在現代晶片設計中，連線延遲是晶片效能的重要因素。繞線是負責做出實體連線的階段，因此是大型體電路設計中重要的階段。在全域繞線中，典型的繞線器會反覆的使用迷宮繞線演算法試圖得到最佳解。在多數的狀況下，過程中繞線器找到的新路徑的成本若是比舊的路徑差就會被拋棄。在這篇論文，我們試圖去預測繞線器在給定資源與限制下，進行繞線的結果會不會得到比較好的結果。這個問題被視為一個分類問題，將繞線資源轉成影像後，訓練一個卷積神經網路來預測結果。訓練完成後，在C++環境下測試網路來得到最終的結果。我們使用一個學術界的全域繞線器NTHU-Route 2.0執行ISPD08的測資來產生我們的資料集。訓練的模型可以在NTHU-Route 2.0 main stage和refinement stage分別達到78%準確率和85%準確率。模型預測的時間比約略一半的測資繞線所需時間快且在某些測資有顯著的優勢。

Interconnect delay is the major factor of the chip performance in modern design. Routing is responsible for making the physical connection and thus is a crucial step in very large-scale integration (VLSI) design. In global routing, a typical router will use a maze routing algorithm to reroute the congested nets iteratively to approximate the optimal solution. The new path found by the router in this process will in general be abandoned if the path does not have an advantage over the old one. In this thesis, we aim to predict if the router can generate a feasible path under the given routing resources and constrains. The routing resources are transformed into an image and the problem is cast into a classification problem. A convolutional neural network (CNN) is trained to make the prediction. After training, the model is executed under a C++ environment to collect the final results. We extract the datasets from the result of an academic global router NTHU-Route 2.0 on ISPD08 testcases. Our model can reach 78% accuracy on the main stage and 85% accuracy on the refinement stage of NTHU-Route 2.0. The prediction time of our model is faster than the running time of performing maze routing on about half of the testcases and is notably faster on some of the testcases.

Acknowledgements
摘要 . . . i
Abstract . . . ii
1 Introduction . . . 1
2 Preliminaries . . . 5
2.1 Maze routing in NTHU-Route 2.0 . . . 5
2.2 Convolutional neural network . . . 7
3 Methodology . . . 9
3.1 Problem formulation . . . 9
3.2 Overview of the proposed work . . . 9
3.3 Data extraction . . . 10
3.3.1 Refinement stage . . . 10
3.3.2 Main stage . . . 10
3.4 Pre-processing . . . 11
3.5 Image generation . . . 11
3.6 Model structure . . . 14
3.7 Training . . . 15
3.8 Tensorflow C API . . . 17
4 Experimental Results . . . 19
5 Conclusions . . . 23
References . . . 25

[1]C. Y. Lee, “An algorithm for path connections and its applications,” IRE Transactions onElectronic Computers, vol. EC-10, no. 3, pp. 346–365, 1961.
[2]R. Kastner, E. Bozorgzadeh, and M. Sarrafzadeh, “Pattern routing: use and theory forincreasing predictability and avoiding coupling,” IEEE Transactions on Computer-AidedDesign of Integrated Circuits and Systems, vol. 21, no. 7, pp. 777–790, 2002.
[3]Yue Xu, Yanheng Zhang, and Chris Chu, “Fastroute 4.0: Global router with efficient viaminimization,” in 2009 Asia and South Pacific Design Automation Conference, pp. 576–581, 2009.[4]C.Albrecht,“Globalroutingbynewapproximationalgorithmsformulticommodityflow,”IEEETransactionsonComputer-AidedDesignofIntegratedCircuitsandSystems,vol.20,no. 5, pp. 622–632, 2001.
[5]J.He,M.Burtscher,R.Manohar,andK.Pingali,“Sproute: Ascalableparallelnegotiation-based global router,” in International Conference on Computer-Aided Design, pp. 1–8,2019.
[6]C. Yu and Z. Zhang, “Painting on piacement: Forecasting routing congestion using con-ditional generative adversarial nets,” in Design Automation Conference, pp. 1–6, 2019.
[7]R. Netto, S. Fabre, T. A. Fontana, V. Livramento, L. Pilla, and J. L. Güntzel, “How deeplearning can drive physical synthesis towards more predictable legalization,” in Interna-tional Symposium on Physical Design, pp. 3–10, 2019.
[8]M. B. Alawieh, Y. Lin, Z. Zhang, M. Li, Q. Huang, and D. Z. Pan, “Gan-sraf: Sub-resolution assist feature generation using conditional generative adversarial networks,” inDesign Automation Conference, pp. 1–6, 2019.
[9]Q. Zhou, X. Wang, Z. Qi, Z. Chen, Q. Zhou, and Y. Cai, “An accurate detailed routingroutability prediction model in placement,” in Asia Symposium on Quality ElectronicDesign, pp. 119–122, 2015.
[10]Z. Zhou, S. Chahal, T.-Y. Ho, and A. Ivanov, “Supervised-learning congestion predic-tor for routability-driven global routing,” International Symposium on VLSI Design, Au-tomation and Test, pp. 1–4, 2019.
[11]Y. Chang, Y. Lee, J. Gao, P. Wu, and T. Wang, “Nthu-route 2.0: A robust global router formodern designs,” IEEE Transactions on Computer-Aided Design of Integrated Circuitsand Systems, vol. 29, no. 12, pp. 1931–1944, 2010.25
[12]A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convo-lutional neural networks,” in Proceedings of the 25th International Conference on NeuralInformation Processing Systems - Volume 1, NIPS’12, p. 1097–1105, 2012.
[13]J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” CoRR,vol. abs/1804.02767, 2018.
[14]Y. Kim, “Convolutional neural networks for sentence classification,” Conference on Em-pirical Methods in Natural Language Processing (EMNLP), pp. 1746–1751, Oct. 2014.
[15]ISPD 2008 Global Routing Contest.[Online] Available:http://www.ispd.cc/contests/ispd08rc.html.