SPIDER: Unleashing Sparse Tensor Cores for Stencil Computation via Strided Swapping (PPoPP'26)

This repository contains the source code of SPIDER, a novel system to converting stencil computation into sparse matrix multiplication using Sparse Tensor Core (SpTC).

• Hardware
  • 1x NVIDIA A100 GPU
• Software
  • CUDA (12.8 Tested). Should support all CUDA version above 12.5 (supporting mma.sp.ordered_metadata)
  • GCC (11.4 Tested) or Clang.
  • cuDNN (9.8.0 Tested).

You can use the Dockerfile to quickly prepare the environment.

First clone the code repository by:

git clone --recurse-submodules https://github.com/KevinWu2017/SPIDER.git SPIDER
cd SPIDER

# Please make sure you nvidia driver supports cuda-12.8 (driver version newer than 570.26).
docker build -t spider_ae .
docker run --gpus all --name spider_ae -it spider_ae /bin/bash

Then create the virtual environment and install packages:

python3 -m venv spider_venv
source spider_venv/bin/activate
pip install -r requirements.txt

In project root, you can build the SPIDER and baselines by using:

./scripts/build_SPIDER.sh # build SPIDER
./scripts/build_baseline.sh # build baselines.

The executable program is listed in ./build/bin/

# e.g. To run a 2d stencil with SpTC and half precision, you can use
./build/bin/stencil_2d_half_sparse -M 10240 -N 10240 -T 1000 --profile
# you can check the acutal parameter input by running
./build/bin/stencil_2d_half_sparse -h

Please first make sure the python venv is activated (by running source spider_venv/bin/active)

./scripts/Figure10_run.sh
python3 scripts/Figure10_draw.py

The output figure is saved in ./outputs/Figure10.pdf

./scripts/Figure11_run.sh
python3 scripts/Figure11_draw.py

The output figure is saved in ./outputs/Figure11.pdf

./scripts/Figure12_run.sh
python3 scripts/Figure12_draw.py

The output figure is saved in ./outputs/Figure12.pdf

If you have any questions, please send an email to the author at qiqi.gu@sjtu.edu.cn or cpwu_sjtu@sjtu.edu.cn.

If you use our code, please cite our paper:

@misc{gu2025sptcstencilusingsparsetensor,
      title={SPTCStencil: Using Sparse Tensor Cores for Stencil Computation}, 
      author={Qiqi GU and Chenpeng Wu and Heng Shi and Jianguo Yao},
      year={2025},
      eprint={2506.22035},
      archivePrefix={arXiv},
      primaryClass={cs.DC},
      url={https://arxiv.org/abs/2506.22035}, 
}

This codebase uses ConvStencil, LoRAStencil and TCStencil as baseline . We sincerely thank the authors for their excellent work!