Neural network calculation graph processing method, computer storage medium and electronic equipment

Abstract

The invention provides a neural network calculation graph processing method, a computer storage medium and electronic equipment. The neural network calculation graph processing method comprises the steps of obtaining a program code for creating a required neural network model based on a demand of a user for creating the neural network model; generating a calculation graph of the required neural network model according to the program code; carrying out format conversion on the calculation graph to obtain a target calculation graph, wherein the target calculation graph comprises a plurality of operators and connection relations among the operators; according to the characteristics of the operators, mapping the operators to operation equipment matched with the characteristics of the operatorsto obtain a mapping relation table of the operation equipment and the operators; and according to the mapping relation table, loading each operator into operation equipment matched with the operatorfor calculation. According to the mapping relation table, the operators are loaded into the arithmetic devices matched with the operators for calculation, the advantages of each arithmetic device canbe brought into full play, and the arithmetic efficiency of the arithmetic devices is improved.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence algorithms, in particular to a processing method of a neural network calculation graph, a computer storage medium and electronic equipment. Background technique [0002] In the existing methods for processing neural network calculation graphs, each operator in the neural network calculation graph can only run on an ASIC (Application Specific Integrated Circuit, ASIC), Graphic Processing Unit (Graphic Processing Unit) within the same running time. , GPU), tensor processor (Tensor Processing Unit, TPU) computing equipment, or run in the central processing unit (Central Processing Unit, CPU), programmable logic gate array (Field-Programmable Gate Array, FPGA) computing equipment . For the edge computing inference platform with heterogeneous architecture, when there are computing devices such as central processing units, graphics processing units, tensor processors, application-specif...

AI Technical Summary

Problems solved by technology

For the edge computing inference platform with heterogeneous architecture, when there are computing devices such as central processing units, graphics processing units, tensor processors, application-specific integrated circuits, and field programmable logic gate arrays in the computing devices, the existing neural network computing graph The processing method cannot make full use of them to cooperate to complete the overall task

Specifically, for example, it is impossible to map each operator to a computing device that matches the characteristics of each operator according to the characteristics of the control logic complexity and calculation amount of each operator in the neural network calculation graph, resulting in the computational efficiency of the computing device relatively low

Images