Arithmetic Processing Device for Deep Learning Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for performing convolution operations in deep learning, such as those used in neural networks, face inefficiencies due to high calculation costs and processing speed limitations when adjusting data access to avoid reading the same data multiple times, particularly when the filter movement exceeds two columns and rows, leading to decreased processing speed and potential data access issues.
Innovation Solution
An arithmetic processing device that optimizes data storage and access by using a data storing unit to store input and weight data as matrices and performs row portion operations based on the number of columns of the weight data, allowing for efficient data reuse and reduced calculation overhead through strategic data movement and storage in registers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If data access is adjusted to avoid reading the same data multiple times, then calculation cost is reduced, but processing speed decreases when filter movement exceeds two columns and rows
Solution Approach 1:
The patent segments the convolution operation into multiple processing stages: pre-processing stage that loads and stores data in memory, and main processing stage that performs calculations using stored data. This segmentation allows the system to optimize data access patterns by preparing data structures in advance, reducing redundant reads during the main processing phase while maintaining high processing speed through efficient parallel computation.
Solution Approach 2:
The patent implements preliminary action by performing data pre-processing before the main convolution operation. Data is loaded from external memory into high-speed memory, pre-processed to eliminate redundant information, and stored in optimized data structures. This preliminary preparation reduces the need for repeated data access during the main processing stage, thereby reducing calculation cost without compromising processing speed.
2Area of stationary object
If filter movement exceeds two columns and rows, then convolution coverage is improved, but data access issues and processing speed degradation occur
Solution Approach 1:
The patent transitions from processing data in traditional row-major or column-major order to a block-based two-dimensional data structure. By organizing data into blocks that can be accessed in multiple dimensions, the system efficiently handles large filter movements (exceeding 2x2) without degrading performance. This dimensional reorganization allows parallel access patterns that maintain high processing speed while covering larger convolution areas.
Solution Approach 2:
The patent introduces an intermediary data structure (high-speed memory with pre-processed data blocks) between external memory and the computation units. This intermediary layer buffers data for large filter movements, providing fast access to multiple data elements simultaneously. The intermediary structure eliminates data access bottlenecks that would otherwise occur with large filter movements, maintaining processing speed while enabling extended convolution coverage.
3Ease of manufacture
If conventional data access methods are used, then implementation is simple, but redundant calculations occur and processing efficiency decreases
Solution Approach 1:
The patent changes the parameter of data organization from conventional linear arrays to block-based two-dimensional structures. This parameter change enables more efficient memory access patterns and data reuse during convolution operations. The new data structure allows the system to process larger filter movements and reduce redundant calculations while maintaining implementation feasibility through systematic data organization and processing stages.
Data Source
AI summary
A memory 11 stores therein first data and second data each of which has element data that forms a matrix. Arithmetic units 51 to 53 repeat, for each of a first predetermined row of the first data and a second predetermined row of the second data that are stored in the memory 11, by using the element data included in the first predetermined row and the element data included in the second predetermined row, a row portion operation based on the number of columns in the second data and performs, by using results of the row portion operations, an arithmetic operation process that acquires the operation results of the operation that uses the first data and the second data.