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

v3.23.2

Browse Source
This commit is contained in:
Jay D Dee
2023-09-21 12:34:06 -04:00
parent d6b5750362
commit be88afc349
113 changed files with 3349 additions and 2920 deletions
Download Patch File Download Diff File Expand all files Collapse all files

323
algo/sha/sha256d-4way.c
View File

@@ -4,7 +4,6 @@
#include <string.h>
#include <stdio.h>
#include "sha256-hash.h"
#include "sha-hash-4way.h"
static const uint32_t sha256_iv[8] __attribute__ ((aligned (32))) =
{
@@ -17,11 +16,15 @@ static const uint32_t sha256_iv[8] __attribute__ ((aligned (32))) =
int scanhash_sha256d_sha( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t block0[16] __attribute__ ((aligned (64)));
uint32_t block1[16] __attribute__ ((aligned (64)));
uint32_t hash0[8] __attribute__ ((aligned (32)));
uint32_t hash1[8] __attribute__ ((aligned (32)));
uint32_t mstate[8] __attribute__ ((aligned (32)));
uint32_t block1a[16] __attribute__ ((aligned (64)));
uint32_t block1b[16] __attribute__ ((aligned (64)));
uint32_t block2a[16] __attribute__ ((aligned (64)));
uint32_t block2b[16] __attribute__ ((aligned (64)));
uint32_t hasha[8] __attribute__ ((aligned (32)));
uint32_t hashb[8] __attribute__ ((aligned (32)));
uint32_t mstatea[8] __attribute__ ((aligned (32)));
uint32_t mstateb[8] __attribute__ ((aligned (32)));
uint32_t sstate[8] __attribute__ ((aligned (32)));
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t first_nonce = pdata[19];
@@ -32,56 +35,60 @@ int scanhash_sha256d_sha( struct work *work, uint32_t max_nonce,
const __m128i shuf_bswap32 =
_mm_set_epi64x( 0x0c0d0e0f08090a0bULL, 0x0405060700010203ULL );
// hash first 64 bytes of data
sha256_opt_transform_le( mstate, pdata, sha256_iv );
// hash first 64 byte block of data
sha256_opt_transform_le( mstatea, pdata, sha256_iv );
// fill & pad second bock without nonce
memcpy( block1a, pdata + 16, 12 );
memcpy( block1b, pdata + 16, 12 );
block1a[ 3] = 0;
block1b[ 3] = 0;
block1a[ 4] = block1b[ 4] = 0x80000000;
memset( block1a + 5, 0, 40 );
memset( block1b + 5, 0, 40 );
block1a[15] = block1b[15] = 80*8; // bit count
sha256_ni_prehash_3rounds( mstateb, block1a, sstate, mstatea);
// Pad third block
block2a[ 8] = block2b[ 8] = 0x80000000;
memset( block2a + 9, 0, 24 );
memset( block2b + 9, 0, 24 );
block2a[15] = block2b[15] = 32*8; // bit count
do
{
// 1. final 16 bytes of data, with padding
memcpy( block0, pdata + 16, 16 );
memcpy( block1, pdata + 16, 16 );
block0[ 3] = n;
block1[ 3] = n+1;
block0[ 4] = block1[ 4] = 0x80000000;
memset( block0 + 5, 0, 40 );
memset( block1 + 5, 0, 40 );
block0[15] = block1[15] = 80*8; // bit count
sha256_ni2way_transform_le( hash0, hash1, block0, block1,
mstate, mstate );
// Insert nonce for second block
block1a[3] = n;
block1b[3] = n+1;
sha256_ni2way_final_rounds( block2a, block2b, block1a, block1b,
mstateb, mstateb, sstate, sstate );
// 2. 32 byte hash from 1.
memcpy( block0, hash0, 32 );
memcpy( block1, hash1, 32 );
block0[ 8] = block1[ 8] = 0x80000000;
memset( block0 + 9, 0, 24 );
memset( block1 + 9, 0, 24 );
block0[15] = block1[15] = 32*8; // bit count
sha256_ni2way_transform_le( hash0, hash1, block0, block1,
sha256_ni2way_transform_le( hasha, hashb, block2a, block2b,
sha256_iv, sha256_iv );
if ( unlikely( bswap_32( hash0[7] ) <= ptarget[7] ) )
if ( unlikely( bswap_32( hasha[7] ) <= ptarget[7] ) )
{
casti_m128i( hash0, 0 ) =
_mm_shuffle_epi8( casti_m128i( hash0, 0 ), shuf_bswap32 );
casti_m128i( hash0, 1 ) =
_mm_shuffle_epi8( casti_m128i( hash0, 1 ), shuf_bswap32 );
if ( likely( valid_hash( hash0, ptarget ) && !bench ) )
casti_m128i( hasha, 0 ) =
_mm_shuffle_epi8( casti_m128i( hasha, 0 ), shuf_bswap32 );
casti_m128i( hasha, 1 ) =
_mm_shuffle_epi8( casti_m128i( hasha, 1 ), shuf_bswap32 );
if ( likely( valid_hash( hasha, ptarget ) && !bench ) )
{
pdata[19] = n;
submit_solution( work, hash0, mythr );
submit_solution( work, hasha, mythr );
}
}
if ( unlikely( bswap_32( hash1[7] ) <= ptarget[7] ) )
if ( unlikely( bswap_32( hashb[7] ) <= ptarget[7] ) )
{
casti_m128i( hash1, 0 ) =
_mm_shuffle_epi8( casti_m128i( hash1, 0 ), shuf_bswap32 );
casti_m128i( hash1, 1 ) =
_mm_shuffle_epi8( casti_m128i( hash1, 1 ), shuf_bswap32 );
if ( likely( valid_hash( hash1, ptarget ) && !bench ) )
casti_m128i( hashb, 0 ) =
_mm_shuffle_epi8( casti_m128i( hashb, 0 ), shuf_bswap32 );
casti_m128i( hashb, 1 ) =
_mm_shuffle_epi8( casti_m128i( hashb, 1 ), shuf_bswap32 );
if ( likely( valid_hash( hashb, ptarget ) && !bench ) )
{
pdata[19] = n+1;
submit_solution( work, hash1, mythr );
submit_solution( work, hashb, mythr );
}
}
n += 2;
@@ -99,18 +106,16 @@ int scanhash_sha256d_sha( struct work *work, uint32_t max_nonce,
int scanhash_sha256d_16way( struct work *work, const uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
__m512i hash32[8] __attribute__ ((aligned (128)));
__m512i block[16] __attribute__ ((aligned (64)));
__m512i block[16] __attribute__ ((aligned (128)));
__m512i buf[16] __attribute__ ((aligned (64)));
__m512i hash32[8] __attribute__ ((aligned (64)));
__m512i mstate1[8] __attribute__ ((aligned (64)));
__m512i mstate2[8] __attribute__ ((aligned (64)));
__m512i istate[8] __attribute__ ((aligned (64)));
__m512i mexp_pre[8] __attribute__ ((aligned (64)));
uint32_t phash[8] __attribute__ ((aligned (32)));
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
uint32_t *hash32_d7 = (uint32_t*)&(hash32[7]);
const uint32_t targ32_d7 = ptarget[7];
const uint32_t *ptarget = work->target;
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 16;
const __m512i last_byte = _mm512_set1_epi32( 0x80000000 );
@@ -134,7 +139,7 @@ int scanhash_sha256d_16way( struct work *work, const uint32_t max_nonce,
mstate1[6] = _mm512_set1_epi32( phash[6] );
mstate1[7] = _mm512_set1_epi32( phash[7] );
// second message block data, with nonce & padding
// second message block data, with nonce & padding
buf[0] = _mm512_set1_epi32( pdata[16] );
buf[1] = _mm512_set1_epi32( pdata[17] );
buf[2] = _mm512_set1_epi32( pdata[18] );
@@ -142,12 +147,12 @@ int scanhash_sha256d_16way( struct work *work, const uint32_t max_nonce,
n+ 7, n+ 6, n+ 5, n+ 4, n+ 3, n+ 2, n +1, n );
buf[4] = last_byte;
memset_zero_512( buf+5, 10 );
buf[15] = _mm512_set1_epi32( 80*8 ); // bit count
buf[15] = _mm512_set1_epi32( 80*8 ); // bit count
// partially pre-expand & prehash second message block, avoiding the nonces
sha256_16way_prehash_3rounds( mstate2, mexp_pre, buf, mstate1 );
// vectorize IV for 2nd & 3rd sha256
// vectorize IV for second hash
istate[0] = _mm512_set1_epi32( sha256_iv[0] );
istate[1] = _mm512_set1_epi32( sha256_iv[1] );
istate[2] = _mm512_set1_epi32( sha256_iv[2] );
@@ -157,27 +162,26 @@ int scanhash_sha256d_16way( struct work *work, const uint32_t max_nonce,
istate[6] = _mm512_set1_epi32( sha256_iv[6] );
istate[7] = _mm512_set1_epi32( sha256_iv[7] );
// initialize padding for 2nd sha256
// initialize padding for second hash
block[ 8] = last_byte;
memset_zero_512( block + 9, 6 );
memset_zero_512( block+9, 6 );
block[15] = _mm512_set1_epi32( 32*8 ); // bit count
do
{
sha256_16way_final_rounds( block, buf, mstate1, mstate2, mexp_pre );
if ( sha256_16way_transform_le_short( hash32, block, istate, ptarget ) )
if ( unlikely( sha256_16way_transform_le_short(
hash32, block, istate, ptarget ) ) )
{
for ( int lane = 0; lane < 16; lane++ )
if ( bswap_32( hash32_d7[ lane ] ) <= targ32_d7 )
{
extr_lane_16x32( phash, hash32, lane, 256 );
casti_m256i( phash, 0 ) =
_mm256_shuffle_epi8( casti_m256i( phash, 0 ), bswap_shuf );
_mm256_shuffle_epi8( casti_m256i( phash, 0 ), bswap_shuf );
if ( likely( valid_hash( phash, ptarget ) && !bench ) )
{
pdata[19] = n + lane;
submit_solution( work, phash, mythr );
pdata[19] = n + lane;
submit_solution( work, phash, mythr );
}
}
}
@@ -188,92 +192,7 @@ int scanhash_sha256d_16way( struct work *work, const uint32_t max_nonce,
*hashes_done = n - first_nonce;
return 0;
}
/*
int scanhash_sha256d_16way( struct work *work, const uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
__m512i vdata[32] __attribute__ ((aligned (128)));
__m512i block[16] __attribute__ ((aligned (64)));
__m512i hash32[8] __attribute__ ((aligned (64)));
__m512i initstate[8] __attribute__ ((aligned (64)));
__m512i midstate1[8] __attribute__ ((aligned (64)));
__m512i midstate2[8] __attribute__ ((aligned (64)));
__m512i mexp_pre[16] __attribute__ ((aligned (64)));
uint32_t lane_hash[8] __attribute__ ((aligned (64)));
uint32_t *hash32_d7 = (uint32_t*)&( hash32[7] );
uint32_t *pdata = work->data;
const uint32_t *ptarget = work->target;
const uint32_t targ32_d7 = ptarget[7];
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 16;
uint32_t n = first_nonce;
__m512i *noncev = vdata + 19;
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
const __m512i last_byte = _mm512_set1_epi32( 0x80000000 );
const __m512i sixteen = _mm512_set1_epi32( 16 );
for ( int i = 0; i < 19; i++ )
vdata[i] = _mm512_set1_epi32( pdata[i] );
*noncev = _mm512_set_epi32( n+15, n+14, n+13, n+12, n+11, n+10, n+9, n+8,
n+ 7, n+ 6, n+ 5, n+ 4, n+ 3, n+ 2, n+1, n );
vdata[16+4] = last_byte;
memset_zero_512( vdata+16 + 5, 10 );
vdata[16+15] = _mm512_set1_epi32( 80*8 ); // bit count
block[ 8] = last_byte;
memset_zero_512( block + 9, 6 );
block[15] = _mm512_set1_epi32( 32*8 ); // bit count
// initialize state
initstate[0] = _mm512_set1_epi64( 0x6A09E6676A09E667 );
initstate[1] = _mm512_set1_epi64( 0xBB67AE85BB67AE85 );
initstate[2] = _mm512_set1_epi64( 0x3C6EF3723C6EF372 );
initstate[3] = _mm512_set1_epi64( 0xA54FF53AA54FF53A );
initstate[4] = _mm512_set1_epi64( 0x510E527F510E527F );
initstate[5] = _mm512_set1_epi64( 0x9B05688C9B05688C );
initstate[6] = _mm512_set1_epi64( 0x1F83D9AB1F83D9AB );
initstate[7] = _mm512_set1_epi64( 0x5BE0CD195BE0CD19 );
sha256_16way_transform_le( midstate1, vdata, initstate );
// Do 3 rounds on the first 12 bytes of the next block
sha256_16way_prehash_3rounds( midstate2, mexp_pre, vdata+16, midstate1 );
do
{
// 1. final 16 bytes of data, with padding
sha256_16way_final_rounds( block, vdata+16, midstate1, midstate2,
mexp_pre );
// 2. 32 byte hash from 1.
sha256_16way_transform_le( hash32, block, initstate );
// byte swap final hash for testing
mm512_block_bswap_32( hash32, hash32 );
for ( int lane = 0; lane < 16; lane++ )
if ( unlikely( hash32_d7[ lane ] <= targ32_d7 ) )
{
extr_lane_16x32( lane_hash, hash32, lane, 256 );
if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
{
pdata[19] = n + lane;
submit_solution( work, lane_hash, mythr );
}
}
*noncev = _mm512_add_epi32( *noncev, sixteen );
n += 16;
} while ( (n < last_nonce) && !work_restart[thr_id].restart );
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}
*/
#endif
#if defined(SHA256D_8WAY)
@@ -284,15 +203,13 @@ int scanhash_sha256d_8way( struct work *work, const uint32_t max_nonce,
__m256i vdata[32] __attribute__ ((aligned (64)));
__m256i block[16] __attribute__ ((aligned (32)));
__m256i hash32[8] __attribute__ ((aligned (32)));
__m256i initstate[8] __attribute__ ((aligned (32)));
__m256i midstate1[8] __attribute__ ((aligned (32)));
__m256i midstate2[8] __attribute__ ((aligned (32)));
__m256i istate[8] __attribute__ ((aligned (32)));
__m256i mstate1[8] __attribute__ ((aligned (32)));
__m256i mstate2[8] __attribute__ ((aligned (32)));
__m256i mexp_pre[8] __attribute__ ((aligned (32)));
uint32_t lane_hash[8] __attribute__ ((aligned (32)));
uint32_t *hash32_d7 = (uint32_t*)&( hash32[7] );
uint32_t *pdata = work->data;
const uint32_t *ptarget = work->target;
const uint32_t targ32_d7 = ptarget[7];
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 8;
uint32_t n = first_nonce;
@@ -301,6 +218,8 @@ int scanhash_sha256d_8way( struct work *work, const uint32_t max_nonce,
const bool bench = opt_benchmark;
const __m256i last_byte = _mm256_set1_epi32( 0x80000000 );
const __m256i eight = _mm256_set1_epi32( 8 );
const __m256i bswap_shuf = mm256_bcast_m128( _mm_set_epi64x(
0x0c0d0e0f08090a0b, 0x0405060700010203 ) );
for ( int i = 0; i < 19; i++ )
vdata[i] = _mm256_set1_epi32( pdata[i] );
@@ -309,50 +228,47 @@ int scanhash_sha256d_8way( struct work *work, const uint32_t max_nonce,
vdata[16+4] = last_byte;
memset_zero_256( vdata+16 + 5, 10 );
vdata[16+15] = _mm256_set1_epi32( 80*8 ); // bit count
vdata[16+15] = _mm256_set1_epi32( 80*8 );
block[ 8] = last_byte;
memset_zero_256( block + 9, 6 );
block[15] = _mm256_set1_epi32( 32*8 ); // bit count
block[15] = _mm256_set1_epi32( 32*8 );
// initialize state
initstate[0] = _mm256_set1_epi64x( 0x6A09E6676A09E667 );
initstate[1] = _mm256_set1_epi64x( 0xBB67AE85BB67AE85 );
initstate[2] = _mm256_set1_epi64x( 0x3C6EF3723C6EF372 );
initstate[3] = _mm256_set1_epi64x( 0xA54FF53AA54FF53A );
initstate[4] = _mm256_set1_epi64x( 0x510E527F510E527F );
initstate[5] = _mm256_set1_epi64x( 0x9B05688C9B05688C );
initstate[6] = _mm256_set1_epi64x( 0x1F83D9AB1F83D9AB );
initstate[7] = _mm256_set1_epi64x( 0x5BE0CD195BE0CD19 );
// initialize state for second hash
istate[0] = _mm256_set1_epi32( sha256_iv[0] );
istate[1] = _mm256_set1_epi32( sha256_iv[1] );
istate[2] = _mm256_set1_epi32( sha256_iv[2] );
istate[3] = _mm256_set1_epi32( sha256_iv[3] );
istate[4] = _mm256_set1_epi32( sha256_iv[4] );
istate[5] = _mm256_set1_epi32( sha256_iv[5] );
istate[6] = _mm256_set1_epi32( sha256_iv[6] );
istate[7] = _mm256_set1_epi32( sha256_iv[7] );
sha256_8way_transform_le( mstate1, vdata, istate );
sha256_8way_transform_le( midstate1, vdata, initstate );
// Do 3 rounds on the first 12 bytes of the next block
sha256_8way_prehash_3rounds( midstate2, mexp_pre, vdata + 16, midstate1 );
sha256_8way_prehash_3rounds( mstate2, mexp_pre, vdata + 16, mstate1 );
do
{
// 1. final 16 bytes of data, with padding
sha256_8way_final_rounds( block, vdata+16, midstate1, midstate2,
mexp_pre );
// 2. 32 byte hash from 1.
sha256_8way_transform_le( hash32, block, initstate );
// byte swap final hash for testing
mm256_block_bswap_32( hash32, hash32 );
for ( int lane = 0; lane < 8; lane++ )
if ( unlikely( hash32_d7[ lane ] <= targ32_d7 ) )
sha256_8way_final_rounds( block, vdata+16, mstate1, mstate2, mexp_pre );
if ( unlikely( sha256_8way_transform_le_short( hash32, block,
istate, ptarget ) ) )
{
extr_lane_8x32( lane_hash, hash32, lane, 256 );
if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
for ( int lane = 0; lane < 8; lane++ )
{
pdata[19] = n + lane;
submit_solution( work, lane_hash, mythr );
extr_lane_8x32( lane_hash, hash32, lane, 256 );
casti_m256i( lane_hash, 0 ) =
_mm256_shuffle_epi8( casti_m256i( lane_hash, 0 ), bswap_shuf );
if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
{
pdata[19] = n + lane;
submit_solution( work, lane_hash, mythr );
}
}
}
*noncev = _mm256_add_epi32( *noncev, eight );
n += 8;
}
*noncev = _mm256_add_epi32( *noncev, eight );
n += 8;
} while ( (n < last_nonce) && !work_restart[thr_id].restart );
pdata[19] = n;
*hashes_done = n - first_nonce;
@@ -366,12 +282,12 @@ int scanhash_sha256d_8way( struct work *work, const uint32_t max_nonce,
int scanhash_sha256d_4way( struct work *work, const uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
__m128i vdata[32] __attribute__ ((aligned (64)));
__m128i block[16] __attribute__ ((aligned (32)));
__m128i hash32[8] __attribute__ ((aligned (32)));
__m128i initstate[8] __attribute__ ((aligned (32)));
__m128i midstate1[8] __attribute__ ((aligned (32)));
uint32_t lane_hash[8] __attribute__ ((aligned (32)));
__m128i vdata[32] __attribute__ ((aligned (64)));
__m128i block[16] __attribute__ ((aligned (32)));
__m128i hash32[8] __attribute__ ((aligned (32)));
__m128i istate[8] __attribute__ ((aligned (32)));
__m128i mstate[8] __attribute__ ((aligned (32)));
uint32_t lane_hash[8] __attribute__ ((aligned (32)));
uint32_t *hash32_d7 = (uint32_t*)&( hash32[7] );
uint32_t *pdata = work->data;
const uint32_t *ptarget = work->target;
@@ -392,33 +308,30 @@ int scanhash_sha256d_4way( struct work *work, const uint32_t max_nonce,
vdata[16+4] = last_byte;
memset_zero_128( vdata+16 + 5, 10 );
vdata[16+15] = _mm_set1_epi32( 80*8 ); // bit count
vdata[16+15] = _mm_set1_epi32( 80*8 );
block[ 8] = last_byte;
memset_zero_128( block + 9, 6 );
block[15] = _mm_set1_epi32( 32*8 ); // bit count
block[15] = _mm_set1_epi32( 32*8 );
// initialize state
initstate[0] = _mm_set1_epi64x( 0x6A09E6676A09E667 );
initstate[1] = _mm_set1_epi64x( 0xBB67AE85BB67AE85 );
initstate[2] = _mm_set1_epi64x( 0x3C6EF3723C6EF372 );
initstate[3] = _mm_set1_epi64x( 0xA54FF53AA54FF53A );
initstate[4] = _mm_set1_epi64x( 0x510E527F510E527F );
initstate[5] = _mm_set1_epi64x( 0x9B05688C9B05688C );
initstate[6] = _mm_set1_epi64x( 0x1F83D9AB1F83D9AB );
initstate[7] = _mm_set1_epi64x( 0x5BE0CD195BE0CD19 );
istate[0] = _mm_set1_epi32( sha256_iv[0] );
istate[1] = _mm_set1_epi32( sha256_iv[1] );
istate[2] = _mm_set1_epi32( sha256_iv[2] );
istate[3] = _mm_set1_epi32( sha256_iv[3] );
istate[4] = _mm_set1_epi32( sha256_iv[4] );
istate[5] = _mm_set1_epi32( sha256_iv[5] );
istate[6] = _mm_set1_epi32( sha256_iv[6] );
istate[7] = _mm_set1_epi32( sha256_iv[7] );
// hash first 64 bytes of data
sha256_4way_transform_le( midstate1, vdata, initstate );
sha256_4way_transform_le( mstate, vdata, istate );
do
{
// 1. final 16 bytes of data, with padding
sha256_4way_transform_le( block, vdata+16, initstate );
sha256_4way_transform_le( block, vdata+16, mstate );
sha256_4way_transform_le( hash32, block, istate );
// 2. 32 byte hash from 1.
sha256_4way_transform_le( hash32, block, initstate );
// byte swap final hash for testing
mm128_block_bswap_32( hash32, hash32 );
for ( int lane = 0; lane < 4; lane++ )
@@ -440,3 +353,5 @@ int scanhash_sha256d_4way( struct work *work, const uint32_t max_nonce,
}
#endif