// Since the tf.AddV2 op has two uses, we have a _FusedBatchNormEx without the
// Relu activation and we only fuse the add.
// CHECK-NEXT: %[[Y:[a-z0-9]*]], {{.*}}_FusedBatchNormEx
// CHECK-NEXT: %[[relu:[a-z0-9]*]] ={{.*}}Relu"(%[[Y]]
// CHECK-NEXT: return %[[relu]]

//   %0 = "tf.Relu"(%y#0)
// ->
//   %y:6 = "tf._FusedBatchNormEx"(%x, %scale, %offset, %mean, %variance)
//
// Or:
//   %y:6 = "tf.FusedBatchNormV3"(%x, %scale, %offset, %mean, %variance)
//   %0 = "tf.AddV2"(%y#0, %side_input)
//   %1 = "tf.Relu"(%0)
// ->
//  %y:6 = "tf._FusedBatchNormEx"(%x, %scale, %offset, %mean, %variance,
//                                %side_input)

          {"REDUCEWINDOWPW", {"BiasAddGrad", "LRN", "LRNGrad"}},
          {"BN",
           {"FusedBatchNorm", "FusedBatchNormV2", "FusedBatchNormV3",
            "_FusedBatchNormEx", "FusedBatchNormGrad", "FusedBatchNormGradV2",
            "FusedBatchNormGradV3"}},
          {"Conv", {"_FusedConv2D"}},
          {"SORT", {"TopKV2"}},  // XLA version much faster then TF version.

  );

  let results = (outs
    TF_Tensor:$output
  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
}

def TF__FusedBatchNormExOp : TF_Op<"_FusedBatchNormEx", [Pure]> {
  let summary = "Internal FusedBatchNorm operation: reserved for internal use.";

  let description = [{
Do not invoke this operator directly in Python. A fusion optimization is