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

v3.12.2

Browse Source
This commit is contained in:
Jay D Dee
2020-02-09 13:30:40 -05:00
parent dc2f8d81d3
commit 3da2b958cf
39 changed files with 1496 additions and 1518 deletions
Download Patch File Download Diff File Expand all files Collapse all files

89
algo/skein/skein-4way.c
View File

@@ -13,17 +13,21 @@
#if defined (SKEIN_8WAY)
static __thread skein512_8way_context skein512_8way_ctx
__attribute__ ((aligned (64)));
void skeinhash_8way( void *state, const void *input )
{
uint64_t vhash64[8*8] __attribute__ ((aligned (128)));
skein512_8way_context ctx_skein;
memcpy( &ctx_skein, &skein512_8way_ctx, sizeof( ctx_skein ) );
uint32_t vhash32[16*8] __attribute__ ((aligned (128)));
sha256_8way_context ctx_sha256;
skein512_8way_init( &ctx_skein );
skein512_8way_update( &ctx_skein, input, 80 );
skein512_8way_close( &ctx_skein, vhash64 );
skein512_8way_full( &ctx_skein, vhash64, input, 80 );
// skein512_8way_update( &ctx_skein, input + (64*8), 16 );
// skein512_8way_close( &ctx_skein, vhash64 );
rintrlv_8x64_8x32( vhash32, vhash64, 512 );
@@ -36,63 +40,74 @@ int scanhash_skein_8way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t vdata[20*8] __attribute__ ((aligned (128)));
uint32_t hash[16*8] __attribute__ ((aligned (64)));
uint32_t hash[8*8] __attribute__ ((aligned (64)));
uint32_t lane_hash[8] __attribute__ ((aligned (64)));
uint32_t *hash7 = &(hash[7<<3]);
uint32_t *hash_d7 = &(hash[7*8]);
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t Htarg = ptarget[7];
const uint32_t targ_d7 = ptarget[7];
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 8;
uint32_t n = first_nonce;
__m512i *noncev = (__m512i*)vdata + 9; // aligned
int thr_id = mythr->id;
__m512i *noncev = (__m512i*)vdata + 9;
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm512_bswap32_intrlv80_8x64( vdata, pdata );
*noncev = mm512_intrlv_blend_32(
_mm512_set_epi32( n+7, 0, n+6, 0, n+5, 0, n+4, 0,
n+3, 0, n+2, 0, n+1, 0, n , 0 ), *noncev );
// skein512_8way_init( &skein512_8way_ctx );
// skein512_8way_update( &skein512_8way_ctx, vdata, 64 );
do
{
*noncev = mm512_intrlv_blend_32( mm512_bswap_32(
_mm512_set_epi32( n+7, 0, n+6, 0, n+5, 0, n+4, 0,
n+3, 0, n+2, 0, n+1, 0, n , 0 ) ), *noncev );
skeinhash_8way( hash, vdata );
for ( int lane = 0; lane < 8; lane++ )
if ( hash7[ lane ] <= Htarg )
if ( unlikely( hash_d7[ lane ] <= targ_d7 ) && !bench )
{
extr_lane_8x32( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) )
if ( valid_hash( lane_hash, ptarget ) )
{
pdata[19] = n + lane;
pdata[19] = bswap_32( n + lane );
submit_lane_solution( work, lane_hash, mythr, lane );
}
}
*noncev = _mm512_add_epi32( *noncev,
m512_const1_64( 0x0000000800000000 ) );
n += 8;
} while ( (n < max_nonce-8) && !work_restart[thr_id].restart );
} while ( likely( (n < last_nonce) && !work_restart[thr_id].restart ) );
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}
#elif defined (SKEIN_4WAY)
//static __thread skein512_4way_context skein512_4way_ctx
// __attribute__ ((aligned (64)));
void skeinhash_4way( void *state, const void *input )
{
uint64_t vhash64[8*4] __attribute__ ((aligned (128)));
skein512_4way_context ctx_skein;
// memcpy( &ctx_skein, &skein512_4way_ctx, sizeof( 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;
SHA256_CTX ctx_sha256;
#else
uint32_t vhash32[16*4] __attribute__ ((aligned (64)));
sha256_4way_context ctx_sha256;
#endif
skein512_4way_init( &ctx_skein );
skein512_4way_update( &ctx_skein, input, 80 );
skein512_4way_close( &ctx_skein, vhash64 );
skein512_4way_full( &ctx_skein, vhash64, input, 80 );
// skein512_4way_update( &ctx_skein, input + (64*4), 16 );
// skein512_4way_close( &ctx_skein, vhash64 );
#if defined(__SHA__)
dintrlv_4x64( hash0, hash1, hash2, hash3, vhash64, 512 );
@@ -127,38 +142,44 @@ int scanhash_skein_4way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t vdata[20*4] __attribute__ ((aligned (64)));
uint32_t hash[16*4] __attribute__ ((aligned (64)));
uint32_t hash[8*4] __attribute__ ((aligned (64)));
uint32_t lane_hash[8] __attribute__ ((aligned (32)));
uint32_t *hash7 = &(hash[7<<2]);
uint32_t *hash_d7 = &(hash[7<<2]);
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t Htarg = ptarget[7];
const uint32_t targ_d7 = ptarget[7];
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 4;
uint32_t n = first_nonce;
__m256i *noncev = (__m256i*)vdata + 9; // aligned
int thr_id = mythr->id;
__m256i *noncev = (__m256i*)vdata + 9;
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm256_bswap32_intrlv80_4x64( vdata, pdata );
// skein512_4way_init( &skein512_4way_ctx );
// skein512_4way_update( &skein512_4way_ctx, vdata, 64 );
*noncev = mm256_intrlv_blend_32(
_mm256_set_epi32( n+3, 0, n+2, 0, n+1, 0, n, 0 ), *noncev );
do
{
*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 );
for ( int lane = 0; lane < 4; lane++ )
if ( hash7[ lane ] <= Htarg )
if ( unlikely( ( hash_d7[ lane ] <= targ_d7 ) && !bench ) )
{
extr_lane_4x32( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) )
if ( valid_hash( lane_hash, ptarget ) )
{
pdata[19] = n + lane;
pdata[19] = bswap_32( n + lane );
submit_lane_solution( work, lane_hash, mythr, lane );
}
}
*noncev = _mm256_add_epi32( *noncev,
m256_const1_64( 0x0000000400000000 ) );
n += 4;
} while ( (n < max_nonce-4) && !work_restart[thr_id].restart );
} while ( likely( (n < last_nonce) && !work_restart[thr_id].restart ) );
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}

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

@@ -4,14 +4,16 @@
bool register_skein_algo( algo_gate_t* gate )
{
gate->optimizations = AVX2_OPT | AVX512_OPT | SHA_OPT;
#if defined (SKEIN_8WAY)
gate->optimizations = AVX2_OPT | AVX512_OPT;
gate->scanhash = (void*)&scanhash_skein_8way;
gate->hash = (void*)&skeinhash_8way;
#elif defined (SKEIN_4WAY)
gate->optimizations = AVX2_OPT | AVX512_OPT | SHA_OPT;
gate->scanhash = (void*)&scanhash_skein_4way;
gate->hash = (void*)&skeinhash_4way;
#else
gate->optimizations = AVX2_OPT | AVX512_OPT | SHA_OPT;
gate->scanhash = (void*)&scanhash_skein;
gate->hash = (void*)&skeinhash;
#endif

151
algo/skein/skein-hash-4way.c
View File

@@ -654,6 +654,80 @@ skein_big_close_8way( skein512_8way_context *sc, unsigned ub, unsigned n,
memcpy_512( dst, buf, out_len >> 3 );
}
void skein512_8way_full( skein512_8way_context *sc, void *out, const void *data,
size_t len )
{
__m512i h0, h1, h2, h3, h4, h5, h6, h7;
__m512i *vdata = (__m512i*)data;
__m512i *buf = sc->buf;
size_t ptr = 0;
unsigned first;
uint64_t bcount = 0;
const int buf_size = 64; // 64 * _m256i
// Init
h0 = m512_const1_64( 0x4903ADFF749C51CE );
h1 = m512_const1_64( 0x0D95DE399746DF03 );
h2 = m512_const1_64( 0x8FD1934127C79BCE );
h3 = m512_const1_64( 0x9A255629FF352CB1 );
h4 = m512_const1_64( 0x5DB62599DF6CA7B0 );
h5 = m512_const1_64( 0xEABE394CA9D5C3F4 );
h6 = m512_const1_64( 0x991112C71A75B523 );
h7 = m512_const1_64( 0xAE18A40B660FCC33 );
// Update
if ( len <= buf_size - ptr )
{
memcpy_512( buf + (ptr>>3), vdata, len>>3 );
ptr += len;
}
else
{
first = ( bcount == 0 ) << 7;
do {
size_t clen;
if ( ptr == buf_size )
{
bcount ++;
UBI_BIG_8WAY( 96 + first, 0 );
first = 0;
ptr = 0;
}
clen = buf_size - ptr;
if ( clen > len )
clen = len;
memcpy_512( buf + (ptr>>3), vdata, clen>>3 );
ptr += clen;
vdata += (clen>>3);
len -= clen;
} while ( len > 0 );
}
// Close
unsigned et;
memset_zero_512( buf + (ptr>>3), (buf_size - ptr) >> 3 );
et = 352 + ((bcount == 0) << 7);
UBI_BIG_8WAY( et, ptr );
memset_zero_512( buf, buf_size >> 3 );
bcount = 0;
UBI_BIG_8WAY( 510, 8 );
casti_m512i( out, 0 ) = h0;
casti_m512i( out, 1 ) = h1;
casti_m512i( out, 2 ) = h2;
casti_m512i( out, 3 ) = h3;
casti_m512i( out, 4 ) = h4;
casti_m512i( out, 5 ) = h5;
casti_m512i( out, 6 ) = h6;
casti_m512i( out, 7 ) = h7;
}
void
skein256_8way_update(void *cc, const void *data, size_t len)
{
@@ -709,6 +783,7 @@ void skein512_4way_init( skein512_4way_context *sc )
sc->ptr = 0;
}
// Do not use for 128 bt data length
static void
skein_big_core_4way( skein512_4way_context *sc, const void *data,
size_t len )
@@ -794,6 +869,79 @@ skein_big_close_4way( skein512_4way_context *sc, unsigned ub, unsigned n,
memcpy_256( dst, buf, out_len >> 3 );
}
void
skein512_4way_full( skein512_4way_context *sc, void *out, const void *data,
size_t len )
{
__m256i h0, h1, h2, h3, h4, h5, h6, h7;
__m256i *vdata = (__m256i*)data;
__m256i *buf = sc->buf;
size_t ptr = 0;
unsigned first;
const int buf_size = 64; // 64 * __m256i
uint64_t bcount = 0;
h0 = m256_const1_64( 0x4903ADFF749C51CE );
h1 = m256_const1_64( 0x0D95DE399746DF03 );
h2 = m256_const1_64( 0x8FD1934127C79BCE );
h3 = m256_const1_64( 0x9A255629FF352CB1 );
h4 = m256_const1_64( 0x5DB62599DF6CA7B0 );
h5 = m256_const1_64( 0xEABE394CA9D5C3F4 );
h6 = m256_const1_64( 0x991112C71A75B523 );
h7 = m256_const1_64( 0xAE18A40B660FCC33 );
// Update
if ( len <= buf_size - ptr )
{
memcpy_256( buf + (ptr>>3), vdata, len>>3 );
ptr += len;
}
else
{
first = ( bcount == 0 ) << 7;
do {
size_t clen;
if ( ptr == buf_size )
{
bcount ++;
UBI_BIG_4WAY( 96 + first, 0 );
first = 0;
ptr = 0;
}
clen = buf_size - ptr;
if ( clen > len )
clen = len;
memcpy_256( buf + (ptr>>3), vdata, clen>>3 );
ptr += clen;
vdata += (clen>>3);
len -= clen;
} while ( len > 0 );
}
// Close
unsigned et;
memset_zero_256( buf + (ptr>>3), (buf_size - ptr) >> 3 );
et = 352 + ((bcount == 0) << 7);
UBI_BIG_4WAY( et, ptr );
memset_zero_256( buf, buf_size >> 3 );
bcount = 0;
UBI_BIG_4WAY( 510, 8 );
casti_m256i( out, 0 ) = h0;
casti_m256i( out, 1 ) = h1;
casti_m256i( out, 2 ) = h2;
casti_m256i( out, 3 ) = h3;
casti_m256i( out, 4 ) = h4;
casti_m256i( out, 5 ) = h5;
casti_m256i( out, 6 ) = h6;
casti_m256i( out, 7 ) = h7;
}
void
skein256_4way_update(void *cc, const void *data, size_t len)
{
@@ -806,6 +954,9 @@ skein256_4way_close(void *cc, void *dst)
skein_big_close_4way(cc, 0, 0, dst, 32);
}
// Do not use with 128 bit data
void
skein512_4way_update(void *cc, const void *data, size_t len)
{

4
algo/skein/skein-hash-4way.h
View File

@@ -63,6 +63,8 @@ typedef struct
typedef skein_8way_big_context skein512_8way_context;
typedef skein_8way_big_context skein256_8way_context;
void skein512_8way_full( skein512_8way_context *sc, void *out,
const void *data, size_t len );
void skein512_8way_init( skein512_8way_context *sc );
void skein512_8way_update( void *cc, const void *data, size_t len );
void skein512_8way_close( void *cc, void *dst );
@@ -85,6 +87,8 @@ typedef skein_4way_big_context skein512_4way_context;
typedef skein_4way_big_context skein256_4way_context;
void skein512_4way_init( skein512_4way_context *sc );
void skein512_4way_full( skein512_4way_context *sc, void *out,
const void *data, size_t len );
void skein512_4way_update( void *cc, const void *data, size_t len );
void skein512_4way_close( void *cc, void *dst );

97
algo/skein/skein2-4way.c
View File

@@ -5,114 +5,131 @@
#if defined(SKEIN_8WAY)
// static __thread skein512_8way_context skein512_8way_ctx
// __attribute__ ((aligned (64)));
void skein2hash_8way( void *output, const void *input )
{
skein512_8way_context ctx;
uint64_t hash[16*8] __attribute__ ((aligned (128)));
skein512_8way_context ctx;
// memcpy( &ctx, &skein512_8way_ctx, sizeof( ctx ) );
skein512_8way_init( &ctx );
skein512_8way_update( &ctx, input, 80 );
skein512_8way_close( &ctx, hash );
skein512_8way_full( &ctx, hash, input, 80 );
skein512_8way_init( &ctx );
skein512_8way_update( &ctx, hash, 64 );
skein512_8way_close( &ctx, output );
// skein512_8way_update( &ctx, input + (64*8), 16 );
// skein512_8way_close( &ctx, hash );
skein512_8way_full( &ctx, output, hash, 64 );
}
int scanhash_skein2_8way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t hash[16*8] __attribute__ ((aligned (128)));
uint64_t hash[8*8] __attribute__ ((aligned (128)));
uint32_t vdata[20*8] __attribute__ ((aligned (64)));
uint32_t lane_hash[8] __attribute__ ((aligned (64)));
uint32_t *hash7 = &(hash[49]);
uint64_t *hashq3 = &(hash[3*8]);
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t Htarg = ptarget[7];
const uint64_t targq3 = ((uint64_t*)ptarget)[3];
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 8;
uint32_t n = first_nonce;
__m512i *noncev = (__m512i*)vdata + 9; // aligned
int thr_id = mythr->id;
__m512i *noncev = (__m512i*)vdata + 9;
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm512_bswap32_intrlv80_8x64( vdata, pdata );
*noncev = mm512_intrlv_blend_32(
_mm512_set_epi32( n+7, 0, n+6, 0, n+5, 0, n+4, 0,
n+3, 0, n+2, 0, n+1, 0, n , 0 ), *noncev );
// skein512_8way_init( &skein512_8way_ctx );
// skein512_8way_update( &skein512_8way_ctx, vdata, 64 );
do
{
*noncev = mm512_intrlv_blend_32( mm512_bswap_32(
_mm512_set_epi32( n+7, 0, n+6, 0, n+5, 0, n+4, 0,
n+3, 0, n+2, 0, n+1, 0, n , 0 ) ), *noncev );
skein2hash_8way( hash, vdata );
for ( int lane = 0; lane < 8; lane++ )
if ( hash7[ lane<<1 ] <= Htarg )
if ( unlikely( hashq3[ lane ] <= targq3 && !bench ) )
{
extr_lane_8x64( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) && !opt_benchmark )
if ( valid_hash( lane_hash, ptarget ) && !bench )
{
pdata[19] = n + lane;
pdata[19] = bswap_32( n + lane );
submit_lane_solution( work, lane_hash, mythr, lane );
}
}
*noncev = _mm512_add_epi32( *noncev,
m512_const1_64( 0x0000000800000000 ) );
n += 8;
} while ( (n < max_nonce-8) && !work_restart[thr_id].restart );
} while ( likely( (n < last_nonce) && !work_restart[thr_id].restart ) );
*hashes_done = n - first_nonce + 1;
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}
#elif defined(SKEIN_4WAY)
//static __thread skein512_4way_context skein512_4way_ctx
// __attribute__ ((aligned (64)));
void skein2hash_4way( void *output, const void *input )
{
skein512_4way_context ctx;
// memcpy( &ctx, &skein512_4way_ctx, sizeof( ctx ) );
uint64_t hash[16*4] __attribute__ ((aligned (64)));
skein512_4way_init( &ctx );
skein512_4way_update( &ctx, input, 80 );
skein512_4way_close( &ctx, hash );
// skein512_4way_update( &ctx, input + (64*4), 16 );
// skein512_4way_close( &ctx, hash );
skein512_4way_init( &ctx );
skein512_4way_update( &ctx, hash, 64 );
skein512_4way_close( &ctx, output );
skein512_4way_full( &ctx, hash, input, 80 );
skein512_4way_full( &ctx, output, hash, 64 );
}
int scanhash_skein2_4way( struct work *work, uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
uint32_t hash[16*4] __attribute__ ((aligned (64)));
uint64_t hash[8*4] __attribute__ ((aligned (64)));
uint32_t vdata[20*4] __attribute__ ((aligned (64)));
uint32_t lane_hash[8] __attribute__ ((aligned (64)));
uint32_t *hash7 = &(hash[25]);
uint64_t *hash_q3 = &(hash[3*4]);
uint32_t *pdata = work->data;
uint32_t *ptarget = work->target;
const uint32_t Htarg = ptarget[7];
const uint64_t targ_q3 = ((uint64_t*)ptarget)[3];
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 4;
uint32_t n = first_nonce;
__m256i *noncev = (__m256i*)vdata + 9; // aligned
int thr_id = mythr->id; // thr_id arg is deprecated
__m256i *noncev = (__m256i*)vdata + 9;
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
mm256_bswap32_intrlv80_4x64( vdata, pdata );
// skein512_4way_init( &skein512_4way_ctx );
// skein512_4way_update( &skein512_4way_ctx, vdata, 64 );
*noncev = mm256_intrlv_blend_32(
_mm256_set_epi32( n+3, 0, n+2, 0, n+1, 0, n, 0 ), *noncev );
do
{
*noncev = mm256_intrlv_blend_32( mm256_bswap_32(
_mm256_set_epi32( n+3, 0, n+2, 0, n+1, 0, n, 0 ) ), *noncev );
skein2hash_4way( hash, vdata );
for ( int lane = 0; lane < 4; lane++ )
if ( hash7[ lane<<1 ] <= Htarg )
if ( hash_q3[ lane ] <= targ_q3 )
{
extr_lane_4x64( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) && !opt_benchmark )
if ( valid_hash( lane_hash, ptarget ) && !bench )
{
pdata[19] = n + lane;
pdata[19] = bswap_32( n + lane );
submit_lane_solution( work, lane_hash, mythr, lane );
}
}
*noncev = _mm256_add_epi32( *noncev,
m256_const1_64( 0x0000000400000000 ) );
n += 4;
} while ( (n < max_nonce) && !work_restart[thr_id].restart );
} while ( (n < last_nonce) && !work_restart[thr_id].restart );
*hashes_done = n - first_nonce + 1;
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}