DeepCodeGraph: A Language Model for Compile-Time Resource Optimization Using Masked Graph Autoencoders

Analysing source code using deep learning aids compile-time decisions affecting performance in embedded devices. We propose DeepCodeGraph, a general graph-based language model, which learns patterns to identify better compilation strategies, optimal hardware configurations and software transformations. DCG includes i) A large-scale dataset containing over 100k graphs. ii) A graph neural network to implement a graph-based language model. iii) A self-supervised pre-training framework leveraging Masked Graph Autoencoders. The performance of DCG is evaluated on two downstream tasks: heterogeneous device mapping and thread block size prediction. DCG outperforms previous graph-based state-of-the-art improving previous results by 3%.