RISC-V向量指令集的Yolov3移植优化  

Porting Optimization of Yolov3 Based on RISC-V Vector Instruction Set

作　　者：王宇木 潘志铭[1] 吴鹏飞 付维 田乐兰 李桂润 孙轶群 Wang Yumu;Pan Zhiming;Wu Pengfei;Fu Wei;Tian Lelan;Li Guirun;Sun Yiqun(School of Electronic and Information Engineering,Shenzhen University,Shenzhen 518000,China;UC TECH IP)

机构地区：[1]深圳大学电子与信息工程学院,深圳518000 [2]深圳优矽科技有限公司

出　　处：《单片机与嵌入式系统应用》2021年第12期20-25,30,共7页Microcontrollers & Embedded Systems

摘　　要：为研究SIMD在嵌入式领域中对处理器性能的提升效果,选择一种并行化程度较高的图像处理算法Yolov3进行SIMD向量化移植。根据开源指令集RISC-V扩展指令集中的V(Vector)指令集修改Yolov3算法的代码,将其部署到优矽科技自研的WH64处理器的VPU(Vector Processor Unit)中验证;结合Amdahl定律和Yolov3自测程序评估SIMD算法提升的性能。实验结果表明,在Xilinx的Kintex7板上以50 MHz主频运行,在向量化算法占比90%以上时,SIMD处理过后的代码程序达到了标量计算2.25x的加速比。In order to study the performance improvement of SIMD in embedded field,Yolov3 is selected to carry out SIMD vectorization porting.According RISC-V Vector extension instruction set,the Yolov3 algorithm code is modified and deployed to the VPU(Vector Processing Unit)of WH64 processor researched by UC TECHIP.Combining the Amdahl’s law and Yolov3 self-test programs to evaluate the improving performance of SIMD algorithm.The experiment results show that within the frequency of 50MHz used on Xilinx kintex7 board,and the proportion of the directional quantization algorithm more than 90%,the Yolov3 program after SIMD optimized reaches the acceleration ratio of 2.25x scalar calculation.

关 键 词：SIMD优化 Yolov3算法移植 RISC-V向量指令集 AMDAHL定律 WH64处理器 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]