Research on tridiagonal matrix solver design based on a combination of processors. (August 2017)

• Record Type: Journal Article
• Authors: Li, Jingmei; Zheng, Zhigao; Tian, Qiao; Zhang, Guoyin; Zheng, Fangyuan; Pan, Yuanyuan
• Is Part Of: Computers & electrical engineering. Volume 62(2017)
• Page Start: 1
• Page End: 16
• Publication Date: 2017-08
• Subjects: Tridiagonal matrix solvers -- Heterogeneous systems -- Central processing unit -- Graphics processing unit -- Extended algorithms; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2017.07.014
• Languages: English
• ISSNs: 0045-7906

Abstract

Abstract: Large-scale tridiagonal matrix solvers based on heterogeneous systems currently cannot balance computational efficiency and numerical stability when solving a non-diagonally dominant matrix. A tridiagonal solver combined central processing unit with graphics processing unit is proposed, based on SPIKE 2 as a solver framework, a simplified SPIKE algorithm as a central processing unit component, and a diagonal pivot algorithm as a graphics processing unit component. The solver performance is further improved by using a data-layout-transformation mechanism to obtain continuous addresses, reducing memory communication using constant memory to store unchanged data in the calculation process, and employing a kernel-fusion mechanism to reduce power consumption of graphics processing unit. For a diagonally dominant matrix, extended Thomas algorithms and cycle reduction to replace the graphics processing unit component are proposed in the solver. Experimental results show that the tridiagonal matrix solver in this paper can effectively consider both numerical stability and computational efficiency, and reduce total power consumption while improving memory efficiency.