ncu

Solution

1. Install Nsight Compute.
   • Version: 2021.2.1 -> Download .run file -> chmod +x -> ./xxx.run
   • sudo echo options nvidia "NVreg_RestrictProfilingToAdminUsers=0" >> /etc/modprobe.d/nvidia.conf.
   • sudo reboot
2. Generate binaries for code you want to analyse.

   nvcc -G -o nodiv.out ../src/aj_warp_divergence_extra/nodiv.cu
   nvcc -G -o div.out ../src/aj_warp_divergence_extra/div.cu
   

3. Generate reports. (view [4] 6.3 Section to translate metrics from nvprof to ncu)

   /NVIDIA-Nsight-Compute-2021.2/ncu -o profile_test_div -f -k "divergence_code" --target-processes all --section "WarpStateStats" --section "SourceCounters" --launch-count 1 ./div.out
   /NVIDIA-Nsight-Compute-2021.2/ncu -o profile_test_nodiv -f -k "no_divergence_code" --target-processes all --section "WarpStateStats" --section "SourceCounters" --launch-count 1 ./nodiv.out
   

4. View reports through UI.

   /NVIDIA-Nsight-Compute-2021.2/ncu-ui
   

   • Import reports. (File -> Open File -> .ncu-rep)

References

1. Profiling GPU code with NSIGHT systems
2. NVIDIA Development Tools Solutions - ERR_NVGPUCTRPERM: Permission issue with Performance Counters
3. Using Nsight Compute and Nsight Systems
4. Nsight Compute CLI