Utilizing CGAN To Predict The Routing Congestion Of FPGA Design

The design complexity of Field Programmable Gate Arrays(FPGAs)increases continuously following Moore’s Law.Physical design requires a large number of optimization iterations to be achieved,with routing being a particularly time-consuming process in chip physical design.Routing congestion affects performance metrics such as chip area and timing delay,so it is essential to predict and address it accurately and quickly.This article proposes a method for predicting detailed routing congestion maps using deep learning.It extracts features and synthesizes images in the detailed layout phase,using Conditional Generative Adversarial Networks(CGANs)for image-to-image translation,allowing for fast and accurate prediction of routing congestion maps,and intuitive observation of the utilization of all routing channels.The main work is as follows:First,the input image learned by the generation model is generated.Electronic Design Automation(EDA)tool VTR8.0 is used to complete automatic layout and routing.To avoid overfitting in deep learning,parameters with a high impact on routing congestion are selected.After detailed layout,three features are selected for image synthesis:(1)layout information,(2)intra-block pin utilization,and(3)netlist connections.These three features are synthesized into one image,which is used as the input image for the generator learning in the model.A real routing congestion image obtained after detailed routing is used as the real image for the discriminator learning.Secondly,the routing congestion prediction task is transformed into an image translation problem.A model called CBAM-CGAN(Convolutional Block Attention Mechanism-Conditional Generative Adversarial Networks)is proposed to implement pixel-level prediction using a conditional generative adversarial network model.The channel attention mechanism and spatial attention mechanism are introduced based on the conditional generative adversarial network framework,highlighting important feature channels,making the generated image more detailed,reducing detail loss during the convolution process,and improving the quality of the generated image.Experiments are conducted based on the benchmark test circuit VTR benchmark.The performance of the CBAM-CGAN model is tested using two approaches,and the model is compared with the CGAN model to verify the effectiveness of incorporating the attention mechanism into the model.The experimental results show that the method achieved good results in predicting routing congestion maps in the layout phase.Compared with the CGAN model,the average structural similarity index increased by0.89%,the average peak signal-to-noise ratio increased by 1.37%,the average normalized root-mean-square pixel error decreased by 3.8%,and the average pixel accuracy difference decreased by 0.06%.The prediction time for a single image is about 0.1seconds.The experiments demonstrate the accuracy and speed of the CBAM-CGAN model in predicting routing congestion maps in FPGAs.