BBOPlace-miniBench: Mini Benchmarking Black-Box Optimization for Chip Placement

This repository contains the Python code for BBOPlace-miniBench, a mini benchmarking of BBOPlace-Bench without DREAMPlace .

Requirements

• numpy==1.24.4
• ray==2.10.0
• matplotlib==3.7.5
• igraph==0.11.8
• pymoo==0.6.1.2
• torch==1.7.1
• torchvision==0.8.2
• gpytorch==1.4.2
• botorch==0.4.0
• pypop7==0.0.82
• torch==2.4.1
• torchaudio==2.4.1
• torchvision==0.19.1
• gpytorch==1.10
• botorch==0.8.5
• pypop7==0.0.82

File structure

• benchmarks directory stores the benchmarks for running. Please download ISPD2005 and ICCAD2015 benchmarks and move them to benchmarks/ (i.e., benchmarks/ispd2005/adaptec1, benchmark/iccad2015/superblue1).
• config stores the hyperparameters for algorithms.
• script contains scripts for code running.
• src contains the source code of our benchmarking.

Usage

Please first build the environment according to the requirements and download benchmarks via google drive: ISPD2005, ICCAD2015.

Usage Demo and Evaluation APIs (for BBO Users)

We provide a simple demo of how to benchmark BBO algorithms (e.g., CMA-ES) on our proposed BBOPlace-Bench in src/demo_cmaes.py. Specifically, we implement a BBO user-friendly evaluator interface in src/evaluator.py, and you can instantiate it for fitness evaluation as

# Set base_path environment variable
import os, sys
base_path = os.path.abspath(".") # Alternative: base_path should be where BBOPlace-miniBench is located
sys.path.append(base_path)

import numpy as np 
from types import SimpleNamespace
from src.evaluator import Evaluator
from utils.args_parser import parse_args

args = SimpleNamespace(
    **{
        "placer": "gg", # GG in our paper
        "benchmark": "ispd2005/adaptec1", # choose which placement benchmark
    } 
)

# Read config (i.e. benchmark, placer)
args = parse_args(args)

# Instantiate the evaluator
evaluator = Evaluator(args)

# Read problem metadata
dim: int = evaluator.n_dim
xl: np.ndarray = evaluator.xl
xu: np.ndarray = evaluator.xu
assert len(xl) == len(xu) == dim

# Hpwl evaluation API
batch_size = 16
x = np.random.uniform(low=xl, high=xu, size=(batch_size, dim))
hpwl, macor_pos = evaluator.evaluate(x)
print(np.max(hpwl), np.min(hpwl), np.mean(hpwl))

where choices for placer are gg, sp, which refer to GG, SP formulations in our paper, respectively. The choices for benchmark are

benchmarks = ["ispd2005/adaptec1", "ispd2005/adaptec2", "ispd2005/adaptec3", "ispd2005/adaptec4", "ispd2005/bigblue1", "ispd2005/bigblue3",   # ISPD 2005
              "iccad2015/superblue1", "iccad2015/superblue3", "iccad2015/superblue4", "iccad2015/superblue5",
              "iccad2015/superblue7", "iccad2015/superblue10", "iccad2015/superblue16", "iccad2015/superblue18"]   # ICCAD 2015

Search Space Statement

For GG formulation, we formulate it as continuous BBO problems. For SP formulation, it is a permutation-based BBO problem.

Reproduce Paper Results

Parameters Settings

Before running an experiment, you can modify the hyper-parameters of different problem formulations and algorithms in config directory. For example, modifying n_population : 50 in config/algorithm/ea.yaml to change the population size of EA.

Quick Run

Run run.sh file in the script directory.

cd script