Convolution calculation acceleration method and device

Abstract

The embodiment of the invention provides a convolution calculation acceleration method and device, and the method comprises the steps: obtaining a current depth time sequence i when the convolution type is depth-first convolution; inputting the column pixels in the convolution kernel corresponding to the current depth time sequence i and the row pixels in the feature map into a convolution calculation unit, wherein the convolution calculation unit is an array composed of multipliers and adders; if the imodK is equal to 0, translating the first row of pixels of the convolution kernel to the last row, wherein K is the number of rows and columns of the convolution kernel; and if the imodK is equal to 0 and i is greater than or equal to K<2>, taking the (i-k<2>)/k+1 modk th line feature in theconvolution calculation unit as the output feature of the (i-k<2>)/k+1 th line depth first convolution. According to the method and the device provided by the embodiment of the invention, the hardware utilization rate of the convolution calculation unit is effectively improved, the energy waste caused by the idleness of a large number of multipliers is avoided, and the depth-first convolution calculation efficiency is improved.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a convolution calculation acceleration method and device. Background technique [0002] With the development of artificial intelligence technology, Convolutional Neural Network (CNN) has been fully utilized. The current mainstream convolutional neural network model has a complex structure and a large amount of calculation data. Conventional CPUs (Central Processing Units, central processing units) and GPUs (Graphics Processing Units, graphics processors) are difficult to meet the power consumption restrictions and calculations of convolution calculations. Need for speed. [0003] A common solution is to introduce heterogeneous computing technology, which combines the general-purpose computing power of the CPU and the high performance of the convolutional computing accelerator to achieve efficient and low-power convolutional neural network computing. Existing convolution compu...

AI Technical Summary

Problems solved by technology

[0005] Embodiments of the present invention provide a convolution calculation acceleration method and device, which are used to solve the problem that only one line of multiplier-adder is in the operation state when the existing convolution calculation accelerator calculates depth-first convolution, which wastes energy and area and has low calculation efficiency.

Images