CUDA 4.0 visual profiler error: .csv for appllicatin run 0 not found

Hi, everyone,

I want to use Visual Profiler to measure cuda SDK program. But when I run any SDK sample, error accures. The following are my operations: (take clock SDK for example)

step1:In “session setting” launch clock_vs2008.sln, set max execution time to 1000 sec(enough large); then lauch;

Then error accures:


 Profiler data file "J:/.../clock/temp_compute_profiler_0_0.csv" for appilcation run 0 not found.

Since I’m not sure whether I should use .sln or .exe, then I try another:

step2:In “session setting” launch clock.exe, set max execution time to 1000 sec(enough large); then lauch;

Another error appears:

 Execution run #1 of program 'J:/.../clock.exe'failed, exit code:-1

OS: win7;

 GPU: 2 Tesla C2050

 CUDA: 4.0

thanks a lot.

Further I’ve set CUDA_PROFILE=1 CUDA_PROFILE_CSV=1, but still it doesnot work.

you should run the .exe file

Your program clock.exe failed and returned error code -1; Check the source code of the project to see why it fails, i.e. the reason for the -1 return value of main().

Is your hardware correctly set up? You can use view->device in Visual Profiler to confirm that.
Also, does clock.exe require you to press a key before it exists? If yes, setting the time to 1000 seconds will set you waiting for just that long. Though I doubt if visual profiler supports 1000 seconds at all or not. Probably it will bring down the value to a max value that it supports?

The output of the visual profiler is

=== Start profiling for session ‘Session1’ ===

Start program 'C:/ProgramData/NVIDIA Corporation/NVIDIA GPU Computing SDK 4.0/C/bin/win64/Release/clock.exe ’ run #1

d:/ : cudaSafeCall() Runtime API error : all CUDA-capable devices are busy or unavailable.

Program run #1 failed, exit code: -1

Error in program execution.

It seems that the problem returns to “all CUDA-capable devices are busy or unavailable”, which has been discussed several times in the forum, but seems do not get a satisfied solution.

But everything go well with CUDA 3.2 including visual profiler. So I wonder if the problem is the bug of CUDA 4.0.

Your clock.exe, when not using visual profiler, also fails to run correctly, right?

Maybe you can do a dump of your device states? I don’t have the SDK, though I remember there is an SDK example that does just that. … Okay the example is called deviceQuery… I suspect the compute mode is not correct

Anyway you should debug your program first before you run it on the profiler. Have you ever succeeded in building and running any program on your machine (after 4.0’s installed)? Are you using the 270.xx driver? Maybe you could try downgrade driver to the previous one you that you had success with?

It does look like a bug, though.