Optimizing irregular dense operators of heterogeneous GNN models on GPU

GNN models on heterogeneous graphs have achieved state-of-the-art (SOTA) performance in various graph tasks such as link prediction and node classification. Despite their success in providing SOTA results, popular GNN libraries, such as PyG and DGL, fail to provide fast and efficient solutions for heterogeneous GNN models. One common key bottlenecks of models like RGAT, RGCN, and HGT is relation-specific linear projection. In this paper, we propose two high-performing tensor operators: gather-mm and segment-mm to address the issue. We demonstrate the effectiveness of the proposed operators in training two popular heterogeneous GNN models – RGCN and HGT. Our proposed approaches outperform the full-batch training time of RGCN by up to 3× and mini-batch by up to 2×.