I have an Xavier NX with Jetpack 4.5 that I wish to profile CUDA kernels using Nsight Compute. I am connecting from my host computer and went through the process to enable root access over ssh already. When I connect I put the User Name as root and this is the output I get after I launch and attach to one of the CUDA sample applications:
==ERROR== An error occurred while trying to profile.
==WARNING== No kernels were profiled.
==PROF== Attempting to connect to ncu-ui at 10.0.0.72:50152…
==PROF== Connected to ncu-ui at 10.0.0.72:50152.
==WARNING== Note: Running with unmodified GPU clocks. If not controlled otherwise, profiling results may be inconsistent.
==PROF== Connected to process 11495 (/usr/local/cuda-10.2/samples/0_Simple/matrixMul/matrixMul)
==ERROR== Failed to prepare kernel for profiling
==ERROR== Failed to profile kernel “MatrixMulCUDA” in process 11495
==ERROR== UnknownError
==PROF== Disconnected from process 11495
==ERROR== An error occurred while trying to profile.
==WARNING== No kernels were profiled.
I’m also confused about Nsight Compute with Jetpack. I believe it’s part of the CUDA toolkit but I cannot find it anywhere locally on the Xavier? Or is it just a host program?
Per the suggestion, I updated to Nsight Compute 2020.3.1 but get the same issue. I was previously using 2020.1.2. The targeted program will execute and produce the correct output but kernel profiling fails.
Just to be sure, I have my target platform set to Linux (aarch64 sbsa). Is this correct?
After checking internally, please install the Nsight Compute via SDKmanager (within CUDA toolkit).
You can find it under the /opt/nvidia/nsight-compute/2019.5.0 folder.
The tool shared in our webpage is for other system, not for Jetson.
Thank you for the response. Does the host or Jetson need this install? Also, is the /opt/nvidia/nsight-compute/2019.5.0 on the host or Jetson side? I see this path on the host but not the Jetson.
Apologies I’m still having some issues. I suspect my install of Nsight-compute 2019.5.0 is incomplete.
When I go to /opt/nvidia/nsight-compute/2019.5.0 I only two folders - host and target. Comparing to another version I had installed this is only a subset.
I would like to reinstall 2019.5.0 but doing so via the SDKmanager seems to install a newer version. Is there a way to specify that I want to reinstall 2019.5.0 since you mentioned it needed to be installed via the SDKmanager?
I confirmed the solution. The issue was an incomplete install of Nsight compute 2019.5.0 from the SDK. After starting from scratch, installing from the SDK I can now profile kernels on the Xavier NX.