Accelerating sparse matrix computations on tensor cores

Sparse matrix computations such as sparse matrix-vector multiplication (SpMV) and sparse matrix-matrix multiplication (SpGEMM) play a key role in computational science and engineering, graph processing, and machine learning applications. Even though there has been much work devoted to optimizing sparse matrix computations, the latest major hardware features, i.e., tensor core units and their low precision compute power, have not been exploited sufficiently to accelerate sparse matrix computations. This talk will present our two recent studies: (1) DASP: a tensor core-accelerated SpMV algorithm, which makes the irregular data layout in sparse matrices regular for efficiently exploiting tensor core units; (2) AmgT: a new algebraic multigrid solver that utilizes the tensor cores and their mixed precision ability for SpGEMM and SpMV in the entire procedure of AMG. Our experiments show that the two studies break the limitations of the traditional GEMM computation pattern supported by tensor cores and exploit the specific dense matrix multiply-accumulate units to accelerate sparse kernels.