A Self-Adaptive weighted-average wire-length model for VLSI global placement

Abstract

Global placement roughly decides the location of units in the very large-scale integrated (VLSI) and fundamentally determines the quality of physical design. Thus, it’s desirable to find an efficient method to solve the global placement problem. Global placement solves the problem by minimizing the total half-perimeter wirelength (HPWL) under density constraints. However, the non-differentiability of HPWL prevents advanced gradient-based methods from being applied to global placement. Therefore, smooth wirelength models have been proposed to approximate HPWL. Among all the models, weighted-average wirelength (WAWL) performs the best. In this letter, we propose an improved self-adaptive weighted-average wirelength (SaWAWL) model to further fit the HPWL. Instead of setting a generic γ for all nets in the design, the new model enables each net to adaptively adjust their respective γ according to their real length, thus can better approximate HPWL to achieve higher-quality placement results. Based on the SaWAWL and the framework of DREAMPlace, a global placer is implemented. Experimental results show that HPWL on open-source benchmarks is reduced by up to 6.56% with an average of 3.74%, which proves that our model can achieve better performance than the current state-of-the-art WAWL.