A Manycore Processor Based Multilayer Perceptron Feedforward Acceleration Framework for Embedded System

Abstract

Because of the complex architecture and multiple iterations algorithm, neural network is sometimes hard for traditional embedded devices to meet the needs of real-time processing speed in large scale data applications. Manycore processors are directly applicable for parallel implementation of the neural network. In this paper we present a multilayer perception feed forward acceleration framework based on power efficiency manycore processor, including network mapping strategy, data structure design and inter-core communication method. The framework is implemented on a Zynq and Epiphany combined hardware platform with OpenCL. The experimental results show that in a concrete example of character recognition, the framework with Epiphany achieves about four times feed forward acceleration than the dual-core ARM in Zynq with same prediction accuracy level.