# CHECK:      func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME:   [[ARG0:%.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]},
# CHECK-SAME:   [[ARG1:%.*]]: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.bound_input = @[[VAR]]})
# CHECK-SAME:             -> (tensor<3x3xf32> {tf_saved_model.index_path = ["r"]})

func.func @TfBatchMatMulV2(%arg0: tensor<4x128x2xf32>, %arg1:  tensor<2x1xf32>) -> tensor<4x128x1xf32> {
  %0 = "tfl.custom"(%arg0, %arg1) {
    custom_code = "FlexBatchMatMulV2",
    custom_option = #tfl<const_bytes : "0x0D42617463684D61744D756C56320038120D42617463684D61744D756C56321A001A002A070A0154120230012A0B0A0561646A5F78120228002A0B0A0561646A5F791202280032000002493B1414042801">
  } : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>

  func.return
}

// CHECK-LABEL: func @basic
func.func @basic(
    %arg0: tensor<3x1xf32>,
    %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> (tensor<3x3xf32>) {
  %1 = "tf.ReadVariableOp"(%arg1) {_output_shapes = ["tfshape$dim { size: 1 } dim { size: 3 }"], device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>

    %11 = "tf.Pad"(%arg0, %10) : (tensor<2x224x224x3xf32>, tensor<4x2xi32>) -> tensor<2x230x230x3xf32>
    %12 = "tf.Cast"(%arg1) {Truncate = false} : (tensor<2x1xf32>) -> tensor<2x1xi64>
    %13 = "tf.Reshape"(%12, %9) : (tensor<2x1xi64>, tensor<1xi32>) -> tensor<2xi64>
    %14 = "tf.Squeeze"(%arg1) {squeeze_dims = [-1]} : (tensor<2x1xf32>) -> tensor<2xf32>
    // CHECK: "tf.Conv2D"

# CHECK-NEXT: [[R0:%.*]] = "tf.ReadVariableOp"([[ARG1]]) {{{.*}}} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>
# CHECK-NEXT: [[R1:%.*]] = "tf.MatMul"([[ARG0]], [[R0]]) <{{{.*}}}> {{{.*}}} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
# CHECK-NEXT: return [[R1]] : tensor<3x3xf32>


def Test():

  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.compat.v1.get_variable(
      name='y',

    module attributes {} {
      func.func @main(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> attributes {} {
        %0 = stablehlo.constant dense<1.0> : tensor<4x3xf32>
        %1 = stablehlo.dot_general %arg0, %0, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>
        return %1 : tensor<1x3xf32>
      }
    }
  )mlir");
  ASSERT_TRUE(module_op);

  // CHECK-NEXT: [[r1:%.*]] = tfrt_fallback_async.executeop {{.*}} "tf.BiasAdd"([[r0]], [[result]])
  %3 = "tf.BiasAdd"(%2, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], data_format = "NHWC", device = "/device:CPU:0"} : (tensor<3x3xf32>, tensor<3xf32>) -> tensor<3x3xf32>

}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

  %4 = "tf.MatMul"(%arg0, %3) {device = "", transpose_a = false, transpose_b = false} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  %5 = "tf.Identity"(%4) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  %6 = "tf.Identity"(%5) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %6 : tensor<2x4xf32>

  // CHECK-LABEL: QuantDequantTranspose

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir