[Performance] Performance regression in ELU operator between v1.20.0 and v1.21.0 (≈220% slowdown, suspected Eigen update)

[junghyunpark2001]

Describe the issue

Title

Performance regression in ELU operator between v1.20.0 and v1.21.0 (≈220% slowdown, suspected Eigen update)

---

Description

We observed a significant performance regression in the ELU operator between onnxruntime v1.20.0 and v1.21.0.

---

Operator / Test Case Details

Operator

• Type: Elu
• Opset Version: 21

Input

• Name: X
• Shape: [2, 64, 28, 28] (4D tensor)
• Data type: float32
• Value range:
  • Min: 0.100
  • Max: 9.999
  • Mean: 5.040

Output

• Name: output
• Shape: [2, 64, 28, 28] (same as input)
• Data type: float32

Attributes

• alpha: 1.0 (default value per ONNX specification)
• No additional attributes specified

Model Information

• IR Version: 10
• Opset Version: 21

---

Regression Magnitude

• Approximately 220% slowdown in v1.21.0 compared to v1.20.0

---

Reproducibility

• Reproducible across multiple runs and different environments

---

Suspected Cause

Based on commit history analysis, the regression is likely related to the Eigen update introduced in the following commit:

• Suspected commit:

  7c0c6fb (Eigen update)

While the ELU kernel implementation itself does not appear to have been directly modified, the Eigen update may have affected:

• vectorization behavior,
• math kernel selection, or
• underlying execution paths for float32 ELU computation.

To reproduce

1. Download the attached ELU model and benchmark script.
2. Run the benchmark using the following command:

python script.py ./elu 1.20.0 1.21.0

• ./elu: directory containing the ELU model
• 1.20.0: baseline (good) onnxruntime version
• 1.21.0: regressed (bad) onnxruntime version

1. Compare the reported latency between the two versions.

   A significant slowdown (~220%) can be observed in v1.21.0.

reproduce.zip

Urgency

No response

Platform

Linux

OS Version

Ubuntu 24.04.3 LTS

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

7c0c6fb

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