Variable-Length Codeword Compression for Faster Neural Parameter Reuse
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing chips achieve low compression rates for data, leading to inefficient storage and processing, particularly in scenarios like neural network inference where network parameters are repeatedly used.
Innovation Solution
A data compression method that assigns codewords based on occurrence frequency and defined lengths, using indication information to simplify allocation and ensure efficient compression, and a decompression method that utilizes decoder groups for parallel processing to enhance efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional compression algorithms are used, then data can be compressed, but the compression rate is low and compression efficiency is poor
Solution Approach 1:
The patent changes the parameter of codeword length from fixed to variable based on occurrence frequency. By dynamically adjusting codeword lengths according to how frequently data patterns appear, the system achieves higher compression rates while maintaining efficient compression through the structured approach of frequency analysis and codeword assignment.
Solution Approach 2:
The patent applies different codeword lengths to different data patterns based on their local occurrence frequency characteristics. Frequently occurring patterns receive shorter codewords while less frequent patterns receive longer codewords, optimizing the overall compression rate without sacrificing compression efficiency through the systematic local adaptation.
2Quantity of substance
If codeword allocation is performed without indication information, then compression can be achieved, but algorithm complexity increases and compression efficiency decreases
Solution Approach 1:
The patent introduces first indication information as an intermediary element that carries the length of the codeword. This intermediary simplifies the decoding process by explicitly providing length information, thereby reducing algorithm complexity while maintaining the compression rate benefits of variable-length codewords.
Solution Approach 2:
The patent performs preliminary encoding by prepending the codeword length information (first indication information) before the actual codeword. This preliminary action simplifies the decoding process by providing advance knowledge of the codeword length, reducing the complexity of the decoding algorithm while preserving compression efficiency.
3Ease of manufacture
If uniform codeword lengths are used, then allocation is simple, but compression rate is low
Solution Approach 1:
The patent transitions from static uniform codeword lengths to dynamic variable-length codewords based on occurrence frequency. By making the codeword length adaptive to the data characteristics, the system achieves higher compression rates while maintaining reasonable allocation simplicity through the structured frequency-based assignment process.
Solution Approach 2:
The patent changes the parameter of codeword length from uniform to variable based on occurrence frequency analysis. By adjusting the codeword length parameter according to how frequently patterns appear in the data, the system achieves superior compression rates while keeping the allocation process manageable through systematic frequency-based categorization.
Data Source
AI summary
This application discloses a data compression method, a data decompression method, apparatus, and system, and a medium, and pertains to the computer field. The method includes: obtaining a codeword corresponding to to-be-compressed first original data; and obtaining first compressed data corresponding to the first original data, where the first compressed data includes first indication information and the codeword, and the first indication information indicates a length of the codeword. According to this application, a compression rate can be improved.