Need for type casting?

[wacky6]

This came up during review for softmax implementation https://chromium-review.googlesource.com/c/chromium/src/+/3856752/comments/267c926c_b14af87f

Softmax will likely only support float point types: #283

Currently, the spec provides no way to convert between data types, this might make it hard to use softmax with certain computation graphs.

This could impact models doing integer computations (e.g. image recognition models that takes in RGB24 image represented by uint8), and returns softmax-ed predictions (e.g. one-hot encoding).

For example, the following graph:

   Input (uint8)
-> matmul (uint32)
-> softmax

Perhaps WebNN needs a type-casting function?

In ML frameworks:

• tf.cast
• torch.Tensor.to with dtype
• Android NDK
• ONNX cast

[fdwr]

Definitely a gap, and included in v2 ops for transformers: #375 (comment)

partial interface MLGraphBuilder {
  MLOperand cast(MLOperand input, MLOperandDataType operandDataType);
}

[inexorabletash]

Now that cast() landed (PR #478) can this be closed?

[inexorabletash]

Closing this as complete.