Is there any efficient way to debug the kernel function of CUDA?

Question

is there any good way the debug the kernel function when programming on GPU?

The way I use is using printf(...) or copy the result back to CPU and then look on them.

I find it is too time-consuming to debug.

I appreciate your advice. Thank you.

Answer 1

If you're on Windows, be sure to try NVIDIA Parallel Nsight (official debug tool from Nvidia)

Answer 2

What about using a debugger?

If you are on windows you can use the VisualStudio Debugger . If you are on linux (better) you can use cuda-gdb.

The typical way to create a proper kernel is:

first implement the straightforward algorithm for the cpu only. then start with the simplest possible way of creating your kernel, upload/download data and compare the result of this first simple kernel to the results of the correct cpu kernel.

Then start add complexity and create new version of the kernel.