2_prediction

Prediction

In this part, we build a runtime prediction model based on Graph Convolutional Network. The model takes as input a graph represnetation of the design. In this work, we use Deep Graph Library to build the model. Below, we provide instructions for testing out the training process of our model.

Setup

First, install the dependencies of the model.

virtualenv .venv --python=python3
source .venv/bin/activate
pip install -r requirements

Processing Input Designs as Graphs

The first step is that to parse an input design as graph representation suitable for the dgl library. There are two different graphs that we operate on: a desing in RTL, and a design in netlist format.

We provide sample data in the data folder. Note that the data is unnormalized. You may skip this section if you just want to try out our model. However, if you want to use it on your own dataset, we provide the details below.

RTL Designs

First, install [Yosys] and make sure that both yosys and yosys-abc commands are available on your $PATH (i.e. you can call them from directory).

The file aig.py provides wrappers around this functuonality. It uses yosys to read a verilog in RTL format and convert it to and AIG graph. From the AIG graph, we construct a dgl graph. Below is the high-level function that you can use to build your own graphs.

from aig import read_verilog

G = read_verilog('/path/to/design.v')

where G is a graph representation in dgl format. You may save this graph in a binary file for future loading (similar to how we provided the data directory) using:

from dgl.data.utils import save_graphs

save_graphs("./design.dgl", [G], {})

Netlist Designs

NOTE ❗

Parsing netlists depends on the python package opendbpy. This package is only compiled for CentOS 7.

First, install opendbpy using pip install opendbpy.

The file netlist.py provides wrappers around this functionality. It basically uses opendbpy to parse LEF and DEF files (standard formats for representing netlists). Then, it constucts a dgl graph using a star model (converting each net into directed edges from the driving cell to the sinks). Below is the high-level function that you can use to build up your graphs.

from netlist import read_netlist

G = read_netlist('/path/to/libary.lef', '/path/to/design.def')

where G is a graph representation in dgl format. Similarly, you may save this graph in a binary file for future loading (similar to how we provided the data directory) using:

from dgl.data.utils import save_graphs

save_graphs("./design.dgl", [G], {})

Dataset

We provide a dataset for pre-processed designs in data/dgl.zip. Each file has a name and an ID that represents the design itself.

In addition, we provide the runtimes on 8 vCPUs in data/train.csv and data/test.csv for training and testing datasets respectively.

This dataset is loaded in train.py file.

Training

The model is defined in model.py.

To start the training loop, run:

python train.py

To see all training options, run:

python train.py --help