Awesome

dgSPARSE Library

Introdution

The dgSPARSE Library (Deep Geometric Sparse Library) is a high performance library for sparse kernel acceleration on GPUs based on CUDA. Now we aims to provide PyTorch-Based Fast and Efficient Processing kernel for users to have better experience in running applications like GNN, Rec sys and 3D pointcloud detection.

Installation

First, setup the the following environment variables:

export CUDA_HOME=/usr/local/cuda # your cuda path
export LD_LIBRARY_PATH=$CUDA_HOME/lib64 # your cuda lib path

Then, install with conda.

conda install -c dgsparse dgsparse

Or you can build from source

pip install -e .

A demo for SpMM inference time compared to other main-stream library. (Tested on RTX 3090 with feature=64).

Run Examples

Previously we provide C++ examples for SpMM and SDDMM kernels. To run these examples, please build dgsparse through make exp. Then, you could run our kernels in the example folder. Check more details in README under example directory.

Documentation

Our new docs for python API will be coming soon! Now you can refer to dgSPARSE Library Documentation for more details.