Optimized Math Functions for a Fixed-Point DSP Architecture

This paper presents a method for implementing several high-performance math functions through polynomial approximation on the fixed-point Blackfin ADSP-BF533 architecture. We present a strategy to overcome the performance given by the Blackfin's C library using a fast emulated floating-point format. We also discuss the methods applied to generate the polynomial approximation for the sine, logarithmic, and exponential functions, as well as the optimization schemes used to implement these functions. This work contributed to a maximum cycle reduction of 85% over the standard math library.