A Configurable Hardware Accelerator Based on Hybrid Dataflow for Depthwise Separable Convolution

Jiahua Ou, Xiaojin Li, Yabin Sun, Yanling Shi

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

4 Scopus citations

Abstract

As one of the most commonly used neural network models, the lightweight convolutional neural network has been widely adopted in the field of object detection, pattern recognition and so on. In order to achieve the real-time calculation in embedded system, the acceleration of depthwise separable convolution has become the focus of research in recent years. In this paper, a configurable hardware accelerator for depthwise separable convolution is proposed. The accelerator contains a high-speed dedicated processing engine array which can be configured to carry out depthwise convolution and pointwise convolution with high hardware utilization. The storage is fully utilized by applying hybrid scheduling strategies. Moreover, the design of MobileNet-V2 is validated on the platform of Zynq 7020, giving a high rate up to 186 frames/s.

Original languageEnglish
Title of host publicationCTISC 2022 - 2022 4th International Conference on Advances in Computer Technology, Information Science and Communications
EditorsVassilis C. Gerogianni, Yong Yue, Fairouz Kamareddine
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781665458726
DOIs
StatePublished - 2022
Event4th International Conference on Advances in Computer Technology, Information Science and Communications, CTISC 2022 - Suzhou, China
Duration: 22 Apr 202224 Apr 2022

Publication series

NameCTISC 2022 - 2022 4th International Conference on Advances in Computer Technology, Information Science and Communications

Conference

Conference4th International Conference on Advances in Computer Technology, Information Science and Communications, CTISC 2022
Country/TerritoryChina
CitySuzhou
Period22/04/2224/04/22

Keywords

  • configurable
  • depthwise separable convolution
  • hardware accelerator

Fingerprint

Dive into the research topics of 'A Configurable Hardware Accelerator Based on Hybrid Dataflow for Depthwise Separable Convolution'. Together they form a unique fingerprint.

Cite this