一句话：提出面向昇腾架构的组件化 roofline 模型以定位算子瓶颈，并总结可复用的优化策略，端到端提速 1.07–2.15 倍。
随着深度学习的发展，昇腾等面向 AI 负载的领域专用架构（DSA）被广泛采用，但相较于 GPU，昇腾要求用户手动管理算子执行，性能问题复杂且需要精确分析；现有 roofline 模型则存在可视化困难、性能评估不准确的问题。本文提出一种组件化 roofline 模型，将算子抽象为若干可独立度量的组件，从而刻画算子性能并有效识别瓶颈组件。作者通过实际算子优化案例，展示了基于 roofline 分析的完整优化流程，归纳出常见性能问题与对应优化策略。最后，大量端到端优化实验表明模型推理速度可获得 1.07× 至 2.15× 的提升，并给出了来自实践的若干经验洞见。
With the rise of deep learning, many companies have developed domain-specific architectures (DSAs) optimized for AI workloads, with Ascend being a representative. To fully realize the operator performance on Ascend, effective analysis and optimization is urgently needed. Compared to GPU, Ascend requires users to manage operations manually, leading to complex performance issues that require precise analysis. However, existing roofline models face challenges of visualization complexity and inaccurate performance assessment. To address these needs, we introduce a component-based roofline model that abstracts components to capture operator performance, thereby effectively identifying bottleneck components. Furthermore, through practical operator optimization case studies, we illustrate a comprehensive process of optimization based on roofline analysis, summarizing common performance issues and optimization strategies. Finally, extensive end-to-end optimization experiments demonstrate significant model speed improvements, ranging from 1.07× to 2.15×, along with valuable insights from practice.
@inproceedings{ZhouWLLLWWWZHLL25,
title = {Squeezing Operator Performance Potential for the Ascend Architecture},
author = {Yuhang Zhou and Zhibin Wang and Guyue Liu and Shipeng Li and Xi Lin and Zibo Wang and Yongzhong Wang and Fuchun Wei and Jingyi Zhang and Zhiheng Hu and Yanlin Liu and Chunsheng Li and Ziyang Zhang and Yaoyuan Wang and Bin Zhou and Wanchun Dou and Guihai Chen and Chen Tian},
booktitle = {ASPLOS},
year = {2025},
doi = {10.1145/3676641.3716243}
}