Macro Placement by Wire-Mask-Guided Black-Box Optimization
Authors: Yunqi Shi, Ke Xue, Song Lei, Chao Qian
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on multiple popular benchmarks demonstrate that Wire Mask-BBO, when equipped with different BBO algorithms such as random search (RS), Bayesian optimization (BO), and evolutionary algorithms (EA), significantly outperforms the compared representative methods... Extensive experimental results show that Wire Mask-BBO is clearly superior to previous packingbased, analytical, and RL-based methods. |
| Researcher Affiliation | Academia | Yunqi Shi, Ke Xue, Lei Song, Chao Qian National Key Laboratory for Novel Software Technology, Nanjing University, China School of Artificial Intelligence, Nanjing University, China {shiyq, xuek, songl, qianc}@lamda.nju.edu.cn |
| Pseudocode | Yes | Algorithm 1 Objective evaluation of Wire Mask-BBO |
| Open Source Code | Yes | Our code is available at https://github.com/lamda-bbo/Wire Mask-BBO. |
| Open Datasets | Yes | We mainly empirically test our method on the ISPD2005 benchmark [35], which was originally proposed as a standard cell placement benchmark with fixed macros. The ISPD2005 benchmark contains eight chips, i.e., adaptec1-4 and bigblue1-4. |
| Dataset Splits | No | The paper does not explicitly provide details about training, validation, or test dataset splits (e.g., specific percentages, absolute counts, or references to predefined splits with citations) for its experiments. It mentions using the ISPD2005 benchmark and running multiple times with 'five random seeds', but no specific data splitting methodology is detailed. |
| Hardware Specification | Yes | All experiments are run on two Intel Xeon Platinum 8171M CPUs, each with 26 cores and 52 threads. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in its experiments (e.g., Python, PyTorch, TensorFlow, NumPy, etc.). While it mentions using baselines like DREAMPlace and PRNet, no versions for these tools are given. |
| Experiment Setup | Yes | The runtime for each run of Wire Mask-BBO is set as 1,000 minutes. For each chip, the canvas is partitioned into approximately 150 150 grids heuristically. The #Partitions, however, constitutes a crucial hyperparameter, which is determined heuristically based on the heights and widths of the macros and the canvas. We empirically investigate the influence of the the hyper-parameters of Wire Mask-EA, i.e., the number of partitions when discretizing the chip canvas into grids, the order of adjusting the positions of macros in objective evaluation, and the mutation operator. |