Title: Macro Placement Generation via Physics-Guided Geometric Diffusion

Abstract: Macro placement serves as a critical phase in VLSI physical design, exerting a fundamental influence on overall chip performance. While recent data-driven placement approaches have shown considerable promise, they frequently encounter difficulties in managing sequential dependencies and reconciling topological connectivity with physical constraints. To address these challenges, we introduce MacroDiff+, a framework for geometric diffusion guided by physics principles. Our approach features a dual-domain denoising architecture that integrates global geometric context, modeled via a Transformer, with topological connectivity encoded by heterogeneous Graph Neural Networks (GNNs). Additionally, we present Physics-Guided Sampling, an inference mechanism that utilizes explicit gradients to actively direct the generation process, thereby guaranteeing both statistical plausibility and physical validity. Evaluations on the ISPD2005 MMS benchmarks reveal that MacroDiff+ surpasses current state-of-the-art baselines, achieving a 6.1-6.2% decrease in wirelength. Furthermore, the method demonstrates enhanced stability and scalability on large-scale designs, particularly in scenarios where previous methods fail to converge. The source code can be accessed at https://github.com/jhy00n/MacroDiff-plus.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC