How to asynchronously copy memory from the host to the device using thrust and CUDA streams

Question

I would like to copy memory from the host to the device using thrust as in

thrust::host_vector h_vec(1 << 28);
thrust::device_vector

        

Answer 1

Here's a worked example using thrust::cuda::experimental::pinned_allocator:

// Compile with:
// nvcc --std=c++11 mem_async.cu -o mem_async

#include 
#include 
#include 

#include 
#include 
#include 
#include 

#define LEN 1024

int main(int argc, char *argv[]) {
    thrust::host_vector> h_vec(LEN);
    thrust::device_vector d_vec(LEN);

    thrust::fill(d_vec.begin(), d_vec.end(), -1.0);

    cudaMemcpyAsync(thrust::raw_pointer_cast(h_vec.data()),
                    thrust::raw_pointer_cast(d_vec.data()),
                    d_vec.size()*sizeof(float),
                    cudaMemcpyDeviceToHost);

    // Comment out this line to see what happens.
    cudaDeviceSynchronize();

    std::cout << h_vec[0] << std::endl;
}

Comment out the synchronize step and you should get 0 printed to the console due to the async memory transfer.