index b35307bacc1fd4a0c28f8ff5213f6bb370886b9e..049ca5811653975b0bce644c61be20cb0356cf9e 100644 (file)
* THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
-#include "ti_cblas.h"
+#include "ti_cblas_acc.h"
+#include "../../ticblas/ticblas.h"
#ifdef __cplusplus
extern "C" {
@@ -44,20 +45,19 @@ void cblas_zgbmv(const enum CBLAS_ORDER order, const enum CBLAS_TRANSPOSE TransA
if (!ti_cblas_init_done) ti_cblas_init();
TI_CBLAS_DEBUG_PRINT("Intercepted call to %s\n", "cblas_zgbmv");
- TI_CBLAS_PROFILE_START();
-
+ TI_CBLAS_PROFILE_START();
/* Dynamic condtional offload to ARM */
- if ((TI_CBLAS_L2_OFFLOAD == TI_CBLAS_OFFLOAD_NONE)) {
- TI_CBLAS_DEBUG_PRINT("Executing ARM %s\n", "cblas_zgbmv");
- __real_cblas_zgbmv(order,TransA,M,N,KL,KU,alpha,A,lda,X,incX,beta,Y,incY);
- TI_CBLAS_PROFILE_REPORT(" Entire %s call (ARM) took %8.2f us\n","cblas_zgbmv", (float) clock_diff);
- return ;
- }
- /* End ARM offload */
+ if ((TI_CBLAS_L2_OFFLOAD == TI_CBLAS_OFFLOAD_NONE)) {
+ TI_CBLAS_DEBUG_PRINT("Executing ARM %s\n", "cblas_zgbmv");
+ __real_cblas_zgbmv(order,TransA,M,N,KL,KU,alpha,A,lda,X,incX,beta,Y,incY);
+ TI_CBLAS_PROFILE_REPORT(" Entire %s call (ARM) took %8.2f us\n","cblas_zgbmv", (float) clock_diff);
+ return ;
+ }
+ /* End ARM offload */
/******************************************************************/
/* DSP offload WILL be done if control reaches here */
- TI_CBLAS_DEBUG_PRINT("Offloading to DSP %s\n", "cblas_zgbmv");
+ TI_CBLAS_DEBUG_PRINT("Offloading to DSP %s\n", "cblas_zgbmv");
/* Lookup kernel pointer from global table */
#ifdef __cplusplus
#else
cl_kernel __K;
#endif
- __K = ti_cblas_get_kernel(TI_CBLAS_CBLAS_ZGBMV_IDX, "ocl_cblas_zgbmv");
+ __K = ti_cblas_get_kernel(TI_CBLAS_CBLAS_ZGBMV_IDX, "ocl_cblas_zgbmv");
+
#ifdef __cplusplus
try
#else
#endif
{
-
#ifdef __cplusplus
__K->setArg(0, order);
#else
@@ -206,8 +206,8 @@ void cblas_zgbmv(const enum CBLAS_ORDER order, const enum CBLAS_TRANSPOSE TransA
TI_CBLAS_OCL_CHKERROR("clWaitForEvents",err);
err |= clReleaseEvent(e);
TI_CBLAS_OCL_CHKERROR("clReleaseEvent",err);
-
#endif
+
ti_cblas_delete_kernel(__K);
TI_CBLAS_DEBUG_PRINT("Finished executing %s\n", "cblas_zgbmv");