一句话：提出输入自适应 GNN 编译器 Voyager，动态生成评估代数变换并运行时选择，GPU 上平均加速 3.5 倍。
图神经网络应用广泛且结构日益复杂，如何实现高性能执行成为关键问题。传统 DNN 中有效的代数变换（算子重排与融合）虽能提升计算与访存效率，但直接用于 GNN 效果不佳：输入图结构多样导致收益难以预测，图结构带来的复杂依赖也限制了现有优化技术。为此作者提出 Voyager，一个面向 GNN 的输入自适应编译器。它不依赖静态优化规则，而是动态生成并评估大量代数变换候选，包括针对图特征设计的新颖重排与融合策略；借助符号图属性高效估计张量形状并剪除低效候选，再由运行时选择模块依据实时输入特征挑选最优变换。在代表性 GNN 模型与数据集上，Voyager 在 GPU 上相较现有最优工作平均取得 3.5 倍加速。
Graph neural networks (GNNs) are gaining popularity in diverse application domains and growing in complexity. As a result, it is crucial to achieve high-performance GNN execution. Among various techniques, algebraic transformations, including operator reordering and operator fusion, have been successfully applied to improve the computation and memory access efficiencies of DNN models. However, directly applying these methods to GNNs results in suboptimal performance. This is primarily due to the diversity of input graph structures, making it challenging to predict performance benefits. Additionally, the complex dependencies introduced by these input graph structures limit the effectiveness of existing optimization techniques. To address these challenges, we propose Voyager, an input-adaptive compiler designed specifically for GNNs. Unlike existing methods that rely on static optimization rules, Voyager dynamically generates and evaluates a wide range of algebraic transformation candidates, including novel reordering and fusion strategies tailored to graph characteristics. By leveraging symbolic graph attributes, Voyager efficiently estimates tensor shapes and prunes low-performance candidates. A runtime selection module further enhances performance by choosing the optimal transformation based on real-time input features. Our evaluation shows that Voyager achieves an average 3.5 × speedup on GPUs over the state-of-the-art works on representative GNNs and datasets.
@inproceedings{0001SLL0C0XA00Z25,
title = {Voyager: Input-Adaptive Algebraic Transformations for High-Performance Graph Neural Networks},
author = {Yangjie Zhou and Wenting Shen and Jingwen Leng and Shuwen Lu and Zihan Liu and Weihao Cui and Zhendong Zhang and Wencong Xiao and Baole Ai and Yong Li and Wei Lin and Deze Zeng and Yun Liang and Quan Chen and Ning Liu and Minyi Guo},
booktitle = {ASPLOS},
year = {2025},
doi = {10.1145/3676642.3736121}
}