Non-stationary wave relaxation methods for general linear systems of Volterra equations: convergence and parallel GPU implementation (Q6145568)

This paper is concerned with a parallel-in -in -time discretization of linear systems of Volterra Integral Equations (LVIEs) appearing rather frequently in problems of image and video processing, for their numerical solutions. Parallel implementation of the NSWR method for time discretization is proposed and its convergence is studied as well. The problem statement in a functional setting related to the analytical solution is stated first as a theoretical background. Richardson discrete time-point NSWR methods are the subject of the author's study in the paper; particularly if compared with sequential implementation. Purpose (aims and objectives): The purpose of this paper is to employ graphics processing units (GPUs) for the numerical solution of large systems of Linear Volterra Integral Equations (LVIEs), by means of non-stationary waveform relaxation (NSWR) methods in the parallel computation environment. Motivation for the work: The motivation for the present work comes from previously published work related to VIEs by the first and second authors [38 to 40 and 43] and the references therein. The literature survey relevant to the published work by the authors reveals that the numerical solution of large system of LVIEs presents formidable difficulty in terms of computational cost. The issue as pointed out by the authors, is addressed in two ways, Fast and efficient in the framework of convolution quadrature exist already in the literature. But any attempt as pointed out by the authors, to improve the velocity of problem resolution is crucial. More specifically, the motivation for the present work stems from exploration of the possibility of extending the previous approaches in the parallel computing environment for the system of linear integral equations of fractional type to a more general class of systems linear convolution equations and at the same time developing faster and more efficient methods of computation. Research findings: Superiority of parallel approach using NSWR methods by means of GPU is established by test problems in comparison with sequential implementation. Moreover, the convergence analysis confirms the goodness of the proposed approach by demonstrating experimental verification with what is theoretically established. According to the results presented; the test problems provide considerable improvement in performance from executing sequentially as well as parallelly. Research methodology and technical resources involved: The implementation of the parallel numerical scheme proposed for time discretization for systems of LVIEs is theoretically analysed. Details related to the algorithm and technical resources involved in the practical implementation of the algorithm are discussed. The proposed approach exploits a SIMD methodology according to Flynn taxonomy. This methodology takes advantage of GPU and allows using them for general purpose through CUDA. The proposed method implementation through CUDA is based on a one-dimensional grid and one-dimensional blocks. Moreover, the technical resources involved in obtaining numerical results are through Google Colab, among others. Research limitations and relevance: As highlighted by the authors, the paper focuses on the accuracy and speed-up analysis of the proposed approach for general purposes, but not for particular applications. The application of this approach in the field of image and video processing represents future development of the present work as envisaged by the authors shows its relevance to practical applications. Main contribution of the paper: The theoretical analysis related to the convergence of the proposed algorithms allows one to ensure well-behaviour of such algorithms in practical instances, in the opinion of the authors. This is amply demonstrated by numerical experiments performed on test problems of practical interest thereby concluding that reaching accurate numerical solution is feasible with much lower computational effort than with traditional methods /sequential approaches. In summary, the paper focuses on the accuracy and speed-up analysis of the proposed parallelized approach for linear systems of general purpose VIEs. Additionally, the numerical results presented through tables and summarized using computer generated plots are in perfect agreement with theoretical results analyzed related to the convergence. In other words, the empirical estimation of convergence order agrees with the theoretically analyzed results.