Convolution Computing Apparatus Parallel Processing Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning neural network models face inefficiencies in convolution operations due to varying sizes of input feature maps and weights, which affect computing performance across different applications such as image and speech processing.
Innovation Solution
A computing apparatus with a master processing circuit and multiple slave processing circuits that divide input feature maps and weights into blocks, allowing for parallel processing and optimized convolution operations by aligning data according to the lowest storage dimension and output channel dimension, respectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the input feature map and weight are processed as whole without division, then the convolution operation can be performed with simple data structure, but the processing capability of single operation apparatus is insufficient and parallel processing cannot be effectively utilized
Solution Approach 1:
The patent divides the input feature map into multiple feature map blocks and the weight into multiple weight blocks according to their storage dimensions. This segmentation enables the data to be processed in manageable chunks by multiple slave processing circuits simultaneously, resolving the contradiction between processing capability and data size.
Solution Approach 2:
The patent introduces a new dimension of organization by blocking data according to storage dimensions (lowest storage dimension for feature maps, output channel dimension for weights). This dimensional reorganization enables efficient parallel processing while maintaining data integrity, allowing the system to scale productivity without proportionally increasing complexity.
2Productivity
If block division is implemented to enable parallel processing, then processing capability and efficiency are improved, but data throughput increases and data transmission burden increases
Solution Approach 1:
The patent merges multiple data blocks into a unified broadcasting mechanism where feature map blocks are broadcast to scheduled slave processing circuits. This combining approach allows multiple processing units to work simultaneously on different blocks while sharing data efficiently, improving productivity without linearly increasing data throughput requirements.
Solution Approach 2:
The patent performs preliminary division and scheduling of data blocks before the actual convolution operation. By pre-organizing feature map blocks and weight blocks according to storage dimensions and scheduling which blocks go to which slave processing circuits, the system reduces runtime data transmission overhead and optimizes data throughput during processing.
3Adaptability or versatility
If the neural network model is applied to different fields with varying input dimensions, then the model's adaptability is improved, but the convolution operation performance becomes inconsistent across different applications
Solution Approach 1:
The patent implements a dynamic block division strategy where the feature map and weight blocks are organized according to their storage dimensions rather than fixed dimensions. This dynamic approach allows the same processing architecture to efficiently handle varying input sizes across different applications (image processing, speech recognition, machine translation) while maintaining consistent performance through adaptive data blocking and scheduling.
Data Source
AI summary
The present disclosure discloses a computing apparatus, a method for implementing a convolution operation by using a computing apparatus, and related products. The computing apparatus is included in a combined processing apparatus. The combined processing apparatus further includes an interface apparatus and other processing apparatus. The computing apparatus interacts with other processing apparatus to jointly complete a computing operation specified by a user. The combined processing apparatus further includes a storage apparatus. The storage apparatus is respectively connected to the computing apparatus and other processing apparatus and is configured to store data of the computing apparatus and other processing apparatus. A solution of the present disclosure optimizes a convolution operation and improves operation processing efficiency.