Révision 250 BLAS/xGEMM/xGEMM.c
| xGEMM.c (revision 250) | ||
|---|---|---|
| 30 | 30 |
#include "fortran_common.h" |
| 31 | 31 |
#include "fortran_thunking.h" |
| 32 | 32 |
#elif FBLAS |
| 33 |
#include <cblas.h> |
|
| 34 |
#include <cblas_f77.h> |
|
| 33 |
#include <f77blas.h> |
|
| 35 | 34 |
#elif GSL |
| 36 | 35 |
#include <gsl_cblas.h> |
| 37 | 36 |
#elif ACML |
| ... | ... | |
| 43 | 42 |
|
| 44 | 43 |
#ifdef CLBLAS |
| 45 | 44 |
|
| 46 |
#ifdef DOUBLE
|
|
| 45 |
#ifdef FP64
|
|
| 47 | 46 |
#define LENGTH cl_double |
| 48 | 47 |
#else |
| 49 | 48 |
#define LENGTH cl_float |
| ... | ... | |
| 51 | 50 |
|
| 52 | 51 |
#else |
| 53 | 52 |
|
| 54 |
#ifdef DOUBLE
|
|
| 53 |
#ifdef FP64
|
|
| 55 | 54 |
#define LENGTH double |
| 56 | 55 |
#else |
| 57 | 56 |
#define LENGTH float |
| ... | ... | |
| 59 | 58 |
|
| 60 | 59 |
#endif |
| 61 | 60 |
|
| 62 |
#ifdef FBLAS
|
|
| 61 |
/* #ifdef FBLAS */
|
|
| 63 | 62 |
|
| 64 |
#ifdef DOUBLE
|
|
| 63 |
/* #ifdef FP64 */
|
|
| 65 | 64 |
|
| 66 |
void F77_dgemm(FCHAR, FCHAR, FINT, FINT, FINT, const double *, const double *, FINT,
|
|
| 67 |
const double *, FINT, const double *, double *, FINT);
|
|
| 65 |
/* void F77_dgemm(FCHAR, FCHAR, FINT, FINT, FINT, const double *, const double *, FINT, */
|
|
| 66 |
/* const double *, FINT, const double *, double *, FINT); */
|
|
| 68 | 67 |
|
| 69 |
#else
|
|
| 68 |
/* #else */
|
|
| 70 | 69 |
|
| 71 |
void F77_sgemm(FCHAR, FCHAR, FINT, FINT, FINT, const float *, const float *, FINT,
|
|
| 72 |
const float *, FINT, const float *, float *, FINT);
|
|
| 70 |
/* void F77_sgemm(FCHAR, FCHAR, FINT, FINT, FINT, const float *, const float *, FINT, */
|
|
| 71 |
/* const float *, FINT, const float *, float *, FINT); */
|
|
| 73 | 72 |
|
| 74 |
#endif
|
|
| 75 |
#endif
|
|
| 73 |
/* #endif */
|
|
| 74 |
/* #endif */
|
|
| 76 | 75 |
|
| 77 | 76 |
/* Matrix with only defined triangular terms */ |
| 78 | 77 |
/* Even if there are 0 in matrix, must be defined at all ! */ |
| ... | ... | |
| 312 | 311 |
/* Get third timer after memory operation */ |
| 313 | 312 |
gettimeofday(&tv3, &tz); |
| 314 | 313 |
|
| 315 |
#ifdef DOUBLE
|
|
| 314 |
#ifdef FP64
|
|
| 316 | 315 |
|
| 317 | 316 |
for (i=0;i<RUNS;i++) |
| 318 | 317 |
{
|
| ... | ... | |
| 428 | 427 |
/* Get third timer after memory operation */ |
| 429 | 428 |
gettimeofday(&tv3, &tz); |
| 430 | 429 |
|
| 431 |
#ifdef DOUBLE
|
|
| 430 |
#ifdef FP64
|
|
| 432 | 431 |
|
| 433 | 432 |
for (i=0;i<RUNS;i++) |
| 434 | 433 |
{
|
| ... | ... | |
| 473 | 472 |
printf("Using CuBLAS/Thunking: %i iterations for %ix%i matrix\n",
|
| 474 | 473 |
RUNS,dim,dim); |
| 475 | 474 |
|
| 476 |
#ifdef DOUBLE
|
|
| 475 |
#ifdef FP64
|
|
| 477 | 476 |
|
| 478 | 477 |
for (i=0;i<RUNS;i++) |
| 479 | 478 |
{
|
| ... | ... | |
| 503 | 502 |
printf("Using FBLAS: %i iterations for %ix%i matrix\n",
|
| 504 | 503 |
RUNS,dim,dim); |
| 505 | 504 |
|
| 506 |
#ifdef DOUBLE
|
|
| 505 |
#ifdef FP64
|
|
| 507 | 506 |
|
| 508 | 507 |
for (i=0;i<RUNS;i++) |
| 509 | 508 |
{
|
| 510 |
F77_dgemm(&transa,&transa,&dim,&dim,&dim,&alpha,B,&dim,A,&dim,&beta,C,&dim);
|
|
| 511 |
F77_dgemm(&transb,&transb,&dim,&dim,&dim,&alpha,A,&dim,B,&dim,&beta,D,&dim);
|
|
| 509 |
dgemm_(&transa,&transa,&dim,&dim,&dim,&alpha,B,&dim,A,&dim,&beta,C,&dim);
|
|
| 510 |
dgemm_(&transb,&transb,&dim,&dim,&dim,&alpha,A,&dim,B,&dim,&beta,D,&dim);
|
|
| 512 | 511 |
} |
| 513 | 512 |
|
| 514 | 513 |
#else |
| 515 | 514 |
|
| 516 | 515 |
for (i=0;i<RUNS;i++) |
| 517 | 516 |
{
|
| 518 |
F77_sgemm(&transa,&transa,&dim,&dim,&dim,&alpha,B,&dim,A,&dim,&beta,C,&dim);
|
|
| 519 |
F77_sgemm(&transb,&transb,&dim,&dim,&dim,&alpha,A,&dim,B,&dim,&beta,D,&dim);
|
|
| 517 |
sgemm_(&transa,&transa,&dim,&dim,&dim,&alpha,B,&dim,A,&dim,&beta,C,&dim);
|
|
| 518 |
sgemm_(&transb,&transb,&dim,&dim,&dim,&alpha,A,&dim,B,&dim,&beta,D,&dim);
|
|
| 520 | 519 |
} |
| 521 | 520 |
|
| 522 | 521 |
#endif |
| ... | ... | |
| 529 | 528 |
printf("Using ACML: %i iterations for %ix%i matrix\n",
|
| 530 | 529 |
RUNS,dim,dim); |
| 531 | 530 |
|
| 532 |
#ifdef DOUBLE
|
|
| 531 |
#ifdef FP64
|
|
| 533 | 532 |
|
| 534 | 533 |
for (i=0;i<RUNS;i++) |
| 535 | 534 |
{
|
| ... | ... | |
| 558 | 557 |
NonUnit : Matrix is not unit |
| 559 | 558 |
*/ |
| 560 | 559 |
|
| 561 |
#ifdef DOUBLE
|
|
| 560 |
#ifdef FP64
|
|
| 562 | 561 |
|
| 563 | 562 |
for (i=0;i<RUNS;i++) |
| 564 | 563 |
{
|
| ... | ... | |
| 591 | 590 |
NonUnit : Matrix is not unit |
| 592 | 591 |
*/ |
| 593 | 592 |
|
| 594 |
#ifdef DOUBLE
|
|
| 593 |
#ifdef FP64
|
|
| 595 | 594 |
|
| 596 | 595 |
for (i=0;i<RUNS;i++) |
| 597 | 596 |
{
|
Formats disponibles : Unified diff