Nsight-compute over ssh: Failed to prepare kernel for profiling

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
,
1

I’m trying to profile a cuda kernel with nsight-compute over ssh. The host is an x86 debian, the nsight-compute (from https://developer.nvidia.com/gameworksdownload) version is 2021.2 (which according to this resource JetPack SDK | NVIDIA Developer seems to be the matching version). The device is a jetson agx/32 with cuda 10.2 etc.

I’m accessing the device as root. Otherwise I get an “Error: Unknown error” for each kernel in the program.

The program runs in the profiler, but when trying to profile any individual kernel I get “Failed to prepare kernel for profiling”. What could be the reason for that?

I have seen this thread: How to avoid the error ==ERROR== Failed to prepare kernel for profiling I’m already running the nsight-compute version that is presumably closest to jetpack 4.6. I cannot install the specific jetpack/host version of nsight-compute through the sdk manager because the sdk manager refuses to install anything on a debian os.

If the nsight-compute version might be an issue, where can I obtain the correct version?

best regards, Lukas Wirz

3

Hi,

Please install the Nsight Compute from the CUDA package of JetPack.
The compatible version is 2019.5.3 and is located at /opt/nvidia/nsight-compute/.

Thanks.

4

Ok, and as I was asking above, where can I get that version given the constraint that I run an up to date Debian? Or is it a hard requirement to run a four-year-old Ubuntu to profile kernels on Xavier?

5

Hi,

That is the Nsight Compute version, not related to the host OS version.
You can get it by installing the host package with SDK Manager.

Thanks.

6

No, that doesn’t work. The sdk manager refuses to install anything unless the correct version of ubuntu is detected.

Reference for anyone else having this issue: You can obtain the correct *deb package from another computer where the correct ubuntu version is installed. However, it can’t easily be installed because of package dependencies/conflicts, especially if any other cuda packages are already installed. The only straightforward method seems to be to install a four-year-old ubuntu on a spare computer. It would be quite nice if nvidia were documenting this requirement.

7

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

Hi,

Would you mind sharing your host environment and the dependencies error with us?

Thanks.

9

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.