%add = "tf.AddV2"(%cst, %cst) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
        %var_handle = "tf.VarHandleOp"() {container = "", device = "/device:CPU:0", shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
        "tf.AssignVariableOp"(%var_handle, %add) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
        return
      }
    }

    %1 = "tfl.abs"(%arg2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    %2 = stablehlo.add %1, %arg2 : tensor<5xf32>
    %3 = "tfl.abs"(%2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    "stablehlo.return"(%3) : (tensor<5xf32>) -> ()
  }) {dimensions = array<i64: 0>} : (tensor<7x5xf32>, tensor<5xf32>) -> tensor<5xf32>
  func.return %0: tensor<5xf32>
}

// -----

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()

    %cst_1 = "tf.Const"() {value = dense<2> : tensor<4xi32>} : () -> tensor<4xi32>

         data_format = "NCHW",
         epsilon = 1.001000e-05 : f32,
         is_training = false
       } : (tensor<?x64x112x112xf32>, tensor<64xf32>, tensor<64xf32>, tensor<64xf32>, tensor<64xf32>)
        -> (tensor<?x64x112x112xf32>, tensor<64xf32>, tensor<64xf32>, tensor<64xf32>, tensor<64xf32>, tensor<*xf32>)

  // CHECK: "tf.FusedBatchNormV3"
  // CHECK-SAME: data_format = "NHWC"

         exponential_avg_factor = 1.0 : f32,
         is_training = true
       }
        : (tensor<1x28x28x64xf32>, tensor<64xf32>, tensor<64xf32>,
           tensor<64xf32>, tensor<64xf32>)
       -> (tensor<1x28x28x64xf32>, tensor<64xf32>, tensor<64xf32>,
           tensor<64xf32>, tensor<64xf32>, tensor<64xf32>)

  func.return %y : tensor<1x28x28x64xf32>
}

// CHECK-LABEL: func @transposeFusedBatchNormGradV3

///      : (tensor<f32>, tensor<f32>, tensor<f32>) -> tensor<5xf32>
///
/// Output would be:
///   %iota = "mhlo.iota"() {iota_dimension = 0 : i64} : () -> tensor<5xf32>
///   %scaled = "mhlo.multiply"(%iota, %delta)
///       {broadcast_dimensions = dense<[]> : tensor<0xi64>} :
///       (tensor<5xf32>, tensor<f32>) -> tensor<5xf32>
///   %result = "mhlo.add"(%scaled, %offset)

    %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3xf32>} : () -> tensor<2x3xf32>
    %cst_0 = "tf.Const"() {value = dense<4.00000000e-1> : tensor<3xf32>} : () -> tensor<3xf32>
    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<?x2xf32>) -> tensor<?x2xf32>

  %b = arith.constant dense<-1.23697901> : tensor<64xf32>
  %conv = "tfl.conv_2d"(%arg0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<1x224x224x3xf32>, tensor<64x3x3x3xf32>, tensor<64xf32>) -> tensor<1x112x112x64xf32>
  func.return %conv : tensor<1x112x112x64xf32>

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
    %cst_1 = "tf.Const"() {value = dense<2.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

      %a, %control1 = tf_executor.island wraps "tf.A"() {_class = ["loc:@class"]} : () -> (tensor<2xf32>)
      %s:2, %control2 = tf_executor.island wraps "tf.Split"(%c, %a) {num_split = 2 : i32} : (tensor<i32>, tensor<2xf32>) -> (tensor<1xf32>, tensor<1xf32>)
      tf_executor.fetch
    }
    func.return
  }