// CHECK-LABEL: func @init
func.func @init(%arg0: tensor<4xf32>) -> tensor<7xf32> {
    %0 = builtin.unrealized_conversion_cast to tensor<7xf32>
    return %0 : tensor<7xf32>
}

// CHECK-LABEL: func @next
func.func @next(%arg0: tensor<7xf32>, %arg1: tensor<3xf32>) -> tensor<6xf32> {
    %0 = builtin.unrealized_conversion_cast to tensor<6xf32>
    return %0 : tensor<6xf32>
}

// CHECK-LABEL: func @finalize

  %1 = "tf.Mul"(%arg1, %arg1) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  %2 = "tf.Div"(%arg1, %arg1) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>

  // CHECK: "tf.If"{{.+}}else_branch = @tf.IfRegion_else{{.+}}then_branch = @tf.IfRegion_then
  %3 = "tf.IfRegion"(%arg0) ({
     %4 = "tf.Select"(%arg0, %0, %1) : (tensor<i1>, tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
     "tf.Yield"(%4) : (tensor<2xf32>) -> ()

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<6xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<6xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

  %2 = "stablehlo.concatenate"(%0, %1) {dimension = 0 : i64} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
  %3 = "quantfork.stats"(%2) {layerStats = dense<[-0.83811146, 2.4960899]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>
  %4 = "quantfork.stats"(%arg2) {layerStats = dense<[-1.5726943, 1.07351148]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>

// PerChannel-DAG: %[[scale2:.*]] = "tf.Const"() <{value = dense<{{[0-9\.Ee\+\-]+}}> : tensor<6xf32>}> : () -> tensor<6xf32>
// PerChannel-DAG: %[[zp1:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<3xi32>}> : () -> tensor<3xi32>
// PerChannel-DAG: %[[zp2:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<6xi32>}> : () -> tensor<6xi32>

  # CHECK-SAME:   %arg0: tensor<1xf32> {tf._user_specified_name = "d", tf_saved_model.index_path = [0, "a"]},
  # CHECK-SAME:   %arg1: tensor<2xf32> {tf._user_specified_name = "d", tf_saved_model.index_path = [0, "b"]},
  # CHECK-SAME:   %arg2: tensor<3xf32> {tf._user_specified_name = "d", tf_saved_model.index_path = [0, "c"]},
  # CHECK-SAME:   %arg3: tensor<4xf32> {tf._user_specified_name = "d", tf_saved_model.index_path = [0, "x"]},

  func.return %0#0 : tensor<?x?x?x?xf32>
}

// -----

// CHECK-LABEL: fusedBatchNormV3_training_dynamic_unsupported1

  func.return %0#0 : tensor<8x8x8xf32>
}

// -----

// Test invalid tf.FusedBatchNorm
func.func @testFusedBatchNormWrongScaleType(tensor<8x8x8x8xf32>, tensor<8xi32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> tensor<8x8x8x8xf32> {

  // Use other output
  %3:6 = "tf.FusedBatchNormV3"( %2#0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", U = "tfdtype$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>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)