Research Of Deep Learning Application Based On FPGA Platform

In recent years,artificial intelligence is one of the hottest topics in the world,and the impetus for the vigorous development of artificial intelligence technology is the rise of deep learning.Nowadays,deep learning has been widely used in various fields and has made great breakthroughs in computer vision,speech recognition and natural language processing.However,deep learning's powerful simulation prediction ability is inseparable from the support of strong computing power hardware,which is the foundation of deep learning and development.How to meet the increasing demand for high performance computing in deep learning has been a hot topic in many research institutes and commercial companies.With the continuous development of FPGA and the continuous improvement of Open CL heterogeneous computing standards,the new heterogeneous computing platform based on FPGA begin to show its potential in high performance computing area.Compared with large workstations and GPU clusters,FPGA platform has advantages of high performance and low power consumption.Therefore,based on the Open CL heterogeneous computing platform of FPGA,this paper has carried out research and specific implementation on heterogeneous computing acceleration of CNN image classification which is commonly used in deep learning.This paper first introduces the knowledge of convolutional neural network used in image processing and common CNN network model.This paper analyzes the programming framework of the Open CL heterogeneous computing standard,the four model features of Open CL standard platform model,execution model,memory model and programming model are described in detail.Then,this paper has designed the heterogeneous computing acceleration scheme of CNN image classification based on the analysis of the parallelism characteristics in the process of CNN image classification and the hardware architecture high parallelism features of FPGA.According to the specific CNN model Alex Net,the whole network model of FPGA heterogeneous computing accelerated kernel has been designed.According to the calculation similarity between different layer networks in the model,the same kernel design method is adopted for the eight network layers in the network model so that the design can be reused.Five corresponding hardware acceleration kernels are designed for different operation operations such as convolution and pooling in single-layer network.According to the data transmission characteristics in the calculation process,channel technology is used to transmit data between multiple cores,which has improved the data transfer efficiency of the kernel and the memory efficiency of the system.Finally,the development environment of heterogeneous computing platform based on FPGA platform was built using the de5-net development board,and the acceleration plan of CNN heterogeneous computing was realized.During the FPGA kernel program implementation process,some optimization methods have been adopted such as quantitative data,multi copy of calculate pipeline,and use of channels,which improves the resource utilization of FPGA and the parallel computing acceleration performance of FPGA.The system was tested with Image Net data set,and the accuracy of classification was achieved,compared with the execution time of CPU and GPU under Caffe.The test results show that the FPGA based on CNN heterogeneous computing acceleration system has a good performance in acceleration performance and power consumption balance.