// 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>)

  %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>

  %0 = "tfl.conv_2d"(%arg0, %arg1, %cst_0) {dilation_h_factor = 2 : i32, dilation_w_factor = 3 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 4 : i32, stride_w = 5 : i32} : (tensor<256x32x32x3xbf16>, tensor<16x3x3x3xbf16>, tensor<16xbf16>) -> tensor<256x8x7x16xbf16>

  %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

// RUN: tf-opt %s -tfl-prepare-tf -tfl-legalize-tf='run-tfl-runtime-verification=false'  -tfl-optimize | FileCheck %s

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>

    return %1: tensor<1x3xbf16>
  }

  %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>

      return %0 : tensor<1x?x4xbf16>
    }
  })";

  TF_ASSERT_OK(LegalizeModule(kModuleWithNonConstParam));
}

TEST_F(Tf2XlaRewriterTest, ErrorsWithInvalidNumberOfParametersToArgs) {

    %2:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = bf16, value = dense<[4.900000e+01, 8.200000e+02]> : tensor<2xbf16>} : () -> tensor<bf16> loc("const3")
    %3:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = bf16, value = dense<0.000000e+00> : tensor<bf16>} : () -> tensor<bf16> loc("const4")

  EXPECT_THAT(num_bytes, Eq(16));
}

TEST_F(GetSizeInBytesTest, Bfloat16ConstOpSizeInBytes) {
  constexpr absl::string_view kConstOpExpr = R"mlir(
    %cst = "tf.Const"() {value = dense<1.0> : tensor<7xbf16>} : () -> tensor<7xbf16>
  )mlir";

  Block block{};
  TF::ConstOp int_tensor_const_op = ParseConstOp(kConstOpExpr, block, ctx_);

  const int64_t num_bytes = GetSizeInBytes(int_tensor_const_op);