Convolution Apparatus Parallel Product-Sum Calculators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolution operation methods in CNNs face challenges with high computational demands and power consumption due to the large number of convolution operations required, especially when implemented in software on processors, and there is a lack of efficient digital hardware arrangements for executing these operations when filter coefficients are decomposed into lower-dimensional forms.
Innovation Solution
A convolution operation apparatus is designed with a hierarchical structure, featuring multiple product-sum calculators and data suppliers, along with a controller that enables parallel execution of product-sum processing, utilizing ring buffers and shift registers to efficiently handle and process filter kernels, allowing for high-speed convolution operations by decomposing two-dimensional filter coefficients into one-dimensional structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If convolution operations are implemented in software on processors, then flexibility and ease of operation are improved, but computational load and power consumption increase significantly
Solution Approach 1:
The patent replaces software-based convolution operations with a dedicated hardware circuit that performs product-sum calculations. This hardware implementation uses multipliers and accumulators to compute convolution results directly in electrical signals, eliminating the need for software execution on general-purpose processors. The substitution of software with specialized hardware circuitry reduces power consumption while maintaining computational flexibility through configurable filter coefficients and parallel processing capabilities.
2Measurement precision
If the number of convolution operations is increased to improve recognition accuracy, then measurement precision is improved, but computational load and processing time increase
Solution Approach 1:
The patent segments the convolution operation into multiple parallel product-sum calculation units, each handling a portion of the filter coefficients. By decomposing the overall convolution operation into smaller parallel tasks, the system can process multiple filter operations simultaneously, thereby maintaining high recognition accuracy through comprehensive feature extraction while reducing the effective processing time through parallel execution.
Solution Approach 2:
The patent transforms the traditional sequential convolution operation into a parallel hierarchical structure where multiple product-sum calculators operate simultaneously at different levels. This dimensional transformation from sequential to parallel processing allows the system to perform numerous convolution operations with large filter coefficients without proportionally increasing processing time, thus improving both accuracy and efficiency.
3Device complexity
If filter coefficients are decomposed into lower-dimensional forms to reduce computational complexity, then device complexity is reduced, but manufacturing precision and operation reliability may be affected
Solution Approach 1:
The patent segments filter coefficients into multiple smaller coefficient sets that can be processed by individual product-sum calculation units. This segmentation reduces the computational complexity of each individual unit while maintaining the overall filtering capability through parallel processing. The decomposition of large filter coefficients into manageable segments allows for simpler hardware design without compromising the reliability of the convolution operation.
Solution Approach 2:
The patent combines the results from multiple parallel product-sum calculation units to produce the final convolution output. By merging the partial results from simpler, lower-dimensional calculations, the system achieves the equivalent effect of processing large filter coefficients directly. This merging strategy maintains operation reliability while keeping individual computational units simple and manageable.
Data Source
AI summary
There is provided with a convolution operation apparatus. A first product-sum calculator includes a plurality of first multipliers, a plurality of first accumulators, a first data supplier which supplies first data to the plurality of first multipliers, and a second data supplier which supplies a plurality of second data to the plurality of first multipliers. A second product-sum calculator includes a plurality of second multipliers, a plurality of second accumulators, a third data supplier which supplies third data to the plurality of second multipliers, and a fourth data supplier which receives and holds an output from each of the plurality of first accumulators and supply the outputs as fourth data to a corresponding one of the plurality of second multipliers.