[Performance] DropQDQ optimization no longer drops QDQ around MaxPool for opset >= 22

[mcollinswisc]

Describe the issue

QDQ MaxPool, i.e. the sequence:

DequantizeLinear -> MaxPool -> QuantizeLinear

no longer drops the DequantizeLinear and QuantizeLinear in optimization, as of opset 22. It did previously for opset 12 (to 21).

To reproduce

Run attached script:

qdq_maxpool.py

(Docstring has commands)

Urgency

Prevents us from moving to opset 22+ because models that previously got optimized to full integer aren't anymore.

Platform

Linux

OS Version

Ubuntu Questing

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.26.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

Model File

Script attached previously will build an example model

Is this a quantized model?

Yes