cudaMemcpy thread safe

Accelerated Computing CUDA CUDA Programming and Performance
1

I am trying to use cudaMemcpy in a multithreaded environment. Consider the case where there is two executing threads, with no shared resources:

two different arrays allocated on GPU with cudaMalloc (no shared memory): gpuPtr1, gpuPtr2
two different arrays allocated on CPU with same size (no shared memory): cpuPtr1, cpuPtr2.

Is the below code thread safe:?

if (threadId == 0)
ASSERT_CUDA_API_SUCCESS(cudaMemcpy(cpuPtr1, gpuPtr1, size, cudaMemcpyDeviceToHost));
else if (threadId == 1)
ASSERT_CUDA_API_SUCCESS(cudaMemcpy(cpuPtr2, gpuPtr2, size, cudaMemcpyDeviceToHost));

2

The entire CUDA runtime API is threadsafe except where there is a stated exception. For example certain graph functions may be listed as not thread safe.

There shouldn’t be any issues with what you have shown.