摘要
使用CUDA平台,提出在通用图形处理器(GPGPU)上实现并行的全选主元、归一和消去等操作,加速实现并行全选主元高斯-约当消去法求解线性方程组的一种基本方法。该方法在CPU上完成解向量的恢复。根据NVIDIA公司最新Fermi架构图形处理器的特点,通过一系列的优化设计,使通用GPGPU相对Intel最新架构CPU的加速比超过了6.5倍,比Intel上一代CPU的加速比超过了10倍。
On CUDA platform,an elementary method is proposed to implement parallel complete pivoting,normalization,elimination and other operations on Generic Purpose Graphic Process Unit（GPGPU） in order to accelerate the implementation of parallel complete pivoting Gauss-Jordan elimination method to solve linear equations.The method recovers the resolved vector on CPU.Relying on the characteristics of NVIDIA＇s latest Fermi architecture GPU,after a series of optimization design,the accelerating ratio of GPGPU is 6.5 times higher than that of Intel＇s latest architecture CPU,and 10 times higher than that of Intel＇s last generation of CPU.
出处
《计算机应用与软件》
CSCD
2011年第9期269-271,共3页
Computer Applications and Software
基金
国家大学生创新性实验计划项目(101025537)
关键词
CUDA
并行计算
通用图形处理器
全选主元高斯-约当消去法
Compute unified device architecture（CUDA） Parallel computing Generic purpose graphic process unit（GPGPU） Complete pivoting Gauss-Jordan elimination method
