稀疏矩阵向量乘的自动调优  

Auto-tuning for Sparse Matrix-vector Multiplication

作　　者：杜臻 谭光明[2] DU Zhen;TAN Guangming(School of Computer and Control Engineering,University of Chinese Academy of Sciences,Beijing 101408,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院大学计算机科学与技术学院,北京101408 [2]中国科学院计算技术研究所,北京100190

出　　处：《计算物理》2024年第1期33-39,共7页Chinese Journal of Computational Physics

基　　金：国家自然科学基金杰出青年基金项目(T2125013)资助。

摘　　要：分析稀疏矩阵向量乘(SpMV)程序优化的难点,介绍两个自动调优的代表性工作:基于预实现模板的SMAT和从头设计程序的AlphaSparse。详细介绍了它们的设计思路、实现细节、测试结果以及各自的优缺点。最后,对SpMV自动调优的发展趋势进行了分析和预测。SpMV(sparse matrix-vector multiplication)is a widely used kernel in scientific computing.Since the performance of specific SpMV program is closely related to the distribution of non-zero elements in sparse matrices,there is no universal SpMV program design that can achieve high performance in all matrices.Therefore,auto-tuning has become a popular method for high SpMV performance.This paper analyzes the difficulties in optimizing SpMV and introduces two representative works of auto-tuning:SMAT,which is based on pre-implemented templates and AlphaSparse which designs SpMV programs from scratch.This paper introduces their designs,implementations,test results,advantages,and disadvantages.Finally,the trend of SpMV auto-tuning is analyzed and predicted.

关 键 词：高性能科学计算 稀疏矩阵 自动调优 稀疏矩阵向量乘 

分 类 号：O469[理学—凝聚态物理]