Inter-subgroup Data Transfer via Barrier Synchronization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current processing units lack efficient mechanisms for inter-subgroup data transfer between work-items belonging to different subgroups, limiting the reuse of memory resources and hindering parallel processing efficiency.
Innovation Solution
The implementation of inter-subgroup data transfer techniques using barriers and specialized API functions, such as gentype copy, gentype shuffle, and subgroup_exchange, allows for data sharing between subgroups without relying on local memory, enabling simultaneous data transfer between multiple subgroups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data transfer between subgroups is enabled using traditional methods, then data sharing between work-items is possible, but memory resource reuse is limited and processing efficiency is hindered
Solution Approach 1:
The patent merges data transfer operations between subgroups into a unified mechanism that operates at the subgroup level. By combining multiple data transfer operations into single collective operations (such as all-to-all, broadcast, or gather operations), the system enables efficient inter-subgroup communication while reducing the complexity of individual memory management operations. This allows work-items across different subgroups to share data directly through the unified transfer mechanism rather than requiring separate memory management for each transfer.
2Productivity
If local memory is used for data sharing between subgroups, then data transfer is possible, but memory resource reuse is reduced
Solution Approach 1:
The patent extracts the data transfer functionality from local memory operations and creates a separate inter-subgroup data transfer mechanism. Instead of relying on local memory as the intermediary for data sharing, the system implements dedicated transfer operations that move data directly between subgroups. This extraction allows local memory to be used more efficiently for its primary purposes while enabling data sharing between subgroups through the specialized transfer mechanism, thereby increasing overall memory resource reuse.
3Adaptability or versatility
If inter-subgroup data transfer is implemented, then data sharing between work-items of different subgroups is enabled, but the mechanism complexity increases
Solution Approach 1:
The patent implements a universal data transfer mechanism that handles multiple data sharing scenarios through a single set of operations. The inter-subgroup transfer mechanism supports various communication patterns (one-to-one, one-to-many, many-to-one, many-to-many) using unified collective operations. This multi-functional approach enables diverse data sharing capabilities between work-items of different subgroups while maintaining a consistent and manageable operational interface, reducing the apparent complexity despite the versatility provided.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In an example, a method of transferring data may include synchronizing work-items corresponding to a first subgroup and work-items corresponding to a second subgroup with a barrier. The method may include performing an inter-subgroup data transfer between the first subgroup and the second subgroup.