Title:
PARALLEL COMPUTING METHOD AND APPARATUS
Document Type and Number:
WIPO Patent Application WO/2020/107616
Kind Code:
A1
Abstract:
A parallel computing method and apparatus, the method comprising: performing horizontal-direction splicing on N input matrices of a target convolutional neural network layer, and acquiring a first spliced input matrix (S101); according to the described computation, a vector processor identifying the performance of horizontal-direction computing processing on the first spliced input matrix to obtain a spliced output matrix (S102); and screening N output matrices out from the spliced output matrix, and using the N output matrices as N input matrices of a next convolutional neural network layer (S103). By means of performing horizontal-direction splicing on input matrices, the described method reduces additional overhead brought about by required loading and storage when a processor performs line feed processing on a pipeline, and improves the computing performance of a vector processor when processing small-size data.
Inventors:
SU LAN (CN)
GU PENG (CN)
GU PENG (CN)
Application Number:
PCT/CN2018/124831
Publication Date:
June 04, 2020
Filing Date:
December 28, 2018
Export Citation:
Assignee:
SHENZHEN INTELLIFUSION TECH CO LTD (CN)
International Classes:
G06N3/04
Foreign References:
| CN108334910A | 2018-07-27 | |||
| CN107368886A | 2017-11-21 | |||
| CN106156807A | 2016-11-23 | |||
| CN108572850A | 2018-09-25 | |||
| US20180182083A1 | 2018-06-28 |
Attorney, Agent or Firm:
SHENZHEN 51REG INTELLECTUAL PROPERTY CO., LTD. (CN)
Download PDF: