Issuing copies to host from within device code

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

Is it possible for a kernel to issue a copy (possibly DMA) to a host ptr? I guess the answer is probably no, but just want to make sure.

2

You guessed correctly. The driver code that configures the DMA engines runs on the host. So transfers between host and device must be initiated from host code.

3

If the host pointer points to pinned host memory, it is directly accessible from within the kernel.

4

But does the transfer actually happen while the kernel is in flight? Do you have some minimal example?

5

here is a minimal example: