Research on place and route method using deep learning

2020 Fiscal Year Final Research Report

• Project/Area Number: 16K00081
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Computer system
• Research Institution: Hiroshima City University
• Principal Investigator: Hironaka Tetsuo 広島市立大学, 情報科学研究科, 教授 (10253486)
• Project Period (FY): 2016-04-01 – 2021-03-31
• Keywords: 配置配線 / ニューラルネットワーク / 再構成可能デバイス

Outline of Final Research Achievements

We revealed the method to apply a neural network as a cost function used in the SA method. If a neural network is simply applied, the placement and connection information will require an unrealistic number of input nodes. In this study, so we established a method to significantly reduce the number of input nodes without significantly losing the meaning of the input data by converting the placement and connection information to a map marked with placement position and the possibly routed area. Furthermore, we established a method to predict the better placement by comparing two placements generated from the same netlist by the neural network. Furthermore, it was also found that the placement performed by the conventional place and route methods can be improved by applying the SA method using the trained neural network as the cost function.

Free Research Field

コンピュータアーキテクチャ

Academic Significance and Societal Importance of the Research Achievements

SA法のコスト関数は配線配置の可否を判定する数式モデルから作成している．そのため，性能向上を目指して再構成デバイスの構成を複雑にすると簡易な数式モデルで配線配置の可否を精度よく判定することが難しくなる．このコスト関数の精度が低下するとSA法を用いて生成される配置配線の品質が低下する．本研究では，深層学習を用いて訓練データである過去の配置配線結果から自動的にSA法のコスト関数を生成することの実現性を示した．これにより，精度の良い簡易な数式モデルが作成困難な場合においても，過去の配置配線結果を用いてより良いSA法のコスト関数を自動生成できる可能性を示した．