https://github.com/kis-balazs/cuda-research
CUDA Research & Code. Course-style structured. Inspiration from @Infatoshi.
https://github.com/kis-balazs/cuda-research
cuda
Last synced: about 1 year ago
JSON representation
CUDA Research & Code. Course-style structured. Inspiration from @Infatoshi.
- Host: GitHub
- URL: https://github.com/kis-balazs/cuda-research
- Owner: kis-balazs
- License: bsd-3-clause
- Created: 2024-11-16T14:40:33.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2025-01-28T19:18:43.000Z (over 1 year ago)
- Last Synced: 2025-02-17T06:13:07.861Z (over 1 year ago)
- Topics: cuda
- Language: Cuda
- Homepage:
- Size: 10.6 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# CUDA-Research
Research & Code for CUDA components developed/experimented by me.
## Resources Outline
- [cuda-course](https://github.com/Infatoshi/cuda-course)
- [cuda-mnist](https://github.com/Infatoshi/cuda-course)
- [cuda-opencv-examples](https://github.com/evlasblom/cuda-opencv-examples/tree/master)
- [Getting Started with Accelerated Computing in CUDA C/C++ -- NVIDIA DLI Course](https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-AC-04+V1)
- [Holistic Trace Analysis](https://github.com/facebookresearch/HolisticTraceAnalysis/tree/main) - this seems like a very-very good resource for understanding PyTorch GPU traces, for end-to-end ML model evaluation
- [Open Torch Perf Traces](https://reimbar.org/dev/torch-profile-trace/)
## Environment
- [CUDA Installation Guide Ubuntu](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/#ubuntu)
- [CUDNN Installation - Latest](https://docs.nvidia.com/deeplearning/cudnn/installation/latest/index.html)
## Related
- [[repo] kis-balazs/CUDA-Containers-Infrastructure-Repository](https://github.com/kis-balazs/cuda-containers-infra)
- [relevant(+adjacent) CUDADocs.md](99_Docs/CUDADocs.md)
- [[PDF] Multi-GPU Programming @ Supercomputing 2011](https://www.nvidia.com/docs/IO/116711/sc11-multi-gpu.pdf)
## Particularly Handy Gists
- [kis-balazs/macro__check_cuda_return.cu](https://gist.github.com/kis-balazs/03f8023320639632db46523aa6e2bc69)
- Macro to check CUDA Error from Function Returning exit code
- [kis-balazs/verif_last_cuda_error.cu](https://gist.github.com/kis-balazs/3a2590d4bf90f33b0f8776d94da25a92)
- Verify last CUDA Error code, mainly from kernels, to use for straight-forward debugging
---
## Important Notice - GPU-specific nvcc params
When running `nvcc` commands, the `-arch` command is not always synced to the correct physical GPU. This can be fixed by specifying this.
Steps:
- find compute version: find the GPU compute version [here](https://developer.nvidia.com/cuda-gpus)
- specify when compiling: `nvcc -o exec code.cu -arch=compute_XX`
- optionally, `code` can be specified as well: `-code=sm_XX,compute_XX`