When using streams do you need to sync after cudamalloc-CodePudding

When using streams in cuda, is it necessary to perform any synchronization between memory allocations and usage of this memory by a stream (assuming cudaMallocAsync is not available, which it is not for me).

example:

cudaStream_t stream;
cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking);
... Other code

int *a;
gpuErrchk(cudaMalloc((void **)&a, sizeof(int)));
foo<<<1, 1, 0, stream>>>(a);
gpuErrchk(cudaStreamSynchronize(stream));
cudafree(a);

Is there a chance that when calling foo in such a situation, that the memory allocation is not completed?

Would I be forced to do a device synchronize after the allocation to be sure?

CodePudding user response：

There is no chance that the memory allocation will not be completed. You don't need an explicit device synchronization after a memory allocation.

When the cudaMalloc call returns, the memory is allocated and usable.