计算机科学 ›› 2015, Vol. 42 ›› Issue (Z6): 549-552.
周伟,戴宗友,袁广林,陈萍
ZHOU Wei, DAI Zong-you, YUAN Guang-ling and CHEN Ping
摘要: 在目标跟踪应用中,常常采用奇异值分解(SVD)作为基本工具进行动态建库。然而当每秒处理的数据量较大、计算精度要求较高时,SVD的计算耗时往往无法满足应用的实时性能要求。针对这一问题,提出了CPU-GPU协同计算的并行奇异值分解方法。该方法利用GPU与CPU间的异步执行,对奇异值分解过程进行划分从而构造软件流水线,大大挖掘软硬件的并行性。实验表明,该方法比一般的基于GPU的Jacobi方法有约23%的性能提升。相对于CPU上的Intel MKL的奇异值分解函数获得了6.8x的加速比,满足了应用中的实时性能要求。
[1] Golub G,Van Loan C.Matrix Computations.Johns Hopkins University Press(2nd edition)[M].Baltimore,MD,1989 [2] Gu Ming,Demmel J,Dhillon I.Eficient Computation of the Singular Value Decomposition with Applications to Least Squares Problems[R].Technical Report CS-94-257.Department of Computer Science,University of Tennessee,October 1994 [3] Hestenes M R.Inversion of matrices by biorthogonalization andrelated results[J].JSoc.Indust.Appl.,1958,6(1):51-90 [4] Kotas C,Barhen J.Singular value decomposition utilizing parallel algorithms on graphical processors[C]∥OCEANS 2011.Publication Year,2011:1-7 [5] Lahabar S,Narayanan P J.Singular Value Decomposition onGPU using CUDA[C]∥IEEE International Parallel Distributed Processing Symposium.Hyderabad,2009:1-10 [6] Novakovi V,Singer S.A GPU-based hyperbolic SVD algorithm[M].BIT 51,2011:1009-1030 [7] Liu Ding,et al.A divide-and-conquer approach for solving singular value decomposition on a heterogeneous system[C]∥Proceedings of the ACM International Conference on Computing Frontiers.ACM,2013 [8] Brent R P,Luk F.T.The solution of Singular Value and Symmetric Eigen problems on Multiprocessor Arrays[J].Sct Stat Comput,1985,6:69-84 [9] 张舒,窦衡.基于CUDA的矩阵奇异值分解[J].计算机应用研究,2007,24(6) [10] http://software.intel.com/en-us/intel-mkl/ |
No related articles found! |
|