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

v3.9.4

Browse Source
This commit is contained in:
Jay D Dee
2019-06-18 13:15:45 -04:00
parent 71d6b97ee8
commit d6e8d7a46e
75 changed files with 1790 additions and 781 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
algo/skein/skein-4way.c
View File

@@ -2,7 +2,11 @@
#include <string.h>
#include <stdint.h>
#include "skein-hash-4way.h"
#include "algo/sha/sha2-hash-4way.h"
#if defined(__SHA__)
#include <openssl/sha.h>
#else
#include "algo/sha/sha2-hash-4way.h"
#endif
#if defined (SKEIN_4WAY)
@@ -11,53 +15,69 @@ void skeinhash_4way( void *state, const void *input )
uint64_t vhash64[8*4] __attribute__ ((aligned (64)));
uint32_t vhash32[16*4] __attribute__ ((aligned (64)));
skein512_4way_context ctx_skein;
#if defined(__SHA__)
uint32_t hash0[16] __attribute__ ((aligned (64)));
uint32_t hash1[16] __attribute__ ((aligned (64)));
uint32_t hash2[16] __attribute__ ((aligned (64)));
uint32_t hash3[16] __attribute__ ((aligned (64)));
SHA256_CTX ctx_sha256;
#else
sha256_4way_context ctx_sha256;
#endif
skein512_4way_init( &ctx_skein );
skein512_4way( &ctx_skein, input, 80 );
skein512_4way_close( &ctx_skein, vhash64 );
#if defined(__SHA__)
mm256_dintrlv_4x64( hash0, hash1, hash2, hash3, vhash64, 512 );
SHA256_Init( &ctx_sha256 );
SHA256_Update( &ctx_sha256, (unsigned char*)hash0, 64 );
SHA256_Final( (unsigned char*)hash0, &ctx_sha256 );
SHA256_Init( &ctx_sha256 );
SHA256_Update( &ctx_sha256, (unsigned char*)hash1, 64 );
SHA256_Final( (unsigned char*)hash1, &ctx_sha256 );
SHA256_Init( &ctx_sha256 );
SHA256_Update( &ctx_sha256, (unsigned char*)hash2, 64 );
SHA256_Final( (unsigned char*)hash2, &ctx_sha256 );
SHA256_Init( &ctx_sha256 );
SHA256_Update( &ctx_sha256, (unsigned char*)hash3, 64 );
SHA256_Final( (unsigned char*)hash3, &ctx_sha256 );
mm128_intrlv_4x32( state, hash0, hash1, hash2, hash3, 256 );
#else
mm256_rintrlv_4x64_4x32( vhash32, vhash64, 512 );
sha256_4way_init( &ctx_sha256 );
sha256_4way( &ctx_sha256, vhash32, 64 );
sha256_4way_close( &ctx_sha256, state );
mm128_dintrlv_4x32( state, state+32, state+64, state+96,
vhash32, 256 );
#endif
}
int scanhash_skein_4way( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done )
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t vdata[20*4] __attribute__ ((aligned (64)));
uint32_t hash[8*4] __attribute__ ((aligned (64)));
uint32_t lane_hash[8];
uint32_t *hash7 = &(hash[7<<2]);
uint32_t edata[20] __attribute__ ((aligned (64)));
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t Htarg = ptarget[7];
const uint32_t first_nonce = pdata[19];
uint32_t n = first_nonce;
// hash is returned deinterleaved
uint32_t *nonces = work->nonces;
int num_found = 0;
// data is 80 bytes, 20 u32 or 4 u64.
swab32_array( edata, pdata, 20 );
mm256_intrlv_4x64( vdata, edata, edata, edata, edata, 640 );
uint32_t *noncep = vdata + 73; // 9*8 + 1
__m256i *noncev = (__m256i*)vdata + 9; // aligned
/* int */ thr_id = mythr->id; // thr_id arg is deprecated
mm256_bswap_intrlv80_4x64( vdata, pdata );
do
{
be32enc( noncep, n );
be32enc( noncep+2, n+1 );
be32enc( noncep+4, n+2 );
be32enc( noncep+6, n+3 );
*noncev = mm256_intrlv_blend_32( mm256_bswap_32(
_mm256_set_epi32( n+3, 0, n+2, 0, n+1, 0, n, 0 ) ), *noncev );
skeinhash_4way( hash, vdata );
@@ -68,16 +88,14 @@ int scanhash_skein_4way( int thr_id, struct work *work, uint32_t max_nonce,
if ( fulltest( lane_hash, ptarget ) )
{
pdata[19] = n + lane;
nonces[ num_found++ ] = n + lane;
work_set_target_ratio( work, lane_hash );
submit_solution( work, lane_hash, mythr, lane );
}
}
n += 4;
} while ( (num_found == 0) && (n < max_nonce)
&& !work_restart[thr_id].restart );
} while ( (n < max_nonce) && !work_restart[thr_id].restart );
*hashes_done = n - first_nonce + 1;
return num_found;
return 0;
}
#endif

7
algo/skein/skein-gate.h
View File

@@ -3,24 +3,21 @@
#include <stdint.h>
#include "algo-gate-api.h"
// Override multi way on ryzen, SHA is better.
#if !defined(RYZEN_)
#if defined(__AVX2__)
#define SKEIN_4WAY
#endif
#endif
#if defined(SKEIN_4WAY)
void skeinhash_4way( void *output, const void *input );
int scanhash_skein_4way( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done );
uint64_t *hashes_done, struct thr_info *mythr );
#endif
void skeinhash( void *output, const void *input );
int scanhash_skein( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done );
uint64_t *hashes_done, struct thr_info *mythr );
#endif

7
algo/skein/skein.c
View File

@@ -21,8 +21,8 @@ void skeinhash(void *state, const void *input)
memcpy(state, hash, 32);
}
int scanhash_skein(int thr_id, struct work *work,
uint32_t max_nonce, uint64_t *hashes_done)
int scanhash_skein( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
@@ -31,7 +31,8 @@ int scanhash_skein(int thr_id, struct work *work,
const uint32_t Htarg = ptarget[7];
const uint32_t first_nonce = pdata[19];
uint32_t n = first_nonce;
/* int */ thr_id = mythr->id; // thr_id arg is deprecated
swab32_array( endiandata, pdata, 20 );
do {

36
algo/skein/skein2-4way.c
View File

@@ -20,55 +20,43 @@ void skein2hash_4way( void *output, const void *input )
}
int scanhash_skein2_4way( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done )
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t hash[8*4] __attribute__ ((aligned (64)));
uint32_t *hash7 = &(hash[25]);
uint32_t vdata[20*4] __attribute__ ((aligned (64)));
uint32_t endiandata[20] __attribute__ ((aligned (64)));
uint64_t *edata = (uint64_t*)endiandata;
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t Htarg = ptarget[7];
const uint32_t first_nonce = pdata[19];
uint32_t n = first_nonce;
uint32_t *nonces = work->nonces;
int num_found = 0;
swab32_array( endiandata, pdata, 20 );
mm256_interleave_4x64( vdata, edata, edata, edata, edata, 640 );
uint32_t *noncep = vdata + 73; // 9*8 + 1
__m256i *noncev = (__m256i*)vdata + 9; // aligned
/* int */ thr_id = mythr->id; // thr_id arg is deprecated
mm256_bswap_intrlv80_4x64( vdata, pdata );
do
{
be32enc( noncep, n );
be32enc( noncep+2, n+1 );
be32enc( noncep+4, n+2 );
be32enc( noncep+6, n+3 );
*noncev = mm256_intrlv_blend_32( mm256_bswap_32(
_mm256_set_epi32( n+3, 0, n+2, 0, n+1, 0, n, 0 ) ), *noncev );
skein2hash( hash, vdata );
skein2hash_4way( hash, vdata );
for ( int lane = 0; lane < 4; lane++ )
if ( hash7[ lane ] <= Htarg )
if ( hash7[ lane<<1 ] <= Htarg )
{
// deinterleave hash for lane
uint32_t lane_hash[8];
mm256_extract_lane_4x64( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) )
if ( fulltest( lane_hash, ptarget ) && !opt_benchmark )
{
pdata[19] = n + lane;
nonces[ num_found++ ] = n + lane;
work_set_target_ratio( work, lane_hash );
submit_solution( work, lane_hash, mythr, lane );
}
}
n += 4;
} while ( (num_found == 0) && (n < max_nonce)
&& !work_restart[thr_id].restart );
} while ( (n < max_nonce) && !work_restart[thr_id].restart );
*hashes_done = n - first_nonce + 1;
return num_found;
return 0;
}
#endif

3
algo/skein/skein2-gate.c
View File

@@ -10,10 +10,9 @@ int64_t skein2_get_max64 ()
bool register_skein2_algo( algo_gate_t* gate )
{
gate->optimizations = AVX2_OPT;
#if defined (FOUR_WAY) && defined (__AVX2__)
#if defined (SKEIN2_4WAY)
gate->scanhash = (void*)&scanhash_skein2_4way;
gate->hash = (void*)&skein2hash_4way;
four_way_not_tested();
#else
gate->scanhash = (void*)&scanhash_skein2;
gate->hash = (void*)&skein2hash;

4
algo/skein/skein2-gate.h
View File

@@ -10,11 +10,11 @@
#if defined(SKEIN2_4WAY)
void skein2hash_4way( void *output, const void *input );
int scanhash_skein2_4way( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t* hashes_done );
uint64_t* hashes_done, struct thr_info *mythr );
#endif
void skein2hash( void *output, const void *input );
int scanhash_skein2( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done );
uint64_t *hashes_done, struct thr_info *mythr );
#endif

5
algo/skein/skein2.c
View File

@@ -34,8 +34,8 @@ void skein2hash(void *output, const void *input)
}
int scanhash_skein2(int thr_id, struct work *work,
uint32_t max_nonce, uint64_t *hashes_done)
int scanhash_skein2( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
@@ -44,6 +44,7 @@ int scanhash_skein2(int thr_id, struct work *work,
const uint32_t Htarg = ptarget[7];
const uint32_t first_nonce = pdata[19];
uint32_t n = first_nonce;
/* int */ thr_id = mythr->id; // thr_id arg is deprecated
swab32_array( endiandata, pdata, 20 );