NamedOutput mechanism clarification

[anssiko]

10. I don't quite understand the NamedOutput mechanism. What if what is output just a feature?

via w3ctag/design-reviews#570 (comment)

@cynthia "What if what is output just a feature?" part needs a clarification?

[cynthia]

Yeah, sorry. "What if the output is just a feature?" is what I intended to write. (Not great at speaking and writing at the same time.)

[huningxin]

Thanks for the clarification @cynthia !

I agree the NamedOutputs needs more details in the spec. We'll fix that.

Basically, the NamedOutputs is just used to index an Output by string-based name.

For the case that the output is just a feature, I suppose the feature can be represented by an Operand, then the developer might first need to bind a name to the operand when creating a model, e.g. the following sample code binds name "C" to operand C.

// 1. Create a model of the computational graph 'C = 0.2 * A + B'.
const constant = builder.constant(0.2);
const A = builder.input('A', operandType);
const B = builder.input('B', operandType);
const C = builder.add(builder.mul(A, constant), B);
const model = builder.createModel({'C': C});

So after computing, the developer can use name "C" to index the output, e.g. by following code

const outputs = await compilation.compute(inputs);
// The computed result of [[1, 1], [1, 1]] is in the buffer associated with
// the output operand.
console.log('Output shape: ' + outputs.C.dimensions);
console.log('Output value: ' + outputs.C.buffer);

You can try above code in WebNN code editor with webnn-polyfill.

Other usages of NamedOutputs include:

1. support the pre-allocated buffer for compute, this might be useful for Wasm integration, see code example
2. support optional output by specifying the output name when invoking compute, see code example

[huningxin]

As discussed in 18 Feb CG call, I'll craft a PR to add specific text into the spec and ask for review.