Complex Network Based Routing Congestion And Optimization Of FPGA Design

The high complexity of Field Programmable Gate Array(FPGA)circuits increases the time cost of circuit design,it tends to cause routing congestion,which leads to unroutability in practice.The routing congestion prediction in the early stage becomes critical to improving the ultimate quality of result(Qo R).According to the characteristics of Very Large Scale Integration(Vl SI)with high integration density and complex structure,FPGA circuit is a man-made complex system.The layout file can be abstracted as the expression form of complex network,and complex network parameters can be used to predict the congestion degree of FPGA circuit routing.Firstly,this paper establishes the weighted circuit network model of FPGA circuit layout stage,extracts and analyzes the characteristic parameters of the complex network model of FPGA circuit,introduces the classical machine learning regression model to learn the complex network parameters,and proposes a method to predict the congestion degree of circuit routing combined with machine learning method.Secondly,The study developed a prediction model for routing congestion degree using the BP neural network,and optimized it with FHO mechanism to enhance its performance.The research also analyzed the relationship between characteristic parameters and routing congestion degree in the layout of a complex network.Thirdly,The predicted congestion degree in the routing stage is used as feedback to the layout stage.The coordinate information of highly congested areas is confirmed,and the corresponding area in the original layout is expanded by adding blank spaces.This process ultimately improves the circuit routing congestion degree.Experiments are carried out on ISPD2016 benchmark circuits.The experimental results show that:(1)The four classical congestion degree models established have high prediction accuracy.The average predicted speed of the model is much lower than the actual wiring time,and there is no overfitting effect in each model,so the routing congestion degree can be predicted quickly and accurately.By analyzing the importance of characteristic parameters,the results show that degree centrality and degree have the greatest correlation with routing congestion degree.(2)After the BP neural network was optimized by FHO algorithm in this paper,R2,MAE,RMSE and other performance indicators were greatly improved.In the training stage,the model achieves fast convergence and has strong generalization performance.(3)The proposed optimization method can provide useful routing congestion prediction information for EDA tools.After circuit layout optimization,the circuit routing congestion area is reduced,the area with severe routing congestion is obviously alleviated,and the line length after wiring is significantly reduced.