Sparse matrix-vector multiply on the Keystone II Digital Signal Processor

In this paper we describe an implementation of sparse matrix-vector multiply (SpMV) on the Texas Instruments (TI) Keystone II architecture. The Keystone II is an eight core Digital Signal Processor (DSP) that offers floating point performance comparable to a desktop CPU while having a power envelope comparable to a mobile embedded CPU. This, combined with its integrated communication interfaces, potentially make it a scalable and efficient HPC processor technology. For this architecture, the key to achieving high computational efficiency is the careful use of its on-chip scratchpad memory. SpMV is a HPC kernel that is both memory bounded and has an irregular memory access pattern. When tuning this kernel, we found that using scratchpad can provide as much as 50% improvement in effective memory bandwidth as compared to using cache, but only with careful scratchpad allocation and run-time management. This includes selection of tile size, the mapping of arrays to specific on-chip memory structures, and the methods by which the DMA is performed in parallel with computation.