Heterogeneous Computing System Custom Operator Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for heterogeneous computing systems, which combine CPU and AI processing units, are limited in implementing new functions required by users due to the narrow application scope of machine learning libraries, as they rely on calling single or spliced operators, which cannot fully meet user demands.
Innovation Solution
A method for acquiring and running an executable file in a heterogeneous computing system, where an AI processing unit program with an entry function and a CPU program with an interface function set are compiled and linked to a machine learning library, allowing users to customize new operators and expand the library's application scope by enabling the CPU to call a target AI processing unit to execute specific functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a single operator or spliced operator in the machine learning library is called, then the implementation is simple, but the application scope is narrow and cannot fully implement new functions required by users
Solution Approach 1:
The patent creates a universal operator execution mechanism that allows the machine learning library to execute not only preset operators but also user-customized operators. The CPU program includes an interface function set that can call AI processing unit programs with entry functions, enabling the system to handle both standard library operators and custom user-defined operators through a unified interface, thereby expanding the application scope while maintaining ease of implementation.
2Adaptability or versatility
If users are allowed to customize new operators, then the application scope expands, but the system complexity increases due to additional compiling and linking processes
Solution Approach 1:
The patent segments the operator execution process into distinct components: the CPU program with interface function set, the AI processing unit program with entry function, and the linking mechanism. This segmentation allows users to customize operators by providing separate program modules that can be independently developed and then linked to the machine learning library, reducing the perceived complexity while expanding functionality.
Solution Approach 2:
The patent introduces a linking mechanism as an intermediary between user-customized AI processing unit programs and the machine learning library. This intermediary automatically connects the interface function set in the CPU program with the entry function in the AI processing unit program, managing the complexity of operator customization without requiring users to directly handle the integration details.
3Loss of time
If preset operators in the machine learning library are used, then the development time is short, but new functions required by users cannot be fully implemented
Solution Approach 1:
The patent prepares the interface function set in the CPU program in advance, which is designed to accept and execute AI processing unit programs with entry functions. This preliminary preparation allows users to quickly extend functionality by simply providing the custom AI processing unit program without needing to modify the existing library structure, thus maintaining fast development while enabling new functions.
Data Source
AI summary
The invention relates to an executable file acquisition method and operation method of a heterogeneous computing system and a related product. According to the executable file obtaining method of the heterogeneous computing system, through the CPU program and the artificial intelligence processor program, the electronic device allows a user to compile a corresponding program according to needs to achieve needed functions, and the application range of a machine learning library is expanded.