Progressive Lossy Memory Compression for GPU Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computing systems face challenges in increasing image throughput to enhance user interaction, particularly in handling realistic images, due to limitations in graphics processing unit (GPU) performance and memory compression efficiency.
Innovation Solution
A method involving progressive lossy memory compression that performs difference transformation, length selection, prioritized ordering, and packing using varying sized bit fields to minimize memory storage requirements while maintaining image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional memory compression is used, then memory storage capacity is improved, but image throughput and quality are degraded
Solution Approach 1:
The patent applies dynamics by making the bit field sizes variable rather than fixed. The packing structure dynamically adjusts the number of bits allocated to each image sample based on the actual data requirements, allowing the compression scheme to adapt to different image content characteristics. This dynamic allocation enables maintaining higher image throughput while achieving effective memory compression.
Solution Approach 2:
The patent changes the parameter of bit field size from a fixed value to a variable parameter. By utilizing varying sized bit fields, the system can optimize the balance between compression ratio and image quality. This parameter change allows the memory compression to preserve more image detail when necessary while achieving higher compression when possible, thus improving both storage capacity and throughput.
2Device complexity
If fixed bit field packing is used, then memory compression is simplified, but image quality and throughput are degraded
Solution Approach 1:
The patent introduces dynamic bit field allocation that adjusts the number of bits per sample based on the distribution and characteristics of the compressed image data. This dynamic approach allows the system to allocate more bits to important image regions and fewer bits to less critical areas, thereby maintaining higher image quality without requiring a uniformly complex compression structure throughout the entire memory system.
Solution Approach 2:
The patent applies local quality by allowing different regions or samples of the image data to have different bit field sizes. Instead of applying a uniform compression scheme, the system locally adapts the precision of each sample's representation based on its specific characteristics. This local optimization maintains image quality where needed while achieving compression where possible, resolving the contradiction between simplicity and quality.
3Reliability
If lossless compression is used, then image quality is maintained, but memory bandwidth and storage efficiency are reduced
Solution Approach 1:
The patent changes the compression approach from strictly lossless to a controlled lossy model using variable bit fields. By allowing some information loss in exchange for higher compression ratios, the system achieves better memory bandwidth efficiency. The variable bit field structure enables the loss to be distributed and controlled, maintaining acceptable image quality while significantly improving throughput and storage efficiency.
Solution Approach 2:
The patent applies partial action by selectively applying different compression levels to different parts of the image data. Rather than uniformly applying lossless compression to all data, the system uses variable bit fields to apply compression where acceptable and maintain higher fidelity where needed. This partial approach to lossless compression improves overall bandwidth efficiency while maintaining sufficient image quality.
Data Source
AI summary
A method, in one embodiment, can include performing difference transformation of image samples. In addition, the method can also include performing length selection. The method can also include performing a prioritized ordering of difference data. Furthermore, the method can include performing packing that includes utilizing varying sized bit fields to produce a lossy compressed representation.