Research On Neural Network Code Conversion Method For The PIM Platform

With the rapid development of artificial intelligence,big data and other fields,the scale and structural complexity of neural network models are getting higher and higher,which puts forward higher requirements for computing power and flexibility of computing platform.The existing computing architecture can not solve the problem of bandwidth difference between memory and arithmetic,which leads to the storage access operation occupying a lot of time of neural network reasoning,leading to serious storage wall problem,so the PIM technology arises at the right moment.PIM technology has tightly coupled computing unit and storage unit,thus eliminating the bottleneck of access bandwidth and reducing the cost of frequent data migration between the computing unit and storage unit,with a very high storage bandwidth and parallel computing performance.However,there are few researches on how to deploy neural network code efficiently in the PIM platform,which leads to the lack of efficient neural network model deployment method on the PIM platform.Therefore,this paper designs and develops a neural network code conversion method for the PIM platform,which provides a convenient and efficient method for the deployment of neural network model in the PIM platform.The main work of this paper is as follows:A method of rapid deployment of neural network model on the PIM platform was proposed.By designing the neural network operator library based on the PIM platform and the code conversion method,the PB model of Tensorflow is read and analyzed.The neural network computation graph is obtained,and the executable HDL-code is generated,which can be deployed on the PIM platform.Based on the characteristics of neural network model and the PIM platform,two optimization strategies of neural network code conversion are proposed,including optimization of execution flow of computational graph and optimization of operator fusion of computational graph.Based on the characteristics of the PIM platform,an improved breadth-first search algorithm is proposed to realize automatic parallelization of operators without direct data dependence in the computational graph.By analyzing the characteristics of the neural network model,some continuous neural network operators are fused into one operator to reduce the number of operators in the computational graph.Through the above optimization strategy,the reconstruction times of the PIM platform are effectively reduced during the execution of the computational graph.Finally,in order to verify the neural network code conversion method for the memory computing platform proposed in this paper,an artificial neural network classification model was built based on the Iris data set.In this paper,the code conversion method is proposed and the corresponding hardware description language code is generated.Finally,the correctness and effectiveness of the neural network code conversion method proposed in this paper are proved by the simulation software ModelSim.