METHOD FOR DATA COMPUTATION IN NEURAL NETWORK MODEL, IMAGE PROCESSING METHOD, AND DEVICE

摘要

A method for data computation in a neural network model, an image processing method, and a device. The method comprises: reading weight data shared by a set of data computation operations in a data processing layer in a neural network model into a group-shared variable of a thread group of a graphics processor (GPU); dividing input data of the data processing layer on the basis of the number of threads in the thread group; reading, for each set of the divided input data and from the group-shared variable, corresponding weight data of one data processing operation of a set of data processing operations to be performed on the set of input data; and each thread in the thread group performing the data processing operation by using a set of read input data and the corresponding weight data for said set of input data to obtain a computation result corresponding to said set of input data. For input data corresponding to the same thread group, corresponding weight data does not need to be read for each piece of the input data, thereby reducing the number of weight data read operations.