Add WebNN baseline implementation for first-wave ops

[huningxin]

This PR implements 41 WebNN ops for first-wave models in JavaScript double-precision Number and calculations without depending on 3rd party libs. It could be used to generate the baseline results for WebNN conformance tests.

Fix webmachinelearning/webnn#245

@anssiko @dontcallmedom @wchao1115 @pyu10055 , PTAL.

/cc @BruceDai

[anssiko]

Thanks @huningxin and @BruceDai for this significant work. This proposal passes my review from the procedural perspective:

• Per our discussion on WebML WG Teleconference – 2 Dec 2021, webnn-baseline helps the WebML WG demonstrate Web Neural Network API implementation experience, a key requirement for advancing to a Candidate Recommendation.

• Considering the above and per WebML CG Charter: Test Suites and Other Software, webnn-baseline is now proposed as a new CG deliverable. (The CG and WG will continue to work closely together per Coordination.)

[dontcallmedom]

Great work! In order to help with the goal of this codebase (making it easy to review it), I've created separately a couple of additional commits that I think help in that direction.

They take the following steps:

• move any code that isn't directly a WebNN operation (but rather support code for operations) into the src/lib directory
• tersify the definitions of operations by
  • moving as much of the parameters normalization into the function signature using JavaScript deconstruction patterns (and other syntactical shortcuts)
  • moving all the validation of the parameters into a separate module, to have the main module focus only on the actual calculation

An alternative to the latter subbullet might be to get rid of parameters validation altogether - since the code isn't actually meant to be used in any production environment, it's not clear that the code adds a lot of value. The refactoring would make that a fairly easy additional change.

[huningxin]

Thanks much @dontcallmedom !

• move any code that isn't directly a WebNN operation (but rather support code for operations) into the src/lib directory

+1. In the implementation, some big ops are implemented by composing small ops. These small ops are also WebNN ops, like reshape, so please keep them in the src folder.

• moving as much of the parameters normalization into the function signature using JavaScript deconstruction patterns (and other syntactical shortcuts)

+1. I like that.

• moving all the validation of the parameters into a separate module, to have the main module focus only on the actual calculation

+1

An alternative to the latter subbullet might be to get rid of parameters validation altogether - since the code isn't actually meant to be used in any production environment, it's not clear that the code adds a lot of value.

It's a fair point.

However, I suppose the validation code would help test developer to verify the test cases. @BruceDai

Another perspective is it probably could help the spec editors to write the validation steps of a method, as webmachinelearning/webnn#210. @anssiko @wchao1115

[dontcallmedom]

An alternative to the latter subbullet might be to get rid of parameters validation altogether - since the code isn't actually meant to be used in any production environment, it's not clear that the code adds a lot of value.

However, I suppose the validation code would help test developer to verify the test cases. @BruceDai

I can certainly relate to that - they did help mind find and fix bugs I introduced in the process of refactoring the code :)

Another perspective is it probably could help the spec editors to write the validation steps of a method, as webmachinelearning/webnn#210. @anssiko @wchao1115

indeed

[anssiko]

If the editors feel the validation code helps them translate checks into corresponding spec validation steps, it sounds like a good reason to keep that part to facilitate this work. A separate validation module is a good idea too from a code organization perspective.

[huningxin]

huningxin#1 merged. Thanks @dontcallmedom 's contribution that improves the reviewability much.

[huningxin]

@wchao1115, this function implements the steps of conv2d, just for your convivence, thanks!

[anssiko]

@huningxin as per our discussion, please feel free to merge this PR now to unblock webmachinelearning/webnn#265

Thanks to all reviewers and special thanks to @huningxin and @BruceDai for preparing this initial contribution that is an important deliverable to demonstrate implementation experience and satisfy interoperability requirements.