1

I have implemented a program that uses a single GPU using the cudaStreamWaitEvent() function to set dependency within two streams using events.

In order to verify this dependency, is it possible to use the "Dependency Analysis" view on the Nvidia Visual Profiler ?

If not, what does each of the following options in the dependency analysis view provide?

  • Focus Critical Path
  • Highlight Execution Dependencies

detailed information on those options doesn't seem to be available in the nvidia official website and here

BAdhi
  • 420
  • 7
  • 19
  • 1
    This [presentation](http://on-demand.gputechconf.com/gtc/2016/presentation/s6810-swapna-matwankar-optimizing-application-performance-cuda-tools.pdf) may be of interest – Robert Crovella Jul 11 '17 at 14:54

1 Answers1

1

Yes, you should be able to use the dependency analysis feature to verify your usage of most CUDA synchronization APIs, including cudaStreamWaitEvent.

To use either of the two mentioned options, you must have computed the dependencies in your application trace. In order to do that, in NVIDIA Visual Profiler, select "Unguided Analysis" and there "Dependency Analysis".

Now you can enable "Highlight Execution Dependencies", which will highlight the incoming and outgoing dependencies for each analyzed activity on the timeline in red, once you hover over it or select it.

If you use cudaStreamWaitEvent to block one kernel until another kernel in another independent stream has finished, those will be highlighted in red if they are direct dependencies.

FelixS
  • 91
  • 3