RedFuser: An Automatic Operator Fusion Framework for Cascaded Reductions on AI Accelerators Conference: Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2; March 21–26, 2026; Pittsburgh, PA, USA. Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’26), March 21–26, 2026, Pittsburgh, PA, USA ISBN: 979-8-4007-2359-9/2026/03 DOI: 10.1145/3779212.3790209 V2fp3794 CCS: Computer systems organization Parallel architectures CCS: Computing methodologies Parallel computing methodologies CCS: Software and its engineering Compilers CCS: Computing methodologies Artificial intelligence

Xinsheng Tang, Yangcheng Li, Nan Wang, Zhiyi Shu, Xingyu Ling, Junna Xing, Peng Zhou, Qiang Liu · 2026

Read RedFuser: An Automatic Operator Fusion Framework for Cascaded Reductions on AI Accelerators Conference: Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2; March 21–26, 2026; Pittsburgh, PA, USA. Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’26), March 21–26, 2026, Pittsburgh, PA, USA ISBN: 979-8-4007-2359-9/2026/03 DOI: 10.1145/3779212.3790209 V2fp3794 CCS: Computer systems organization Parallel architectures CCS: Computing methodologies Parallel computing methodologies CCS: Software and its engineering Compilers CCS: Computing methodologies Artificial intelligence by Xinsheng Tang, Yangcheng Li, Nan Wang, Zhiyi Shu, Xingyu Ling, Junna Xing, Peng Zhou, Qiang Liu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-10026.txt.

Description

Tang and coauthors present RedFuser, a tool that joins chained shrink steps on AI chips into one loop. Such steps, like safe softmax feeding a matrix product in attention, show up a lot, yet compilers handle them poorly. Hand built fixes lack reach and resist reuse. RedFuser gives a formal way to merge the loops into one with step by step sums, then finds matching steps in code and writes fast fused kernels for them.

Catalog record

CREATORXinsheng Tang, Yangcheng Li, Nan Wang, Zhiyi Shu, Xingyu Ling, Junna Xing, Peng Zhou, Qiang Liu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79063
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.10026

Browse the full library · About the Tower of Babel