Revise graph resource validation

[inexorabletash]

compute() calls validate graph resources with the passed input/output maps and the graph's input/output descriptors. The intent is to ensure that all of a graph's descriptors are matched to a passed resource. However, the steps were only validating that a passed resource had a corresponding descriptor - it wouldn't flag if a resource was missing!

The Chromium prototype implementation made the test reflexive - all resources much have matching descriptors and vice versa. After deliberation we settled on a more helpful behavior: all inputs must be provided, but extra inputs are ignored. A requested output must be present, but dropping outputs is allowed.

Fixes #602

---

Preview | Diff

[inexorabletash]

@huningxin and @fdwr - can you please review? Thanks!

[fdwr]

Before reviewing, I'm just thinking aloud the various cases, given a graph that expects 2 inputs and generates 2 outputs...

• await context.compute(graph, {input1: inbuf1, input2: inbuf2}, {output1: outbuf1, output2: outbuf2}); ✅ Exactly matching. All required inputs satisfied. All generated outputs bound.
• await context.compute(graph, {input1: inbuf1}, {output1: outbuf1, output2: outbuf2}); ❌ Missing a required input.
• await context.compute(graph, {input1: inbuf1, input2: inbuf2, input3: inbuf3}, {output1: outbuf1, output2: outbuf2}); 🤷‍♂️✔ Extra input specified. I foresee useful cases to pass potentially ignorable input so an app can more generically provide inputs regardless of the model flavor. Say one diffusion model doesn't use a mask input, whereas another inpainting model does use the mask, and just having the compute ignore that ignore that parameter for the non-inpainting model is cleaner for the caller than inspecting which ones are actually present and branching between them, especially when you have many different permutations of potential inputs.
• await context.compute(graph, {input1: inbuf1, input2: inbuf2}, {output1: outbuf1}); 🤷‍♂️✔ An output was generated by the graph was ignored and discarded. This seems fine to me too. The caller doesn't always care about all outputs, and needing to create a dummy variable just to discard that dummy seems pointless.
• await context.compute(graph, {input1: inbuf1, input2: inbuf2}, {output1: outbuf1, output2: outbuf2, output3: outbuf3}); 🤷‍♂️❌ If you require an output, and the graph doesn't provide it, that's an error. Though, by my same logic above for optionally unused inputs, there can be a set of closely related model flavors where some provide an output and others do not, and the output in that case would be missing from the return results.outputs. So maybe this is 🤷‍♂️✔ too, where the caller has to check the dictionary. I suppose there isn't an IDL dictionary syntax to express {requiredOutput1: outbuf1, optionalOutput2: outbuf2}. 🤔...

Anyway, we definitely both agree (per https://github.com/webmachinelearning/webnn/issues/602) on the case where it "Should fail, because input2 is missing - but it doesn't" ❌.

[huningxin]

• await context.compute(graph, {input1: inbuf1, input2: inbuf2}, {output1: outbuf1}); 🤷‍♂️✔ An output was generated by the graph was ignored and discarded.

Computing a subset of outputs is supported. There was an example using sync compute API https://www.w3.org/TR/2024/CRD-webnn-20240214/#example-4a21156c. Unfortunately, my PR https://github.com/webmachinelearning/webnn/pull/548 that removes the sync APIs deleted this example together. We may want to bring it back by adapting to async compute API.

[inexorabletash]

there isn't an IDL dictionary syntax to express {requiredOutput1: outbuf1, optionalOutput2: outbuf2}.

Correct.

Based on above ✔🤷‍♂️❌ it sounds like we might want:

• Inputs:
  • every input descriptor must have matching input resource: can't run inference with missing inputs!
  • an input resource without a matching input descriptor is okay: some graphs don't use it
• Outputs:
  • every output resource must have a matching output descriptor: asking for an output that's not provided by the inference is a bug!
  • an output descriptor without a matching output resource is okay: if it's not needed, avoid extra work

Given this we'd want either one algorithm that takes a flag (is input or output?) or two algorithms. The latter is probably easier. And since they're only used in one place this could be in-line in "execute graph", something like this:

5. For each name → descriptor of graph.[[inputDescriptors]]:

   1. If inputs[name] does not exist, then return a new promise rejected with a TypeError.
   2. If validating buffer with descriptor given inputs[name] and descriptor returns false, then return a new promise rejected with a TypeError.

6. For each name → resource of outputs:

   1. If graph.[[outputDescriptors]][name] does not exist, then return a new promise rejected with a TypeError.
   2. If validating buffer with descriptor given resource and graph.[[outputDescriptors]][name] returns false, then return a new promise rejected with a TypeError.

... and then a non-normative note saying something like:

NOTE: Invocations of compute() will fail if any of the graph's inputs are not provided as input resources, or if any requested output resources do not match the graph's outputs.

[huningxin]

The latter is probably easier. And since they're only used in one place this could be in-line in "execute graph", something like this:

SGTM!

[inexorabletash]

1e7e2b988cc8141c347510012d97ccd7b7c9139b captures the algorithm discussed above. We should probably wait for the WG telecon to decide if this is what we want.

[inexorabletash]

I updated this PR's name/description to match the changes. I think we're probably good to merge now?

[huningxin]

@fdwr , do you have any further comments?

[fdwr]

... or two algorithms. The latter is probably easier. And since they're only used in one place this could be in-line in "execute graph", something like this ... I updated this PR's name/description to match the changes. I think we're probably good to merge now?

Yeah, I like it. Thanks for noticing and fixing.

[fdwr]

Minor merge conflict from the transferred inputs/outputs CR. So just resolved.