Developing Tools for Floating-point Debugging

Floating-point arithmetic exhibits subtle behavior and the growing majority of users of numerical software are not experts in roundoff error analysis. Even for experts, locating the source of floating-point exceptions and numerical instabilities in real world programs can be difficult. Further, many of the hardware features required or suggested by the IEEE Standard 754 which aid in the debugging of numerical programs are either not accessible to programming languages or simply not implemented in hardware at all [Kah06]. The objective in this project is to help develop a tool for debugging floating-point programs by using the techniques of “delta debugging” [ZH02] to localize numerically sensitive operations.