FusedMM CUDA

This project was done as part of Ariful Azad’s HIPGraph group. Graph neural networks rely on message passing to propogate information between nodes in the graph. This is a very performance heavy operation and relies on two kernels called the Sampled dense-dense matrix product (SDDMM) and Sparse Matrix-Matrix multiplication (SPMM). Dr. Azad’s lab had developed a method called FusedMM, which take advantage of the fact that certain intermediate steps in the GNN pipeline can be shared, allowing for the fusion of the two kernels. My work was primarily exploring the best methods for implementing this fused kernel on GPU’s using CUDA.

Structure of message passing operations from original paper by Ariful Azad and Khaled Rahman