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

v3.6.5

Browse Source
This commit is contained in:
Jay D Dee
2017-05-19 16:38:26 -04:00
parent e7dbd27636
commit 7544cb956c
8 changed files with 314 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

221
algo/cryptonight/cryptonight-aesni.c
View File

@@ -109,43 +109,43 @@ static __thread cryptonight_ctx ctx;
void cryptonight_hash_aes( void *restrict output, const void *input, int len )
{
#ifndef NO_AES_NI
keccak( (const uint8_t*)input, 76, (char*)&ctx.state.hs.b, 200 );
uint8_t ExpandedKey[256] __attribute__((aligned(64)));
__m128i *longoutput, *expkey, *xmminput;
size_t i, j;
memcpy(ctx.text, ctx.state.init, INIT_SIZE_BYTE);
memcpy(ExpandedKey, ctx.state.hs.b, AES_KEY_SIZE);
ExpandAESKey256(ExpandedKey);
keccak( (const uint8_t*)input, 76, (char*)&ctx.state.hs.b, 200 );
memcpy( ExpandedKey, ctx.state.hs.b, AES_KEY_SIZE );
ExpandAESKey256( ExpandedKey );
memcpy( ctx.text, ctx.state.init, INIT_SIZE_BYTE );
__m128i *longoutput, *expkey, *xmminput;
longoutput = (__m128i *)ctx.long_state;
expkey = (__m128i *)ExpandedKey;
xmminput = (__m128i *)ctx.text;
longoutput = (__m128i*)ctx.long_state;
xmminput = (__m128i*)ctx.text;
expkey = (__m128i*)ExpandedKey;
//for (i = 0; likely(i < MEMORY); i += INIT_SIZE_BYTE)
// aesni_parallel_noxor(&ctx->long_state[i], ctx->text, ExpandedKey);
// prefetch expkey, all of xmminput and enough longoutput for 4 loops
// prefetch expkey, xmminput and enough longoutput for 4 iterations
_mm_prefetch( xmminput, _MM_HINT_T0 );
_mm_prefetch( xmminput + 4, _MM_HINT_T0 );
for ( i = 0; i < 64; i += 16 )
{
_mm_prefetch( longoutput + i, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 4, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 8, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 12, _MM_HINT_T0 );
}
_mm_prefetch( expkey, _MM_HINT_T0 );
_mm_prefetch( expkey + 4, _MM_HINT_T0 );
_mm_prefetch( expkey + 8, _MM_HINT_T0 );
for ( i = 0; likely( i < MEMORY_M128I ); i += INIT_SIZE_M128I )
for ( i = 0; i < 64; i += 16 )
{
// prefetch 4 loops ahead,
__builtin_prefetch( longoutput + i, 1, 0 );
__builtin_prefetch( longoutput + i + 4, 1, 0 );
__builtin_prefetch( longoutput + i + 8, 1, 0 );
__builtin_prefetch( longoutput + i + 12, 1, 0 );
}
// n-4 iterations
for ( i = 0; likely( i < MEMORY_M128I - 4*INIT_SIZE_M128I );
i += INIT_SIZE_M128I )
{
// prefetch 4 iterations ahead.
__builtin_prefetch( longoutput + i + 64, 1, 0 );
__builtin_prefetch( longoutput + i + 68, 1, 0 );
for (j = 0; j < 10; j++ )
for ( j = 0; j < 10; j++ )
{
xmminput[0] = _mm_aesenc_si128( xmminput[0], expkey[j] );
xmminput[1] = _mm_aesenc_si128( xmminput[1], expkey[j] );
@@ -165,84 +165,99 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
_mm_store_si128( &( longoutput[i+6] ), xmminput[6] );
_mm_store_si128( &( longoutput[i+7] ), xmminput[7] );
}
// last 4 iterations
for ( ; likely( i < MEMORY_M128I ); i += INIT_SIZE_M128I )
{
for ( j = 0; j < 10; j++ )
{
xmminput[0] = _mm_aesenc_si128( xmminput[0], expkey[j] );
xmminput[1] = _mm_aesenc_si128( xmminput[1], expkey[j] );
xmminput[2] = _mm_aesenc_si128( xmminput[2], expkey[j] );
xmminput[3] = _mm_aesenc_si128( xmminput[3], expkey[j] );
xmminput[4] = _mm_aesenc_si128( xmminput[4], expkey[j] );
xmminput[5] = _mm_aesenc_si128( xmminput[5], expkey[j] );
xmminput[6] = _mm_aesenc_si128( xmminput[6], expkey[j] );
xmminput[7] = _mm_aesenc_si128( xmminput[7], expkey[j] );
}
_mm_store_si128( &( longoutput[i ] ), xmminput[0] );
_mm_store_si128( &( longoutput[i+1] ), xmminput[1] );
_mm_store_si128( &( longoutput[i+2] ), xmminput[2] );
_mm_store_si128( &( longoutput[i+3] ), xmminput[3] );
_mm_store_si128( &( longoutput[i+4] ), xmminput[4] );
_mm_store_si128( &( longoutput[i+5] ), xmminput[5] );
_mm_store_si128( &( longoutput[i+6] ), xmminput[6] );
_mm_store_si128( &( longoutput[i+7] ), xmminput[7] );
}
// cast_m128i( ctx.a ) = _mm_xor_si128( casti_m128i( ctx.state.k, 0 ) ,
// casti_m128i( ctx.state.k, 2 ) );
// cast_m128i( ctx.b ) = _mm_xor_si128( casti_m128i( ctx.state.k, 1 ),
// casti_m128i( ctx.state.k, 3 ) );
ctx.a[0] = ((uint64_t *)ctx.state.k)[0] ^ ((uint64_t *)ctx.state.k)[4];
ctx.b[0] = ((uint64_t *)ctx.state.k)[2] ^ ((uint64_t *)ctx.state.k)[6];
ctx.a[1] = ((uint64_t *)ctx.state.k)[1] ^ ((uint64_t *)ctx.state.k)[5];
ctx.b[1] = ((uint64_t *)ctx.state.k)[3] ^ ((uint64_t *)ctx.state.k)[7];
ctx.a[0] = ((uint64_t *)ctx.state.k)[0] ^ ((uint64_t *)ctx.state.k)[4];
ctx.b[0] = ((uint64_t *)ctx.state.k)[2] ^ ((uint64_t *)ctx.state.k)[6];
ctx.a[1] = ((uint64_t *)ctx.state.k)[1] ^ ((uint64_t *)ctx.state.k)[5];
ctx.b[1] = ((uint64_t *)ctx.state.k)[3] ^ ((uint64_t *)ctx.state.k)[7];
// for (i = 0; i < 2; i++)
// {
// ctx.a[i] = ((uint64_t *)ctx.state.k)[i] ^ ((uint64_t *)ctx.state.k)[i+4];
// ctx.b[i] = ((uint64_t *)ctx.state.k)[i+2] ^ ((uint64_t *)ctx.state.k)[i+6];
// }
__m128i b_x = _mm_load_si128((__m128i *)ctx.b);
uint64_t a[2] __attribute((aligned(16))), b[2] __attribute((aligned(16)));
uint64_t a[2] __attribute((aligned(16))),
b[2] __attribute((aligned(16))),
c[2] __attribute((aligned(16)));
a[0] = ctx.a[0];
a[1] = ctx.a[1];
for(i = 0; __builtin_expect(i < 0x80000, 1); i++)
__m128i b_x = _mm_load_si128( (__m128i*)ctx.b );
__m128i a_x = _mm_load_si128( (__m128i*)a );
__m128i* lsa = (__m128i*)&ctx.long_state[ a[0] & 0x1FFFF0 ];
__m128i c_x = _mm_load_si128( lsa );
uint64_t *nextblock;
uint64_t hi, lo;
// n-1 iterations
for( i = 0; __builtin_expect( i < 0x7ffff, 1 ); i++ )
{
uint64_t c[2];
__builtin_prefetch( &ctx.long_state[c[0] & 0x1FFFF0], 0, 1 );
__m128i c_x = _mm_load_si128(
(__m128i *)&ctx.long_state[a[0] & 0x1FFFF0]);
__m128i a_x = _mm_load_si128((__m128i *)a);
c_x = _mm_aesenc_si128(c_x, a_x);
_mm_store_si128((__m128i *)c, c_x);
b_x = _mm_xor_si128(b_x, c_x);
_mm_store_si128((__m128i *)&ctx.long_state[a[0] & 0x1FFFF0], b_x);
uint64_t *nextblock = (uint64_t *)&ctx.long_state[c[0] & 0x1FFFF0];
// uint64_t b[2];
c_x = _mm_aesenc_si128( c_x, a_x );
_mm_store_si128( (__m128i*)c, c_x );
b_x = _mm_xor_si128( b_x, c_x );
nextblock = (uint64_t *)&ctx.long_state[c[0] & 0x1FFFF0];
_mm_store_si128( lsa, b_x );
b[0] = nextblock[0];
b[1] = nextblock[1];
{
uint64_t hi, lo;
// hi,lo = 64bit x 64bit multiply of c[0] and b[0]
// hi,lo = 64bit x 64bit multiply of c[0] and b[0]
__asm__( "mulq %3\n\t"
: "=d" ( hi ),
"=a" ( lo )
: "%a" ( c[0] ),
"rm" ( b[0] )
: "cc" );
__asm__("mulq %3\n\t"
: "=d" (hi),
"=a" (lo)
: "%a" (c[0]),
"rm" (b[0])
: "cc" );
a[0] += hi;
a[1] += lo;
}
uint64_t *dst = (uint64_t*)&ctx.long_state[c[0] & 0x1FFFF0];
// __m128i *dst = (__m128i*)&ctx.long_state[c[0] & 0x1FFFF0];
// *dst = cast_m128i( a );
dst[0] = a[0];
dst[1] = a[1];
// cast_m128i( a ) = _mm_xor_si128( cast_m128i( a ), cast_m128i( b ) );
a[0] ^= b[0];
a[1] ^= b[1];
b_x = c_x;
__builtin_prefetch( &ctx.long_state[a[0] & 0x1FFFF0], 0, 3 );
b_x = c_x;
nextblock[0] = a[0] + hi;
nextblock[1] = a[1] + lo;
a[0] = b[0] ^ nextblock[0];
a[1] = b[1] ^ nextblock[1];
lsa = (__m128i*)&ctx.long_state[ a[0] & 0x1FFFF0 ];
a_x = _mm_load_si128( (__m128i*)a );
c_x = _mm_load_si128( lsa );
}
// abreviated nth iteration
c_x = _mm_aesenc_si128( c_x, a_x );
_mm_store_si128( (__m128i*)c, c_x );
b_x = _mm_xor_si128( b_x, c_x );
nextblock = (uint64_t *)&ctx.long_state[c[0] & 0x1FFFF0];
_mm_store_si128( lsa, b_x );
b[0] = nextblock[0];
b[1] = nextblock[1];
__asm__( "mulq %3\n\t"
: "=d" ( hi ),
"=a" ( lo )
: "%a" ( c[0] ),
"rm" ( b[0] )
: "cc" );
nextblock[0] = a[0] + hi;
nextblock[1] = a[1] + lo;
memcpy( ctx.text, ctx.state.init, INIT_SIZE_BYTE );
memcpy( ExpandedKey, &ctx.state.hs.b[32], AES_KEY_SIZE );
ExpandAESKey256( ExpandedKey );
//for (i = 0; likely(i < MEMORY); i += INIT_SIZE_BYTE)
// aesni_parallel_xor(&ctx->text, ExpandedKey, &ctx->long_state[i]);
memcpy( ctx.text, ctx.state.init, INIT_SIZE_BYTE );
// prefetch expkey, all of xmminput and enough longoutput for 4 loops
_mm_prefetch( xmminput, _MM_HINT_T0 );
_mm_prefetch( xmminput + 4, _MM_HINT_T0 );
for ( i = 0; i < 64; i += 16 )
@@ -256,9 +271,11 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
_mm_prefetch( expkey + 4, _MM_HINT_T0 );
_mm_prefetch( expkey + 8, _MM_HINT_T0 );
for ( i = 0; likely( i < MEMORY_M128I ); i += INIT_SIZE_M128I )
// n-4 iterations
for ( i = 0; likely( i < MEMORY_M128I - 4*INIT_SIZE_M128I );
i += INIT_SIZE_M128I )
{
// stay 4 loops ahead,
// stay 4 iterations ahead.
_mm_prefetch( longoutput + i + 64, _MM_HINT_T0 );
_mm_prefetch( longoutput + i + 68, _MM_HINT_T0 );
@@ -283,10 +300,34 @@ void cryptonight_hash_aes( void *restrict output, const void *input, int len )
xmminput[7] = _mm_aesenc_si128( xmminput[7], expkey[j] );
}
}
// last 4 iterations
for ( ; likely( i < MEMORY_M128I ); i += INIT_SIZE_M128I )
{
xmminput[0] = _mm_xor_si128( longoutput[i ], xmminput[0] );
xmminput[1] = _mm_xor_si128( longoutput[i+1], xmminput[1] );
xmminput[2] = _mm_xor_si128( longoutput[i+2], xmminput[2] );
xmminput[3] = _mm_xor_si128( longoutput[i+3], xmminput[3] );
xmminput[4] = _mm_xor_si128( longoutput[i+4], xmminput[4] );
xmminput[5] = _mm_xor_si128( longoutput[i+5], xmminput[5] );
xmminput[6] = _mm_xor_si128( longoutput[i+6], xmminput[6] );
xmminput[7] = _mm_xor_si128( longoutput[i+7], xmminput[7] );
for( j = 0; j < 10; j++ )
{
xmminput[0] = _mm_aesenc_si128( xmminput[0], expkey[j] );
xmminput[1] = _mm_aesenc_si128( xmminput[1], expkey[j] );
xmminput[2] = _mm_aesenc_si128( xmminput[2], expkey[j] );
xmminput[3] = _mm_aesenc_si128( xmminput[3], expkey[j] );
xmminput[4] = _mm_aesenc_si128( xmminput[4], expkey[j] );
xmminput[5] = _mm_aesenc_si128( xmminput[5], expkey[j] );
xmminput[6] = _mm_aesenc_si128( xmminput[6], expkey[j] );
xmminput[7] = _mm_aesenc_si128( xmminput[7], expkey[j] );
}
}
memcpy( ctx.state.init, ctx.text, INIT_SIZE_BYTE);
keccakf( (uint64_t*)&ctx.state.hs.w, 24 );
extra_hashes[ctx.state.hs.b[0] & 3](&ctx.state, 200, output);
#endif
}