I am relatively new to GPU programming. I want to use nvidia profiler (Version 3.2 for visual profiler) for CUDA and OpenCL profiling on Mac. I have two problems:
The NVIDIA command-line profiler works fine for CUDA but not the visual one as it says that “Unable to initialize profiling in start/stop mode”, This problem is already asked but no solution is known to me.
The NVIDIA Profiler is not working for OpenCL even the command line one (no log file is created at all and there is no permission issue). Other possible causes mentioned are with deallocation of resources but that should not be the case with openCL SDK code samples (e.g. VectorAdd). Following is my environment variable setting:
CUDA_PROFILE=1
CUDA_PROFILE_CSV=1
CUDA_PROFILE_CONFIG=<path to .cofig file> which contains 4 entries: gld_coherent, gld_incoherent, gst_coherent, gst_incoherent
I have exactly the same problem. I did not try the command line profiler, but I get the same error message when trying to use the visual profiler with the sample projects (CUDA and OpenCL). Me too, I am new to GPU programming, and I have no idea how to find out what’s going wrong here.
By mere accident, I found some kind of workaround:
I click the “Abort Application” button just before profiling ends (i.e. when the status bar on the bottom right is at about 90%),
it asks me whether to “abort remaining 3 runs” (whatever that means precisely),
I click “Yes”,
I click away three error messages saying that some profiler data files could not be found.
Then, despite all the errors, I get a session containing profiler data.
Naturally I have no idea whether these data are correct and complete. The output I got from the oclBoxFilter example seems plausible, but I have not enough experience to really tell.
So, if someone has found a real solution to the problem in the meantime, I would be really grateful.