With Nsight Systems 2024.5.1 I’m trying to profile a C# exe (with no cuda) that loads a C++ dll (with no cuda) and then that dll loads another C++ dll where all the cuda code is. It’s working at a basic level (my exe process is showing with tons of threads, I can see my GPU activity and it looks right), but I can’t see the “CUDA API” timeline. I have checked Collect CUDA trace/Collect GPU Memory Usage. Running Nsight Systems as admin. Nsight Compute works fine, it sees all the API events correctly.
@liuyi can you look into this? Seems likely to be something about how we are using CUPTI.
Is there any progress on getting this issue resolved, or at least explained?
Oops … looks like the whole problem here was failing to call cudaDeviceReset() at the proper time. After sticking that in I’m seeing the CUDA API timelines.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.