Abstract
A closed-loop GPU kernel optimization technique collects hardware performance counters to form a normalized counter-profile vector and classifies the kernel into a bottleneck class. Candidate structured transforms are selected using probabilistic ranking based on stored successes and failures for transforms conditioned on bottleneck class and hardware. After each transform, the kernel is re-profiled (in addition to benchmarking), re-classified, and a bottleneck shift is detected when the bottleneck class changes. Upon a shift, remaining transforms are re-ranked for the new bottleneck class. Optimization trajectories are recorded as sequences of transforms, bottleneck-shift pairs, and speedups, enabling multi-step lookahead scoring. Cross-kernel transfer is supported by retrieving prior outcomes from kernels with cosine-similar counter profiles within the same bottleneck class and estimating expected speedups using similarity-weighted aggregation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Closed-Loop Profile-Guided Kernel Optimization with Bottleneck-Shift Detection and Cross-Kernel Transfer Learning", Technical Disclosure Commons, (June 30, 2026)
https://www.tdcommons.org/dpubs_series/10765