A three-dimensional deconvolution acceleration method and three-dimensional deconvolution hardware acceleration architecture
The proposed method and architecture for 3D deconvolution transform 3D operations into efficient 2D steps, addressing the inefficiency of current methods by reducing multiplications and enhancing computational efficiency.
CN114742215BActive Publication Date: 2025-03-28NANJING UNIV
0 Cites 0 Cited by
Patent Information
- Application Number
- CN202210393362.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-14
- Publication Date
- 2025-03-28
- Estimated Expiration
- 2042-04-14
AI Technical Summary
Technical Problem
Current methods for accelerating 3D deconvolution in 3D-GANs are inefficient as they fail to fully leverage the third dimension, resulting in low computational efficiency for 3D deconvolution operations.
Method used
A novel method and hardware architecture that transforms 3D deconvolution into a series of 2D operations using pre-defined matrices and rotations, reducing the number of multiplications by converting the 3D deconvolution process into pre-processing, element-wise multiplication, and post-processing steps.
Benefits of technology
This approach significantly enhances computational efficiency by reducing the number of multiplications and lowering the complexity of calculations, thereby improving the overall performance of 3D deconvolution.
✦ Generated by Eureka AI based on patent content.
Abstract
The present application provides a three-dimensional deconvolution acceleration method and a three-dimensional deconvolution hardware acceleration architecture. The method includes: using the size of the candidate deconvolution kernel, the preset transformation strength and the deconvolution step size to predetermine the size of the candidate input block, the size of the data block to be operated and the size of the output data block, respectively using the corresponding two-dimensional preset pre-processing matrix and the pre-processing transpose matrix to transform the candidate input block and the candidate deconvolution kernel, and multiplying the pre-processing data blocks obtained by each element to obtain the data block to be operated, and then using the two-dimensional preset post-processing matrix and the post-processing transpose matrix to transform the data block to be operated, and finally obtain the output data block corresponding to the candidate input block. The whole method uses the correlation between the data before and after the deconvolution calculation to convert the multiplication and accumulation operation of the conventional three-dimensional deconvolution into pre- and post-processing and element-by-element multiplication, which can reduce the number of multiplications and reduce the computational complexity, thereby greatly improving the computational efficiency.
Need to check novelty before this filing date? Find Prior Art