BADGE: Boundary-Aware Dirichlet Graph Embedding for Initialization in DREAMPlace

초록

Initialization is critical in analytical placement to prevent optimizers from being trapped in suboptimal local minima. We introduce BADGE, a Boundary-Aware Dirichlet Graph Embedding that leverages fixed macros as boundary conditions. By minimizing a weighted quadratic energy under Dirichlet constraints, BADGE generates a globally coherent initial placement. We further design boundary-attraction weighting and boundary-biased filler initialization to preserve embedding quality. Experiments on MMS benchmarks demonstrate that BADGE consistently reduces half-perimeter wirelength (HPWL) and global placement iterations compared to random and GiFt initializations. These results highlight the efficiency of boundary-aware initialization for scalable placers.

키워드