Multilevel Large-Scale Modules Placement with Refined Neighborhood Exchange

Abstract

In SoC era, it is necessary to have a good and efficient large-scale modules placement for better performance estimation in chip implementation or rapid prototyping. Good representation for non-slicing floorplan/placement and ability to solve large-scale modules packing are key components. MB*-tree (Lee et al., 2003) adopted very good and well-known representation B*-tree and modern multilevel framework to handle large-scale modules floorplanning/placement. However the simulated annealing approach in declustering stage paid more time to find candidate solutions with lower cost. In this paper, we transform the epsi-neighborhood and lambda-exchange (Goto, 1981) to fit in the large-scale modules placement and use it in the refinement stage of MB*-tree algorithm. The results are encouraging. We have obtained comparable or better results in area and wirelength metrics in less time spent (up to 30% improvement), compared with original MB*-tree framework