Floating-Point Divide Operation without Special Hardware Supports

Takashi AMISAKI  Umpei NAGASHIMA  Kazutoshi TANABE  

IEICE TRANSACTIONS on Fundamentals of Electronics, Communications and Computer Sciences   Vol.E82-A   No.1   pp.173-177
Publication Date: 1999/01/25
Online ISSN: 
Print ISSN: 0916-8508
Type of Manuscript: LETTER
Category: Numerical Analysis and Optimization
computer arithmetic,  division,  FPU,  error analysis,  special-purpose,  

Full Text: PDF>>
Buy this Article

Three multiplicative algorithms for the floating-point divide operation are compared: the Newton-Raphson method, Goldschmidt's algorithm, and a naive method that simply calculates a form of the Taylor series expansion of a reciprocal. The series also provides a theoretical basis for Goldschmidt's algorithm. It is well known that, of the Newton-Raphson method and Goldschmidt's algorithm, the former is the more accurate while the latter is the faster on a pipelined unit. However, little is reported about the naive method. In this report, we analyze the speed and accuracy of each method and present the results of numerical tests, which we conducted to confirm the validity of the accuracy analysis. Basically, the comparison are made in the context of software implementation (e. g. , a macro library) and compliance with the IEEE Standard 754 rounding is not considered. It is shown that the naive method is useful in a realistic setting where the number of iterations is small and the method is implemented on a pipelined floating-point unit with a multiply-accumulate configuration. In such a situation, the naive method gives a more accurate result with a slightly lower latency, as compared with Goldschmidt's algorithm, and is much faster than but slightly inferior in accuracy to the Newton-Raphson method.