Function Approximation based on Estimated Arithmetic Operators
Abstract
This article reports preliminary results on hardware operators for function evaluation with moderate average accuracy. The obtained operators provide a small average error on the input domain but a few large errors may occur for some very infrequent arguments. The proposed approximation method combines two methods: i) approximation using low-degree polynomials with small and sparse coefficients; ii) operators for addition based on estimated arithmetic and integer power operators (x^2, x^3) with reduced accuracy. Typical delay reduction is about 20-50% and 15-30% for silicon area reduction compared to previous results.