There are two ways to get the dataset.
We provide converted files that are dervied from the Graph Challenge website over a Box. This can be downloaded from (Total of ~90GB): BoxFolder
After download is complete, untar all the files present inside the dataset.
Automatic download and compilation (requires gzip and tar support). Space Required: ~200GB. Post processing ~90GB.
We assume you have set PROJREPO environment variable to the repo home.
git clone https://github.com/merthidayetoglu/SpDNN_Challenge2020.git
cd SpDNN_Challenge2020
export PROJREPO=$PWD
mkdir dataset
cd dataset
bash $PROJREPO/utils/download.sh
- Latest version of CUDA.
- g++ compiler
# For CentOS/RedHat system
sudo dnf install mpich mpich-devel
# For Ubuntu system
sudo apt-get install -y mpich
export the installed mpich binary path and lib paths to $PATH and $LD_LIBRARY_PATH variables.
After clearing dependencies and setting PROJREPO environment variable, run the following.
cd singlegpu
bash run_ampere.sh > output.log // you can change to your version of GPU. Ensure you set correct SM and COMPUTE Arch in the makefile settings.
Do let us know if you get any errors in output.log. Ideally it should work without any issues.
Tutorial Link : Here
Session 2 --- Modeling and Performance of Tiled SpMM Slides: Here
HPEC'20 Graph Challenge (Champion): Here
If you use our work in your experiments, please cite with the following bibtex
@inproceedings{hidayetouglu2020scale,
title={At-scale sparse deep neural network inference with efficient GPU implementation},
author={Hidayetoglu, Mert and Pearson, Carl and Mailthody, Vikram Sharma and Ebrahimi, Eiman and Xiong, Jinjun and Nagi, Rakesh and Hwu, Wen-mei},
booktitle={2020 IEEE High Performance Extreme Computing Conference (HPEC)},
pages={1--7},
year={2020},
organization={IEEE}
}
MIT License