pytorch-spice-cnn

This repository contains a conventional neural network model designed to predict the IR-drop of macro-block circuits in integrated circuits. The purpose of this project is to develop an accurate and efficient model for estimating the IR-drop, which is an important factor in analyzing the performance and reliability of integrated circuits.

Testing

To evaluate the performance of the neural network model, testing was conducted using the OpenLane project. OpenLane was utilized to generate real spice-netlist files from Verilog descriptions. The generated spice-netlists were then used as input for the neural network model, allowing for accurate prediction of the IR-drop.

The testing results for various integrated circuits are as follows:

Circuit	Error (%)	NN Time	go-spice Time	Time Difference
APU	0.18553	0.515s	18.33s	35.5
GCD	0.79802	0.1585s	53.21s	335.7
mem_v1r1	0.07101	0.0987s	8.66s	87.7
wbqspiflash	0.5811	0.11789s	11.7s	100
xtea	0.12282	0.08216s	9.53s	116.8
zipdiv	0.17582	0.13105s	6.88s	52.4

The predictions obtained from the neural network model were compared with the results obtained from a program called go-spice, which is used for calculating IR-drop. This allows for a comprehensive analysis of the accuracy and reliability of the neural network model.

Example

APU - Real	APU - Prediction

Repository Contents

• utils - project utilities such as models, readers, etc.
• analysis.py - analyze generated data on IR-drop (minimum, maximum, and mean) depending on threshold value (maximum IR-drop from voltage source value in %)
• draw.py - draw spice-netlist image
• generate.py - generate training dataset of integrated circuits
• predict.py - predict spice-netlist IR-drop
• test.py - perform tests on the test dataset, which includes spice-netlists and IR-drop solutions in CSV format
• train.py - perform training on the train dataset, which includes spice-netlists and IR-drop solutions in CSV format

Usage

Generate

To generate training dataset, execute generate.py:

python generate.py -p ./assets/train -w [300000, 350000, 400000] -d1 [30, 35, 40, 45, 55] -d45 [2, 3, 4, 6, 8] -s [15, 20, 25, 30] -p [15, 20, 25, 30]

Where:

• -p, --path - path to the training dataset folder
• -w, --width - array of circuit widths/heights
• -d1, `--density1

` - array of the number of rails in the 1st metal layer

• -d45, --density45 - array of the number of rails in the 4th and 5th layers at the same time
• -s, --split_level - array of the maximum number of current sources on each of the 1st layer's rails
• -p, --padding - array of padding values for the circuit

Train

To train the model, you need a training dataset that contains spice-netlists and IR-drop solutions in CSV format, along with a CSV file that contains all the training data - [path_to_netlist, path_to_solution]. Then, execute train.py:

python train.py -ft ./assets/train.csv -e 1500 -bt 128 -r

Where:

• -ft, --file_train - path to the train CSV file
• -fv, --file_valid - path to the validation (if needed) CSV file
• -e, --epochs - number of training epochs
• -bt, --batch_size_train - size of the training batch
• -bv, --batch_size_valid - size of the validation batch
• -r, --resave - resave .pt files

Test

To test the model on the test dataset, execute test.py:

python test.py -p --path ./assets/test -s --scaler ./dict/scaler/scaler.pkl -m ./dict/dnn/best.pt

Where:

• -p, --path - path to the test folder
• -s, --scaler - path to the dataset scaler file (created during training)
• -m, --model - path to the model file

Contribution

Contributions to this project are welcome and appreciated. If you find any issues, have suggestions for improvements, or want to add new features, please feel free to submit a pull request.

Please ensure that your contributions align with the project's coding style and follow best practices. Also, make sure to test your changes thoroughly before submitting a pull request.

License

This project is licensed under the MIT License. Feel free to use and modify the code for your own purposes.