Rounding in Redundant Digit Floating Point Systems

摘要

Redundant representations are used to increase the performance of arithmetic units. If redundancy is eliminated, the bits needed to represent a number may increase or decrease depending on the type of redundancy used. In such a redundant representation, finding the exact location and correct decision for rounding without eliminating the redundancy or loosing its performance gains is difficult. This paper discusses the different issues related to rounding in redundant systems. It also presents a solution that was used to maintain the gains of redundancy in a floating point unit while correctly implementing the IEEE rounding modes.