基于GPU的稀疏矩阵压缩存储格式研究  

Sparse matrix compressed storage format based on GPU

作　　者：陈闽昊 边浩东 Chen Minhao;Bian Haodong(School of Computer Technology and Application,Qinghai University,Xining 810016,China)

机构地区：[1]青海大学计算机技术与应用学院,青海西宁810016

出　　处：《电子技术应用》2024年第11期1-8,共8页Application of Electronic Technique

摘　　要：稀疏矩阵向量乘法(Sparse Matrix-Vector Multiplication,SpMV)是矩阵数值计算领域重要的线性代数子程序。通过对SpMV算法的负载均衡以及访存频度这两个关键性能瓶颈的研究,提出了一种VCSR(Vectorized Com‐pressed Sparse Row)稀疏矩阵压缩存储格式。该格式根据各行非零元素分布的统计特性调整各个线程的数据负载来防止线程发散的问题,并且基于快速分段求和的策略以及使用矢量化的方法来提高SpMV流程的计算性能。通过使用佛罗里达大学的稀疏矩阵作为测试集,在GPU上进行性能测试,获得了相较CSR5(Compressed Sparse Row 5)格式平均10%到30%,最高50%的性能提升。Sparse Matrix-Vector Multiplication(SpMV)is an important linear algebraic subroutine in Matrix numerical computa‐tion.Vectorized Compressed Sparse Row(VCSR)sparse matrix compression format is proposed by studying the load balancing and memory access frequency of SpMV algorithm.This format adjusts the data load of each thread according to the statistical characteristics of the distribution of each line of non-zero elements to prevent the problem of thread divergence,and improves the computational performance of SpMV flow based on the strategy of fast segmented summation and the vectorization method.By using the Sparse matrix of the University of Florida as the test set,the performance of the GPU is tested,and the average perfor‐mance improvement is 10%to 30%,and the maximum performance is 50%compared to the CSR5(Compressed Sparse Row 5)format.

关 键 词：稀疏矩阵向量乘 负载均衡 存储格式 分段求和 浮点运算 矢量化 图形处理器 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]