func.func @concat_v2_1d_axis(%arg0: tensor<3x3xf32>, %arg1: tensor<3x3xf32>) -> tensor<3x6xf32> {
  // CHECK: "mhlo.concatenate"({{.*}}) <{dimension = 1 : i64}> : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x6xf32>

  %axis = "tf.Const"() { value = dense<[1]> : tensor<1xi64> } : () -> tensor<1xi64>
  %1 = "tf.ConcatV2"(%arg0, %arg1, %axis) : (tensor<3x3xf32>, tensor<3x3xf32>, tensor<1xi64>) -> tensor<3x6xf32>
  func.return %1 : tensor<3x6xf32>
}

  %1 = "tf.AddV2"(%arg0, %0) : (tensor<4x1xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  %2 = "tf.AddV2"(%0, %arg0) : (tensor<1x2xf32>, tensor<4x1xf32>) -> tensor<4x2xf32>

  // If operand has the same shape as a result, we can fold it.
  %3 = "tf.AddV2"(%arg1, %0) : (tensor<4x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  %4 = "tf.AddV2"(%0, %arg1) : (tensor<1x2xf32>, tensor<4x2xf32>) -> tensor<4x2xf32>

  // CHECK: %[[CONST:.*]] = "tf.Const"()

    "tf.AssignVariableOp"(%arg0, %partitioned_output#0) : (tensor<!tf_type.resource<tensor<3x2xf32>>>, tensor<3x2xf32>) -> ()
    "tf.AssignVariableOp"(%arg1, %partitioned_output#1) : (tensor<!tf_type.resource<tensor<3x2xf32>>>, tensor<3x2xf32>) -> ()
    func.return
  }
  func.func @computation(%arg0: tensor<3x4xf32>) -> tensor<3x4xf32> {
    func.return %arg0: tensor<3x4xf32>
  }
}

// -----

  %1 = "tfl.reshape"(%0, %cst3) : (tensor<4x2xf32>, tensor<2xi32>) -> tensor<1x8xf32>
  %2 = "tfl.mul"(%0, %cst2) {fused_activation_function = "RELU6"} : (tensor<4x2xf32>, tensor<2xf32>) -> tensor<4x2xf32>

  func.return %1, %2 : tensor<1x8xf32>, tensor<4x2xf32>

  func.return %24 : tensor<1x4xf32>

// `tfl.concatenation`.

func.func @concatenate_float(%arg0: tensor<3x2xf32>, %arg1: tensor<1x2xf32>) -> tensor<4x2xf32> {
  %0 = "stablehlo.concatenate"(%arg0, %arg1) {dimension = 0 : i64} : (tensor<3x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  return %0 : tensor<4x2xf32>
}
// CHECK-LABEL: concatenate_float
// CHECK-NOT: tfl.concatenation
// CHECK: stablehlo.concatenate

// -----

        tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,
        none, none, none, none) -> tensor<*xf32>

  %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> {

  func.return %2 : tensor<3x3xf32>

  // 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>