H-GNN

Introduction

H-GNN is a novel approach designed to enhance high-level synthesis (HLS) by providing efficient post-route quality of results (QoR) predictions. HLS significantly accelerates the hardware design process by eliminating the need for RTL programming. However, considering post-route QoR during optimization can lead to increased turnaround times. To address this challenge, we propose a hierarchical post-route QoR prediction approach specifically for FPGA HLS. Our method features:

1. Modeling Flow: Direct estimation of latency and post-route resource usage from C/C++ programs.
2. Graph Construction: An effective method that captures the control and data flow graph of source code along with the effects of HLS pragmas.
3. Hierarchical GNN Training: A training and prediction mechanism that effectively captures the impact of loop hierarchies.

Experimental results demonstrate that our method achieves a prediction error of less than 10% across various QoR metrics, marking a significant improvement over existing GNN methods.

Our paper, "Hierarchical Source-to-Post-Route QoR Prediction in High-Level Synthesis with GNNs," is presented at DATE '24. Please refer to our paper for more details:

@inproceedings{h-gnn,
  author    = {Mingzhe Gao and Jieru Zhao and Zhe Lin and Minyi Guo},
  title     = {Hierarchical Source-to-Post-Route QoR Prediction in High-Level Synthesis with GNNs},
  journal   = {Design, Automation and Test in Europe Conference},
  year      = {2024},
}

Overview

In the training stage, we implement the C-to-bitstream flow to derive ground-truth QoR using various HLS pragma combinations. We gather resource stats from implementation reports and latency from HLS reports. Starting with C/C++ source code, we convert it to LLVM IR and construct a graph representation with different pragma configurations. Features are extracted and annotated in the graph, leading to dataset generation. Finally, we train GNN models for accurate post-route QoR prediction, utilizing loop hierarchies. In the inference phase, predictions are made without HLS or RTL flows, enabling efficient DSE to find optimal pragma configurations for performance.

Requirement

You can install the required packages for running this project using:

pip3 install --upgrade pip
pip3 install -r requirements.txt

This project is built on LLVM 9.0 that can be installed by:

wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh <version number>

Project File Tree

The project file structure is as below:

.
├── GNNg
│   ├── data               # Dataset for GNNg
│   └── src                # Training workflow for GNNg
├── GNNnp
│   ├── data               # Dataset for GNNnp
│   └── src                # Training workflow for GNNnp
├── GNNp
│   ├── data               # Dataset for GNNp
│   └── src                # Training workflow for GNNp
├── benchmark
│   ├── auto_directive.py   # Script for automatically adding combinations of different pragmas
│   ├── kernel4graph        # Kernel code used to generate graphs
│   ├── kernel4hls          # Kernel code for extracting post-route ground truth data
│   └── run_script.py       # Script to execute benchmarking
├── requirements.txt        # List of project dependencies
└── utils
    ├── llvm-pass           # Tools for calculating iteration interval
    └── visualization        # Tools for graph visualization

Build Datasets

To build the GNNp dataset, navigate to the directory and run the following commands:

cd GNNp/data/datasets/
python run.py

The generated data files (.pt) will be saved in GNNp/data/processed/pt, while the corresponding graphs for visualization will be stored in GNNp/data/processed/graph. The dataset construction for GNNnp and GNNg follows a similar process.

Train GNN Model

To train the GNNp model, navigate to the directory and run the following commands:

cd GNNp/data/src/
python main.py

You can modify the training hyperparameters (e.g., batch size) and objectives (i.e., LUT, FF, DSP, latency) in GNNp/data/src/config.py. The training process and results will be saved in the logs folder. The model training for GNNnp and GNNg follows a similar process.