%2 = "tf.Reshape"(%1, %cst_0) : (tensor<1x10xf32>, tensor<2xi32>) -> tensor<?x10xf32> loc(fused["Reshape:", "model/reshape"])
    %3 = "tf.AddV2"(%2, %cst) : (tensor<?x10xf32>, tensor<10xf32>) -> tensor<?x10xf32> loc(fused["AddV2:", "model/add"])
    func.return %3 : tensor<?x10xf32>
  }
}

// CHECK-LABEL: func @matmul_with_reshape_disable_quantization
// CHECK: "tf.PartitionedCall"

    %5 = stablehlo.constant dense<2.000000e+02> : tensor<1x1xf32>  // Output inverse scale.
    %6 = stablehlo.constant dense<2> : tensor<1x1xi8>  // Output zero point.
    %7 = call @uniform_quantize_0(%arg0, %0, %1) : (tensor<1x2xf32>, tensor<1x1xf32>, tensor<1x1xi8>) -> tensor<1x2xi8>
    %8 = stablehlo.convert %7 : (tensor<1x2xi8>) -> tensor<1x2xf32>

  %1 = "tf.MatMul"(%arg0, %0) {T = f32, device = "/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  func.return %1 : tensor<3x3xf32>
}

// CHECK-LABEL: func @gpu_device
func.func @gpu_device(%arg0: tensor<3x1xf32>, %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<3x3xf32> {
  // CHECK: {{%.*}} = corert.get_op_handler %arg0 "/device:GPU:0"

// -----

// CHECK-LABEL: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

func.func @move_transpose_handle_broadcast(%arg0:tensor<8x64xf32>, %arg1:tensor<8x64x64xf32>) -> tensor<512x64xf32> {
  %cst = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
  %cst_1 = "tf.Const"() {value = dense<[2, 0, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
  %cst_2 = "tf.Const"() {value = dense<[512, 64]> : tensor<2xi32>} : () -> tensor<2xi32>
  %0 = "tf.ExpandDims"(%arg0, %cst) {device = ""} : (tensor<8x64xf32>, tensor<i32>) -> tensor<8x64x1xf32>

      fused_activation_function = "TANH",
      proj_clip = 0.000000e+00 : f32, time_major = false} : (
        tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,

// RUN: tf-tfrt-opt -tf-to-tfrt %s | FileCheck %s --dump-input=fail

// CHECK-LABEL: func @device_test
func.func @device_test(
    %arg0: tensor<3x1xf32> {tf_saved_model.index_path = [0]},
    %arg1: tensor<1x3xf32> {tf_saved_model.index_path = [0]})
      -> (tensor<3x3xf32> {tf_saved_model.index_path = []}) {
  // CHECK: {{%.*}} = corert.get_op_handler %arg0 "/device:GPU:0"

func.func @device_test(%arg0: tensor<3x1xf32>) -> (tensor<3x3xf32>) {

  // CHECK: device = "gpu"
  %0 = "tf.Const"() {value = dense<[[1.0, 2.0, 3.0]]> : tensor<1x3xf32>} : () -> tensor<1x3xf32>
  // CHECK: device = "gpu"
  %1 = "tf.MatMul"(%arg0, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>

  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %3 : tensor<2x1xf32>
})";
  const std::string kExpectedFB = CreateRuntimeMetadata();
  mlir::DialectRegistry registry;
  registry.insert<mlir::TFL::TensorFlowLiteDialect, mlir::arith::ArithDialect,

      %3 = stablehlo.reshape %2 : (tensor<1x3xf32>) -> tensor<3x1xf32>
      %4 = "quantfork.qcast"(%3) {volatile} : (tensor<3x1xf32>) -> tensor<3x1x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
      %5 = "quantfork.dcast"(%4) : (tensor<3x1x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x1xf32>
      return %5 : tensor<3x1xf32>
    }
  }
)mlir";