int main()
{
__m512d t0,t1;
double d1[8] __attribute__ ((aligned(64)));
double d2[8] __attribute__ ((aligned(64)));
double d3[8] __attribute__ ((aligned(64)));
for(int i=0; i<8; i++)
{
d1[i]= i*1.0;
d2[i]= 0.0;
d3[i] = d1[i];
}
//printf("testing intialization of registers\n");
//_mm512_store_pd(d1,t0);
//printf("d1=t0: %f %f %f %f %f %f %f %f \n",d1[0],d1[1],d1[2],d1[3],d1[4],d1[5],d1[6],d1[7]);
//_mm512_store_pd(d1,t1);
//printf("d1=t1: %f %f %f %f %f %f %f %f \n",d1[0],d1[1],d1[2],d1[3],d1[4],d1[5],d1[6],d1[7]);
t0 = _mm512_load_pd(d1);
printf("permute backward\n");
t1 = (__m512d) _mm512_permute4f128_epi32 ( (__m512i) t0, 0b00111001);
_mm512_store_pd(d2,t1);
printf("d1: %f %f %f %f %f %f %f %f \n",d1[0],d1[1],d1[2],d1[3],d1[4],d1[5],d1[6],d1[7]);
printf("d2: %f %f %f %f %f %f %f %f \n",d2[0],d2[1],d2[2],d2[3],d2[4],d2[5],d2[6],d2[7]);
printf("permute forward\n");
t1 = (__m512d) _mm512_permute4f128_epi32 ( (__m512i) t0, 0b10010011);
_mm512_store_pd(d2,t1);
printf("d1: %f %f %f %f %f %f %f %f \n",d1[0],d1[1],d1[2],d1[3],d1[4],d1[5],d1[6],d1[7]);
printf("d2: %f %f %f %f %f %f %f %f \n",d2[0],d2[1],d2[2],d2[3],d2[4],d2[5],d2[6],d2[7]);
int __attribute__((aligned(64))) order[16]={0,1,0,1,4,5,6,7,8,9,10,11,12,13,14,15};
__m512i morder = _mm512_load_epi32(order);
printf("permuting doubles\n");
t1 = (__m512d) _mm512_permutevar_epi32 (morder, (__m512i) t0);
_mm512_store_pd(d2,t1);
printf("d1: %f %f %f %f %f %f %f %f \n",d1[0],d1[1],d1[2],d1[3],d1[4],d1[5],d1[6],d1[7]);
printf("d2: %f %f %f %f %f %f %f %f \n",d2[0],d2[1],d2[2],d2[3],d2[4],d2[5],d2[6],d2[7]);
return 0;
}
void mad12(int num, DT* data, int repeat, DT vv1, DT vv2)
{
int gid, j;
#ifndef MIC
#pragma omp parallel for
for (gid = 0; gid<num; gid++)
{
register DT s = (DT)(0.999f);
register DT v1 = vv1;
register DT v2 = vv2;
for (j=0; j<repeat; ++j)
{
MADD1_MOP2
}
data[gid] = s;
}
#else
#pragma omp parallel for
for (gid = 0; gid<num; gid=gid+STEP)
{
__m512d s = _mm512_set1_pd(0.999f);
__m512d v1 = _mm512_set1_pd(vv1);
__m512d v2 = _mm512_set1_pd(vv2);
for (j=0; j<repeat; ++j)
{
MADD1_MOP2
}
_mm512_store_pd(&(data[gid]), s);
}
#endif
return ;
}
LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE
void stream_vector_set( const double i_scalar,
double* io_c,
const int i_length) {
int l_n = 0;
int l_trip_prolog = 0;
int l_trip_stream = 0;
/* init the trip counts */
stream_init( i_length, (size_t)io_c, &l_trip_prolog, &l_trip_stream );
/* run the prologue */
for ( ; l_n < l_trip_prolog; l_n++ ) {
io_c[l_n] = i_scalar;
}
/* run the bulk, hopefully using streaming stores */
#if defined(__SSE3__) && defined(__AVX__) && !defined(__AVX512F__)
{
/* we need manual unrolling as the compiler otherwise generates
too many dependencies */
const __m256d vec_scalar = _mm256_broadcast_sd(&i_scalar);
for ( ; l_n < l_trip_stream; l_n+=8 ) {
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n]), vec_scalar );
_mm256_store_pd( &(io_c[l_n+4]), vec_scalar );
#else
_mm256_stream_pd( &(io_c[l_n]), vec_scalar );
_mm256_stream_pd( &(io_c[l_n+4]), vec_scalar );
#endif
}
}
#elif defined(__SSE3__) && defined(__AVX__) && defined(__AVX512F__)
{
const __m512d vec_scalar = _mm512_broadcastsd_pd(_mm_load_sd(&i_scalar));
for ( ; l_n < l_trip_stream; l_n+=8 ) {
#ifdef DISABLE_NONTEMPORAL_STORES
_mm512_store_pd( &(io_c[l_n]), vec_scalar );
#else
_mm512_stream_pd( &(io_c[l_n]), vec_scalar );
#endif
}
}
#else
for ( ; l_n < l_trip_stream; l_n++ ) {
io_c[l_n] = i_scalar;
}
#endif
/* run the epilogue */
for ( ; l_n < i_length; l_n++ ) {
io_c[l_n] = i_scalar;
}
}
void extern
avx512f_test (void)
{
x1 = _mm512_mask_mov_pd (x1, m, x2);
x1 = _mm512_maskz_mov_pd (m, x2);
x1 = _mm512_load_pd (p);
x1 = _mm512_mask_load_pd (x1, m, p);
x1 = _mm512_maskz_load_pd (m, p);
_mm512_store_pd (p, x1);
_mm512_mask_store_pd (p, m, x1);
}
__inline void mic_broadcast16x64(const double* inv, double* outv)
{
__mmask8 k1 = _mm512_int2mask(0x0F);
__mmask8 k2 = _mm512_int2mask(0xF0);
for(int l = 0; l < 16; l += 2)
{
__m512d t = _mm512_setzero_pd();
t = _mm512_mask_extload_pd(t, k1, &inv[(l%4)*4 + l/4], _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, _MM_HINT_NONE);
t = _mm512_mask_extload_pd(t, k2, &inv[((l+1)%4)*4 + (l+1)/4], _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, _MM_HINT_NONE);
_mm512_store_pd(&outv[l*4], t);
}
}
void ks_rank_k_int_d16x14(
int k,
double *a,
double *b,
double *c,
int ldc,
aux_t *aux
)
{
int i;
double neg2 = -2.0;
v8df_t c007_0, c007_1, c007_2, c007_3, c007_4;
v8df_t c007_5, c007_6, c007_7, c007_8, c007_9;
v8df_t c007_10, c007_11, c007_12, c007_13;
v8df_t c815_0, c815_1, c815_2, c815_3, c815_4;
v8df_t c815_5, c815_6, c815_7, c815_8, c815_9;
v8df_t c815_10, c815_11, c815_12, c815_13;
v8df_t a007, a815, b_tmp;
int k_iter = k;
// TODO: need to clean the c buffer.
for ( i = 0; i < k_iter; ++ i ) {
a007.v = _mm512_load_pd( a );
a815.v = _mm512_load_pd( a + 8 );
//printf( "%lf, %lf, %lf, %lf, %lf, %lf, %lf, %lf\n",
// a007.d[ 0 ], a007.d[ 1 ], a007.d[ 2 ], a007.d[ 3 ],
// a007.d[ 4 ], a007.d[ 5 ], a007.d[ 6 ], a007.d[ 7 ] );
b_tmp.v = _mm512_extload_pd( b, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
//printf( "b[ 0 ] = %lf\n", b[ 0 ] );
//printf( "%lf, %lf, %lf, %lf, %lf, %lf, %lf, %lf\n",
// b_tmp.d[ 0 ], b_tmp.d[ 1 ], b_tmp.d[ 2 ], b_tmp.d[ 3 ],
// b_tmp.d[ 4 ], b_tmp.d[ 5 ], b_tmp.d[ 6 ], b_tmp.d[ 7 ] );
c007_0.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_0.v );
c815_0.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_0.v );
b_tmp.v = _mm512_extload_pd( b + 1, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_1.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_1.v );
c815_1.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_1.v );
b_tmp.v = _mm512_extload_pd( b + 2, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_2.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_2.v );
c815_2.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_2.v );
b_tmp.v = _mm512_extload_pd( b + 3, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_3.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_3.v );
c815_3.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_3.v );
b_tmp.v = _mm512_extload_pd( b + 4, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_4.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_4.v );
c815_4.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_4.v );
b_tmp.v = _mm512_extload_pd( b + 5, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_5.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_5.v );
c815_5.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_5.v );
b_tmp.v = _mm512_extload_pd( b + 6, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_6.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_6.v );
c815_6.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_6.v );
b_tmp.v = _mm512_extload_pd( b + 7, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_7.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_7.v );
c815_7.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_7.v );
b_tmp.v = _mm512_extload_pd( b + 8, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_8.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_8.v );
c815_8.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_8.v );
b_tmp.v = _mm512_extload_pd( b + 9, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_9.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_9.v );
c815_9.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_9.v );
b_tmp.v = _mm512_extload_pd( b + 10, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_10.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_10.v );
c815_10.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_10.v );
b_tmp.v = _mm512_extload_pd( b + 11, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_11.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_11.v );
c815_11.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_11.v );
b_tmp.v = _mm512_extload_pd( b + 12, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_12.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_12.v );
c815_12.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_12.v );
b_tmp.v = _mm512_extload_pd( b + 13, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, 0 );
c007_13.v = _mm512_fmadd_pd( a007.v, b_tmp.v, c007_13.v );
c815_13.v = _mm512_fmadd_pd( a815.v, b_tmp.v, c815_13.v );
a += 16;
b += 16;
}
// simulate kernel summation
c007_0.v = _mm512_add_pd( c007_0.v, c007_1.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_1.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_2.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_2.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_3.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_3.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_4.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_4.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_5.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_5.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_6.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_6.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_7.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_7.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_8.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_8.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_9.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_9.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_10.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_10.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_11.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_11.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_12.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_12.v );
c007_0.v = _mm512_add_pd( c007_0.v, c007_13.v );
c815_0.v = _mm512_add_pd( c815_0.v, c815_13.v );
// if ( aux->pc != 0 ) {
//
// // packed
// tmpc03_0.v = _mm256_load_pd( (double*)( c ) );
// tmpc47_0.v = _mm256_load_pd( (double*)( c + 4 ) );
//
// tmpc03_1.v = _mm256_load_pd( (double*)( c + 8 ) );
// tmpc47_1.v = _mm256_load_pd( (double*)( c + 12 ) );
//
// tmpc03_2.v = _mm256_load_pd( (double*)( c + 16 ) );
// tmpc47_2.v = _mm256_load_pd( (double*)( c + 20 ) );
//
// tmpc03_3.v = _mm256_load_pd( (double*)( c + 24 ) );
// tmpc47_3.v = _mm256_load_pd( (double*)( c + 28 ) );
//
//
// c03_0.v = _mm256_add_pd( tmpc03_0.v, c03_0.v );
// c47_0.v = _mm256_add_pd( tmpc47_0.v, c47_0.v );
//
// c03_1.v = _mm256_add_pd( tmpc03_1.v, c03_1.v );
// c47_1.v = _mm256_add_pd( tmpc47_1.v, c47_1.v );
//
// c03_2.v = _mm256_add_pd( tmpc03_2.v, c03_2.v );
// c47_2.v = _mm256_add_pd( tmpc47_2.v, c47_2.v );
//
// c03_3.v = _mm256_add_pd( tmpc03_3.v, c03_3.v );
// c47_3.v = _mm256_add_pd( tmpc47_3.v, c47_3.v );
// }
//
//
// packed
_mm512_store_pd( c , c007_0.v );
_mm512_store_pd( c + 8 , c815_0.v );
// _mm512_store_pd( c + 16, c007_1.v );
// _mm512_store_pd( c + 24, c815_1.v );
//
// _mm512_store_pd( c + 32, c007_2.v );
// _mm512_store_pd( c + 40, c815_2.v );
//
// _mm512_store_pd( c + 48, c007_3.v );
// _mm512_store_pd( c + 56, c815_3.v );
//
// _mm512_store_pd( c + 64, c007_4.v );
// _mm512_store_pd( c + 72, c815_4.v );
//
// _mm512_store_pd( c + 80, c007_5.v );
// _mm512_store_pd( c + 88, c815_5.v );
//
// _mm512_store_pd( c + 96, c007_6.v );
// _mm512_store_pd( c + 104, c815_6.v );
//
// _mm512_store_pd( c + 112, c007_7.v );
// _mm512_store_pd( c + 120, c815_7.v );
//
// _mm512_store_pd( c + 128, c007_8.v );
// _mm512_store_pd( c + 136, c815_8.v );
//
// _mm512_store_pd( c + 144, c007_9.v );
// _mm512_store_pd( c + 152, c815_9.v );
//
// _mm512_store_pd( c + 160, c007_10.v );
// _mm512_store_pd( c + 168, c815_10.v );
//
// _mm512_store_pd( c + 176, c007_11.v );
// _mm512_store_pd( c + 184, c815_11.v );
//
// _mm512_store_pd( c + 192, c007_12.v );
// _mm512_store_pd( c + 200, c815_12.v );
//
// _mm512_store_pd( c + 208, c007_13.v );
// _mm512_store_pd( c + 216, c815_13.v );
//printf( "ldc = %d\n", ldc );
// printf( "%lf, %lf, %lf, %lf, %lf, %lf, %lf, %lf\n",
// c007_0.d[ 0 ], c007_0.d[ 1 ], c007_0.d[ 2 ], c007_0.d[ 3 ],
// c007_0.d[ 4 ], c007_0.d[ 5 ], c007_0.d[ 6 ], c007_0.d[ 7 ] );
//
//printf( "%lf, %lf, %lf, %lf\n", c[1], c[ ldc + 1], c[ ldc * 2 + 1], c[ ldc * 3 + 1] );
//printf( "%lf, %lf, %lf, %lf\n", c[2], c[ ldc + 2], c[ ldc * 2 + 2], c[ ldc * 3 + 2] );
//printf( "%lf, %lf, %lf, %lf\n", c[3], c[ ldc + 3], c[ ldc * 2 + 3], c[ ldc * 3 + 3] );
//printf( "%lf, %lf, %lf, %lf\n", c[4], c[ ldc + 4], c[ ldc * 2 + 4], c[ ldc * 3 + 4] );
//printf( "%lf, %lf, %lf, %lf\n", c[5], c[ ldc + 5], c[ ldc * 2 + 5], c[ ldc * 3 + 5] );
//printf( "%lf, %lf, %lf, %lf\n", c[6], c[ ldc + 6], c[ ldc * 2 + 6], c[ ldc * 3 + 6] );
//printf( "%lf, %lf, %lf, %lf\n", c[7], c[ ldc + 7], c[ ldc * 2 + 7], c[ ldc * 3 + 7] );
}
void newviewGTRGAMMA_MIC(int tipCase,
double *x1, double *x2, double *x3, double *extEV, double *tipVector,
int *ex3, unsigned char *tipX1, unsigned char *tipX2,
int n, double *left, double *right, int *wgt, int *scalerIncrement, const pllBoolean fastScaling)
{
__m512d minlikelihood_MIC = _mm512_set1_pd(PLL_MINLIKELIHOOD);
__m512d twotothe256_MIC = _mm512_set1_pd(PLL_TWOTOTHE256);
__m512i absMask_MIC = _mm512_set1_epi64(0x7fffffffffffffffULL);
int addScale = 0;
double aEV[64] __attribute__((align(PLL_BYTE_ALIGNMENT)));
#pragma ivdep
for (int l = 0; l < 64; ++l)
{
aEV[l] = extEV[(l / 16) * 4 + (l % 4)];
}
switch(tipCase)
{
case PLL_TIP_TIP:
{
/* multiply all possible tip state vectors with the respective P-matrices
*/
double umpX1[256] __attribute__((align(PLL_BYTE_ALIGNMENT)));
double umpX2[256] __attribute__((align(PLL_BYTE_ALIGNMENT)));
for(int k = 0; k < 256; ++k)
{
umpX1[k] = 0.0;
umpX2[k] = 0.0;
}
for(int i = 0; i < maxStateValue; ++i)
{
for(int l = 0; l < states; ++l)
{
#pragma ivdep
for(int k = 0; k < span; ++k)
{
umpX1[16 * i + k] += tipVector[i * 4 + l] * left[k * 4 + l];
umpX2[16 * i + k] += tipVector[i * 4 + l] * right[k * 4 + l];
}
}
}
double auX[64] __attribute__((align(64)));
for(int i = 0; i < n; ++i)
{
_mm_prefetch((const char*) (const char*) &x3[span*(i+8)], _MM_HINT_ET1);
_mm_prefetch((const char*) &x3[span*(i+8) + 8], _MM_HINT_ET1);
_mm_prefetch((const char*) &x3[span*(i+1)], _MM_HINT_ET0);
_mm_prefetch((const char*) &x3[span*(i+1) + 8], _MM_HINT_ET0);
const double *uX1 = &umpX1[16 * tipX1[i]];
const double *uX2 = &umpX2[16 * tipX2[i]];
double uX[16] __attribute__((align(PLL_BYTE_ALIGNMENT)));
double* v = &x3[i * 16];
#pragma ivdep
#pragma vector aligned
for(int l = 0; l < 16; ++l)
{
uX[l] = uX1[l] * uX2[l];
v[l] = 0.;
}
mic_broadcast16x64(uX, auX);
for (int j = 0; j < 4; ++j)
{
#pragma ivdep
#pragma vector aligned
#pragma vector nontemporal
for(int k = 0; k < 16; ++k)
{
v[k] += auX[j*16 + k] * aEV[j*16 + k];
}
}
// init scaling counter for the site
if (!fastScaling)
ex3[i] = 0;
} // sites loop
}
break;
case PLL_TIP_INNER:
{
/* we do analogous pre-computations as above, with the only difference that we now do them
only for one tip vector */
double umpX1[256] __attribute__((align(PLL_BYTE_ALIGNMENT)));
/* precompute P and left tip vector product */
for(int k = 0; k < 256; ++k)
{
umpX1[k] = 0.0;
}
for(int i = 0; i < 16; ++i)
{
for(int l = 0; l < 4; ++l)
{
#pragma ivdep
for(int k = 0; k < 16; ++k)
{
umpX1[16 * i + k] += tipVector[i * 4 + l] * left[k * 4 + l];
}
}
}
// re-arrange right matrix for better memory layout
double aRight[64] __attribute__((align(PLL_BYTE_ALIGNMENT)));
for(int j = 0; j < 4; j++)
{
for(int l = 0; l < 16; l++)
{
aRight[j*16 + l] = right[l*4 + j];
}
}
for (int i = 0; i < n; i++)
{
_mm_prefetch((const char*) &x2[span*(i+16)], _MM_HINT_T1);
_mm_prefetch((const char*) &x2[span*(i+16) + 8], _MM_HINT_T1);
_mm_prefetch((const char*) &x3[span*(i+16)], _MM_HINT_ET1);
_mm_prefetch((const char*) &x3[span*(i+16) + 8], _MM_HINT_ET1);
_mm_prefetch((const char*) &x2[span*(i+1)], _MM_HINT_T0);
_mm_prefetch((const char*) &x2[span*(i+1) + 8], _MM_HINT_T0);
_mm_prefetch((const char*) &x3[span*(i+1)], _MM_HINT_ET0);
_mm_prefetch((const char*) &x3[span*(i+1) + 8], _MM_HINT_ET0);
/* access pre-computed value based on the raw sequence data tipX1 that is used as an index */
double* uX1 = &umpX1[span * tipX1[i]];
double uX2[16] __attribute__((align(PLL_BYTE_ALIGNMENT)));
double uX[16] __attribute__((align(PLL_BYTE_ALIGNMENT)));
#pragma vector aligned
for(int l = 0; l < 16; ++l)
{
uX2[l] = 0.;
}
double aV2[64] __attribute__((align(PLL_BYTE_ALIGNMENT)));
const double* v2 = &(x2[16 * i]);
mic_broadcast16x64(v2, aV2);
for(int j = 0; j < 4; j++)
{
#pragma ivdep
#pragma vector aligned
for(int l = 0; l < 16; l++)
{
uX2[l] += aV2[j*16 + l] * aRight[j*16 + l];
}
}
double* v3 = &(x3[span * i]);
#pragma ivdep
#pragma vector aligned
for(int l = 0; l < 16; ++l)
{
uX[l] = uX1[l] * uX2[l];
v3[l] = 0.;
}
double auX[64] __attribute__((align(PLL_BYTE_ALIGNMENT)));
mic_broadcast16x64(uX, auX);
for (int j = 0; j < 4; ++j)
{
#pragma ivdep
#pragma vector aligned
for(int k = 0; k < 16; ++k)
{
v3[k] += auX[j*16 + k] * aEV[j*16 + k];
}
}
__m512d t1 = _mm512_load_pd(&v3[0]);
t1 = _mm512_castsi512_pd(_mm512_and_epi64(_mm512_castpd_si512(t1), absMask_MIC));
double vmax1 = _mm512_reduce_gmax_pd(t1);
__m512d t2 = _mm512_load_pd(&v3[8]);
t2 = _mm512_castsi512_pd(_mm512_and_epi64(_mm512_castpd_si512(t2), absMask_MIC));
double vmax2 = _mm512_reduce_gmax_pd(t2);
if(vmax1 < PLL_MINLIKELIHOOD && vmax2 < PLL_MINLIKELIHOOD)
{
t1 = _mm512_mul_pd(t1, twotothe256_MIC);
_mm512_store_pd(&v3[0], t1);
t2 = _mm512_mul_pd(t2, twotothe256_MIC);
_mm512_store_pd(&v3[8], t2);
if(!fastScaling)
ex3[i] += 1;
else
addScale += wgt[i];
}
} // site loop
}
break;
case PLL_INNER_INNER:
{
/* same as above, without pre-computations */
// re-arrange right matrix for better memory layout
double aLeft[64] __attribute__((align(PLL_BYTE_ALIGNMENT)));
double aRight[64] __attribute__((align(PLL_BYTE_ALIGNMENT)));
for(int j = 0; j < 4; j++)
{
for(int l = 0; l < 16; l++)
{
aLeft[j*16 + l] = left[l*4 + j];
aRight[j*16 + l] = right[l*4 + j];
}
}
for (int i = 0; i < n; i++)
{
_mm_prefetch((const char*) &x1[span*(i+8)], _MM_HINT_T1);
_mm_prefetch((const char*) &x1[span*(i+8) + 8], _MM_HINT_T1);
_mm_prefetch((const char*) &x2[span*(i+8)], _MM_HINT_T1);
_mm_prefetch((const char*) &x2[span*(i+8) + 8], _MM_HINT_T1);
_mm_prefetch((const char*) &x3[span*(i+8)], _MM_HINT_ET1);
_mm_prefetch((const char*) &x3[span*(i+8) + 8], _MM_HINT_ET1);
_mm_prefetch((const char*) &x1[span*(i+1)], _MM_HINT_T0);
_mm_prefetch((const char*) &x1[span*(i+1) + 8], _MM_HINT_T0);
_mm_prefetch((const char*) &x2[span*(i+1)], _MM_HINT_T0);
_mm_prefetch((const char*) &x2[span*(i+1) + 8], _MM_HINT_T0);
_mm_prefetch((const char*) &x3[span*(i+1)], _MM_HINT_ET0);
_mm_prefetch((const char*) &x3[span*(i+1) + 8], _MM_HINT_ET0);
double uX1[16] __attribute__((align(64)));
double uX2[16] __attribute__((align(64)));
double uX[16] __attribute__((align(64)));
for(int l = 0; l < 16; l++)
{
uX1[l] = 0.;
uX2[l] = 0.;
}
double aV1[64] __attribute__((align(64)));
double aV2[64] __attribute__((align(64)));
const double* v1 = &(x1[span * i]);
const double* v2 = &(x2[span * i]);
mic_broadcast16x64(v1, aV1);
mic_broadcast16x64(v2, aV2);
for(int j = 0; j < 4; j++)
{
#pragma ivdep
#pragma vector aligned
for(int l = 0; l < 16; l++)
{
uX1[l] += aV1[j*16 + l] * aLeft[j*16 + l];
uX2[l] += aV2[j*16 + l] * aRight[j*16 + l];
}
}
double* v3 = &(x3[span * i]);
#pragma ivdep
#pragma vector aligned
for(int l = 0; l < 16; ++l)
{
uX[l] = uX1[l] * uX2[l];
v3[l] = 0.;
}
double auX[64] __attribute__((align(64)));
mic_broadcast16x64(uX, auX);
for(int j = 0; j < 4; ++j)
{
#pragma ivdep
#pragma vector aligned
for(int k = 0; k < 16; ++k)
{
v3[k] += auX[j*16 + k] * aEV[j*16 + k];
}
}
__m512d t1 = _mm512_load_pd(&v3[0]);
t1 = _mm512_castsi512_pd(_mm512_and_epi64(_mm512_castpd_si512(t1), absMask_MIC));
double vmax1 = _mm512_reduce_gmax_pd(t1);
__m512d t2 = _mm512_load_pd(&v3[8]);
t2 = _mm512_castsi512_pd(_mm512_and_epi64(_mm512_castpd_si512(t2), absMask_MIC));
double vmax2 = _mm512_reduce_gmax_pd(t2);
if(vmax1 < PLL_MINLIKELIHOOD && vmax2 < PLL_MINLIKELIHOOD)
{
t1 = _mm512_mul_pd(t1, twotothe256_MIC);
_mm512_store_pd(&v3[0], t1);
t2 = _mm512_mul_pd(t2, twotothe256_MIC);
_mm512_store_pd(&v3[8], t2);
if(!fastScaling)
ex3[i] += 1;
else
addScale += wgt[i];
}
}
} break;
default:
// assert(0);
break;
}
/* as above, increment the global counter that counts scaling multiplications by the scaling multiplications
carried out for computing the likelihood array at node p */
if (fastScaling)
{
*scalerIncrement = addScale;
}
}
inline void store(double *r, const F64vec8 v)
{
_mm512_store_pd(r, v);
}
LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE
void stream_update_helmholtz_no_h2( const double* i_g1,
const double* i_g2,
const double* i_g3,
const double* i_tm1,
const double* i_tm2,
const double* i_tm3,
double* io_c,
const double i_h1,
const int i_length) {
int l_n = 0;
int l_trip_prolog = 0;
int l_trip_stream = 0;
/* init the trip counts */
stream_init( i_length, (size_t)io_c, &l_trip_prolog, &l_trip_stream );
/* run the prologue */
for ( ; l_n < l_trip_prolog; l_n++ ) {
io_c[l_n] = i_h1*(i_g1[l_n]*i_tm1[l_n] + i_g2[l_n]*i_tm2[l_n] + i_g3[l_n]*i_tm3[l_n]);
}
/* run the bulk, hopefully using streaming stores */
#if defined(__SSE3__) && defined(__AVX__) && !defined(__AVX512F__)
{
const __m256d vec_h1 = _mm256_broadcast_sd(&i_h1);
/* we need manual unrolling as the compiler otherwise generates
too many dependencies */
for ( ; l_n < l_trip_stream; l_n+=8 ) {
__m256d vec_g1_1, vec_g2_1, vec_g3_1, vec_tm1_1, vec_tm2_1, vec_tm3_1;
__m256d vec_g1_2, vec_g2_2, vec_g3_2, vec_tm1_2, vec_tm2_2, vec_tm3_2;
vec_g1_1 = _mm256_loadu_pd(&(i_g1[l_n]));
vec_tm1_1 = _mm256_loadu_pd(&(i_tm1[l_n]));
vec_g1_2 = _mm256_loadu_pd(&(i_g1[l_n+4]));
vec_tm1_2 = _mm256_loadu_pd(&(i_tm1[l_n+4]));
vec_g1_1 = _mm256_mul_pd(vec_g1_1, vec_tm1_1);
vec_g2_1 = _mm256_loadu_pd(&(i_g2[l_n]));
vec_g1_2 = _mm256_mul_pd(vec_g1_2, vec_tm1_2);
vec_g2_2 = _mm256_loadu_pd(&(i_g2[l_n+4]));
vec_tm2_1 = _mm256_loadu_pd(&(i_tm2[l_n]));
vec_g2_1 = _mm256_mul_pd(vec_g2_1, vec_tm2_1);
vec_tm2_2 = _mm256_loadu_pd(&(i_tm2[l_n+4]));
vec_g2_2 = _mm256_mul_pd(vec_g2_2, vec_tm2_2);
vec_g3_1 = _mm256_loadu_pd(&(i_g3[l_n]));
vec_tm3_1 = _mm256_loadu_pd(&(i_tm3[l_n]));
vec_g3_2 = _mm256_loadu_pd(&(i_g3[l_n+4]));
vec_tm3_2 = _mm256_loadu_pd(&(i_tm3[l_n+4]));
vec_g3_1 = _mm256_mul_pd(vec_g3_1, vec_tm3_1);
vec_g3_2 = _mm256_mul_pd(vec_g3_2, vec_tm3_2);
vec_g1_1 = _mm256_add_pd(vec_g1_1, vec_g2_1);
vec_g1_2 = _mm256_add_pd(vec_g1_2, vec_g2_2);
vec_g1_1 = _mm256_add_pd(vec_g1_1, vec_g3_1);
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n]), _mm256_mul_pd(vec_g1_1, vec_h1) );
#else
_mm256_stream_pd( &(io_c[l_n]), _mm256_mul_pd(vec_g1_1, vec_h1) );
#endif
vec_g1_2 = _mm256_add_pd(vec_g1_2, vec_g3_2);
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n+4]), _mm256_mul_pd(vec_g1_2, vec_h1) );
#else
_mm256_stream_pd( &(io_c[l_n+4]), _mm256_mul_pd(vec_g1_2, vec_h1) );
#endif
}
}
#elif defined(__SSE3__) && defined(__AVX__) && defined(__AVX512F__)
{
const __m512d vec_h1 = _mm512_broadcastsd_pd(_mm_load_sd(&i_h1));
for ( ; l_n < l_trip_stream; l_n+=8 ) {
__m512d vec_g1, vec_g2, vec_g3, vec_tm1, vec_tm2, vec_tm3;
vec_g1 = _mm512_loadu_pd(&(i_g1[l_n]));
vec_tm1 = _mm512_loadu_pd(&(i_tm1[l_n]));
vec_g1 = _mm512_mul_pd(vec_g1, vec_tm1);
vec_g2 = _mm512_loadu_pd(&(i_g2[l_n]));
vec_tm2 = _mm512_loadu_pd(&(i_tm2[l_n]));
vec_g2 = _mm512_mul_pd(vec_g2, vec_tm2);
vec_g3 = _mm512_loadu_pd(&(i_g3[l_n]));
vec_tm3 = _mm512_loadu_pd(&(i_tm3[l_n]));
vec_g3 = _mm512_mul_pd(vec_g3, vec_tm3);
vec_g1 = _mm512_add_pd(vec_g1, vec_g2);
vec_g1 = _mm512_add_pd(vec_g1, vec_g3);
#ifdef DISABLE_NONTEMPORAL_STORES
_mm512_store_pd( &(io_c[l_n]), _mm512_mul_pd(vec_g1, vec_h1) );
#else
_mm512_stream_pd( &(io_c[l_n]), _mm512_mul_pd(vec_g1, vec_h1) );
#endif
}
}
#else
for ( ; l_n < l_trip_stream; l_n++ ) {
io_c[l_n] = i_h1*(i_g1[l_n]*i_tm1[l_n] + i_g2[l_n]*i_tm2[l_n] + i_g3[l_n]*i_tm3[l_n]);
}
#endif
/* run the epilogue */
for ( ; l_n < i_length; l_n++ ) {
io_c[l_n] = i_h1*(i_g1[l_n]*i_tm1[l_n] + i_g2[l_n]*i_tm2[l_n] + i_g3[l_n]*i_tm3[l_n]);
}
}
LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE
void stream_vector_compscale( const double* i_a,
const double* i_b,
double* io_c,
const int i_length) {
int l_n = 0;
int l_trip_prolog = 0;
int l_trip_stream = 0;
/* init the trip counts */
stream_init( i_length, (size_t)io_c, &l_trip_prolog, &l_trip_stream );
/* run the prologue */
for ( ; l_n < l_trip_prolog; l_n++ ) {
io_c[l_n] = i_a[l_n]*i_b[l_n];
}
/* run the bulk, hopefully using streaming stores */
#if defined(__SSE3__) && defined(__AVX__) && !defined(__AVX512F__)
{
/* we need manual unrolling as the compiler otherwise generates
too many dependencies */
for ( ; l_n < l_trip_stream; l_n+=8 ) {
__m256d vec_a_1, vec_b_1;
__m256d vec_a_2, vec_b_2;
vec_a_1 = _mm256_loadu_pd(&(i_a[l_n]));
vec_a_2 = _mm256_loadu_pd(&(i_a[l_n+4]));
vec_b_1 = _mm256_loadu_pd(&(i_b[l_n]));
vec_b_2 = _mm256_loadu_pd(&(i_b[l_n+4]));
#ifdef DISABLE_NONTEMPORAL_STORES
_mm256_store_pd( &(io_c[l_n]), _mm256_mul_pd( vec_a_1, vec_b_1 ) );
_mm256_store_pd( &(io_c[l_n+4]), _mm256_mul_pd( vec_a_2, vec_b_2 ) );
#else
_mm256_stream_pd( &(io_c[l_n]), _mm256_mul_pd( vec_a_1, vec_b_1 ) );
_mm256_stream_pd( &(io_c[l_n+4]), _mm256_mul_pd( vec_a_2, vec_b_2 ) );
#endif
}
}
#elif defined(__SSE3__) && defined(__AVX__) && defined(__AVX512F__)
{
for ( ; l_n < l_trip_stream; l_n+=8 ) {
__m512d vec_a, vec_b;
vec_a = _mm512_loadu_pd(&(i_a[l_n]));
vec_b = _mm512_loadu_pd(&(i_b[l_n]));
#ifdef DISABLE_NONTEMPORAL_STORES
_mm512_store_pd( &(io_c[l_n]), _mm512_mul_pd( vec_a, vec_b ) );
#else
_mm512_stream_pd( &(io_c[l_n]), _mm512_mul_pd( vec_a, vec_b ) );
#endif
}
}
#else
for ( ; l_n < l_trip_stream; l_n++ ) {
io_c[l_n] = i_a[l_n]*i_b[l_n];
}
#endif
/* run the epilogue */
for ( ; l_n < i_length; l_n++ ) {
io_c[l_n] = i_a[l_n]*i_b[l_n];
}
}
inline
void store_aligned(double *data) const
{
SHORTVEC_ASSERT_ALIGNED(data, 64);
_mm512_store_pd(data, val);
}
