I am trying to profile some PyTorch programs on my Jetson TX2 (8GB) Devkit w/ Jetpack 4.6.2.
JetPack 4.6.2 Missing NSIGHT Compute
I have seen other users mention that the NVIDIA NSight Compute executable should be located somewhere in /opt/nvidia/, but that is not the case for me.
I used the SDK Manager to reinstall the JetPack 4.6.2 development kit components on my Ubuntu 18.04 machine, but I still do not see the NSight Compute executable in /opt/nvidia/
Remote Launch + Attach
I have also attempted to remotely launch + attach with NSight Compute 2019.5.3 on my Ubuntu 18.04 host machine, but I encounter a warning that indicates I do not have the correct permission to remotely profile.
To fix this issue, I copy pasted the CLI code from NSight Compute and ran it directly on the Jetson TX2 with sudo. This also results in a problem, because I receive an error that my GPU is unsupported.
It would be great to understand why NSight Compute is not installed on my TX2, and how to fix that.
I have also attached the output of the deviceQuery sample just in case it is useful deviceQuery.txt (2.4 KB)
I have NSight Compute 2019.5.3 on my host system, however I encounter two problems when trying to remotely launch + attach on the TX2:
(1) NSight Compute returns error code 103, and notes that I do not have permission to profile on the remote device (2) To resolve this, I launched the CLI command supplied by NSight Compute directly on the Jetson TX2 prefixed with sudo. This results in an error that says the TX2 GPU is unsupported.
Do you have any suggestions on how to resolve this? Let me know any further information I can supply to assist you.
For TX2, could you try if nvprof can meet your requirement?
You can find it in the below folder:
$ /usr/local/cuda-10.2/bin/nvprof -hnvidia@tegra-ubuntu:/usr/local/cuda-10.2/tools$ ls
Usage: nvprof [options] [application] [application-arguments]
Turn on/off aggregate mode for events and metrics specified
by subsequent "--events" and "--metrics" options. Those
event/metric values will be collected for each domain instance,
instead of the whole device. Allowed values:
on - turn on aggregate mode (default)
off - turn off aggregate mode
Collect profiling data that can be imported to Visual Profiler's
"analysis" mode. Note: Use "--export-profile" to specify
an export file.