Research on optimization of convolution back propagation on domestic heterogeneous platforms

Zhan Yang; Fushuai Li; Yongqing Chen; Jingde Bu; Rong Dai

doi:10.1117/12.3005093

9 October 2023 Research on optimization of convolution back propagation on domestic heterogeneous platforms

Zhan Yang, Fushuai Li, Yongqing Chen, Jingde Bu, Rong Dai

Proceedings Volume 12791, Third International Conference on Advanced Algorithms and Neural Networks (AANN 2023); 127910X (2023) https://doi.org/10.1117/12.3005093
Event: Third International Conference on Advanced Algorithms and Neural Networks (AANN 2023), 2023, Qingdao, SD, China

Abstract

Convolution neural network is widely used in various fields. The convolution layer is the core layer of the convolution neural network. The back propagation speed of the convolution layer will directly affect the training speed of the whole network, thus affecting the whole performance. For the convolution layer with stride ≥ 2, the error transmission phase of back propagation will carry out a large amount of padding in the feature graph, resulting in a large amount of additional overhead in access and calculation. In this case, we propose a new optimization method, which can reduce the overhead caused by padding to almost zero, and implement it by implicit convolution on domestic heterogeneous platforms. The experiment shows that the performance of the operator optimized by this method is nearly 50% higher than that of the original operator of the platform, and the average performance reaches 90% of that of NVIDIA V100 operator.

(2023) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Zhan Yang, Fushuai Li, Yongqing Chen, Jingde Bu, and Rong Dai "Research on optimization of convolution back propagation on domestic heterogeneous platforms", Proc. SPIE 12791, Third International Conference on Advanced Algorithms and Neural Networks (AANN 2023), 127910X (9 October 2023); https://doi.org/10.1117/12.3005093

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available