Data Compression and Vectorization of Matrix Multiplication on HXDSP

Abstract

In order to solve the problem of low efficiency of hardware resources and low data processing ability of vector processors, this paper uses data compression and vectorization method to realize matrix multiplication based on the HXDSP platform with the DCT algorithm in HEVC. It can make full use of the hardware resources of DSP to achieve the optimal optimization. The experimental results show that this method can achieve 32GMACS which is the peak-point multiply-accumulate capability of HXDSP. It can achieve to 2Gpixel/s for the data processing capability, which meets the performance requirements of HEVC coding standard and provides a reference for hardware implementation of HEVC.