Bivariate function approximation with encoded gradients

Abstract

Efficient signal processing is one of the most important challenges for modern digital circuits and systems. For applications that base on complex algorithms, multivariate numeric functions with approximate computational accuracy have become a promising approach. Though, State-of-the Art approximation techniques perform quite well in terms of throughput, the area and energy requirements still need major improvement. In this paper we present a novel design technique for the efficient hardware realization of bivariate numeric functions. A sophisticated gradient encoding scheme is deployed that reduces the size of the multiplexer tree evidently. For evaluation, three different function approximations are generated and compared to actual references. The results highlight our approach to be a powerful extension for high-performance bivariate numeric function approximation.