TC-GVF: Tensor Core GPU-Based Vector Fitting via Accelerated Tall-Skinny QR Solvers

Abstract

QR decomposition and solution of linear least-squares-based large system of equations form the backbone of computational flow in many scientific applications. Usually, these account for the bulk of the computational cost in these applications, such as in vector fitting (VF) methods, which are widely used for system identification via rational function approximation from tabulated data of high-speed modules. Since the VF algorithm is iterative in nature, minimizing its computational cost and increasing its parallel efficiency on mixed CPU and GPU environments are critical in reducing the time needed for each iteration. In this article, a novel tensor core-based QR (TC-QR) decomposition method and tensor core-based linear least-squares-based solver (TC-LLS) are introduced to speed up the computationally expensive steps of QR factorization and solution to a set of linear least-squares equations, exploiting the emerging GPU platforms with tensor core (TC) architectures. These modules are utilized in developing the TC GPU-based VF (TC-GVF) algorithm, providing significant speedup compared with the state-of-the-art GVF implementations in the literature.