Remove superfluous cudaThreadSynchronize() calls.

src/datawizard/interfaces/bcsr_interface.c

@@ -437,8 +437,6 @@ static int copy_cuda_common(void *src_interface, unsigned src_node __attribute__
 	if (STARPU_UNLIKELY(cures))
 		STARPU_CUDA_REPORT_ERROR(cures);
 
-	cudaThreadSynchronize();
-
 	STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*r*c*elemsize + (nnz+nrow+1)*sizeof(uint32_t));
 
 	return 0;

src/datawizard/interfaces/block_interface.c

@@ -429,8 +429,6 @@ static int copy_cuda_common(void *src_interface, unsigned src_node __attribute__
 		}
 	}
 
-	cudaThreadSynchronize();
-
 	STARPU_TRACE_DATA_COPY(src_node, dst_node, src_block->nx*src_block->ny*src_block->elemsize*src_block->elemsize);
 
 	return 0;
@@ -465,7 +463,6 @@ static int copy_cuda_async_common(void *src_interface, unsigned src_node __attri
 					nx*ny*nz*elemsize, kind);
 				if (STARPU_UNLIKELY(cures))
 					STARPU_CUDA_REPORT_ERROR(cures);
-				cudaThreadSynchronize();
 
 				ret = 0;
 			}
@@ -486,7 +483,6 @@ static int copy_cuda_async_common(void *src_interface, unsigned src_node __attri
 						nx*ny*elemsize, nz, kind);
 				if (STARPU_UNLIKELY(cures))
 					STARPU_CUDA_REPORT_ERROR(cures);
-				cudaThreadSynchronize();
 
 				ret = 0;
 			}
@@ -539,7 +535,6 @@ no_async_default:
 		if (STARPU_UNLIKELY(cures))
 			STARPU_CUDA_REPORT_ERROR(cures);
 	}
-	cudaThreadSynchronize();
 
 	STARPU_TRACE_DATA_COPY(src_node, dst_node, src_block->nx*src_block->ny*src_block->nz*src_block->elemsize);
 	return 0;

src/datawizard/interfaces/csr_interface.c

@@ -401,8 +401,6 @@ static int copy_cuda_common(void *src_interface, unsigned src_node __attribute__
 	if (STARPU_UNLIKELY(cures))
 		STARPU_CUDA_REPORT_ERROR(cures);
 
-	cudaThreadSynchronize();
-
 	STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*elemsize + (nnz+nrow+1)*sizeof(uint32_t));
 
 	return 0;

src/datawizard/interfaces/matrix_interface.c

@@ -406,11 +406,6 @@ static int copy_cuda_to_ram_async(void *src_interface, unsigned src_node __attri
 		if (STARPU_UNLIKELY(cures))
 			STARPU_CUDA_REPORT_ERROR(cures);
 
-		cures = cudaThreadSynchronize();
-		if (STARPU_UNLIKELY(cures))
-			STARPU_CUDA_REPORT_ERROR(cures);
-		
-
 		return 0;
 	}
 
@@ -436,7 +431,6 @@ static int copy_ram_to_cuda_async(void *src_interface, unsigned src_node __attri
 		cures = cudaMemcpy2D((char *)dst_matrix->ptr, dst_matrix->ld*elemsize,
 				(char *)src_matrix->ptr, src_matrix->ld*elemsize,
 				src_matrix->nx*elemsize, src_matrix->ny, cudaMemcpyHostToDevice);
-		cudaThreadSynchronize();
 
 		if (STARPU_UNLIKELY(cures))
 			STARPU_CUDA_REPORT_ERROR(cures);

src/datawizard/interfaces/variable_interface.c

@@ -273,7 +273,6 @@ static int copy_cuda_common(void *src_interface, unsigned src_node __attribute__
 
 	cudaError_t cures;
 	cures = cudaMemcpy((char *)dst_variable->ptr, (char *)src_variable->ptr, src_variable->elemsize, kind);
-	cudaThreadSynchronize();
 
 	if (STARPU_UNLIKELY(cures))
 		STARPU_CUDA_REPORT_ERROR(cures);
@@ -315,7 +314,6 @@ static int copy_cuda_async_common(void *src_interface, unsigned src_node __attri
 	{
 		/* do it in a synchronous fashion */
 		cures = cudaMemcpy((char *)dst_variable->ptr, (char *)src_variable->ptr, src_variable->elemsize, kind);
-		cudaThreadSynchronize();
 
 		if (STARPU_UNLIKELY(cures))
 			STARPU_CUDA_REPORT_ERROR(cures);

src/datawizard/interfaces/vector_interface.c

@@ -308,8 +308,6 @@ static int copy_cuda_common(void *src_interface, unsigned src_node __attribute__
 
 	cudaError_t cures;
 	cures = cudaMemcpy((char *)dst_vector->ptr, (char *)src_vector->ptr, src_vector->nx*src_vector->elemsize, kind);
-	cudaThreadSynchronize();
-
 	if (STARPU_UNLIKELY(cures))
 		STARPU_CUDA_REPORT_ERROR(cures);
 
@@ -350,8 +348,6 @@ static int copy_cuda_async_common(void *src_interface, unsigned src_node __attri
 	{
 		/* do it in a synchronous fashion */
 		cures = cudaMemcpy((char *)dst_vector->ptr, (char *)src_vector->ptr, src_vector->nx*src_vector->elemsize, kind);
-		cudaThreadSynchronize();
-
 		if (STARPU_UNLIKELY(cures))
 			STARPU_CUDA_REPORT_ERROR(cures);