[Performance] Performance regression in Sign operator (int64) between v1.20.0 and v1.21.0 (≈23-26% slowdown) #27115
Description
Describe the issue
Description
We observed a significant performance regression in the Sign operator for int64 data type between ONNXRuntime v1.20.0 and v1.21.0. This regression affects all int64 Sign operations tested, regardless of input data characteristics.
Affected Operator
Sign
- Opset Version: 13
- Data Type: int64
- Regression: +22.58% to +25.79% kernel time slowdown (approximately 1.2-1.3x slower)
- Status: Confirmed regression, persists to v1.23.0 (latest)
Data Type Specificity
Important Note: This regression appears to be specific to int64 data type:
- No similar regressions reported for float32/float64 Sign operations
- Sign is typically a simple element-wise operation that should perform similarly across data types
To reproduce
- Download zip file
- Run benchmark using the provided script:
python script_profiling.py sign_13_v2_sign_int64_3d_boundary_values 1.20.0 1.21.0
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