[Performance] ScatterElements operator regression 1.20.0 → 1.21.0 #27139
Description
Describe the issue
Summary
ScatterElements shows significant regression on large tensors across all data types.
Test Results
| Test Case | Input Shape | Type | Reduction | Kernel Change |
|---|---|---|---|---|
v2_axis0_none |
(2,64,56,56) | float32 | none | +21.51% |
v2_axis1_add |
(2,64,56,56) | int32 | add | +25.94% |
v2_axis_neg2_mul |
(2,64,56,56) | int64 | mul | +32.74% |
v2_axis2_max |
(2,64,56,56) | float32 | max | +15.80% |
axis0_basic |
(2,4) | float32 | none | +2.82% |
axis0_add_overlap |
(3,3) | float32 | add | +5.60% |
Regression Pattern
Regressed (all types):
- Large 4D tensors (2,64,56,56)
- All reduction modes affected (none, add, mul, max)
- int64 worst affected (+32.74%)
Stable:
- Small 2D tensors
To reproduce
cd evaluation/report
python script_profiling.py scatterelements_18_v2_scatterelements_float32_axis0_none 1.20.0 1.21.0
python script_profiling.py scatterelements_18_v2_scatterelements_int64_axis_neg2_mul_reduction 1.20.0 1.21.0Urgency
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