Depth separable convolutional neural network acceleration method and accelerator

• Main Authors: ZHANG XING, WANG XIN'AN, GUO PENGFEI, YONG SHANSHAN, LIU HUANSHUANG, LI QIUPING, GAO JINXIAO, LI XIAOFEI
• Format: Patent
• Language: Chinese; English
• Published: 24.08.2021
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention provides a depth separable convolutional neural network acceleration method, which comprises the following steps that: deep convolution is carried out on input neurons, and when deep convolution calculation is carried out, same M rows of C input channels are independently and parallelly calculated in a three-dimensional processing unit PE array to obtain same N rows of output neurons of the C channels, and N is less than M; and point convolution is performed on the output neurons obtained by the deep convolution, and during point convolution calculation, independent parallel calculation is performed on each row of data of the C channel. Through reasonable distribution of deep convolution and point convolution hardware resources, efficient support can be provided for a lightweight neural network model adopting deep separable convolution; data multiplexing is fully explored to reduce access to an external memory, and a basic calculation unit capable of jumping zero is adopted, so that power consum