NeurIPS'24 Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer

Official implementation of NeurIPS'24 paper "Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer"

This repository contains the Python code for MaskRegulate, a reinforcement learning implementation for macro placement. Formulated as a regulator rather than a placer and equipped with RegularMask, MaskRegulate empirically achieves significant improvements over previous methods.

Requirements

• python==3.8.5
• torch==1.7.1
• torchvision==0.8.2
• torchaudio==0.7.2
• pyyaml==5.3.1
• gym==0.22.0
• Shapely==2.0.4
• matplotlib==3.4.3
• cairocffi==1.7.0
• tqdm==4.61.2
• tensorboard==2.14.0
• scikit_learn==1.3.2
• numpy==1.21.2

File structure

• benchmark directory stores the benchmarks for running. Please download the ICCAD2015 benchmark and move it to benchmark/ (i.e., benchmark/superblue1).
• config stores the hyperparameters for our algorithm.
• DREAMPlace_source serves as a thirdparty standard cell placer borrowed from DREAMPlace.
• policy stores a pretrained policy trained on superblue1, superblue3, superblue4 and superblue5.
• src contains the source code of MaskRegulate.
• utils defines some functions to be used for optimization.

Usage

Please first download the docker image from Baidu Netdisk or pull it from the cloud duketomlist/macro-regulator.

docker pull duketomlist/macro-regulator:cuda

Then, please compile DREAMPlace_source in the docker container following the below commands:

cd DREAMPlace_source
mkdir build
cd build
cmake .. -DCMAKE_INSTALL_PREFIX=../../DREAMPlace
make
make install

After that, please download the ICCAD2015 benchmark via Google Drive.

Parameters

• --seed random seed for running.
• --gpu GPU ID used by the algorithm execution.
• --episode number of episodes for training.
• --checkpoint_path the saved model to be loaded.
• --eval_policy only evalute the policy given by --checkpoint_path.
• --dataset_path the placement file to regulate. MaskRegulate will improve the chip layout obtained from DREAMPlace if --dataset_path is not provided. Currently, MaskRegulate only supports training on a single benchmark when a --dataset_path is provided (i.e., if superblue1_reference.def is given, please set --benchmark_train=[superblue1] and --dataset_path=./superblue1_reference.def).

Run a training task

Please first navigate to the src directory.

python main.py --benchmark_train=[Benchmark1,Benchmark2] --benchmark_eval=[Benchmark1',Benchmark2']

• --benchmark_train contains the benchmarks to train on.
• --benchmark_eval contains the benchmarks to evaluate on.

For example, if you want to train MaskRegulate on benchmark superblue1, superblue3 and evaluate the performance on superblue1, superblue3, superblue5, run our command as shown bellow:

python main.py --benchmark_train=[superblue1,superblue3] --benchmark_eval=[superblue1,superblue3,superblue5]

Script run_train.sh is provided for a quick start.

Run a testing task

We also provide a pre-trained model trained on superblue1, superblue3, superblue4 and superblue5 in policy/pretrained_model.pkl, which can be loaded and evaluated. For example, run the following command to test our policy on superblue1:

python main.py --benchmark_train=[] --benchmark_eval=[superblue1] --check_point_path=../policy/pretrained_model.pkl --eval_policy=True

Script run_test.sh is provided for a quick start.

Citation

@inproceedings{macro-regulator,
    author = {Ke Xue, Ruo-Tong Chen, Xi Lin, Yunqi Shi, Shixiong Kai, Siyuan Xu, Chao Qian.},
    title = {Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer},
    booktitle = {Advances in Neural Information Processing Systems 38 (NeurIPS’24)},
    year = {2024},
    pages = {140565--140588},
    address={Vancouver, Canada}
}