基于GPU的矩阵求逆性能测试和分析_刘丽.pdfVIP

( ) Vol. 36 No.6 812 Journal of East China University of Science and T echnology ( Natural Science Edition) 2 1-12 : 1 6-3 8 ( 2 1 ) 6- 812- 6 GPU 1 2 2 刘 丽, 沈 杰 , 李洪林 ( 华东理工大学 1.信息科学与工程学院; 2. 药学院, 上海2 237) : 在CPU 串行运算模式下实 大规模矩阵求逆是一个非常耗时的过程为了解决这 一 问题, 基于NVIDIA 公司专为GPU( 图形处理器) 提供的CU DA(计算统 一设备架构) , 从新的编程 角度出发, 利用GPU 多线程并行处理技术, 将矩阵求逆过程中大量的数据实 并行运算, 从而获 得了较大的加速比同时, 根据程序的执行结果, 分析了GPU 的单精度与双精度的浮点运算能力 及其优劣势最后, 通过分析数据传输时间对GPU 性能的影响, 总结出适合GPU 的算法特征 : 图形处理器( GPU ) ; 计算统一设备架构( CUDA) ; CPU; 并行运算; 矩阵求逆 : TP311. 11 : A Performance Testing and Analysis for Matrix Inversion Base on GPU 1 2 2 LI U L i , SH EN J ie , L I H ong-lin ( 1. S chool of I nf or mation S cience and Eng ineer ing ; 2. School of Phar macy , E ast China Uni ersity of S cience and T echnology , S hanghai 2 237, China) Abstract: For the CPU serial operation mode, it is a very time-consuming process to obtain the inverse of large-scale matrix. Aiming at the above shortcoming, this paper proposes a new programming method based on the common platform CU DA for GPU designed by NV IDIA. By using the mult-i threaded parallel processing technology of GPU, a large scale of data during solving the inverse matrix are parallelly computed such that a higher speedup may be obtained. Moreover, both the single- precision and the double- precision FLOPS of GPU are analyzed according to the results of this program . Finally , some characteristics of the proposed algorithms are summarized by analyzing the effect of the data transmission time on th