Skip to content

Latest commit

 

History

History
147 lines (99 loc) · 6.98 KB

README.md

File metadata and controls

147 lines (99 loc) · 6.98 KB
Raw

CircuitOps

(Notice: This Github repository is currently in the progress of being developed and some files are still missing. We will try to complete it ASAP. Thanks for your patience!)

Introduction

CircuitOps is a data infrastructure to facilitate dataset generation and model deployment in Generative AI (GAI)-based circuit optimization tasks. It mainly has the following contributes:

  1. Shared Intermediate Representation: labeled property graphs (LPGs) backed by relational tables, a flexible way to represent detailed netlist information and suitable for parallel processing;
  2. Tools-agnostic IR generation: parsing standard EDA files and attributes tables to generate LPGs for given netlists, which can be reused for many downstream GAI applications;
  3. Customizable dataset generation: generating a customized dataset for each GAI application by performing on the AI-friendly LPGs;
  4. Inference with gRPC-based data transfer: facilitating the deployment of GAI models into production.

Figure.1 depicts the overview of CircuitOps. Based on the Intermediate Representation of labeled property graphs, CircuitOps consists of two main modules: IR generation and dataset generation. The IR generation module transforms standard EDA files into LPGs that store netlist information and are reused across tasks. The taskspecific dataset is constructed with the dataset generation module using its AI-friendly data structures and interfaces. CircuitOps also provides a gRPC-based data transfer method facilitating inference of GAI models in production deployment.

Markdown Monster icon

Fig. 1: CircuitOps overview. (a) shows the structure of CircuitOps; (b) illustrates the netlist labeled property graph backed by relational tables.

Getting Started

Download the CircuitOps repository as shown below:

git clone --recursive https://github.com/NVlabs/CircuitOps.git
cd CircuitOps

Install CircuitOps

Dependencies

The following dependencies are needed by CircuitOps. OpenROAD is required for EDA tools file parsing and generating properties.

  • python3.7
  • pip3
  • OpenROAD

Install OpenROAD

Refer to the dependencies of the OpenROAD Project and instrcutions here.

We use OpenROAD to read in standard EDA files and generate relational tables as IRs.

TLDR instructions to build OpenROAD is listed below:

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

Install CircuitOps in Bash

From the IRs, CircuitOps uses the relational tables generated from OpenROAD and creates LPGs and datasets. Installation of Python scripts of Circuit ops in described below through a virtual environment and pip. From the CircuitOps top level directory run the following commands:

python3 -m venv circuitops
source circuitops/bin/activate
pip3 install -r requirements.txt

Use CircuitOps

Generate IRs from OpenROAD using TCL

Set design and platform

Modify set_design.tcl to name the design and platform. If you need to add more designs, add them to the designs directory and modify the set_design.tcl file appropriately.

Also modify the fixed_load_cell in set_design.tcl, which provides the type of cell that should be used to calculate the fixed load delay for cells.

Run OpenROAD and TCL scripts to generate relational tables

The following command to generate the relations tables in the ./IRs/ directory.

./path/to/binary/openroad ./src/tcl/generate_tables.tcl

Generate IRs from OpenROAD using Python

Run the following command to generate the relations tables in the ./IRs/ directory.

./path/to/binary/openroad -python ./src/python/generate_tables.py -w 1 -d <design_name>  -t <tech_node>

Arguments of python script:
-w --> [0 | 1] Store IR tables into csv files. Default: 0
-d --> To provide the design name for which IR table should be generated. Default: "gcd"
-t --> To provide the technology node. Default: "nangate45"

Sample IR tables

There are IR tables available for a number of designs in Nangate45, asap7 and sky130hd tech nodes in this git repo. This can be used by engineers for ML applications.

The list of designs available are given in the table below along with post filler instance count and runtime to generate IR tables using python script for these designs.

Technode Design # of instances IR generation runtime (mins) Core utilisation
asap7 gcd 1387 0.15 Default ORFS
asap7 uart 1679 0.20 Default ORFS
asap7 mock-array_Element 7994 0.60 Default ORFS
asap7 ibex 48237 42.14 Default ORFS
asap7 NV_NVDLA_partition_m 65353 28.78 30
asap7 NV_NVDLA_partition_a 111207 104.04 30
asap7 jpeg 169095 197.99 Default ORFS
asap7 NV_NVDLA_partition_p 215140 305.55 30
asap7 NV_NVDLA_partition_c 499581 1755.31 30
nangate45 gcd 752 0.11 Default ORFS
nangate45 aes 30202 20.78 Default ORFS
nangate45 ibex 32111 23.44 Default ORFS
nangate45 bp_fe 96150 59.80 Default ORFS
nangate45 bp_be 141468 129.14 Default ORFS
nangate45 jpeg 141651 248.85 Default ORFS
nangate45 swerv 193054 608.56 Default ORFS
sky130hd gcd 1181 0.21 Default ORFS
sky130hd riscv32i 20104 7.41 Default ORFS
sky130hd ibex 42487 32.68 Default ORFS
sky130hd aes 64389 17.27 Default ORFS
sky130hd jpeg 140975 178.49 Default ORFS

Generate Datasets

cd src/python

python BT_sampling_OpenROAD.py ../../IRs/nangate45/gcd/ ../../datasets/

gRPC-based Data Transfer

Cite this work

  • R. Liang, A. Agnesina, G. Pradipta, V. A. Chhabria and H. Ren, "CircuitOps: An ML Infrastructure Enabling Generative AI for VLSI Circuit Optimization", 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD). (preprint)