Bibliographic Details
| Title: |
VMOC: MLIR-based memory optimization for CNN deployment on vector accelerators. |
| Authors: |
Chen, Xiaorong1 (AUTHOR) chenxiaorong23@nudt.edu.cn, Liu, Zhong1 (AUTHOR) zhongliu@nudt.edu.cn |
| Source: |
Journal of Systems Architecture. Oct2026, Vol. 179, pN.PAG-N.PAG. 1p. |
| Subject Terms: |
Computer memory management, Code generators |
| Abstract: |
Vector accelerators are increasingly prominent platforms for deploying Convolutional Neural Networks (CNNs) due to their high energy efficiency and parallelism. On such architectures, efficient memory utilization is critical for successful deployment. However, existing graph fission methods exhibit several limitations on vector accelerators, including poor hardware adaptability, inaccurate fission cost modeling, low compilation efficiency, and heavy reliance on manual intervention. This paper proposes VMOC, a novel MLIR-based memory optimization method. VMOC automatically identifies memory hotspot subgraphs, constructs a Hardware-aware Fission Cost Model (HFCM) to quantify the memory benefits and execution overhead introduced by fission, and employs an Adaptive Graph Fission Algorithm (AGFA) to dynamically search for fission schemes that satisfy both memory constraints and positive benefit constraints. Meanwhile, a Graph Encapsulation-based Scheduling Optimization (GESO) is introduced to alleviate the problem of increased scheduling complexity caused by graph fission. These optimization strategies are integrated into the MLIR framework, enabling end-to-end automatic compilation from ONNX models to target executable code and generating high-performance code for vector accelerators. Experimental results demonstrate the effectiveness of VMOC on GoogleNet, ResNet18, VGG16, and YOLOv5s. VMOC achieves memory compression ratios of 52.05% ∼ 96.18%, with compilation speedups of 2.34 × ∼ 152.11 × over MAGIS and 12.63 × ∼ 100.93 × over TVM. On FT-M7032 and FT-XT4000 vector accelerators, it achieves computational efficiencies of 62.30% ∼ 74.05% and 50.25% ∼ 71.01%, consistently outperforming VCNN. These results suggest that VMOC has the potential to become an efficient and practical approach for the automatic deployment of CNNs on vector accelerators. [Display omitted] • A memory-hotspot method identifies critical operators affecting peak memory. • The HFCM model evaluates the benefits of fissioning under GSM capacity constraints. • The AGFA dynamically finds the minimum fission granularity under GSM constraints. • The GESO encapsulates local subgraphs into supernodes to suppress graph expansion. • An end-to-end MLIR framework efficiently deploys CNN models on vector accelerators. [ABSTRACT FROM AUTHOR] |
|
Copyright of Journal of Systems Architecture is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) |
| Database: |
Business Source Index |