SparseP: Efficient Sparse Matrix Vector Multiplication on Real Processing-In-Memory Architectures

Sparse Matrix Vector Multiplication (SpMV) is one of the most thoroughly studied scientific computation kernels, be-cause it lies at the heart of many important applications from the scientific computing, machine learning, and graph analyt-ics domains. SpMV performs indirect memory references as a result of storing the sparse matrix in a compressed format, and irregular memory accesses to the input vector due to the spar-sity pattern of the input matrix [1]–[3]. Thus, in CPU and GPU systems, SpMV is a primarily memory-bandwidth-bound ker-nel for the majority of real sparse matrices, and is bottlenecked by data movement between memory and processors [3]–[6].