GPU Combined Instructions with Dual Data Pipes for Shader Bottlenecks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional GPU architectures face performance bottlenecks due to data dependencies among instructions, leading to inefficient use of Common Register (CR) bandwidth and wasted SP capacity, particularly in shader operations.

Innovation Solution

Introduce a novel GPU architecture with two data paths: a main-pipe for Algorithm, Compare, Logic, and Selection operations, and a bypass-pipe for data or constant reading and passing, allowing for combined instructions to be executed in parallel.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If multiple instructions are executed using multiple data pipes in parallel, then processing speed is improved, but instruction complexity and compilation difficulty increase

Engineering Contradiction:
Improveprocessing speedVSAvoidinstruction complexity
Core Design Contradiction:
SpeedVSDevice complexity

Solution Approach 1:

The patent segments instructions into different types (calculation, comparison, logic, selection, branching, LD/ST, SMP, and complicated mathematics operations) and assigns them to different data pipes based on their specific requirements. This segmentation allows each data pipe to be optimized for its designated instruction type, improving parallel processing speed while maintaining manageable instruction complexity through systematic classification.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a compiler as an intermediary that automatically analyzes data dependencies between instructions and combines compatible instructions into combined instructions. This intermediary handles the complexity of parallel execution planning, allowing multiple instructions to be executed in parallel through the combined instruction mechanism without requiring manual intervention to manage the complexity.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If instructions are combined based on data dependencies, then execution efficiency is improved, but compilation complexity increases

Engineering Contradiction:
Improveexecution efficiencyVSAvoidcompilation complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent implements a self-service mechanism where the compiler automatically performs dependency analysis and instruction combination without external intervention. The system analyzes data dependencies between instructions, identifies compatible instruction pairs, and combines them into unified combined instructions that can be executed efficiently in parallel, thereby improving execution efficiency while managing compilation complexity through automation.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent changes the parameter of instruction representation by combining multiple instructions into a single combined instruction structure. This parameter change allows the system to represent complex parallel execution plans in a simplified form that maintains execution efficiency while reducing the apparent compilation complexity, as the combined instructions encapsulate the dependency relationships internally.

Inventive Principle:
Principle #35Parameter changes

3Device complexity

If different data types are processed using the same data pipe, then device simplicity is maintained, but processing precision may be compromised

Engineering Contradiction:
Improvedevice simplicityVSAvoidprocessing precision
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by assigning specific data types to specific data pipes based on their processing requirements. Different data pipes are optimized for different data types (e.g., floating-point operations in one pipe, integer operations in another), ensuring that each data type receives the appropriate processing precision while maintaining overall device simplicity through a systematic division of labor rather than requiring a single complex processing unit.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4258110B1Methods for combining instructions and apparatuses having multiple data pipes
Publication Date: 2026.05.13 GLENFLY TECH CO LTD
  • EP4258110B1 patent drawingFigure 1
  • EP4258110B1 patent drawingFigure 2
  • EP4258110B1 patent drawingFigure 3A

AI summary

A method for combining instructions, performed by a compiler, containing at least the following steps. First instructions are obtained, where each performs one of a calculation operation, a comparison operation, a logic operation, a selection operation, a branching operation, a LD/ST (Load/Store) operation, a SMP (sampling) operation and a complicated mathematics operation. The first instructions are combined as one combined instruction according to data dependencies between the first instructions. The combined instruction is sent to a SP (Stream Processor).