popov/cpuminer-opt-gpu
1
0
Fork 0
mirror of https://github.com/JayDDee/cpuminer-opt.git synced 2025-09-17 23:44:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

v3.5.8

Browse Source
This commit is contained in:
Jay D Dee
2017-02-16 10:51:20 -05:00
parent 9ee5965e3f
commit 4521b324e3
9 changed files with 94 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
algo/cryptonight/cryptonight-aesni.c
View File

@@ -126,25 +126,24 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
// aesni_parallel_noxor(&ctx->long_state[i], ctx->text, ExpandedKey);
// prefetch expkey, all of xmminput and enough longoutput for 4 loops
_mm_prefetch( expkey, _MM_HINT_T0 );
_mm_prefetch( expkey + 4, _MM_HINT_T0 );
_mm_prefetch( expkey + 8, _MM_HINT_T0 );
_mm_prefetch( xmminput, _MM_HINT_T0 );
_mm_prefetch( xmminput + 4, _MM_HINT_T0 );
for ( i = 0; i < 64; i += 8 )
for ( i = 0; i < 64; i += 16 )
{
_mm_prefetch( longoutput + i, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 4, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 8, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 12, _MM_HINT_T0 );
}
_mm_prefetch( expkey, _MM_HINT_T0 );
_mm_prefetch( expkey + 4, _MM_HINT_T0 );
_mm_prefetch( expkey + 8, _MM_HINT_T0 );
for ( i = 0; likely( i < MEMORY_M128I ); i += INIT_SIZE_M128I )
{
// prefetch 4 loops ahead,
_mm_prefetch( longoutput + i + 64, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 68, _MM_HINT_T0 );
__builtin_prefetch( longoutput + i + 64, 1, 0 );
__builtin_prefetch( longoutput + i + 68, 1, 0 );
for (j = 0; j < 10; j++ )
{
@@ -191,7 +190,7 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
for(i = 0; __builtin_expect(i < 0x80000, 1); i++)
{
uint64_t c[2];
_mm_prefetch( &ctx.long_state[c[0] & 0x1FFFF0], _MM_HINT_T0 );
__builtin_prefetch( &ctx.long_state[c[0] & 0x1FFFF0], 0, 1 );
__m128i c_x = _mm_load_si128(
(__m128i *)&ctx.long_state[a[0] & 0x1FFFF0]);
@@ -232,7 +231,7 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
a[0] ^= b[0];
a[1] ^= b[1];
b_x = c_x;
_mm_prefetch( &ctx.long_state[a[0] & 0x1FFFF0], _MM_HINT_T0 );
__builtin_prefetch( &ctx.long_state[a[0] & 0x1FFFF0], 0, 3 );
}
memcpy( ctx.text, ctx.state.init, INIT_SIZE_BYTE );
@@ -243,6 +242,7 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
// aesni_parallel_xor(&ctx->text, ExpandedKey, &ctx->long_state[i]);
// prefetch expkey, all of xmminput and enough longoutput for 4 loops
_mm_prefetch( xmminput, _MM_HINT_T0 );
_mm_prefetch( xmminput + 4, _MM_HINT_T0 );
for ( i = 0; i < 64; i += 16 )

32
algo/cryptonight/cryptonight.c
View File

@@ -5,6 +5,7 @@
// Modified for CPUminer by Lucas Jones
#include "miner.h"
#include <memory.h>
#if defined(__arm__) || defined(_MSC_VER)
#ifndef NOASM
@@ -175,11 +176,27 @@ void cryptonight_hash_ctx(void* output, const void* input, int len)
{
hash_process(&ctx.state.hs, (const uint8_t*) input, len);
ctx.aes_ctx = (oaes_ctx*) oaes_alloc();
__builtin_prefetch( ctx.text, 0, 3 );
__builtin_prefetch( ctx.text + 64, 0, 3 );
__builtin_prefetch( ctx.long_state, 1, 0 );
__builtin_prefetch( ctx.long_state + 64, 1, 0 );
__builtin_prefetch( ctx.long_state + 128, 1, 0 );
__builtin_prefetch( ctx.long_state + 192, 1, 0 );
__builtin_prefetch( ctx.long_state + 256, 1, 0 );
__builtin_prefetch( ctx.long_state + 320, 1, 0 );
__builtin_prefetch( ctx.long_state + 384, 1, 0 );
__builtin_prefetch( ctx.long_state + 448, 1, 0 );
size_t i, j;
memcpy(ctx.text, ctx.state.init, INIT_SIZE_BYTE);
oaes_key_import_data(ctx.aes_ctx, ctx.state.hs.b, AES_KEY_SIZE);
for (i = 0; likely(i < MEMORY); i += INIT_SIZE_BYTE) {
__builtin_prefetch( ctx.long_state + i + 512, 1, 0 );
__builtin_prefetch( ctx.long_state + i + 576, 1, 0 );
aesb_pseudo_round_mut(&ctx.text[AES_BLOCK_SIZE * 0], ctx.aes_ctx->key->exp_data);
aesb_pseudo_round_mut(&ctx.text[AES_BLOCK_SIZE * 1], ctx.aes_ctx->key->exp_data);
aesb_pseudo_round_mut(&ctx.text[AES_BLOCK_SIZE * 2], ctx.aes_ctx->key->exp_data);
@@ -213,9 +230,24 @@ void cryptonight_hash_ctx(void* output, const void* input, int len)
mul_sum_xor_dst(ctx.b, ctx.a, &ctx.long_state[e2i(ctx.b)]);
}
__builtin_prefetch( ctx.text, 0, 3 );
__builtin_prefetch( ctx.text + 64, 0, 3 );
__builtin_prefetch( ctx.long_state, 1, 0 );
__builtin_prefetch( ctx.long_state + 64, 1, 0 );
__builtin_prefetch( ctx.long_state + 128, 1, 0 );
__builtin_prefetch( ctx.long_state + 192, 1, 0 );
__builtin_prefetch( ctx.long_state + 256, 1, 0 );
__builtin_prefetch( ctx.long_state + 320, 1, 0 );
__builtin_prefetch( ctx.long_state + 384, 1, 0 );
__builtin_prefetch( ctx.long_state + 448, 1, 0 );
memcpy(ctx.text, ctx.state.init, INIT_SIZE_BYTE);
oaes_key_import_data(ctx.aes_ctx, &ctx.state.hs.b[32], AES_KEY_SIZE);
for (i = 0; likely(i < MEMORY); i += INIT_SIZE_BYTE) {
__builtin_prefetch( ctx.long_state + i + 512, 1, 0 );
__builtin_prefetch( ctx.long_state + i + 576, 1, 0 );
xor_blocks(&ctx.text[0 * AES_BLOCK_SIZE], &ctx.long_state[i + 0 * AES_BLOCK_SIZE]);
aesb_pseudo_round_mut(&ctx.text[0 * AES_BLOCK_SIZE], ctx.aes_ctx->key->exp_data);
xor_blocks(&ctx.text[1 * AES_BLOCK_SIZE], &ctx.long_state[i + 1 * AES_BLOCK_SIZE]);