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

v3.11.1

Browse Source
This commit is contained in:
Jay D Dee
2020-01-06 13:55:17 -05:00
parent 3572cb53c4
commit 82c2605d77
84 changed files with 1546 additions and 14296 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
algo/echo/aes_ni/hash.c
View File

@@ -179,53 +179,53 @@ void Compress(hashState_echo *ctx, const unsigned char *pmsg, unsigned int uBloc
for(b = 0; b < uBlockCount; b++)
{
ctx->k = _mm_add_epi64(ctx->k, ctx->const1536);
ctx->k = _mm_add_epi64(ctx->k, ctx->const1536);
// load message
for(j = ctx->uHashSize / 256; j < 4; j++)
{
for(i = 0; i < 4; i++)
// load message
for(j = ctx->uHashSize / 256; j < 4; j++)
{
_state[i][j] = _mm_load_si128((__m128i*)pmsg + 4 * (j - (ctx->uHashSize / 256)) + i);
for(i = 0; i < 4; i++)
{
_state[i][j] = _mm_load_si128((__m128i*)pmsg + 4 * (j - (ctx->uHashSize / 256)) + i);
}
}
}
// save state
SAVESTATE(_statebackup, _state);
// save state
SAVESTATE(_statebackup, _state);
k1 = ctx->k;
k1 = ctx->k;
for(r = 0; r < ctx->uRounds / 2; r++)
{
ECHO_ROUND_UNROLL2;
}
for(r = 0; r < ctx->uRounds / 2; r++)
{
ECHO_ROUND_UNROLL2;
}
if(ctx->uHashSize == 256)
{
for(i = 0; i < 4; i++)
if(ctx->uHashSize == 256)
{
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][1]);
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][2]);
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][3]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][0]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][1]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][2]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][3]);
for(i = 0; i < 4; i++)
{
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][1]);
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][2]);
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][3]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][0]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][1]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][2]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][3]);
}
}
}
else
{
for(i = 0; i < 4; i++)
{
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][2]);
_state[i][1] = _mm_xor_si128(_state[i][1], _state[i][3]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][0]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][2]);
_state[i][1] = _mm_xor_si128(_state[i][1], _statebackup[i][1]);
_state[i][1] = _mm_xor_si128(_state[i][1], _statebackup[i][3]);
}
}
pmsg += ctx->uBlockLength;
else
{
for(i = 0; i < 4; i++)
{
_state[i][0] = _mm_xor_si128(_state[i][0], _state[i][2]);
_state[i][1] = _mm_xor_si128(_state[i][1], _state[i][3]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][0]);
_state[i][0] = _mm_xor_si128(_state[i][0], _statebackup[i][2]);
_state[i][1] = _mm_xor_si128(_state[i][1], _statebackup[i][1]);
_state[i][1] = _mm_xor_si128(_state[i][1], _statebackup[i][3]);
}
}
pmsg += ctx->uBlockLength;
}
SAVESTATE(ctx->state, _state);

33
algo/echo/echo-hash-4way.c
View File

@@ -277,41 +277,40 @@ int echo_4way_update_close( echo_4way_context *state, void *hashval,
{
echo_4way_compress( state, data, 1 );
state->processed_bits = 1024;
remainingbits = m512_zero;
remainingbits = m512_const2_64( 0, -1024 );
vlen = 0;
}
else
{
vlen = databitlen / 128; // * 4 lanes / 128 bits per lane
memcpy_512( state->buffer, data, vlen );
state->processed_bits += (unsigned int)( databitlen );
remainingbits = _mm512_set4_epi32( 0, 0, 0, databitlen );
}
state->buffer[ vlen ] = _mm512_set4_epi32( 0, 0, 0, 0x80 );
memset_zero_512( state->buffer + vlen + 1, vblen - vlen - 2 );
state->buffer[ vblen-2 ] =
state->buffer[ vlen ] = _mm512_set4_epi32( 0, 0, 0, 0x80 );
memset_zero_512( state->buffer + vlen + 1, vblen - vlen - 2 );
state->buffer[ vblen-2 ] =
_mm512_set4_epi32( (uint32_t)state->uHashSize << 16, 0, 0, 0 );
state->buffer[ vblen-1 ] =
state->buffer[ vblen-1 ] =
_mm512_set4_epi64( 0, state->processed_bits,
0, state->processed_bits );
state->k = _mm512_add_epi64( state->k, remainingbits );
state->k = _mm512_sub_epi64( state->k, state->const1536 );
state->k = _mm512_add_epi64( state->k, remainingbits );
state->k = _mm512_sub_epi64( state->k, state->const1536 );
echo_4way_compress( state, state->buffer, 1 );
echo_4way_compress( state, state->buffer, 1 );
_mm512_store_si512( (__m512i*)hashval + 0, state->state[ 0 ][ 0] );
_mm512_store_si512( (__m512i*)hashval + 1, state->state[ 1 ][ 0] );
_mm512_store_si512( (__m512i*)hashval + 0, state->state[ 0 ][ 0] );
_mm512_store_si512( (__m512i*)hashval + 1, state->state[ 1 ][ 0] );
if ( state->uHashSize == 512 )
{
_mm512_store_si512( (__m512i*)hashval + 2, state->state[ 2 ][ 0 ] );
_mm512_store_si512( (__m512i*)hashval + 3, state->state[ 3 ][ 0 ] );
}
return 0;
if ( state->uHashSize == 512 )
{
_mm512_store_si512( (__m512i*)hashval + 2, state->state[ 2 ][ 0 ] );
_mm512_store_si512( (__m512i*)hashval + 3, state->state[ 3 ][ 0 ] );
}
return 0;
}
#endif