Thanks for the review, updated the PR with addressed feedback.

shall we enhance this op to support reverse sequences of different lengths? in tf, it is called tf.reversesequence().

@pengwa If you are referring to the case where sequential data of different lengths has been padded, it may be cleaner to have a separate op to cover such scenario.

@linkerzhang, @houseroad: can you please review the updated PR so we can move ahead with this op? thanks.

@pk-g can you update the PR by fixing the conflicts?

@houseroad any more comments on this op please?

Thank you!

I will do the review later this week, but please don't rush.

Will revert it here: #1882

Please resubmit the changes.

Will revert it here: #1882

Please resubmit the changes.

Was this op reverted and then simply forgotten to re-add? It's found in the following libraries...

• PyTorch flip
  TensorFlow reverse
• CoreML tensor_transformation.reverse
• TFLite tfl.reverse_v2
• StableHLO reverse
• tosa.reverse (mlir::tosa::ReverseOp)
• ANEURALNETWORKS_REVERSE
• PyTorch prims rev

...and we're looking to add it to WebNN for a voice to text model. It can be achieved in ONNX via the overloaded interpretation of setting ONNX Slice's to -1, but loses semantics and is more awkward (needing to specify starts: [2147483647, 2147483647], ends: [-2147483648, -2147483648], steps: [-1,-1], axes: [0,1] rather than just axes: [0,1]).