高性能稀疏矩阵向量乘的程序设计综述

A survey of high-performance sparse matrix-vector multiplication programming

稀疏矩阵向量乘(SpMV)广泛应用于科学计算、图计算、数据分析等领域,是自现代计算机诞生以来经久不衰且挑战依旧的研究热点。本文系统回顾了20世纪70年代以来稀疏矩阵向量乘程序设计的发展脉络和各阶段的代表性工作;分析比较了这一领域4条技术路线,即人工程序设计、自动调优器、稀疏编译器和自动程序设计器,在当今的流行方法;并在此基础上对高性能稀疏矩阵向量乘程序设计的研究趋势做出预测,力图给学习者和研究者带来有益的知识与启示。

Sparse matrix-vector multiplication(SpMV)are fundamental operations in scientific computing,graph compu-tation,and data analysis.They have been an enduring and challenging research topic since the birth of modern computing.This paper systematically reviews the development of SpMV from 1970s and the representative work at each stage.It analyzes and compares four technical routes in this field:manual programming,automatic tuners,sparse compilers,and automatic programmers.These are the popular approaches today.On this basis,the paper makes predictions on the future trends of research on SpMV programs.It aims to provide useful insights to learners and researchers.