Disable starpu_malloc_on_node when CUDA does not support peer access

src/datawizard/malloc.c

@@ -404,7 +404,14 @@ _starpu_malloc_on_node(unsigned dst_node, size_t size)
 			STARPU_ASSERT(last[dst_node] >= addr);
 			STARPU_PTHREAD_MUTEX_UNLOCK(&cuda_alloc_mutex);
 #else
-			starpu_cuda_set_device(_starpu_memory_node_get_devid(dst_node));
+			struct _starpu_worker *worker = _starpu_get_local_worker_key();
+			unsigned devid = _starpu_memory_node_get_devid(dst_node);
+			if (!worker || worker->arch != STARPU_CUDA_WORKER || worker->devid != devid)
+#if defined(HAVE_CUDA_MEMCPY_PEER)
+				starpu_cuda_set_device(devid);
+#else
+				STARPU_ASSERT_MSG(0, "CUDA peer access is not available with this version of CUDA");
+#endif
 			status = cudaMalloc((void **)&addr, size);
 			if (!addr || (status != cudaSuccess))
 			{
@@ -506,7 +513,14 @@ _starpu_free_on_node(unsigned dst_node, uintptr_t addr, size_t size)
 			STARPU_PTHREAD_MUTEX_UNLOCK(&cuda_alloc_mutex);
 #else
 			cudaError_t err;
-			starpu_cuda_set_device(_starpu_memory_node_get_devid(dst_node));
+			struct _starpu_worker *worker = _starpu_get_local_worker_key();
+			unsigned devid = _starpu_memory_node_get_devid(dst_node);
+			if (!worker || worker->arch != STARPU_CUDA_WORKER || worker->devid != devid)
+#if defined(HAVE_CUDA_MEMCPY_PEER)
+				starpu_cuda_set_device(devid);
+#else
+				STARPU_ASSERT_MSG(0, "CUDA peer access is not available with this version of CUDA");
+#endif
 			err = cudaFree((void*)addr);
 			if (STARPU_UNLIKELY(err != cudaSuccess))
 				STARPU_CUDA_REPORT_ERROR(err);