[Performance] Performance regression in Cast operator for float32 to double conversion between v1.18.0 and v1.19.0

[junghyunpark2001]

Describe the issue

Description

We observed a performance regression in the Cast operator when converting float32 to double (float64) between ONNXRuntime v1.18.0 and v1.19.0.

Affected Operator

Cast

• Opset Version: 21
• Source Type: float32
• Target Type: double (float64)
• Attribute: to=11 (DOUBLE), saturate=1
• Regression: +10.4% kernel slowdown

Test Case Details

Test Case: cast_cast_21_cast_float32_to_double

Inputs:

• input tensor:
  • Data type: float32 (type=1)
  • Shape: [4, 64, 100] (25,600 elements)

Attributes:

• to: 11 (DOUBLE)
• saturate: 1

Output:

• Data type: double (float64)
• Shape: [4, 64, 100]

Performance:

• v1.18.0: 0.0055 ms (kernel time)
• v1.19.0: 0.0060 ms (kernel time)
• Kernel regression: +10.4% slowdown
• Confirmation: 5/10 validation runs confirmed

Regression Characteristics

Affected Configuration

• Source type: float32
• Target type: double (float64)
• Tensor size: Medium (25K elements)

Key Characteristics

• Type conversion specific: float32 → double
• Opset version: 21
• Saturate attribute: Enabled (saturate=1)

To reproduce

python script_profiling.py  cast_cast_21_cast_float32_to_double 1.18.0 1.19.0
``

[Archive.zip](https://github.com/user-attachments/files/24907619/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

1.19.0

### 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