自动向量化:近期进展与展望  

Auto-vectorization:recent development and prospect

作　　者：冯竞舸 贺也平[1,2,3] 陶秋铭 FENG Jingge;HE Yeping;TAO Qiuming(National Engineering Research Center for Fundamental Software,Institute of Software Chinese Academy of Sciences,Beijing 100190,China;Graduate University,University of Chinese Academy of Sciences,Beijing 100049,China;China State Key Laboratory of Computer Science,Institute of Software Chinese Academy of Sciences,Beijing 100090,China)

机构地区：[1]中国科学院软件研究所基础软件国家工程研究中心,北京100190 [2]中国科学院大学研究生院,北京100049 [3]中国科学院软件研究所计算机科学国家重点实验室,北京100090

出　　处：《通信学报》2022年第3期180-195,共16页Journal on Communications

基　　金：中国科学院战略性先导科技专项基金资助项目(No.XDA-Y01-01,No.XDC02010600)。

摘　　要：随着单指令流多数据流(SIMD)技术的迅速发展,近年来许多面向SIMD扩展部件的自动向量化编译方法被提出,有效缓解了程序员手写向量程序的压力,并发挥了SIMD扩展部件的加速效能。基于此,分析总结了自动向量化领域近10年的研究成果,从保义分析和变换、向量化分组分析和变换、面向处理器支持特性的分析和变换以及性能评估分析这4个方面分类归纳了自动向量化的关键问题和主要突破,进而对4个方面的发展趋势和研究方向进行了展望。The technology of SIMD is developing rapidly,and quite a few auto-vectorization methods have been pro-posed.Auto-vectorization can automatically translate scalar programs into vector programs based on SIMD extension,decrease workload of the programmers in coding vector programs,and effectively improve performance of programs.Based on that,the research achievements in the field of automatic vectorization in recent 10 years were analyzed and summarized.The key problems and major breakthroughs in automatic vectorization were classified from four aspects:semantic-maintaining analysis and transformation,vectorization grouping analysis and transformation,processor-oriented analysis and transformation,and performance evaluation analysis.Furtherly,the development trends and research direc-tions of the four aspects were prospected.

关 键 词：自动向量化 SIMD扩展 编译技术 数据级并行 性能优化 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]