AI Layer Task Scheduling With Dependency-Aware Register Configuration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The low computing efficiency of AI engines in deep learning algorithms is due to overheads such as configuration time, data transmission time, and latency caused by incoordination between modules, particularly in convolutional neural networks (CNNs) with multiple convolution combination layers.
Innovation Solution
A computing task scheduling method that involves acquiring network description data, determining dependency relationships, detecting idle register groups, generating target register configurations, and scheduling next tasks based on these dependencies to optimize the use of AI engine resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a fixed hardware structure is used for the AI engine, then the device complexity is reduced and ease of manufacture is improved, but the computing efficiency deteriorates due to configuration time, data transmission time, and latency
Solution Approach 1:
The patent implements dynamic task scheduling that adapts to different neural network computation patterns. The scheduler dynamically allocates computing resources, adjusts data transmission timing, and optimizes configuration operations based on real-time workload characteristics, transforming the static AI engine into a dynamically adaptable system that maintains high efficiency across diverse workloads
2Power
If multiple convolution combination layers are added to increase computing capability, then the processing power is improved, but the computing efficiency deteriorates due to increased overhead from configuration time and data transmission
Solution Approach 1:
The patent implements continuous task scheduling that eliminates idle periods between convolution layers. By maintaining a ready queue of pre-processed tasks and continuously allocating work to available computing units, the system ensures that computing resources remain actively engaged throughout the multi-layer processing, preventing configuration and data transmission overhead from interrupting the computational flow
Solution Approach 2:
The patent performs preliminary task generation and queuing before actual computation begins. Tasks for subsequent convolution layers are prepared in advance and organized in dependency-order queues, so that when computing units become available, they can immediately execute pre-prepared tasks without waiting for configuration or data preparation, thereby reducing overhead time
3Productivity
If module coordination is improved to reduce latency, then the computing efficiency is improved, but the device complexity increases
Solution Approach 1:
The patent introduces a task scheduler as an intermediary layer between the neural network computation graph and the physical AI computing units. This scheduler mediates task allocation, coordinates data transmission timing, and manages configuration operations, thereby improving module coordination and reducing latency without requiring complex hardware modifications to the underlying computing units
Data Source
AI summary
A computing task scheduling method and apparatus, an electronic device, and a computer readable storage medium are disclosed. The method includes: acquiring network description data; determining a dependency relationship between network layer tasks on the basis of the network description data; detecting an idle register group on an artificial intelligence accelerated computing component; generating a target register configuration on the basis of the dependency relationship, the network description data, and a type of the idle register group, and writing the target register configuration into the idle register group; and if a target network layer task completion notification is detected, determining, on the basis of the dependency relationship, a next task of a target network layer task corresponding to the target network layer task completion notification, and starting the next task on a corresponding first function module on the artificial intelligence accelerated computing component so as to execute the next task.