[Performance] Performance regression in Einsum operator for int64 data type between v1.22.0 and v1.23.0

[junghyunpark2001]

Describe the issue

Description

We observed a performance regression in the Einsum operator when using int64 data type inputs between ONNXRuntime v1.22.0 and v1.23.0. This regression is specific to int64 type - other data types (e.g., int32) are not affected.

Affected Operator

Einsum

• Opset Version: 12
• Data Type: int64 (regressed)
• Regression: +14% kernel slowdown

Test Case Details

Test Case: einsum_12_v2_einsum_elementwise_multiplication_broadcast_int64

Inputs:

• input_0 tensor:

  • Data type: int64 (type=7)
  • Shape: [2, 64, 56, 56]

• input_1 tensor:

  • Data type: int64 (type=7)
  • Shape: [64]

Attributes:

• equation: "nchw,c->nchw" (elementwise multiplication with broadcast)

Output:

• Data type: int64
• Shape: [2, 64, 56, 56]
• Elementwise multiplication with channel broadcast

Performance:

• v1.22.0: 26.62 ms (kernel time)
• v1.23.0: 30.45 ms (kernel time)
• Kernel regression: +14.4% slowdown
• Total time regression: +23.9% slowdown

To reproduce

1. Download zip file

Archive.zip

2. Run benchmark using the provided script:

   python script_profiling.py einsum_12_v2_einsum_elementwise_multiplication_broadcast_int64 1.22.0 1.23.0

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