A Neural Network Heterogeneous Acceleration Method and System Based on Asynchronous Events

Abstract

The embodiments of the invention provide a neural-network heterogeneous acceleration method. The method comprises the following steps that a main control unit completes the basic configuration of an accelerator during an initialization phase; the main control unit stores data that needs to be processed by the accelerator in a system memory; the main control unit stores a command descriptor in thesystem memory; the main control unit stores command words in the system memory according to a queue mode; the main control unit notifies the accelerator of a command article number which needs to be processed; the accelerator reads the command words from the system memory based on the configuration of the initialization phase and completes command word parsing, and simultaneously reads the data which needs to be processed from the system memory; the accelerator stores a calculation result in the first storage position of the system memory; and the main control unit directly reads the first storage position of the system memory at an execution interval so as to obtain the calculation result of the accelerator.

Description

technical field [0001] The invention relates to the field of computers, in particular to a method and system for accelerating neural network heterogeneity based on asynchronous events. Background technique [0002] The field of neural networks is very broad and involves a variety of disciplines, attracting the interest of researchers in many different fields, and has broad application prospects in various industries, such as engineering, physics, neurology, psychology, medicine, etc. , mathematics, computer science, chemistry and economics. At the same time, it is also a very important core technology in the field of artificial intelligence. Using neural computing methods to solve certain problems has many advantages, such as strong fault tolerance and self-learning ability. [0003] At present, neural network computing data models are various and complex in structure. The mainstream model contains dozens of hidden layers, each layer contains tens of thousands of neurons, ...

AI Technical Summary

Problems solved by technology

[0008] In the prior art, due to the difference in design goals and computing performance between the host and the accelerator (such as a programmable device), the sending and moving of the computing data in the acceleration process of the heterogeneous computing system needs to be performed by the host, so it takes a lot of money. Host time and wait time

In addition, after the calculation is completed, the host obtains the calculation results from the internal storage of the accelerator system and saves them, which will also bring a lot of system time consumption, which seriously affects the calculation efficiency of the entire heterogeneous system.

Images