[Performance] ORT-WebGPU Average Pooling is working too long in edge case

[grazder]

Describe the issue

I'm using ORT with WebGPU for inference. I did profiling and found that 4 AveragePoolings takes almost half of model inference speed.

Here is example of operation:

AvgPool inference time has linear dependence on kernel size

Shapes // Time

• (20, 32) = 640 // 0.2ms;
• (40, 64) = 2560 (x4) // 0.8ms (x4)
• (80, 128) = 10240 (x16) // 3.2ms (x16)

Problem here is that it is an edge case, because it is literally ReduceMean on dims=(2,3). I changed the operation and got giant speed boost.

I think that it is an edge case (that AvgPool == ReduceMean) and you can create a special handler for it.

To reproduce

Just create AvgPool with big kernel_size

Urgency

Not urgent, workaround was found

Platform

Mac

OS Version

15.2

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.20.1

ONNX Runtime API

JavaScript

Architecture

ARM64

Execution Provider

Other / Unknown

Execution Provider Library Version

No response

Model File

No response

Is this a quantized model?

No

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.