[2603.23566] AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization

Computer Science > Machine Learning arXiv:2603.23566 (cs) [Submitted on 24 Mar 2026] Title:AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization Authors:Jiehao Wu, Zixiao Huang, Wenhao Li, Chuyun Shen, Junjie Sheng, Xiangfeng Wang View a PDF of the paper titled AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization, by Jiehao Wu and 5 other authors View PDF HTML (experimental) Abstract:AscendC (Ascend C) operator optimization on Huawei Ascend neural processing units (NPUs) faces a two-fold knowledge bottleneck: unlike the CUDA ecosystem, there are few public reference implementations to learn from, and performance hinges on a coupled two-part artifact - a host-side tiling program that orchestrates data movement and a kernel program that schedules and pipelines instructions. We present AscendOptimizer, an episodic agent that bootstraps this missing expertise by turning execution into experience. On the host side, AscendOptimizer performs profiling-in-the-loop evolutionary search to discover valid and high-performing tiling and data-movement configurations directly from hardware feedback. On the kernel side, it mines transferable optimization motifs by rewinding optimized kernels - systematically de-optimizing them to synthesize instructive "bad-to-good" trajectories - and distills these motifs into a retrievable experience bank for guided rewriting. By alternating host tuning and kernel rewriting in a closed loop, AscendOptimizer steadily expand...