Memory Migration to Host

[hubdw]

Would doing cudaMemcpyAsync or cudaMemPrefetchAsync be destructive without using synchronization?

Let’s say I do something like this:

cudaMalloc(a);
kernel<<<...>>>(a)
cudaMemcpyAsync(a, ToHost);

// then doing check on host on a
checkFunction(a);

or this:

cudaMalloc(a);
kernel<<<...>>>(a)
cudaMemPrefetchAsync(a, ToHost);

// then doing check on host on a
checkFunction(a);

I’m still a beginner in this. Trying to wrap my head around memory migration.

[striker159]

If an operation on the host depends on the result of an asynchronous operation, you need to wait until the asynchronous operation is finished.

[hubdw]

Though I tested out by creating some kernel to do some operations. I did something like this:

cudaMallocManaged(a)
kernel<<<...>>>()
cudaDeviceSynchronize()
cudaMemPrefetchAsync(a)
hostFunc(a)

hostFunc() does some checks

and it works. Could Unified Memory be the cause of this program to not crash?