Is there a way for the nvprof option of nsys to output the full kernel names? This would really help especially with the thrust related calls which have incredibly long signatures:
2.0 172368360 91 1894157.8 1895315.0 1857651 1906739 7486.6 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrus…
Or do I need to run nsys twice instead: once to generate the nsys-rep file and a second time to extract the stats?
The “nsys nvprof” is intended to be able to give users of nvprof a way to transition to nsys. It isn’t intended to be fully funtionally with all the new options.
If you use “nsys profile” with the --stats option, it will give you a simple output on the first pass, but that will also have the shortened names.
So yes, you will need to collect the result and then run “nsys stats” to get the stats the way you want. See User Guide — nsight-systems 2024.5 documentation for details on stats scripts and options. (That’s an exact link, but the forum software munges the text).
Note that the ability to show demangled names requires having used a relatively recent version of Nsys to collect the data.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.