Export CUDA nvprof output to the Visual Profiler

6
votes

I would like to extract the data from my GPU application in order to check its limits. I have to use nvprof because the application runs on a remote server, so I should create a file to import locally in the Visual Profiler. I've tried to create the file with nvprof -o file_name <app> <params> and with nvprof --analysis-metrics --output-profile file_name <app> <params> but when I import these files on the Visual Profiler, in the Analysis section some fields are empty: "insufficient global memory load data", "insufficient global memory store data", "insufficient kernel SM data"... . How could I generate a file (or more) in order to have all the information for the Analysis section? I compile the cuda code with nvcc with the flags -lineinfo -arch compute_20 -code sm_20 --ptxas-options=-v. These are some examples of empty fields: enter image description here

1
cudanvvpnvprof
I ran your nvprof --analysis-metrics ... command on the cuda vectorAdd sample. I had no trouble with analysis data both in the analysis tab and details tab of the visual profiler. Perhaps you'll need to be more specific about your exact test case and the exact places you are looking for data. I believe the "insufficient data" messages are a result of the actual kernel/code you are profiling, not anything to do with the profiler itself. – Robert Crovella
Thanks @RobertCrovella for the reply. I've added an image with an example of empty fields. – Stefano Sandonà
The number of times I get this "warning" message without any guidance from the UI on how to resolve it is fairly infuriating. This is on 9.0... how is it not fixed by now? Nvidia sure isn't broke! – interestedparty333

1 Answers

1
votes

You can try to add a session instead of importing prof file into the visual profiler. I run into the similar problem. what I did is adding a session according to the instructions in here, and you will be able to see all the information.