FPGA Implementation of a Fixed-Point Arctangent Function Using Harmonized Parabolic Synthesis

Harmonized Parabolic Synthesis (HPS) is a method that combines parabolic synthesis with seconddegree interpolation. In this work, we implement an arctangent function for Neural Networks on FPGAS, with high accuracy using this method (error $\lt 1$ ULP). Experiments are carried out for the activation function on a Xilinx FPGA, and the implementation demonstrates lower resource usage, lower latency, high clock rates, and higher throughput compared to other implementations, such as CORDIC. The method is implemented in unsigned fixed-point format (FxP) of 16 bits and is suitable for highperformance embedded systems due to its small size, low latency, and high throughput. It can be used as a standalone accelerator or integrated into an ALU or FPU.