Why the cuda kernel and copy do not overlap?

[Robert_Crovella]

I think some overlap between these two:

@CudaCodec16K

    cudaMemcpyAsync(hostData1, devData1, numElements * sizeof(int), cudaMemcpyDeviceToHost, stream1);

    kernel << <1, 1024, 0, stream2 >> > (devData2, 2, iterations);

should be possible. When in a WDDM setting, it’s possible that WDDM is causing issues. I sometimes suggest that people try both settings of Hardware Accelerated GPU Scheduling setting, to see if either setting results in observing the desired overlap. You can simply take a google search of that term (Hardware Accelerated GPU Scheduling), take the first blog hit from Microsoft, and use that to guide your study.