BBOPlace-Bench: Benchmarking Black-Box Optimization for Chip Placement

Official implementation of paper "BBOPlace-Bench: Benchmarking Black-Box Optimization for Chip Placement".

This repository contains the Python code for BBOPlace-Bench, a benchmarking containing BBO algorithms for chip placement tasks.

Installation

Important Reminder:

If you find it difficult to install DREAMPlace, we have a basic version available that works out of the box without DREAMPlace. Check it out at: BBOPlace-miniBench

Option 1: Using pip

# Create a virtual environment (recommended)
python -m venv bboplace-env
source bboplace-env/bin/activate

# Install requirements
pip install -r requirements.txt

Option 2: Using Docker

Please first install docker and docker cuda toolkit.

# Load our docker image
docker load --input bboplace-bench.tar

# Find our docker image and rename it
docker images
docker tag <IMAGE ID> <Your Name>/bboplace-bench:latest

# Run a docker container
docker run --gpus all -it -v $(pwd):/workspace <Your Name>/bboplace-bench:latest bash

File Structure

• benchmarks/ - Stores the benchmarks for running. Please download ISPD2005 and ICCAD2015 benchmarks and move them here.
• config/ - Stores the hyperparameters for algorithms.
• script/ - Contains scripts for code running.
• src/ - Contains the source code of our benchmarking.
• thirdparty/ - Serves as a thirdparty standard placer borrowed from DREAMPlace.
• demo/ - Contains example usage demos.

Dataset Preparation

Download the required benchmark datasets:

• ISPD2005 Benchmarks
• ICCAD2015 Benchmarks

Extract and place them in the benchmarks/ directory:

benchmarks/
├── ispd2005/
│   ├── adaptec1/
│   ├── adaptec2/
│   └── ...
└── iccad2015/
    ├── superblue1/
    ├── superblue3/
    └── ...

Usage

Quick Start Demo

We provide a simple demo of how to benchmark BBO algorithms (e.g., CMA-ES) in demo/demo_cmaes.py:

from src.evaluator import Evaluator
import numpy as np

# Initialize evaluator
args = SimpleNamespace(
    placer="grid_guide",  # Options: grid_guide, sp, dmp
    benchmark="adaptec1", 
    eval_gp_hpwl=False
)
evaluator = Evaluator(args)

# Get problem dimensions and bounds
dim = evaluator.n_dim
xl, xu = evaluator.xl, evaluator.xu

# Evaluate solutions
x = np.random.uniform(low=xl, high=xu, size=(128, dim))
hpwl = evaluator.evaluate(x)

Available Benchmarks

benchmarks = [
    # ISPD 2005
    "adaptec1", "adaptec2", "adaptec3", "adaptec4",
    "bigblue1", "bigblue3",
    # ICCAD 2015   
    "superblue1", "superblue3", "superblue4", "superblue5",
    "superblue7", "superblue10", "superblue16", "superblue18"
]

Problem Formulations

• GG (Grid Guide): Continuous optimization
• SP (Sequence Pair): Permutation-based optimization
• HPO (DMP): Continuous optimization

Configuration

Modify algorithm hyperparameters in config/ directory:

# config/algorithm/ea.yaml example
n_population: 50
n_generation: 100

Wandb Integration

To use Weights & Biases logging:

# config/default.yaml
use_wandb: True
wandb_offline: True  # Set True for offline mode
wandb_api: "<Your API KEY>"

Note: Set n_cpu_max=1 when using dmp placer or evaluating global placement HPWL.

Running Experiments

Execute the shell scripts in script/ directory:

cd script
./run_experiment.sh

License

This project is licensed under the MIT License - see the LICENSE file for details.