An improved micro-architecture for function approximation using piecewise quadratic interpolation

Shai Erez*, Guy Even

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

We present a new micro-architecture for evaluating functions based on piecewise quadratic interpolation. The micro-architecture consists mainly of a look-up table and two multiply-accumulate units. Previous micro-architectures based on piecewise quadratic interpolation have been shown to be efficient for small precision (e.g., single precision) computations. Moreover, they are as fast as piecewise linear interpolation while requiring smaller tables. Our main contribution is in circumventing the need for the additional squaring unit that appears in previous micro-architectures. Based on the proposed micro-architecture, we present a detailed design of single precision reciprocal approximation (1/x) . Our design is based on two multiply-accumulate units that contain truncated Booth radix 4 multipliers. The number of partial products in this design is reduced by over 20% compared to previous designs using quadratic interpolation. The latency of this design is roughly the delay of 19 full-adder gates, and it can be easily pipelined into two stages each with a delay of 10 full-adder gates.

Original languageEnglish
Title of host publication26th IEEE International Conference on Computer Design 2008, ICCD
Pages422-426
Number of pages5
DOIs
StatePublished - 2008
Event26th IEEE International Conference on Computer Design 2008, ICCD - Lake Tahoe, CA, United States
Duration: 12 Oct 200815 Oct 2008

Publication series

Name26th IEEE International Conference on Computer Design 2008, ICCD

Conference

Conference26th IEEE International Conference on Computer Design 2008, ICCD
Country/TerritoryUnited States
CityLake Tahoe, CA
Period12/10/0815/10/08

Fingerprint

Dive into the research topics of 'An improved micro-architecture for function approximation using piecewise quadratic interpolation'. Together they form a unique fingerprint.

Cite this