HA-SpMM: A Hybrid Feature-Based Adaptive SpMM Algorithm on GPU

Sparse Matrix-Matrix Multiplication (SpMM) plays a key role in practical applications such as machine learning and neural networks. With the rapid development of parallel computing, accelerating SpMM on GPUs has become a current research focus. However, existing acceleration methods have not simultaneously considered the effects of matrix sparsity and GPU hardware parameters. In this article, we propose an efficient adaptive SpMM algorithm called HASpMM, which can automatically select suitable storage formats and load-balancing algorithms based on the hybrid feature of matrix sparsity and GPU hardware parameters. We first use CSR or COO format to store the matrix, and then we implement the load-balancing SpMM algorithm. Finally, we design a novel framework based on the hybrid feature to select the optimal storage format and SpMM algorithm for matrices. The experimental results demonstrate the effectiveness of our method. On all matrices in the Suite Sparse Matrix Collection, our method performs better than state-of-the-art SpMV methods (i.e. cuSPARSE SpMM, ASpT, GE-SpMM, and HPSpMM) on more than half of the matrices.