%ch0 = tfrt.new.chain

  %a1 = tfrt_fallback_async.const_dense_tensor dense<[[2, 2], [2, 2]]> : tensor<2x2xi32>
  %a2 = tfrt_fallback_async.const_dense_tensor dense<[[3, 3], [3, 3]]> : tensor<2x2xi32>
  %b = tfrt_fallback_async.const_dense_tensor dense<[[1, 1], [1, 1]]> : tensor<2x2xi32>

  // Two batch_size=2 batches get concatenated.

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>
  // CHECK: device = "cpu"

  func.return %0 : tensor<*xf32>
}

func.func @xla_func_0(%arg0: tensor<7xf32>, %arg1: tensor<f32>) -> tensor<10xf32> {
  %cst = "tf.Const"() {value = dense<0> : tensor<1x2xi32>} : () -> tensor<1x2xi32>
  %cst_0 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
  %cst_1 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>

# 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"]})

// 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>

  // CHECK: {{%.*}} = tfrt_fallback_async.executeop {{.*}} device("/device:CPU:0") "tf.MatMul"

    %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"
  %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>

  // 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>
}

    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);

  func.return
}

func.func private @write_chain_func(%arg0: tensor<1xf32>) -> (tensor<1xf32>) {
  %cst = "tf.Const"() <{value = dense<[[0, 1]]> : tensor<1x2xi32>}> : () -> tensor<1x2xi32>
  %0 = "tf.XlaAllReduce"(%arg0, %cst) <{mode = "CrossReplica", reduce_op = "Add"}> : (tensor<1xf32>, tensor<1x2xi32>) -> tensor<1xf32>
  return %0 : tensor<1xf32>

  %2 = "tf.Transpose"(%1, %cst_1) {device = ""} : (tensor<8x64x64xf32>, tensor<3xi32>) -> tensor<64x8x64xf32>
  %3 = "tf.Reshape"(%2, %cst_2) {device = ""} : (tensor<64x8x64xf32>, tensor<2xi32>) -> tensor<512x64xf32>

  func.return %3 : tensor<512x64xf32>

  // CHECK-DAG: %[[CST_0:.*]] = "tf.Const"() <{value = dense<[2, 0, 1]> : tensor<3xi32>}> : () -> tensor<3xi32>