Deep neural network model compiling optimization method based on SW processor

The invention relates to a deep neural network model compilation optimization method based on an SW processor, and the method comprises the steps: carrying out the corresponding packaging of an automatically generated AOT (Ahead-of-Time) code according to the code specification of the SW processor, and carrying out the correct initialization and calling in a main function, a local data storage management (LDM) method is used for transmitting data related to partial calculation to a local data cache in batches so as to improve the transmission rate, and a direct memory access (DMA) code insertion method is used for realizing efficient data transmission of a main memory and the LDM. The method disclosed by the invention is suitable for a hardware system structure of the SW processor, high-performance codes can be quickly generated, and the development efficiency of programmers is improved. 本发明涉及一种基于申威处理器的深度神经网络模型编译优化方法，其方法依照申威处理器的代码规范，对自动生成的AOT(Ahead-of-Time)代码进行相应的封装，并在主函数中进行正确的初始化和调用，并利用局部数据存储管理(LDM)方法将部分计算涉及