OpenABC-D: A Large-Scale Dataset For Machine Learning Guided Integrated Circuit Synthesis

Update

1. Full dataset (Initial, intermediate and final optimized AIGs) is hosted at OpenABC-D. All relevant bench, graphml, label CSVs and run logs are stored here. The entire zip is 1.4TB so it has been equally divided into 14 chunks, each of 107GB. For downloading and unzipping, minimum of 3TB disk space is required.

2. ML-ready dataset (Pytorch files for initial AIG and synthesis recipe information + target labels as pickled file ~ 19GB) is hosted here. All one need is to configure the paths as documented and run the models for QoR prediction.

3. Pytorch-geometric has newly released versions 2.0.* and is not backward compatible. Thus, to use the already existing pytorch data, please install Pytorch-geometric version < 2.0.* or regenerate the pt data files using the dumped graphml files.

4. Original AIGs of designs have been added in the repository under bench_openabcd.

Overview

OpenABC-D is a large-scale labeled dataset generated by synthesizing open source hardware IPs using state-of-art logic synthesis tool yosys-abc. We consider 29 open-source hardware IP designs collected from various sources (MIT-CEP, IWLS, OpenROAD, OpenPiton etc) and synthesized them with 1500 random synthesis flows (we call them synthesis recipes).

Each synthesis flow has a predefined length L (L=20, in our case). We preserved all AIGs: starting, intermediate and final AIGs with labels like number of nodes, longest path, sequence of atomic synthesis transofrmations (rewrite, refactor, balance etc.) along with graph statistics, area and delay of final AIG.

We converted the AIGs in pytorch data format that can be directly used by a machine learning engineer lessening the effort of costly labeled data generation and pre-processing. OpenABC-D can be used for a variety of learning tasks on logic synthesis such as

1. Predicting quality of result (QoR) performance of a synthesis recipe on a hardware IP.
2. Area and delay prediction post techonolgy mapping.
3. Learn functional and structural features of AIG using self-supervised labels (useful for tasks like RL-based logic synthesis)

Our dataset can easily be used for graph-based machine learning framework like Pytorch-Geometric. The data generation pipeline of OpenABC-D is shown as follows:

Installing dependencies

We recommend using venv or Anaconda environment to install pre-requisites packages for running our framework and models. We list down the packages which we used on our side for experimentations. We recommend installing the packages using requirements.txt file provided in our repository.

• cudatoolkit = 10.1
• numpy >= 1.20.1
• pandas >= 1.2.2
• pickleshare >= 0.7.5
• python >=3.9
• pytorch = 1.8.1
• scikit-learn = 0.24.1
• torch-geometric=1.7.0
• tqdm >= 4.56
• seaborn >= 0.11.1
• networkx >= 2.5
• joblib >= 1.1.0

Here are few resources to install the packages (if not using requirements.txt)

• Pytorch
• Torch-geometric
• Networkx

Make sure that that the cudatoolkit version in the gpu matches with the pytorch-geometric's (and dependencies) CUDA version.

Organisation

Dataset directory structure

├── OPENABC_DATASET
│   ├── bench			# Original and synthesized bench files. Log reports post technology mapping
│   ├── graphml                # Graphml files
│   ├── lib			# Nangate 15nm technology library
│   ├── ptdata			# pytorch-geometric compatible data
│   ├── statistics		# Area, delay, number of nodes, depth of final AIGs for all designs
│   └── synScripts		# 1500 synthesis scripts customized for each design

1. In bench directory, each design has a subfolder containing original bench file: design_orig.bench, a log folder containing log runs of 1500 synthesis recipes, and syn.zip file containing all bench files synthesized with synthesis recipe N.

2. In graphml directory, each design has subfolder containing zipped graphml files corresponding to the bench files created for each synthesis runs.

3. In lib directory, Nangate15nm.lib file is present. This is used for technology mapping post logic minimization.

4. In ptdata directory, we have subfolders for each design having zipped pytorch file of the format designIP_synthesisID_stepID.pt. Also, we kept train-test split csv files for each learning tasks in subdirectories with naming convention lr_ID.

5. In statistics diretcory, we have two subfolders: adp and finalAig. In adp, we have csv files for all designs with information about area and delay of final AIG post tech-mapping. In finalAig, csv files have information about graph characteristics of final AIGs obtained post optimization. Also, there is another file named synthesisstastistics.pickle which have all the above information in dictionary format. This file is used for labelling purpose in ML pipeline for various tasks.

6. In synScripts directory, we have subfolders of each design having 1500 synthesis scripts.

Data generation

├── datagen
│   ├── automation 			      # Scripts for automation (Bulk/parallel runs for synthesis, AIG2Graph conversions etc.)
│   │   ├── automate_bulkSynthesis.py         # Shell script for each design to perform 1500 synthesis runs
│   │   ├── automate_finalDataCollection.py   # Script file to collect graph statistics, area and delay of final AIG
│   │   ├── automate_synbench2Graphml.py      # Shell script file generation to involking andAIG2Graphml.py
│   │   └── automate_synthesisScriptGen.py    # Script to generate 1500 synthesis script customized for each design
│   └── utilities
│       ├── andAIG2Graphml.py		      # Python utility to convert AIG BENCH file to graphml format
│       ├── collectAreaAndDelay.py            # Python utility to parse log and collect area and delay numbers
│       ├── collectGraphStatistics.py         # Python utility to for computing final AIG statistics
│       ├── pickleStatsForML.py               # Pickled file containing labels of all designs (to be used to assign labels in ML pipeline)
│       ├── PyGDataAIG.py		      # Python utility to convert synthesized graphml files to pytorch data format
│       └── synthID2SeqMapping.py	      # Python utility to annotate synthesis recipe using numerical encoding and dump in pickle form

1. automation directory contains python scripts for automating bulk data generation (e.g. synthesis runs, graphml conversion, pytorch data generation etc.). utilities folder have utility scripts performing various tasks and called from automation scripts.

2. Step 1: Run automate_synthesisScriptGen.py to generate customized synthesis script for 1500 synthesis recipes. One can see the template of a synthesis recipe in referenceScripts under synScripts folder.

3. Step 2: Run automate_bultkSynthesis.py to generate a shell script for a design. Run the shell script to perform the synthesis runs. Make sure yosys-abc is available in PATH.

4. Step 3: Run automate_synbench2Graphml.py to generate a shell script for generating graphml files. The shell script invokes andAIG2Graphml.py using 21 parallel threads processing data of each synthesis runs in sequence.

5. Step 4: Run PyGDataAIG.py to generate pytorch data for each graphml file of the format designIP_synthesisID_stepID.pt.

6. Step 5: Run collectAreaAndDelay.py and collectGraphStatistics.py to collect information about final AIG's statistics. Post that, run pickleStatsForML.py which will output synthesisStatistics.pickle file.

7. Step 6: Run synthID2SeqMapping.py utility to generate synthID2Vec.pickle file containing numerically encoded data of synthesis recipes.

Benchmarking models: Training and evaluation

├── models
│   ├── classification
│   │   └── ClassNetV1
│   │       ├── model.py			# Graph convolution network based architecture model
│   │       ├── netlistDataset.py		# Dataset loader
│   │       ├── train.py			# Train and evaluation utility
│   │       └── utils.py			# Utitility functions
│   └── qor
│       ├── NetV1
│       │   ├── evaluate.py
│       │   ├── model.py
│       │   ├── netlistDataset.py
│       │   ├── train.py
│       │   └── utils.py
│       ├── NetV2
│       │   ├── evaluate.py
│       │   ├── model.py
│       │   ├── netlistDataset.py
│       │   ├── train.py
│       │   └── utils.py
│       └── NetV3
│           ├── evaluate.py
│           ├── model.py
│           ├── netlistDataset.py
│           ├── train.py
│           └── utils.py

models directory contains the benchmarked model described in details in our paper. The names of the python utilities are self explainatory.

Case 1: Prediction QoR of a synthesis recipe

We recommend creating a following folder hierarchy before training/evaluating a model using our dataset and model codes:

├── OPENABC-D
│   ├── lp1
│   │   ├── test_data_set1.csv
│   │   ├── test_data_set2.csv
│   │   ├── test_data_set3.csv
│   │   ├── train_data_set1.csv
│   │   ├── train_data_set2.csv
│   │   └── train_data_set3.csv
│   ├── lp2
│   │   ├── test_data_set1.csv
│   │   └── train_data_set1.csv
│   ├── processed
│   ├── synthesisStatistics.pickle
│   └── synthID2Vec.pickle

OPENABC-D is the top level directory containing the dataset, train-test split files, and labeled data available. Transfer all the relevant zipped pytorch data in the subdirectory processed.

The user can now go the models directory and run codes for training and evaluation. An example run for dataset split strategy 1 (Train on first 1000 synthesis recipe, predict QoR of next 500 recipe)

python train.py --datadir $HOME/OPENABC-D --rundir $HOMEDIR/NETV1_set1 --dataset set1 --lp 1 --lr 0.001 --epochs 60 --batch-size 32

Setting lp=1 and dataset=set1 will pick appropriate train-test split strategy dataset for QoR regression problem. The model will run for 60 epochs and report the training, validation and test performance on the dataset outputing appropriate plots.

Similarly for split-strategy 2 and 3, one can set the dataset as set2 and set3 respectively.

For evaluating performance of specific model on a custom curated dataset, a user can create appropriate csv file with dataset instances and add it to dictionary entry in train.py. For evaluating existing dataset split, one can run the following code.

python evaluate.py --datadir $HOME/OPENABC-D --rundir $HOMEDIR/NETV1_set1 --dataset set1 --lp 1 --model "gcn-epoch20-loss-0.813.pt" --batch-size 32

The test-MSE performance we obtained on our side are as follows:

Net Type	Case-I	Case-II	Case-III
NetV1	0.648+-0.05	10.59+-2.78	0.588+-0.04
NetV2	0.815+-0.02	1.236+-0.15	0.538+-0.01
NetV3	0.579+-0.02	1.470+-0.14	0.536+-0.03

How to cite

If you use this code/dataset, please cite:

@misc{chowdhury2021openabcd,
      title={OpenABC-D: A Large-Scale Dataset For Machine Learning Guided Integrated Circuit Synthesis}, 
      author={Animesh Basak Chowdhury and Benjamin Tan and Ramesh Karri and Siddharth Garg},
      year={2021},
      eprint={2110.11292},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}