Three-precision iterative refinement with parameter regularization and prediction for solving large sparse linear systems

TL;DR

This paper introduces GADI-IR, a mixed-precision iterative refinement method that uses low-precision arithmetic and parameter prediction to efficiently solve large sparse linear systems with high accuracy.

Contribution

It develops a novel mixed-precision iterative refinement algorithm incorporating regularization and Gaussian process regression for improved efficiency and robustness.

Findings

Accelerates convergence using FP16 arithmetic.

Maintains high accuracy with regularization and backward error analysis.

Reduces computational time through low-precision parameter prediction.

Abstract

This study presents a novel mixed-precision iterative refinement algorithm, GADI-IR, within the general alternating-direction implicit (GADI) framework, designed for efficiently solving large-scale sparse linear systems. By employing low-precision arithmetic, particularly half-precision (FP16), for computationally intensive inner iterations, the method achieves substantial acceleration while maintaining high numerical accuracy. Key challenges such as overflow in FP16 and convergence issues for low precision are addressed through careful backward error analysis and the application of a regularization parameter $\alpha$. Furthermore, the integration of low-precision arithmetic into the parameter prediction process, using Gaussian process regression (GPR), significantly reduces computational time without degrading performance. The method is particularly effective for large-scale linear systems arising from discretized partial differential equations and other high-dimensional problems, where both accuracy and efficiency are critical. Numerical experiments demonstrate that the use of FP16 and mixed-precision strategies not only accelerates computation but also ensures robust convergence, making the approach advantageous for various applications. The results highlight the potential of leveraging lower-precision arithmetic to achieve superior computational efficiency in high-performance computing.