attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "training/LossScaleOptimizer/gradients/bn5c_branch2c_1/FusedBatchNormV2_grad/FusedBatchNormGradV2"
  op: "FusedBatchNormGradV2"
  input: "training/LossScaleOptimizer/gradients/activation_48_1/Relu_grad/ReluGrad"
  input: "res5c_branch2c_1/Conv2D"
  input: "bn5c_branch2c_1/ReadVariableOp"

cluster 0 size 1910
 Add 16
 AddN 71
 ArgMax 1
 AssignAddVariableOp 1
 BiasAdd 1
 BiasAddGrad 1
 Cast 115
 Const 407
 Conv2D 53
 Conv2DBackpropFilter 53
 Conv2DBackpropInput 52
 Equal 1
 FusedBatchNormGradV2 53
 FusedBatchNormV2 53
 MatMul 3
 MaxPool 1
 MaxPoolGrad 1
 Mean 1
 Mul 218
 Pad 2
 ReadVariableOp 538
 Relu 49
 ReluGrad 49
 Reshape 2
 ResourceApplyKerasMomentum 161
 Slice 1

  // CHECK-NEXT: %[[x_backprop:.*]] = mhlo.convert %[[mul3]] : tensor<8x8x8x8xf32>
  // CHECK-NEXT: return %[[x_backprop]] : tensor<8x8x8x8xf32>

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

in FusedBatchNorm.}]>:$reserve_space_4
  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
}

def TF_FusedBatchNormGradV2Op : TF_Op<"FusedBatchNormGradV2", [Pure]> {
  let summary = "Gradient for batch normalization.";

  let description = [{
Note that the size of 4D Tensors are defined by either "NHWC" or "NCHW".