I’m trying to run Nsight Systems 2022.1.1 on a x64 Windows app running on Server 2019. The app uses libtorch with cuda and a Tesla v100 gpu.
I’m launching it from Visual Studio 2019 integration. I’m using the default settings except for turning on cuda trace, shown below (removed sreenshot due to “new user” limitation of one graphic per topic), as well as the error message.
Usually that message indicates that the application ended without data collection (before the profiler got started. Were you launching the application through a script? Do you need to set to trace the entire process tree?
Could you then also uncheck “start profiling manually” - that should result in a report though the app may still crash. But we can check the stdout/err from the report for the app.
Somehow the report didn’t catch the std outputs. We can try another method, using Nsys CLI to start the profiling session to see if it works.
To do this firstly you need to find where Nsight Systems was installed on your system. After that, use the following command (please use a terminal that has Administrative privilege):
cd <your working dir>
<Nsys install folder path>/target-windows-x64/nsys profile -t none --capture-range=hotkey <your app>
(Note that this command line should provide the same setting as shown in your project)
Sorry but I’ll need to wrap up today soon. In case the CLI and 2022.2 both don’t work - is it possible to share a reproducer for me to debug at my side? (If that’s possible, just share it and I’ll look into it tomorrow)