NetlistTomography

This repository contains the implementation for the paper: "Netlist Tomography-Based Soft Guide Generation with Refinement Across P&R Stages for Improved PPA".

Overview

This codebase provides:

• Complete synthesis and place-and-route (P&R) flows for benchmark designs
• Netlist tomography feature extraction tools
• TomoGNN architecture and training pipeline
• Soft guide generation for P&R tools
• Soft guide refinement algorithms

Methodology Flow

The figure above illustrates our complete methodology. The following table maps each component to its implementation in this repository:

Flow Component	Scripts/Directories	Description
Synthesis & Initial P&R	flow/<design>/run.sh flow/<design>/run_genus_hybrid.tcl flow/<design>/run_invs.tcl	RTL synthesis and baseline placement & routing
Netlist Tomography	scripts/tomo/gen_netlist_tomo.sh scripts/tomo/run.sh scripts/tomo/syn_map_*.py	Extract placement-to-synthesis mappings
TomoGNN	tomoGNN/main_pipeline.py tomoGNN/train.py tomoGNN/models/losses.py	Self-supervised training of tomoGNN with PPA-aware contrastive loss
Soft Guide Generation	tomoGNN/utils/cluster_hdbscan.py	Generate initial soft guides from tomoGNN predictions
Soft Guide Refinement	flow/util/optimize_clustering.py flow/util/analyze_path_clustering.py	Refine soft guides using timing criticality and physical location
DEF Generation	tomoGNN/generate_def.py flow/util/generate_cluster_def.py	Create soft guide DEF files for P&R tools
P&R with Guides	flow/<design>/run_invs.tcl	Apply soft guides and re-run place & route for improved PPA

Directory Structure

NetlistTomography/
├── flow/                   # EDA tool flows and testcases
│   ├── ariane/             # Ariane RISC-V CPU design
│   ├── bp_quad/            # BlackParrot quad-core processor
│   ├── jpeg_encoder/       # JPEG encoder accelerator
│   ├── swerv_wrapper/      # SweRV RISC-V core wrapper
│   ├── pdk_ng45/           # NanGate45 PDK files
│   └── util/               # Shared utility scripts
├── scripts/
│   └── tomo/               # Netlist tomography extraction scripts
├── tomoGNN/                # tomoGNN architecture and training pipeline
│   ├── data/               # Graph construction and data loading
│   ├── models/             # tomoGNN architecture
│   └── utils/              # Soft guide generation and sampling utilities
└── paper/                  # LaTeX source for the paper

See flow/README.md for detailed information about running the EDA flows.

Quick Start

Prerequisites

• EDA Tools: Cadence Genus (synthesis) and Innovus (P&R)
• Python: 3.8+ with PyTorch and PyTorch Geometric
• Environment: Linux with modules support (for tool loading)

Running the Flow

1. Navigate to a design directory:

   cd flow/ariane

2. Load required modules:

   module load genus
   module load innovus

3. Execute the flow:

   bash run.sh

See flow/README.md for detailed instructions.

Generating Netlist Tomography Features

cd scripts/tomo
bash gen_netlist_tomo.sh <design_name> <run_directory>

Training tomoGNN Model

cd tomoGNN
# Setup environment
bash setup_env.sh
# Run training
bash run.sh --config config.py

See tomoGNN/README.md for detailed training options.

Dependencies

EDA Tools

• Cadence Genus (synthesis)
• Cadence Innovus (place and route)

Python Packages

See tomoGNN/requirements.txt for the complete list. Key dependencies:

• PyTorch >= 1.10
• PyTorch Geometric >= 2.0
• NetworkX
• NumPy, Pandas
• scikit-learn
• HDBSCAN, Leidenalg

PDK

The flows use NanGate45 open-source PDK (included in flow/pdk_ng45/).

License

This project is licensed under the BSD 3-Clause License - see the LICENSE file for details.