3D Convolution Device Dimension Transposing Data Access
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current three-dimensional convolution devices face inefficiencies due to dispersed data storage in the depth and channel dimensions, leading to complex calculations and poor data access efficiency during three-dimensional convolutions.
Innovation Solution
A three-dimensional convolution device and method that performs dimension transposing to consecutively arrange data in the depth and channel dimensions, allowing for block-wise convolution operations and subsequent rearrangement of computed data to enhance processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If data values in depth dimension and channel dimension are stored in a dispersed manner in memory, then the original data format is preserved, but data access efficiency deteriorates and calculation complexity increases
Solution Approach 1:
The patent applies preliminary action by performing dimension transposing on input data before the convolution operation. The data is pre-arranged in a consecutive format in memory, so that when convolution is performed, the data access efficiency is already optimized and no additional complexity is introduced during the actual computation phase.
Solution Approach 2:
The patent changes the dimensional arrangement of data in memory from a dispersed multi-dimensional format to a consecutive one-dimensional format by transposing dimensions. This dimensional transformation allows the convolution operation to access data sequentially, improving access efficiency while maintaining the logical three-dimensional convolution structure through software层面的 dimension management.
2Measurement precision
If multiple data values are accumulated in depth dimension and channel dimension for three-dimensional convolution, then the convolution operation is performed correctly, but more time is needed to read dispersed data values
Solution Approach 1:
The patent pre-arranges all necessary data values in a consecutive manner before the convolution operation. This preliminary arrangement ensures that during the accumulation process for three-dimensional convolution, all required data values can be read continuously without seeking operations, thereby maintaining convolution accuracy while minimizing data reading time.
Solution Approach 2:
By storing data values consecutively in memory according to the transposed dimension format, the patent enables continuous reading of data during the convolution accumulation process. This eliminates gaps and interruptions in data access, allowing the useful action of data reading to proceed continuously without wasted time.
Data Source
AI summary
A three-dimensional convolution method includes performing a dimension transposing operation on input data to consecutively arrange elements of the input data in depth and channel dimensions to further generate first data, performing in blocks a convolution on the first data and second data that corresponds to first weight data to generate computed data, and rearranging the computed data according to an original dimensional format of the input data to generate output data.