Support reverse operator

[huningxin]

This operator reverses the order of the input tensor along specified axes. There are some pytorch models using flip operator. When they target WebNN, the flip/reverse operator get fallback to other execution providers and cause performance drop.

Frameworks' support

• PyTorch flip
• TensorFlow reverse
• ONNX Slice support reverse slicing by setting step to -1

Native ML API's support

• DML DML_SLICE1_OPERATOR_DESC supports reverse slicing by setting step to -1
• CoreML tensor_transformation.reverse
• TFLite tfl.reverse_v2

Proposal

dictionary MLReverseOptions : MLOperatorOptions {
  sequence<[EnforceRange] unsigned long> axes;
};
 
MLOperand reverse(MLOperand input, optional MLReverseOptions options = {});

Discussed with @fdwr offline, Dwayne recommended treating the axes similar to reduce operators. So if not provided at all, then all axes are reversed; but if explicitly passed as empty, then no axes are reversed.

[fdwr]

Signature looks good to me Ningxin. 👍

Also related:

• StableHLO reverse
• tosa.reverse (mlir::tosa::ReverseOp)
• ANEURALNETWORKS_REVERSE
• PyTorch prims rev

(historical trivia) There was actually an ONNX Reverse operator here briefly, but it was rushed before all the feedback was addressed and reverted, and then it appears Peyman never followed up with an updated version. 🤷‍♂️. (I've also thought about adding a DML_OPERATOR_REVERSE before, but didn't get to it as I didn't know a clear model that used it)