Fractional Matrix Division for Flexible Signal Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks face limitations in design flexibility and power consumption due to fractional processing methods that fix the shapes of fractional signal and weight matrices, leading to inefficient matrix multiplication operations.
Innovation Solution
A signal processing method that divides signal and weight matrices into non-square fractional matrices, allowing for more flexible matrix multiplication and accumulation operations, thereby optimizing power consumption by reducing data reuse and improving matrix division flexibility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If fractional processing is performed based on the condition |row - columns| ≤ 1 to obtain square-shaped fractional matrices, then matrix multiplication operation can be performed, but design flexibility is insufficient and power consumption is fixed
Solution Approach 1:
The patent applies dynamics by making the fractional matrix shapes adjustable rather than fixed. The method allows selecting different division strategies (first strategy: |row - columns| ≤ 1; second strategy: other conditions) to obtain fractional matrices with varying shapes, enabling the system to adapt to different computational scenarios and optimize power consumption dynamically based on specific processing requirements
Solution Approach 2:
The patent changes the parameter of matrix shape flexibility by allowing fractional matrices to have different dimensional characteristics. Instead of being constrained to square shapes only, the system can generate fractional matrices with various row-column relationships, thereby changing the operational parameters to achieve both flexibility and power optimization
2Productivity
If the signal matrix and weight matrix are divided into fractional matrices to reduce processing cost, then matrix multiplication can be performed on smaller matrices, but the fixed shape constraint limits optimization capability
Solution Approach 1:
The patent makes the matrix division process dynamic by providing multiple division strategies that can be selected based on the specific matrices being processed. The system can adaptively choose between dividing into square-shaped fractional matrices or other configurations, enabling optimal processing efficiency for different matrix dimensions and data types
3Loss of energy
If data reuse is performed during calculation of matrix blocks to reduce power consumption, then reading overhead is reduced, but the fixed fractional matrix shapes prevent further optimization
Solution Approach 1:
The patent enables dynamic optimization of data reuse strategies by allowing flexible fractional matrix shapes. The system can adapt the matrix division and data reuse patterns based on the specific computational tasks, thereby optimizing the balance between data reuse benefits and processing requirements
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
This application provides a signal processing method and apparatus, and relates to the field of computer technologies, so as to improve flexibility of a fractional matrix. The method is applied to a device including a processor, and the method includes: obtaining a signal matrix and a weight matrix, where both the signal matrix and the weight matrix are two-dimensional matrices, and a column quantity of the signal matrix is equal to a row quantity of the weight matrix; dividing the signal matrix to obtain a plurality of first fractional signal matrices of X rows and H columns, and dividing the weight matrix to obtain a plurality of first fractional weight matrices of H rows and Y columns, where neither the first fractional signal matrices nor the first fractional weight matrices approximate to a square shape; and performing matrix multiplication and accumulation operation on the plurality of first fractional signal matrices and the plurality of first fractional weight matrices, to obtain a plurality of matrix operation results, where the plurality of matrix operation results are used to form a signal processing result.