Hello
As shown in the figure, when looking at the CUDA API as Events View, I want to know the duration values about ‘Kernel issue (or kernel issue)’ as a .csv file.
What should I do?
sudo nsys stats --report gputrace --format csv,column --output .,- app.qdrep
I’ve already done it.
But, the result is about kernel execution not call to kernel.
Please tell me in detail what to do.
You can run nsys stats --help-reports to see a list of all the available reports, and nsys stats --help-reports <report name> to see details about the data provided by a specific report.
Are you looking for the start time of the kernel?
No.
I want to get the duration of each kernel launcher as a csv file for all kernels.
What should I do?
You may use the report cudaapitrace to extract the duration time for kernel launchers.
E.g.,
nsys stats --report cudaapitrace --format csv,column --output .,- app.qdrep
Thanks to you, I succeeded in obtaining duration of ‘cudaLaunchKernel’.
However, as shown in the figure, only the api name ‘cudaLaunchKernel’ appears, but it is still unknown what kind of kernel the kernel is.
Can I get the kernel’s name and that’s duration as a csv file at the same time?
Have a look at the kernexecsum report, and see if that’s what you want. It is a summary report, so it shows aggregate averages, rather than each individual kernel launch, but it links CUDA API calls to Kernels, showing the API time, the Kernel time, and the “queue time” between the API call and kernel launch. It would be a good idea to have a look at the help docs (nsys stats --help-reports kernexecsum) to be sure you understand exactly what the different time values represent.
This is a fairly new report, so if you’re not running a recent release, it might not be available. If you’d rather have this as a trace (where each individual API/Kernel is reported, rather than a summary), let us know.
