Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
TENSORCORE-BASED INT4 DATA TYPE PROCESSING METHOD AND SYSTEM, DEVICE, AND MEDIUM
Document Type and Number:
WIPO Patent Application WO/2022/057459
Kind Code:
A1
Abstract:
Disclosed in the present invention are a Tensorcore-based int4 data type processing method and system, a device, and a storage medium. The method comprises: in response to received data of which the data type is int4, determining, according to the input data dimension, weight dimension, and offset dimension of the data, whether the batch processing size, input dimension number, and output dimension number of the data satisfy requirements; in response to the fact that the batch processing size, input dimension number, and output dimension number of the data satisfy the requirements, writing the input data of the data into a first shared memory from a global memory, and writing the weight data of the data into a second shared memory from the global memory; storing, into a third shared memory, a first calculation result that is obtained on the basis of the first shared memory and the second shared memory, so as to be added with offset data to obtain a second calculation result; and returning the second calculation result to the global memory. The present invention achieves the support of a TVM full connection layer for the int4 data type, and compared with int8, greatly improves the performance.

Inventors:
SONG XIAOMEI (CN)
Application Number:
PCT/CN2021/109214
Publication Date:
March 24, 2022
Filing Date:
July 29, 2021
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
SUZHOU INSPUR INTELLIGENT TECH CO LTD (CN)
International Classes:
G06N3/063; G06F9/54
Foreign References:
CN112232496A2021-01-15
CN111860838A2020-10-30
CN111859270A2020-10-30
CN111539526A2020-08-14
CN111124656A2020-05-08
Attorney, Agent or Firm:
LIAN & LIEN IP ATTORNEYS (CN)
Download PDF: