mirror of
https://github.com/JayDDee/cpuminer-opt.git
synced 2025-09-17 23:44:27 +00:00
v3.10.6
This commit is contained in:
@@ -623,3 +623,303 @@ void ripemd160_8way_close( ripemd160_8way_context *sc, void *dst )
|
||||
|
||||
#endif // __AVX2__
|
||||
|
||||
#if defined(__AVX512F__) && defined(__AVX512VL__) && defined(__AVX512DQ__) && defined(__AVX512BW__)
|
||||
|
||||
// RIPEMD-160 16 way
|
||||
|
||||
|
||||
#define F16W_1(x, y, z) \
|
||||
_mm512_xor_si512( _mm512_xor_si512( x, y ), z )
|
||||
|
||||
#define F16W_2(x, y, z) \
|
||||
_mm512_xor_si512( _mm512_and_si512( _mm512_xor_si512( y, z ), x ), z )
|
||||
|
||||
#define F16W_3(x, y, z) \
|
||||
_mm512_xor_si512( _mm512_or_si512( x, mm512_not( y ) ), z )
|
||||
|
||||
#define F16W_4(x, y, z) \
|
||||
_mm512_xor_si512( _mm512_and_si512( _mm512_xor_si512( x, y ), z ), y )
|
||||
|
||||
#define F16W_5(x, y, z) \
|
||||
_mm512_xor_si512( x, _mm512_or_si512( y, mm512_not( z ) ) )
|
||||
|
||||
#define RR_16W(a, b, c, d, e, f, s, r, k) \
|
||||
do{ \
|
||||
a = _mm512_add_epi32( mm512_rol_32( _mm512_add_epi32( _mm512_add_epi32( \
|
||||
_mm512_add_epi32( a, f( b ,c, d ) ), r ), \
|
||||
m512_const1_64( k ) ), s ), e ); \
|
||||
c = mm512_rol_32( c, 10 );\
|
||||
} while (0)
|
||||
|
||||
#define ROUND1_16W(a, b, c, d, e, f, s, r, k) \
|
||||
RR_16W(a ## 1, b ## 1, c ## 1, d ## 1, e ## 1, f, s, r, K1 ## k)
|
||||
|
||||
#define ROUND2_16W(a, b, c, d, e, f, s, r, k) \
|
||||
RR_16W(a ## 2, b ## 2, c ## 2, d ## 2, e ## 2, f, s, r, K2 ## k)
|
||||
|
||||
static void ripemd160_16way_round( ripemd160_16way_context *sc )
|
||||
{
|
||||
const __m512i *in = (__m512i*)sc->buf;
|
||||
__m512i *h = (__m512i*)sc->val;
|
||||
register __m512i A1, B1, C1, D1, E1;
|
||||
register __m512i A2, B2, C2, D2, E2;
|
||||
__m512i tmp;
|
||||
|
||||
A1 = A2 = h[0];
|
||||
B1 = B2 = h[1];
|
||||
C1 = C2 = h[2];
|
||||
D1 = D2 = h[3];
|
||||
E1 = E2 = h[4];
|
||||
|
||||
ROUND1_16W( A, B, C, D, E, F16W_1, 11, in[ 0], 1 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_1, 14, in[ 1], 1 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_1, 15, in[ 2], 1 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_1, 12, in[ 3], 1 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_1, 5, in[ 4], 1 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_1, 8, in[ 5], 1 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_1, 7, in[ 6], 1 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_1, 9, in[ 7], 1 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_1, 11, in[ 8], 1 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_1, 13, in[ 9], 1 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_1, 14, in[10], 1 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_1, 15, in[11], 1 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_1, 6, in[12], 1 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_1, 7, in[13], 1 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_1, 9, in[14], 1 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_1, 8, in[15], 1 );
|
||||
|
||||
ROUND1_16W( E, A, B, C, D, F16W_2, 7, in[ 7], 2 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_2, 6, in[ 4], 2 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_2, 8, in[13], 2 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_2, 13, in[ 1], 2 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_2, 11, in[10], 2 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_2, 9, in[ 6], 2 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_2, 7, in[15], 2 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_2, 15, in[ 3], 2 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_2, 7, in[12], 2 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_2, 12, in[ 0], 2 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_2, 15, in[ 9], 2 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_2, 9, in[ 5], 2 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_2, 11, in[ 2], 2 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_2, 7, in[14], 2 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_2, 13, in[11], 2 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_2, 12, in[ 8], 2 );
|
||||
|
||||
ROUND1_16W( D, E, A, B, C, F16W_3, 11, in[ 3], 3 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_3, 13, in[10], 3 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_3, 6, in[14], 3 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_3, 7, in[ 4], 3 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_3, 14, in[ 9], 3 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_3, 9, in[15], 3 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_3, 13, in[ 8], 3 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_3, 15, in[ 1], 3 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_3, 14, in[ 2], 3 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_3, 8, in[ 7], 3 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_3, 13, in[ 0], 3 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_3, 6, in[ 6], 3 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_3, 5, in[13], 3 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_3, 12, in[11], 3 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_3, 7, in[ 5], 3 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_3, 5, in[12], 3 );
|
||||
|
||||
ROUND1_16W( C, D, E, A, B, F16W_4, 11, in[ 1], 4 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_4, 12, in[ 9], 4 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_4, 14, in[11], 4 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_4, 15, in[10], 4 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_4, 14, in[ 0], 4 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_4, 15, in[ 8], 4 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_4, 9, in[12], 4 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_4, 8, in[ 4], 4 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_4, 9, in[13], 4 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_4, 14, in[ 3], 4 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_4, 5, in[ 7], 4 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_4, 6, in[15], 4 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_4, 8, in[14], 4 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_4, 6, in[ 5], 4 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_4, 5, in[ 6], 4 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_4, 12, in[ 2], 4 );
|
||||
|
||||
ROUND1_16W( B, C, D, E, A, F16W_5, 9, in[ 4], 5 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_5, 15, in[ 0], 5 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_5, 5, in[ 5], 5 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_5, 11, in[ 9], 5 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_5, 6, in[ 7], 5 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_5, 8, in[12], 5 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_5, 13, in[ 2], 5 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_5, 12, in[10], 5 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_5, 5, in[14], 5 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_5, 12, in[ 1], 5 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_5, 13, in[ 3], 5 );
|
||||
ROUND1_16W( A, B, C, D, E, F16W_5, 14, in[ 8], 5 );
|
||||
ROUND1_16W( E, A, B, C, D, F16W_5, 11, in[11], 5 );
|
||||
ROUND1_16W( D, E, A, B, C, F16W_5, 8, in[ 6], 5 );
|
||||
ROUND1_16W( C, D, E, A, B, F16W_5, 5, in[15], 5 );
|
||||
ROUND1_16W( B, C, D, E, A, F16W_5, 6, in[13], 5 );
|
||||
|
||||
ROUND2_16W( A, B, C, D, E, F16W_5, 8, in[ 5], 1 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_5, 9, in[14], 1 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_5, 9, in[ 7], 1 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_5, 11, in[ 0], 1 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_5, 13, in[ 9], 1 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_5, 15, in[ 2], 1 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_5, 15, in[11], 1 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_5, 5, in[ 4], 1 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_5, 7, in[13], 1 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_5, 7, in[ 6], 1 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_5, 8, in[15], 1 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_5, 11, in[ 8], 1 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_5, 14, in[ 1], 1 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_5, 14, in[10], 1 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_5, 12, in[ 3], 1 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_5, 6, in[12], 1 );
|
||||
|
||||
ROUND2_16W( E, A, B, C, D, F16W_4, 9, in[ 6], 2 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_4, 13, in[11], 2 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_4, 15, in[ 3], 2 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_4, 7, in[ 7], 2 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_4, 12, in[ 0], 2 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_4, 8, in[13], 2 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_4, 9, in[ 5], 2 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_4, 11, in[10], 2 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_4, 7, in[14], 2 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_4, 7, in[15], 2 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_4, 12, in[ 8], 2 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_4, 7, in[12], 2 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_4, 6, in[ 4], 2 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_4, 15, in[ 9], 2 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_4, 13, in[ 1], 2 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_4, 11, in[ 2], 2 );
|
||||
|
||||
ROUND2_16W( D, E, A, B, C, F16W_3, 9, in[15], 3 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_3, 7, in[ 5], 3 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_3, 15, in[ 1], 3 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_3, 11, in[ 3], 3 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_3, 8, in[ 7], 3 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_3, 6, in[14], 3 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_3, 6, in[ 6], 3 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_3, 14, in[ 9], 3 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_3, 12, in[11], 3 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_3, 13, in[ 8], 3 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_3, 5, in[12], 3 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_3, 14, in[ 2], 3 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_3, 13, in[10], 3 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_3, 13, in[ 0], 3 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_3, 7, in[ 4], 3 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_3, 5, in[13], 3 );
|
||||
|
||||
ROUND2_16W( C, D, E, A, B, F16W_2, 15, in[ 8], 4 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_2, 5, in[ 6], 4 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_2, 8, in[ 4], 4 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_2, 11, in[ 1], 4 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_2, 14, in[ 3], 4 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_2, 14, in[11], 4 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_2, 6, in[15], 4 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_2, 14, in[ 0], 4 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_2, 6, in[ 5], 4 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_2, 9, in[12], 4 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_2, 12, in[ 2], 4 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_2, 9, in[13], 4 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_2, 12, in[ 9], 4 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_2, 5, in[ 7], 4 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_2, 15, in[10], 4 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_2, 8, in[14], 4 );
|
||||
|
||||
ROUND2_16W( B, C, D, E, A, F16W_1, 8, in[12], 5 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_1, 5, in[15], 5 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_1, 12, in[10], 5 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_1, 9, in[ 4], 5 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_1, 12, in[ 1], 5 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_1, 5, in[ 5], 5 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_1, 14, in[ 8], 5 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_1, 6, in[ 7], 5 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_1, 8, in[ 6], 5 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_1, 13, in[ 2], 5 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_1, 6, in[13], 5 );
|
||||
ROUND2_16W( A, B, C, D, E, F16W_1, 5, in[14], 5 );
|
||||
ROUND2_16W( E, A, B, C, D, F16W_1, 15, in[ 0], 5 );
|
||||
ROUND2_16W( D, E, A, B, C, F16W_1, 13, in[ 3], 5 );
|
||||
ROUND2_16W( C, D, E, A, B, F16W_1, 11, in[ 9], 5 );
|
||||
ROUND2_16W( B, C, D, E, A, F16W_1, 11, in[11], 5 );
|
||||
|
||||
tmp = _mm512_add_epi32( _mm512_add_epi32( h[1], C1 ), D2 );
|
||||
h[1] = _mm512_add_epi32( _mm512_add_epi32( h[2], D1 ), E2 );
|
||||
h[2] = _mm512_add_epi32( _mm512_add_epi32( h[3], E1 ), A2 );
|
||||
h[3] = _mm512_add_epi32( _mm512_add_epi32( h[4], A1 ), B2 );
|
||||
h[4] = _mm512_add_epi32( _mm512_add_epi32( h[0], B1 ), C2 );
|
||||
h[0] = tmp;
|
||||
}
|
||||
|
||||
void ripemd160_16way_init( ripemd160_16way_context *sc )
|
||||
{
|
||||
sc->val[0] = m512_const1_64( 0x6745230167452301 );
|
||||
sc->val[1] = m512_const1_64( 0xEFCDAB89EFCDAB89 );
|
||||
sc->val[2] = m512_const1_64( 0x98BADCFE98BADCFE );
|
||||
sc->val[3] = m512_const1_64( 0x1032547610325476 );
|
||||
sc->val[4] = m512_const1_64( 0xC3D2E1F0C3D2E1F0 );
|
||||
sc->count_high = sc->count_low = 0;
|
||||
}
|
||||
|
||||
void ripemd160_16way( ripemd160_16way_context *sc, const void *data,
|
||||
size_t len )
|
||||
{
|
||||
__m512i *vdata = (__m512i*)data;
|
||||
size_t ptr;
|
||||
const int block_size = 64;
|
||||
|
||||
ptr = (unsigned)sc->count_low & (block_size - 1U);
|
||||
while ( len > 0 )
|
||||
{
|
||||
size_t clen;
|
||||
uint32_t clow, clow2;
|
||||
|
||||
clen = block_size - ptr;
|
||||
if ( clen > len )
|
||||
clen = len;
|
||||
memcpy_512( sc->buf + (ptr>>2), vdata, clen>>2 );
|
||||
vdata = vdata + (clen>>2);
|
||||
ptr += clen;
|
||||
len -= clen;
|
||||
if ( ptr == block_size )
|
||||
{
|
||||
ripemd160_16way_round( sc );
|
||||
ptr = 0;
|
||||
}
|
||||
clow = sc->count_low;
|
||||
clow2 = clow + clen;
|
||||
sc->count_low = clow2;
|
||||
if ( clow2 < clow )
|
||||
sc->count_high++;
|
||||
}
|
||||
}
|
||||
|
||||
void ripemd160_16way_close( ripemd160_16way_context *sc, void *dst )
|
||||
{
|
||||
unsigned ptr, u;
|
||||
uint32_t low, high;
|
||||
const int block_size = 64;
|
||||
const int pad = block_size - 8;
|
||||
|
||||
ptr = (unsigned)sc->count_low & ( block_size - 1U);
|
||||
sc->buf[ ptr>>2 ] = m512_const1_32( 0x80 );
|
||||
ptr += 4;
|
||||
|
||||
if ( ptr > pad )
|
||||
{
|
||||
memset_zero_512( sc->buf + (ptr>>2), (block_size - ptr) >> 2 );
|
||||
ripemd160_16way_round( sc );
|
||||
memset_zero_512( sc->buf, pad>>2 );
|
||||
}
|
||||
else
|
||||
memset_zero_512( sc->buf + (ptr>>2), (pad - ptr) >> 2 );
|
||||
|
||||
low = sc->count_low;
|
||||
high = (sc->count_high << 3) | (low >> 29);
|
||||
low = low << 3;
|
||||
sc->buf[ pad>>2 ] = _mm512_set1_epi32( low );
|
||||
sc->buf[ (pad>>2) + 1 ] = _mm512_set1_epi32( high );
|
||||
ripemd160_16way_round( sc );
|
||||
for (u = 0; u < 5; u ++)
|
||||
casti_m512i( dst, u ) = sc->val[u];
|
||||
}
|
||||
|
||||
#endif // AVX512
|
||||
|
Reference in New Issue
Block a user