Remove sequential filling overload of constant()

[a-sully]

Discussion on #492 (specifically, this comment #492 (comment)) concluded that this overload was not needed. Its behavior can be emulated by filling a buffer in JavaScript to be passed in with the buffer-based constant() overload

---

Preview | Diff

[fdwr]

Thanks Mr. Sully.

I originally proposed adding it, but model analysis (~700 models, 19 of which have the op) shows the:

• (a) tiny tensors being generated, not worth the beefy GPU effort.
• (b) increasing sequences are usually used for shape computation which can actually make it worse when the sequential fill is done on the GPU only to be copied back to the CPU for shape computation (meaning it would be better to keep that computation on the CPU to begin with). These cases should be ultimately optimized away by constant folding after shape inference too.
• (c) very few occurrences, typically once or twice (and in the cases with many more occurrences, 30+, those are all shape related).

Additionally this sidesteps any of the implementation difference issues we encountered between TF and PyTorch and validation ambiguities about infinities in the data - if the caller wants infinity in the buffer, they upload their buffer with infinity (and WebNN doesn't stop them).

This operator was not integrated yet into Chromium and is not used currently by known callers (e.g. here).

[fdwr]

👍

[fdwr]

@inexorabletash I can't find your comment about the spec missing the IDL definition, but removing this overload makes that earlier comment moot:

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLGraphBuilder {
  ...
  // Create an operand for a graph constant.
  MLOperand constant(MLOperandDescriptor descriptor, ArrayBufferView bufferView);

  // Create a single-value operand from the specified number of the specified type.
  MLOperand constant(double value, optional MLOperandDataType type = "float32");
  ...
};

https://www.w3.org/TR/webnn/#idl-index

[inexorabletash]

@inexorabletash I can't find your comment, but removal of this overload entirely make moot its current absence in the IDL definition:

Indeed. For posterity:

• I'd filed WebIDL definition for constant(start, end, step, type) is missing #571 but ended up duplicating that into Constant sequential filling operation needs output shape parameter #492 which this will handily resolve.
• Introduce MLNumber for specifying numeric inputs of any type #647 will be simplified too.

[huningxin]

LGTM, thanks!

[a-sully]

Thanks for the reviews! Please merge at your convenience :)