Gamora: Graph Learning based Symbolic Reasoning for Large-Scale Boolean Networks (DAC'23)

Paper

https://arxiv.org/pdf/2303.08256.pdf

@article{wu2023gamora,
  title={Gamora: Graph Learning based Symbolic Reasoning for Large-Scale Boolean Networks},
  author={Wu, Nan and Li, Yingjie and Hao, Cong and Dai, Steve and Yu, Cunxi and Xie, Yuan},
  journal={Design Automation Conference (DAC'23)},
  year={2023}
}

A more generic framework for Boolean/RTL Graph Learning -- V2PYG

We are developing a more generic V2PYG framework to provide versatile capabilities for converting RTL designs into foundational graph representations, while also facilitating RTL-equivalent argumentation. https://ycunxi.github.io/Verilog-to-PyG

Installation

Prereq:

1) Python packages: torch and torch_geometric

2) Package: readline (follow ABC installation requirement)

sudo apt-get install libreadline6 libreadline6-dev (Ubuntu)

sudo yum install readline-devel (CentOS, RedHat)

Installation

compile ABC customized for graph learning

cd abc;make clean;make -j4

Pre-processed Dataset - CSA (techmapped) and Booth Multipliers

Check at Huggingface: https://huggingface.co/datasets/yucx0626/Gamora-CSA-Multiplier/tree/main

Implementation

1) Data generator

dataset_prep/dataset_generator.py
class ABCGenDataset

• gentype =0 CSA-array Multiplier generation and labeling
• gentype =1 CPA Adder generation and labeling
• gentype =2 Read a design and generate dataset
• gentype =3 Generate Booth-encoded multiplier (tbd)

Note: ABC is required (../abc) (make sure to create a link symbol of abc binary in this folder)

ln -s ../abc/abc .

ABC implementation for data generation.

// abc/src/proof/acec/acecXor.c
class Gia_EdgelistMultiLabel()

2) Train-Test Demo - training on 8-bit CSA and predicting on 32-bit CSA

• Dataset generation

  python ABC_dataset_generation.py --bits 8
  # generate an 8-bit CSA multiplier

  python ABC_dataset_generation.py --bits 32
  # generate a 32-bit CSA multiplier

• Training and inference

  python gnn_multitask.py --bits 8 --bits_test 32
  # training with mult8, and testing with mult32

• Inference with pre-trained model

  python gnn_multitask_inference.py --model_path SAGE_mult8 --bits_test 32 --design_copies 1
  # load the pre-trained model "SAGE_mult8", and test with mult32

Training INPUT: 8-bit CSA-Mult

Testing INPUT: 32-bit CSA-Mult

# training
Highest Train: 99.45
Highest Valid: 100.00
  Final Train: 98.90
   Final Test: 99.12

# testing
mult32
Highest Train: 0.00 ± nan
Highest Valid: 0.00 ± nan
  Final Train: 0.00 ± nan
   Final Test: 99.95 ± nan

New commands for ABC

    abc 01> edgelist -h
    usage: edgelist : Generate pre-dataset for graph learning (MPNN,GraphSAGE, dense graph matrix)
    -F : Edgelist file name (*.el)
    -c : Class map for corresponding edgelist (Only for GraphSAGE; must has -F -c -f all enabled)
    -f : Features of nodes (Only for GraphSAGE; must has -F -c -f all enabled)
    -L : Switch to logic netlist without labels (such as AIG and LUT-netlist)
    Example 1 (GraphSAGE dataset)
         read your.aig; edgelist -F test.el -c test-class-map.json -f test-feats.csv 
    Example 2 (Generate dataset for LUT-mapping netlist; unsupervised)
          read your.blif; strash; if -K 6; edgelist -L -F lut-test.el 
    Example 3 (Generate dataset for abstraction; supervised for FA/HA extraction  - GraphSAGE)
          read your.blif; strash; &get; &edgelist -F test.el -c test-class_map.json -f test-feats.csv