Unable to profile application

When i use Visual Profiler to profile cuda application, i always get error:
Unable to profile application.
Unknown nvprof return value: 143

the cuda aplication i use: /usr/local/cuda/samples/2_Graphics/Mandelbrot
ubuntu 16.04
driver: 415
cuda version 8.0

Thanks a lot for help.

Hi chenka,

We are unable to reproduce the issue at our end using same components. Can you please provide more information for further investigation?

  1. Which GPU you are using?
  2. Did you try any other application? Are seeing same error for other applications?
  3. Are you able to run standalone CUDA application (without profiling)?
  4. Can you please send whole output from Visual Profiler console (or from nvprof)?


Thanks,
Ramesh

Thanks for looking into it.

the sample i uses is usr/local/cuda/samples/2_Graphics/Mandelbrot . The other samples or other application are the same.
The GPU is GTX 1060.
I also tried another computer with same OS, different GPU GTX 1080, and it is same problem.

the output console window:
==28685== Warning: Some profiling data are not recorded. Make sure cudaProfilerStop() or cuProfilerStop() is called before application exit to flush profile data.

But even if i manually add these function in cuda sample in the render loop, it still give the same output and error.

I also had an issue that it complains about
======== Error: unified memory profiling failed.
So i disabled unified memory profiling in setting.

I can run the stand alone cuda sample without problem.

It is okay to ignore the warning:
==28685== Warning: Some profiling data are not recorded. Make sure cudaProfilerStop() or cuProfilerStop() is called before application exit to flush profile data.

  1. Do you see any profiler output (after disabling unified memory profiling)?

  2. Can you try profiling with nvprof:

nvprof --unified-memory-profiling off -o prof.out /usr/local/cuda/samples/2_Graphics/Mandelbrot
Do you get any errors or warnings?
Can you please share the profiler output file “prof.out”?

  1. Can you please try profiling some simpler sample such as “vectorAdd”?

Or you can try using a newer version of CUDA Toolkit.