Scalable Streaming Accelerator of Graph Convolution Network

We develop a streaming-based accelerator to deal with large scale Graph Convolution Network.

Introduction

Graph convolutional network (GCN) is a kind of convolutional neural network that has the ability to directly working with graphs and their structural information. And it can be generalize to two matrix operation: GEMM and SpMM. We seek to design an accelerator which can deal with real world graph (which is irregular and large).

Major Optimizations

In order to compute large graph on the limited FPGA resource, we use the following strategy and method to utilize the on-chip memory and optimize the dataflow.

1. Matrix Tiling
2. Streaming Dataflow

For GEMM implementation, we use the code from https://spcl.inf.ethz.ch/Teaching/hls-tutorial/

As the result, We now can control the Bram utilization used by the FPGA and scale-up the problem-set.

Fold Structure

.
├── docs                    # Documentation files 
├── src                     # Source files 
├── tb                      # Automated tests 
├── LICENSE
└── README.md

Rebuild the Project

Requirements

• Test on Windows 10
• Vitis HLS 2021.2 and Vivado 2021.2

Procedure

In Vitis HLS

choose the board you use(We test on pynq-z2)

Add src files into source, Add tb files into Testbench

In project settings choose GCN.cpp as Top function

---

parameter setting

In GCN.h, for SPMM
set SIZE for square adjcency matrix A,
set SIZE_K for the H's col size ,
set SIZE_N for the W's col size ,
set NNZ for the number of non-zero

For partition, the value to set need to satisfy below equation
BLOCK_SIZE = SIZE / PARTITION_NUM
which SIZE must be divided by PARTITION_NUM

In GEMM.h for GEMM
set M for matrix a's row
set K for matrix a's col
set N for matrix b's col

---

Finally, go through sim -> synthesis -> cosim

Testing

Testcase generation: python3 testcase.py

run sim for functional testing run cosim for rtl testing