No measurements recorded for long-running kernel

Hi folks -

I’m trying to use nvvp to profile some code I wrote for the GPU. If the kernel completes in a few seconds or less, I get kernel measurements (as expected) as a row on the timeline when the profiling is done. However, if the kernel takes longer than a few seconds to run, no measurements are shown, with no entry for the kernel on the timeline.

Is this expected behavior?

Tested with multiple programs on multiple machines. Running with Cuda 7.5 on CentOS 7.

Hi, agstockton

This should not be expected.
NVVP should be able to profile these long kernels.

  1. Can you try with latest toolkit ?
  2. Can you try with nvprof ./sample to see if it works ?
  3. If nvprof command line works, you can try nvvp again with disable serveral options in “Profiling Options” page
  4. If all above not work for you, please provide us the application to have a check.

Thanks!