Nsight_Compute_Tutorial

Docker container

To build the Docker container, follow these steps:

1. Open a terminal and navigate to the Nsight_Compute_Tutorial/docker directory.
2. Run the following command to build the container without caching:

docker build -t cuda_nsight:v0.1 --no-cache --rm --file Dockerfile.nsight .

To run the Docker container, execute the following command:

cd Nsight_Compute_Tutorial
docker run -it --rm --gpus all --cap-add=SYS_ADMIN --volume="$PWD:/workspace" cuda_nsight:v0.1

Make sure to replace cuda_nsight:v0.1 with the appropriate image name and version.

How to use nvprof

Basic commands:

1. Simple analysis:

nvprof ./your_cuda_application

2. Output to .nvprof file for further analysis with nvprof or nvvp tools:

nvprof --output-profile my_profile.nvprof ./your_cuda_application

3. Output to .nvvp file for visualization analysis in NVIDIA Visual Profiler (nvvp):

nvprof --export-profile my_timeline_report.nvvp ./your_cuda_application

Event and Metrics

1. all events

nvprof --events all ./your_cuda_application

2. certain events.

nvprof --events event1,event2 ./your_cuda_application

3. certain metrics and output to .nvvp

nvprof --metrics metric1,metric2 --export-profile my_timeline_report.nvvp  ./your_cuda_application

How to use Nsight system

User guide

https://docs.nvidia.com/nsight-systems/UserGuide/index.html

1. Simple Analysis

 nsys profile ./your_cuda_application
 nsys profile --stats=true ./your_cuda_application

2. save report as my_profile

 nsys profile --stats=true -o my_profile ./your_cuda_application
 nsys profile --stats=true --trace=cuda -o my_profile ./your_cuda_application
 nsys profile --stats=true --trace=osrt,cuda,nvtx --show-output=true --output=my_profile ./your_cuda_application

How to use Nsight compute

User guide

https://docs.nvidia.com/nsight-compute/NsightCompute/index.html#

1. save report as my_profile

ncu --export=my_profile ./your_cuda_application
ncu --target-processes=all --export=/workspace/report/sgemm --set=full --force-overwrite ./sgemm

2. show profiling result in terminal

ncu --import my_profile.ncu-rep --page=detail

Facing problem: ERR_NVGPUCTRPERM: Permission issue with Performance Counters

• Solution: add --cap-add=SYS_ADMIN in docker run.

Remote Profiling on Windows Visual Profiler (Host)

1. Download and install cuda-tool-kit with same version of device. https://developer.nvidia.com/cuda-11-8-0-download-archive
2. Download JRE1.8 https://www.java.com/zh-TW/download/
3. Search Visual Profiler in Host.

Remote Profiling on Windows Nsight System and Compute (Host)

1. Download and install cuda-tool-kit with same version of device. https://developer.nvidia.com/cuda-11-8-0-download-archive. Nsight system is included.
2. Download JRE1.8 https://www.java.com/zh-TW/download/
3. Download latest Nsignt Compute https://developer.nvidia.com/nsight-compute

• In Device
  • install Nsight system https://docs.nvidia.com/nsight-systems/InstallationGuide/index.html
  • install Nsight compute https://developer.nvidia.com/blog/using-nsight-compute-in-containers/

Reference

CUDA Profiler
CUDA Programming Guild
Learn-CUDA-Programming
CUDA Parallel Reduction