Exploiting fast hardware floating point in high precision computation
Keith O. Geddes, Wei Wei Zheng · 2003
We apply an iterative refinement method based on a linear Newton iteration to solve a particular group of high precision computation problems. The method generates an initial solution at hardware floating point precision using a traditional method and then repeatedly refines this solution to higher precision, exploiting hardware floating point computation in each iteration. This is in contrast to direct solution of the high precision problem completely in software floating point. Theoretical cost analysis, as well as experimental evidence, shows a significant reduction in computational cost is achieved by the iterative refinement method on this group of problems.