Asynchronous processing method and device based on neural network accelerator

Abstract

The invention discloses an asynchronous processing method and device based on a neural network accelerator, and the method comprises the steps: generating an asynchronous request through a main thread according to an algorithm service demand, generating a corresponding handle number, generating a group of command words according to the generated service request, adding a command word queue, and querying the command word queue, and if only one group of command words exist currently, setting the group of command words to be in an execution state, starting a neural network accelerator hardcore, otherwise, setting the group of command words to be in an initial state and returning. And after the execution of the hard core is finished, generating hardware interruption, starting soft interruption, querying a command word queue, taking out command words from the queue if the command words exist, judging whether the command words are in an initial state, if so, setting the command words to be in an execution state and starting the hard core, and otherwise, deleting the group of command words from the queue. And the main thread queries whether the hard core of the operation is completed or not according to the handle number. According to the operation, the asynchronous operation process of a plurality of network algorithms is realized, the asynchronous operation is processed through hard core interruption, the network performance is accelerated, and the bandwidth occupied by a processor is reduced.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a processing method of a neural network accelerator, and an asynchronous processing method and device based on a neural network accelerator. Background technique [0002] With the development of artificial intelligence, the design of deep learning network is becoming more and more complex. Without neural network accelerator hardware, it cannot meet the requirements of embedded performance. In the security field, the main detection objects are people, vehicles, and objects. While detecting people, vehicles, and objects, it is necessary to track people, vehicles, and objects. Then it is necessary to integrate multiple deep learning algorithms to solve the detection and tracking of people. When a frame of image screen When there are many people, it is necessary to track them one by one, and then use the feature matching strategy to ensure that the same person is tracked; the e...

AI Technical Summary

Problems solved by technology

[0002] With the development of artificial intelligence, the design of deep learning network is becoming more and more complex. Without neural network accelerator hardware, it cannot meet the needs of embedded performance.

In the security field, the main detection objects are people, vehicles, and objects. While detecting people, vehicles, and objects, it is necessary to track people, vehicles, and objects. Then it is necessary to integrate multiple deep learning algorithms to solve the detection and tracking of people. When a frame of image screen When there are many people, it is necessary to track them one by one, and then use the feature matching strategy to ensure that the same person is tracked; the existing invention patent with the publication number of CN111679815A discloses an asynchronous operation method and device, storage medium and related equipment. The invention patent solves the problem of multi-layer nesting, separates the main logic and processing logic, solves the problem of poor flexibility and small scope of application, an asynchronous processing method, device, computer equipment and storage medium for Lua voice, adopts the callback and Promise scheme, and aims at the traditional Asynchronous callback method solves the problem of bloated callback code

Images