Algorithm Graph Rewriting to Reduce CPU-GPU Data Copying

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data processing methods require frequent switching between CPU and GPU, leading to significant computing and memory overheads due to multiple data copies, which affects processing efficiency.

Innovation Solution

A data processing method that involves obtaining an algorithm directed graph, re-editing algorithms meeting specific conditions to support GPU execution, and invoking processors accordingly to minimize data copying by executing algorithms on the GPU where possible.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If data is copied multiple times between CPU and GPU during algorithm execution, then the task can be completed using both processors, but the computing and memory overheads increase significantly

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidcomputing and memory overheads

Core Design Contradiction:

ProductivityVSLoss of energy

Solution Approach 1:

The algorithm directed graph is segmented into CPU-executable nodes and GPU-executable nodes based on computational characteristics. This segmentation allows each processor type to handle tasks最适合其 capabilities, reducing unnecessary data copying between processors and optimizing the balance between processing efficiency and resource overhead.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Different algorithms within the directed graph are assigned to different processors based on their specific computational requirements. GPU-accelerated algorithms are identified and rewritten to run on GPU, while others remain on CPU. This local optimization reduces the frequency of CPU-GPU switching and associated data copying overheads.

Inventive Principle:

Principle #3Local quality

2Loss of energy

If algorithms are rewritten to execute on GPU, then data copying between CPU and GPU is reduced, but the complexity of algorithm processing increases

Engineering Contradiction:

Improvedata copying overheadsVSAvoidalgorithm processing complexity

Core Design Contradiction:

Loss of energyVSDevice complexity

Solution Approach 1:

The system performs preliminary analysis of the algorithm directed graph to identify which algorithms are suitable for GPU execution. Algorithms meeting specific conditions are pre-rewritten and marked for GPU execution before the actual task processing begins. This preliminary preparation reduces the complexity burden during runtime by establishing the execution plan in advance.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary mechanism that automatically identifies and rewrites algorithms suitable for GPU execution. This intermediary layer handles the complexity of determining which algorithms to rewrite and how to rewrite them, shielding the user from the complexity while enabling optimized GPU execution.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20260017090A1Data processing method and apparatus, device, and storage medium

Publication Date: 2026.01.15 LEMON INC(GB)

AI summary

Embodiments of the present disclosure disclose a data processing method and apparatus, a device, and a storage medium. The method includes: obtaining an algorithm directed graph corresponding to a target task, where the algorithm directed graph includes a plurality of algorithm nodes, and one algorithm node corresponds to one processing algorithm; re-editing a processing algorithm that meets a set condition, to obtain a rewritten processing algorithm; and invoking a first processor to execute a processing algorithm that is not re-edited, and invoking a second processor to execute the rewritten processing algorithm, according to an execution sequence of the algorithm directed graph.