[Feature Request] Restore XNNPACK Execution Provider for ONNX Runtime Web Backend

[fanchenkong1]

Describe the feature request

We would like to request the re-introduction of the XNNPACK Execution Provider (EP) for ONNX Runtime Web.

Motivation

We’ve benchmarked default Wasm and XNNPACK-Wasm execution providers for browser-based inference, and observed that XNNPACK EP outperforms the current WebAssembly-based EP for Mobilenetv3-like models, on various laptop devices.

Thread1

Device	Execution Time Reduction by Switching to XNNPACK
TGL / i7-1165G7	54%
MTL/Ultra7 155H	76%
LNL/Ultra7 268V	79%

In our benchmarks, XNNPACK EP achieves over 50% lower latency than the MLAS Wasm EP on all laptop devices.

We have a local build for adding xnnpack ep back on the latest onnxruntime and are willing to contribute. Would this change be welcomed?

Describe scenario use case

Browser-based fp32 inference could benefit from the XNNPack ep.

[fanchenkong1]

By the way, it seems that XNNPACK ep has better support for quantized models in QDQ format. We have created a PR for updating the xnnpack.cmake file based on XNNPACK version used by latest onnxruntime (#24860).

@fs-eire, Do you have any comments on this issue?

[huningxin]

also /cc @guschmue

[guschmue]

ok, if it is that much we agree that it might be worth turning re-adding xnnpack support.
Last time I measured it I belief I saw 7%.
Going to capture some numbers across a 90 model test.

[fanchenkong1]

Thanks! We are looking forward to the results of the 90 model test.