TC-GNN (Running Sparse GNN on Dense Tensor Core on Ampere GPU)
- Cite this project and paper.
@inproceedings{TC-GNN,
title={TC-GNN: Accelerating Sparse Graph Neural Network Computation Via Dense Tensor Core on GPUs},
author={Yuke Wang and Boyuan Feng and Yufei Ding},
booktitle={Arxiv},
year={2022}
}
- Clone this project.
git clone [email protected]:YukeWang96/TCGNN-Pytorch.git
- OS & Compiler:
Ubuntu 16.04+gcc >= 7.5cmake >= 3.14CUDA >= 11.0andnvcc >= 11.0
Files and Directories.
config.py: the configuration file for the shape of a TC block.bench.py: the benchmark file for invokingmain_tcgnn.pyfor various datasets and models.main_tcgnn.py: the main entry for running TC-GNN.count_TC_blocks.py: counting the total number of TC blocks without sparse-graph translation.proc_prof.py: get the detailed GPU kernel metrics from the ncu csv output.TCGNN_conv/: the directory for core TC-GNN implementations, includingTCGNN_kernel.cuandTCGNN.cpp.
Environment Setup.
[Method-1] Install via Docker (Recommended).
- Go to
Docker/ - Run
./build.sh - Run
./launch.sh
[Method-2] Install via Conda.
- Install
condaon system Toturial. - Create a
condaenvironment:
conda create -n env_name python=3.6
- Install
Pytorch:
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge
or using pip [Note that make sure the pip you use is the pip from current conda environment. You can check this by which pip]
pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
- Install
Deep Graph Library (DGL).
conda install -c dglteam dgl-cuda11.0
pip install torch requests tqdm
- Install
Pytorch-Geometric (PyG).
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-geometric
Install TC-GNN.
Go to TCGNN_conv/, then run
./build.sh
to install the TCGNN_conv modules with Pytorch binding. Note that this step is required for both Docker and Conda setup.
Download graph datasets.
Get the preprocessed datasets in .npy at here, then run
tar -zxvf tcgnn-ae-graphs.tar.gz
Running PyG baseline.
- Go to
pyg_baseline/directory;- Pass the
--modelparameter inpyg_main.pywithgcnandginto profile the example GCN and GIN model, respectively;./0_bench.py| tee run_pyg.logto run the script and the report 10 epoch runtime for all evaluated datasets../1_log2csv.pyto convert therun_pyg.logtorun_pyg.csvfor ease of analysis.
Running DGL baseline.
- Go to
dgl_baseline/directory- Pass the
--modelparameter indgl_main.pywithgcnandginto profile the example GCN and GIN model, respectively;./0_bench.py| tee run_dgl.logto run the script and the report 10 epoch runtime for all evaluated datasets../1_log2csv.pyto convert therun_dgl.logtorun_dgl.csvfor ease of visualization.
Running TC-GNN.
- Under the current project directory
./0_bench.py| tee run_TCGNN.logto run the script and the report 10 epoch runtime for all evaluated datasets../1_log2csv.pyto convert therun_TCGNN.logtorun_TCGNN.csvfor ease of analysis.
92 Dec 19, 2022
149 Dec 22, 2022
27 Dec 16, 2022
45 Dec 8, 2022
216 Dec 26, 2022
143 Dec 29, 2022
48 Dec 23, 2022
0 Mar 11, 2022
39 Nov 10, 2022
288 Oct 28, 2022
45 Nov 29, 2022
17 Nov 14, 2022
3 Dec 13, 2021
78 Dec 19, 2022
454 Jan 03, 2023
76 Dec 27, 2022
7 Aug 04, 2022
1 Feb 24, 2022
298 Dec 22, 2022
46 Dec 26, 2022
17 Dec 28, 2022
31 Dec 29, 2022
427 Dec 25, 2022
8 Jan 22, 2022
70 Jan 03, 2023
343 Jan 03, 2023
33 Nov 28, 2022
29 Oct 06, 2022
45 Nov 10, 2022