ApproxiMath: Approximating Math Functions with Polynomial Series to Improve Performance on Accurate Hardware
J. Castro Godinez, Tanfer Alan, Jörg Henkel · Zenodo (CERN European Organization for Nuclear Research) · 2022
We propose approximating transcendental math functions with polynomial Taylor series and Remez algorithm. These functions are by-definition approximation-tolerant, operating on floating-point data, and commonly used between applications. Although they occupy just a single code line, they are computed through a routine, taking a significant amount of cycles. We build towards an approximate math library with multiple accuracies that we obtain by changing the series order, and consequently reducing the number of instructions and required execution time on accurate, off-the-shelf processors. Our work is enabling a large and fine-grained design space with trade-offs between input range, accuracy, and performance.