dgSPARSE
Open-source sparse kernels and GNN operators for the GPU.
dgSPARSE is a collection of GPU sparse kernels aimed at graph learning workloads — SpMM, SDDMM, and the GNN operators built on top of them, shipped as a drop-in library rather than a research artifact.
My contributions here span several papers: online locality and reduction parallelism for SDDMM (ICCD ‘21), heuristic adaptability to input dynamics for SpMM (DAC ‘22), and the computation-graph analysis behind dgNN (MLSys ‘22).