Detecting Data Races in OpenMP with Deep Learning and Large Language Models

Transformer-based neural network models are increasingly employed to handle software engineering issues, such as bug localization and program repair. These models, equipped with a self-attention mechanism, excel at understanding source code context and semantics. Recently, large language models (LLMs) have emerged as a promising alternative for analyzing and understanding code structure. In this paper, we propose two novel methods for detecting data race bugs in OpenMP programs. The first method is based on a transformer encoder trained from scratch. The second method leverages LLMs, specifically extending GPT-4 Turbo through the use of prompt engineering and fine-tuning techniques. For training and testing our approach, we utilized two datasets comprising different OpenMP directives. Our experiments show that the transformer encoder achieves competitive accuracy compared to LLMs, whether through fine-tuning or prompt engineering techniques. This performance may be attributed to the complexity of many OpenMP directives and the limited availability of labeled datasets.