Hello,
I’am trying to use NVIDIA Compute (11.2) for the first time on the local host (Windows),
but, unfortunately, as soon as NVIDIA Compute connects to the application,
its GUI freezes and I can not step through the workflow which I want to profile.
Launched process: ncu.exe (pid: 38216)
Launch succeeded.
Profiling…
==PROF== Connected to process 34460 (D:\Builds\xxx.exe)
NVIDIA Visual Profiler works fine for me.
What can I do wrong? is there a tutorial, how to get started?
As a general recommendation, you should use the latest available version of the tool, since it has a lot of bug fixes and of course new functionality compared to the one bundled with CUDA 11.2. You can download it separately from the toolkit from here and it’s backwards compatible with older toolkits and driver versions.
As for the hang, can you confirm which activity within the tool you are using? As you can find in the quickstart guides, it is expected that the application is paused when the Interactive Profile activity connects, since it works very similar to a CPU debugger (but on CUDA API calls and kernel launches). The (non-interactive) Profile activity, which just launches the ncu command line interface, should however not show this behavior.
It is possible though that it may hang on specific kernels if they are mandatory concurrent, since all kernel launches are serialized for profiling in the tool.
You can also find links to further training material here.
thank you very much for the quick response and usefull information.
I’ve already installed the latest CUDA 11.2, which is used in my company.
The interactive debugging works well, I have an issue with the non-interactive profiling.
I assume, the application which I try to profile is launching multiple CUDA activities, which are not relevant for me. I am interested in a certain part of the application workflow.
so, is it possible to disable profiling at the application start to avoid “hanging on specific kernels” and switch it manually on when needed?
as I already wrote, NVIDIA Visual profiler works fine with no GUI freeze…
I am afraid that most all the training materials are about using the NVIDIA Comput, but not about starting it :(
Just tried the latest Nsight Compute (2023.1.0), it makes no difference, GUI freezes…
There are kernel filtering options in the Profile > Filter tab of the GUI or from the CLI that allow you to choose which kernel to profile by name or ID including skip counts etc… and supporting regular expressions. Nsight Compute CLI :: Nsight Compute Documentation
If you know what kernel you want to profile, you can use those options to automatically skip profiling the other kernels.
unfortunately, this does not help. GUI is still frozen…
And if I terminate the application, I get a message saying that no kernels were profiled…
There are a few things you can do to try and narrow down the issue. Can you try running the profile from the CLI? The connection dialog in the GUI should provide a command line in the Activity>Profile>Common section. You can check if the CLI works okay to determine if this is actually related to the GUI. If that runs, see if you can open the generated report in the GUI. If it doesn’t run, please share the command and output you see on the CLI.
I have the same issue starting the profiling from the CLI: the application GUI freezes.
After one minute of “profiling” I’ve terminated the application process, so that one can see an error and warning messages.
I’ve created a simple console application which is starts execution of the kernels of interest immediately. Now I get a different kind of error:
==ERROR== Profiling is not supported on device 0. To find out supported GPUs refer --list-chips option.
Does this mean I can not use Nsight Compute on my developer PC with Quadro P3200 installed?
Visual Profiler is than the only profiling option for me?
That’s correct. Nsight Compute supports Volta and newer GPUs. See the list here:
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.
