Optimization Implementation Of Deep Learning Algorithm Based On GPDSP Vector Processing

With the rapid development of deep learning algorithm,object recognition technology and object detection algorithm based on convolution neural network have been widely used in image recognition,image segmentation and other fields.The application of neural network algorithm in various real-time environments is also a concern of academia and industry.Generally,neural network algorithm has a huge amount of computation.How to optimize the algorithm and satisfy the real-time requirement is a hot research topic.At the same time,it is becoming more and more urgent to find and develop a suitable hardware platform to support in deep learning algorithm.Because of the huge amount of computation,the practical application of the deep learning algorithm depends on the high performance computing platform.The main work of this paper is as follows: exploring how to map the deep learning algorithm on GPDSP,calculating the typical layer of convolution,pooling,normalization and other depth learning algorithms on vector processors,mapping the algorithm to the processor efficiently with vectorization method;According to the conclusion obtained by mapping method,the deep learning algorithms Alex Net and YOLO are implemented on FT vector processor,and the camera is connected with GPDSP to process the real-time captured images and build a complete system.To further improve the processor utilization,the application on GPDSP is optimized in mapping method,programming language,algorithm and other aspects.In the future work,we will further improve the structure of GPDSP to provide better support for deep learning algorithm and neural network structure.And for the general convolution neural network,develop a standard convolution library,so that in the future research and use,can be more convenient to call,fast implementation,simplify the operation,can be better promoted.