Work Item Coordinate Decompression for Interface-Limited GPU Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The rate at which work item information can be sent across the interface between a computation requesting unit and a computation sequencing unit in a computing system is limited, restricting the scheduling and execution rate of computational instances.
Innovation Solution
Compress work item coordinate data using swizzle masks to reduce the data sent across the interface, allowing the computation sequencing unit to decompress and execute the instances efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If work item coordinate data is sent uncompressed across the interface, then the computation sequencing unit can execute instances, but the interface transaction rate becomes the limiting factor for scheduling speed
Solution Approach 1:
The patent extracts only the essential information needed for coordinate reconstruction (swizzle masks and valid work item masks) rather than transmitting complete coordinate data. This selective extraction reduces the data volume transmitted across the interface while preserving all necessary information for accurate coordinate recovery at the computation sequencing unit.
Solution Approach 2:
Instead of the conventional approach of transmitting complete coordinate data and having the execution unit process it, the patent inverts the approach by transmitting compressed mask data and having the sequencing unit perform the decomposition. This reversal of the processing location enables more efficient use of the interface bandwidth.
2Productivity
If the interface bandwidth is increased to send more work item data, then the scheduling rate improves, but the device complexity and cost increase
Solution Approach 1:
The patent changes the parameter representation from full coordinate values to compressed mask indicators. By encoding work item validity and coordinate information as binary masks rather than full numerical coordinates, the data transmission requirement is dramatically reduced, allowing high scheduling rates without increasing interface bandwidth.
3Loss of substance
If compressed data formats are used, then the data transmission efficiency improves, but the decompression complexity increases
Solution Approach 1:
The compression of work item coordinate data is performed in advance at the computation requesting unit before transmission. By pre-processing the data into compressed mask format and transmitting only the essential indicators, the actual decompression work is simplified to basic bit manipulation operations at the sequencing unit, reducing the complexity burden.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for receiving compressed work item coordinate data for work items and decompressing it, each work item being identifiable with a swizzle index, the method comprising receiving a work item valid mask, the work item valid mask indicating valid work items; computing the swizzle index for each valid work item, as indicated by the work item valid mask; receiving a first swizzle mask, indicating which bits of the swizzle index for each work item correspond to the value of a first coordinate for that work item; receiving a second swizzle mask, indicating which bits of the swizzle index for each work item correspond to the value of a second coordinate for that work item; determining a first and a second coordinate for each valid work item in dependence on the first and second swizzle masks and the swizzle index computed for that valid work item.