Nsight compute from CUDA.jl and source annotation

General Usage Performance
1

Hi,

I try to build a training session on GPU programming (including kernel programming) and I want to use Julia and CUDA.jl as a development environment. Profiling is an essential tool for GPGPU and I think that profiling tools and profile based development is a very important topic to be covered in a GPU lecture.

I have followed this (excellent) video by @maleadt

and I have two questions:

  • Is their other (newest) video for GPU profiling for CUDA.jl ?
  • Is it possible to get source related profile with CUDA.jl similar to what you can get via the -lineinfo nvcc’s option and nsight compute profiling software ?
2

There’s an updated version from last January on Webinars - JuliaHub

There was, but due to a bug in CUDA we currently can’t emit the necessary debug info (the CUDA driver crashes). This has been filed with NVIDIA, and I’m told this may be fixed in the next driver release.

1 Like
3

Thank you Tim !

4

This updated version is great.
The cuda-memcheck --tools racecheck julia /tmp/test.jl (@ 32’28) is super useful !