Using the latest Jetpack on AGX Xavier and accompanying Host tools from the SDK Manager, the following happens:
I use the PVA via VPI and want to trace its behavior. I also need to profile the GPU via plain CUDA. Now, profiling each individually works fine. However both at the same time does not work. I attached some screenshots to showcase the behavior.
I would like to have both the CUDA kernels and the PVA Tasks in one profiling session. In the future I also expect the DLA and possibly VIC to be used as well as the NVENC hardware.
Is the scenario of showing all of those at the same time supported? For now I would be already happy if the PVA and CUDA traces would show up simultaneously though.
Here some additional info to aid diagnosis: # R32 (release), REVISION: 4.4, GCID: 23942405, BOARD: t186ref, EABI: aarch64, DATE: Fri Oct 16 19:37:08 UTC 2020
CUDA Version 10.2.89
Nsight systems about:
If further information is required I am happy to provide it.
Just checking back if there are any new insights on this. Is this issue reproducible for Nvidia? Or could it be a misconfigured system? If so, what could be possible causes? Thanks in advance :)
Hello tobias.fischer1. Sorry for the delay, Nsys dev team has started investigating your issue. Thank you for so detailed information. It really helps to get to the root cause faster.
FYI, I did not change anything specifically, but both the accelerator trace and the gpu trace show up now. So maybe it was some configuration issue on my side. I don’t want you to chase ghosts :)
I’m running into the same / a similar issue on a TX2 (still running an old R28.2.1, can’t change this ATM) and Nsight Systems 2020.3.2 No combination of options has produced a CUDA trace so far