作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2019, Vol. 45 ›› Issue (4): 25-29. doi: 10.19678/j.issn.1000-3428.0050025

• 体系结构与软件技术 • 上一篇    下一篇

HXDSP平台上矩阵乘法的实现与性能分析

刘余福1a,1b,郎文辉1a,1b,贾光帅2   

  1. 1.合肥工业大学 a.计算机与信息学院; b.工业安全与应急技术安徽省重点实验室,合肥 230009; 2.中国电子科技集团公司第三十八研究所,合肥 230088
  • 收稿日期:2018-01-08 出版日期:2019-04-15 发布日期:2019-04-15
  • 作者简介:刘余福(1990—),男,硕士研究生,主研方向为数字信号处理技术、嵌入式系统;郎文辉,副教授;贾光帅,工程师。
  • 基金资助:

    安徽省科技专项资金“高清摄像机产业化及核心处理器芯片技术开发”(13Z02010)

Realization and Performance Analysis of Matrix Multiplication on HXDSP Platform

LIU Yufu 1a,1b,LANG Wenhui 1a,1b,JIA Guangshuai 2   

  1. 1a.School of Computer Science and Information Engineering; 1b.Anhui Province Key Laboratory of Industry Safety and Emergency Technology,Hefei University of Technology,Hefei 230009,China; 2.The 38th Research Institute of China Electronic Technology Group Corporation,Hefei 230088,China
  • Received:2018-01-08 Online:2019-04-15 Published:2019-04-15

摘要:

在向量处理器上进行矩阵运算时硬件资源利用率与数据处理能力较低。为此,基于魂芯数字信号处理器(HXDSP)平台,结合高效视频编码(HEVC)标准中的离散余弦变换算法,采用数据压缩式向量法实现矩阵乘法,以发挥向量处理器的硬件资源。实验结果表明,该方法可达到HXDSP的定点乘累加运算能力峰值32 GMACs,数据处理能力可达2 GPixel/s,满足HEVC编码标准的性能要求。

关键词: 魂芯数字信号处理器, 矩阵乘法, 数据压缩, 向量法, 高效视频编码标准, 离散余弦变换算法

Abstract:

Matrix operations on vector processors face the problems of low utilization of hardware resources and low data processing capacity.Therefore,based on HX Digital Signal Processor(HXDSP) platform,combined with Discrete Cosine Transform(DCT) algorithm in High Efficiency Video Coding(HEVC) standard,the matrix multiplication is realized by data compression vector method,in order to make full use of the hardware resources of vector processor.Experimental results show that this method can reach the peak value of 32 GMACs of fixed-point multiplication and accumulation operation ability of HXDSP,and the data processing ability can reach 2 GPixel/s,which meets the performance requirements of HEVC coding standard.

Key words: HX Digital Signal Processor(HXDSP), matrix multiplication, data compression, vector method, High Efficiency Video Coding(HEVC) standard, Discrete Cosine Transform(DCT) algorithm

中图分类号: