Nsight Compute profiles the GPU side of the Python application, meaning the CUDA kernels launched by Python or through the frameworks called by it. How to enable line information for this depends on the framework and library used. For numba.cuda, you can find that info here. For other libraries, please contact the library developer or its documentation on how to enable this.
Hi @felix_dt,
thanks for your answer.
I’m currently using NVIDIA Warp as framework. I was able to perform a general system profiling using Nsight System as described here (Profiling — Warp 1.5.1). Now I wanted to profile a specific kernel of the code written in Warp using Nsight Compute.
Do you know how to enable line information for Warp?
The ability to “Resolve” requires line info to start with. Once you have line info, to avoid the need to resolve, you can also profile with --import-source yes. I am not aware that Nvidia Warp supports lineinfo, but I sent a message to the team to check.
I confirmed with the team that Debugging — Warp 1.5.1 does in fact work and produce lineinfo for GPU code. It should map to the main CUDA file generated from your Warp Python code though, not to the original Python.
It’s hard to say why it doesn’t work in your particular case. That said, there is no dedicated lineinfo option, and using the debug mode will make the profile result not actionable due to the debug overhead. We will look into adding such an option in the future, including mapping to the original Python.
