Abstract: In applications of tracking,singular value decomposition(SVD) is often used as a basic tool for dynamic library construction.However,facing the large amount of dataset per second and high accuracy requirements,the calculation of SVD is too time-consuming to satisfy the real-time constraint in the applications.This paper put forward the CPU-GPU collaborative method of parallel singular value decomposition.It utilizes the feature of asynchronous execution of GPU,and organizes the process of SVD into pipeline-style,with which we can largely exploit the parallelism.Experiments show that this method outperforms better than normal parallelized one-sided Jacobi method for SVD on GPU by 20%.Compared with the Intel MKL SGESVD on CPU,out approach can achieve a 6.8x performance improvement,which makes it satisfy the requirements of real-time applications.

Key words: GPU,Collaborative computing,Jacobi,SVD