func.func @reshape_removeIdentity(tensor<4x4x4xf32>) -> tensor<4x4x4xf32> {
^bb0(%arg0: tensor<4x4x4xf32>) :
  %cst = arith.constant dense<[4, 4, 4]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<4x4x4xf32>, tensor<3xi32>) -> tensor<4x4x4xf32>
  func.return %0 : tensor<4x4x4xf32>

// CHECK-LABEL: func @reshape_removeIdentity
// CHECK:  return %arg0 : tensor<4x4x4xf32>
}

  %3 = "tfl.dequantize"(%2) : (tensor<1x3x3x!quant.uniform<i8:f32, 0.004:-128>>) -> tensor<1x3x3xf32>
  %4 = "tfl.div"(%arg0, %3) {fused_activation_function = "NONE"} : (tensor<1x3x3xf32>, tensor<1x3x3xf32>) -> tensor<1x3x3xf32>
  func.return %4 : tensor<1x3x3xf32>
//  CHECK: %[[logistic:.*]] = "tfl.logistic"
//  CHECK: %[[dq:.*]] = "tfl.dequantize"(%[[logistic]])

  constexpr absl::string_view
      kModuleXlaCallModuleOpWithDefaultQuantizationMethod = R"mlir(
    func.func @xla_call_module_default_quantization_method(%arg0: tensor<1x1x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<1x1x4xf32> {

  // CHECK-NEXT: [[out_ch:%.*]] = tfrt.merge.chains [[ch]], [[ch1]] : !tfrt.chain, !tfrt.chain
  // CHECK-NEXT: tfrt.return [[out_ch]], [[r2]] : !tfrt.chain, !tfrt_fallback.tf_tensor
  func.return %6#0 : tensor<3x3xf32>
}

    %cst_1 = "tf.Const"() <{value = dense<[[-0.795477629, 0.581315517, 0.921566545], [0.138622552, 0.463866323, 0.95474267], [-0.143770888, -0.796835303, 0.899996876], [0.0989735424, -0.483384758, -7.277030e-01]]> : tensor<4x3xf32>}> {device = ""} : () -> tensor<4x3xf32>

  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"

      -> (tensor<3x3xf32> {tf_saved_model.index_path = []}) {
  // CHECK: {{%.*}} = corert.get_op_handler %arg0 "/device:GPU:0"
  %2 = "tf.MatMul"(%arg0, %arg1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "/device:GPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  func.return %2 : tensor<3x3xf32>

// CHECK-LABEL: @add_f32(%arg0: tensor<3x3xbf16>, %arg1: tensor<3x3xbf16>) -> tensor<3x3xbf16>
func.func @add_f32(%arg0: tensor<3x3xf32>, %arg1: tensor<3x3xf32>) -> tensor<3x3xf32> {
  // CHECK-NOT: f32
  // CHECK: stablehlo.add
  %0 = stablehlo.add %arg0, %arg1: (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
  return %0 : tensor<3x3xf32>
}

// -----

// CHECK-LABEL: @add_f64(%arg0: tensor<3x3xbf16>, %arg1: tensor<3x3xbf16>) -> tensor<3x3xbf16>

  // attribute `_quantization_method` set to `"no_quantization {}"`.
  constexpr absl::string_view kXlaCallModuleOpWithQuantizationMethodAttr =
      R"mlir(
    func.func @main(%arg0: tensor<1x1x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<1x1x4xf32> {

    %0 = stablehlo.constant dense<1.000000e+00> : tensor<4x3xf32>
    %1 = "quantfork.qcast"(%0) {volatile} : (tensor<4x3xf32>) -> tensor<4x3x!quant.uniform<i8<-127:127>:f32:1, {5.000000e-03, 5.000000e-03, 5.000000e-03}>>
    %2 = "quantfork.dcast"(%1) : (tensor<4x3x!quant.uniform<i8<-127:127>:f32:1, {5.000000e-03, 5.000000e-03, 5.000000e-03}>>) -> tensor<4x3xf32>