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.