func.func @broadcast_to_bf16(%arg0: tensor<3xbf16>, %arg1: tensor<2xi64>) -> tensor<3x3xbf16> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<3xbf16>, tensor<2xi64>) -> tensor<3x3xbf16>
  func.return %0: tensor<3x3xbf16>

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>

func.func @SparseSoftmaxCrossEntropyWithLogits_with_dynamic(%features: tensor<*xbf16>, %labels: tensor<*xi64>) -> (tensor<2xbf16>, tensor<*xbf16>) {
  // CHECK: tf.SparseSoftmaxCrossEntropyWithLogits
  %0:2 = "tf.SparseSoftmaxCrossEntropyWithLogits"(%features, %labels) : (tensor<*xbf16>, tensor<*xi64>) -> (tensor<2xbf16>, tensor<*xbf16>)
  func.return %0#0, %0#1 : tensor<2xbf16>, tensor<*xbf16>
}

// CHECK-LABEL: func @tanhgrad_float

  %0 = "tf.Const"() {value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : () -> tensor<4xui64>
  // CHECK: tfrt_fallback_async.const_dense_tensor  dense<1.000000e+00> : tensor<1xbf16>
  %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<[1.0]> : tensor<1xbf16>} : () -> tensor<4xbf16>
  // CHECK: corert.executeop({{.*}}) "tf.Const"() {dtype = ui64, value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : 1

// CHECK:  return %[[V1]] : tensor<32x1x256xbf16>
}

// CHECK-LABEL: Relu_bf16
func.func @Relu_bf16(%arg0: tensor<2x3xbf16>) -> tensor<2x3xbf16> {
  %cst = arith.constant dense<0.0> : tensor<2x3xbf16>
  %0 = "tfl.maximum"(%arg0, %cst) : (tensor<2x3xbf16>, tensor<2x3xbf16>) -> tensor<2x3xbf16>
  func.return %0 : tensor<2x3xbf16>

  // CHECK: %[[RESULT:.*]] = "tfl.relu"(%arg0)
  // CHECK: return %[[RESULT]]
}

// Tests that an AddV2 op accepting two bf16 operands is transformed into
// an AddV2 op that accepts two fp32 operands.
func.func @cast_bf16_add_v2_to_fp32(%arg0: tensor<2xbf16>, %arg1: tensor<2xbf16>) -> tensor<2xf32> {
  %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<2xbf16>, tensor<2xbf16>) -> tensor<2xbf16>
  %1 = "tf.Cast"(%0) {Truncate = false} : (tensor<2xbf16>) -> tensor<2xf32>
  return %1 : tensor<2xf32>
}

  %cst_1 = arith.constant dense<[1, -1, 6, 1, 6, 1, 1, -1]> : tensor<8xi64>
  %97 = "tf.BroadcastTo"(%arg0, %cst_1) : (tensor<1x?x1x1x1x1x1x?xbf16>, tensor<8xi64>) -> tensor<1x?x6x1x6x1x1x?xbf16>
  %cst_2 = arith.constant dense<[8, 6, 6, 18]> : tensor<4xi64>
  %98 = "tf.Reshape"(%97, %cst_2) : (tensor<1x?x6x1x6x1x1x?xbf16>, tensor<4xi64>) -> tensor<8x6x6x18xbf16>

  }

  func.func private @composite_dequantize_uniform(%arg0: tensor<*xi8>) -> tensor<*xbf16> {
    %cst = "tf.Const"() {value = dense<1.574710e-02> : tensor<bf16>} : () -> tensor<bf16>
    %0 = "tf.Cast"(%arg0) {Truncate = false} : (tensor<*xi8>) -> tensor<*xbf16>
    %1 = "tf.Mul"(%0, %cst) : (tensor<*xbf16>, tensor<bf16>) -> tensor<*xbf16>
    return %1 : tensor<*xbf16>
  }

// CHECK-LABEL: func @propagate_xladotv2_bf16

  %0 = "tf.AddV2"(%arg0, %arg1) {device = ""} : (tensor<*xbf16>, tensor<*xbf16>) -> tensor<*xbf16>
  func.return %0 : tensor<*xbf16>
}

// -----

func.func @testXlaReduceWindowAttrs(%arg0: tensor<7xf32>, %arg1: tensor<f32>) -> tensor<?xf32> {
  %cst = "tf.Const"() {value = dense<0> : tensor<1x2xi32>} : () -> tensor<1x2xi32>

    // CHECK: %[[CAST_1:.*]] = "tf.Cast"(%[[ARG_1]]) <{Truncate = false}> : (tensor<10xf32>) -> tensor<10xbf16>
    // CHECK: %[[RESULT:.*]]:2 = "tf.XlaCallModule"(%[[CAST_0]], %[[CAST_1]], %[[ARG_2]])
    // CHECK-SAME: (tensor<10xbf16>, tensor<10xbf16>, tensor<6xi32>) -> (tensor<10xbf16>, tensor<6xi32>)

  %8 = "tfl.mul"(%2, %3) {fused_activation_function = "NONE"} : (tensor<4xbf16>, tensor<4xbf16>) -> tensor<4xbf16>

  func.return %5, %6, %7, %8 : tensor<bf16>, tensor<4xbf16>, tensor<4xbf16>, tensor<4xbf16>
}

// CHECK-LABEL: @mul_f16
func.func @mul_f16() -> (tensor<f16>, tensor<4xf16>, tensor<4xf16>, tensor<4xf16>) {
  %0 = arith.constant dense<4.5> : tensor<f16>