LOOPerSet
收藏资源简介:
LOOPerSet是一个大规模的公共数据集,旨在支持数据驱动代码优化的研究和开发。该数据集包含超过2800万个有标签的数据点,这些数据点来自大约22万个独特的、合成的多面体程序。每个数据点将程序和特定的优化序列映射到物理硬件上测量的真实性能提升。该数据集的核心是一个程序集合,每个数据点将程序和特定的优化序列映射到物理硬件上测量的真实性能提升。该数据集的核心是一个程序集合,每个数据点将程序和特定的优化序列映射到物理硬件上测量的真实性能提升。该数据集的设计旨在为编译器社区提供一个多功能工具,其最直接的应用是作为训练和评估机器学习模型的大规模基准。研究人员还可以直接分析数据以发现新的数据驱动编译器启发式算法。最后,它为解决硬件可移植性问题提供了基础,使模型能够在很大程度上减少数据收集的情况下预先训练,并为新的架构进行微调。
LOOPerSet is a large-scale public dataset designed to support research and development in data-driven code optimization. It contains over 28 million labeled data points sourced from approximately 220,000 unique synthetic polyhedral programs. Each data point maps a program and a specific optimization sequence to the real performance gains measured on physical hardware. At the core of this dataset lies a program collection, with every data point establishing the correspondence between a program, a targeted optimization sequence, and the actual measured performance gains on physical hardware. This dataset is crafted to serve as a versatile tool for the compiler community. Its most immediate application is as a large-scale benchmark for training and evaluating machine learning models. Researchers can also directly analyze the dataset to uncover new data-driven compiler heuristics. Finally, it provides a foundation for addressing hardware portability issues, allowing models to be pre-trained with substantially reduced data collection efforts and fine-tuned for novel architectures.

- 1LOOPerSet: A Large-Scale Dataset for Data-Driven Polyhedral Compiler Optimization纽约大学阿布扎比分校 · 2025年



