Guard for prefetchAsync GPU capability. #470 #451 #477

TimDettmers · TimDettmers · commit 7be5f2c7b353 · 2023-07-16T21:12:03.000-07:00
diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
@@ -385,6 +385,11 @@ extern "C"
 
 	void cprefetch(void *ptr, size_t bytes, int device)
 	{
+
+		int hasPrefetch = 0;
+		CUDA_CHECK_RETURN(cudaDeviceGetAttribute(&hasPrefetch, cudaDevAttrConcurrentManagedAccess, device)); // 40ns overhead
+		if (hasPrefetch == 0) return;
+ 
 		CUDA_CHECK_RETURN(cudaMemPrefetchAsync(ptr, bytes, device, 0));
 		CUDA_CHECK_RETURN(cudaPeekAtLastError());
 	}

Original file line number	Diff line number	Diff line change
`@@ -385,6 +385,11 @@ extern "C"`
`385`	`385`
`386`	`386`	`void cprefetch(void *ptr, size_t bytes, int device)`
`387`	`387`	`{`
	`388`	`+`
	`389`	`+ int hasPrefetch = 0;`
	`390`	`+ CUDA_CHECK_RETURN(cudaDeviceGetAttribute(&hasPrefetch, cudaDevAttrConcurrentManagedAccess, device)); // 40ns overhead`
	`391`	`+ if (hasPrefetch == 0) return;`
	`392`	`+`
`388`	`393`	`CUDA_CHECK_RETURN(cudaMemPrefetchAsync(ptr, bytes, device, 0));`
`389`	`394`	`CUDA_CHECK_RETURN(cudaPeekAtLastError());`
`390`	`395`	`}`