cudaMemcpy during kernel execution asynchronous kernel launch

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I tried to copy data from host to device exactly after I started a kernel on the device to test asynchronous kernel launch functionality.

Here is the code:

...

myKernel<<<...>>>(...);

// cudaThreadSynchronize();

cudaMemcpy(....);

...

I did one run with and one without “cudaThreadSynchronize()”. I hoped to decrease the total running time of my program when I comment “cudaThreadSynchronize()” out but there was no change to the run time.

I asume that “cudaMemcpy()” does not work with asynchronous kernel launches. Am I right? I would only save time if I would compute other things on the host exactly after a asynchronous kernel launch, right?

Thanks in advance for your answers.

Best regards,

Christoph

2

There won’t be any change in time - cudaMemcpy synchronizes with the GPU. The call to cudaThreadSynchronize is redundant.

As mentioned in other posts, the current hardware doesn’t support cudaMemcpy and kernel execution at the same time.

3

Memory transactions, GL/DX resource registration, and a few other CUDA calls are blocking (Programming Guide, section 4.5.1.5).

Paulius