Skip to content

GitLab

elpa
elpa

There is a maintenance of MPCDF Gitlab on Thursday, April 22st 2020, 9:00 am CEST - Expect some service interruptions during this time

Commit 863d1b97 authored by Andreas Marek's avatar Andreas Marek
Browse files
Options

Start to unfiy complex single/double SSE block 1 kernel

parent dc90a3b8
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 953 additions and 1013 deletions
Makefile.am
View file @ 863d1b97
......@@ -595,6 +595,7 @@ EXTRA_DIST = \
src/elpa2/kernels/elpa2_kernels_real_avx-avx2_2hv_template.Xc \
src/elpa2/kernels/elpa2_kernels_real_avx-avx2_4hv_template.Xc \
src/elpa2/kernels/elpa2_kernels_real_avx-avx2_6hv_template.Xc \
src/elpa2/kernels/elpa2_kernels_complex_sse_1hv_template.Xc \
src/elpa2/redist_band.X90 \
src/elpa2/pack_unpack_cpu.X90 \
src/elpa2/pack_unpack_gpu.X90 \
......
src/elpa2/kernels/elpa2_kernels_complex_sse_1hv_double_precision.c
View file @ 863d1b97
......@@ -42,514 +42,14 @@
// any derivatives of ELPA under the same license that we chose for
// the original distribution, the GNU Lesser General Public License.
//
//
// --------------------------------------------------------------------------------------------------
//
// This file contains the compute intensive kernels for the Householder transformations.
// It should be compiled with the highest possible optimization level.
//
// On Intel Nehalem or Intel Westmere or AMD Magny Cours use -O3 -msse3
// On Intel Sandy Bridge use -O3 -mavx
//
// Copyright of the original code rests with the authors inside the ELPA
// consortium. The copyright of any additional modifications shall rest
// with their original authors, but shall adhere to the licensing terms
// distributed along with the original code in the file "COPYING".
//
// Author: Alexander Heinecke (alexander.heinecke@mytum.de)
// Adapted for building a shared-library by Andreas Marek, MPCDF (andreas.marek@mpcdf.mpg.de)
// --------------------------------------------------------------------------------------------------
// Author: Andreas Marek, MPCDF
#include "config-f90.h"
#include <complex.h>
#include <x86intrin.h>
#define __forceinline __attribute__((always_inline))
#ifdef HAVE_SSE_INTRINSICS
#undef __AVX__
#endif
//Forward declaration
static __forceinline void hh_trafo_complex_kernel_6_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq);
static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq);
static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq);
/*
!f>#ifdef WITH_COMPLEX_SSE_BLOCK1_KERNEL
!f> interface
!f> subroutine single_hh_trafo_complex_sse_1hv_double(q, hh, pnb, pnq, pldq) &
!f> bind(C, name="single_hh_trafo_complex_sse_1hv_double")
!f> use, intrinsic :: iso_c_binding
!f> integer(kind=c_int) :: pnb, pnq, pldq
!f> ! complex(kind=c_double_complex) :: q(*)
!f> type(c_ptr), value :: q
!f> complex(kind=c_double_complex) :: hh(pnb,2)
!f> end subroutine
!f> end interface
!f>#endif
*/
void single_hh_trafo_complex_sse_1hv_double(double complex* q, double complex* hh, int* pnb, int* pnq, int* pldq)
{
int i;
int nb = *pnb;
int nq = *pldq;
int ldq = *pldq;
//int ldh = *pldh;
for (i = 0; i < nq-4; i+=6)
{
hh_trafo_complex_kernel_6_SSE_1hv_double(&q[i], hh, nb, ldq);
}
if (nq-i == 0) {
return;
} else {
if (nq-i > 2)
{
hh_trafo_complex_kernel_4_SSE_1hv_double(&q[i], hh, nb, ldq);
}
else
{
hh_trafo_complex_kernel_2_SSE_1hv_double(&q[i], hh, nb, ldq);
}
}
}
static __forceinline void hh_trafo_complex_kernel_6_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq)
{
double* q_dbl = (double*)q;
double* hh_dbl = (double*)hh;
__m128d x1, x2, x3, x4, x5, x6;
__m128d q1, q2, q3, q4, q5, q6;
__m128d h1_real, h1_imag;
__m128d tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
int i=0;
__m128d sign = (__m128d)_mm_set_epi64x(0x8000000000000000, 0x8000000000000000);
x1 = _mm_load_pd(&q_dbl[0]);
x2 = _mm_load_pd(&q_dbl[2]);
x3 = _mm_load_pd(&q_dbl[4]);
x4 = _mm_load_pd(&q_dbl[6]);
x5 = _mm_load_pd(&q_dbl[8]);
x6 = _mm_load_pd(&q_dbl[10]);
for (i = 1; i < nb; i++)
{
h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
#ifndef __ELPA_USE_FMA__
// conjugate
h1_imag = _mm_xor_pd(h1_imag, sign);
#endif
q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);
q5 = _mm_load_pd(&q_dbl[(2*i*ldq)+8]);
q6 = _mm_load_pd(&q_dbl[(2*i*ldq)+10]);
tmp1 = _mm_mul_pd(h1_imag, q1);
#ifdef __ELPA_USE_FMA__
x1 = _mm_add_pd(x1, _mm_msubadd_pd(h1_real, q1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
x1 = _mm_add_pd(x1, _mm_addsub_pd( _mm_mul_pd(h1_real, q1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
tmp2 = _mm_mul_pd(h1_imag, q2);
#ifdef __ELPA_USE_FMA__
x2 = _mm_add_pd(x2, _mm_msubadd_pd(h1_real, q2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
x2 = _mm_add_pd(x2, _mm_addsub_pd( _mm_mul_pd(h1_real, q2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
tmp3 = _mm_mul_pd(h1_imag, q3);
#ifdef __ELPA_USE_FMA__
x3 = _mm_add_pd(x3, _mm_msubadd_pd(h1_real, q3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
x3 = _mm_add_pd(x3, _mm_addsub_pd( _mm_mul_pd(h1_real, q3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
tmp4 = _mm_mul_pd(h1_imag, q4);
#ifdef __ELPA_USE_FMA__
x4 = _mm_add_pd(x4, _mm_msubadd_pd(h1_real, q4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
x4 = _mm_add_pd(x4, _mm_addsub_pd( _mm_mul_pd(h1_real, q4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
tmp5 = _mm_mul_pd(h1_imag, q5);
#ifdef __ELPA_USE_FMA__
x5 = _mm_add_pd(x5, _mm_msubadd_pd(h1_real, q5, _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#else
x5 = _mm_add_pd(x5, _mm_addsub_pd( _mm_mul_pd(h1_real, q5), _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#endif
tmp6 = _mm_mul_pd(h1_imag, q6);
#ifdef __ELPA_USE_FMA__
x6 = _mm_add_pd(x6, _mm_msubadd_pd(h1_real, q6, _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#else
x6 = _mm_add_pd(x6, _mm_addsub_pd( _mm_mul_pd(h1_real, q6), _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#endif
}
h1_real = _mm_loaddup_pd(&hh_dbl[0]);
h1_imag = _mm_loaddup_pd(&hh_dbl[1]);
h1_real = _mm_xor_pd(h1_real, sign);
h1_imag = _mm_xor_pd(h1_imag, sign);
tmp1 = _mm_mul_pd(h1_imag, x1);
#ifdef __ELPA_USE_FMA__
x1 = _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#else
x1 = _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#endif
tmp2 = _mm_mul_pd(h1_imag, x2);
#ifdef __ELPA_USE_FMA__
x2 = _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#else
x2 = _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#endif
tmp3 = _mm_mul_pd(h1_imag, x3);
#ifdef __ELPA_USE_FMA__
x3 = _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#else
x3 = _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#endif
tmp4 = _mm_mul_pd(h1_imag, x4);
#ifdef __ELPA_USE_FMA__
x4 = _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#else
x4 = _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#endif
tmp5 = _mm_mul_pd(h1_imag, x5);
#ifdef __ELPA_USE_FMA__
x5 = _mm_maddsub_pd(h1_real, x5, _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1)));
#else
x5 = _mm_addsub_pd( _mm_mul_pd(h1_real, x5), _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1)));
#endif
tmp6 = _mm_mul_pd(h1_imag, x6);
#ifdef __ELPA_USE_FMA__
x6 = _mm_maddsub_pd(h1_real, x6, _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1)));
#else
x6 = _mm_addsub_pd( _mm_mul_pd(h1_real, x6), _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1)));
#endif
q1 = _mm_load_pd(&q_dbl[0]);
q2 = _mm_load_pd(&q_dbl[2]);
q3 = _mm_load_pd(&q_dbl[4]);
q4 = _mm_load_pd(&q_dbl[6]);
q5 = _mm_load_pd(&q_dbl[8]);
q6 = _mm_load_pd(&q_dbl[10]);
q1 = _mm_add_pd(q1, x1);
q2 = _mm_add_pd(q2, x2);
q3 = _mm_add_pd(q3, x3);
q4 = _mm_add_pd(q4, x4);
q5 = _mm_add_pd(q5, x5);
q6 = _mm_add_pd(q6, x6);
_mm_store_pd(&q_dbl[0], q1);
_mm_store_pd(&q_dbl[2], q2);
_mm_store_pd(&q_dbl[4], q3);
_mm_store_pd(&q_dbl[6], q4);
_mm_store_pd(&q_dbl[8], q5);
_mm_store_pd(&q_dbl[10], q6);
for (i = 1; i < nb; i++)
{
h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);
q5 = _mm_load_pd(&q_dbl[(2*i*ldq)+8]);
q6 = _mm_load_pd(&q_dbl[(2*i*ldq)+10]);
tmp1 = _mm_mul_pd(h1_imag, x1);
#ifdef __ELPA_USE_FMA__
q1 = _mm_add_pd(q1, _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
q1 = _mm_add_pd(q1, _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
tmp2 = _mm_mul_pd(h1_imag, x2);
#ifdef __ELPA_USE_FMA__
q2 = _mm_add_pd(q2, _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
q2 = _mm_add_pd(q2, _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
tmp3 = _mm_mul_pd(h1_imag, x3);
#ifdef __ELPA_USE_FMA__
q3 = _mm_add_pd(q3, _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
q3 = _mm_add_pd(q3, _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
tmp4 = _mm_mul_pd(h1_imag, x4);
#ifdef __ELPA_USE_FMA__
q4 = _mm_add_pd(q4, _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
q4 = _mm_add_pd(q4, _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
tmp5 = _mm_mul_pd(h1_imag, x5);
#ifdef __ELPA_USE_FMA__
q5 = _mm_add_pd(q5, _mm_maddsub_pd(h1_real, x5, _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#else
q5 = _mm_add_pd(q5, _mm_addsub_pd( _mm_mul_pd(h1_real, x5), _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#endif
tmp6 = _mm_mul_pd(h1_imag, x6);
#ifdef __ELPA_USE_FMA__
q6 = _mm_add_pd(q6, _mm_maddsub_pd(h1_real, x6, _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#else
q6 = _mm_add_pd(q6, _mm_addsub_pd( _mm_mul_pd(h1_real, x6), _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#endif
_mm_store_pd(&q_dbl[(2*i*ldq)+0], q1);
_mm_store_pd(&q_dbl[(2*i*ldq)+2], q2);
_mm_store_pd(&q_dbl[(2*i*ldq)+4], q3);
_mm_store_pd(&q_dbl[(2*i*ldq)+6], q4);
_mm_store_pd(&q_dbl[(2*i*ldq)+8], q5);
_mm_store_pd(&q_dbl[(2*i*ldq)+10], q6);
}
}
static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq)
{
double* q_dbl = (double*)q;
double* hh_dbl = (double*)hh;
__m128d x1, x2, x3, x4;
__m128d q1, q2, q3, q4;
__m128d h1_real, h1_imag;
__m128d tmp1, tmp2, tmp3, tmp4;
int i=0;
__m128d sign = (__m128d)_mm_set_epi64x(0x8000000000000000, 0x8000000000000000);
x1 = _mm_load_pd(&q_dbl[0]);
x2 = _mm_load_pd(&q_dbl[2]);
x3 = _mm_load_pd(&q_dbl[4]);
x4 = _mm_load_pd(&q_dbl[6]);
for (i = 1; i < nb; i++)
{
h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
#ifndef __ELPA_USE_FMA__
// conjugate
h1_imag = _mm_xor_pd(h1_imag, sign);
#endif
q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);
tmp1 = _mm_mul_pd(h1_imag, q1);
#ifdef __ELPA_USE_FMA__
x1 = _mm_add_pd(x1, _mm_msubadd_pd(h1_real, q1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
x1 = _mm_add_pd(x1, _mm_addsub_pd( _mm_mul_pd(h1_real, q1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
tmp2 = _mm_mul_pd(h1_imag, q2);
#ifdef __ELPA_USE_FMA__
x2 = _mm_add_pd(x2, _mm_msubadd_pd(h1_real, q2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
x2 = _mm_add_pd(x2, _mm_addsub_pd( _mm_mul_pd(h1_real, q2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
tmp3 = _mm_mul_pd(h1_imag, q3);
#ifdef __ELPA_USE_FMA__
x3 = _mm_add_pd(x3, _mm_msubadd_pd(h1_real, q3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
x3 = _mm_add_pd(x3, _mm_addsub_pd( _mm_mul_pd(h1_real, q3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
tmp4 = _mm_mul_pd(h1_imag, q4);
#ifdef __ELPA_USE_FMA__
x4 = _mm_add_pd(x4, _mm_msubadd_pd(h1_real, q4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
x4 = _mm_add_pd(x4, _mm_addsub_pd( _mm_mul_pd(h1_real, q4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
}
h1_real = _mm_loaddup_pd(&hh_dbl[0]);
h1_imag = _mm_loaddup_pd(&hh_dbl[1]);
h1_real = _mm_xor_pd(h1_real, sign);
h1_imag = _mm_xor_pd(h1_imag, sign);
tmp1 = _mm_mul_pd(h1_imag, x1);
#ifdef __ELPA_USE_FMA__
x1 = _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#else
x1 = _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#endif
tmp2 = _mm_mul_pd(h1_imag, x2);
#ifdef __ELPA_USE_FMA__
x2 = _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#else
x2 = _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#endif
tmp3 = _mm_mul_pd(h1_imag, x3);
#ifdef __ELPA_USE_FMA__
x3 = _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#else
x3 = _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#endif
tmp4 = _mm_mul_pd(h1_imag, x4);
#ifdef __ELPA_USE_FMA__
x4 = _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#else
x4 = _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#endif
q1 = _mm_load_pd(&q_dbl[0]);
q2 = _mm_load_pd(&q_dbl[2]);
q3 = _mm_load_pd(&q_dbl[4]);
q4 = _mm_load_pd(&q_dbl[6]);
q1 = _mm_add_pd(q1, x1);
q2 = _mm_add_pd(q2, x2);
q3 = _mm_add_pd(q3, x3);
q4 = _mm_add_pd(q4, x4);
_mm_store_pd(&q_dbl[0], q1);
_mm_store_pd(&q_dbl[2], q2);
_mm_store_pd(&q_dbl[4], q3);
_mm_store_pd(&q_dbl[6], q4);
for (i = 1; i < nb; i++)
{
h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);
tmp1 = _mm_mul_pd(h1_imag, x1);
#ifdef __ELPA_USE_FMA__
q1 = _mm_add_pd(q1, _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
q1 = _mm_add_pd(q1, _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
tmp2 = _mm_mul_pd(h1_imag, x2);
#ifdef __ELPA_USE_FMA__
q2 = _mm_add_pd(q2, _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
q2 = _mm_add_pd(q2, _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
tmp3 = _mm_mul_pd(h1_imag, x3);
#ifdef __ELPA_USE_FMA__
q3 = _mm_add_pd(q3, _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
q3 = _mm_add_pd(q3, _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
tmp4 = _mm_mul_pd(h1_imag, x4);
#ifdef __ELPA_USE_FMA__
q4 = _mm_add_pd(q4, _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
q4 = _mm_add_pd(q4, _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
_mm_store_pd(&q_dbl[(2*i*ldq)+0], q1);
_mm_store_pd(&q_dbl[(2*i*ldq)+2], q2);
_mm_store_pd(&q_dbl[(2*i*ldq)+4], q3);
_mm_store_pd(&q_dbl[(2*i*ldq)+6], q4);
}
}
static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv_double(double complex* q, double complex* hh, int nb, int ldq)
{
double* q_dbl = (double*)q;
double* hh_dbl = (double*)hh;
__m128d x1, x2;
__m128d q1, q2;
__m128d h1_real, h1_imag;
__m128d tmp1, tmp2;
int i=0;
__m128d sign = (__m128d)_mm_set_epi64x(0x8000000000000000, 0x8000000000000000);
x1 = _mm_load_pd(&q_dbl[0]);
x2 = _mm_load_pd(&q_dbl[2]);
for (i = 1; i < nb; i++)
{
h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
#ifndef __ELPA_USE_FMA__
// conjugate
h1_imag = _mm_xor_pd(h1_imag, sign);
#endif
q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
tmp1 = _mm_mul_pd(h1_imag, q1);
#ifdef __ELPA_USE_FMA__
x1 = _mm_add_pd(x1, _mm_msubadd_pd(h1_real, q1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
x1 = _mm_add_pd(x1, _mm_addsub_pd( _mm_mul_pd(h1_real, q1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
tmp2 = _mm_mul_pd(h1_imag, q2);
#ifdef __ELPA_USE_FMA__
x2 = _mm_add_pd(x2, _mm_msubadd_pd(h1_real, q2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
x2 = _mm_add_pd(x2, _mm_addsub_pd( _mm_mul_pd(h1_real, q2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
}
h1_real = _mm_loaddup_pd(&hh_dbl[0]);
h1_imag = _mm_loaddup_pd(&hh_dbl[1]);
h1_real = _mm_xor_pd(h1_real, sign);
h1_imag = _mm_xor_pd(h1_imag, sign);
tmp1 = _mm_mul_pd(h1_imag, x1);
#ifdef __ELPA_USE_FMA__
x1 = _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#else
x1 = _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#endif
tmp2 = _mm_mul_pd(h1_imag, x2);
#ifdef __ELPA_USE_FMA__
x2 = _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#else
x2 = _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#endif
q1 = _mm_load_pd(&q_dbl[0]);
q2 = _mm_load_pd(&q_dbl[2]);
q1 = _mm_add_pd(q1, x1);
q2 = _mm_add_pd(q2, x2);
_mm_store_pd(&q_dbl[0], q1);
_mm_store_pd(&q_dbl[2], q2);
for (i = 1; i < nb; i++)
{
h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
tmp1 = _mm_mul_pd(h1_imag, x1);
#ifdef __ELPA_USE_FMA__
q1 = _mm_add_pd(q1, _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
q1 = _mm_add_pd(q1, _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
tmp2 = _mm_mul_pd(h1_imag, x2);
#ifdef __ELPA_USE_FMA__
q2 = _mm_add_pd(q2, _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
q2 = _mm_add_pd(q2, _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
#define COMPLEXCASE 1
#define DOUBLE_PRECISION 1
#include "../../general/precision_macros.h"
#include "elpa2_kernels_complex_sse_1hv_template.Xc"
#undef DOUBLE_PRECISION
#undef COMPLEXCASE
_mm_store_pd(&q_dbl[(2*i*ldq)+0], q1);
_mm_store_pd(&q_dbl[(2*i*ldq)+2], q2);
}
}
src/elpa2/kernels/elpa2_kernels_complex_sse_1hv_single_precision.c
View file @ 863d1b97
......@@ -42,512 +42,14 @@
// any derivatives of ELPA under the same license that we chose for
// the original distribution, the GNU Lesser General Public License.
//
// Author: Andreas Marek, MPCDF, based on the double precision case of A. Heinecke
//
#include "config-f90.h"
#include <complex.h>
#include <x86intrin.h>
#define __forceinline __attribute__((always_inline))
#ifdef HAVE_SSE_INTRINSICS
#undef __AVX__
#endif
//Forward declaration
static __forceinline void hh_trafo_complex_kernel_6_SSE_1hv_single(float complex* q, float complex* hh, int nb, int ldq);
static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv_single(float complex* q, float complex* hh, int nb, int ldq);
static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv_single(float complex* q, float complex* hh, int nb, int ldq);
/*
!f>#ifdef HAVE_SSE_INTRINSICS
!f> interface
!f> subroutine single_hh_trafo_complex_sse_1hv_single(q, hh, pnb, pnq, pldq) &
!f> bind(C, name="single_hh_trafo_complex_sse_1hv_single")
!f> use, intrinsic :: iso_c_binding
!f> integer(kind=c_int) :: pnb, pnq, pldq
!f> ! complex(kind=c_float_complex) :: q(*)
!f> type(c_ptr), value :: q
!f> complex(kind=c_float_complex) :: hh(pnb,2)