The contest of the High-Performance Graph Analytics course aimed to enhance the performance of a sequential CPU implementation of a Graph Convolutional Network (GCN) in C++ for Semi-Supervised Classification. The provided code was based on the original GCN model developed by Kipf et al. The goal was to leverage GPU acceleration without using CUDA libraries while maintaining a satisfactory level of accuracy compared to the baseline implementation.
The contest consisted of two main steps:
- GPU acceleration: Focus on the acceleration of the workload;
- Model optimization: Focus on exploring various approaches to modify the model or experiment with different hyperparameters to improve the accuracy of the predictions.
In /original-code you can can find the c++ code that we had to optimize, in /gpu_track there is the code for the first part, and in /model-optimize our solution for the second part.
You can build and execute the existing implementation by running the following commands:
make
./exec/gcn-seq cora # dataset name from the dataset folder: [cora, pubmed, citeseer, reddit]In every folder there is a colab.ipynb file: click on the "open in Colab" button to execute the code on Colab.
In the .\src\ folder of /original-code, /gpu_track, /model-optimize you will find the main components of the implementation.
As usual, the core is the main.cpp file that parses the selected dataset and creates and object of type GCN.
During the initialization phase all the layers for the model are constructed.
Consequently, the model is then run by calling the function GCN::run().
This function, based on the parameters set by GCN::get_default(), will execute a predefined number of epoch during the training phase.
In addition, in /gpu_track and /model-optimize we have kernels.cu where we have the implementation of CUDA kernels functions called by the host.