A computing device and method for data reuse

Abstract

The embodiment of the invention relates to a computing device and a method for data reuse, and relates to the field of computers. The computing device includes: a universal register; an arithmetic unit that is coupled with the universal register, the arithmetic unit comprising a data reuse unit, the data reuse unit being coupled with a plurality of dot product data units, and the data reuse unit being configured to read a data set used for multiple convolution operations from the universal register and temporarily store the data set and to determine a plurality of data subsets from the data set for respectively inputting the plurality of dot product data units, wherein two data subsets for inputting the adjacent dot product data units comprise partially same data; and a plurality of dot product data units, each dot product data unit in the plurality of dot product data units being configured to perform a dot product operation on the input data subset to generate a dot product operation result. Consequently, convolution operation efficiency can be improved.

Description

technical field [0001] Embodiments of the present disclosure generally relate to the field of computers, and in particular to computing devices and methods for data reuse. Background technique [0002] Convolutional neural network is a kind of feed-forward neural network with convolution calculation and deep structure, and it is one of the representative algorithms of deep learning. Typically, a convolutional neural network consists of several convolutional layers, pooling layers, and fully connected layers. Among them, the convolutional layer and the pooling layer occupy most of the computing, storage and bandwidth resources. Therefore, the execution efficiency of the convolutional layer and the pooling layer largely determines the efficiency of the entire convolutional neural network. [0003] refer to figure 1 , which shows a schematic block diagram of a stream processor 100 according to the prior art. Such as figure 1 As shown, the stream processor 100 includes a ge...

AI Technical Summary

Problems solved by technology

[0011] However, the existing technology has the following problems: 1) each cycle operation can only complete single-pixel mathematical operations, resulting in a large number of cycles required for a convolution and pooling operation, and low execution efficiency; 2) each cycle The operation includes multiple steps, each step requires multiple instructions, and the execution efficiency is low; 3) Between loop operations, data is not shared or reused, requiring a large number of memory read and write operations, and consumes a lot of power

Images