[MLBuffer] Support for MLBuffer in graph execution

[bbernhar]

Purpose/Motivation

Provides a means to execute a MLGraph using MLBuffer. This is a sub-issue of https://github.com/webmachinelearning/webnn/issues/482.

Proposed API

dictionary MLBufferView {
  required MLBuffer buffer;
  MLSize64 offset = 0;
  MLSize64 size;
};

typedef record<DOMString, MLBufferView> MLNamedMLBufferViews;

[Exposed=(Window, DedicatedWorker), SecureContext]
partial interface MLContext {
  undefined dispatch(
        MLGraph graph, MLNamedMLBufferViews inputs, MLNamedMLBufferViews outputs);
}

Example JS

const bufferA = new MLBuffer({size:4});
const bufferB = new MLBuffer({size:4});
const inputs = {'A': {bufferA, 0, bufferA.size()};
const outputs = {'B': {bufferB, 0, bufferB.size()};
context.dispatch(graph, inputs, outputs);

• Enqueues a request to compute the graph onto some WebNN timeline
• Execution cannot start until all input and output MLBuffers are available
• All input and output MLBuffers are unavailable while execution is in progress
• All work submitted after this dispatch() call which relies on an input or output MLBuffer will be queued behind this execution

Alternative API proposals

N/A

Opens

1. Should this method be on the MLGraph (related to https://github.com/webmachinelearning/webnn/issues/303)? @a-sully
2. Is it valid to pass the same MLBuffer as both an input and output of the same dispatch() call? @a-sully
3. If the approach is flexible enough to allow for graph execution on all backends. Do we need a separate compute() method? @a-sully
4. Can dispatch be exclusive to MLBuffer bindings? @bbernhar

[bbernhar]

@anssiko FYI, this issue should be tagged as non-interop

[inexorabletash]

I was interpreting https://github.com/webmachinelearning/webnn/labels/webgpu%20interop to apply to anything easing interop between WebNN and WebGPU, which would implicitly cover anything related to MLBuffer.

Is there another interpretation we should be using? And whatever we decide, can @anssiko apply it to the label so it shows at https://github.com/webmachinelearning/webnn/labels ? (I don't have permission to do so.)

[bbernhar]

Is there another interpretation we should be using?

Good question. Early on, MLBuffer was just for interop. But now MLBuffer will be used without interop too [no WebGPU] like any other WebNN primitive (ex. MLGraph). Since we have a dedicated issue for the interop part now, I don't think the webgpu interop label is as useful.

[inexorabletash]

Thanks @bbernhar that helps! I'll drop the label, we can always re-add it later.