Reformer: A Deep Learning Model for Runtime Selection of Convolution Kernels

Muhammad Saud Ul Hassan, Devanshu Mantri, Jehandad Khan, Christopher Erb, Alexandra Bara, Reid Kawaja, Jun Liu

Keywords: Reformer, Kernel Selection, Kernel Optimization

TL;DR: This paper proposes a novel transformer-inspired AI model to optimize GPU kernel selection and lower the training and inference time of deep learning models

Abstract: As neural networks grow larger, optimizing GPU kernel selection becomes increasingly essential to minimizing the time, cost, and energy demands of model training and inference. Current methods rely on hand-written rules-based heuristics, which often yield suboptimal performance, are labor-intensive to develop, and are difficult to adapt across hardware architectures and firmware releases. In this paper, we frame kernel selection as a sequence classification problem solved on the CPU, thereby leaving GPU resources free for user training and inference tasks. Traditional transformers are less effective in this context because CPU deployment limits the advantages of parallelism in attention mechanisms. In this regard, we propose the $\Gamma$-block, which performs only three matmul operations compared to the six required by a transformer block, while maintaining the same depth in terms of learnable layers. Our experiments on the IMDB and Reuters datasets demonstrate that a small model based on the $\Gamma$-block delivers comparable sequence classification accuracy to a similar model based on transformer blocks, while also providing faster inference times on the CPU. By stacking multiple $\Gamma$-blocks, we develop a lightweight model for kernel selection, named Reformer. To train the model, we propose a novel approach that assigns optimality probabilities to kernels based on their runtimes, offering a more robust alternative to one-hot probabilities. We demonstrate the effectiveness of Reformer by integrating it into MIOpen for convolution kernel selection, achieving an average speed-up of approximately 3x in convolution operations on the AMD Instinct$\texttrademark$ MI100 GPU.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13039