How to debug OpenACC code with NVIDIA GPU?

Hi. First of all, I cannot use English very well. I’m sorry for that.

I have some trouble using OpenACC with NVIDIA GPU.
I tried to make a acceled code with OpenACC.
So I need to debug GPU status or values.

How can I debug OpenACC code?

I tried Nsight(in CUDA), PGI Eclipse-plugin.
But I couldn’t get debug infos from kernel region where made from OpenACC.
Can I use Nsight or PGI Eclipse-plugin for debug some OpenACC code??
If not, How can I debug that??

I really thank you for your reply.

We are working on better debug access to GPUs, but the main
way in which to debug OpenACC programs is to compare the execution
at each step between running the code on the CPU, and running the
code on OpenACC, and comparing the results.

-acc -ta=tesla,host

will create codes that run on the GPU or on the CPU, depending
on environment at runtime.

You can also compile the program without -acc, and you can run pgdbg
on that code with a great deal of control.

If we assume the CPU only code runs correctly, OpenACC directives are added incrementally to those sections of code that have high computation


You can use CUDA 7.5’s cuda debugger, but it’s experimental and doesn’t support Fortran very well.

Commercially, Totalview and Alinea DDT both support OpenACC debugging.

  • MAt