A GPU-Accelerated SVD Algorithm, Based on QR Factorization and Givens Rotations, for DWI Denoising

In this work, we present a parallel implementation of the Singular Value Decomposition (SVD) method on Graphics Processing Units (GPUs) using CUDA programming model. Our approach is based on an iterative parallel version of the QR factorization by means Givens plane rotations using the Sameh and Kuck scheme. The parallel algorithm is driven by an outer loop executed on the CPU. Therefore, threads and blocks configuration is organized in order to use the shared memory and avoid multiple accesses to global memory. However, the main kernel provides coalesced accesses to global memory using contiguous indices. As case study, we consider the application of the SVD in the Overcomplete Local Principal Component Analysis (OLPCA) algorithm for the Diffusion Weighted Imaging (DWI) denoising process. Our results show significant improvements in terms of performances with respect to the CPU version that encourage its usability for this expensive application.