An Analytical Global Placement Algorithm With A New Self-Adaptive Wirelength Model

Analytical-Based placement algorithms consistently deliver high-quality solutions in global placement. One common approach for global placement is minimizing the half-perimeter wirelength (HPWL), which is, however, non-differentiable. To this end, we employ Moreau-Yosida regularization to smooth the HPWL and propose a new self-adaptive wirelength model in this paper. Particularly, the self-adaptive smoothing parameter value is dynamically updated according to the overlaps between instances. In addition, we propose to incorporate the local density function into the global density function to mitigate the unnecessary wirelength overhead. Experimental results show that, compared to the commonly used log-sum-exp (LSE) wire-length model and weighted-average (WA) wirelength model, our proposed method reduces the HPWL by over 1% on average in both ISPD2018 and ISPD2019 benchmarks.