//   %cst_2 = "tf.Const"() {value = dense<[0.3, 0.4, 0.3, 0.4, 0.3, 0.4, 0.3, 0.4]> : tensor<8xf32>} : () -> tensor<8xf32>

// CHECK:  %[[ADD2:.*]] = "tf.Add"(%[[MUL2]], %[[SUB]])
// CHECK:  %[[BATCHNORM1_a:[^,]+]], {{.*}} = "tf.FusedBatchNormV3"(%[[ADD2]], %[[CONSTANT1]], %[[ARG2]], %[[ARG3]], %[[ARG4]])
// CHECK:  %[[BATCHNORM1_b:[^,]+]], {{.*}} = "tf.FusedBatchNormV3"(%[[BATCHNORM1_a]], %[[CONSTANT1]], %[[ARG2]], %[[ARG3]], %[[ARG4]])
// CHECK:  "tf.FusedBatchNormV3"(%[[BATCHNORM1_b]], %[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG4]])
}

  %cst_0 = "tf.Const"() {value = dense<0.500000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

  (TF_CheckNumericsOp $arg, $msg),
  (replaceWithValue $arg)>;

// Remove StopGradient op
def RemoveStopGradient : Pat<
  (TF_StopGradientOp $arg),
  (replaceWithValue $arg)>;

// Converts tf.FusedBatchNormV3 into a sequence of more primitive arithmetic
// operations. Specifically, performs the following calculation:
//
//   (x - mean) * scale / sqrt(variance + epsilon) + offset
//

// the checks on intermediate values, ideally we should find equivalent
// checks that guarantees the resultant ops are valid.
// The extra conditions are the broadcasting conditions.
//
// The pattern lower FusedBatchNormV3 to arithmetic ops.
// Specifically, performs the following calculation:
//
//   (x - mean) * scale / sqrt(variance + epsilon) + offset
//
// Let multiplier = scale / sqrt(variance + epsilon),
// to compute

    %0:6 = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>,...

func.func @fusedBatchNormV3_training_dynamic_unsupported1(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>, %arg3: tensor<?xf32>, %arg4: tensor<?xf32>) -> (tensor<?x?x?x?xf32>) {
  // CHECK: tf.FusedBatchNormV3

      self.assertTrue(self._contains_op(output_graphdef, 'XlaCallModule'))
    else:
      self.assertTrue(self._contains_quantized_function_call(output_graphdef))
    self.assertFalse(self._contains_op(output_graphdef, 'FusedBatchNormV3'))

  @parameterized.named_parameters(
      ('to_tf_with_int32_input_type', dtypes.int32, quant_opts_pb2.TF),
      ('to_xla_with_int32_input_type', dtypes.int32, quant_opts_pb2.XLA),

            "SparseSoftmaxCrossEntropyWithLogits", "LogSoftmax"}},
          {"REDUCEWINDOWPW", {"BiasAddGrad", "LRN", "LRNGrad"}},
          {"BN",
           {"FusedBatchNorm", "FusedBatchNormV2", "FusedBatchNormV3",
            "_FusedBatchNormEx", "FusedBatchNormGrad", "FusedBatchNormGradV2",
            "FusedBatchNormGradV3"}},
          {"Conv", {"_FusedConv2D"}},

func.func @testFusedBatchNormToBatchNormV3(%arg0: tensor<8x8x8x8xf32>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<8xf32>) -> (tensor<8x8x8x8xf32>) {
  // CHECK: "tf.FusedBatchNormV3"
  // CHECK-SAME: device = "/job:localhost/replica:0/task:0/device:GPU:0"