Selective Attribute Distribution in Parallel Graphics Processors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In conventional graphics processing systems with parallel processors, efficiently distributing attributes to the correct processing units is challenging, leading to congestion, timing issues, and potential deadlock conditions, as all processors receive attributes regardless of their workload.
Innovation Solution
A method where attributes are selectively distributed to processing units only when the end of a primitive is reached, using a primitive pipeline to deliver tiles and an attribute pipeline to withhold attributes from uninvolved units, preventing unnecessary attribute delivery and reducing congestion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If attributes are distributed to all parallel processors, then all processors receive necessary information, but congestion and timing problems occur due to unnecessary attribute delivery
Solution Approach 1:
The patent applies local quality by selectively distributing attributes only to the specific processing units that need them, rather than uniformly to all processors. The attribute distribution is localized to match the actual workload distribution, where each processing unit receives attributes only for primitives assigned to it, eliminating unnecessary attribute delivery and reducing congestion.
Solution Approach 2:
The patent implements preliminary action by determining which processing units will handle which primitives before attribute distribution occurs. The system pre-calculates the assignment of primitives to processing units based on tile ownership, then uses this information to guide attribute distribution, ensuring that attributes are delivered only to the correct recipients from the outset.
2Reliability
If attributes are distributed to all processing units, then no attribute is withheld, but deadlock conditions and congestion arise
Solution Approach 1:
The system performs preliminary determination of primitive-to-processing-unit assignments before attribute distribution. By pre-establishing which processing units will handle which primitives, the system can then selectively distribute attributes without risking deadlock, as the distribution pattern is predetermined and matches the actual processing workload.
Solution Approach 2:
The patent employs feedback mechanisms where the system monitors which processing units are assigned to handle which primitives, then uses this information to control attribute distribution. The feedback loop ensures that attribute distribution decisions are based on actual workload requirements, preventing both congestion and deadlock while maintaining system reliability.
3Productivity
If selective attribute distribution is implemented, then congestion is reduced, but complex control logic is required
Solution Approach 1:
The patent reduces control complexity by performing all necessary determination of attribute distribution targets in advance, before the actual attribute distribution occurs. The system pre-calculates which processing units will need which attributes based on primitive assignment, then executes the distribution according to this predetermined plan, simplifying the control logic during the actual distribution phase.
Solution Approach 2:
The patent segments the attribute distribution process into distinct phases: determination of primitive-to-processing-unit assignment, identification of required attributes, and actual attribute distribution. This segmentation allows each phase to be handled independently with specialized logic, reducing overall system complexity while maintaining selective distribution benefits.
Data Source
AI summary
A graphics processing unit has a set of parallel processing units. A primitive pipeline delivers tiles of a primitive to selected processing units of the set of processing units. An attribute pipeline distributes attributes to the selected processing units when the end of the primitive is reached, while withholding attributes from the remaining processing units of the set of processing units.