Prediction of detailed routing errors based on a placed layout via machine learning.

The aim of this project is to predict shorts that occur after detailed routing as part of an electronic design automation (EDA) flow.

Several software tools needed to be scripted to accomplish that goal. There are scripts that control the layout viewer and editor Klayout, the placement tool Eh?Placer, the (commercial) place & route suite Cadence ® Encounter ® Digital Implementation System and the Keras neural network framework.

Additional information to the one below can be found in the comments at the start of the individual scripts.

Klayout

Klayout/klayout_scripts.py provides functionality to import a GDSII (.gds) or OASIS (.oas) layout file, separately save its layers, and save square sectors of it as image (.png) files.

Eh?Placer and Encounter

The Benchmarks from the ISPD 2015 Blockage-Aware Detailed Routing-Driven Placement Contest were placed with the Eh?Placer with the following command executed in the folder of each benchmark:

~/bin/EhPlacer_ispd01.placer -tech_lef ./tech.lef -cell_lef ./cells.lef -floorplan_def ./floorplan.def -output ./EhPlacerOutput.def -placement_constraints ./placement.constraints -verilog ./design.v -cpu 8

The commands in the shell script Encounter/place_route_get_info.sh execute the command above and the TCL scripts that control the Encounter software to import and save the LEF/DEF file output of Eh?Placer as an Encounter file (import_save.tcl), route the design ('route.tcl') and then extract information from the placed (to use as features) and routed (to use as labels for the neural network) design to save the info as CSV files (get_info.tcl).

Keras

Feature Extraction

extract_data.py extracts and processes the information saved as .csv files by get_info.tcl and should therefore be used in the same folder. The generated data should be saved in the Data folder. The output is a pickled list of numpy arrays with the training data (e.g. design_placed_training_data_mgc_fft_1.pickle) and the labels (e.g. design_routed_labels_shorts_mgc_fft_1.pickle).

Configuring and training the neural network

The data extracted from the layouts via the TCL scripts is imported and processed by neural_network.py.

Evaluation

The history data how the loss and accuracy metrics have changed during training is plotted with metrics_history.py. metrics.py calculates different custom metrics based on the predictions of the fitted (trained) neural network They help with the evaluation of binary predictors for unbalanced (e.g. way more 0s than 1s) data. Its output is also saved as CSV files.

Google Colab

There are variants of neural_networks.py for use as Jupyter Notebooks in Google Colab with support for either GPU or TPU (Tensor Processing Unit) acceleration. The GPU acceleration works flawlessly already, while the TPU version will probably work once Tensorflow 2.1 is released.

Custom metrics

There is currently the problem that the functions metrics_K() and binary_confusion_matrix_K() do not calculate correct results. Nevertheless, they are still included in neural_network-experimental_keras_binary_predictor_quality_metrics.py to give the opportunity to fix these metrics. They only use the keras.backend functions and can since be called during model.fit() training process.

Draft to use Talos for automated optimization of hyper parameters

With more processing power, hyperparameter_optimization_talos.py can be used to experiment with parameters such as different batch sizes, optimization algorithms, layer shapes and training epochs.

Data

Empty folders that correspond to the data structures in the above scripts are set up in the Data folder. Due to licensing issues with the Encounter Digital Implementation Suite no output derived from it can be included in this public repository.

Style guide

The Google Python style guide was used.

Project origins

This project was created as part of a Studienarbeit (German equivalent of a Bachelor's thesis) by Matthias von Wachter with supervision by Dr. Robert Fischbach at the Institute of Electromechanical and Electronic Design at the Dresden University of Technology.

License

This software is licensed under the MIT License. The PDF with the (German) paper MLinEDA.pdf "Entwicklung eines maschinellen Lernverfahrens basierend auf neuronalen Netzen zur Vorhersage der Verdrahtbarkeit platzierter Schaltungslayouts" is licensed under the Creative Commons Attribution-ShareAlike 4.0 License.