WebML WG Teleconference – 27 June 2024 - Open issues and PRs

[anssiko]

We'll discuss open issues and review PRs based on your feedback:

• ⨀ All open issues
• ⛙ All open pull requests

Please suggest issue(s) and PR(s) you'd like to see discussed on this teleconference in this issue.

[a-sully]

I'd like to discuss webmachinelearning/webnn#567

[inexorabletash]

While most feedback will probably require async review, there was some interest in making progress on webmachinelearning/webnn#647 and specifically webmachinelearning/webnn#647 (comment) so I can give a quick overview of the PR and its implications, and what the proposed next steps are. Hopefully that will generate some feedback.

[inexorabletash]

Async update for webmachinelearning/webnn#647 "MLNumber" proposal

• Some places that took a float or double argument/option now take MLNumber or MLFiniteNumber - a "variant" (union) of either a JS Number (equivalent to a C++ double) or BigInt (for full precision int64/uint64 use)
  • constant() (scalar overload) - MLNumber
  • clamp() (min/max options) - MLNumber
  • pad() (padding value) - MLFiniteNumber
  • Fixes Type of some parameters should match the input data type webnn#442
• Places that took a float argument/option now take a double. These are all ops that only operate on float types, so no need for MLNumber. This follows web API best practices, and allows full precision conversion to float16 (float64 to float32 to float16 is lossy)
  • batchNormalization(), instanceNormalization(), layerNormalization() - epsilon
  • elu(), gemm(), hardSigmoid(), leakyRelu(), linear() - alpha, beta
  • Fixes Clarify the usage of 32 bit floating point type and consider using double webnn#325
• In all of these places, the input number is cast to the input/output operand data type when it is used.
  Casting algorithms are spelled out; latest discussion proposes we always have "clamp" semantics (i.e. no weird modulus wrapping), and never fail
  • For MLOperand-vending methods, the conversion can be done eagerly.
  • For MLActivations, the conversion is done at "fusion time"; the same MLActivation instance could be fused with multiple ops, and cast differently for each.
  • Fixes Specifies scalar values casted to match input type. webnn#678
• Using "Number or BigInt" is novel here, but WebIDL experts agree it's reasonable. Note that all existing WebIDL/JS APIs that are designed to take a Number, or to take a BigInt, will throw if the other thing is passed - there is intentionally no silent conversion. So we are in novel territory here and should look for implementation and developer feedback.
• Prototyping in https://chromium-review.googlesource.com/c/chromium/src/+/5466806 and some discussion about error cases, incorporated into the PR.
• Note that Clarify the casting behavior from floating-point / signed integers <-> unsigned integers webnn#489 looks similar, but is about the explicit cast() op.