Deconvolution calculation method, hardware accelerator, equipment and readable storage medium

Abstract

The invention discloses a deconvolution calculation method, a hardware accelerator, equipment and a readable storage medium. The method comprises the following steps: acquiring a plurality of input blocks; transforming data in the deconvolution kernel according to the first preposed matrix to obtain a deconvolution kernel matrix; transforming data in the plurality of input blocks according to the second preposed matrix to obtain a plurality of input matrixes; then multiplying the deconvolution kernel matrix by a plurality of input matrixes to obtain a plurality of intermediate matrixes; accumulating data of all layers of data matrixes in any intermediate matrix according to channels to obtain a plurality of accumulation matrixes; according to the post-matrix, respectively transforming the data in the plurality of accumulation matrixes to obtain a plurality of output blocks; and sequentially arranging the plurality of output blocks into an output feature map, and obtaining a deconvolution calculation result. In the calculation process, a large number of zeros are not inserted into the original input feature map, so that the calculation efficiency is effectively improved.

Description

Technical field [0001] The present application relates to the depth of neural network technology, particularly to a method of deconvolution calculation, a hardware accelerator, and device-readable storage medium. Background technique [0002] FIG using hardware accelerators perform a deconvolution calculation, usually inserted in a large number of zero data of the original input feature map, the extension to give a new input to the input feature map feature, and then use the same size as a window size of new nuclear deconvolution FIG input pan sliding characteristics according to a preset step length, each time the slide, and the data in the window and then multiplying covered summed data deconvolution kernel, the output summation result, the window slides over the other new after entering all the data on the characteristic diagram, the summation of all the results obtained compared with the final result of the deconvolution calculation. [0003] See calculation figure 1 , The in...

AI Technical Summary

Problems solved by technology

[0003] Calculation process see figure 1 , a large number of zeros have been inserted into the data of the original input feature map, the blank cells represent the inserted zeros, the black cells represent the data of the original input feature map, and the gray cells represent the window. It can be seen that the window currently only contains two The data in the original input feature map, and the rest are all zero. When multiplying the data in the window with the data in the deconvolution kernel, these zeros will cause many multiplication operations to be invalid operations. In the entire deconvolution calculation In the process, there will be a large number of invalid operations, which will seriously affect the computing efficiency of the hardware accelerator.

Images