Skip to content

README.md

Latest commit

 

History

History
134 lines (91 loc) · 4.79 KB

README.md

File metadata and controls

134 lines (91 loc) · 4.79 KB

Machine Learning-based Multi-Input Switching (MIS) Support in OpenSTA

Overview

This project extends OpenSTA to support Multi-Input Switching (MIS) timing analysis using machine learning.

By default, OpenSTA only supports Single Input Switching (SIS). This modification enables dynamic delay annotation for MIS conditions by hooking a trained ML model into the STA flow — no core rebuilding needed for adding or updating models.

sta -version = 2.6.0

🛠️ Build Instructions

No changes to the build system are required. This version builds exactly like vanilla OpenSTA. Please refer to the OpenSTA build instructions for details.

📦 Dependencies

All dependencies remain the same as the original OpenSTA project.

🔧 Configuration

A new configuration file models_config.json is introduced to define the ML models used for MIS delay prediction.

Example:

{
  "lazyLoad": true,
  "models": [
    {
      "name": "nand2x1_sc",
      "path": "nand2x1_model.json",
      "modify_annotation": true,
      "inputFormat": ["load", "slew_a", "slew_b", "skew_ab"],
      "outputFormat": ["rise_delay", "rise_slew"],
      "scaleInput": [
        7e-16, 4.65e-14,
        5e-12, 3.2e-10,
        5e-12, 3.2e-10,
        -1e-9, 1e-9
      ],
      "scaleOutput": [
        7.59e-13, 2.96e-10,
        4.56e-12, 5.34e-10
      ]
    }
  ]
}

Key Points:

  • modify_annotation: true enables delay override for that gate.
  • Input/output formats and scaling ranges align with your ML model's expectations.
  • You can add models for additional gates by simply modifying the JSON — no recompilation needed.

🧩 Source Code Changes

The following new classes are introduced under the modification/ directory:

  • StaInterface: Interface layer between STA engine and ML models.
  • MlModel: Handles loading and querying the model JSONs.
  • DataToModel: Extracts input features from OpenSTA data.

Additionally, a new public method is added to the Graph class:

void Graph::changeArrivals(Vertex* vertex, Arrival* new_arrivals);

This enables controlled modification of arrival times for specific gates during analysis.

Integration Point:

The hook for model-based annotation is invoked from:

ArrivalVisitor::visit()  // Located in search/Search.cc

Approach

Assumption: the ML model modifies the annotation (arrival) of the fanin pins (input pins) of each gate to capture MIS effects on the output pin. We integrate this within the SIS flow of OpenSTA by hijacking the fanin arrival before computing the output arrival.

  1. ArrivalVisitor Hook
    In search/Search.cc, the ArrivalVisitor::visit method drives the STA graph traversal. We insert our modification just before OpenSTA computes the output-pin arrivals by calling hackModelUpdate in modification/StaInterface.cc.

  2. Identify Output Pin
    In StaInterface::hackModelUpdate(Vertex*), we check if the current vertex corresponds to an output pin (direction->isOutput()) and has fanout. If so, and an ML model is available for its cell, we proceed.

  3. Gather Fanin Data
    updateAnnotation_fanin_from_fanin(DataToModel*) builds a DataToModel object containing:

    • The output vertex Zn
    • Fanin vertices A and B
    • Original arrival arrays (getAnnotationArray)
    • Slew information (getSlew)
    • Load capacitance on the output pin (getLoadCapacitance)

  4. ML Inference
    MlModel::Modify(DataToModel*) takes the original fanin arrivals and slews, constructs two input vectors (one for A→Y, one for B→Y, swapping A/B in the “slew” and “skew” features), invokes the FDeep model, and decodes the scaled outputs back to physical arrival/slew values.

  5. Apply Modified Arrivals
    Back in updateAnnotation_fanin_from_fanin, we write the modified arrival arrays into the graph via a new public Graph::changeArrivals helper (a friend of Vertex), updating A and B’s arrival annotations.

  6. Resume SIS Flow
    With fanin arrivals modified, OpenSTA’s existing combinational traversal computes the output-pin arrivals (for Y) including the ML-captured MIS effect.

Assumptions

  • ML-model configuration is per-cell (per-model) in models_config.json.
  • If a model is missing or modify_annotation is false, we leave original SIS arrivals untouched.
  • modifying the Fanout path cornologicaly from start

for theoratical Approach look into app/README.md

MIS Analysis Plots

Delay

Delay Plot

Slack

Slack Plot

we can see the MIS is removing Pessimistisum from the SIS model. we could get more complex MIS bheaviour using other models and adding them to app/models_config.json