MM-FSM: A High-Efficiency General Nonlinear Function Generator for Stochastic Computation
Xincheng Feng, Ke Hu, Kaining Han
Abstract
AbstractNonlinear function calculation is widely used in numerous science and technology fields. Stochastic computation is a novel high-efficiency value representation and calculation scheme for information and signal processing. The main challenges of stochastic computation based nonlinear function implementation lie on poor generalization, low accuracy and high latency. In this paper, we propose a multiple driving and multiple dimension finite state machine (MM-FSM) to realize major single variable nonlinear functions used in information and signal processing areas on common platforms with low complexity and latency. We provide corresponding synthesis method of the activation parameters and conditional parameters of MM-FSM. In order to improve the calculation accuracy, we further propose an adaptive scaling algorithm for MM-FSM. The most salient feature of MM-FSM is that we can configure different types of nonlinear functions with the same MM-FSM structure. Thus, MM-FSM can be used in a wide range of stochastic based applications. Compared with the traditional stochastic scheme and Coordinate Rotation Digital Computer (CORDIC) algorithm, simulation results show that the proposed MM-FSM nonlinear function generator has significantly lower complexity while guaranteeing the calculation accuracy.