Setting the kernel is done in the codelet

cudablas/compute/cuda_zgemm.c

@@ -34,10 +34,6 @@ int CUDA_zgemm(MORSE_enum transa, MORSE_enum transb,
                cuDoubleComplex *C, int ldc,
                CUBLAS_STREAM_PARAM)
 {
-
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
     cublasZgemm(CUBLAS_HANDLE
                 morse_lapack_const(transa), morse_lapack_const(transb),
                 m, n, k,
@@ -46,6 +42,5 @@ int CUDA_zgemm(MORSE_enum transa, MORSE_enum transb,
                 CUBLAS_VALUE(beta),  C, ldc);
 
     assert( CUBLAS_STATUS_SUCCESS == cublasGetError() );
-
     return MORSE_SUCCESS;
 }

cudablas/compute/cuda_zhemm.c

@@ -34,10 +34,6 @@ int CUDA_zhemm(MORSE_enum side, MORSE_enum uplo,
                cuDoubleComplex *C, int ldc,
                CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZhemm(CUBLAS_HANDLE
                 morse_lapack_const(side), morse_lapack_const(uplo),
                 m, n,
@@ -46,6 +42,5 @@ int CUDA_zhemm(MORSE_enum side, MORSE_enum uplo,
                 CUBLAS_VALUE(beta),  C, ldc);
 
     assert( CUBLAS_STATUS_SUCCESS == cublasGetError() );
-
     return MORSE_SUCCESS;
 }

cudablas/compute/cuda_zher2k.c

@@ -34,10 +34,6 @@ int CUDA_zher2k(MORSE_enum uplo, MORSE_enum trans,
                 cuDoubleComplex *C, int ldc,
                 CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZher2k(CUBLAS_HANDLE
                  morse_lapack_const(uplo), morse_lapack_const(trans),
                  n, k,

cudablas/compute/cuda_zherk.c

@@ -33,10 +33,6 @@ int CUDA_zherk( MORSE_enum uplo, MORSE_enum trans,
                 cuDoubleComplex *B, int ldb,
                 CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZherk(
         morse_lapack_const(uplo), morse_lapack_const(trans),
         n, k,

cudablas/compute/cuda_zlarfb.c

@@ -49,10 +49,6 @@ CUDA_zlarfb(MORSE_enum side, MORSE_enum trans,
 
     MORSE_enum transT, uplo, notransV, transV;
 
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     /* Check input arguments */
     if ((side != MorseLeft) && (side != MorseRight)) {
         return -1;

cudablas/compute/cuda_zsymm.c

@@ -34,10 +34,6 @@ int CUDA_zsymm(MORSE_enum side, MORSE_enum uplo,
                cuDoubleComplex *C, int ldc,
                CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZsymm(CUBLAS_HANDLE
                 morse_lapack_const(side), morse_lapack_const(uplo),
                 m, n,

cudablas/compute/cuda_zsyr2k.c

@@ -35,10 +35,6 @@ int CUDA_zsyr2k(
         cuDoubleComplex *C, int ldc,
         CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZsyr2k(CUBLAS_HANDLE
                  morse_lapack_const(uplo), morse_lapack_const(trans),
                  n, k,

cudablas/compute/cuda_zsyrk.c

@@ -33,10 +33,6 @@ int CUDA_zsyrk(MORSE_enum uplo, MORSE_enum trans,
                cuDoubleComplex *C, int ldc,
                CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZsyrk(CUBLAS_HANDLE
                 morse_lapack_const(uplo), morse_lapack_const(trans),
                 n, k,

cudablas/compute/cuda_ztrmm.c

@@ -34,10 +34,6 @@ int CUDA_ztrmm(
         cuDoubleComplex *B, int ldb,
         CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZtrmm(CUBLAS_HANDLE
         morse_lapack_const(side), morse_lapack_const(uplo),
         morse_lapack_const(transa), morse_lapack_const(diag),

cudablas/compute/cuda_ztrsm.c

@@ -33,10 +33,6 @@ int CUDA_ztrsm(MORSE_enum side, MORSE_enum uplo,
                cuDoubleComplex *B, int ldb,
                CUBLAS_STREAM_PARAM)
 {
-#if !defined(CHAMELEON_USE_CUBLAS_V2)
-    cublasSetKernelStream( stream );
-#endif
-
     cublasZtrsm(CUBLAS_HANDLE
         morse_lapack_const(side), morse_lapack_const(uplo),
         morse_lapack_const(transa), morse_lapack_const(diag),