Files
cpuminer-opt-gpu/algo/sha/sha256t-4way.c
Jay D Dee d6b5750362 v3.23.1
2023-09-13 11:48:52 -04:00

290 lines
10 KiB
C

#include "sha256t-gate.h"
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <stdio.h>
#include "sha256-hash.h"
#include "sha-hash-4way.h"
#if defined(SHA256T_16WAY)
int scanhash_sha256t_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 buf[16] __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)));
static const uint32_t IV[8] __attribute__ ((aligned (32))) =
{
0x6A09E667, 0xBB67AE85, 0x3C6EF372, 0xA54FF53A,
0x510E527F, 0x9B05688C, 0x1F83D9AB, 0x5BE0CD19
};
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 first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 16;
const __m512i last_byte = _mm512_set1_epi32( 0x80000000 );
uint32_t n = first_nonce;
const int thr_id = mythr->id;
const __m512i sixteen = _mm512_set1_epi32( 16 );
const bool bench = opt_benchmark;
const __m256i bswap_shuf = mm256_bcast_m128( _mm_set_epi64x(
0x0c0d0e0f08090a0b, 0x0405060700010203 ) );
// prehash first block directly from pdata
sha256_transform_le( phash, pdata, IV );
// vectorize block 0 hash for second block
mstate1[0] = _mm512_set1_epi32( phash[0] );
mstate1[1] = _mm512_set1_epi32( phash[1] );
mstate1[2] = _mm512_set1_epi32( phash[2] );
mstate1[3] = _mm512_set1_epi32( phash[3] );
mstate1[4] = _mm512_set1_epi32( phash[4] );
mstate1[5] = _mm512_set1_epi32( phash[5] );
mstate1[6] = _mm512_set1_epi32( phash[6] );
mstate1[7] = _mm512_set1_epi32( phash[7] );
// 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] );
buf[3] = _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 );
buf[4] = last_byte;
memset_zero_512( buf+5, 10 );
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
istate[0] = _mm512_set1_epi32( IV[0] );
istate[1] = _mm512_set1_epi32( IV[1] );
istate[2] = _mm512_set1_epi32( IV[2] );
istate[3] = _mm512_set1_epi32( IV[3] );
istate[4] = _mm512_set1_epi32( IV[4] );
istate[5] = _mm512_set1_epi32( IV[5] );
istate[6] = _mm512_set1_epi32( IV[6] );
istate[7] = _mm512_set1_epi32( IV[7] );
// initialize padding for 2nd & 3rd sha256
block[ 8] = last_byte;
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 );
sha256_16way_transform_le( block, block, istate );
if ( 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 );
if ( likely( valid_hash( phash, ptarget ) && !bench ) )
{
pdata[19] = n + lane;
submit_solution( work, phash, mythr );
}
}
}
buf[3] = _mm512_add_epi32( buf[3], 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(SHA256T_8WAY)
int scanhash_sha256t_8way( struct work *work, const uint32_t max_nonce,
uint64_t *hashes_done, struct thr_info *mythr )
{
__m256i vdata[32] __attribute__ ((aligned (64)));
__m256i block[16] __attribute__ ((aligned (32)));
__m256i hash32[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 *pdata = work->data;
const uint32_t *ptarget = work->target;
const uint32_t first_nonce = pdata[19];
const uint32_t last_nonce = max_nonce - 8;
uint32_t n = first_nonce;
__m256i *noncev = vdata + 19;
const int thr_id = mythr->id;
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] );
*noncev = _mm256_set_epi32( n+ 7, n+ 6, n+ 5, n+ 4, n+ 3, n+ 2, n+1, n );
vdata[16+4] = last_byte;
memset_zero_256( vdata+16 + 5, 10 );
vdata[16+15] = _mm256_set1_epi32( 80*8 ); // bit count
block[ 8] = last_byte;
memset_zero_256( block + 9, 6 );
block[15] = _mm256_set1_epi32( 32*8 ); // bit count
// initialize state
istate[0] = _mm256_set1_epi64x( 0x6A09E6676A09E667 );
istate[1] = _mm256_set1_epi64x( 0xBB67AE85BB67AE85 );
istate[2] = _mm256_set1_epi64x( 0x3C6EF3723C6EF372 );
istate[3] = _mm256_set1_epi64x( 0xA54FF53AA54FF53A );
istate[4] = _mm256_set1_epi64x( 0x510E527F510E527F );
istate[5] = _mm256_set1_epi64x( 0x9B05688C9B05688C );
istate[6] = _mm256_set1_epi64x( 0x1F83D9AB1F83D9AB );
istate[7] = _mm256_set1_epi64x( 0x5BE0CD195BE0CD19 );
sha256_8way_transform_le( mstate1, vdata, istate );
// Do 3 rounds on the first 12 bytes of the next block
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, mstate1, mstate2,
mexp_pre );
// 2. 32 byte hash from 1.
sha256_8way_transform_le( block, block, istate );
// 3. 32 byte hash from 2.
if ( unlikely( sha256_8way_transform_le_short(
hash32, block, istate, ptarget ) ) )
{
for ( int lane = 0; lane < 8; lane++ )
{
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;
} while ( (n < last_nonce) && !work_restart[thr_id].restart );
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}
#endif
#if defined(SHA256T_4WAY)
int scanhash_sha256t_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 istate[8] __attribute__ ((aligned (32)));
__m128i mstate[8] __attribute__ ((aligned (32)));
// __m128i mstate2[8] __attribute__ ((aligned (32)));
// __m128i 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 - 4;
uint32_t n = first_nonce;
__m128i *noncev = vdata + 19;
const int thr_id = mythr->id;
const bool bench = opt_benchmark;
const __m128i last_byte = _mm_set1_epi32( 0x80000000 );
const __m128i four = _mm_set1_epi32( 4 );
for ( int i = 0; i < 19; i++ )
vdata[i] = _mm_set1_epi32( pdata[i] );
*noncev = _mm_set_epi32( n+ 3, n+ 2, n+1, n );
vdata[16+4] = last_byte;
memset_zero_128( vdata+16 + 5, 10 );
vdata[16+15] = _mm_set1_epi32( 80*8 ); // bit count
block[ 8] = last_byte;
memset_zero_128( block + 9, 6 );
block[15] = _mm_set1_epi32( 32*8 ); // bit count
// initialize state
istate[0] = _mm_set1_epi64x( 0x6A09E6676A09E667 );
istate[1] = _mm_set1_epi64x( 0xBB67AE85BB67AE85 );
istate[2] = _mm_set1_epi64x( 0x3C6EF3723C6EF372 );
istate[3] = _mm_set1_epi64x( 0xA54FF53AA54FF53A );
istate[4] = _mm_set1_epi64x( 0x510E527F510E527F );
istate[5] = _mm_set1_epi64x( 0x9B05688C9B05688C );
istate[6] = _mm_set1_epi64x( 0x1F83D9AB1F83D9AB );
istate[7] = _mm_set1_epi64x( 0x5BE0CD195BE0CD19 );
// hash first 64 bytes of data
sha256_4way_transform_le( mstate, vdata, istate );
// sha256_4way_prehash_3rounds( mstate2, mexp_pre, vdata + 16, mstate1 );
do
{
// sha256_4way_final_rounds( block, vdata+16, mstate1, mstate2,
// mexp_pre );
sha256_4way_transform_le( block, vdata+16, mstate );
sha256_4way_transform_le( block, block, istate );
sha256_4way_transform_le( hash32, block, istate );
// if ( unlikely( sha256_4way_transform_le_short(
// hash32, block, initstate, ptarget ) ))
// {
mm128_block_bswap_32( hash32, hash32 );
for ( int lane = 0; lane < 4; lane++ )
if ( unlikely( hash32_d7[ lane ] <= targ32_d7 ) )
{
extr_lane_4x32( lane_hash, hash32, lane, 256 );
if ( likely( valid_hash( lane_hash, ptarget ) && !bench ) )
{
pdata[19] = n + lane;
submit_solution( work, lane_hash, mythr );
}
}
// }
*noncev = _mm_add_epi32( *noncev, four );
n += 4;
} while ( (n < last_nonce) && !work_restart[thr_id].restart );
pdata[19] = n;
*hashes_done = n - first_nonce;
return 0;
}
#endif