lvchenyangai / chipformer

The release for ICML 2023 paper

ChiPFormer: Transferable Chip Placement via Offline Decision Transformer

An offline RL placement method ChiPFormer, significantly improving the design quality and efficiency.

Publication

Yao Lai, Jinxin Liu, Zhentao Tang, Bin Wang, Jianye Hao, Ping Luo. "ChiPFormer: Transferable Chip Placement via Offline Decision Transformer." International Conference on Machine Learning, ICML (2023): 18346-18364.

paper | dataset | website | video(English) | video(Mandarin)

Usage

Download the offline placement dataset

Download the offline placement dataset from Google Drive. (We provide the placement data including 12 benchmarks: adaptec1-4, bigblue1-4, ibm01-04, but you can just download a single benchmark file for training.) (24/03/02 Updated)

For a quick start, you can directly unzip the adaptec1_small.pkl data in this repo.

tar -zxvf adaptec1_small.pkl.tar.gz

Download placement benchmark files

Our test dataset includes ISPD05 and ICCAD04 datasets. We provide the adaptec1 dataset for a quick start. For all benchmarks, they can also be downloaded from the placement_bench.zip file in Google Drive.

Quick start

Our ChiPFormer includes the pretraining and finetuning two parts.

• For pretraining:

python run_dt_place.py

The dataset file for training can be modified in create_dataset.py. The saved models are in the folder save_models.

• For finetuning:

python odt.py --benchmark=adaptec1

The model path for funetuning can be modified in odt.py. Typically, the optimal model files obtained from pre-training are used to fine-tune.

VGAE circuit feature extracture

(2024/05/25 updated)

For using VAGE, you need to firstly train the model, and use this model to evaluate the given benchmarks.

python graph_train.py

Then the trained model would be saved in the directory save_graph_models/.

python graph_eval.py

The circuit embeddings are saved in the file circuit_g_token*.pkl.

Then, you should modify the file mingpt/place_db.py for the line

def __init__(self, benchmark = None, offset = 0, is_graph = False):

to

def __init__(self, benchmark = None, offset = 0, is_graph = True):

We also provide the example model and embedding file which you can directly use.

Parameter

For run_dt_place.py:

• seed Random seed.
• context_length Maximum length of decision transformer.
• epochs Maximum training epochs.
• batch_size Batch size.
• cuda GPU label for use.
• is_eval_only Whether to evaluate function. (In evaluation function, it will place all macros rather than the maximum length number of macros)
• test_all_macro Whether to place all existing macros.

For odt.py:

• replay_size Replay buffer size for finetuning.
• traj_len Maximum length of decision transformer.
• batch_size Batch size.
• benchmark Circuit benchmark for finetuning.
• max_online_iters Maximum number of iterations for finetuning.
• eval_interval Evaluation every how many iterations.
• exploration_rtg Return-to-go value for exploration.
• is_fifo Whether to use fifo buffer or priority queue buffer.
• cuda GPU label for use.

Dependency

• Python >= 3.9

• Pytorch >= 1.10

  • Other versions may also work, but not tested

• tqdm

Reference code

The code refers to the following open source repos:

• decision-transformer
• online-dt

Citation

If you find our paper/code useful in your research, please cite

@inproceedings{lai2023chipformer,
  author       = {Lai, Yao and Liu, Jinxin and Tang, Zhentao and Wang, Bin and Hao, Jianye and Luo, Ping},
  title        = {ChiPFormer: Transferable Chip Placement via Offline Decision Transformer},
  booktitle    = {International Conference on Machine Learning, {ICML} 2023, 23-29 July
                  2023, Honolulu, Hawaii, {USA}},
  series       = {Proceedings of Machine Learning Research},
  volume       = {202},
  pages        = {18346--18364},
  publisher    = {{PMLR}},
  year         = {2023},
  url          = {https://proceedings.mlr.press/v202/lai23c.html},
}