南方科技大学知识苑(SUSTech KC): 基于权重预编码的稀疏卷积神经网络加速器和计算设备

Title	基于权重预编码的稀疏卷积神经网络加速器和计算设备
Alternative Title	Sparse convolutional neural network accelerator based on weight precoding and computing device
Author	余浩; 毛伟; 程全; 李博宇; 代柳瑶
First Inventor	余浩
Original applicant	南方科技大学
First applicant	南方科技大学
Address of First applicant	518055 广东省深圳市南山区桃源街道学苑大道1088号
Current applicant	南方科技大学
Address of Current applicant	518055 广东省深圳市南山区桃源街道学苑大道1088号 (广东,深圳,南山区)
First Current Applicant	南方科技大学
Address of First Current Applicant	518055 广东省深圳市南山区桃源街道学苑大道1088号 (广东,深圳,南山区)
Application Number	CN202210963479.3
Application Date	2022-08-11
Open (Notice) Number	CN115456152A
Date Available	2022-12-09
Status of Patent	实质审查
Legal Date	2022-12-27
Subtype	发明申请
SUSTech Authorship	First
Abstract	本发明公开了基于权重预编码的稀疏卷积神经网络加速器和计算设备，稀疏卷积神经网络加速器包括若干引擎和预编码模块，每一引擎包括若干块；预编码模块用于对每一引擎的权重数据进行编码，得到该引擎中各块的编码权重数据；每一块包括：激活选择器用于获取掩码数据和若干激活数据，根据掩码数据从各激活数据中确定若干目标激活数据；乘法模块用于根据编码权重数据和各目标激活数据，确定部分积数据；加法树用于对各部分积数据进行累加，得到乘加运算数据。本发明通过掩码数据对激活数据进行筛选，实现了各种稀疏度场景下的卷积神经网络加速。解决了可变稀疏密度约束块受权重稀疏度限制，导致低权重稀疏度场景下难以实现卷积神经网络加速的问题。
Other Abstract	The invention discloses a sparse convolutional neural network accelerator based on weight precoding and computing equipment, the sparse convolutional neural network accelerator comprises a plurality of engines and a precoding module, and each engine comprises a plurality of blocks; the pre-coding module is used for coding the weight data of each engine to obtain coding weight data of each block in the engine; each block comprises an activation selector used for obtaining mask data and a plurality of activation data, and determining a plurality of target activation data from the activation data according to the mask data; the multiplication module is used for determining partial product data according to the coding weight data and the target activation data; and the adder tree is used for accumulating the partial product data to obtain multiply-add operation data. According to the method, the activation data is screened through the mask data, so that the convolutional neural network acceleration in various sparseness scenes is realized. The problem that a variable sparse density constraint block is limited by weight sparseness, so that convolutional neural network acceleration is difficult to realize in a low-weight sparseness scene is solved.
CPC Classification Number	Y02D10/00
IPC Classification Number	G06N3/063 ; G06N3/04
INPADOC Legal Status	（ENTRY INTO FORCE OF REQUEST FOR SUBSTANTIVE EXAMINATION）[2022-12-27][CN]
INPADOC Patent Family Count	1
Extended Patent Family Count	1
Priority date	2022-08-11
Patent Agent	陈专 ; 吴志益
Agency	深圳市君胜知识产权代理事务所(普通合伙)
URL	[Source Record]
Data Source	PatSnap
Document Type	Patent
Identifier	http://kc.sustech.edu.cn/handle/2SGJ60CL/531841
Department	SUSTech Institute of Microelectronics 南方科技大学-香港科技大学深港微电子学院筹建办公室
Recommended Citation GB/T 7714	余浩,毛伟,程全,等. 基于权重预编码的稀疏卷积神经网络加速器和计算设备.

Files in This Item:
There are no files associated with this item.

If you have any objections to this item, please fill out the form below and the administrator will contact you as soon as possible.
Content:
Email：	*
Affiliation Sequential No:
Verification Code:	Refresh

Any comments and suggestions are welcomed.
Title:	*
Content:
Email：	*
Verification Code:	Refresh