Tensor Processing Descriptors for Direct Data Addressing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing methods require setting parameters for data access, which reduces processing efficiency as processors need to determine data addresses based on specified parameters, increasing complexity.
Innovation Solution
A data processing apparatus and method that utilizes a descriptor to indicate the shape of tensor data, allowing the processor to determine data addresses directly from the descriptor, reducing the need for explicit address input during data access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If parameters for data access are explicitly set by programmers, then data access control is precise, but processing efficiency decreases due to increased complexity in determining data addresses
Solution Approach 1:
The patent introduces a descriptor as an intermediary data structure that contains pre-calculated data address information. Instead of programmers directly calculating addresses using complex parameters, the descriptor serves as a mediator that stores the relationship between logical data identifiers and physical memory addresses. This resolves the contradiction by maintaining precise data access control through the descriptor while simplifying the programming process and improving processing efficiency.
Solution Approach 2:
The patent applies preliminary action by pre-calculating and storing data address information in the descriptor before the actual data processing occurs. The descriptor is prepared in advance with all necessary address mapping information, so that during execution, the processor can directly use the stored address information without performing complex address calculations at runtime. This improves processing efficiency while maintaining access precision.
2Measurement precision
If complex parameter setting is required for data access, then data access control is accurate, but device complexity increases
Solution Approach 1:
The descriptor acts as an intermediary layer between the simple logical data identifiers used in programming and the complex physical memory addresses. This intermediary structure contains all the complex address mapping information, allowing programmers to work with simple identifiers while the descriptor handles the complexity of address calculation and management, thus reducing overall system complexity.
Solution Approach 2:
The patent creates a simplified copy of the address space information in the descriptor. Instead of requiring programmers to work with complex physical address calculations, the descriptor contains a simplified representation (logical to physical address mapping) that captures all necessary address information in an easily accessible format, reducing the complexity of the data access mechanism.
3Adaptability or versatility
If data addresses are determined based on specified parameters, then data access is flexible, but processing time increases
Solution Approach 1:
The descriptor is prepared in advance with all data address mapping information, performing the complex address determination work before execution. During actual data processing, the processor simply retrieves pre-calculated address information from the descriptor, eliminating runtime address calculation overhead and reducing processing time while maintaining flexible data access capabilities.
Solution Approach 2:
The descriptor contains self-contained address mapping information that allows the system to determine data addresses autonomously without requiring external parameter specification during execution. The descriptor serves itself by providing all necessary address information internally, enabling fast and flexible data access without time-consuming parameter processing.
Data Source
AI summary
A data processing method includes obtaining content of a descriptor when an operand of a first processing instruction includes the descriptor, where the descriptor is configured to indicate a shape of tensor data and to indicate data address of the tensor data, and executing the first processing instruction according to the content of the descriptor by determining the data address of the tensor data corresponding to the operand of the first processing instruction in a data storage space, according to the content of the descriptor, and according to the data address, executing data processing corresponding to the first processing instruction.