%cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>
  %1 = "tf.MatMul"(%0, %cst) {device = "", transpose_a = false, transpose_b = false} : (tensor<1x10xbf16>, tensor<10x2xbf16>) -> tensor<1x2xbf16>
  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2xbf16>) -> tensor<1x2xf32>

    // 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>)

// RUN: tf-mlir-translate -mlir-to-graph %s -o - | FileCheck %s

func.func @main() -> (tensor<1x2xf16>, tensor<2xf16>) {
  %graph:2 = tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<1.0> : tensor<1x2xf16>} : () -> tensor<1x2xf16> loc("const1")
// CHECK: node {
// CHECK-NEXT: name: "const1"
// CHECK-NEXT: op: "Const"
// CHECK: dtype: DT_HALF
// CHECK: half_val: 15360

// RUN: tf-mlir-translate -mlir-to-graphdef %s -o - | FileCheck %s

func.func @main() -> (tensor<1x2xf16>, tensor<2xf16>) {
  %graph:2 = tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<1.0> : tensor<1x2xf16>} : () -> tensor<1x2xf16> loc("const1")

}

func.func @RemoveTrivialAddBf16RHS(%arg0: tensor<2x2xbf16>) -> tensor<2x2xbf16> {
  %cst = arith.constant dense<0.0> : tensor<2x2xbf16>
  %0 = "tf.Add"(%arg0, %cst) : (tensor<2x2xbf16>, tensor<2x2xbf16>) -> tensor<2x2xbf16>
  func.return %0 : tensor<2x2xbf16>

  // CHECK-LABEL: RemoveTrivialAdd
  // CHECK-NEXT: return %arg0 : tensor<2x2xbf16>
}

TEST_F(MlirTranslationTest, TranslatesMlirToGraph) {
  static constexpr char kMlirSource[] = R"(
func.func @main() -> (tensor<1x2xf16>, tensor<2xf16>) {
  %graph:2 = tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<1.0> : tensor<1x2xf16>} : () -> tensor<1x2xf16> loc("const1")

    return %1: tensor<1x3xbf16>
  }

  %8 = mhlo.and %5, %7 : tensor<10x10xi1>
  %9 = mhlo.subtract %arg0, %4 : tensor<10x10xbf16>
  %10 = mhlo.divide %9, %1 : tensor<10x10xbf16>
  %11 = mhlo.add %10, %3 : tensor<10x10xbf16>
  %12 = "mhlo.select"(%8, %11, %10) : (tensor<10x10xi1>, tensor<10x10xbf16>, tensor<10x10xbf16>) -> tensor<10x10xbf16>
  %13 = "mhlo.round_nearest_afz"(%12) : (tensor<10x10xbf16>) -> tensor<10x10xbf16>

  return %0 : tensor<3x3xf64>
}

// -----

// CHECK-LABEL: @constant_f32() -> tensor<2x2xbf16>
func.func @constant_f32() -> tensor<2x2xf32> {
  // CHECK-NOT: f32
  // CHECK{LITERAL}: stablehlo.constant dense<[[1.398440e+00, 0.000000e+00], [3.093750e+00, -2.001950e-01]]> : tensor<2x2xbf16>
  %0 = stablehlo.constant dense<[[1.4, 0.0], [3.1, -0.2]]> : tensor<2x2xf32>
  return %0 : tensor<2x2xf32>
}

// -----

      shape {
        unknown_rank: true
      }
    }
  }
  attr {
    key: "dtype"
    value {
      type: DT_BOOL
    }
  }
}

# CHECK-DAG: tf.IteratorGetNext{{.+}}-> (tensor<1x8xbf16>, tensor<2x?x16xf32>, tensor<*xf64>)
# CHECK-DAG: tf.IteratorGetNextSync{{.+}}-> (tensor<*xi16>, tensor<3x24xi32>, tensor<?x4x32xi64>)