Failed to profile CUDA.jl with Nsight Systems 2024.1.1

Specific Domains GPU
1

I am following the doc to learn profiling GPU in Julia. When I followed
$ nsys launch julia
It started a Julia terminal but then failed when I run
using CUDA
and return errors:

Please submit a bug report with steps to reproduce this fault, and any error messages that follow (in their entirety). Thanks.
Exception: EXCEPTION_ACCESS_VIOLATION at 0x6be45c40 -- nvtxGlobals_v3 at C:\Users\ddt00\.julia\artifacts\b4eeaf094ffb6aacf1b20ee5d2ac9aa1818fc732\bin\libnvToolsExt.dll (unknown line)
in expression starting at REPL[1]:1
nvtxGlobals_v3 at C:\Users\ddt00\.julia\artifacts\b4eeaf094ffb6aacf1b20ee5d2ac9aa1818fc732\bin\libnvToolsExt.dll (unknown line)
Allocations: 1476205 (Pool: 1475266; Big: 939); GC: 2

My CUDA settings are:

julia> CUDA.versioninfo()
CUDA runtime 12.3, artifact installation
CUDA driver 12.3
NVIDIA driver 546.12.0

CUDA libraries:
- CUBLAS: 12.3.4
- CURAND: 10.3.4
- CUFFT: 11.0.12
- CUSOLVER: 11.5.4
- CUSPARSE: 12.2.0
- CUPTI: 21.0.0
- NVML: 12.0.0+546.12

Julia packages:
- CUDA: 5.1.2
- CUDA_Driver_jll: 0.7.0+1
- CUDA_Runtime_jll: 0.10.1+0

Toolchain:
- Julia: 1.9.2
- LLVM: 14.0.6

Environment:
- JULIA_CUDA_NSYS: C:\Program Files\NVIDIA Corporation\Nsight Systems 2024.1.1\target-windows-x64\nsys.exe

1 device:
  0: NVIDIA GeForce GTX 1070 with Max-Q Design (sm_61, 7.886 GiB / 8.000 GiB available)
2

An issue exists already, Crash on Windows · Issue #37 · JuliaGPU/NVTX.jl · GitHub. If you’re familiar with Windows development, it would be great to chime in there.

3

Hi @maleadt

Sorry that I’m not familiar with it. Wish someone could solve it and it would be a great help for the development.

I also tried nsight computing. It works for code of several line but it got stuck if I want to test a large project (containing mixed use of both CPU and GPU code). Is there any alternative way I can do for profiling in Julia right now? Or can I only profile my GPU code, which is just a small portion of my code in the project?

4

If you need to profile both CPU and CUDA GPU code, NSight Systems is really the tool to use. Maybe you could try using Linux, or WSL?

5

Thanks! I would have a try in Linux.