Optimization of SVD over Graphic Processor

Abstract

Many of the engineering applications employ linear algebra to furnish the analysis. Image Processing deals with Eigen values/vectors, Machine Design requires principal component analysis of stress, Statistics and Data compression requires minimization of dimensionality in data. Singular Value Decomposition serves as the answer to all these varied needs. This method alone serves many computational & analytical purposes. Although the computation of SVD of a matrix is bulky, the process involves a sequence of vector operations. This makes it a good candidate for parallelization of over Graphic Processors. This paper proposes parallelization of SVD modules in LAPACK over GPGPU using OpenCL. OpenCL is crucial for making the implementation platform independent. Narayanan[1] too considers parallelization of SVD over CUDA using CUBLAS. This work proposes a scheme which is platform independent, and focuses on routines beyond BLAS.