Computational Graph Transformation for Cross-System Code Portability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computation systems face inefficiencies in data movement and application overheads due to the labor-intensive process of manually selecting computational graph transformations, which hinders optimization efficiency and performance portability.
Innovation Solution
A data processing apparatus and method that performs multi-level IR decomposition on original application code to extract abundant computation and data flow features, followed by graph and parameter transformation processing to generate target application code optimized for different systems, enhancing performance portability and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If manual selection of computational graph transformation is used, then user control over optimization is improved, but optimization efficiency and productivity deteriorate
Solution Approach 1:
The system performs self-service by automatically selecting and applying computational graph transformation methods without requiring user intervention. The apparatus autonomously analyzes the computational graph, identifies optimization opportunities, and applies appropriate transformation methods, thereby eliminating the need for users to manually select transformation methods while maintaining high optimization efficiency.
Solution Approach 2:
The patent replaces the manual mechanical process of user selection with an automated computational system. The apparatus uses algorithms and data structures to automatically determine optimal transformation methods, substituting human manual operation with an automated mechanical/computational process that achieves both ease of operation and high productivity.
2Loss of information
If multi-level IR decomposition is performed, then extraction of computation and data flow features is improved, but device complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the intermediate representation into multiple levels (e.g., high-level IR, low-level IR). Each level captures different aspects of computation and data flow features, allowing comprehensive feature extraction without overwhelming complexity. The segmented approach processes information hierarchically, managing complexity through structured decomposition.
Solution Approach 2:
The patent introduces an additional dimension by creating multiple levels of IR representation rather than a single flat structure. This multi-level dimensionality allows the system to capture features at different abstraction levels simultaneously, improving feature extraction capability while organizing complexity in a structured hierarchical manner.
3Productivity
If automated graph transformation is implemented, then optimization efficiency is improved, but adaptability to different systems deteriorates
Solution Approach 1:
The patent implements universality by designing the apparatus to handle multiple types of computational graphs and support various transformation methods within a single unified system. The multi-level IR decomposition and automated selection mechanism can adapt to different computational patterns and target systems, providing both high optimization efficiency and broad adaptability through a universal framework.
Data Source
AI summary
A data processing apparatus and method performs multi-level IR decomposition on original application code, extract more abundant computation and data flow features to obtain an initial computational graph, and perform graph transformation processing on the initial computational graph to obtain a target computational graph, allowing an application corresponding to original application code to run in different systems and implement performance portability.