Skip to content

GitLab

elpa
elpa
Commit e54b7632 authored by Andreas Marek's avatar Andreas Marek
Browse files
Options

Start to unfiy double/single real sse block2 kernel

parent ebb92453
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 881 additions and 976 deletions
Makefile.am
View file @ e54b7632
......@@ -589,6 +589,7 @@ EXTRA_DIST = \
src/elpa2/kernels/elpa2_kernels_real_template.X90 \
src/elpa2/kernels/elpa2_kernels_complex_template.X90 \
src/elpa2/kernels/elpa2_kernels_simple_template.X90 \
src/elpa2/kernels/elpa2_kernels_real_sse_2hv_template.Xc \
src/elpa2/redist_band.X90 \
src/elpa2/pack_unpack_cpu.X90 \
src/elpa2/pack_unpack_gpu.X90 \
......
generated_headers.am
View file @ e54b7632
......@@ -28,7 +28,7 @@ test/shared/generated.h: $(wildcard $(top_srcdir)/test/shared/*.*90) | test/shar
@rm -f $@
$(call extract_interface,!c>)
elpa/elpa_generated_fortran_interfaces.h: $(wildcard $(top_srcdir)/src/elpa2/kernels/*.c $(top_srcdir)/src/elpa2/kernels/*.s $(top_srcdir)/src/*.[ch]) | elpa
elpa/elpa_generated_fortran_interfaces.h: $(wildcard $(top_srcdir)/src/elpa2/kernels/*.*c $(top_srcdir)/src/elpa2/kernels/*.s $(top_srcdir)/src/*.[ch]) | elpa
@rm -f $@
$(call extract_interface,!f>)
$(call extract_interface,#!f>)
......
src/elpa2/kernels/elpa2_kernels_real_sse_2hv_double_precision.c
View file @ e54b7632
......@@ -42,532 +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 <x86intrin.h>
#define __forceinline __attribute__((always_inline)) static
#ifdef HAVE_SSE_INTRINSICS
#undef __AVX__
#endif
//Forward declaration
__forceinline void hh_trafo_kernel_4_SSE_2hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s);
__forceinline void hh_trafo_kernel_8_SSE_2hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s);
__forceinline void hh_trafo_kernel_12_SSE_2hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s);
void double_hh_trafo_real_sse_2hv_double(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh);
/*
!f>#ifdef HAVE_SSE_INTRINSICS
!f> interface
!f> subroutine double_hh_trafo_real_sse_2hv_double(q, hh, pnb, pnq, pldq, pldh) &
!f> bind(C, name="double_hh_trafo_real_sse_2hv_double")
!f> use, intrinsic :: iso_c_binding
!f> integer(kind=c_int) :: pnb, pnq, pldq, pldh
!f> type(c_ptr), value :: q
!f> real(kind=c_double) :: hh(pnb,6)
!f> end subroutine
!f> end interface
!f>#endif
*/
void double_hh_trafo_real_sse_2hv_double(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh)
{
int i;
int nb = *pnb;
int nq = *pldq;
int ldq = *pldq;
int ldh = *pldh;
// calculating scalar product to compute
// 2 householder vectors simultaneously
double s = hh[(ldh)+1]*1.0;
#pragma ivdep
for (i = 2; i < nb; i++)
{
s += hh[i-1] * hh[(i+ldh)];
}
// Production level kernel calls with padding
for (i = 0; i < nq-8; i+=12)
{
hh_trafo_kernel_12_SSE_2hv_double(&q[i], hh, nb, ldq, ldh, s);
}
if (nq == i)
{
return;
}
else
{
if (nq-i > 4)
{
hh_trafo_kernel_8_SSE_2hv_double(&q[i], hh, nb, ldq, ldh, s);
}
else if (nq-i > 0)
{
hh_trafo_kernel_4_SSE_2hv_double(&q[i], hh, nb, ldq, ldh, s);
}
}
}
/**
* Unrolled kernel that computes
* 12 rows of Q simultaneously, a
* matrix Vector product with two householder
* vectors + a rank 2 update is performed
*/
__forceinline void hh_trafo_kernel_12_SSE_2hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s)
{
/////////////////////////////////////////////////////
// Matrix Vector Multiplication, Q [12 x nb+1] * hh
// hh contains two householder vectors, with offset 1
/////////////////////////////////////////////////////
int i;
// Needed bit mask for floating point sign flip
__m128d sign = (__m128d)_mm_set1_epi64x(0x8000000000000000LL);
__m128d x1 = _mm_load_pd(&q[ldq]);
__m128d x2 = _mm_load_pd(&q[ldq+2]);
__m128d x3 = _mm_load_pd(&q[ldq+4]);
__m128d x4 = _mm_load_pd(&q[ldq+6]);
__m128d x5 = _mm_load_pd(&q[ldq+8]);
__m128d x6 = _mm_load_pd(&q[ldq+10]);
__m128d h1 = _mm_loaddup_pd(&hh[ldh+1]);
__m128d h2;
__m128d q1 = _mm_load_pd(q);
__m128d y1 = _mm_add_pd(q1, _mm_mul_pd(x1, h1));
__m128d q2 = _mm_load_pd(&q[2]);
__m128d y2 = _mm_add_pd(q2, _mm_mul_pd(x2, h1));
__m128d q3 = _mm_load_pd(&q[4]);
__m128d y3 = _mm_add_pd(q3, _mm_mul_pd(x3, h1));
__m128d q4 = _mm_load_pd(&q[6]);
__m128d y4 = _mm_add_pd(q4, _mm_mul_pd(x4, h1));
__m128d q5 = _mm_load_pd(&q[8]);
__m128d y5 = _mm_add_pd(q5, _mm_mul_pd(x5, h1));
__m128d q6 = _mm_load_pd(&q[10]);
__m128d y6 = _mm_add_pd(q6, _mm_mul_pd(x6, h1));
for(i = 2; i < nb; i++)
{
h1 = _mm_loaddup_pd(&hh[i-1]);
h2 = _mm_loaddup_pd(&hh[ldh+i]);
q1 = _mm_load_pd(&q[i*ldq]);
x1 = _mm_add_pd(x1, _mm_mul_pd(q1,h1));
y1 = _mm_add_pd(y1, _mm_mul_pd(q1,h2));
q2 = _mm_load_pd(&q[(i*ldq)+2]);
x2 = _mm_add_pd(x2, _mm_mul_pd(q2,h1));
y2 = _mm_add_pd(y2, _mm_mul_pd(q2,h2));
q3 = _mm_load_pd(&q[(i*ldq)+4]);
x3 = _mm_add_pd(x3, _mm_mul_pd(q3,h1));
y3 = _mm_add_pd(y3, _mm_mul_pd(q3,h2));
q4 = _mm_load_pd(&q[(i*ldq)+6]);
x4 = _mm_add_pd(x4, _mm_mul_pd(q4,h1));
y4 = _mm_add_pd(y4, _mm_mul_pd(q4,h2));
q5 = _mm_load_pd(&q[(i*ldq)+8]);
x5 = _mm_add_pd(x5, _mm_mul_pd(q5,h1));
y5 = _mm_add_pd(y5, _mm_mul_pd(q5,h2));
q6 = _mm_load_pd(&q[(i*ldq)+10]);
x6 = _mm_add_pd(x6, _mm_mul_pd(q6,h1));
y6 = _mm_add_pd(y6, _mm_mul_pd(q6,h2));
}
h1 = _mm_loaddup_pd(&hh[nb-1]);
q1 = _mm_load_pd(&q[nb*ldq]);
x1 = _mm_add_pd(x1, _mm_mul_pd(q1,h1));
q2 = _mm_load_pd(&q[(nb*ldq)+2]);
x2 = _mm_add_pd(x2, _mm_mul_pd(q2,h1));
q3 = _mm_load_pd(&q[(nb*ldq)+4]);
x3 = _mm_add_pd(x3, _mm_mul_pd(q3,h1));
q4 = _mm_load_pd(&q[(nb*ldq)+6]);
x4 = _mm_add_pd(x4, _mm_mul_pd(q4,h1));
q5 = _mm_load_pd(&q[(nb*ldq)+8]);
x5 = _mm_add_pd(x5, _mm_mul_pd(q5,h1));
q6 = _mm_load_pd(&q[(nb*ldq)+10]);
x6 = _mm_add_pd(x6, _mm_mul_pd(q6,h1));
/////////////////////////////////////////////////////
// Rank-2 update of Q [12 x nb+1]
/////////////////////////////////////////////////////
__m128d tau1 = _mm_loaddup_pd(hh);
__m128d tau2 = _mm_loaddup_pd(&hh[ldh]);
__m128d vs = _mm_loaddup_pd(&s);
h1 = _mm_xor_pd(tau1, sign);
x1 = _mm_mul_pd(x1, h1);
x2 = _mm_mul_pd(x2, h1);
x3 = _mm_mul_pd(x3, h1);
x4 = _mm_mul_pd(x4, h1);
x5 = _mm_mul_pd(x5, h1);
x6 = _mm_mul_pd(x6, h1);
h1 = _mm_xor_pd(tau2, sign);
h2 = _mm_mul_pd(h1, vs);
y1 = _mm_add_pd(_mm_mul_pd(y1,h1), _mm_mul_pd(x1,h2));
y2 = _mm_add_pd(_mm_mul_pd(y2,h1), _mm_mul_pd(x2,h2));
y3 = _mm_add_pd(_mm_mul_pd(y3,h1), _mm_mul_pd(x3,h2));
y4 = _mm_add_pd(_mm_mul_pd(y4,h1), _mm_mul_pd(x4,h2));
y5 = _mm_add_pd(_mm_mul_pd(y5,h1), _mm_mul_pd(x5,h2));
y6 = _mm_add_pd(_mm_mul_pd(y6,h1), _mm_mul_pd(x6,h2));
q1 = _mm_load_pd(q);
q1 = _mm_add_pd(q1, y1);
_mm_store_pd(q,q1);
q2 = _mm_load_pd(&q[2]);
q2 = _mm_add_pd(q2, y2);
_mm_store_pd(&q[2],q2);
q3 = _mm_load_pd(&q[4]);
q3 = _mm_add_pd(q3, y3);
_mm_store_pd(&q[4],q3);
q4 = _mm_load_pd(&q[6]);
q4 = _mm_add_pd(q4, y4);
_mm_store_pd(&q[6],q4);
q5 = _mm_load_pd(&q[8]);
q5 = _mm_add_pd(q5, y5);
_mm_store_pd(&q[8],q5);
q6 = _mm_load_pd(&q[10]);
q6 = _mm_add_pd(q6, y6);
_mm_store_pd(&q[10],q6);
h2 = _mm_loaddup_pd(&hh[ldh+1]);
q1 = _mm_load_pd(&q[ldq]);
q1 = _mm_add_pd(q1, _mm_add_pd(x1, _mm_mul_pd(y1, h2)));
_mm_store_pd(&q[ldq],q1);
q2 = _mm_load_pd(&q[ldq+2]);
q2 = _mm_add_pd(q2, _mm_add_pd(x2, _mm_mul_pd(y2, h2)));
_mm_store_pd(&q[ldq+2],q2);
q3 = _mm_load_pd(&q[ldq+4]);
q3 = _mm_add_pd(q3, _mm_add_pd(x3, _mm_mul_pd(y3, h2)));
_mm_store_pd(&q[ldq+4],q3);
q4 = _mm_load_pd(&q[ldq+6]);
q4 = _mm_add_pd(q4, _mm_add_pd(x4, _mm_mul_pd(y4, h2)));
_mm_store_pd(&q[ldq+6],q4);
q5 = _mm_load_pd(&q[ldq+8]);
q5 = _mm_add_pd(q5, _mm_add_pd(x5, _mm_mul_pd(y5, h2)));
_mm_store_pd(&q[ldq+8],q5);
q6 = _mm_load_pd(&q[ldq+10]);
q6 = _mm_add_pd(q6, _mm_add_pd(x6, _mm_mul_pd(y6, h2)));
_mm_store_pd(&q[ldq+10],q6);
for (i = 2; i < nb; i++)
{
h1 = _mm_loaddup_pd(&hh[i-1]);
h2 = _mm_loaddup_pd(&hh[ldh+i]);
q1 = _mm_load_pd(&q[i*ldq]);
q1 = _mm_add_pd(q1, _mm_add_pd(_mm_mul_pd(x1,h1), _mm_mul_pd(y1, h2)));
_mm_store_pd(&q[i*ldq],q1);
q2 = _mm_load_pd(&q[(i*ldq)+2]);
q2 = _mm_add_pd(q2, _mm_add_pd(_mm_mul_pd(x2,h1), _mm_mul_pd(y2, h2)));
_mm_store_pd(&q[(i*ldq)+2],q2);
q3 = _mm_load_pd(&q[(i*ldq)+4]);
q3 = _mm_add_pd(q3, _mm_add_pd(_mm_mul_pd(x3,h1), _mm_mul_pd(y3, h2)));
_mm_store_pd(&q[(i*ldq)+4],q3);
q4 = _mm_load_pd(&q[(i*ldq)+6]);
q4 = _mm_add_pd(q4, _mm_add_pd(_mm_mul_pd(x4,h1), _mm_mul_pd(y4, h2)));
_mm_store_pd(&q[(i*ldq)+6],q4);
q5 = _mm_load_pd(&q[(i*ldq)+8]);
q5 = _mm_add_pd(q5, _mm_add_pd(_mm_mul_pd(x5,h1), _mm_mul_pd(y5, h2)));
_mm_store_pd(&q[(i*ldq)+8],q5);
q6 = _mm_load_pd(&q[(i*ldq)+10]);
q6 = _mm_add_pd(q6, _mm_add_pd(_mm_mul_pd(x6,h1), _mm_mul_pd(y6, h2)));
_mm_store_pd(&q[(i*ldq)+10],q6);
}
h1 = _mm_loaddup_pd(&hh[nb-1]);
q1 = _mm_load_pd(&q[nb*ldq]);
q1 = _mm_add_pd(q1, _mm_mul_pd(x1, h1));
_mm_store_pd(&q[nb*ldq],q1);
q2 = _mm_load_pd(&q[(nb*ldq)+2]);
q2 = _mm_add_pd(q2, _mm_mul_pd(x2, h1));
_mm_store_pd(&q[(nb*ldq)+2],q2);
q3 = _mm_load_pd(&q[(nb*ldq)+4]);
q3 = _mm_add_pd(q3, _mm_mul_pd(x3, h1));
_mm_store_pd(&q[(nb*ldq)+4],q3);
q4 = _mm_load_pd(&q[(nb*ldq)+6]);
q4 = _mm_add_pd(q4, _mm_mul_pd(x4, h1));
_mm_store_pd(&q[(nb*ldq)+6],q4);
q5 = _mm_load_pd(&q[(nb*ldq)+8]);
q5 = _mm_add_pd(q5, _mm_mul_pd(x5, h1));
_mm_store_pd(&q[(nb*ldq)+8],q5);
q6 = _mm_load_pd(&q[(nb*ldq)+10]);
q6 = _mm_add_pd(q6, _mm_mul_pd(x6, h1));
_mm_store_pd(&q[(nb*ldq)+10],q6);
}
/**
* Unrolled kernel that computes
* 8 rows of Q simultaneously, a
* matrix Vector product with two householder
* vectors + a rank 2 update is performed
*/
__forceinline void hh_trafo_kernel_8_SSE_2hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s)
{
/////////////////////////////////////////////////////
// Matrix Vector Multiplication, Q [8 x nb+1] * hh
// hh contains two householder vectors, with offset 1
/////////////////////////////////////////////////////
int i;
// Needed bit mask for floating point sign flip
__m128d sign = (__m128d)_mm_set1_epi64x(0x8000000000000000LL);
__m128d x1 = _mm_load_pd(&q[ldq]);
__m128d x2 = _mm_load_pd(&q[ldq+2]);
__m128d x3 = _mm_load_pd(&q[ldq+4]);
__m128d x4 = _mm_load_pd(&q[ldq+6]);
__m128d h1 = _mm_loaddup_pd(&hh[ldh+1]);
__m128d h2;
__m128d q1 = _mm_load_pd(q);
__m128d y1 = _mm_add_pd(q1, _mm_mul_pd(x1, h1));
__m128d q2 = _mm_load_pd(&q[2]);
__m128d y2 = _mm_add_pd(q2, _mm_mul_pd(x2, h1));
__m128d q3 = _mm_load_pd(&q[4]);
__m128d y3 = _mm_add_pd(q3, _mm_mul_pd(x3, h1));
__m128d q4 = _mm_load_pd(&q[6]);
__m128d y4 = _mm_add_pd(q4, _mm_mul_pd(x4, h1));
for(i = 2; i < nb; i++)
{
h1 = _mm_loaddup_pd(&hh[i-1]);
h2 = _mm_loaddup_pd(&hh[ldh+i]);
q1 = _mm_load_pd(&q[i*ldq]);
x1 = _mm_add_pd(x1, _mm_mul_pd(q1,h1));
y1 = _mm_add_pd(y1, _mm_mul_pd(q1,h2));
q2 = _mm_load_pd(&q[(i*ldq)+2]);
x2 = _mm_add_pd(x2, _mm_mul_pd(q2,h1));
y2 = _mm_add_pd(y2, _mm_mul_pd(q2,h2));
q3 = _mm_load_pd(&q[(i*ldq)+4]);
x3 = _mm_add_pd(x3, _mm_mul_pd(q3,h1));
y3 = _mm_add_pd(y3, _mm_mul_pd(q3,h2));
q4 = _mm_load_pd(&q[(i*ldq)+6]);
x4 = _mm_add_pd(x4, _mm_mul_pd(q4,h1));
y4 = _mm_add_pd(y4, _mm_mul_pd(q4,h2));
}
h1 = _mm_loaddup_pd(&hh[nb-1]);
q1 = _mm_load_pd(&q[nb*ldq]);
x1 = _mm_add_pd(x1, _mm_mul_pd(q1,h1));
q2 = _mm_load_pd(&q[(nb*ldq)+2]);
x2 = _mm_add_pd(x2, _mm_mul_pd(q2,h1));
q3 = _mm_load_pd(&q[(nb*ldq)+4]);
x3 = _mm_add_pd(x3, _mm_mul_pd(q3,h1));
q4 = _mm_load_pd(&q[(nb*ldq)+6]);
x4 = _mm_add_pd(x4, _mm_mul_pd(q4,h1));
/////////////////////////////////////////////////////
// Rank-2 update of Q [8 x nb+1]
/////////////////////////////////////////////////////
__m128d tau1 = _mm_loaddup_pd(hh);
__m128d tau2 = _mm_loaddup_pd(&hh[ldh]);
__m128d vs = _mm_loaddup_pd(&s);
h1 = _mm_xor_pd(tau1, sign);
x1 = _mm_mul_pd(x1, h1);
x2 = _mm_mul_pd(x2, h1);
x3 = _mm_mul_pd(x3, h1);
x4 = _mm_mul_pd(x4, h1);
h1 = _mm_xor_pd(tau2, sign);
h2 = _mm_mul_pd(h1, vs);
y1 = _mm_add_pd(_mm_mul_pd(y1,h1), _mm_mul_pd(x1,h2));
y2 = _mm_add_pd(_mm_mul_pd(y2,h1), _mm_mul_pd(x2,h2));
y3 = _mm_add_pd(_mm_mul_pd(y3,h1), _mm_mul_pd(x3,h2));
y4 = _mm_add_pd(_mm_mul_pd(y4,h1), _mm_mul_pd(x4,h2));
q1 = _mm_load_pd(q);
q1 = _mm_add_pd(q1, y1);
_mm_store_pd(q,q1);
q2 = _mm_load_pd(&q[2]);
q2 = _mm_add_pd(q2, y2);
_mm_store_pd(&q[2],q2);
q3 = _mm_load_pd(&q[4]);
q3 = _mm_add_pd(q3, y3);
_mm_store_pd(&q[4],q3);
q4 = _mm_load_pd(&q[6]);
q4 = _mm_add_pd(q4, y4);
_mm_store_pd(&q[6],q4);
h2 = _mm_loaddup_pd(&hh[ldh+1]);
q1 = _mm_load_pd(&q[ldq]);
q1 = _mm_add_pd(q1, _mm_add_pd(x1, _mm_mul_pd(y1, h2)));
_mm_store_pd(&q[ldq],q1);
q2 = _mm_load_pd(&q[ldq+2]);
q2 = _mm_add_pd(q2, _mm_add_pd(x2, _mm_mul_pd(y2, h2)));
_mm_store_pd(&q[ldq+2],q2);
q3 = _mm_load_pd(&q[ldq+4]);
q3 = _mm_add_pd(q3, _mm_add_pd(x3, _mm_mul_pd(y3, h2)));
_mm_store_pd(&q[ldq+4],q3);
q4 = _mm_load_pd(&q[ldq+6]);
q4 = _mm_add_pd(q4, _mm_add_pd(x4, _mm_mul_pd(y4, h2)));
_mm_store_pd(&q[ldq+6],q4);
for (i = 2; i < nb; i++)
{
h1 = _mm_loaddup_pd(&hh[i-1]);
h2 = _mm_loaddup_pd(&hh[ldh+i]);
q1 = _mm_load_pd(&q[i*ldq]);
q1 = _mm_add_pd(q1, _mm_add_pd(_mm_mul_pd(x1,h1), _mm_mul_pd(y1, h2)));
_mm_store_pd(&q[i*ldq],q1);
q2 = _mm_load_pd(&q[(i*ldq)+2]);
q2 = _mm_add_pd(q2, _mm_add_pd(_mm_mul_pd(x2,h1), _mm_mul_pd(y2, h2)));
_mm_store_pd(&q[(i*ldq)+2],q2);
q3 = _mm_load_pd(&q[(i*ldq)+4]);
q3 = _mm_add_pd(q3, _mm_add_pd(_mm_mul_pd(x3,h1), _mm_mul_pd(y3, h2)));
_mm_store_pd(&q[(i*ldq)+4],q3);
q4 = _mm_load_pd(&q[(i*ldq)+6]);
q4 = _mm_add_pd(q4, _mm_add_pd(_mm_mul_pd(x4,h1), _mm_mul_pd(y4, h2)));
_mm_store_pd(&q[(i*ldq)+6],q4);
}
h1 = _mm_loaddup_pd(&hh[nb-1]);
q1 = _mm_load_pd(&q[nb*ldq]);
q1 = _mm_add_pd(q1, _mm_mul_pd(x1, h1));
_mm_store_pd(&q[nb*ldq],q1);
q2 = _mm_load_pd(&q[(nb*ldq)+2]);
q2 = _mm_add_pd(q2, _mm_mul_pd(x2, h1));
_mm_store_pd(&q[(nb*ldq)+2],q2);
q3 = _mm_load_pd(&q[(nb*ldq)+4]);
q3 = _mm_add_pd(q3, _mm_mul_pd(x3, h1));
_mm_store_pd(&q[(nb*ldq)+4],q3);
q4 = _mm_load_pd(&q[(nb*ldq)+6]);
q4 = _mm_add_pd(q4, _mm_mul_pd(x4, h1));
_mm_store_pd(&q[(nb*ldq)+6],q4);
}
/**
* Unrolled kernel that computes
* 4 rows of Q simultaneously, a
* matrix Vector product with two householder
* vectors + a rank 2 update is performed
*/
__forceinline void hh_trafo_kernel_4_SSE_2hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s)
{
/////////////////////////////////////////////////////
// Matrix Vector Multiplication, Q [4 x nb+1] * hh
// hh contains two householder vectors, with offset 1
/////////////////////////////////////////////////////
int i;
// Needed bit mask for floating point sign flip
__m128d sign = (__m128d)_mm_set1_epi64x(0x8000000000000000LL);
__m128d x1 = _mm_load_pd(&q[ldq]);
__m128d x2 = _mm_load_pd(&q[ldq+2]);
__m128d h1 = _mm_loaddup_pd(&hh[ldh+1]);
__m128d h2;
__m128d q1 = _mm_load_pd(q);
__m128d y1 = _mm_add_pd(q1, _mm_mul_pd(x1, h1));
__m128d q2 = _mm_load_pd(&q[2]);
__m128d y2 = _mm_add_pd(q2, _mm_mul_pd(x2, h1));
for(i = 2; i < nb; i++)
{
h1 = _mm_loaddup_pd(&hh[i-1]);
h2 = _mm_loaddup_pd(&hh[ldh+i]);
q1 = _mm_load_pd(&q[i*ldq]);
x1 = _mm_add_pd(x1, _mm_mul_pd(q1,h1));
y1 = _mm_add_pd(y1, _mm_mul_pd(q1,h2));
q2 = _mm_load_pd(&q[(i*ldq)+2]);
x2 = _mm_add_pd(x2, _mm_mul_pd(q2,h1));
y2 = _mm_add_pd(y2, _mm_mul_pd(q2,h2));
}
h1 = _mm_loaddup_pd(&hh[nb-1]);
q1 = _mm_load_pd(&q[nb*ldq]);
x1 = _mm_add_pd(x1, _mm_mul_pd(q1,h1));
q2 = _mm_load_pd(&q[(nb*ldq)+2]);
x2 = _mm_add_pd(x2, _mm_mul_pd(q2,h1));
/////////////////////////////////////////////////////
// Rank-2 update of Q [12 x nb+1]
/////////////////////////////////////////////////////
__m128d tau1 = _mm_loaddup_pd(hh);
__m128d tau2 = _mm_loaddup_pd(&hh[ldh]);
__m128d vs = _mm_loaddup_pd(&s);
h1 = _mm_xor_pd(tau1, sign);
x1 = _mm_mul_pd(x1, h1);
x2 = _mm_mul_pd(x2, h1);
h1 = _mm_xor_pd(tau2, sign);
h2 = _mm_mul_pd(h1, vs);
y1 = _mm_add_pd(_mm_mul_pd(y1,h1), _mm_mul_pd(x1,h2));
y2 = _mm_add_pd(_mm_mul_pd(y2,h1), _mm_mul_pd(x2,h2));
q1 = _mm_load_pd(q);
q1 = _mm_add_pd(q1, y1);
_mm_store_pd(q,q1);
q2 = _mm_load_pd(&q[2]);
q2 = _mm_add_pd(q2, y2);
_mm_store_pd(&q[2],q2);
h2 = _mm_loaddup_pd(&hh[ldh+1]);
q1 = _mm_load_pd(&q[ldq]);
q1 = _mm_add_pd(q1, _mm_add_pd(x1, _mm_mul_pd(y1, h2)));
_mm_store_pd(&q[ldq],q1);
q2 = _mm_load_pd(&q[ldq+2]);
q2 = _mm_add_pd(q2, _mm_add_pd(x2, _mm_mul_pd(y2, h2)));
_mm_store_pd(&q[ldq+2],q2);
for (i = 2; i < nb; i++)
{
h1 = _mm_loaddup_pd(&hh[i-1]);
h2 = _mm_loaddup_pd(&hh[ldh+i]);
q1 = _mm_load_pd(&q[i*ldq]);
q1 = _mm_add_pd(q1, _mm_add_pd(_mm_mul_pd(x1,h1), _mm_mul_pd(y1, h2)));
_mm_store_pd(&q[i*ldq],q1);
q2 = _mm_load_pd(&q[(i*ldq)+2]);
q2 = _mm_add_pd(q2, _mm_add_pd(_mm_mul_pd(x2,h1), _mm_mul_pd(y2, h2)));
_mm_store_pd(&q[(i*ldq)+2],q2);
}
h1 = _mm_loaddup_pd(&hh[nb-1]);
#define REALCASE 1
#define DOUBLE_PRECISION 1
#include "../../general/precision_macros.h"
#include "elpa2_kernels_real_sse_2hv_template.Xc"
#undef REALCASE
#undef DOUBLE_PRECISION
q1 = _mm_load_pd(&q[nb*ldq]);
q1 = _mm_add_pd(q1, _mm_mul_pd(x1, h1));
_mm_store_pd(&q[nb*ldq],q1);
q2 = _mm_load_pd(&q[(nb*ldq)+2]);
q2 = _mm_add_pd(q2, _mm_mul_pd(x2, h1));
_mm_store_pd(&q[(nb*ldq)+2],q2);
}
src/elpa2/kernels/elpa2_kernels_real_sse_2hv_single_precision.c
View file @ e54b7632
......@@ -42,455 +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 <x86intrin.h>
#define __forceinline __attribute__((always_inline)) static
#ifdef HAVE_SSE_INTRINSICS
#undef __AVX__
#endif
//Forward declaration
__forceinline void hh_trafo_kernel_4_SSE_2hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s);
__forceinline void hh_trafo_kernel_8_SSE_2hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s);
__forceinline void hh_trafo_kernel_12_SSE_2hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s);
void double_hh_trafo_real_sse_2hv_single_(float* q, float* hh, int* pnb, int* pnq, int* pldq, int* pldh);
/*
!f>#ifdef HAVE_SSE_INTRINSICS