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

In OpenMP threading, preallocate the thread buffer instead of allocating the buffer every time. This patch improved the performance slightly.

This commit is contained in:
Zhang Xianyi 2013-03-01 14:36:47 +08:00
parent 3cc6ae793e
commit d744c9590a
1 changed files with 45 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

44
driver/others/blas_server_omp.c
View File

@ -49,8 +49,12 @@
int blas_server_avail = 0; int blas_server_avail = 0;
static void * blas_thread_buffer[MAX_CPU_NUMBER];
void goto_set_num_threads(int num_threads) { void goto_set_num_threads(int num_threads) {
int i=0;
if (num_threads < 1) num_threads = blas_num_threads; if (num_threads < 1) num_threads = blas_num_threads;
if (num_threads > MAX_CPU_NUMBER) num_threads = MAX_CPU_NUMBER; if (num_threads > MAX_CPU_NUMBER) num_threads = MAX_CPU_NUMBER;
@ -63,6 +67,18 @@ void goto_set_num_threads(int num_threads) {
omp_set_num_threads(blas_cpu_number); omp_set_num_threads(blas_cpu_number);
//adjust buffer for each thread
for(i=0; i<blas_cpu_number; i++){
if(blas_thread_buffer[i]==NULL){
blas_thread_buffer[i]=blas_memory_alloc(2);
}
}
for(; i<MAX_CPU_NUMBER; i++){
if(blas_thread_buffer[i]!=NULL){
blas_memory_free(blas_thread_buffer[i]);
blas_thread_buffer[i]=NULL;
}
}
#if defined(ARCH_MIPS64) #if defined(ARCH_MIPS64)
//set parameters for different number of threads. //set parameters for different number of threads.
blas_set_parameter(); blas_set_parameter();
@ -76,17 +92,33 @@ void openblas_set_num_threads(int num_threads) {
int blas_thread_init(void){ int blas_thread_init(void){
int i=0;
blas_get_cpu_number(); blas_get_cpu_number();
blas_server_avail = 1; blas_server_avail = 1;
for(i=0; i<blas_num_threads; i++){
blas_thread_buffer[i]=blas_memory_alloc(2);
}
for(; i<MAX_CPU_NUMBER; i++){
blas_thread_buffer[i]=NULL;
}
return 0; return 0;
} }
int BLASFUNC(blas_thread_shutdown)(void){ int BLASFUNC(blas_thread_shutdown)(void){
int i=0;
blas_server_avail = 0; blas_server_avail = 0;
for(i=0; i<MAX_CPU_NUMBER; i++){
if(blas_thread_buffer[i]!=NULL){
blas_memory_free(blas_thread_buffer[i]);
blas_thread_buffer[i]=NULL;
}
}
return 0; return 0;
} }
@ -177,6 +209,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
static void exec_threads(blas_queue_t *queue){ static void exec_threads(blas_queue_t *queue){
void *buffer, *sa, *sb; void *buffer, *sa, *sb;
int pos=0, release_flag=0;
buffer = NULL; buffer = NULL;
sa = queue -> sa; sa = queue -> sa;
@ -189,7 +222,14 @@ static void exec_threads(blas_queue_t *queue){
if ((sa == NULL) && (sb == NULL) && ((queue -> mode & BLAS_PTHREAD) == 0)) { if ((sa == NULL) && (sb == NULL) && ((queue -> mode & BLAS_PTHREAD) == 0)) {
pos = omp_get_thread_num();
buffer = blas_thread_buffer[pos];
//fallback
if(buffer==NULL) {
buffer = blas_memory_alloc(2); buffer = blas_memory_alloc(2);
release_flag=1;
}
if (sa == NULL) sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A); if (sa == NULL) sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A);
@ -242,7 +282,7 @@ static void exec_threads(blas_queue_t *queue){
} }
if (buffer != NULL) blas_memory_free(buffer); if (release_flag) blas_memory_free(buffer);
} }