A Dynamic Mapping Model for General CNN Accelerator Based on FPGA - Network and Parallel Computing Access content directly
Conference Papers Year : 2021

A Dynamic Mapping Model for General CNN Accelerator Based on FPGA

Abstract

As the application scenarios of convolutional neural network (CNN) become more and more complex, the general CNN accelerator based on matrix multiplication has become a new research focus. The existing mapping methods for converting convolution calculation into matrix multiplication need to be improved. This paper proposes a new dynamic mapping model to improve the flexibility and versatility of matrix multiplication. The dynamic mapping model implements two algorithms: dynamic residue processing mapping algorithm (DRPMA) and dilated convolution mapping algorithm (DCMA). The former can dynamically adjust the mapping method according to the number of output channels of the convolution layer, improve the utilization of the multiply-accumulate (MAC) array. The latter extends the efficient support for Dilated CNNs. For demonstration, we implement an accelerator with Verilog on Xilinx VC709 FPGA board and test some typical CNN models. Experimental results show that the general accelerator achieves high performance and energy efficiency.
Fichier principal
Vignette du fichier
511910_1_En_2_Chapter.pdf (537.7 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03768741 , version 1 (04-09-2022)

Licence

Attribution

Identifiers

Cite

Xiaoqiang Zhao, Jingfei Jiang, Zhe Han, Jinwei Xu, Zhiqiang Liu. A Dynamic Mapping Model for General CNN Accelerator Based on FPGA. 17th IFIP International Conference on Network and Parallel Computing (NPC), Sep 2020, Zhengzhou, China. pp.17-29, ⟨10.1007/978-3-030-79478-1_2⟩. ⟨hal-03768741⟩
21 View
4 Download

Altmetric

Share

Gmail Facebook X LinkedIn More