I am trying to use nvprof on a cuda/c++ program, but I get the output:
======== Warning: unable to locate profiling library, GPU profiling skipped
... my output ...
======== Warning: No CUDA application was profiled, exiting
My command:
nvprof.exe -o prof.nvvp ./debug/my.exe args
It's a pascal GPU, drivers are up to date, I can run CUDA programmes without issue. I have CUDA 9.0, 10.1 and 11.0 installed. Path and CUDA_PATH link to 11.0. I have not found any answers to this question online.
There are some similar questions to bad installations, but only this with the same warning. Found nothing for windows.
I suspect some kind of bad installation, but I have reinstalled the CUDA Toolkit multiple times already and the examples from the install guide run fine.
What should I try next? My goal is to find inefficiencies in my program for optimization (serialization, memory usage, occupancy etc.). nprof seems to be the best/only option for this.
The error is NOT the same as nvprof command error: cupti64_102.dll was not found. This error (and it's solution) is the " similar error" I mentioned. I already have the path to cupti in PATH.