3D Convolution Architecture Using Parallel Systolic Arrays
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Three-dimensional convolution calculations in video data processing are inefficient due to high computational demands, leading to slow processing times.
Innovation Solution
A four-dimensional systolic calculation architecture is constructed using multiple three-dimensional systolic arrays for parallel calculations on feature and weight value matrices, reducing dimensionality and enhancing parallelism.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If three-dimensional data is reduced to two-dimensional or one-dimensional for parallel calculation, then the calculation can be performed, but the processing speed is slow due to huge computational amount
Solution Approach 1:
The patent transforms the traditional approach by maintaining three-dimensional data structure and introducing a fourth dimension (time dimension) to create a 4D systolic array architecture. This allows parallel computation of 3D convolution operations without reducing data dimensions, thereby improving both processing speed and computational efficiency simultaneously.
Solution Approach 2:
The patent divides the 3D convolution operation into multiple 2D convolution operations that can be executed in parallel across multiple systolic arrays. By segmenting the computational task and distributing it across parallel processing units, the system achieves high-speed processing while maintaining 3D data integrity.
2Measurement precision
If three-dimensional convolution is performed directly, then accurate feature extraction is achieved, but the calculation time is long due to huge computational amount
Solution Approach 1:
The patent performs preliminary data preparation and reorganization into 3D data structures before entering the systolic array. This pre-processing step optimizes data layout for parallel computation, enabling accurate 3D convolution to be executed faster by eliminating computational bottlenecks during the actual feature extraction process.
Solution Approach 2:
The patent creates multiple copies of 3D data across parallel systolic arrays, allowing simultaneous computation of convolution operations. Each array processes a portion of the computation, and the results are aggregated to produce the final feature extraction output, significantly reducing calculation time while maintaining accuracy.
3Productivity
If multiple three-dimensional systolic arrays are used for parallel calculations, then processing efficiency is improved, but the device complexity increases
Solution Approach 1:
The patent designs a universal 3D systolic array architecture that can handle various 3D convolution operations through a single standardized interface. Multiple arrays use the same structure and control logic, allowing parallel processing while maintaining manageable system complexity through modularity and reusability of the same computational unit.
Data Source
AI summary
A video data processing method, including: obtaining three-dimensional feature data and three-dimensional weight data corresponding to video data; separately preprocessing the three-dimensional feature data and the three-dimensional weight data to obtain a feature value matrix and a weight value matrix; and inputting the feature value matrix and the weight value matrix into a plurality of three-dimensional systolic arrays for parallel computing to obtain a video data processing result. The present method can fully expand the degree of parallelism of computation and a four-dimensional systolic computation architecture is constructed by using multiple three-dimensional systolic arrays, so as to perform parallel computing on a three-dimensional feature value matrix and a three-dimensional weight value matrix, thereby shortening the computation time of three-dimensional convolution, and improving the video data processing efficiency.