US 11,966,583 B2
Data pre-processing method and device, and related computer device and storage medium
Shaoli Liu, Beijing (CN); and Xiaofu Meng, Beijing (CN)
Assigned to CAMBRICON TECHNOLOGIES CORPORATION LIMITED, Beijing (CN)
Appl. No. 16/622,503
Filed by CAMBRICON TECHNOLOGIES CORPORATION LIMITED, Beijing (CN)
PCT Filed Jun. 27, 2019, PCT No. PCT/CN2019/093144
§ 371(c)(1), (2) Date Dec. 13, 2019,
PCT Pub. No. WO2020/042739, PCT Pub. Date Mar. 5, 2020.
Claims priority of application No. 201810987293.5 (CN), filed on Aug. 28, 2018; and application No. 201810987343.X (CN), filed on Aug. 28, 2018.
Prior Publication US 2021/0334007 A1, Oct. 28, 2021
This patent is subject to a terminal disclaimer.
Int. Cl. G06F 12/00 (2006.01); G06F 3/06 (2006.01); G06F 13/16 (2006.01); G06N 3/02 (2006.01)
CPC G06F 3/0611 (2013.01) [G06F 3/0619 (2013.01); G06F 3/0631 (2013.01); G06F 3/0683 (2013.01); G06F 13/1668 (2013.01); G06N 3/02 (2013.01); G06F 2213/16 (2013.01)] 18 Claims
OG exemplary drawing
 
11. A data pre-processing device, comprising:
a storage capacity obtaining circuit configured to obtain an available storage capacity of a first memory, and a target operation;
an input determination circuit configured to determine target input data corresponding to the target operation according to the target operation and the available storage capacity of the first memory;
an output determination circuit configured to determine target output data corresponding to the target operation according to the target operation and the target input data; and
a storage allocation module configured to store the intermediate computation result of the current operation into the first memory when an intermediate computation result output by a current operation in the target operation is required as the input data of another operation in the target operation;
wherein the target operation includes one or more operations, each operation represents an operation layer in a neural network, the target operation is obtained by fusing a plurality of operation layers having data dependence relationship in a depth direction of the neural network; and
wherein a number of operation layers in the plurality of operation layers to be fused in the target operation is determined based on the available storage capacity of the first memory and a data volume of the target input data and on a requirement that an intermediate computation result output by any operation in one or more operations fused in the target operation can be stored in the first memory.