Compacted Memory Transactions for LDS Data Transfer

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing processing systems copy empty elements from local memory to system memory, leading to increased memory usage and inefficiency due to unnecessary data transfer.

Innovation Solution

Implement compacted memory transactions using compaction masks to selectively write and read data from local data shares (LDS) to system memory, reducing empty elements and optimizing memory usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If all elements from local memory data structure are copied to system memory, then data completeness is maintained, but system memory footprint increases due to empty elements

Engineering Contradiction:
Improvesystem memory footprintVSAvoiddata completeness
Core Design Contradiction:
Quantity of substanceVSLoss of information

Solution Approach 1:

The patent extracts only the necessary data elements from the local memory data structure to system memory by using compaction masks to identify and transfer only non-empty elements, excluding empty elements from the transfer process

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the state of data transfer by introducing compaction masks that modify which elements are transferred, transforming the transfer process from copying all elements to selectively copying only necessary elements based on their content state

Inventive Principle:
Principle #35Parameter changes

2Productivity

If all elements including empty ones are transferred between memory and local data share, then data structure integrity is maintained, but processing efficiency decreases due to unnecessary data transfer

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidunnecessary data transfer
Core Design Contradiction:
ProductivityVSLoss of energy

Solution Approach 1:

The patent extracts the compaction mask information from the data structure to identify which elements contain valid data, then uses this extracted information to guide the selective transfer process, eliminating unnecessary transfer of empty elements

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent introduces compaction masks as an intermediary mechanism that mediates between the local memory data structure and system memory, controlling which elements are transferred based on their content state without requiring full data structure replication

Inventive Principle:
Principle #24Intermediary (Mediator)

3Quantity of substance

If compaction masks are used to selectively transfer data, then memory footprint is reduced, but device complexity increases

Engineering Contradiction:
Improvememory footprintVSAvoiddevice complexity
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent segments the data transfer process into two distinct phases: first generating compaction masks that identify non-empty elements, then using these masks to guide selective data transfer, dividing the complex operation into manageable segments

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary generation of compaction masks before the actual data transfer operation, preparing the selection criteria in advance to streamline the subsequent transfer process and reduce overall complexity

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentUS12608130B2Compacted memory transactions for local data shares
Publication Date: 2026.04.21 ADVANCED MICRO DEVICES INC
  • US12608130B2 patent drawing
  • US12608130B2 patent drawing
  • US12608130B2 patent drawing

AI summary

To load compacted memory between a system memory and a local data share (LDS), a processing system includes an accelerator unit (AU) connected to a memory unit. The memory unit is configured to identify that compacted data in the system memory is to be written to elements of an LDS based on two or more compaction masks. The memory unit is configured to then determine sources within the memory from which to load compacted data into the elements of the LDS by determining pre-fix sums based on the two or more compaction masks. The memory unit is configured to then load the compacted data from the identified sources to corresponding elements of the LDS.