Deep learning operator optimization method and device, equipment and storage medium

Abstract

The invention relates to a deep learning operator optimization method and device, equipment and a storage medium. The method comprises the steps: in response to target data detected in a first-layer high-speed memory L1 cache of a graphics processor, calling a method for reading an Image object to read the target data into the processor from the first-layer high-speed memory; performing secondary quantization operation on the target data in the processor to obtain an operation result; and writing the operation result into a main memory of the graphics processor. According to the invention, the speed of reading the target data can be improved by using an optimization acceleration mechanism of an L1cache of the graphics processor, so that the speed of subsequently quantifying a neural network based on the read target data can be improved, the interaction efficiency of a deep learning framework and a memory can be improved, the execution speed of the deep learning framework can be improved, resource occupation of a terminal system is reduced, and power consumption of the terminal system is reduced.

Description

technical field [0001] The present disclosure relates to the technical field of artificial neural networks, and in particular to a deep learning operator optimization method, device, equipment, and storage medium. Background technique [0002] With the rapid development of artificial intelligence technology, deep learning neural networks have achieved great success in pattern recognition such as image classification, object detection, image segmentation, speech recognition, and machine translation. However, a neural network with better performance usually has larger model parameters, which makes its calculation complexity higher, mainly reflected in two aspects: space storage and time calculation. Therefore, quantization of neural networks becomes particularly important, especially for applications running on devices such as terminal devices or embedded chips. [0003] In related technologies, the neural network on the terminal device can be accelerated by using a fixed-poi...

AI Technical Summary

Problems solved by technology

However, when the existing Buffer method is used to perform fixed-point quantization operations, data is loaded from the second-level high-speed memory (L2cache) of the processor, resulting in a slow quantization speed and poor performance of the deep learning framework for reasoning neural network models. In turn, it will increase the system resource usage of the terminal device and increase the power consumption

Images