32_t *) (m + 76); \
*(uint32_t *) (c + 64) = in##A; \
*(uint32_t *) (c + 68) = in##B; \
*(uint32_t *) (c + 72) = in##C; \
*(uint32_t *) (c + 76) = in##D; \
\
in##A = _mm_extract_epi32(z##A, 2); \
in##B = _mm_extract_epi32(z##B, 2); \
in##C = _mm_extract_epi32(z##C, 2); \
in##D = _mm_extract_epi32(z##D, 2); \
\
in##A ^= *(uint32_t *) (m + 128); \
in##B ^= *(uint32_t *) (m + 132); \
in##C ^= *(uint32_t *) (m + 136); \
in##D ^= *(uint32_t *) (m + 140); \
*(uint32_t *) (c + 128) = in##A; \
*(uint32_t *) (c + 132) = in##B; \
*(uint32_t *) (c + 136) = in##C; \
*(uint32_t *) (c + 140) = in##D; \
\
in##A = _mm_extract_epi32(z##A, 3); \
in##B = _mm_extract_epi32(z##B, 3); \
in##C = _mm_extract_epi32(z##C, 3); \
in##D = _mm_extract_epi32(z##D, 3); \
\
in##A ^= *(uint32_t *) (m + 192); \
in##B ^= *(uint32_t *) (m + 196); \
in##C ^= *(uint32_t *) (m + 200); \
in##D ^= *(uint32_t *) (m + 204); \
*(uint32_t *) (c + 192) = in##A; \
*(uint32_t *) (c + 196) = in##B; \
*(uint32_t *) (c + 200) = in##C; \
*(uint32_t *) (c + 204) = in##D
/* store data ; this macro first transpose data in-registers, and then store
* them in memory. much faster with icc. */
#define ONEQUAD_TRANSPOSE(A, B, C, D) \
z##A = _mm_add_epi32(z##A, orig##A); \
z##B = _mm_add_epi32(z##B, orig##B); \
z##C = _mm_add_epi32(z##C, orig##C); \
z##D = _mm_add_epi32(z##D, orig##D); \
y##A = _mm_unpacklo_epi32(z##A, z##B); \
y##B = _mm_unpacklo_epi32(z##C, z##D); \
y##C = _mm_unpackhi_epi32(z##A, z##B); \
y##D = _mm_unpackhi_epi32(z##C, z##D); \
z##A = _mm_unpacklo_epi64(y##A, y##B); \
z##B = _mm_unpackhi_epi64(y##A, y##B); \
z##C = _mm_unpacklo_epi64(y##C, y##D); \
z##D = _mm_unpackhi_epi64(y##C, y##D); \
y##A = _mm_xor_si128(z##A, _mm_loadu_si128((const __m128i *) (m + 0))); \
_mm_storeu_si128((__m128i *) (c + 0), y##A); \
y##B = _mm_xor_si128(z##B, _mm_loadu_si128((const __m128i *) (m + 64))); \
_mm_storeu_si128((__m128i *) (c + 64), y##B); \
y##C = _mm_xor_si128(z##C, _mm_loadu_si128((const __m128i *) (m + 128))); \
_mm_storeu_si128((__m128i *) (c + 128), y##C); \
y##D = _mm_xor_si128(z##D, _mm_loadu_si128((const __m128i *) (m + 192))); \
_mm_storeu_si128((__m128i *) (c + 192), y##D)
#define ONEQUAD(A, B, C, D) ONEQUAD_TRANSPOSE(A, B, C, D)
ONEQUAD(0, 1, 2, 3);
m += 16;
c += 16;
ONEQUAD(4, 5, 6, 7);
m += 16;
c += 16;
ONEQUAD(8, 9, 10, 11);
m += 16;
c += 16;
ONEQUAD(12, 13, 14, 15);
m -= 48;
c -= 48;
#undef ONEQUAD
#undef ONEQUAD_TRANSPOSE
#undef ONEQUAD_EXTRACT
#undef ONEQUAD_SHUFFLE
bytes -= 256;
c += 256;
m += 256;
}
}

477
src/crypto/astrobwt/xmm6int/u8.h Normal file
View file

@ -0,0 +1,477 @@
if (bytes >= 512) {
__m256i y0, y1, y2, y3, y4, y5, y6, y7, y8, y9, y10, y11, y12, y13, y14,
y15;
/* the naive way seems as fast (if not a bit faster) than the vector way */
__m256i z0 = _mm256_set1_epi32(x[0]);
__m256i z5 = _mm256_set1_epi32(x[1]);
__m256i z10 = _mm256_set1_epi32(x[2]);
__m256i z15 = _mm256_set1_epi32(x[3]);
__m256i z12 = _mm256_set1_epi32(x[4]);
__m256i z1 = _mm256_set1_epi32(x[5]);
__m256i z6 = _mm256_set1_epi32(x[6]);
__m256i z11 = _mm256_set1_epi32(x[7]);
__m256i z8; /* useless */
__m256i z13 = _mm256_set1_epi32(x[9]);
__m256i z2 = _mm256_set1_epi32(x[10]);
__m256i z7 = _mm256_set1_epi32(x[11]);
__m256i z4 = _mm256_set1_epi32(x[12]);
__m256i z9; /* useless */
__m256i z14 = _mm256_set1_epi32(x[14]);
__m256i z3 = _mm256_set1_epi32(x[15]);
__m256i orig0 = z0;
__m256i orig1 = z1;
__m256i orig2 = z2;
__m256i orig3 = z3;
__m256i orig4 = z4;
__m256i orig5 = z5;
__m256i orig6 = z6;
__m256i orig7 = z7;
__m256i orig8;
__m256i orig9;
__m256i orig10 = z10;
__m256i orig11 = z11;
__m256i orig12 = z12;
__m256i orig13 = z13;
__m256i orig14 = z14;
__m256i orig15 = z15;
uint32_t in8;
uint32_t in9;
int i;
while (bytes >= 512) {
/* vector implementation for z8 and z9 */
/* faster than the naive version for 8 blocks */
const __m256i addv8 = _mm256_set_epi64x(3, 2, 1, 0);
const __m256i addv9 = _mm256_set_epi64x(7, 6, 5, 4);
const __m256i permute = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
__m256i t8, t9;
uint64_t in89;
in8 = x[8];
in9 = x[13]; /* see arrays above for the address translation */
in89 = ((uint64_t) in8) | (((uint64_t) in9) << 32);
z8 = z9 = _mm256_broadcastq_epi64(_mm_cvtsi64_si128(in89));
t8 = _mm256_add_epi64(addv8, z8);
t9 = _mm256_add_epi64(addv9, z9);
z8 = _mm256_unpacklo_epi32(t8, t9);
z9 = _mm256_unpackhi_epi32(t8, t9);
t8 = _mm256_unpacklo_epi32(z8, z9);
t9 = _mm256_unpackhi_epi32(z8, z9);
/* required because unpack* are intra-lane */
z8 = _mm256_permutevar8x32_epi32(t8, permute);
z9 = _mm256_permutevar8x32_epi32(t9, permute);
orig8 = z8;
orig9 = z9;
in89 += 8;
x[8] = in89 & 0xFFFFFFFF;
x[13] = (in89 >> 32) & 0xFFFFFFFF;
z5 = orig5;
z10 = orig10;
z15 = orig15;
z14 = orig14;
z3 = orig3;
z6 = orig6;
z11 = orig11;
z1 = orig1;
z7 = orig7;
z13 = orig13;
z2 = orig2;
z9 = orig9;
z0 = orig0;
z12 = orig12;
z4 = orig4;
z8 = orig8;
for (i = 0; i < ROUNDS; i += 2) {
/* the inner loop is a direct translation (regexp search/replace)
* from the amd64-xmm6 ASM */
__m256i r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13,
r14, r15;
y4 = z12;
y4 = _mm256_add_epi32(y4, z0);
r4 = y4;
y4 = _mm256_slli_epi32(y4, 7);
z4 = _mm256_xor_si256(z4, y4);
r4 = _mm256_srli_epi32(r4, 25);
z4 = _mm256_xor_si256(z4, r4);
y9 = z1;
y9 = _mm256_add_epi32(y9, z5);
r9 = y9;
y9 = _mm256_slli_epi32(y9, 7);
z9 = _mm256_xor_si256(z9, y9);
r9 = _mm256_srli_epi32(r9, 25);
z9 = _mm256_xor_si256(z9, r9);
y8 = z0;
y8 = _mm256_add_epi32(y8, z4);
r8 = y8;
y8 = _mm256_slli_epi32(y8, 9);
z8 = _mm256_xor_si256(z8, y8);
r8 = _mm256_srli_epi32(r8, 23);
z8 = _mm256_xor_si256(z8, r8);
y13 = z5;
y13 = _mm256_add_epi32(y13, z9);
r13 = y13;
y13 = _mm256_slli_epi32(y13, 9);
z13 = _mm256_xor_si256(z13, y13);
r13 = _mm256_srli_epi32(r13, 23);
z13 = _mm256_xor_si256(z13, r13);
y12 = z4;
y12 = _mm256_add_epi32(y12, z8);
r12 = y12;
y12 = _mm256_slli_epi32(y12, 13);
z12 = _mm256_xor_si256(z12, y12);
r12 = _mm256_srli_epi32(r12, 19);
z12 = _mm256_xor_si256(z12, r12);
y1 = z9;
y1 = _mm256_add_epi32(y1, z13);
r1 = y1;
y1 = _mm256_slli_epi32(y1, 13);
z1 = _mm256_xor_si256(z1, y1);
r1 = _mm256_srli_epi32(r1, 19);
z1 = _mm256_xor_si256(z1, r1);
y0 = z8;
y0 = _mm256_add_epi32(y0, z12);
r0 = y0;
y0 = _mm256_slli_epi32(y0, 18);
z0 = _mm256_xor_si256(z0, y0);
r0 = _mm256_srli_epi32(r0, 14);
z0 = _mm256_xor_si256(z0, r0);
y5 = z13;
y5 = _mm256_add_epi32(y5, z1);
r5 = y5;
y5 = _mm256_slli_epi32(y5, 18);
z5 = _mm256_xor_si256(z5, y5);
r5 = _mm256_srli_epi32(r5, 14);
z5 = _mm256_xor_si256(z5, r5);
y14 = z6;
y14 = _mm256_add_epi32(y14, z10);
r14 = y14;
y14 = _mm256_slli_epi32(y14, 7);
z14 = _mm256_xor_si256(z14, y14);
r14 = _mm256_srli_epi32(r14, 25);
z14 = _mm256_xor_si256(z14, r14);
y3 = z11;
y3 = _mm256_add_epi32(y3, z15);
r3 = y3;
y3 = _mm256_slli_epi32(y3, 7);
z3 = _mm256_xor_si256(z3, y3);
r3 = _mm256_srli_epi32(r3, 25);
z3 = _mm256_xor_si256(z3, r3);
y2 = z10;
y2 = _mm256_add_epi32(y2, z14);
r2 = y2;
y2 = _mm256_slli_epi32(y2, 9);
z2 = _mm256_xor_si256(z2, y2);
r2 = _mm256_srli_epi32(r2, 23);
z2 = _mm256_xor_si256(z2, r2);
y7 = z15;
y7 = _mm256_add_epi32(y7, z3);
r7 = y7;
y7 = _mm256_slli_epi32(y7, 9);
z7 = _mm256_xor_si256(z7, y7);
r7 = _mm256_srli_epi32(r7, 23);
z7 = _mm256_xor_si256(z7, r7);
y6 = z14;
y6 = _mm256_add_epi32(y6, z2);
r6 = y6;
y6 = _mm256_slli_epi32(y6, 13);
z6 = _mm256_xor_si256(z6, y6);
r6 = _mm256_srli_epi32(r6, 19);
z6 = _mm256_xor_si256(z6, r6);
y11 = z3;
y11 = _mm256_add_epi32(y11, z7);
r11 = y11;
y11 = _mm256_slli_epi32(y11, 13);
z11 = _mm256_xor_si256(z11, y11);
r11 = _mm256_srli_epi32(r11, 19);
z11 = _mm256_xor_si256(z11, r11);
y10 = z2;
y10 = _mm256_add_epi32(y10, z6);
r10 = y10;
y10 = _mm256_slli_epi32(y10, 18);
z10 = _mm256_xor_si256(z10, y10);
r10 = _mm256_srli_epi32(r10, 14);
z10 = _mm256_xor_si256(z10, r10);
y1 = z3;
y1 = _mm256_add_epi32(y1, z0);
r1 = y1;
y1 = _mm256_slli_epi32(y1, 7);
z1 = _mm256_xor_si256(z1, y1);
r1 = _mm256_srli_epi32(r1, 25);
z1 = _mm256_xor_si256(z1, r1);
y15 = z7;
y15 = _mm256_add_epi32(y15, z11);
r15 = y15;
y15 = _mm256_slli_epi32(y15, 18);
z15 = _mm256_xor_si256(z15, y15);
r15 = _mm256_srli_epi32(r15, 14);
z15 = _mm256_xor_si256(z15, r15);
y6 = z4;
y6 = _mm256_add_epi32(y6, z5);
r6 = y6;
y6 = _mm256_slli_epi32(y6, 7);
z6 = _mm256_xor_si256(z6, y6);
r6 = _mm256_srli_epi32(r6, 25);
z6 = _mm256_xor_si256(z6, r6);
y2 = z0;
y2 = _mm256_add_epi32(y2, z1);
r2 = y2;
y2 = _mm256_slli_epi32(y2, 9);
z2 = _mm256_xor_si256(z2, y2);
r2 = _mm256_srli_epi32(r2, 23);
z2 = _mm256_xor_si256(z2, r2);
y7 = z5;
y7 = _mm256_add_epi32(y7, z6);
r7 = y7;
y7 = _mm256_slli_epi32(y7, 9);
z7 = _mm256_xor_si256(z7, y7);
r7 = _mm256_srli_epi32(r7, 23);
z7 = _mm256_xor_si256(z7, r7);
y3 = z1;
y3 = _mm256_add_epi32(y3, z2);
r3 = y3;
y3 = _mm256_slli_epi32(y3, 13);
z3 = _mm256_xor_si256(z3, y3);
r3 = _mm256_srli_epi32(r3, 19);
z3 = _mm256_xor_si256(z3, r3);
y4 = z6;
y4 = _mm256_add_epi32(y4, z7);
r4 = y4;
y4 = _mm256_slli_epi32(y4, 13);
z4 = _mm256_xor_si256(z4, y4);
r4 = _mm256_srli_epi32(r4, 19);
z4 = _mm256_xor_si256(z4, r4);
y0 = z2;
y0 = _mm256_add_epi32(y0, z3);
r0 = y0;
y0 = _mm256_slli_epi32(y0, 18);
z0 = _mm256_xor_si256(z0, y0);
r0 = _mm256_srli_epi32(r0, 14);
z0 = _mm256_xor_si256(z0, r0);
y5 = z7;
y5 = _mm256_add_epi32(y5, z4);
r5 = y5;
y5 = _mm256_slli_epi32(y5, 18);
z5 = _mm256_xor_si256(z5, y5);
r5 = _mm256_srli_epi32(r5, 14);
z5 = _mm256_xor_si256(z5, r5);
y11 = z9;
y11 = _mm256_add_epi32(y11, z10);
r11 = y11;
y11 = _mm256_slli_epi32(y11, 7);
z11 = _mm256_xor_si256(z11, y11);
r11 = _mm256_srli_epi32(r11, 25);
z11 = _mm256_xor_si256(z11, r11);
y12 = z14;
y12 = _mm256_add_epi32(y12, z15);
r12 = y12;
y12 = _mm256_slli_epi32(y12, 7);
z12 = _mm256_xor_si256(z12, y12);
r12 = _mm256_srli_epi32(r12, 25);
z12 = _mm256_xor_si256(z12, r12);
y8 = z10;
y8 = _mm256_add_epi32(y8, z11);
r8 = y8;
y8 = _mm256_slli_epi32(y8, 9);
z8 = _mm256_xor_si256(z8, y8);
r8 = _mm256_srli_epi32(r8, 23);
z8 = _mm256_xor_si256(z8, r8);
y13 = z15;
y13 = _mm256_add_epi32(y13, z12);
r13 = y13;
y13 = _mm256_slli_epi32(y13, 9);
z13 = _mm256_xor_si256(z13, y13);
r13 = _mm256_srli_epi32(r13, 23);
z13 = _mm256_xor_si256(z13, r13);
y9 = z11;
y9 = _mm256_add_epi32(y9, z8);
r9 = y9;
y9 = _mm256_slli_epi32(y9, 13);
z9 = _mm256_xor_si256(z9, y9);
r9 = _mm256_srli_epi32(r9, 19);
z9 = _mm256_xor_si256(z9, r9);
y14 = z12;
y14 = _mm256_add_epi32(y14, z13);
r14 = y14;
y14 = _mm256_slli_epi32(y14, 13);
z14 = _mm256_xor_si256(z14, y14);
r14 = _mm256_srli_epi32(r14, 19);
z14 = _mm256_xor_si256(z14, r14);
y10 = z8;
y10 = _mm256_add_epi32(y10, z9);
r10 = y10;
y10 = _mm256_slli_epi32(y10, 18);
z10 = _mm256_xor_si256(z10, y10);
r10 = _mm256_srli_epi32(r10, 14);
z10 = _mm256_xor_si256(z10, r10);
y15 = z13;
y15 = _mm256_add_epi32(y15, z14);
r15 = y15;
y15 = _mm256_slli_epi32(y15, 18);
z15 = _mm256_xor_si256(z15, y15);
r15 = _mm256_srli_epi32(r15, 14);
z15 = _mm256_xor_si256(z15, r15);
}
/* store data ; this macro first transpose data in-registers, and then store
* them in memory. much faster with icc. */
#define ONEQUAD_TRANSPOSE(A, B, C, D) \
{ \
__m128i t0, t1, t2, t3; \
z##A = _mm256_add_epi32(z##A, orig##A); \
z##B = _mm256_add_epi32(z##B, orig##B); \
z##C = _mm256_add_epi32(z##C, orig##C); \
z##D = _mm256_add_epi32(z##D, orig##D); \
y##A = _mm256_unpacklo_epi32(z##A, z##B); \
y##B = _mm256_unpacklo_epi32(z##C, z##D); \
y##C = _mm256_unpackhi_epi32(z##A, z##B); \
y##D = _mm256_unpackhi_epi32(z##C, z##D); \
z##A = _mm256_unpacklo_epi64(y##A, y##B); \
z##B = _mm256_unpackhi_epi64(y##A, y##B); \
z##C = _mm256_unpacklo_epi64(y##C, y##D); \
z##D = _mm256_unpackhi_epi64(y##C, y##D); \
t0 = _mm_xor_si128(_mm256_extracti128_si256(z##A, 0), \
_mm_loadu_si128((const __m128i*) (m + 0))); \
_mm_storeu_si128((__m128i*) (c + 0), t0); \
t1 = _mm_xor_si128(_mm256_extracti128_si256(z##B, 0), \
_mm_loadu_si128((const __m128i*) (m + 64))); \
_mm_storeu_si128((__m128i*) (c + 64), t1); \
t2 = _mm_xor_si128(_mm256_extracti128_si256(z##C, 0), \
_mm_loadu_si128((const __m128i*) (m + 128))); \
_mm_storeu_si128((__m128i*) (c + 128), t2); \
t3 = _mm_xor_si128(_mm256_extracti128_si256(z##D, 0), \
_mm_loadu_si128((const __m128i*) (m + 192))); \
_mm_storeu_si128((__m128i*) (c + 192), t3); \
t0 = _mm_xor_si128(_mm256_extracti128_si256(z##A, 1), \
_mm_loadu_si128((const __m128i*) (m + 256))); \
_mm_storeu_si128((__m128i*) (c + 256), t0); \
t1 = _mm_xor_si128(_mm256_extracti128_si256(z##B, 1), \
_mm_loadu_si128((const __m128i*) (m + 320))); \
_mm_storeu_si128((__m128i*) (c + 320), t1); \
t2 = _mm_xor_si128(_mm256_extracti128_si256(z##C, 1), \
_mm_loadu_si128((const __m128i*) (m + 384))); \
_mm_storeu_si128((__m128i*) (c + 384), t2); \
t3 = _mm_xor_si128(_mm256_extracti128_si256(z##D, 1), \
_mm_loadu_si128((const __m128i*) (m + 448))); \
_mm_storeu_si128((__m128i*) (c + 448), t3); \
}
#define ONEQUAD(A, B, C, D) ONEQUAD_TRANSPOSE(A, B, C, D)
#define ONEQUAD_UNPCK(A, B, C, D) \
{ \
z##A = _mm256_add_epi32(z##A, orig##A); \
z##B = _mm256_add_epi32(z##B, orig##B); \
z##C = _mm256_add_epi32(z##C, orig##C); \
z##D = _mm256_add_epi32(z##D, orig##D); \
y##A = _mm256_unpacklo_epi32(z##A, z##B); \
y##B = _mm256_unpacklo_epi32(z##C, z##D); \
y##C = _mm256_unpackhi_epi32(z##A, z##B); \
y##D = _mm256_unpackhi_epi32(z##C, z##D); \
z##A = _mm256_unpacklo_epi64(y##A, y##B); \
z##B = _mm256_unpackhi_epi64(y##A, y##B); \
z##C = _mm256_unpacklo_epi64(y##C, y##D); \
z##D = _mm256_unpackhi_epi64(y##C, y##D); \
}
#define ONEOCTO(A, B, C, D, A2, B2, C2, D2) \
{ \
ONEQUAD_UNPCK(A, B, C, D); \
ONEQUAD_UNPCK(A2, B2, C2, D2); \
y##A = _mm256_permute2x128_si256(z##A, z##A2, 0x20); \
y##A2 = _mm256_permute2x128_si256(z##A, z##A2, 0x31); \
y##B = _mm256_permute2x128_si256(z##B, z##B2, 0x20); \
y##B2 = _mm256_permute2x128_si256(z##B, z##B2, 0x31); \
y##C = _mm256_permute2x128_si256(z##C, z##C2, 0x20); \
y##C2 = _mm256_permute2x128_si256(z##C, z##C2, 0x31); \
y##D = _mm256_permute2x128_si256(z##D, z##D2, 0x20); \
y##D2 = _mm256_permute2x128_si256(z##D, z##D2, 0x31); \
y##A = _mm256_xor_si256(y##A, \
_mm256_loadu_si256((const __m256i*) (m + 0))); \
y##B = _mm256_xor_si256( \
y##B, _mm256_loadu_si256((const __m256i*) (m + 64))); \
y##C = _mm256_xor_si256( \
y##C, _mm256_loadu_si256((const __m256i*) (m + 128))); \
y##D = _mm256_xor_si256( \
y##D, _mm256_loadu_si256((const __m256i*) (m + 192))); \
y##A2 = _mm256_xor_si256( \
y##A2, _mm256_loadu_si256((const __m256i*) (m + 256))); \
y##B2 = _mm256_xor_si256( \
y##B2, _mm256_loadu_si256((const __m256i*) (m + 320))); \
y##C2 = _mm256_xor_si256( \
y##C2, _mm256_loadu_si256((const __m256i*) (m + 384))); \
y##D2 = _mm256_xor_si256( \
y##D2, _mm256_loadu_si256((const __m256i*) (m + 448))); \
_mm256_storeu_si256((__m256i*) (c + 0), y##A); \
_mm256_storeu_si256((__m256i*) (c + 64), y##B); \
_mm256_storeu_si256((__m256i*) (c + 128), y##C); \
_mm256_storeu_si256((__m256i*) (c + 192), y##D); \
_mm256_storeu_si256((__m256i*) (c + 256), y##A2); \
_mm256_storeu_si256((__m256i*) (c + 320), y##B2); \
_mm256_storeu_si256((__m256i*) (c + 384), y##C2); \
_mm256_storeu_si256((__m256i*) (c + 448), y##D2); \
}
ONEOCTO(0, 1, 2, 3, 4, 5, 6, 7);
m += 32;
c += 32;
ONEOCTO(8, 9, 10, 11, 12, 13, 14, 15);
m -= 32;
c -= 32;
#undef ONEQUAD
#undef ONEQUAD_TRANSPOSE
#undef ONEQUAD_UNPCK
#undef ONEOCTO
bytes -= 512;
c += 512;
m += 512;
}
}