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

v3.13.1.1

Browse Source
This commit is contained in:
Jay D Dee
2020-05-17 19:21:37 -04:00
parent 1c0719e8a4
commit 9653bca1e2
27 changed files with 143 additions and 183 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
algo/skein/skein-hash-4way.c
View File

@@ -731,7 +731,7 @@ void skein512_8way_full( skein512_8way_context *sc, void *out, const void *data,
void
skein512_8way_prehash64( skein512_8way_context *sc, const void *data )
{
__m512i *vdata = (__m512*)data;
__m512i *vdata = (__m512i*)data;
__m512i *buf = sc->buf;
buf[0] = vdata[0];
buf[1] = vdata[1];

23
algo/x16/minotaur.c
View File

@@ -7,7 +7,6 @@
#include <stdio.h>
#include "algo/blake/sph_blake.h"
#include "algo/bmw/sph_bmw.h"
#include "algo/groestl/sph_groestl.h"
#include "algo/jh/sph_jh.h"
#include "algo/keccak/sph_keccak.h"
#include "algo/skein/sph_skein.h"
@@ -15,7 +14,6 @@
#include "algo/luffa/luffa_for_sse2.h"
#include "algo/cubehash/cubehash_sse2.h"
#include "algo/simd/nist.h"
#include "algo/echo/sph_echo.h"
#include "algo/hamsi/sph_hamsi.h"
#include "algo/fugue/sph_fugue.h"
#include "algo/shabal/sph_shabal.h"
@@ -24,9 +22,11 @@
#if defined(__AES__)
#include "algo/echo/aes_ni/hash_api.h"
#include "algo/groestl/aes_ni/hash-groestl.h"
#else
#include "algo/echo/sph_echo.h"
#include "algo/groestl/sph_groestl.h"
#endif
// Config
#define MINOTAUR_ALGO_COUNT 16
@@ -39,8 +39,8 @@ struct TortureGarden {
hashState_echo echo;
hashState_groestl groestl;
#else
sph_groestl512_context groestl;
sph_echo512_context echo;
sph_groestl512_context groestl;
#endif
sph_blake512_context blake;
sph_bmw512_context bmw;
@@ -62,13 +62,13 @@ struct TortureGarden {
TortureNode *childLeft;
TortureNode *childRight;
} nodes[22];
};
} __attribute__ ((aligned (64)));
// Get a 64-byte hash for given 64-byte input, using given TortureGarden contexts and given algo index
static void get_hash( void *output, const void *input, TortureGarden *garden,
unsigned int algo )
{
unsigned char _ALIGN(64) hash[64];
unsigned char hash[64] __attribute__ ((aligned (64)));
switch (algo) {
case 0:
@@ -170,7 +170,7 @@ static void get_hash( void *output, const void *input, TortureGarden *garden,
static void traverse_garden( TortureGarden *garden, void *hash,
TortureNode *node )
{
unsigned char _ALIGN(64) partialHash[64];
unsigned char partialHash[64] __attribute__ ((aligned (64)));
get_hash(partialHash, hash, garden, node->algo);
if ( partialHash[63] % 2 == 0 )
@@ -195,9 +195,9 @@ static inline void link_nodes( TortureNode *parent, TortureNode *childLeft,
parent->childRight = childRight;
}
static TortureGarden garden;
static __thread TortureGarden garden;
void initialize_torture_garden()
bool initialize_torture_garden()
{
// Create torture garden nodes. Note that both sides of 19 and 20 lead to 21, and 21 has no children (to make traversal complete).
link_nodes(&garden.nodes[0], &garden.nodes[1], &garden.nodes[2]);
@@ -223,12 +223,13 @@ void initialize_torture_garden()
link_nodes(&garden.nodes[20], &garden.nodes[21], &garden.nodes[21]);
garden.nodes[21].childLeft = NULL;
garden.nodes[21].childRight = NULL;
return true;
}
// Produce a 32-byte hash from 80-byte input data
int minotaur_hash( void *output, const void *input )
{
unsigned char _ALIGN(64) hash[64];
unsigned char hash[64] __attribute__ ((aligned (64)));
// Find initial sha512 hash
SHA512_Init( &garden.sha512 );
@@ -251,7 +252,7 @@ bool register_minotaur_algo( algo_gate_t* gate )
{
gate->hash = (void*)&minotaur_hash;
gate->optimizations = SSE2_OPT | AES_OPT | AVX2_OPT | AVX512_OPT;
initialize_torture_garden();
gate->miner_thread_init = (void*)&initialize_torture_garden;
return true;
};

8
algo/x17/sonoa-gate.c
View File

@@ -4,14 +4,14 @@ bool register_sonoa_algo( algo_gate_t* gate )
{
#if defined (SONOA_8WAY)
gate->scanhash = (void*)&scanhash_sonoa_8way;
// gate->hash = (void*)&sonoa_8way_hash;
gate->hash = (void*)&sonoa_8way_hash;
#elif defined (SONOA_4WAY)
gate->scanhash = (void*)&scanhash_sonoa_4way;
// gate->hash = (void*)&sonoa_4way_hash;
gate->hash = (void*)&sonoa_4way_hash;
#else
init_sonoa_ctx();
gate->scanhash = (void*)&scanhash_sonoa;
// gate->hash = (void*)&sonoa_hash;
// gate->scanhash = (void*)&scanhash_sonoa;
gate->hash = (void*)&sonoa_hash;
#endif
gate->optimizations = SSE2_OPT | AES_OPT | AVX2_OPT | AVX512_OPT | VAES_OPT;
return true;

30
algo/x17/sonoa.c
View File

@@ -569,34 +569,4 @@ int sonoa_hash( void *state, const void *input, int thrid )
return 1;
}
int scanhash_sonoa( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr)
{
uint32_t edata[20] __attribute__((aligned(64)));
uint32_t hash64[8] __attribute__((aligned(64)));
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
uint32_t n = pdata[19];
const uint32_t first_nonce = pdata[19];
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm128_bswap32_80( edata, pdata );
do
{
edata[19] = n;
if ( sonoa_hash( hash64, edata, thr_id ) )
if ( unlikely( valid_hash( hash64, ptarget ) && !bench ) )
{
pdata[19] = bswap_32( n );
submit_solution( work, hash64, mythr );
}
n++;
} while ( n < max_nonce && !work_restart[thr_id].restart );
*hashes_done = n - first_nonce;
pdata[19] = n;
return 0;
}
#endif

8
algo/x17/x17-4way.c
View File

@@ -57,7 +57,7 @@ union _x17_8way_context_overlay
} __attribute__ ((aligned (64)));
typedef union _x17_8way_context_overlay x17_8way_context_overlay;
void x17_8way_hash( void *state, const void *input )
int x17_8way_hash( void *state, const void *input )
{
uint64_t vhash[8*8] __attribute__ ((aligned (128)));
uint64_t vhashA[8*8] __attribute__ ((aligned (64)));
@@ -230,6 +230,8 @@ void x17_8way_hash( void *state, const void *input )
haval256_5_8way_init( &ctx.haval );
haval256_5_8way_update( &ctx.haval, vhashA, 64 );
haval256_5_8way_close( &ctx.haval, state );
return 1;
}
int scanhash_x17_8way( struct work *work, uint32_t max_nonce,
@@ -300,7 +302,7 @@ union _x17_4way_context_overlay
};
typedef union _x17_4way_context_overlay x17_4way_context_overlay;
void x17_4way_hash( void *state, const void *input )
int x17_4way_hash( void *state, const void *input )
{
uint64_t vhash[8*4] __attribute__ ((aligned (64)));
uint64_t vhashA[8*4] __attribute__ ((aligned (64)));
@@ -399,6 +401,8 @@ void x17_4way_hash( void *state, const void *input )
haval256_5_4way_init( &ctx.haval );
haval256_5_4way_update( &ctx.haval, vhashB, 64 );
haval256_5_4way_close( &ctx.haval, state );
return 1;
}
int scanhash_x17_4way( struct work *work, uint32_t max_nonce,

1
algo/x17/x17-gate.c
View File

@@ -9,7 +9,6 @@ bool register_x17_algo( algo_gate_t* gate )
gate->scanhash = (void*)&scanhash_x17_4way;
gate->hash = (void*)&x17_4way_hash;
#else
gate->scanhash = (void*)&scanhash_x17;
gate->hash = (void*)&x17_hash;
#endif
gate->optimizations = SSE2_OPT | AES_OPT | AVX2_OPT | AVX512_OPT | VAES_OPT;

8
algo/x17/x17-gate.h
View File

@@ -14,20 +14,18 @@ bool register_x17_algo( algo_gate_t* gate );
#if defined(X17_8WAY)
void x17_8way_hash( void *state, const void *input );
int x17_8way_hash( void *state, const void *input );
int scanhash_x17_8way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr );
#elif defined(X17_4WAY)
void x17_4way_hash( void *state, const void *input );
int x17_4way_hash( void *state, const void *input );
int scanhash_x17_4way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr );
#endif
void x17_hash( void *state, const void *input );
int scanhash_x17( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr );
int x17_hash( void *state, const void *input, int thr_id );
#endif

32
algo/x17/x17.c
View File

@@ -56,7 +56,7 @@ union _x17_context_overlay
};
typedef union _x17_context_overlay x17_context_overlay;
void x17_hash(void *output, const void *input)
int x17_hash(void *output, const void *input, int thr_id )
{
// unsigned char hash[64 * 4] __attribute__((aligned(64))) = {0};
unsigned char hash[64] __attribute__((aligned(64)));
@@ -143,36 +143,8 @@ void x17_hash(void *output, const void *input)
sph_haval256_5_init(&ctx.haval);
sph_haval256_5( &ctx.haval, (const void*)hash, 64 );
sph_haval256_5_close( &ctx.haval, output );
}
int scanhash_x17( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr)
{
uint32_t edata[20] __attribute__((aligned(64)));
uint32_t hash64[8] __attribute__((aligned(64)));
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
uint32_t n = pdata[19] - 1;
const uint32_t first_nonce = pdata[19];
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm128_bswap32_80( edata, pdata );
do
{
edata[19] = n;
x17_hash( hash64, edata );
if ( unlikely( valid_hash( hash64, ptarget ) && !bench ) )
{
pdata[19] = bswap_32( n );
submit_solution( work, hash64, mythr );
}
n++;
} while ( n < max_nonce && !work_restart[thr_id].restart );
*hashes_done = n - first_nonce;
pdata[19] = n;
return 0;
return 1;
}
#endif

8
algo/x17/xevan-4way.c
View File

@@ -57,7 +57,7 @@ union _xevan_8way_context_overlay
} __attribute__ ((aligned (64)));
typedef union _xevan_8way_context_overlay xevan_8way_context_overlay;
void xevan_8way_hash( void *output, const void *input )
int xevan_8way_hash( void *output, const void *input )
{
uint64_t vhash[16<<3] __attribute__ ((aligned (128)));
uint64_t vhashA[16<<3] __attribute__ ((aligned (64)));
@@ -395,6 +395,8 @@ void xevan_8way_hash( void *output, const void *input )
haval256_5_8way_init( &ctx.haval );
haval256_5_8way_update( &ctx.haval, vhashA, dataLen );
haval256_5_8way_close( &ctx.haval, output );
return 1;
}
int scanhash_xevan_8way( struct work *work, uint32_t max_nonce,
@@ -465,7 +467,7 @@ union _xevan_4way_context_overlay
};
typedef union _xevan_4way_context_overlay xevan_4way_context_overlay;
void xevan_4way_hash( void *output, const void *input )
int xevan_4way_hash( void *output, const void *input )
{
uint64_t hash0[16] __attribute__ ((aligned (64)));
uint64_t hash1[16] __attribute__ ((aligned (64)));
@@ -666,6 +668,8 @@ void xevan_4way_hash( void *output, const void *input )
haval256_5_4way_init( &ctx.haval );
haval256_5_4way_update( &ctx.haval, vhashA, dataLen );
haval256_5_4way_close( &ctx.haval, output );
return 1;
}
int scanhash_xevan_4way( struct work *work, uint32_t max_nonce,

1
algo/x17/xevan-gate.c
View File

@@ -10,7 +10,6 @@ bool register_xevan_algo( algo_gate_t* gate )
gate->hash = (void*)&xevan_4way_hash;
#else
init_xevan_ctx();
gate->scanhash = (void*)&scanhash_xevan;
gate->hash = (void*)&xevan_hash;
#endif
gate->optimizations = SSE2_OPT | AES_OPT | AVX2_OPT | AVX512_OPT | VAES_OPT;

12
algo/x17/xevan-gate.h
View File

@@ -14,14 +14,12 @@ bool register_xevan_algo( algo_gate_t* gate );
#if defined(XEVAN_8WAY)
void xevan_8way_hash( void *state, const void *input );
int xevan_8way_hash( void *state, const void *input );
int scanhash_xevan_8way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr );
#elif defined(XEVAN_4WAY)
void xevan_4way_hash( void *state, const void *input );
int xevan_4way_hash( void *state, const void *input );
int scanhash_xevan_4way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr );
@@ -29,11 +27,7 @@ int scanhash_xevan_4way( struct work *work, uint32_t max_nonce,
#else
void xevan_hash( void *state, const void *input );
int scanhash_xevan( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr );
int xevan_hash( void *state, const void *input, int trh_id );
void init_xevan_ctx();
#endif

32
algo/x17/xevan.c
View File

@@ -83,7 +83,7 @@ void init_xevan_ctx()
#endif
};
void xevan_hash(void *output, const void *input)
int xevan_hash(void *output, const void *input, int thr_id )
{
uint32_t _ALIGN(64) hash[32]; // 128 bytes required
const int dataLen = 128;
@@ -218,36 +218,8 @@ void xevan_hash(void *output, const void *input)
sph_haval256_5_close(&ctx.haval, hash);
memcpy(output, hash, 32);
}
int scanhash_xevan( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr)
{
uint32_t edata[20] __attribute__((aligned(64)));
uint32_t hash64[8] __attribute__((aligned(64)));
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
uint32_t n = pdata[19];
const uint32_t first_nonce = pdata[19];
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm128_bswap32_80( edata, pdata );
do
{
edata[19] = n;
xevan_hash( hash64, edata );
if ( unlikely( valid_hash( hash64, ptarget ) && !bench ) )
{
pdata[19] = bswap_32( n );
submit_solution( work, hash64, mythr );
}
n++;
} while ( n < max_nonce && !work_restart[thr_id].restart );
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
return 1;
}
#endif

2
algo/yescrypt/yescrypt.c
View File

@@ -445,7 +445,7 @@ bool register_yescrypt_algo( algo_gate_t* gate )
YESCRYPT_P = 1;
applog( LOG_NOTICE,"Yescrypt parameters: N= %d, R= %d.", YESCRYPT_N,
applog( LOG_NOTICE,"Yescrypt parameters: N= %d, R= %d", YESCRYPT_N,
YESCRYPT_R );
if ( yescrypt_client_key )
applog( LOG_NOTICE,"Key= \"%s\"\n", yescrypt_client_key );

4
algo/yespower/yespower-gate.c
View File

@@ -139,7 +139,7 @@ bool register_yespower_algo( algo_gate_t* gate )
yespower_params.perslen = 0;
}
applog( LOG_NOTICE,"Yespower parameters: N= %d, R= %d.", yespower_params.N,
applog( LOG_NOTICE,"Yespower parameters: N= %d, R= %d", yespower_params.N,
yespower_params.r );
if ( yespower_params.pers )
applog( LOG_NOTICE,"Key= \"%s\"\n", yespower_params.pers );
@@ -264,7 +264,7 @@ bool register_power2b_algo( algo_gate_t* gate )
yespower_params.pers = "Now I am become Death, the destroyer of worlds";
yespower_params.perslen = 46;
applog( LOG_NOTICE,"yespower-b2b parameters: N= %d, R= %d.", yespower_params.N,
applog( LOG_NOTICE,"yespower-b2b parameters: N= %d, R= %d", yespower_params.N,
yespower_params.r );
applog( LOG_NOTICE,"Key= \"%s\"", yespower_params.pers );
applog( LOG_NOTICE,"Key length= %d\n", yespower_params.perslen );

2
algo/yespower/yespower.h
View File

@@ -76,7 +76,7 @@ typedef struct {
unsigned char uc[32];
} yespower_binary_t __attribute__ ((aligned (64)));
yespower_params_t yespower_params;
extern yespower_params_t yespower_params;
//SHA256_CTX sha256_prehash_ctx;
extern __thread SHA256_CTX sha256_prehash_ctx;