MiniGun: Light-weight GPU kernel interface for graph operations

The project is heavily influenced by the Gunrock project. The goal is to define a general kernel library that can cover a wide range of graph operations used in DGL.

Current roadmap:

• Port basic advance operator (all edge policy).
• Port advance operator with dynamic frontiers.
• Port more advance kernel policies
  • TWC
  • Inbound edge partition
  • outbound edge partition
• CPU implementation for advance
• Integration with DGL
• Filter

Build

Require CUDA.

mkdir build
cmake ..
make -j4

Play

See codes in samples. Currently, there are four:

• spmv
• spmm
• masked_mm
• edge_softmax

You could run sample by (in the build folder):

./samples/spmv/spmv

---

Disclaimer: Minigun project is not related to the "mini-gunrock" project.