*   Add `UnifiedGRU` as the new GRU implementation for tf2.0. Change the default
    recurrent activation function for GRU from `hard_sigmoid` to `sigmoid`, and
    `reset_after` to True in 2.0. Historically recurrent activation is
    `hard_sigmoid` since it is fast than 'sigmoid'. With new unified backend
    between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we change

    QuantizableResult,
    PredOpTrait<"input and output must have same element type",
      TFL_TCresVTEtIsSameAsOp<0, 0>>]> {
  let summary = "Hardswish activation function.";
  let description = [{
    Computes hard-swish activation function
      f(x) -> (x * relu6(x+3))/6
    element-wise.
  }];

  let arguments = (ins TFL_TensorOf<[F32, QUI8, QI8]>:$input);

      Option<"quantize_signed_", "quantize-signed", "bool", "false",
             "signed inference type. Only used in tests">,
      Option<"activation_number_of_bits_", "activation-number-of-bits", "int", "8",
             "number of bits for inference type. Only used in tests">,
      Option<"post_training_quantize_", "post-training-quantize", "bool", "false",

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true activation-number-of-bits=16" -cse | FileCheck %s

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>