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

Merge pull request #564 from wernsaar/develop 

Use only 1 thread in trsm if m or n < 2*GEMM_MULTITHREAD_THRESHOLD
This commit is contained in:
wernsaar 2015-05-06 11:10:31 +02:00
parent e6f57db846 f8f2e261fe
commit e52d36450a
2 changed files with 39 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
benchmark/trsm.c
View File

@ -130,11 +130,21 @@ int main(int argc, char *argv[]){
char trans='N';
char diag ='U';
int l;
int loops = 1;
double timeg;
if ((p = getenv("OPENBLAS_SIDE"))) side=*p;
if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p;
if ((p = getenv("OPENBLAS_TRANS"))) trans=*p;
if ((p = getenv("OPENBLAS_DIAG"))) diag=*p;
p = getenv("OPENBLAS_LOOPS");
if ( p != NULL )
loops = atoi(p);
blasint m, i, j;
int from = 1;
@ -150,7 +160,7 @@ int main(int argc, char *argv[]){
if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++;}
if (argc > 0) { step = atol(*argv); argc--; argv++;}
fprintf(stderr, "From : %3d To : %3d Step = %3d Side = %c Uplo = %c Trans = %c Diag = %c\n", from, to, step,side,uplo,trans,diag);
fprintf(stderr, "From : %3d To : %3d Step = %3d Side = %c Uplo = %c Trans = %c Diag = %c Loops = %d\n", from, to, step,side,uplo,trans,diag,loops);
if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
fprintf(stderr,"Out of Memory!!\n");exit(1);
@ -171,8 +181,14 @@ int main(int argc, char *argv[]){
for(m = from; m <= to; m += step)
{
timeg=0.0;
fprintf(stderr, " %6d : ", (int)m);
for (l=0; l<loops; l++)
{
for(j = 0; j < m; j++){
for(i = 0; i < m * COMPSIZE; i++){
a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
@ -188,11 +204,12 @@ int main(int argc, char *argv[]){
time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
gettimeofday( &start, (struct timezone *)0);
timeg += time1;
}
fprintf(stderr,
" %10.2f MFlops\n",
COMPSIZE * COMPSIZE * 1. * (double)m * (double)m * (double)m / time1 * 1.e-6);
time1 = timeg/loops;
fprintf(stderr, " %10.2f MFlops\n", COMPSIZE * COMPSIZE * 1. * (double)m * (double)m * (double)m / time1 * 1.e-6);
}

6
interface/trsm.c
View File

@ -362,6 +362,12 @@ void CNAME(enum CBLAS_ORDER order,
mode |= (side << BLAS_RSIDE_SHIFT);
args.nthreads = num_cpu_avail(3);
if ( args.m < 2*GEMM_MULTITHREAD_THRESHOLD )
args.nthreads = 1;
else
if ( args.n < 2*GEMM_MULTITHREAD_THRESHOLD )
args.nthreads = 1;
if (args.nthreads == 1) {
#endif