This is the updated version of the methods proposed in "Fast Sparse GPU Kernels for Accelerated Training of Graph Neural Networks", accepted by IPDPS 2023.
The code contains high-performance FP32 SpMM implementations (for Ampere and Hopper Arch).
Please cite:
@INPROCEEDINGS{10177444,
author={Fan, Ruibo and Wang, Wei and Chu, Xiaowen},
booktitle={2023 IEEE International Parallel and Distributed Processing Symposium (IPDPS)},
title={Fast Sparse GPU Kernels for Accelerated Training of Graph Neural Networks},
year={2023},
volume={},
number={},
pages={501-511},
doi={10.1109/IPDPS54959.2023.00057}}
Please make sure you are running on Ampere (A100, A800) or Hopper (H100, H800) GPUs.
Please use NVCC >= 11.8.
git clone [email protected]:fan1997/HP-SpMM-SDDMM.git
mkdir build
cd build
cmake .. && make -j
cd ./dataset
chmod +x download.sh
source download.sh
source run.sh
We compare with cuSPARSE-12.2 and GE-SpMM (https://github.com/hgyhungry/ge-spmm.git). We set K to be 32, 64, and 128, and the average GFLOPS are reported.