use WebGPU EP instead of JSEP in WebAssembly

[fs-eire]

Description

This PR allows to use WebGPU EP in onnxruntime-web NPM package.

Migration Plan

Currently, there are 2 different EPs implemented the WebGPU backend of onnxruntime-web. They are JSEP and WebGPU EP. The migration plan is to replace the JSEP with WebGPU EP and eventually remove the JSEP. The plan contains the following stages:

• STAGE 1: enable WebGPU EP on onnxruntime-web in local build. (Done)
• STAGE 2: enable WebGPU EP on onnxruntime-web in the public package. (This PR)
• STAGE 3: remove JSEP from onnxruntime-web.

Package consumption changes

• Default import (import 'onnxruntime-web') and CPU only import (import 'onnxruntime-web/wasm') keeps their previous behaviors.

• WebGPU import (import 'onnxruntime-web/webgpu') will now use WebGPU EP instead of JSEP. Previously it was the same as default import.

  • WebGPU import will use a different suffix for the .mjs and .wasm name (which was .jsep):

    • ort-wasm-simd-threaded.asyncify.mjs
    • ort-wasm-simd-threaded.asyncify.wasm

  • The suffix .asyncify is used as .jspi is planned for future. They are 2 different ways of emscripten's async implementation (sync C++ function calls async JS function)

[fs-eire]

TODO: resolve conflicts after #25022 is merged.

[guschmue]

both jsep and webgpu-ep builds for for me.

[fs-eire]

@Honry could you please help to test/validate if this change breaks WebNN?

This change is expected to have no behavior change for WebNN. Specifically:

• importing from "onnxruntime-web" (import * as ort from 'onnxruntime-web';) should be using the old JSEP + WebNN EP.
• importing from "onnxruntime-web/webgpu" (import * as ort from 'onnxruntime-web/webgpu';) should be using the new WebGPU EP + WebNN EP

Both should work for WebNN

[Honry]

@Honry could you please help to test/validate if this change breaks WebNN?

This change is expected to have no behavior change for WebNN. Specifically:

• importing from "onnxruntime-web" (import * as ort from 'onnxruntime-web';) should be using the old JSEP + WebNN EP.
• importing from "onnxruntime-web/webgpu" (import * as ort from 'onnxruntime-web/webgpu';) should be using the new WebGPU EP + WebNN EP

Both should work for WebNN

Sure, I will verify it with some demos.

[Honry]

@Honry could you please help to test/validate if this change breaks WebNN?
This change is expected to have no behavior change for WebNN. Specifically:

• importing from "onnxruntime-web" (import * as ort from 'onnxruntime-web';) should be using the old JSEP + WebNN EP.
• importing from "onnxruntime-web/webgpu" (import * as ort from 'onnxruntime-web/webgpu';) should be using the new WebGPU EP + WebNN EP

Both should work for WebNN

Sure, I will verify it with some demos.

@fs-eire, both work! BTW, will we have a build specific for WebNN? e.g. ort.webnn.min.js, onnxruntime-web/webnn

[fs-eire]

@Honry could you please help to test/validate if this change breaks WebNN?
This change is expected to have no behavior change for WebNN. Specifically:

• importing from "onnxruntime-web" (import * as ort from 'onnxruntime-web';) should be using the old JSEP + WebNN EP.
• importing from "onnxruntime-web/webgpu" (import * as ort from 'onnxruntime-web/webgpu';) should be using the new WebGPU EP + WebNN EP

Both should work for WebNN

Sure, I will verify it with some demos.

@fs-eire, both work! BTW, will we have a build specific for WebNN? e.g. ort.webnn.min.js, onnxruntime-web/webnn

Will not. This is the plan:

currently both onnxruntime-web and onnxruntime-web/webgpu supports WebNN.
we keep both for testing and validation purpose.

once we think the time is mature, we will use WebGPU EP in onnxruntime-web and onnxruntime-web/webgpu will be deprecated. eventually everyone should use onnxruntime-web (WebGPU EP + WebNN EP) and there is going to be no JSEP.

[Honry]

@Honry could you please help to test/validate if this change breaks WebNN?
This change is expected to have no behavior change for WebNN. Specifically:

• importing from "onnxruntime-web" (import * as ort from 'onnxruntime-web';) should be using the old JSEP + WebNN EP.
• importing from "onnxruntime-web/webgpu" (import * as ort from 'onnxruntime-web/webgpu';) should be using the new WebGPU EP + WebNN EP

Both should work for WebNN

Sure, I will verify it with some demos.

@fs-eire, both work! BTW, will we have a build specific for WebNN? e.g. ort.webnn.min.js, onnxruntime-web/webnn

Will not. This is the plan:

currently both onnxruntime-web and onnxruntime-web/webgpu supports WebNN. we keep both for testing and validation purpose.

once we think the time is mature, we will use WebGPU EP in onnxruntime-web and onnxruntime-web/webgpu will be deprecated. eventually everyone should use onnxruntime-web (WebGPU EP + WebNN EP) and there is going to be no JSEP.

That make sense, thank you for the explanation!