0c1c903f1e
In current code, no matter what number of threads specified, all available CPU count is used when invoking OMP, which leads to very bad performance if the workload is small while all available CPUs are big. Lots of time are wasted on inter-thread sync. Fix this issue by really using the number specified by the variable 'num' from calling API. Signed-off-by: Chen, Guobing <guobing.chen@intel.com>
358 lines
10 KiB
C
358 lines
10 KiB
C
/*********************************************************************/
|
|
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
/* 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. */
|
|
/* */
|
|
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
/* AUSTIN ``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 UNIVERSITY OF TEXAS AT */
|
|
/* AUSTIN 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. */
|
|
/* */
|
|
/* The views and conclusions contained in the software and */
|
|
/* documentation are those of the authors and should not be */
|
|
/* interpreted as representing official policies, either expressed */
|
|
/* or implied, of The University of Texas at Austin. */
|
|
/*********************************************************************/
|
|
|
|
#include <stdbool.h>
|
|
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
//#include <sys/mman.h>
|
|
#include "common.h"
|
|
|
|
#ifndef USE_OPENMP
|
|
|
|
#include "blas_server.c"
|
|
|
|
#else
|
|
|
|
#ifndef OMP_SCHED
|
|
#define OMP_SCHED static
|
|
#endif
|
|
|
|
int blas_server_avail = 0;
|
|
|
|
static void * blas_thread_buffer[MAX_PARALLEL_NUMBER][MAX_CPU_NUMBER];
|
|
#ifdef HAVE_C11
|
|
static atomic_bool blas_buffer_inuse[MAX_PARALLEL_NUMBER];
|
|
#else
|
|
static _Bool blas_buffer_inuse[MAX_PARALLEL_NUMBER];
|
|
#endif
|
|
|
|
void goto_set_num_threads(int num_threads) {
|
|
|
|
int i=0, j=0;
|
|
|
|
if (num_threads < 1) num_threads = blas_num_threads;
|
|
|
|
if (num_threads > MAX_CPU_NUMBER) num_threads = MAX_CPU_NUMBER;
|
|
|
|
if (num_threads > blas_num_threads) {
|
|
blas_num_threads = num_threads;
|
|
}
|
|
|
|
blas_cpu_number = num_threads;
|
|
|
|
omp_set_num_threads(blas_cpu_number);
|
|
|
|
//adjust buffer for each thread
|
|
for(i=0; i<MAX_PARALLEL_NUMBER; i++) {
|
|
for(j=0; j<blas_cpu_number; j++){
|
|
if(blas_thread_buffer[i][j]==NULL){
|
|
blas_thread_buffer[i][j]=blas_memory_alloc(2);
|
|
}
|
|
}
|
|
for(; j<MAX_CPU_NUMBER; j++){
|
|
if(blas_thread_buffer[i][j]!=NULL){
|
|
blas_memory_free(blas_thread_buffer[i][j]);
|
|
blas_thread_buffer[i][j]=NULL;
|
|
}
|
|
}
|
|
}
|
|
#if defined(ARCH_MIPS64)
|
|
//set parameters for different number of threads.
|
|
blas_set_parameter();
|
|
#endif
|
|
|
|
}
|
|
void openblas_set_num_threads(int num_threads) {
|
|
|
|
goto_set_num_threads(num_threads);
|
|
}
|
|
|
|
int blas_thread_init(void){
|
|
|
|
int i=0, j=0;
|
|
|
|
blas_get_cpu_number();
|
|
|
|
blas_server_avail = 1;
|
|
|
|
for(i=0; i<MAX_PARALLEL_NUMBER; i++) {
|
|
for(j=0; j<blas_num_threads; j++){
|
|
blas_thread_buffer[i][j]=blas_memory_alloc(2);
|
|
}
|
|
for(; j<MAX_CPU_NUMBER; j++){
|
|
blas_thread_buffer[i][j]=NULL;
|
|
}
|
|
}
|
|
|
|
return 0;
|
|
}
|
|
|
|
int BLASFUNC(blas_thread_shutdown)(void){
|
|
int i=0, j=0;
|
|
blas_server_avail = 0;
|
|
|
|
for(i=0; i<MAX_PARALLEL_NUMBER; i++) {
|
|
for(j=0; j<MAX_CPU_NUMBER; j++){
|
|
if(blas_thread_buffer[i][j]!=NULL){
|
|
blas_memory_free(blas_thread_buffer[i][j]);
|
|
blas_thread_buffer[i][j]=NULL;
|
|
}
|
|
}
|
|
}
|
|
|
|
return 0;
|
|
}
|
|
|
|
static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
|
|
|
|
if (!(mode & BLAS_COMPLEX)){
|
|
#ifdef EXPRECISION
|
|
if (mode & BLAS_XDOUBLE){
|
|
/* REAL / Extended Double */
|
|
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, xdouble,
|
|
xdouble *, BLASLONG, xdouble *, BLASLONG,
|
|
xdouble *, BLASLONG, void *) = func;
|
|
|
|
afunc(args -> m, args -> n, args -> k,
|
|
((xdouble *)args -> alpha)[0],
|
|
args -> a, args -> lda,
|
|
args -> b, args -> ldb,
|
|
args -> c, args -> ldc, sb);
|
|
} else
|
|
#endif
|
|
if (mode & BLAS_DOUBLE){
|
|
/* REAL / Double */
|
|
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double,
|
|
double *, BLASLONG, double *, BLASLONG,
|
|
double *, BLASLONG, void *) = func;
|
|
|
|
afunc(args -> m, args -> n, args -> k,
|
|
((double *)args -> alpha)[0],
|
|
args -> a, args -> lda,
|
|
args -> b, args -> ldb,
|
|
args -> c, args -> ldc, sb);
|
|
} else {
|
|
/* REAL / Single */
|
|
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float,
|
|
float *, BLASLONG, float *, BLASLONG,
|
|
float *, BLASLONG, void *) = func;
|
|
|
|
afunc(args -> m, args -> n, args -> k,
|
|
((float *)args -> alpha)[0],
|
|
args -> a, args -> lda,
|
|
args -> b, args -> ldb,
|
|
args -> c, args -> ldc, sb);
|
|
}
|
|
} else {
|
|
#ifdef EXPRECISION
|
|
if (mode & BLAS_XDOUBLE){
|
|
/* COMPLEX / Extended Double */
|
|
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble,
|
|
xdouble *, BLASLONG, xdouble *, BLASLONG,
|
|
xdouble *, BLASLONG, void *) = func;
|
|
|
|
afunc(args -> m, args -> n, args -> k,
|
|
((xdouble *)args -> alpha)[0],
|
|
((xdouble *)args -> alpha)[1],
|
|
args -> a, args -> lda,
|
|
args -> b, args -> ldb,
|
|
args -> c, args -> ldc, sb);
|
|
} else
|
|
#endif
|
|
if (mode & BLAS_DOUBLE){
|
|
/* COMPLEX / Double */
|
|
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, double,
|
|
double *, BLASLONG, double *, BLASLONG,
|
|
double *, BLASLONG, void *) = func;
|
|
|
|
afunc(args -> m, args -> n, args -> k,
|
|
((double *)args -> alpha)[0],
|
|
((double *)args -> alpha)[1],
|
|
args -> a, args -> lda,
|
|
args -> b, args -> ldb,
|
|
args -> c, args -> ldc, sb);
|
|
} else {
|
|
/* COMPLEX / Single */
|
|
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, float,
|
|
float *, BLASLONG, float *, BLASLONG,
|
|
float *, BLASLONG, void *) = func;
|
|
|
|
afunc(args -> m, args -> n, args -> k,
|
|
((float *)args -> alpha)[0],
|
|
((float *)args -> alpha)[1],
|
|
args -> a, args -> lda,
|
|
args -> b, args -> ldb,
|
|
args -> c, args -> ldc, sb);
|
|
}
|
|
}
|
|
}
|
|
|
|
static void exec_threads(blas_queue_t *queue, int buf_index){
|
|
|
|
void *buffer, *sa, *sb;
|
|
int pos=0, release_flag=0;
|
|
|
|
buffer = NULL;
|
|
sa = queue -> sa;
|
|
sb = queue -> sb;
|
|
|
|
#ifdef CONSISTENT_FPCSR
|
|
__asm__ __volatile__ ("ldmxcsr %0" : : "m" (queue -> sse_mode));
|
|
__asm__ __volatile__ ("fldcw %0" : : "m" (queue -> x87_mode));
|
|
#endif
|
|
|
|
if ((sa == NULL) && (sb == NULL) && ((queue -> mode & BLAS_PTHREAD) == 0)) {
|
|
|
|
pos = omp_get_thread_num();
|
|
buffer = blas_thread_buffer[buf_index][pos];
|
|
|
|
//fallback
|
|
if(buffer==NULL) {
|
|
buffer = blas_memory_alloc(2);
|
|
release_flag=1;
|
|
}
|
|
|
|
if (sa == NULL) {
|
|
sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A);
|
|
queue->sa=sa;
|
|
}
|
|
|
|
if (sb == NULL) {
|
|
if (!(queue -> mode & BLAS_COMPLEX)){
|
|
#ifdef EXPRECISION
|
|
if (queue -> mode & BLAS_XDOUBLE){
|
|
sb = (void *)(((BLASLONG)sa + ((QGEMM_P * QGEMM_Q * sizeof(xdouble)
|
|
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
} else
|
|
#endif
|
|
if (queue -> mode & BLAS_DOUBLE){
|
|
sb = (void *)(((BLASLONG)sa + ((DGEMM_P * DGEMM_Q * sizeof(double)
|
|
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
|
|
} else {
|
|
sb = (void *)(((BLASLONG)sa + ((SGEMM_P * SGEMM_Q * sizeof(float)
|
|
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
}
|
|
} else {
|
|
#ifdef EXPRECISION
|
|
if (queue -> mode & BLAS_XDOUBLE){
|
|
sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * 2 * sizeof(xdouble)
|
|
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
} else
|
|
#endif
|
|
if (queue -> mode & BLAS_DOUBLE){
|
|
sb = (void *)(((BLASLONG)sa + ((ZGEMM_P * ZGEMM_Q * 2 * sizeof(double)
|
|
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
} else {
|
|
sb = (void *)(((BLASLONG)sa + ((CGEMM_P * CGEMM_Q * 2 * sizeof(float)
|
|
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
|
|
}
|
|
}
|
|
queue->sb=sb;
|
|
}
|
|
}
|
|
|
|
if (queue -> mode & BLAS_LEGACY) {
|
|
legacy_exec(queue -> routine, queue -> mode, queue -> args, sb);
|
|
} else
|
|
if (queue -> mode & BLAS_PTHREAD) {
|
|
void (*pthreadcompat)(void *) = queue -> routine;
|
|
(pthreadcompat)(queue -> args);
|
|
|
|
} else {
|
|
int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine;
|
|
|
|
(routine)(queue -> args, queue -> range_m, queue -> range_n, sa, sb, queue -> position);
|
|
|
|
}
|
|
|
|
if (release_flag) blas_memory_free(buffer);
|
|
|
|
}
|
|
|
|
int exec_blas(BLASLONG num, blas_queue_t *queue){
|
|
|
|
BLASLONG i, buf_index;
|
|
|
|
if ((num <= 0) || (queue == NULL)) return 0;
|
|
|
|
#ifdef CONSISTENT_FPCSR
|
|
for (i = 0; i < num; i ++) {
|
|
__asm__ __volatile__ ("fnstcw %0" : "=m" (queue[i].x87_mode));
|
|
__asm__ __volatile__ ("stmxcsr %0" : "=m" (queue[i].sse_mode));
|
|
}
|
|
#endif
|
|
|
|
while(true) {
|
|
for(i=0; i < MAX_PARALLEL_NUMBER; i++) {
|
|
#ifdef HAVE_C11
|
|
_Bool inuse = false;
|
|
if(atomic_compare_exchange_weak(&blas_buffer_inuse[i], &inuse, true)) {
|
|
#else
|
|
if(blas_buffer_inuse[i] == false) {
|
|
blas_buffer_inuse[i] = true;
|
|
#endif
|
|
buf_index = i;
|
|
break;
|
|
}
|
|
}
|
|
if(i != MAX_PARALLEL_NUMBER)
|
|
break;
|
|
}
|
|
|
|
#pragma omp parallel for num_threads(num) schedule(OMP_SCHED)
|
|
for (i = 0; i < num; i ++) {
|
|
|
|
#ifndef USE_SIMPLE_THREADED_LEVEL3
|
|
queue[i].position = i;
|
|
#endif
|
|
|
|
exec_threads(&queue[i], buf_index);
|
|
}
|
|
|
|
#ifdef HAVE_C11
|
|
atomic_store(&blas_buffer_inuse[buf_index], false);
|
|
#else
|
|
blas_buffer_inuse[buf_index] = false;
|
|
#endif
|
|
|
|
return 0;
|
|
}
|
|
|
|
#endif
|