On the Impact of Numerical Accuracy Optimization on General Performances of Programs

Abstract

The floating-point numbers used in computer programs are a finite approximation of real numbers. In practice, this approximation may introduce round-off errors and this can lead to catastrophic results. In previous work, we have proposed intraprocedural and interprocedural program transformations for numerical accuracy optimization. All these transformations have been implemented in our tool, Salsa. The experimental results applied on various programs either coming from embedded systems or numerical methods, show the efficiency of the transformation in terms of numerical accuracy improvement but also in terms of other criteria such as execution time and code size. This article studies the impact of program transformations for numerical accuracy specially in embedded systems on other efficiency parameters such as execution time, code size and accuracy of the other variables (these which are not chosen for optimization).