MaskPlace: Fast Chip Placement via Reinforced Visual Representation Learning
A new chip placement method based on visual representation learning.
Publication
Lai, Yao, Yao Mu, and Ping Luo. “Maskplace: Fast chip placement via reinforced visual representation learning.” Advances in Neural Information Processing Systems 35 (2022): 24019-24030. (NeurIPS 2022, spotlight)
Usage
You can start easily by using the following script.
cd maskplace
python PPO2.py
Parameter
- gamma Decay factor.
- seed Random seed.
- disable_tqdm Whether to disable the progress bar.
- lr Learning rate.
- log-interval Interval between training status logs.
- pnm Number of place modules for each placement trajectory.
- benchmark Circuit benchmark.
- soft_coefficient Whether to constriant the actions based on the wiremask.
- batch_size Batch size.
- is_test Testing mode based on the trained agent.
- save_fig Whether to save placement figures.
Benchmark
The repo has provided the benchmark adaptec1 and ariane. For other benchmarks, you can download them by the following the link:
http://www.cerc.utexas.edu/~zixuan/ispd2005dp.tar.xz
Dependency
-
Python >= 3.9
-
Pytorch >= 1.10
- Other versions may also work, but not tested
- gym >= 0.21.0
- matplotlib >= 3.7.1
- tqdm
- protobuf (for benchmark ariane)
Citation
If you find our paper/code useful in your research, please cite
@article{lai2022maskplace,
title={Maskplace: Fast chip placement via reinforced visual representation learning},
author={Lai, Yao and Mu, Yao and Luo, Ping},
journal={Advances in Neural Information Processing Systems},
volume={35},
pages={24019--24030},
year={2022}
}
The placement process animation
Benchmark: Bigblue3
| Placement | Pos Mask t | Wire Mask t |
|---|---|---|
 |
 |
 |
 |
 |
 |
Standard Cell Placement
Fix macros and use DREAMPlace (classic optimization-based method) to place standard cells.
 |
 |
 |
Full Benchmark demonstration
| Benchmark | DREAMPlace | Graph | DeepPR | MaskPlace |
|---|---|---|---|---|
| adaptec1 |  |
 |
 |
 |
| HPWL (105) | 17.94 | 26.05 | 21.36 | 6.57 |
| Wirel (105) | 19.24 | 28.54 | 25.64 | 7.36 |
| Overlap | 0.34% | 1.89% | 32.03% | 0 |
| adaptec2 |  |
 |
 |
 |
| HPWL (105) | 135.32 | 359.35 | 197.13 | 79.98 |
| Wirel (105) | 140.91 | 381.64 | 205.78 | 83.59 |
| Overlap | 0.16% | 1.54% | 49.10% | 0 |
| adaptec3 |  |
 |
 |
 |
| HPWL (105) | 112.28 | 392.66 | 340.29 | 79.33 |
| Wirel (105) | 119.23 | 409.37 | 372.02 | 85.28 |
| Overlap | 0 | 1.26% | 29.10% | 0 |
| adaptec4 |  |
 |
 |
 |
| HPWL (105) | 37.77 | 152.89 | 243.12 | 75.75 |
| Wirel (105) | 47.90 | 179.43 | 290.14 | 88.87 |
| Overlap | 0 | 7.43% | 19.29% | 0 |
| bigblue1 |  |
 |
 |
 |
| HPWL (105) | 2.50 | 8.32 | 20.49 | 2.42 |
| Wirel (105) | 3.41 | 10.00 | 25.68 | 3.14 |
| Overlap | 0 | 2.48% | 9.33% | 0 |
| bigblue3 |  |
 |
 |
 |
| HPWL (105) | 104.05 | 345.49 | 439.09 | 82.61 |
| Wirel (105) | 107.58 | 373.33 | 517.86 | 88.51 |
| Overlap | 8.06% | 0.80% | 85.23% | 0 |
| ariane |  |
 |
 |
 |
| HPWL (105) | 20.30 | 16.83 | 51.43 | 14.86 |
| Wirel (105) | 21.72 | 18.48 | 55.85 | 15.80 |
| Overlap | 0.78% | 3.72% | 38.91% | 1.94% |