Work Item Coordinate Compression for Faster 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, affecting the scheduling and execution of compute tasks, particularly in GPUs, due to the need to transmit large amounts of coordinate data per cycle.
Innovation Solution
A method involving the creation of a work item valid mask and computation of swizzle masks to compress work item coordinate data, allowing for the transmission of compressed data across the interface, which is then decompressed for execution by the computation sequencing unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If work item coordinate data is transmitted across the interface between computation requesting unit and computation sequencing unit, then work items can be scheduled and executed, but the data transmission rate limits the scheduling rate
Solution Approach 1:
The patent segments the coordinate data transmission by separating valid work item indicators from actual coordinate values. The computation requesting unit transmits only valid work item masks and coordinate data, while the computation sequencing unit reconstructs full coordinate information locally, dividing the transmission burden and improving scheduling rate.
Solution Approach 2:
The computation requesting unit performs preliminary action by pre-calculating and transmitting only the essential valid work item masks and coordinate data before the computation sequencing unit reconstructs the full coordinate information. This preliminary preparation reduces the data transmission requirement and increases the instance scheduling rate.
2Reliability
If large amounts of coordinate data are transmitted per cycle, then complete work item information is provided, but the interface data width requirement increases
Solution Approach 1:
The patent extracts only the essential components (valid work item masks and coordinate data) from the complete coordinate information and transmits them across the interface. The computation sequencing unit then reconstructs the full coordinate information locally, reducing interface data width requirements while maintaining work item information completeness.
Solution Approach 2:
The valid work item mask acts as an intermediary that enables the computation sequencing unit to reconstruct complete coordinate information from partial data. This intermediary mechanism allows reliable work item information to be obtained without transmitting all coordinate data directly across the interface.
3Measurement precision
If all work item coordinates are transmitted, then accurate scheduling is achieved, but redundant data transmission occurs
Solution Approach 1:
The patent applies local quality by transmitting different types of data to different units based on their specific needs. The computation requesting unit transmits valid work item masks and coordinate data, while the computation sequencing unit reconstructs full coordinate information locally for accurate scheduling, avoiding redundant transmission of data that can be generated locally.
Data Source
AI summary
A method for compressing work item coordinate data for work items in a work group and sending the data across an interface between a computation requesting unit and a computation sequencing unit. A work item valid mask is created in dependence on the number and positions of work items in the work group, the work item valid mask indicating valid work items in the work group. A first swizzle mask indicates which bits of a swizzle index for each work item in the work group correspond to the value of a first coordinate for that work item. A second swizzle mask indicates which bits of the swizzle index for each work item in the work group correspond to the value of a second coordinate for that work item.