Selective Dummy Thread Instruction Omission in Data Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Data processing systems face inefficiencies in handling thread groups due to the presence of dummy threads required to resolve inter-thread dependencies, which consume significant resources and increase power consumption without contributing to the final result data.

Innovation Solution

A data processing apparatus that generates thread groups with identification of dummy threads, using execution flow modification circuitry to selectively omit instructions for dummy threads based on control information associated with the program, reducing resource consumption and improving performance without additional execution resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If dummy threads are executed to resolve inter-thread dependencies, then the correctness of result data is ensured, but resource consumption and power consumption increase significantly

Engineering Contradiction:

Improvecorrectness of result dataVSAvoidpower consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts and identifies dummy threads from thread groups using analysis circuitry. Once identified, these dummy threads are separated from the execution flow through selective omission of instructions, allowing the system to maintain correctness for active threads while eliminating unnecessary resource consumption from dummy threads.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different execution treatments to different threads within the same thread group based on their quality/type. Active threads receive full instruction execution while dummy threads have instructions selectively omitted. This local differentiation allows the system to optimize power consumption without compromising the correctness of active thread execution.

Inventive Principle:

Principle #3Local quality

2Reliability

If dummy threads are executed to resolve inter-thread dependencies, then the correctness of result data is ensured, but data processing performance decreases

Engineering Contradiction:

Improvecorrectness of result dataVSAvoiddata processing performance

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

Inventive Principle:

Principle #3Local quality

3Reliability

If all instructions are executed for each thread in a thread group, then complete processing is ensured, but resource usage increases for threads that do not contribute to result data

Engineering Contradiction:

Improvecompleteness of processingVSAvoidresource usage

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

Inventive Principle:

Principle #3Local quality

4Productivity

If multiple execution pipelines are used to process threads in parallel, then throughput is improved, but power consumption increases

Engineering Contradiction:

ImprovethroughputVSAvoidpower consumption

Core Design Contradiction:

ProductivityVSUse of energy by moving object

Solution Approach 1:

The patent applies partial action by executing only the necessary subset of instructions for dummy threads rather than complete execution. The execution flow modification circuitry selectively omits instructions for identified dummy threads, performing just enough processing to resolve inter-thread dependencies without the full overhead of complete thread execution, thereby reducing power consumption while maintaining throughput.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS8978038B2Data processing apparatus and method for processing a received workload in order to generate result data

Publication Date: 2015.03.10 ARM LTD

AI summary

A thread group generator generates from a received workload a plurality of thread groups. Each thread group consists of a plurality of threads, and at least one thread group has an interthread dependency existing between the plurality of threads. Each thread may be either an active thread whose output is required to form the result data, or a dummy thread required to resolve the inter-thread dependency for one of the active threads but whose output is not required to form the result data. A thread execution unit then executes each thread within a thread group received from the generator by executing a predetermined program. Execution flow modification circuitry is responsive to the received thread group having at least one dummy thread, to cause the unit to selectively omit at least part of the execution of at least one of the plurality of instructions when executing each dummy thread.