Direct Memory Access Apparatus for Neural Network Data Movement

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing Direct Memory Access (DMA) systems for neural-network processors require complex hardware structures to manage the mapping relationships of input feature maps and filters, leading to inflexibility and increased complexity.

Innovation Solution

A method for controlling a direct memory access apparatus to write instructions, where a coprocessor executes algorithms to cut data into segments and outputs instructions that the DMA decodes and executes, simplifying the DMA hardware structure and improving flexibility.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If complex hardware structures are used to manage mapping relationships in DMA systems, then data movement capability is improved, but device complexity increases

Engineering Contradiction:

Improvedata movement capabilityVSAvoidhardware structure complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent introduces an intermediary module between the CPU and DMA controller that generates mapping relationship instructions. This intermediary acts as a mediator that handles the complex mapping logic through software/algorithms rather than hardwired hardware, thereby maintaining data movement capability while reducing hardware complexity. The intermediary translates high-level data movement requirements into detailed DMA instructions without requiring complex hardware structures in the DMA controller itself.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces the mechanical/hardware-based mapping management system with a software-based instruction generation approach. Instead of using complex hardware circuits to manage mapping relationships, the system uses algorithms executed by a CPU or coprocessor to generate mapping instructions, which are then transmitted to the DMA controller. This substitution of mechanical hardware logic with software algorithms significantly reduces device complexity while preserving functionality.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Productivity

If complex hardware structures are used to manage mapping relationships, then data movement capability is improved, but adaptability decreases

Engineering Contradiction:

Improvedata movement capabilityVSAvoidsystem flexibility

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamic adaptability by generating mapping relationship instructions through executable algorithms rather than fixed hardware configurations. The mapping relationships can be dynamically adjusted by modifying the instruction-generating algorithms, allowing the system to adapt to different neural network architectures and data formats without hardware changes. This dynamic approach maintains data movement capability while significantly improving system flexibility and versatility.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent enables parameter changes in mapping relationships by controlling them through software instructions rather than fixed hardware parameters. The mapping relationships can be modified by changing instruction parameters such as source addresses, destination addresses, transfer sizes, and timing parameters, allowing the same hardware to adapt to various neural network processing requirements without structural modifications.

Inventive Principle:

Principle #35Parameter changes

3Device complexity

If hardware structures are simplified by using instruction-based control, then device complexity is reduced, but manufacturing precision requirements increase

Engineering Contradiction:

ImproveDMA hardware structureVSAvoidinstruction decoding precision

Core Design Contradiction:

Device complexityVSManufacturing precision

Solution Approach 1:

The patent uses instruction templates and standardized instruction formats that can be precisely replicated and decoded. The mapping relationship instructions follow defined formats with specific fields for source address, destination address, transfer size, and control parameters. This standardized instruction copying approach ensures high decoding precision while keeping the hardware structure simple, as the decoder only needs to recognize and interpret standardized instruction formats rather than handle complex hardware signals.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS12216600B2Neural network processing system for data moving method using direct memory access apparatus

Publication Date: 2025.02.04 BEIJING ESWIN COMPUTING TECH CO LTD

AI summary

A data moving method for a direct memory access apparatus is disclosed, and the data moving method for the direct memory access apparatus includes: receiving, by the direct memory access apparatus, an object data moving instruction and decoding the object data moving instruction, wherein the object data moving instruction includes a first field, and the first field of the object data moving instruction indicates a data moving operation for a neural-network processor; and executing, by the direct memory access apparatus, the object data moving instruction.