计算机工程与应用 ›› 2009, Vol. 45 ›› Issue (34): 53-55.DOI: 10.3778/j.issn.1002-8331.2009.34.017

• 研发、设计、测试 • 上一篇    下一篇

雅可比迭代法在图形处理器上实现的研究

张 健1,2,涂永明3,涂晓明4   

  1. 1.南京工程学院 通信工程学院,南京 211167
    2.东南大学 电子科学与工程学院,南京 210096
    3.东南大学 混凝土及预应力混凝土结构教育部重点实验室,南京 210096
    4.中国人民解放军 91960部队 71分队
  • 收稿日期:2009-04-20 修回日期:2009-05-26 出版日期:2009-12-01 发布日期:2009-12-01
  • 通讯作者: 张 健

Investigation of implement of Jacobi iteration on GPU

ZHANG Jian1,2,TU Yong-ming3,TU Xiao-ming4   

  1. 1.School of Communication,Nanjing Institute of Technology,Nanjing 211167,China
    2.School of Electronic Science and Engineering,Southeast University,Nanjing 210096,China
    3.Key Lab of Concrete and Prestressed Concrete Structure of Ministry of Education,Southeast University,Nanjing 210096,China
    4.71 Unit of 91960 Army Force of the Chinese PLA
  • Received:2009-04-20 Revised:2009-05-26 Online:2009-12-01 Published:2009-12-01
  • Contact: ZHANG Jian

摘要: 雅可比迭代法是求解大型线性方程组的基本方法。利用GPU(Graphics Processing Unit,图形处理器)的并行处理能力,将雅可比迭代求解线性方程组过程中运算量较大的部分移植到GPU上执行,以提高运算速度。并分析了影响运算速度的两个因素:CPU-GPU数据交换和共享变量的访问;实验结果表明采用单个thread访问共享变量判断迭代是否收敛时,线性方程组的阶数为500,速度可以提高45倍以上。

关键词: 雅可比迭代, 图形处理器, CPU-GPU数据交换, 共享变量

Abstract: Jacobi iteration is the basic method to solve the large-scale linear equations.In this paper,the parts of Jacobi iteration,which needs a lot of time for calculation,are transplanted on Graphics Processing Unit(GPU) to run,in order to improve the speed to solve the large-scale linear equations with the parallel processing capabilities of GPU.The two factors,which effect the speed,are analyzed:The data exchange between CPU and GPU and the accessing of shared memory.The experiment shows when the order of linear equations is 500,the speed of calculation improves more than 45 times.

Key words: Jacobi iteration, Graphics Processing Unit(GPU), data exchange between CPU and GPU, shared memory

中图分类号: