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

v3.9.0

Browse Source
This commit is contained in:
Jay D Dee
2019-05-19 13:39:45 -04:00
parent bfd1c002f9
commit e1aead3c76
139 changed files with 10907 additions and 4218 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
algo/skein/skein-4way.c
View File

@@ -21,9 +21,10 @@ void skeinhash_4way( void *state, const void *input )
sha256_4way_init( &ctx_sha256 );
sha256_4way( &ctx_sha256, vhash32, 64 );
sha256_4way_close( &ctx_sha256, vhash32 );
sha256_4way_close( &ctx_sha256, state );
mm_deinterleave_4x32( state, state+32, state+64, state+96, vhash32, 256 );
mm128_deinterleave_4x32( state, state+32, state+64, state+96,
vhash32, 256 );
}
int scanhash_skein_4way( int thr_id, struct work *work, uint32_t max_nonce,
@@ -31,6 +32,8 @@ int scanhash_skein_4way( int thr_id, struct work *work, uint32_t max_nonce,
{
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;
@@ -58,12 +61,16 @@ int scanhash_skein_4way( int thr_id, struct work *work, uint32_t max_nonce,
skeinhash_4way( hash, vdata );
for ( int i = 0; i < 4; i++ )
if ( (hash+(i<<3))[7] <= Htarg && fulltest( hash+(i<<3), ptarget ) )
for ( int lane = 0; lane < 4; lane++ )
if ( hash7[ lane ] <= Htarg )
{
pdata[19] = n+i;
nonces[ num_found++ ] = n+i;
work_set_target_ratio( work, hash+(i<<3) );
mm128_extract_lane_4x32( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) )
{
pdata[19] = n + lane;
nonces[ num_found++ ] = n + lane;
work_set_target_ratio( work, lane_hash );
}
}
n += 4;
} while ( (num_found == 0) && (n < max_nonce)

23
algo/skein/skein2-4way.c
View File

@@ -9,7 +9,6 @@ void skein2hash_4way( void *output, const void *input )
{
skein512_4way_context ctx;
uint64_t hash[8*4] __attribute__ ((aligned (64)));
uint64_t *out64 = (uint64_t*)output;
skein512_4way_init( &ctx );
skein512_4way( &ctx, input, 80 );
@@ -17,15 +16,14 @@ void skein2hash_4way( void *output, const void *input )
skein512_4way_init( &ctx );
skein512_4way( &ctx, hash, 64 );
skein512_4way_close( &ctx, hash );
mm256_deinterleave_4x64( out64, out64+4, out64+8, out64+12, hash, 256 );
skein512_4way_close( &ctx, output );
}
int scanhash_skein2_4way( int thr_id, struct work *work, uint32_t max_nonce,
uint64_t *hashes_done )
{
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;
@@ -34,7 +32,6 @@ int scanhash_skein2_4way( int thr_id, struct work *work, uint32_t max_nonce,
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;
@@ -53,12 +50,18 @@ int scanhash_skein2_4way( int thr_id, struct work *work, uint32_t max_nonce,
skein2hash( hash, vdata );
for ( int i = 0; i < 4; i++ )
if ( (hash+(i<<3))[7] <= Htarg && fulltest( hash+(i<<3), ptarget ) )
for ( int lane = 0; lane < 4; lane++ )
if ( hash7[ lane ] <= Htarg )
{
pdata[19] = n+i;
nonces[ num_found++ ] = n+i;
work_set_target_ratio( work, hash+(i<<3) );
// deinterleave hash for lane
uint32_t lane_hash[8];
mm256_extract_lane_4x64( lane_hash, hash, lane, 256 );
if ( fulltest( lane_hash, ptarget ) )
{
pdata[19] = n + lane;
nonces[ num_found++ ] = n + lane;
work_set_target_ratio( work, lane_hash );
}
}
n += 4;
} while ( (num_found == 0) && (n < max_nonce)