Confidence-Based RAW Hazard Predictor for Load Blocking

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional RAW predictors often cause performance losses by blocking non-critical loads unnecessarily, even when no significant RAW resynchronization exceptions (RREs) would occur, due to overtraining on initial RREs in repetitive instruction code loops.

Innovation Solution

The Confidence/Sampling Read-After-Write Resynchronization Predictor (CSRRP) uses a confidence score and dynamic sampling to determine whether to treat an instruction as a resynchronization predicted store (RPS), avoiding unnecessary blocking of non-critical loads by periodically updating and deallocating entries based on execution outcomes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a conventional RRP is used to avoid RRE penalties, then RRE penalties are reduced, but non-critical loads are blocked unnecessarily causing performance loss

Engineering Contradiction:

ImproveRRE penalty avoidanceVSAvoidnon-critical load execution

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent applies local quality by differentiating between critical loads (those that would cause RRE if executed ahead of RPS) and non-critical loads (NCL). The system selectively blocks only critical loads while allowing NCL to execute freely, thereby avoiding unnecessary performance loss from blocking non-critical operations while maintaining protection against RRE penalties.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent implements dynamics by making the RRP entries dynamically valid or invalid based on confidence scores. Entries are marked invalid when confidence drops below a threshold, allowing the system to adapt its prediction behavior over time. This dynamic adjustment prevents the system from persistently blocking loads based on outdated or inaccurate predictions, thus reducing unnecessary performance loss.

Inventive Principle:

Principle #15Dynamics

2Reliability

If RRP trains on initial RREs in repetitive instruction code loops, then RRE penalties are avoided in subsequent iterations, but performance is lost due to unnecessary blocking of NCL

Engineering Contradiction:

ImproveRRE prediction accuracyVSAvoidinstruction execution throughput

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent applies parameter changes by monitoring confidence scores associated with RRP entries and adjusting their validity state accordingly. When confidence drops below a threshold (indicating the prediction may be incorrect due to repetitive code patterns), the entry is marked invalid, allowing the system to change its predictive behavior parameter and stop blocking loads unnecessarily.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent implements feedback by continuously monitoring whether RRE actually occurs when an RPS is executed. If no RRE occurs despite the RPS prediction, the system updates the confidence score downward. This feedback mechanism allows the system to learn from actual execution outcomes and adjust its predictions, preventing unnecessary blocking in repetitive code scenarios where RRE does not actually occur.

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS9286073B2Read-after-write hazard predictor employing confidence and sampling

Publication Date: 2016.03.15 SAMSUNG ELECTRONICS CO LTD

AI summary

Dynamically predicting a Read-After-Write (RAW) hazard by employing a variable confidence score attributed to a RAW Resynchronization Predictor (RRP) for sampling the RRP at timing periods dynamically adjusted based on the confidence score to optimize prediction of the RAW hazard.