The design and use of a floating-point (software) simulator for testing the arithmetic behavior of mathematical software

An important aspect of any evaluative procedure for developing high quality mathematical software is testing the effects of arithmetic behavior on algorithmic implementations. This paper describes a proposed design approach and various applications of a high-level language floating-point simulator which has two inputs: the program to be tested and a description of the floating-point arithmetic under which the routine is to be executed. A brief discussion of the motivation for this approach is given along with a review of existing efforts to study the influences of computer arithmetic on the accuracy and reliability of mathematical software. An overview of the simulator's structure is presented as well as suggestions for experiments to assist in determining the effects of floatingpoint behavior across several different computer architectures. Present and future uses of the simulator are also indicated.