Research On The Automated Software Development Environment Of Programmable Neural Network Processor

In recent years,the Convolutional Neural Network(CNN)development is rapid,programmable Neural Network processor as its product became one of research hot spots.However,in the process of applying programmable neural network processor,convolutional neural network is difficult to find appropriate data structure for dynamic storage and management of data due to its multiple types and large scale parameter data.In addition,the operation of programmable neural network processor is inseparable from the control of software programs,which require developers to write manually on the basis of being familiar with the underlying principles of hardware.On the one hand,software developers need to be capable of both software and hardware,making software development more difficult.On the other hand,tens of thousands of and even more software programs written by humans are not only inefficient but also error-prone.In view of the above problems,this thesis realizes the automated software development environment of programmable neural network processor.In this environment,users can input the parameters of convolutional neural network and the parameters of hardware resources through the graphical interfaces.Then by parsing these parameters,the automated environment can achieve the integration of different types of data storage structures,the automatically built instruction programs,and network structure configuration information of automatically generated processor level.In this thesis,the main work and contributions are as follows:(1)In the case of limited storage resources,the vertical segmentation storage method is proposed for the image storage data structure and the address structure of layer,column and row splicing is designed to simplify the calculation process of transmission address.At the same time,other types of data storage structures in the convolutional neural network are configured to realize the dynamic storage management of data.(2)Transportn trigger architecture instruction for programmable neural network processor is studied.(3)By parsing the parameters of neural network model and the parameters of hardware configuration,this thesis realizes the automatic construction of software program and network structure configuration information of processor-level,which shortens the software development cycle of programmable neural network processor.(4)The user-oriented visual graphic interface is designed and implemented.The automated software development environment of programmable neural network processor realized in this thesis shortens the time cost of the programmable neural network processor software program development,improves the efficiency of software program development and is beneficial to the application and promotion of programmable neural network processor.