.png){kind=icon}
.png){kind=icon}
Return
Discovering software parallelization points using deep neural networks
.png){kind=icon}
OA
AI
DOI：10.1016/j.array.2026.100781.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This study investigates whether lightweight deep learning models can distinguish parallelizable loops using only token-based representations of source code under controlled experimental conditions. To address the scarcity of labeled data for loop parallelizability analysis, we employ a genetic algorithm to generate synthetic Python loops belonging to two simplified categories: independent loops without data dependencies and dependent loops whose dependency structure prevents parallel execution. The generated code is tokenized and preprocessed to construct a dataset of approximately 4000 samples. Two neural architectures-a Deep Neural Network (DNN) and a Convolutional Neural Network (CNN)-are evaluated on this task, with performance stability assessed across 30 independent training runs and compared against classical linear baselines under the same representational constraints. While the CNN achieves marginally higher mean accuracy, both models exhibit comparable variability. Additional experiments with reduced datasets highlight the influence of dataset diversity on classification performance, and preliminary validation on a manually curated real-code dataset suggests that the learned token-level patterns extend to structurally simple real-world loop instances. Rather than proposing a new parallelization technique, this work provides empirical evidence on the minimum representational and architectural complexity required to capture loop-level parallelizability signals from lexical code features, thereby clarifying the potential and limitations of token-based learning approaches in this domain.
Keywords:
Loop classification
Parallelization detection
Deep neural networks
Convolutional neural networks
Genetic algorithm
Code tokenization
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4.5
Papers:
753
Citations:
1.2K