| The physical design optimization of the integrated circuit(IC)has been a hotspot of recent research.How to predict the performance of the circuit in the higher-level stage of the circuit design has been concerned by domestic and foreign researchers.It is also crucial to research the importance of the complex network characteristic parameters in the optimization of integrated circuit physical design.The traditional method of obtaining circuit performance is implemented the complete physical design of the circuit,that is,finish the placement and routing by running traditional electronic design automation(EDA)tools,which consumes a lot of time and resources.Integrating machine learning and complex network into the physical design optimization of integrated circuit provides a new perspective for the predicting of circuit performance.Firstly,we establish a model of a complex network of integrated circuits,regard the modules in the integrated circuit as the nodes in the network,the interconnection of the modules as the edges of the network,and use the half-perimeter wire-length(HPWL)as the weight of the edge between interconnected modules.In such a way,we construct a directed weighted network for an integrated circuit.We propose a method of extracting the complex network features suitable for predicting the circuit performance.Secondly,this paper proposes a machine learning framework for predicting circuit performance using complex network characteristics.We construct a large-volume dataset using the different complex network properties due to the different performance of IC physical designs.We can predict the performance of an unseen circuit using the trained machine learning models.Finally,based on machine learning,we explore the correlation between integrated circuit performance and complex network characteristics of placement.We propose a method to evaluate the importance of complex network characteristics by using machine learning.Through experiments on benchmark circuits,we can conclude:(1)Among the four different machine learning models,the random forest regression(RF)model represents the most efficient due to the accuracy for predicting the performance of the circuit in the placement stage and the accuracy reaches 93.71%.Compared with traditional electronic design automation tools,the method proposed in this paper has great advantages in running time,which runs 2.76 to 43.81 times faster than the EDA routing tool and on average 4.17 times faster the EDA routing tool;(2)The importance of feature parameters estimated in this paper is in the order following: the number of nodes,the average edge weight,the average degree,the average betweenness,the average strength,and the average weighted clustering coefficient,which is consistent with previous work. |
