Abstract

The work presented in this paper was initiated as part of a study on software alternatives to the hardware implementations of floating-point operations such as divide and square root. The results of the study proved the viability of software implementations, and showed that certain proposed algorithms are comparable in performance to current hardware implementations. This paper discusses two components of that study: (1) A methodology for proving the IEEE correctness of the result of iterative algorithms that implement the floating-point square root, divide, or remainder operation. (2) Identification of operands for the floating-point divide and square root operations that lead to results representing difficult cases for IEEE rounding. Some general properties of floating-point computations are presented first. The IEEE correctness of the floatingpoint square root operation is discussed next. We show how operands for the floating-point square root that lead to difficult cases for rounding can be generated, and how to use this knowledge in proving the IEEE correctness of the result of iterative algorithms that calculate the square root of a floating-point number. Similar aspects are analyzed for the floating-point divide operation, and we present a method for generating difficult cases for rounding. In the case of the floating-point divide operation, however, it is more difficult to use this information in proving the IEEE correctness of the result of an iterative algorithm than it is for the floating-point square root operation. We examine the restrictions on the method used for square root. Finally, we present possible limitations due to the finite exponent range.

Citations

505	Mathematica: A System for Doing Mathematics by Computer - Wolfram - 1988
33	An Introduction to Number Theory - Stark - 1978