When I was using pgprof, I clicked under the analysis window.
Guided analysis ->performance-critical kernels ->perform additional analysis .
Then the program starts the profile and then pops up “profiling failed:unable to collect metric and event values.” the application being profied received a signal.
======== Error: Application received signal 135
==8397== Warning: Some profiling data are not recorded. Make sure cudaProfilerStop() or cuProfilerStop() is called before application exit to flush profile data.
I now look at GPU usage through nvidia-smi. Probably only 15% (K40), so I want to use the further profile to view the GPU usage information, but I have encountered this problem.
When I was building, I used:
Pgfortran -acc -fast -ta=tesla:deepcopy -ta=tesla:pinned -Minfo *.F90 -g -o cnm
So why is this? what should I do?
ThX
Can you provide more information, what is the CUDA driver version number on your system? Can you successfully collect a ‘Timeline’ profile?
It might also help us debug this problem if you could provide the source code of this application or a test program that also exhibits this same issue.
The amount of source code is too large, I don’t know how to send it to you, probably about 100M. If you can give me your email, I can send it to you.
Cuda version: cudatoolkit 9.0.
I can generate timeline information, then in the analysis options below, I chose perform additional analysis. The program will then rerun, then fail, and then output the above information.
Hi wanghr323,
Please send a note to PGI Customer Service (trs@pgroup.com). Your app is probably too large to send via email so we’ll need to arrange another way for you sent it to us such as NVIDIA’s secure FTP.
-Mat