对角线稀疏矩阵的SpMV自适应性能优化.pdf

计箅机研 究 与发展 ISSNCN 111777／1’P JournalofCompulerResearchandDevelopment 50(3)：648—656，2O13 对角线稀疏矩阵的SpMV 自适应性能优化 孙相征 张云泉。。 王 婷 李 焱 袁 良 ’。 (中同科学院软件所并行软件与计算科学实验室 北京 100190) (计算机科学国家重点实验室 (中周科学 院软件研究所) 北京 100190) (中国科学院大学 北京 l00190) (xiangzheng．sun@ gmail．corn) Auto—TuningofSpM V forDiagonalSparseM atrices SunXiangzheng，。 。ZhangYunquan一，W angTing一，LiYan ，。。。，andYuanLiang’’。 (Laboratoryo，ParallelSoftwareand ComputationalScience，Instituteo、Software，ChineseAcademyu_『Sciences， Beijing100190) (StateKeyLaboratory(jl，ComputerScience(InstitateoJSo，tzaare，ChineseAtademyo，Sciences)，Beijing100190) (Uniuersityol，ChineseA(ademyofSciemes，Beijing100190) Abstract SpMV (sparse matrix—vector multiplication) is an important computationalkernel in scientificapplications． Its performance is mainly affected by the nonzero distribution of sparse matrices．Inthispaper，weproposeanew storageformatfordiagonalsparsematrices，calledCRSD (compressedrow segmentedwith diagonal—pattern)．W edesign adiagonalpattern torepresentthe diagonaldistribution．Unlike generalpurposemethods，ourCRSD basedSpMV implementation is applicationspecificsinceitneedstosamplethesparsematrix ofonegiven application．Theoptimal implementationisselectedautomaticallyforagivenhardwareplatform duringthesoftwareinstallation phase．Theinformationcollectedduringauto—tuningprocessisalsoutilizedforparallelimplementation to maintain load—balance．Theimplementationisalsotunedon GPU platform．Experimentalresults demonstratethatthespeedupreachesupto2．37X(1．7X onaverage)incomparisonwithDIA and4． 6X(2．1X onaverage)in comparison withCSR underthesamenumberofthreadsonIntelandAM D platforms． Keywords compressedrow segmentedwithdiagonal—pattern(CRSD)；auto—tuning SpMV ；diagonal— pattern；diagonalsparsematrix；GPU ；scientificapplication 摘 要 稀疏矩阵向量乘 (SpMV)是科学计算 中常用的 内核之一，其运行速率跟非