[Performance] Performance regression in Clip operator for int64 inputs between v1.20.0 and v1.21.0

[junghyunpark2001]

Describe the issue

Description

We observed a performance regression in the Clip operator for int64 data type with only minimum bound specified (no maximum bound) between ONNXRuntime v1.20.0 and v1.21.0.

Affected Operator

Clip

• Opset Version: 13
• Data Type: int64
• Configuration: min bound only (no max bound)
• Input Shape: [2, 3, 32, 32] (4D tensor)
• Regression: Performance degradation observed

Test Case Details

Test Case: clip_13_v2_clip_int64_no_max

Input:

• X (input tensor):

  • Shape: [2, 3, 32, 32] (4D tensor)
  • Data type: int64
  • Total elements: 6,144

• min (minimum bound):

  • Shape: [] (scalar)
  • Data type: int64
  • Value: -100

• max (maximum bound):

  • Not provided (unbounded on upper end)

Output:

• Name: output
• Shape: [2, 3, 32, 32]
• Data type: int64

Operation:
Clips input values to be >= -100 (no upper bound).

Performance:

• v1.20.0: baseline
• v1.21.0: regression observed

Regression Characteristics

• Data type specific: int64 input and bounds
• Partial bounds: Only minimum bound specified (no maximum)
• Moderate tensor size: 6,144 elements (2×3×32×32)
• 4D input tensor: Standard tensor shape for vision-related operations

To reproduce

1. Download zip

2. Run benchmark using the provided script:

python profile_operator.py clip_13_v2_clip_int64_no_max 1.20.0 1.21.0

[Archive.zip](https://github.com/user-attachments/files/24695593/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

1.21.0

### 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

[microsoft-github-policy-service]

Applying stale label due to no activity in 30 days. Please react to the issue by assigning it or labeling it as 'contributions welcome', 'documentation', 'feature request', 'regression', or 'no stale'.