Method, computing device, and computer-readable storage medium for convolution computation

Abstract

The present disclosure relates to a method for convolution calculation, a computing device and a computer-readable storage medium. The method includes: if it is determined that the current period is the initial period, loading data of n*n pixels from the input buffer to the systolic array, the systolic array includes a processing unit with a first predetermined number of rows multiplied by a second predetermined number of columns, the initial period and the initial weight Associated, n is a natural number; if it is determined that the current cycle is another cycle after the initial cycle, based on the sequence of the current cycle, load the data of a column of pixels, the data of a row of pixels or the data of one pixel to the systolic array; and if it is determined that the current cycle is completed For the corresponding data loading, a group of n*n pixel data is selected from one or more groups of n*n pixel data in the systolic array via the output multiplexing unit, so as to be provided to the general matrix multiplication unit. According to the embodiments of the present disclosure, bandwidth occupation and power consumption can be effectively reduced.

Description

technical field [0001] Embodiments of the present disclosure generally relate to the field of machine learning, and more specifically relate to a method for convolution calculation, a computing device, and a computer-readable storage medium. Background technique [0002] A typical operation in machine learning is convolution calculation. Convolution calculation is to multiply and add the weight in the convolution kernel and the corresponding input pixel block to generate the value of the corresponding output feature. Block convolution requires the img2col (image-to-column) function to convert pixel blocks into matrices. Traditional convolution calculation methods, such as block convolution calculations based on the img2col function, require each cycle to load complete pixels of each weight. For example, 3x3 block convolution (9 weights, from weight 0 to weight 8, corresponding to the initial period to the eighth period , namely cycle 0 to cycle 8), each cycle needs to load ...

AI Technical Summary

Problems solved by technology

Since the same data exists for the pixel data that needs to be loaded for adjacent cycles, reloading the complete 2x2 pixel data will result in wasted bandwidth and power

[0003] To sum up, the traditional convolution calculation scheme needs to reload the complete pixel corresponding to each weight every cycle, thus resulting in more bandwidth and higher power consumption

Images