floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #419 from wernsaar/develop 

added optimized sgemv kernels for Sandy Bridge, Haswell, Bullldozer, and Piledriver.
This commit is contained in:
Zhang Xianyi 2014-07-20 23:35:17 +08:00
parent e6668dd83b 6acbafe45b
commit 80bf3e6a35
18 changed files with 2502 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Makefile
View File

@ -247,7 +247,7 @@ ifndef NOFORTRAN
-@echo "SUFFIX = $(SUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc -@echo "SUFFIX = $(SUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
-@echo "PSUFFIX = $(PSUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc -@echo "PSUFFIX = $(PSUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
-@echo "CEXTRALIB = $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc -@echo "CEXTRALIB = $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
ifeq ($(FC), GFORTRAN) ifeq ($(FC), gfortran)
-@echo "TIMER = INT_ETIME" >> $(NETLIB_LAPACK_DIR)/make.inc -@echo "TIMER = INT_ETIME" >> $(NETLIB_LAPACK_DIR)/make.inc
ifdef SMP ifdef SMP
-@echo "LOADER = $(FC) -pthread" >> $(NETLIB_LAPACK_DIR)/make.inc -@echo "LOADER = $(FC) -pthread" >> $(NETLIB_LAPACK_DIR)/make.inc

77
benchmark/Makefile
View File

@ -2,12 +2,12 @@ TOPDIR = ..
include $(TOPDIR)/Makefile.system include $(TOPDIR)/Makefile.system
# ACML standard # ACML standard
ACML=/opt/acml5.3.1/gfortran64_mp/lib #ACML=/opt/acml5.3.1/gfortran64_mp/lib
LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm #LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
# ACML custom # ACML custom
#ACML=/opt/pb/acml-5-3-1-gfortran-64bit/gfortran64_fma4_mp/lib ACML=/opt/pb/acml-5-3-1-gfortran-64bit/gfortran64_fma4_mp/lib
#LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
# Atlas Ubuntu # Atlas Ubuntu
#ATLAS=/usr/lib/atlas-base #ATLAS=/usr/lib/atlas-base
@ -37,6 +37,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
chemm.goto zhemm.goto \ chemm.goto zhemm.goto \
cherk.goto zherk.goto \ cherk.goto zherk.goto \
cher2k.goto zher2k.goto \ cher2k.goto zher2k.goto \
sgemv.goto dgemv.goto cgemv.goto zgemv.goto \
ssymm.goto dsymm.goto csymm.goto zsymm.goto ssymm.goto dsymm.goto csymm.goto zsymm.goto
acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
@ -49,6 +50,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
chemm.acml zhemm.acml \ chemm.acml zhemm.acml \
cherk.acml zherk.acml \ cherk.acml zherk.acml \
cher2k.acml zher2k.acml \ cher2k.acml zher2k.acml \
sgemv.acml dgemv.acml cgemv.acml zgemv.acml \
ssymm.acml dsymm.acml csymm.acml zsymm.acml ssymm.acml dsymm.acml csymm.acml zsymm.acml
atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
@ -61,6 +63,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
chemm.atlas zhemm.atlas \ chemm.atlas zhemm.atlas \
cherk.atlas zherk.atlas \ cherk.atlas zherk.atlas \
cher2k.atlas zher2k.atlas \ cher2k.atlas zher2k.atlas \
sgemv.atlas dgemv.atlas cgemv.atlas zgemv.atlas \
ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas
mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
@ -73,6 +76,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
chemm.mkl zhemm.mkl \ chemm.mkl zhemm.mkl \
cherk.mkl zherk.mkl \ cherk.mkl zherk.mkl \
cher2k.mkl zher2k.mkl \ cher2k.mkl zher2k.mkl \
sgemv.mkl dgemv.mkl cgemv.mkl zgemv.mkl \
ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl
all :: goto atlas acml mkl all :: goto atlas acml mkl
@ -601,6 +605,61 @@ zher2k.atlas : zher2k.$(SUFFIX)
zher2k.mkl : zher2k.$(SUFFIX) zher2k.mkl : zher2k.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
##################################### Sgemv ####################################################
sgemv.goto : sgemv.$(SUFFIX) ../$(LIBNAME)
$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
sgemv.acml : sgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
sgemv.atlas : sgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
sgemv.mkl : sgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
##################################### Dgemv ####################################################
dgemv.goto : dgemv.$(SUFFIX) ../$(LIBNAME)
$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
dgemv.acml : dgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
dgemv.atlas : dgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
dgemv.mkl : dgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
##################################### Cgemv ####################################################
cgemv.goto : cgemv.$(SUFFIX) ../$(LIBNAME)
$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
cgemv.acml : cgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
cgemv.atlas : cgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
cgemv.mkl : cgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
##################################### Zgemv ####################################################
zgemv.goto : zgemv.$(SUFFIX) ../$(LIBNAME)
$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
zgemv.acml : zgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
zgemv.atlas : zgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
zgemv.mkl : zgemv.$(SUFFIX)
-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
################################################################################################### ###################################################################################################
slinpack.$(SUFFIX) : linpack.c slinpack.$(SUFFIX) : linpack.c
@ -717,7 +776,17 @@ cher2k.$(SUFFIX) : her2k.c
zher2k.$(SUFFIX) : her2k.c zher2k.$(SUFFIX) : her2k.c
$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
sgemv.$(SUFFIX) : gemv.c
$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
dgemv.$(SUFFIX) : gemv.c
$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
cgemv.$(SUFFIX) : gemv.c
$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
zgemv.$(SUFFIX) : gemv.c
$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
clean :: clean ::

229
benchmark/gemv.c Normal file
View File

@ -0,0 +1,229 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
#include <stdio.h>
#include <stdlib.h>
#ifdef __CYGWIN32__
#include <sys/time.h>
#endif
#include "common.h"
#undef GEMV
#ifndef COMPLEX
#ifdef DOUBLE
#define GEMV BLASFUNC(dgemv)
#else
#define GEMV BLASFUNC(sgemv)
#endif
#else
#ifdef DOUBLE
#define GEMV BLASFUNC(zgemv)
#else
#define GEMV BLASFUNC(cgemv)
#endif
#endif
#if defined(__WIN32__) || defined(__WIN64__)
#ifndef DELTA_EPOCH_IN_MICROSECS
#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
#endif
int gettimeofday(struct timeval *tv, void *tz){
FILETIME ft;
unsigned __int64 tmpres = 0;
static int tzflag;
if (NULL != tv)
{
GetSystemTimeAsFileTime(&ft);
tmpres |= ft.dwHighDateTime;
tmpres <<= 32;
tmpres |= ft.dwLowDateTime;
/*converting file time to unix epoch*/
tmpres /= 10; /*convert into microseconds*/
tmpres -= DELTA_EPOCH_IN_MICROSECS;
tv->tv_sec = (long)(tmpres / 1000000UL);
tv->tv_usec = (long)(tmpres % 1000000UL);
}
return 0;
}
#endif
#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
static void *huge_malloc(BLASLONG size){
int shmid;
void *address;
#ifndef SHM_HUGETLB
#define SHM_HUGETLB 04000
#endif
if ((shmid =shmget(IPC_PRIVATE,
(size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
printf( "Memory allocation failed(shmget).\n");
exit(1);
}
address = shmat(shmid, NULL, SHM_RND);
if ((BLASLONG)address == -1){
printf( "Memory allocation failed(shmat).\n");
exit(1);
}
shmctl(shmid, IPC_RMID, 0);
return address;
}
#define malloc huge_malloc
#endif
int MAIN__(int argc, char *argv[]){
FLOAT *a, *x, *y;
FLOAT alpha[] = {1.0, 1.0};
FLOAT beta [] = {1.0, 1.0};
char trans='N';
blasint m, i, j;
blasint inc_x=1,inc_y=1;
blasint n=0;
int has_param_n = 0;
int loops = 1;
int l;
char *p;
int from = 1;
int to = 200;
int step = 1;
struct timeval start, stop;
double time1,timeg;
argc--;argv++;
if (argc > 0) { from = atol(*argv); argc--; argv++;}
if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;}
if (argc > 0) { step = atol(*argv); argc--; argv++;}
if ((p = getenv("OPENBLAS_LOOPS"))) loops = atoi(p);
if ((p = getenv("OPENBLAS_INCX"))) inc_x = atoi(p);
if ((p = getenv("OPENBLAS_INCY"))) inc_y = atoi(p);
if ((p = getenv("OPENBLAS_TRANS"))) trans=*p;
if ((p = getenv("OPENBLAS_PARAM_N"))) {
n = atoi(p);
if ((n>0) && (n<=to)) has_param_n = 1;
}
if ( has_param_n == 1 )
fprintf(stderr, "From : %3d To : %3d Step = %3d Trans = '%c' N = %d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,trans,n,inc_x,inc_y,loops);
else
fprintf(stderr, "From : %3d To : %3d Step = %3d Trans = '%c' Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,trans,inc_x,inc_y,loops);
if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
fprintf(stderr,"Out of Memory!!\n");exit(1);
}
if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
fprintf(stderr,"Out of Memory!!\n");exit(1);
}
if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
fprintf(stderr,"Out of Memory!!\n");exit(1);
}
#ifdef linux
srandom(getpid());
#endif
fprintf(stderr, " SIZE Flops\n");
for(m = from; m <= to; m += step)
{
timeg=0;
if ( has_param_n == 0 ) n = m;
fprintf(stderr, " %6dx%d : ", (int)m,(int)n);
for(j = 0; j < m; j++){
for(i = 0; i < n * COMPSIZE; i++){
a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
}
}
for (l=0; l<loops; l++)
{
for(i = 0; i < n * COMPSIZE * abs(inc_x); i++){
x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
}
for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){
y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
}
gettimeofday( &start, (struct timezone *)0);
GEMV (&trans, &m, &n, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
gettimeofday( &stop, (struct timezone *)0);
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
timeg += time1;
}
timeg /= loops;
fprintf(stderr,
" %10.2f MFlops\n",
COMPSIZE * COMPSIZE * 2. * (double)m * (double)n / timeg * 1.e-6);
}
return 0;
}
void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

4
interface/gemm.c
View File

@ -405,11 +405,11 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
#ifndef COMPLEX #ifndef COMPLEX
double MNK = (double) args.m * (double) args.n * (double) args.k; double MNK = (double) args.m * (double) args.n * (double) args.k;
if ( MNK <= (1024.0 * (double) GEMM_MULTITHREAD_THRESHOLD) ) if ( MNK <= (16.0 * 1024.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
nthreads_max = 1; nthreads_max = 1;
else else
{ {
if ( MNK <= (65536.0 * (double) GEMM_MULTITHREAD_THRESHOLD) ) if ( MNK <= (2.0 * 65536.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
{ {
nthreads_max = 4; nthreads_max = 4;
if ( args.m < 16 * GEMM_MULTITHREAD_THRESHOLD ) if ( args.m < 16 * GEMM_MULTITHREAD_THRESHOLD )

13
interface/gemv.c
View File

@ -211,7 +211,18 @@ void CNAME(enum CBLAS_ORDER order,
buffer = (FLOAT *)blas_memory_alloc(1); buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP #ifdef SMP
nthreads = num_cpu_avail(2);
int nthreads_max = num_cpu_avail(2);
int nthreads_avail = nthreads_max;
double MNK = (double) m * (double) n;
if ( MNK <= (500.0 * 100.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )
nthreads = nthreads_avail;
else
nthreads = nthreads_max;
if (nthreads == 1) { if (nthreads == 1) {
#endif #endif

14
interface/zgemv.c
View File

@ -233,7 +233,19 @@ void CNAME(enum CBLAS_ORDER order,
buffer = (FLOAT *)blas_memory_alloc(1); buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP #ifdef SMP
nthreads = num_cpu_avail(2);
int nthreads_max = num_cpu_avail(2);
int nthreads_avail = nthreads_max;
double MNK = (double) m * (double) n;
if ( MNK <= (80.0 * 20.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )
nthreads = nthreads_avail;
else
nthreads = nthreads_max;
if (nthreads == 1) { if (nthreads == 1) {
#endif #endif

8
kernel/x86_64/KERNEL.BULLDOZER
View File

@ -1,3 +1,11 @@
ifdef OS_WINDOWS
SGEMVNKERNEL = ../arm/gemv_n.c
SGEMVTKERNEL = ../arm/gemv_t.c
else
SGEMVNKERNEL = sgemv_n_avx.c
SGEMVTKERNEL = sgemv_t_avx.c
endif
ZGEMVNKERNEL = zgemv_n_dup.S ZGEMVNKERNEL = zgemv_n_dup.S
ZGEMVTKERNEL = zgemv_t.S ZGEMVTKERNEL = zgemv_t.S

8
kernel/x86_64/KERNEL.HASWELL
View File

@ -1,3 +1,11 @@
ifdef OS_WINDOWS
SGEMVNKERNEL = ../arm/gemv_n.c
SGEMVTKERNEL = ../arm/gemv_t.c
else
SGEMVNKERNEL = sgemv_n_avx.c
SGEMVTKERNEL = sgemv_t_avx.c
endif
SGEMMKERNEL = sgemm_kernel_16x4_haswell.S SGEMMKERNEL = sgemm_kernel_16x4_haswell.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c SGEMMINCOPY = ../generic/gemm_ncopy_16.c

8
kernel/x86_64/KERNEL.PILEDRIVER
View File

@ -1,3 +1,11 @@
ifdef OS_WINDOWS
SGEMVNKERNEL = ../arm/gemv_n.c
SGEMVTKERNEL = ../arm/gemv_t.c
else
SGEMVNKERNEL = sgemv_n_avx.c
SGEMVTKERNEL = sgemv_t_avx.c
endif
ZGEMVNKERNEL = zgemv_n_dup.S ZGEMVNKERNEL = zgemv_n_dup.S
ZGEMVTKERNEL = zgemv_t.S ZGEMVTKERNEL = zgemv_t.S

8
kernel/x86_64/KERNEL.SANDYBRIDGE
View File

@ -1,3 +1,11 @@
ifdef OS_WINDOWS
SGEMVNKERNEL = ../arm/gemv_n.c
SGEMVTKERNEL = ../arm/gemv_t.c
else
SGEMVNKERNEL = sgemv_n_avx.c
SGEMVTKERNEL = sgemv_t_avx.c
endif
SGEMMKERNEL = sgemm_kernel_16x4_sandy.S SGEMMKERNEL = sgemm_kernel_16x4_sandy.S
SGEMMINCOPY = ../generic/gemm_ncopy_16.c SGEMMINCOPY = ../generic/gemm_ncopy_16.c

218
kernel/x86_64/sgemv_n_avx.c Normal file
View File

@ -0,0 +1,218 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
#include "common.h"
#if defined(BULLDOZER) || defined(PILEDRIVER)
#include "sgemv_n_microk_bulldozer.c"
#elif defined(HASWELL)
#include "sgemv_n_microk_haswell.c"
#else
#include "sgemv_n_microk_sandy.c"
#endif
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
dest++;
src += inc_src;
}
}
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest += *src;
src++;
dest += inc_dest;
}
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
BLASLONG j;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
BLASLONG n1;
BLASLONG m1;
BLASLONG register m2;
BLASLONG register n2;
FLOAT *xbuffer,*ybuffer;
xbuffer = buffer;
ybuffer = xbuffer + 2048 + 256;
n1 = n / 512 ;
n2 = n % 512 ;
m1 = m / 64;
m2 = m % 64;
y_ptr = y;
x_ptr = x;
for (j=0; j<n1; j++)
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(512,x_ptr,xbuffer,inc_x);
a_ptr = a + j * 512 * lda;
y_ptr = y;
for(i = 0; i<m1; i++ )
{
sgemv_kernel_64(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(64,ybuffer,y_ptr,inc_y);
y_ptr += 64 * inc_y;
a_ptr += 64;
}
if ( m2 & 32 )
{
sgemv_kernel_32(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(32,ybuffer,y_ptr,inc_y);
y_ptr += 32 * inc_y;
a_ptr += 32;
}
if ( m2 & 16 )
{
sgemv_kernel_16(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(16,ybuffer,y_ptr,inc_y);
y_ptr += 16 * inc_y;
a_ptr += 16;
}
if ( m2 & 8 )
{
sgemv_kernel_8(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(8,ybuffer,y_ptr,inc_y);
y_ptr += 8 * inc_y;
a_ptr += 8;
}
if ( m2 & 4 )
{
sgemv_kernel_4(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(4,ybuffer,y_ptr,inc_y);
y_ptr += 4 * inc_y;
a_ptr += 4;
}
if ( m2 & 2 )
{
sgemv_kernel_2(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(2,ybuffer,y_ptr,inc_y);
y_ptr += 2 * inc_y;
a_ptr += 2;
}
if ( m2 & 1 )
{
sgemv_kernel_1(512,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(1,ybuffer,y_ptr,inc_y);
}
x_ptr += 512 * inc_x;
}
if ( n2 > 0 )
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(n2,x_ptr,xbuffer,inc_x);
a_ptr = a + n1 * 512 * lda;
y_ptr = y;
for(i = 0; i<m1; i++ )
{
sgemv_kernel_64(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(64,ybuffer,y_ptr,inc_y);
y_ptr += 64 * inc_y;
a_ptr += 64;
}
if ( m2 & 32 )
{
sgemv_kernel_32(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(32,ybuffer,y_ptr,inc_y);
y_ptr += 32 * inc_y;
a_ptr += 32;
}
if ( m2 & 16 )
{
sgemv_kernel_16(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(16,ybuffer,y_ptr,inc_y);
y_ptr += 16 * inc_y;
a_ptr += 16;
}
if ( m2 & 8 )
{
sgemv_kernel_8(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(8,ybuffer,y_ptr,inc_y);
y_ptr += 8 * inc_y;
a_ptr += 8;
}
if ( m2 & 4 )
{
sgemv_kernel_4(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(4,ybuffer,y_ptr,inc_y);
y_ptr += 4 * inc_y;
a_ptr += 4;
}
if ( m2 & 2 )
{
sgemv_kernel_2(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(2,ybuffer,y_ptr,inc_y);
y_ptr += 2 * inc_y;
a_ptr += 2;
}
if ( m2 & 1 )
{
sgemv_kernel_1(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
add_y(1,ybuffer,y_ptr,inc_y);
}
}
return(0);
}

451
kernel/x86_64/sgemv_n_microk_bulldozer.c Normal file
View File

@ -0,0 +1,451 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
"vxorps %%ymm14, %%ymm14, %%ymm14\n\t" // set to zero
"vxorps %%ymm15, %%ymm15, %%ymm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmaddps %%ymm8 , 0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vfmaddps %%ymm9 , 8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
"prefetcht0 128(%%r8)\n\t" // Prefetch
"vfmaddps %%ymm10, 16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm11, 24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
"prefetcht0 192(%%r8)\n\t" // Prefetch
"vfmaddps %%ymm12, 32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm13, 40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm14, 48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm15, 56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmulps %%ymm14, %%ymm1, %%ymm14\n\t" // scale by alpha
"vmulps %%ymm15, %%ymm1, %%ymm15\n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm12, 32*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 40*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm14, 48*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm15, 56*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%xmm8 , %%xmm8 , %%xmm8 \n\t" // set to zero
"vxorps %%xmm9 , %%xmm9 , %%xmm9 \n\t" // set to zero
"vxorps %%xmm10, %%xmm10, %%xmm10\n\t" // set to zero
"vxorps %%xmm11, %%xmm11, %%xmm11\n\t" // set to zero
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmaddps %%xmm8 , 0*4(%%rsi), %%xmm0, %%xmm8 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vfmaddps %%xmm9 , 4*4(%%rsi), %%xmm0, %%xmm9 \n\t" // multiply a and c and add to temp
"vfmaddps %%xmm10, 8*4(%%rsi), %%xmm0, %%xmm10\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm11, 12*4(%%rsi), %%xmm0, %%xmm11\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm12, 16*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm13, 20*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm14, 24*4(%%rsi), %%xmm0, %%xmm14\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm15, 28*4(%%rsi), %%xmm0, %%xmm15\n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%xmm8 , %%xmm1, %%xmm8 \n\t" // scale by alpha
"vmulps %%xmm9 , %%xmm1, %%xmm9 \n\t" // scale by alpha
"vmulps %%xmm10, %%xmm1, %%xmm10\n\t" // scale by alpha
"vmulps %%xmm11, %%xmm1, %%xmm11\n\t" // scale by alpha
"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulps %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha
"vmulps %%xmm14, %%xmm1, %%xmm14\n\t" // scale by alpha
"vmulps %%xmm15, %%xmm1, %%xmm15\n\t" // scale by alpha
"vmovups %%xmm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%xmm9 , 4*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm10, 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm11, 12*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm12, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm13, 20*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm14, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%xmm15, 28*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
);
}
static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmaddps %%ymm12, 0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"vfmaddps %%ymm13, 8*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmovups %%ymm12, (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 8*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"vfmaddps %%ymm12, 0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmovups %%ymm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"vfmaddps %%xmm12, 0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmovups %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"vfmaddss %%xmm12, 0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
"vfmaddss %%xmm13, 1*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulss %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
"vmovss %%xmm13, 4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"vfmaddss %%xmm12, 0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

461
kernel/x86_64/sgemv_n_microk_haswell.c Normal file
View File

@ -0,0 +1,461 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*2;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
"vxorps %%ymm14, %%ymm14, %%ymm14\n\t" // set to zero
"vxorps %%ymm15, %%ymm15, %%ymm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vfmadd231ps 0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
"vfmadd231ps 8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vfmadd231ps 16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
"vfmadd231ps 24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
"prefetcht0 128(%%r8)\n\t" // Prefetch
"vfmadd231ps 32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
"vfmadd231ps 40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
"prefetcht0 192(%%r8)\n\t" // Prefetch
"vfmadd231ps 48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
"vfmadd231ps 56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmulps %%ymm14, %%ymm1, %%ymm14\n\t" // scale by alpha
"vmulps %%ymm15, %%ymm1, %%ymm15\n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm12, 32*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 40*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm14, 48*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm15, 56*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vmulps 16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);
}
static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);
}
static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);
}
static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c
"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmovups %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vmulps 1*4(%%rsi), %%xmm0, %%xmm5 \n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"vaddps %%xmm13, %%xmm5, %%xmm13 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulss %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
"vmovss %%xmm13, 4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"vmulss 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddss %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

473
kernel/x86_64/sgemv_n_microk_sandy.c Normal file
View File

@ -0,0 +1,473 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
static void sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*2;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
"vxorps %%ymm12, %%ymm12, %%ymm12\n\t" // set to zero
"vxorps %%ymm13, %%ymm13, %%ymm13\n\t" // set to zero
"vxorps %%ymm14, %%ymm14, %%ymm14\n\t" // set to zero
"vxorps %%ymm15, %%ymm15, %%ymm15\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vmulps 16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"prefetcht0 128(%%r8)\n\t" // Prefetch
"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
"prefetcht0 192(%%r8)\n\t" // Prefetch
"vmulps 32*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 40*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vmulps 48*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 56*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
"vaddps %%ymm12, %%ymm4, %%ymm12\n\t" // multiply a and c and add to temp
"vaddps %%ymm13, %%ymm5, %%ymm13\n\t" // multiply a and c and add to temp
"vaddps %%ymm14, %%ymm6, %%ymm14\n\t" // multiply a and c and add to temp
"vaddps %%ymm15, %%ymm7, %%ymm15\n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmulps %%ymm12, %%ymm1, %%ymm12\n\t" // scale by alpha
"vmulps %%ymm13, %%ymm1, %%ymm13\n\t" // scale by alpha
"vmulps %%ymm14, %%ymm1, %%ymm14\n\t" // scale by alpha
"vmulps %%ymm15, %%ymm1, %%ymm15\n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm12, 32*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm13, 40*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm14, 48*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm15, 56*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
"vxorps %%ymm10, %%ymm10, %%ymm10\n\t" // set to zero
"vxorps %%ymm11, %%ymm11, %%ymm11\n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vmulps 16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
"vmulps 24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmulps %%ymm10, %%ymm1, %%ymm10\n\t" // scale by alpha
"vmulps %%ymm11, %%ymm1, %%ymm11\n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm10, 16*4(%%rdx) \n\t" // store temp -> y
"vmovups %%ymm11, 24*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);
}
static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{
float *pre = a + lda*3;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"movq %6, %%r8\n\t" // address for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"prefetcht0 64(%%r8)\n\t" // Prefetch
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
"vxorps %%ymm9 , %%ymm9 , %%ymm9 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"nop \n\t"
"leaq (%%r8 , %%rcx, 4), %%r8 \n\t" // add lda to pointer for prefetch
"prefetcht0 (%%r8)\n\t" // Prefetch
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmulps %%ymm9 , %%ymm1, %%ymm9 \n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
"vmovups %%ymm9 , 8*4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y), // 5
"m" (pre) // 6
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);
}
static void sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%ymm1\n\t" // alpha -> ymm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%ymm8 , %%ymm8 , %%ymm8 \n\t" // set to zero
".align 16 \n\t"
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%ymm0 \n\t" // load values of c
"vmulps 0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%ymm8 , %%ymm1, %%ymm8 \n\t" // scale by alpha
"vmovups %%ymm8 , (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"memory"
);
}
static void sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vbroadcastss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vbroadcastss (%%rdi), %%xmm0 \n\t" // load values of c
"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulps %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmovups %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"vmulps 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vmulps 1*4(%%rsi), %%xmm0, %%xmm5 \n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"vaddps %%xmm13, %%xmm5, %%xmm13 \n\t" // multiply a and c and add to temp
"addq $4 , %%rdi \n\t" // increment pointer of c
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmulss %%xmm13, %%xmm1, %%xmm13\n\t" // scale by alpha
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
"vmovss %%xmm13, 4(%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}
static void sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
{
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
".L01LOOP%=: \n\t"
"vmovss (%%rdi), %%xmm0 \n\t" // load values of c
"addq $4 , %%rdi \n\t" // increment pointer of c
"vmulss 0*4(%%rsi), %%xmm0, %%xmm4 \n\t" // multiply a and c and add to temp
"vaddss %%xmm12, %%xmm4, %%xmm12 \n\t" // multiply a and c and add to temp
"leaq (%%rsi, %%rcx, 4), %%rsi \n\t" // add lda to pointer of a
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t" // scale by alpha
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

232
kernel/x86_64/sgemv_t_avx.c Normal file
View File

@ -0,0 +1,232 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
#include "common.h"
#if defined(BULLDOZER) || defined(PILEDRIVER)
#include "sgemv_t_microk_bulldozer.c"
#elif defined(HASWELL)
#include "sgemv_t_microk_haswell.c"
#else
#include "sgemv_t_microk_sandy.c"
#endif
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
{
BLASLONG i;
for ( i=0; i<n; i++ )
{
*dest = *src;
dest++;
src += inc_src;
}
}
static void sgemv_kernel_1( BLASLONG n, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, FLOAT *y)
{
FLOAT register temp0 = 0.0;
BLASLONG i;
for ( i=0; i<n ; i++)
{
temp0 += a[i] * x[i];
}
temp0 *= alpha ;
*y += temp0;
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
BLASLONG j;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *a_ptrl;
BLASLONG m1;
BLASLONG register m2;
FLOAT *xbuffer;
xbuffer = buffer;
BLASLONG register Mblock;
m1 = m / 1024 ;
m2 = m % 1024 ;
x_ptr = x;
a_ptr = a;
for (j=0; j<m1; j++)
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(1024,x_ptr,xbuffer,inc_x);
y_ptr = y;
a_ptrl = a_ptr;
for(i = 0; i<n; i++ )
{
sgemv_kernel_16(1024,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += 1024;
x_ptr += 1024 * inc_x;
}
if ( m2 == 0 ) return(0);
Mblock = 512;
while ( Mblock >= 16 )
{
if ( m2 & Mblock)
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(Mblock,x_ptr,xbuffer,inc_x);
y_ptr = y;
a_ptrl = a_ptr;
for(i = 0; i<n; i++ )
{
sgemv_kernel_16(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;
}
Mblock /= 2;
}
if ( m2 & Mblock)
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(Mblock,x_ptr,xbuffer,inc_x);
y_ptr = y;
a_ptrl = a_ptr;
for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;
}
Mblock /= 2;
if ( m2 & Mblock)
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(Mblock,x_ptr,xbuffer,inc_x);
y_ptr = y;
a_ptrl = a_ptr;
for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;
}
Mblock /= 2;
if ( m2 & Mblock)
{
if ( inc_x == 1 )
xbuffer = x_ptr;
else
copy_x(Mblock,x_ptr,xbuffer,inc_x);
y_ptr = y;
a_ptrl = a_ptr;
for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
a_ptr += Mblock;
x_ptr += Mblock * inc_x;
}
Mblock /= 2;
if ( m2 & Mblock)
{
xbuffer = x_ptr;
y_ptr = y;
a_ptrl = a_ptr;
for(i = 0; i<n; i++ )
{
sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
y_ptr += inc_y;
a_ptrl += lda;
}
}
return(0);
}

99
kernel/x86_64/sgemv_t_microk_bulldozer.c Normal file
View File

@ -0,0 +1,99 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{
//n = n / 16;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"leaq (, %%rcx,4), %%rcx \n\t" // scale lda by size of float
"leaq (%%rsi,%%rcx,1), %%r8 \n\t" // pointer to next line
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero
"sarq $4, %%rax \n\t" // n = n / 16
".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 512(%%rsi) \n\t"
"prefetcht0 (%%r8) \n\t" //prefetch next line of a
"vmovups (%%rsi), %%xmm4 \n\t"
"vmovups 4*4(%%rsi), %%xmm5 \n\t"
"vmovups 8*4(%%rsi), %%xmm6 \n\t"
"vmovups 12*4(%%rsi), %%xmm7 \n\t"
"vfmaddps %%xmm12, 0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm13, 4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm14, 8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
"vfmaddps %%xmm15, 12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp
"addq $16*4 , %%r8 \n\t" // increment prefetch pointer
"addq $16*4 , %%rsi \n\t" // increment pointer of a
"addq $16*4 , %%rdi \n\t" // increment pointer of c
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vaddps %%xmm12, %%xmm14, %%xmm12\n\t"
"vaddps %%xmm13, %%xmm15, %%xmm13\n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vfmaddss (%%rdx), %%xmm12, %%xmm1, %%xmm12\n\t"
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

100
kernel/x86_64/sgemv_t_microk_haswell.c Normal file
View File

@ -0,0 +1,100 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{
//n = n / 16;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"leaq (, %%rcx,4), %%rcx \n\t" // scale lda by size of float
"leaq (%%rsi,%%rcx,1), %%r8 \n\t" // pointer to next line
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero
"sarq $4, %%rax \n\t" // n = n / 16
".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 512(%%rsi) \n\t"
"prefetcht0 (%%r8) \n\t" //prefetch next line of a
"vmovups (%%rsi), %%xmm4 \n\t"
"vmovups 4*4(%%rsi), %%xmm5 \n\t"
"vmovups 8*4(%%rsi), %%xmm6 \n\t"
"vmovups 12*4(%%rsi), %%xmm7 \n\t"
"vfmadd231ps 0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
"vfmadd231ps 4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
"vfmadd231ps 8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
"vfmadd231ps 12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp
"addq $16*4 , %%r8 \n\t" // increment prefetch pointer
"addq $16*4 , %%rsi \n\t" // increment pointer of a
"addq $16*4 , %%rdi \n\t" // increment pointer of c
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vaddps %%xmm12, %%xmm14, %%xmm12\n\t"
"vaddps %%xmm13, %%xmm15, %%xmm13\n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12\n\t"
"vaddss (%%rdx), %%xmm12,%%xmm12\n\t"
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}

106
kernel/x86_64/sgemv_t_microk_sandy.c Normal file
View File

@ -0,0 +1,106 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
static void sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
{
//n = n / 16;
__asm__ __volatile__
(
"movq %0, %%rax\n\t" // n -> rax
"vmovss %1, %%xmm1\n\t" // alpha -> xmm1
"movq %2, %%rsi\n\t" // adress of a -> rsi
"movq %3, %%rcx\n\t" // value of lda > rcx
"movq %4, %%rdi\n\t" // adress of x -> rdi
"movq %5, %%rdx\n\t" // adress of y -> rdx
"leaq (, %%rcx,4), %%rcx \n\t" // scale lda by size of float
"leaq (%%rsi,%%rcx,1), %%r8 \n\t" // pointer to next line
"vxorps %%xmm12, %%xmm12, %%xmm12\n\t" // set to zero
"vxorps %%xmm13, %%xmm13, %%xmm13\n\t" // set to zero
"vxorps %%xmm14, %%xmm14, %%xmm14\n\t" // set to zero
"vxorps %%xmm15, %%xmm15, %%xmm15\n\t" // set to zero
"sarq $4, %%rax \n\t" // n = n / 16
".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 512(%%rsi) \n\t"
"prefetcht0 (%%r8) \n\t" //prefetch next line of a
"vmovups (%%rsi), %%xmm4 \n\t"
"vmovups 4*4(%%rsi), %%xmm5 \n\t"
"vmovups 8*4(%%rsi), %%xmm6 \n\t"
"vmovups 12*4(%%rsi), %%xmm7 \n\t"
"vmulps 0*4(%%rdi), %%xmm4, %%xmm8 \n\t" // multiply a and c and add to temp
"vmulps 4*4(%%rdi), %%xmm5, %%xmm9 \n\t" // multiply a and c and add to temp
"vmulps 8*4(%%rdi), %%xmm6, %%xmm10\n\t" // multiply a and c and add to temp
"vmulps 12*4(%%rdi), %%xmm7, %%xmm11\n\t" // multiply a and c and add to temp
"vaddps %%xmm12, %%xmm8 , %%xmm12\n\t"
"vaddps %%xmm13, %%xmm9 , %%xmm13\n\t"
"vaddps %%xmm14, %%xmm10, %%xmm14\n\t"
"vaddps %%xmm15, %%xmm11, %%xmm15\n\t"
"addq $16*4 , %%r8 \n\t" // increment prefetch pointer
"addq $16*4 , %%rsi \n\t" // increment pointer of a
"addq $16*4 , %%rdi \n\t" // increment pointer of c
"dec %%rax \n\t" // n = n -1
"jnz .L01LOOP%= \n\t"
"vaddps %%xmm12, %%xmm14, %%xmm12\n\t"
"vaddps %%xmm13, %%xmm15, %%xmm13\n\t"
"vaddps %%xmm12, %%xmm13, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vhaddps %%xmm12, %%xmm12, %%xmm12\n\t"
"vmulss %%xmm12, %%xmm1, %%xmm12 \n\t"
"vaddss (%%rdx), %%xmm12, %%xmm12\n\t"
"vmovss %%xmm12, (%%rdx) \n\t" // store temp -> y
:
:
"m" (n), // 0
"m" (alpha), // 1
"m" (a), // 2
"m" (lda), // 3
"m" (x), // 4
"m" (y) // 5
: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
"%xmm0", "%xmm1",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
}