Pixel Block Cache Reading for Neural Network Back Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for computing in deep neural networks require complex and expensive dedicated hardware for padding and insert-zero operations during back propagation, leading to high system power consumption and performance issues.
Innovation Solution
A method that determines a first pixel block in a cache, composed of a 2m row×2n column pixel matrix, reads it, and uses related pixel data for padding, allowing for simultaneous execution of insert-zero and padding operations, thereby reducing the need for complex hardware and improving system performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If dedicated hardware is used for padding and insert-zero operations, then computation accuracy is improved, but device complexity and power consumption increase
Solution Approach 1:
The patent combines padding operations and insert-zero operations into a single unified processing step. By determining a pixel block that includes both original pixel data and related pixel data, and performing both operations simultaneously during back propagation, the system eliminates the need for separate dedicated hardware modules for each operation, thereby reducing device complexity while maintaining computation accuracy.
Solution Approach 2:
The pixel block determination mechanism serves multiple functions: it provides original pixel data for computation, supplies related pixel data for padding operations, and enables insert-zero operations all in one structure. This multi-functional approach replaces multiple specialized hardware components with a single versatile data structure, reducing overall hardware complexity.
2Reliability
If dedicated hardware is used for padding and insert-zero operations, then computation reliability is improved, but power consumption increases
Solution Approach 1:
By merging padding and insert-zero operations into a single unified process that operates on a pre-determined pixel block, the system reduces the total number of computational steps and hardware operations required. This consolidation decreases the active time of computational units and reduces overall power consumption while maintaining the accuracy benefits of both operations.
Solution Approach 2:
The pixel block is determined in advance to include both original pixel data and related pixel data before the actual computation begins. This preliminary preparation allows the main computation to proceed more efficiently with fewer additional operations, reducing the total energy consumption during the critical computation phase while ensuring accuracy is maintained.
3Manufacturing precision
If complex hardware is used for padding operations, then processing accuracy is improved, but productivity decreases
Solution Approach 1:
The patent merges padding operations and insert-zero operations into a single unified processing step that operates on a pre-determined pixel block. This consolidation reduces the total number of processing stages and data transfers required, improving system throughput and productivity while maintaining the processing accuracy of both individual operations through the unified approach.
Data Source
AI summary
A method for computing, a computing device, and a computer-readable storage medium are provided. The method includes determining a first pixel block in a cache. The first pixel block is composed of a 2m row×2n column pixel matrix and includes original pixel data and pixel data related to the original pixel data. The first pixel block is read from the cache. At least part of the pixel data related to the original pixel data is used for padding related to the original pixel data. The original pixel data includes pixel data from the (n+1)th column to the 2nth column in the (m+1)th row to the 2mth row in the 2m row×2 n column pixel matrix. When reading data from the cache, pixel data that needs to be obtained after insert-zero and padding operations on the original pixel data in back propagation can be read at one time.