面向集成电路拥塞预测的版图数据扩充方法  

Layout data augmentation method for congestion prediction in integrated circuits

作　　者：莫桂棋 夏益民[1] 邢延[1] 李卫军[1] 蔡述庭[1] MO Guiqi;XIA Yimin;XING Yan;LI Weijun;CAI Shuting(College of Integrated Circuits,Guangdong University of Technology,Guangzhou Guangdong 510006,China)

机构地区：[1]广东工业大学集成电路学院,广州510006

出　　处：《计算机应用》2023年第S02期261-267,共7页journal of Computer Applications

基　　金：广东省重点领域研发计划项目(2022B0701180001);广东工业大学大学生创新训练项目(xj2023118450109)。

摘　　要：针对深度学习领域中集成电路(IC)版图数据集不充足导致预测模型准确度有限的问题,基于双生成对抗网络(DoubleGAN),提出了一种面向拥塞预测的版图数据扩充和自动标注方法。首先,利用5层反卷积网络结构搭建特征图生成模型,利用U-net结构搭建标签自动标注模型;然后,以Wasserstein距离为损失函数,训练以上模型;最后,原数据集通过以上模型扩充一定倍数后作为训练集,代入拥塞预测模型,提升预测的准确度。在拥塞预测数据集上进行数据生成实验,DoubleGAN生成数据的FID(Fréchet Inception Distance)的平均值为165.943,质量较好。与传统、深度卷积生成对抗网络(DCGAN)扩充方法进行对比实验,使用DoubleGAN扩充2倍时的预测方法,在归一化均方根误差(NRMS)、峰值信噪比(PSNR)、结构相似衡量(SSIM)指标上均优于传统方法和DCGAN扩充方法;对比数据扩充前的拥塞预测模型,各指标均有1.34%~17.98%的改善效果,实验结果表明所提扩充方法在总体上能够提高预测模型的准确度。In response to the insufficient dataset of Integrated Circuit(IC)layout in the field of deep learning,a method for layout data augmentation and automatic labeling for congestion prediction based on DoubleGAN(Double Generative Adversarial Network)was proposed.Firstly,a five-layer deconvolutional network structure was used to build up feature map generation model,and a U-net model was used to build up automatic labeling model.Then,using the Wasserstein distance as the loss function,the two trained models were used to expand the original dataset by a certain multiple,and the accuracy of the congestion prediction model was improved.The experiments on congestion prediction dataset showed that the average FID(Fréchet Inception Distance)value of the data generated by DoubleGAN was 165.943,indicating good quality.Normalized Root Mean Square error(NRMS),Peak Signal-to-Noise Ratio(PSNR),and Structural Similarity Index Measure(SSIM)were used as evaluation indexes for the prediction model.Comparative analysis with traditional and DCGAN(Deep Convolutional GAN)data augmentation methods shows that the proposed method is superior when the augmentation factor is 2,and all indexes improve by 1.34%to 17.98%compared to the congestion prediction model before data augmentation,demonstrating that the proposed augmentation algorithm can optimize the accuracy of the prediction model.

关 键 词：深度学习 集成电路 双生成对抗网络 拥塞预测 数据扩充 自动标注 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TN40[自动化与计算机技术—控制科学与工程]