Depthwise Convolution Multiplier Distribution for Resource Utilization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Depthwise convolution in neural networks leads to resource waste due to the underutilization of multipliers in computing systems, as each convolution kernel is only responsible for one channel, resulting in inefficient computation and idle resources.
Innovation Solution
A method and apparatus that equally distribute the multipliers used in standard convolution across two parts of the depthwise convolution, allowing for simultaneous computation of quantified results for multiple block units, thereby optimizing the use of resources and increasing computing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If depthwise convolution is adopted for quantitative computation, then the computation complexity is reduced and each convolution kernel is only responsible for one channel, but the multipliers are underutilized causing resource waste
Solution Approach 1:
The patent makes the multipliers universal by enabling them to handle both depthwise convolution operations (where each multiplier processes one channel) and standard convolution operations (where multiple multipliers process multiple channels simultaneously). The system dynamically configures the multipliers to perform different functions based on the computation mode, thereby eliminating resource waste while maintaining computational efficiency.
Solution Approach 2:
The patent merges the computation of multiple block units by combining depthwise convolution with standard convolution operations. By distributing n multipliers to process both the first part (depthwise convolution) and second part (standard convolution) simultaneously, the system utilizes all available multipliers effectively, converting idle resources into productive computational capacity.
2Productivity
If depthwise convolution is used, then only one multiplier is needed per channel, but other multipliers remain idle causing resource waste
Solution Approach 1:
The multipliers are designed to be universal resources that can adapt between depthwise convolution mode (high computing efficiency with fewer active multipliers) and standard convolution mode (full resource utilization with all multipliers active). This universality allows the system to maintain high productivity while achieving full resource utilization by switching between operational modes.
Solution Approach 2:
The system dynamically adjusts the operational mode of the multipliers based on computational requirements. During depthwise convolution, the system activates only the necessary number of multipliers for optimal efficiency, while during standard convolution, all multipliers are activated to maximize resource utilization. This dynamic configuration resolves the contradiction between productivity and resource utilization.
3Productivity
If standard convolution is used with n multipliers, then all channels are processed simultaneously, but the same multipliers cannot be efficiently used for depthwise convolution
Solution Approach 1:
The patent creates universal multipliers that can adapt to different convolution types. The same set of n multipliers used for standard convolution (processing all channels simultaneously) can be reconfigured for depthwise convolution (processing one channel at a time). This universality maintains parallel processing capability while adding adaptability to handle different convolution operations efficiently.
Solution Approach 2:
The system dynamically reconfigures the multipliers between standard convolution mode (where all n multipliers process multiple channels in parallel) and depthwise convolution mode (where the same n multipliers process channels sequentially or in smaller groups). This dynamic adaptability allows the system to maintain high productivity across different convolution types while fully utilizing available resources.
Data Source
AI summary
The present application provides a quantitative computation method and apparatus applied to depthwise convolution. The method includes: determining n multipliers adopted for standard convolution in a preset part of quantitative computation; equally distributing the n multipliers to a first part and a second part of depthwise convolution in the quantitative computation; in the depthwise convolution, computing a first result of a target pixel point in a target block unit in the first part by one multiplier in the first part, and computing a second result of the target pixel point in the second part by one multiplier in the second part; and obtaining quantified results of the target block unit specific to the first part and the second part according to the first result and the second result of each target pixel point. According to the present application, resources are utilized to the maximum extent.