search

ASU-VDA-Lab / ASP-DAC24-Tutorial

This GitHub repo is for the OpenROAD and CircuitOps Tutorial at ASP-DAC 2024
BSD 3-Clause "New" or "Revised" License
22 stars 6 forks source link

readme

ASP-DAC24 Tutorial-8: CircuitOps and OpenROAD: Unleashing ML EDA for Research and Education

This is a GitHub repository that has the scripts being demoed at the tutorial. The scripts highlight example use cases of ML EDA infrastructure.

Background

Over the last decade, there has been a significant rise in machine learning (ML)--based electronic design automation (ML-EDA) research. However, a notable challenge lies in the interaction between existing EDA tools and ML frameworks. Researchers often use TCL scripts for interaction with EDA tools and Python scripts for ML frameworks. They rely on file I/O for interaction between the EDA world and ML world or reinvent the wheel and implement the complete EDA tools in Python. Both these approaches have challenges. The first is extremely slow, and it's impossible to send data/information between EDA tools iteratively, which is a barrier to entry for non-chip designers. The second slows down research and makes it challenging to make apples-to-apples comparisons between various ML EDA algorithms. As a remedy, there is a pressing need for an ML-friendly EDA tool that seamlessly integrates with ML methodologies, reducing reliance on TCL programming and file-based information extraction.

This tutorial introduces the CircuitOps and OpenROAD Python APIs, representing a significant milestone in ML-EDA research. CircuitOps is an ML-friendly data infrastructure utilizing labeled property graphs (LPGs) backed by relational tables (IR tables) to generate datasets for ML-EDA applications. The Python-compatible LPG minimizes the developmental effort required for ML-EDA research. The IR tables are generated using OpenROAD Python APIs, which offer numerous advantages over TCL APIs. One key advantage is that OpenROAD can now be directly imported into a Python environment, this not only means we can incorporate OpenROAD with other Python libraries, but we can also interact with OpenROAD in real-time, providing unprecedented flexibility in ML-EDA application development. Additionally, we can get the information directly via the Python API without file IO, which increases ML-EDA performance.

Tutorial Contents

This tutorial is composed of two hands-on sessions, each with two demos:

Getting Started

Live at ASP-DAC 24

If attending the 2024 ASP-DAC tutorial live and have emailed your public SSH key to us for Google Cloud Computing resources, you will have received an email with a user-name and IP address please login from a Linux/MacOS terminal or Window CMD using the following command:

ssh -Y <user-name>@<ip-address>

We have already cloned the repository in your user directory, and installed the required software dependencies and built the OpenROAD application. You can directly jump to running the scripts as described in session 1 and session 2.

Everywhere else (Not attending the ASP-DAC tutorial)

If you are not attending the tutorial, you are required to clone this repository, install the required software dependencies, and build OpenROAD and CircuitOps.

Clone the repository

git clone --recursive https://github.com/ASU-VDA-Lab/ASP-DAC24-Tutorial

Build OpenROAD and CircuitOps

Option 1: Build using Docker

The following technique assumes you have docker installed on your machine. If you do not have then install docker from here. Build the docker image and run using the following commands:

docker build -t <image_name> .
docker run -it --name <container_name> <image_name>
Option 2: Build locally

The following technique assumes you have a machine with the required Ubuntu OS prerequisite of OpenROAD and CircuitOps.

Install dependencies for OpenROAD:

sudo ./OpenROAD/etc/DependencyInstaller.sh

Install dependencies for CircuitOps and ML EDA applications:

sudo apt-get install -y python3-matplotlib
sudo apt-get install -y nvidia-cuda-toolkit
sudo apt-get update
sudo apt-get install -y python3-graph-tool
sudo apt-get update && apt-get install -y gnupg2 ca-certificates
sudo apt-get install -y python3-pip
pip3 install torch
pip3 install dgl
pip3 install pycairo
pip3 install pandas
pip3 install scikit-learn

Once packages have been installed, build OpenROAD:

cd ./OpenROAD/
mkdir build
cd build
cmake ..
make -j

Running the Demo Scripts

The following script assumes that we work from the ASPDAC2024-Tutorial directory. Before testing the script please select the tutorial directory.

cd <Path to ASP-DAC-24-Tutorial Directoy>

Session 1

This session demonstrates OpenROAD Python APIs in ML-EDA applications. The session contains two demos:

Demo 1

The first demo shows the Python APIs to run EDA tools flow and APIs to query the OpenROAD database. These APIs can work within a regular Python shell and easily integrate with ML-EDA applications. 

./OpenROAD/build/src/openroad -python session1/demo1_flow.py

OpenROAD Python APIs circuit properties query example:

./OpenROAD/build/src/openroad -python session1/demo1_query.py

Demo 2

The second demo shows the Python APIs used in two ML EDA applications. The first is for IR drop prediction. The script creates a power map within OpenROAD Python interpreter and performance inference using U-Net model. The second shows the use of Python APIs with an RL gate sizing framework. The GNN takes an action to update the size of the netlist in the OpenROAD database (OpenDB) and uses OpenROAD timer (OpenSTA) to estimate the reward.

To run the image-based static IR Drop prediction prediction using OpenROAD Python:

./OpenROAD/build/src/openroad -python session1/demo2_IR.py

To run the RL-based gate sizing example using OpenROAD Python:

./OpenROAD/build/src/openroad -python session1/demo2_gate_sizing.py

Session 2

This session demonstrates CircuitOps data structure which is an Intermediate Representation (IR) and includes labeled property graph (LPG) backed by relational tables.

Demo 3

The first demo in this session highlights LPG creation and IR table generation using OpenROAD and it is followed by examples of using Python APIs for querying LPG.

IR Tables generation using OpenROAD: 

cd CircuitOps
../OpenROAD/build/src/openroad -python ./src/python/generate_tables.py
cd ../

There are two ways to create LPGs:

(1) Create LPG via IR Tables

python3 session2/demo3_LPG_query_example.py

(2) Create LPG via OpenROAD Python API

./OpenROAD/build/src/openroad -python session2/demo3_LPG_query_example.py --use_pd

Demo 4

The second demo in this session showcases an example of using CircuitOps data representation format for easy data generation for net delay prediction.

To run this example:

./OpenROAD/build/src/openroad -python session2/demo4_preroute_net_delay_prediction.py