Set the CUDA stream before calling cublas.

Signed-off-by: George Bosilca <[email protected]>
ICLDisco · Aug 2, 2023 · cf33780 · cf33780
1 parent 77c8e9f
commit cf33780
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 0 deletions.
diff --git a/src/dtd_wrappers/zgemm.c b/src/dtd_wrappers/zgemm.c
@@ -82,6 +82,8 @@ parsec_core_zgemm_cuda(parsec_device_gpu_module_t* gpu_device,
 
     handles = parsec_info_get(&gpu_stream->infos, CuHI);
 
+    parsec_cuda_exec_stream_t* cuda_stream = (parsec_cuda_exec_stream_t*)gpu_stream;
+    cublasSetStream( handles->cublas_handle, cuda_stream->cuda_stream );
     status = cublasZgemm(handles->cublas_handle, transA, transB,
                           n, m, k,
                           &alphag, (cuDoubleComplex*)Ag, lda,

diff --git a/src/dtd_wrappers/zherk.c b/src/dtd_wrappers/zherk.c
@@ -71,6 +71,8 @@ parsec_core_zherk_cuda(parsec_device_gpu_module_t* gpu_device,
     }
 #endif /* defined(PARSEC_DEBUG_NOISIER) */
 
+    parsec_cuda_exec_stream_t* cuda_stream = (parsec_cuda_exec_stream_t*)gpu_stream;
+    cublasSetStream( handles->cublas_handle, cuda_stream->cuda_stream );
     status = cublasZherk(handles->cublas_handle, uplo, trans,
                           m, n,
                           &alpha, (cuDoubleComplex*)Ag, lda,

diff --git a/src/dtd_wrappers/ztrsm.c b/src/dtd_wrappers/ztrsm.c
@@ -79,6 +79,8 @@ parsec_core_ztrsm_cuda(parsec_device_gpu_module_t* gpu_device,
     }
 #endif /* defined(PARSEC_DEBUG_NOISIER) */
 
+    parsec_cuda_exec_stream_t* cuda_stream = (parsec_cuda_exec_stream_t*)gpu_stream;
+    cublasSetStream( handles->cublas_handle, cuda_stream->cuda_stream );
     status = cublasZtrsm(handles->cublas_handle,
                           side, uplo, trans, diag,
                           m, n, &alphag,