Optimization Technology in SIMD Mathematical Functions Based on Vector Register Reuse

Abstract

To efficiently solve the problem low degree of SIMD (Single Instruction Multiple Data) vectorization caused by the lower speed of memory access and the lack of vector registers in the SIMD mathematical function library, an optimizing method based on vector register data reuse is introduced aiming at relative address access mode. Furthermore, to improve the access frequency of vector registers, a novel vector register instruction reuse method is proposed. We build these methods on the SIMD architecture equipped with n-dimensional vectors, and implement the corresponding theoretical feasibility analysis and experimental verification for each method. The experimental results show that these optimizing methods can significantly improve the efficiency of the functions and the access frequency of vector registers.