Design And Implementation Of Embedded Neural Networks Computing Framework Based On DSP

In recent years,the research and application of Deep Learning has entered an explosive phase.There have been several neural networks suitable for mobile devices and some model compression algorithms.The neural network has begun to be applied to embedded platforms,which has led to the emergence of more and more neural network computing frameworks for embedded platforms.These computing frameworks mainly use ARM CPU,GPU(Graphics Processing Unit)or FPGA(Field Programmable Gate Arrays)as the computational acceleration hardware,but these hardware cannot make a good balance among performance,cost and power consumption,which stopped the application of these frameworks in the embedded platforms on a large scale.In contrast,DSP(Digital Signal Processor)has better performance and lower cost and power consumption.Therefore,this thesis investigated the problems existing in the current embedded platform neural networks acceleration framework,and designed and implemented a DSP-based neural networks computing framework-Lightweight Accelerator for Neural Networks on Embedded System,referred to as LANNES.This computing framework has good performance on the embedded platform,and can better control cost and power consumption.It is a solution suitable for large-scale application of neural networks on embedded platforms.This thesis designed the architecture and underlying acceleration mechanism of the neural network computing framework for DSP platform.The framework architecture consists of three main parts: the LANNES model and network files,the LANNES infrastructure and the LANNES accelerator.This thesis implemented the model,network file conversion and loading,and the various basic components of the framework,making full use of the special architecture of DSP,with some memory optimization algorithms and CPU instruction-level parallel optimization technology to maximize the performance of the DSP accelerator.In this thesis,the performance and accuracy of the implemented LANNES framework are tested and verified.The tests show that the LANNES framework based on DSP is more cost-effective than the GPU-based framework,and the performance is significantly ahead of the framework based on more expensive quad-core ARM CPU.This framework also has the advantages of very lightweight,high ease of use and low power consumption,in addition to ensure that the accuracy of the calculation results will not be affected by the framework itself.The lightweight neural network computing framework based on DSP proposed by this research can be used as an excellent solution for large-scale application of neural networks in embedded platforms.