Convolutional calculation accelerator, convolutional calculation method and convolutional calculation equipment

Abstract

The invention relates to a convolution calculation accelerator, a convolution calculation method and convolution calculation equipment, and relates to the technical field of electronic circuits. The convolution calculation accelerator comprises a controller, a calculation matrix and a first cache, the calculation matrix comprises at least one row of calculation units, and each row of calculation units comprises at least two calculation units; the controller is used for controlling input data loaded to the first cache to be input into the computing units of the corresponding rows, and the computing units of the corresponding rows transmit the input data in the computing units of the corresponding rows; each computing unit in the computing units of the corresponding row performs convolutioncomputation on the received input data and a pre-stored convolution kernel; at least two computing units in the same row multiplex the same input data, and only one input channel is needed, so that the cache capacity and input bandwidth requirements of the computing matrix are reduced, and the expandability of the computing matrix is improved.

Description

technical field [0001] The present application relates to the technical field of electronic circuits, in particular to a convolution computing accelerator, a convolution computing method, and a convolution computing device. Background technique [0002] In the field of data processing, many data processing processes, such as image processing through a convolutional neural network (CNN), require a large number of convolution calculation steps. [0003] In related technologies, computer equipment can accelerate the calculation speed of convolution calculation by means of hardware acceleration. Specifically, for example, the computer equipment can be connected to a dedicated convolution calculation accelerator, which includes a calculation matrix, control The calculation matrix contains several calculation units, wherein the input data corresponding to each calculation unit is independently cached, and when convolution calculation is performed, it is directly controlled by the ...

AI Technical Summary

Problems solved by technology

[0004] In related technologies, since the input data of each computing unit in the computing matrix is ​​independently cached and directly controlled by the controller to be sent to the corresponding computing unit, it is necessary to design a higher buffer capacity and input bandwidth for the convolution computing accelerator , affecting the scalability of the calculation matrix

Images