README.md

DiffPlace - A Conditional Diffusion Framework for Simultaneous VLSI Placement Beyond Sequential Paradigms

Paper:

Experimental Results

Below are the placement results on ISPD2005 benchmarks using DiffPlace.

Adaptec1	Adaptec2	Adaptec3	Adaptec4

BigBlue1	BigBlue2	BigBlue3	BigBlue4

Project Structure

diffplace/
├── train.py                 # Training script
├── scripts/
│   └── deploy.py            # Inference
├── engine/
│   ├── diffplace.py         # Main model (DiffPlace)
│   ├── models.py            # Legacy models
│   ├── utils.py
│   ├── diffusion/           # Diffusion components
│   ├── training/            # Training utilities
│   ├── networks/            # Neural networks
│   ├── datasets/            # Data loading
│   └── conf/                # Configs
└── data/
    └── ispd2005/            # Benchmarks (see README inside)

Quick Start

Training

python train.py --config engine/conf/pretrain.yaml --data_dir path/to/synthetic/data
python train.py --config engine/conf/finetune.yaml --dataset_type ispd --data_dir data/ispd2005

Inference

python scripts/deploy_nangate45.py \
  --checkpoint path/to/checkpoint.pt \
  --design_dir path/to/ariane133 \
  --output_pkl results/ariane133.pkl

Data

Download ISPD2005 benchmarks from DREAMPlace

Additional modern SoC benchmarks (Ariane, NVDLA, MemPool, BlackParrot) available from:

• TILOS MacroPlacement - Macro placement benchmarks and flows
• OpenROAD - Open-source RTL-to-GDS flow

Requirements

pip install -r requirements.txt