• This repository is the implementation of Analog Placement Quality Prediction.

Towards Predicting the Art of Analog Layout: Placement Quality Prediction via Transfer Learning
Mingjie Liu, Keren Zhu, Jiaqi Gu, Linxiao Shen, Xiyuan Tang, Nan Sun, David Z. Pan
2020 Design, Automation Test in Europe Conference Exhibition (DATE)

This code has been tested under Python 2.7.13. Please install the dependancies first:

pip install -r requirements.txt

Layouts are generated with MAGICAL.
The analytical analog placement optimizes the weighted wirelength and layout area. Different net weighting would result in different placement solutions and post layout performance results.
For more information on the circuits and layout generation go under benchmark.

Placement results are represented into 2D images and device type and locations are embedded into different image channels according to the circuit topology. For more details go under extract_place.

The initial placement images and labeled performance are stored in hdf5 format under data.
To embed coordinate channels, run:

python prepare_dataset.py

If you would like to explore the dataset with different performance metrics and labeling percentiles, you can customize the script extract_feature.py. Remove ./data/data.hdf5 and run the script:

source extract.sh

Also see compress_dataset for details.

Details on the CNN models are under cnn. The models is the implementaion of training and testing in tensorflow. Although feat_3D demonstrate the best ability to generalize, feat_2D is easier to train with only minor performance degradations. The pretrained model weights are stored under weights.
See for more details on how these are generated.

For options on training, run:

python train.py --help

For transfer learning with the pretrained model on other dataset, the following demonstrates an example:

python train.py --design OTA2 --finetune --alpha 0.1 --load_weights ./weights/OTA1_feat2D_pretrain.ckpt

Make sure you have the correct pretraining weights loaded.