SmartMem: Layout Transformation Elimination and Adaptation for Efficient DNN Execution on Mobile
Open Access
- 27 April 2024
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 916-931
Abstract
This work is motivated by recent developments in Deep Neural Networks, particularly the Transformer architectures underlying applications such as ChatGPT, and the need for performing inference on mobile devices. Focusing on emerging transformers (specifically the ones with computationally efficient Swin-like architectures) and large models (e.g., Stable Diffusion and LLMs) based on transformers, we observe that layout transformations between the computational operators cause a significant slowdown in these applications. This paper presents SmartMem, a comprehensive framework for eliminating most layout transformations, with the idea that multiple operators can use the same tensor layout through careful choice of layout and implementation of operations. Our approach is based on classifying the operators into four groups, and considering combinations of producer-consumer edges between the operators. We develop a set of methods for searching such layouts. Another component of our work is developing efficient memory layouts for 2.5 dimensional memory commonly seen in mobile devices. Our experimental results show that SmartMem outperforms 5 state-of-the-art DNN execution frameworks on mobile devices across 18 varied neural networks, including CNNs, Transformers with both local and global attention, as well as LLMs. In particular, compared to DNNFusion, SmartMem achieves an average speedup of 2.8×, and outperforms TVM and MNN with speedups of 6.9× and 7.9×, respectively, on average.Keywords
All Related Versions
Funding Information
- DOE U.S. Department of Energy (DE-SC0021293)
- NSF (National Science Foundation) (CNS-2312207, CCF-2047516 (CAREER), CCF-2146873, CNS-2230944, III-2008557, CNS-2341378, CCF-2333895, OAC-2333899, CCF-2334273)
This publication has 44 references indexed in Scilit:
- High-Resolution Image Synthesis with Latent Diffusion Models2022
- A ConvNet for the 2020s2022
- ViViT: A Video Vision Transformer2021
- Co-Scale Conv-Attentional Image Transformers2021
- Designing Network Design Spaces2020
- Aggregated Residual Transformations for Deep Neural Networks2017
- Perceptual Losses for Real-Time Style Transfer and Super-Resolution2016
- Deep Residual Learning for Image Recognition2016
- ImageNet: A large-scale hierarchical image database2009
- Integrating Loop and Data Transformations for Global OptimizationJournal of Parallel and Distributed Computing, 2002