[Performance] ThresholdedRelu operator regression 1.20.0 → 1.21.0

[junghyunpark2001]

Describe the issue

Summary

ThresholdedRelu shows regression on 4D tensors with high channel count.

Test Results

Test Case	Input Shape	Kernel Change
v21_default_alpha_float32_4d	(2,64,28,28)	+19.66%
v13_float32	(2,3,32,32)	+4.40%
v21_custom_alpha_float32_2d	(8,128)	-1.93%
v13_edge_case	(1,2,3)	-1.89%
v21_scalar_edge	(1,1)	+4.15%

Regression Pattern

Regressed (float32):

• 4D tensor + high channel count (64ch)

Stable (float32):

• 4D tensor + low channel count (3ch)
• 2D tensor
• Scalar/small tensors

To reproduce

python script_profiling.py thresholdedrelu_thresholdedrelu_21_thresholdedrelu_default_alpha_float32_4d 1.20.0 1.21.0

Archive.zip

Urgency

No response

Platform

Linux

OS Version

Ubuntu 24.04.3 LTS

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.21

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

Model File

No response

Is this a quantized model?

Yes