Report Number: CSL-TR-91-459
Institution: Stanford University, Computer Systems Laboratory
Title: On fast IEEE rounding
Author: Quach, Nhon
Author: Takagi, Naofumi
Author: Flynn, Michael J.
Date: January 1991
Abstract: A systematic general rounding procedure is proposed. This procedure consists of 2 steps: constructing a rounding table and selecting a prediction scheme. Optimization guidelines are given in each step to minimize the hardware used. This procedure-based rounding method has the additional advantage that verification and generalization are trivial. Two rounding hardware models are described. The first is shown to be identical to that reported by Santoro, et al. The second is more powerful, providing solutions where the first fails. Applying this approach to the IEEE rounding modes for high-speed conventional binary multipliers reveals that round to infinity is more difficult to implement than the round to nearest mode; more adders are potentially needed. Round to zero requires the least amount of hardware. A generalization of this procedure to redundant binary multipliers reveals two major advantages over conventional binary multipliers. First, the computation of the sticky bit consumes considerably less hardware. Second, implementing round to positive and minus infinity modes does not require the examination of the sticky bit, removing a possible worst-case path. A generalization of this approach to addition produces a similar solution to that reported by Quach and Flynn. Although generalizable to other kinds of rounding as well as other arithmetic operations, we only treat the case of IEEE rounding for addition and multiplication; IEEE rounding because it is the current standard on rounding, addition and multiplication because they are the most frequently used arithmetic operations in a typical scientific computation.
http://i.stanford.edu/pub/cstr/reports/csl/tr/91/459/CSL-TR-91-459.pdf