Method and system for calculating neural network convolution operation, and hardware device

Abstract

The invention discloses a method and system for calculating neural network convolution operation and a hardware device, belongs to the technical field of convolution operation methods, and solves the technical problems that the method in the prior art occupies a large amount of resources for processing data read-write bandwidth, cannot realize variable precision multiplication operation, and is low in hardware utilization rate, so that the operation efficiency is low. The method comprises the steps that: a processor is configured to execute a preset calculation precision; at least one pair of multiplier and multiplicand is obtained, digits are encoded into a plurality of fixed-point basic precision data, and fixed-point multiplication is executed according to a program instruction; and the processor is configured to divide the multiplier and the multiplicand into high-order data and low-order data represented by low-precision digits during first calculation precision multiplication, and calculation is performed according to a preset variable-precision multiplication instruction. The invention is used for perfecting the function of multi-precision convolution operation and improving the operation rate of the processor.

Description

technical field [0001] The invention belongs to the technical field of convolution operation methods, and in particular relates to a method, system and hardware device for computing neural network convolution operations. Background technique [0002] Due to the limitations of storage resources, computing power resources, real-time performance, power consumption and other factors on the device side of the mobile network or the Internet of Things, it is necessary to dynamically allocate resources and computing power according to tasks to achieve a balance between performance and power consumption resources. Using low-precision convolutional neural network to obtain higher performance by sacrificing a small amount of recognition accuracy, the data bit width used is relatively low, which can be realized by simple logic, saving computing and storage resources. Under the condition that a certain output precision is satisfied, the precision of the data used in the calculation of th...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to provide a method for computing neural network convolution operations, which solves the technical problems that the prior art method takes up a large amount of resources for processing data read and write bandwidth, fails to realize variable precision multiplication operations, and wastes hardware resources.

Images