Scheduling method and device based on deep learning node calculation and storage medium

Abstract

The embodiment of the invention discloses a scheduling method and a device based on deep learning node calculation and a storage medium. The scheduling method comprises the following steps: obtainingto-be-calculated nodes of a preset neural network calculation graph; judging the node type of the node to be calculated, wherein the node type comprises a hardware calculation node and a software calculation node; scheduling the hardware computing node to a first queue under the condition that the node type is the hardware computing node, and judging whether a hardware computing power module corresponding to the hardware computing node is occupied or not; and inputting the hardware computing node into the hardware computing power module for computing under the condition that the hardware computing power module is not occupied. According to the embodiment of the invention, asynchronization of a graph reasoning process is realized, and resources of software and hardware are fully utilized.

Description

technical field [0001] Embodiments of the present invention relate to deep learning technologies, such as a scheduling method, device, and storage medium based on deep learning node calculations. Background technique [0002] Deep learning is a branch of machine learning. It is an algorithm that uses artificial neural networks as the framework to perform representation learning on data. Deep learning networks are usually trained by algorithms. In most cases, algorithm developers tend to use existing public deep learning frameworks for model training, and most public deep learning frameworks are for central processing unit (Central Processing Unit, CPU) / image processor (Graphics Processing Unit, Designed for computing devices such as GPUs. CPU / GPU adopts traditional instruction set architecture, which has low architecture efficiency and high flexibility. [0003] However, with the development of deep learning related technologies, the requirements for computing power in ...

AI Technical Summary

Problems solved by technology

However, there is a big difference between the data expression of the data flow architecture and the instruction set architecture: the operator granularity of the data flow architecture is much larger than that of the instruction set architecture; the operator of the data flow architecture needs to determine the arrangement order of the calculation modules according to the data dependence before calculation

This difference leads to the fact that the data flow chip is only suitable for deep learning operators, and some places with a high degree of customization still need general computing equipment for auxiliary processing.

The existence of software nodes leads to the following situations: software nodes have no hardware acceleration effect and run slowly; when running graph calculations in a single thread, the operation of software nodes will cause hardware to be idle, which is not conducive to acceleration

Images