How to profile OpenCL application with CUDA 8.0 nvprof

9
votes

I'm trying to profile OpenCL application, a.out, in a system with NVIDIA TITAN X and CUDA 8.0.

If it was CUDA application, nvprof ./a.out would be enough. But I found this does not work with OpenCL application, with a message "No kernels were profiled."

Until CUDA 7.5, I successfully used COMPUTE_PROFILE=1 following this. Unfortunately, the documentation says "The support for command-line profiler using the environment variable COMPUTE_PROFILE has been dropped in the CUDA 8.0 release."

The question is, is there any way other than downgrading CUDA to profile OpenCL application with nvprof?

1
cudaprofilingopenclnvprof
Are you calling clReleaseContext within your application? – talonmies
@talonmies Yes, it properly calls release functions for context, queue, program, kernel, and memory object. – csehydrogen

1 Answers

4
votes

To the best of my knowledge, nvprof has never supported OpenCL profiling.

Running code with COMPUTE_PROFILE=1 invokes a driver based profiling mechanism which predates the introduction of nvprof. That driver based mechanism was deprecated a while ago and has now been removed as of CUDA 8 in favour of using nvprof.

As a result, it would appear that there is no way to profile OpenCL code running on NVIDIA hardware using the CUDA toolkit.