def testBiasVec(self):
   with self.assertRaises(ValueError):
     nn_ops.bias_add(
         array_ops.reshape(
             [1, 2], shape=[1, 2]),
         array_ops.reshape(
             [1, 2], shape=[1, 2]))

  def _testGradient(self, np_input, bias, dtype, data_format, use_gpu):
    with self.test_session(use_gpu=use_gpu):
      if data_format == "NCHW":
        np_input = self._NHWCToNCHW(np_input)
      input_tensor = constant_op.constant(
          np_input, shape=np_input.shape, dtype=dtype)
      bias_tensor = constant_op.constant(bias, shape=bias.shape, dtype=dtype)
      output_tensor = nn_ops.bias_add(
          input_tensor, bias_tensor, data_format=data_format)
      tensor_jacob_t, tensor_jacob_n = gradient_checker.compute_gradient(
          input_tensor, np_input.shape, output_tensor, np_input.shape)
      bias_jacob_t, bias_jacob_n = gradient_checker.compute_gradient(
          bias_tensor, bias.shape, output_tensor, np_input.shape)

      # Test gradient of BiasAddGrad
      bias_add_grad = gradients_impl.gradients(
          nn_ops.l2_loss(output_tensor), bias_tensor)[0]
      grad_jacob_t, grad_jacob_n = gradient_checker.compute_gradient(
          output_tensor, np_input.shape, bias_add_grad, bias.shape)

      if dtype == np.float16:
        # Compare fp16 theoretical gradients to fp32 numerical gradients,
        # since fp16 numerical gradients are too imprecise unless great
        # care is taken with choosing the inputs and the delta. This is
        # a weaker check (in particular, it does not test the op itself,
        # only its gradient), but it's much better than nothing.
        input_tensor = constant_op.constant(
            np_input, shape=np_input.shape, dtype=np.float32)
        bias_tensor = constant_op.constant(
            bias, shape=bias.shape, dtype=np.float32)
        output_tensor = nn_ops.bias_add(
            input_tensor, bias_tensor, data_format=data_format)
        _, tensor_jacob_n = gradient_checker.compute_gradient(input_tensor,
                                                              np_input.shape,
                                                              output_tensor,
                                                              np_input.shape)
        _, bias_jacob_n = gradient_checker.compute_gradient(bias_tensor,
                                                            bias.shape,
                                                            output_tensor,
                                                            np_input.shape)

        bias_add_grad = gradients_impl.gradients(
            nn_ops.l2_loss(output_tensor), bias_tensor)[0]
        _, grad_jacob_n = gradient_checker.compute_gradient(output_tensor,
                                                            np_input.shape,
                                                            bias_add_grad,
                                                            bias.shape)

      threshold = 2e-3
      if dtype == dtypes.float64:
        threshold = 1e-10
      self.assertAllClose(tensor_jacob_t, tensor_jacob_n, threshold, threshold)
      # TODO(annarev): Re-add assertion for float16, float32 dtypes and NCHW
      # once we figure out why this check started failing with cuda mavx.
      if dtype == dtypes.float64 or data_format != "NCHW":
        self.assertAllClose(bias_jacob_t, bias_jacob_n, threshold, threshold)
        self.assertAllClose(grad_jacob_t, grad_jacob_n, threshold, threshold)

 def _testBiasNCHW(self, np_inputs, np_bias, use_gpu):
   np_val = self._npBias(np_inputs, np_bias)
   np_inputs = self._NHWCToNCHW(np_inputs)
   with self.cached_session(use_gpu=use_gpu):
     tf_val = nn_ops.bias_add(np_inputs, np_bias, data_format="NCHW").eval()
   tf_val = self._NCHWToNHWC(tf_val)
   self.assertAllCloseAccordingToType(self._AtLeast3d(np_val), tf_val)

def SimulateFusedConv2dBiasActivationInt8(conv_input_scale, conv_input, kernel,
                                          padding, strides, side_input_scale,
                                          side_input, biases):
  """Simulates the int8 fused 2-D convolution op using separate float ops.

    The arguments and return values have the same format, meanings and
    restrictions as the actual op.
  Args:
    conv_input_scale: A scalar 'float'.
    conv_input: A `Tensor` of type `qint8` in NCHW_VECT_C layout.
    kernel: A `Tensor` of type `qint8` in OIHW_VECT_I layout.
    padding: A `string` from: `"SAME", "VALID"`.
    strides: A list of `ints`.
    side_input_scale: A scalar 'float'.
    side_input: A `Tensor` of type `qint8` in NCHW_VECT_C layout.
    biases: A `Tensor` of type `float32` in NCHW layout.
  Returns:
    A `Tensor` of type `qint8` in NCHW_VECT_C layout.
  """
  conv_result = nn_ops.conv2d(
      NchwVectCToNchw(gen_array_ops.dequantize(conv_input, -128, 127)),
      OihwVectIToHwio(gen_array_ops.dequantize(kernel, -128, 127)),
      strides=strides,
      padding=padding,
      data_format="NCHW") * conv_input_scale

  conv_and_side_inputs = conv_result + side_input_scale * NchwVectCToNchw(
      gen_array_ops.dequantize(side_input, -128, 127))

  logit = nn_ops.bias_add(conv_and_side_inputs, biases, data_format="NCHW")

  result, _, _ = gen_array_ops.quantize_v2(
      NchwToNchwVectC(nn_ops.relu(logit)), -128, 127, dtypes.qint8)
  return result

  def call(self, inputs, state):
    """Most basic RNN: output = new_state = act(W * input + U * state + B)."""

    gate_inputs = math_ops.matmul(
        array_ops.concat([inputs, state], 1), self._kernel)
    gate_inputs = nn_ops.bias_add(gate_inputs, self._bias)
    output = self._activation(gate_inputs)
    return output, output

def _linear(args, output_size, bias, bias_initializer=None,
            kernel_initializer=None):
  """Linear map: sum_i(args[i] * W[i]), where W[i] is a variable.

  Args:
    args: a 2D Tensor or a list of 2D, batch x n, Tensors.
    output_size: int, second dimension of W[i].
    bias: boolean, whether to add a bias term or not.
    bias_initializer: starting value to initialize the bias; None by default.
    kernel_initializer: starting value to initialize the weight; None by default.

  Returns:
    A 2D Tensor with shape [batch x output_size] equal to
    sum_i(args[i] * W[i]), where W[i]s are newly created matrices.

  Raises:
    ValueError: if some of the arguments has unspecified or wrong shape.
  """
  if args is None or (nest.is_sequence(args) and not args):
    raise ValueError("`args` must be specified")
  if not nest.is_sequence(args):
    args = [args]

  # Calculate the total size of arguments on dimension 1.
  total_arg_size = 0
  shapes = [a.get_shape() for a in args]
  for shape in shapes:
    if shape.ndims != 2:
      raise ValueError("linear is expecting 2D arguments: %s" % shapes)
    if shape[1].value is None:
      raise ValueError("linear expects shape[1] to be provided for shape %s, "
                       "but saw %s" % (shape, shape[1]))
    else:
      total_arg_size += shape[1].value

  dtype = [a.dtype for a in args][0]

  # Now the computation.
  scope = vs.get_variable_scope()
  with vs.variable_scope(scope) as outer_scope:
    weights = vs.get_variable(
        _WEIGHTS_VARIABLE_NAME, [total_arg_size, output_size], dtype=dtype,
        initializer=kernel_initializer)
    if len(args) == 1:
      res = math_ops.matmul(args[0], weights)
    else:
      res = math_ops.matmul(array_ops.concat(args, 1), weights)
    if not bias:
      return res
    with vs.variable_scope(outer_scope) as inner_scope:
      inner_scope.set_partitioner(None)
      if bias_initializer is None:
        bias_initializer = init_ops.constant_initializer(0.0, dtype=dtype)
      biases = vs.get_variable(
          _BIAS_VARIABLE_NAME, [output_size],
          dtype=dtype,
          initializer=bias_initializer)
    return nn_ops.bias_add(res, biases)

 def _linear(self, args, copy):
   out_size = copy * self._num_units
   proj_size = args.get_shape()[-1]
   weights = vs.get_variable("kernel", [proj_size, out_size])
   out = math_ops.matmul(args, weights)
   if not self._layer_norm:
     bias = vs.get_variable("bias", [out_size])
     out = nn_ops.bias_add(out, bias)
   return out

  def _SetupValuesForDevice(self, tensor_in_sizes, filter_in_sizes, bias,
                            strides, padding, activation_mode, data_format,
                            dtype):
    """Verifies the output values of the convolution function.

    Args:
      tensor_in_sizes: Input tensor dimensions in
        [batch, input_rows, input_cols, input_depth].
      filter_in_sizes: Filter tensor dimensions in
        [kernel_rows, kernel_cols, input_depth, output_depth].
      bias: 1-D bias tensor of length output_depth.
      strides: Stride: [col_stride, row_stride]
      padding: Padding type.
      activation_mode: Activation mode.
      data_format: Format of the data tensors.
      dtype: Data type for inputs and outputs.
    Returns:
      Symbolic tensor value and reference value that can be used to
      execute the computation and verify the results.
    """
    input_size = np.prod(tensor_in_sizes)
    filter_size = np.prod(filter_in_sizes)
    bias_size = filter_in_sizes[-1]  # equals to output depth
    # Initializes the input tensor with array containing incrementing
    # numbers from 1.
    x1 = [f * 1.0 for f in range(1, input_size + 1)]
    x2 = [f * 1.0 for f in range(1, filter_size + 1)]
    # This is to guarantee that there is always negative values after
    # bias add so that we can test whether relu works correctly.
    x3 = bias
    with self.test_session(use_gpu=True):
      t1 = constant_op.constant(x1, shape=tensor_in_sizes, dtype=dtype)
      t2 = constant_op.constant(x2, shape=filter_in_sizes, dtype=dtype)
      t3 = constant_op.constant(x3, shape=[bias_size], dtype=dtype)
      strides = [1] + strides + [1]
      if data_format == "NCHW":
        t1 = test_util.NHWCToNCHW(t1)
        strides = test_util.NHWCToNCHW(strides)
      output = fused_conv2d_bias_activation_op.fused_conv2d_bias_activation(
          t1,
          t2,
          t3,
          strides=strides,
          padding=padding,
          data_format=data_format,
          activation_mode=activation_mode)
      ref_conv_output = nn_ops.conv2d(
          t1, t2, strides=strides, padding=padding, data_format=data_format)
      ref_bias_output = nn_ops.bias_add(
          ref_conv_output, t3, data_format=data_format)
      ref_output = nn_ops.relu(ref_bias_output)
      if data_format == "NCHW":
        output = test_util.NCHWToNHWC(output)
        ref_output = test_util.NCHWToNHWC(ref_output)

      return output, ref_output

def build_conv_bias_relu_graph(device, input_shape, filter_shape, strides,
                               padding, num_iters, data_format):
  """builds a graph containing a sequence of conv2d operations.

  Args:
    device: String, the device to run on.
    input_shape: Shape of the input tensor.
    filter_shape: Shape of the filter tensor.
    strides: A list of ints. 1-D of length 4. The stride of sliding
             window for each dimension of input.
    padding: A string from: "SAME", "VALID". The type of padding
             algorithm to use.
    num_iters: number of iterations to run conv2d.
    data_format: data format string of input, 'NHWC' and 'NCHW' are
    supported.

  Returns:
    An array of tensors to run()
  """
  if data_format == "NCHW":
    input_shape = [
        input_shape[0], input_shape[3], input_shape[1], input_shape[2]
    ]
  with ops.device("/%s:0" % device):
    inp = variables.Variable(random_ops.truncated_normal(input_shape))
    filt = variables.Variable(random_ops.truncated_normal(filter_shape))
    bias_shape = [filter_shape[-1]]
    bias = variables.Variable(random_ops.truncated_normal(bias_shape))

    outputs = []
    conv2d_out = nn_ops.conv2d(
        inp, filt, strides, padding, data_format=data_format)
    bias_out = nn_ops.bias_add(conv2d_out, bias, data_format=data_format)
    relu_out = nn_ops.relu(bias_out)
    outputs.append(relu_out)
    for _ in range(1, num_iters):
      with ops.control_dependencies([relu_out]):
        conv2d_out = nn_ops.conv2d(
            inp, filt, strides, padding, data_format=data_format)
        bias_out = nn_ops.bias_add(conv2d_out, bias, data_format=data_format)
        relu_out = nn_ops.relu(bias_out)
        outputs.append(relu_out)
    return control_flow_ops.group(*outputs)

  def __call__(self, args):
    if not self._is_sequence:
      args = [args]

    if len(args) == 1:
      res = math_ops.matmul(args[0], self._weights)
    else:
      res = math_ops.matmul(array_ops.concat(args, 1), self._weights)
    if self._build_bias:
      res = nn_ops.bias_add(res, self._biases)
    return res

 def testGradients(self):
   with ops.Graph().as_default():
     inp = constant(1.0, shape=[32, 100], name="in")
     w = constant(1.0, shape=[100, 10], name="w")
     b = constant(1.0, shape=[10], name="b")
     xw = math_ops.matmul(inp, w, name="xw")
     h = bias_add(xw, b, name="h")
     w_grad = gradients.gradients(h, w)[0]
   self.assertEquals("MatMul", w_grad.op.type)
   self.assertEquals(w_grad.op._original_op, xw.op)
   self.assertTrue(w_grad.op.get_attr("transpose_a"))
   self.assertFalse(w_grad.op.get_attr("transpose_b"))

  def __call__(self, args):
    if not self._is_sequence:
      args = [args]

    if len(args) == 1:
      res = math_ops.matmul(args[0], self._weights)
    else:
      # Explicitly creating a one for a minor performance improvement.
      one = constant_op.constant(1, dtype=dtypes.int32)
      res = math_ops.matmul(array_ops.concat(args, one), self._weights)
    if self._build_bias:
      res = nn_ops.bias_add(res, self._biases)
    return res

  def call(self, inputs, state):
    """Long short-term memory cell (LSTM) with masks for pruning.

    Args:
      inputs: `2-D` tensor with shape `[batch_size, input_size]`.
      state: An `LSTMStateTuple` of state tensors, each shaped
        `[batch_size, self.state_size]`, if `state_is_tuple` has been set to
        `True`.  Otherwise, a `Tensor` shaped
        `[batch_size, 2 * self.state_size]`.

    Returns:
      A pair containing the new hidden state, and the new state (either a
        `LSTMStateTuple` or a concatenated state, depending on
        `state_is_tuple`).
    """
    sigmoid = math_ops.sigmoid
    one = constant_op.constant(1, dtype=dtypes.int32)
    # Parameters of gates are concatenated into one multiply for efficiency.
    if self._state_is_tuple:
      c, h = state
    else:
      c, h = array_ops.split(value=state, num_or_size_splits=2, axis=one)

    gate_inputs = math_ops.matmul(
        array_ops.concat([inputs, h], 1), self._masked_kernel)
    gate_inputs = nn_ops.bias_add(gate_inputs, self._bias)

    # i = input_gate, j = new_input, f = forget_gate, o = output_gate
    i, j, f, o = array_ops.split(
        value=gate_inputs, num_or_size_splits=4, axis=one)

    forget_bias_tensor = constant_op.constant(self._forget_bias, dtype=f.dtype)
    # Note that using `add` and `multiply` instead of `+` and `*` gives a
    # performance improvement. So using those at the cost of readability.
    add = math_ops.add
    multiply = math_ops.multiply
    new_c = add(
        multiply(c, sigmoid(add(f, forget_bias_tensor))),
        multiply(sigmoid(i), self._activation(j)))
    new_h = multiply(self._activation(new_c), sigmoid(o))

    if self._state_is_tuple:
      new_state = tf_rnn.LSTMStateTuple(new_c, new_h)
    else:
      new_state = array_ops.concat([new_c, new_h], 1)
    return new_h, new_state

def relu_layer(x, weights, biases, name=None):
  """Computes Relu(x * weight + biases).

  Args:
    x: a 2D tensor.  Dimensions typically: batch, in_units
    weights: a 2D tensor.  Dimensions typically: in_units, out_units
    biases: a 1D tensor.  Dimensions: out_units
    name: A name for the operation (optional).  If not specified
      "nn_relu_layer" is used.

  Returns:
    A 2-D Tensor computing relu(matmul(x, weights) + biases).
    Dimensions typically: batch, out_units.
  """
  with ops.op_scope([x, weights, biases], name, "relu_layer") as name:
    x = ops.convert_to_tensor(x, name="x")
    weights = ops.convert_to_tensor(weights, name="weights")
    biases = ops.convert_to_tensor(biases, name="biases")
    xw_plus_b = nn_ops.bias_add(math_ops.matmul(x, weights), biases)
    return nn_ops.relu(xw_plus_b, name=name)

def xw_plus_b(x, weights, biases, name=None):
  """Computes matmul(x, weights) + biases.

  Args:
    x: a 2D tensor.  Dimensions typically: batch, in_units
    weights: a 2D tensor.  Dimensions typically: in_units, out_units
    biases: a 1D tensor.  Dimensions: out_units
    name: A name for the operation (optional).  If not specified
      "wx_plus_b" is used.

  Returns:
    A 2-D Tensor computing matmul(x, weights) + biases.
    Dimensions typically: batch, out_units.
  """
  with ops.op_scope([x, weights, biases], name, "xw_plus_b") as name:
    x = ops.convert_to_tensor(x, name="x")
    weights = ops.convert_to_tensor(weights, name="weights")
    biases = ops.convert_to_tensor(biases, name="biases")
    mm = math_ops.matmul(x, weights)
    return nn_ops.bias_add(mm, biases, name=name)

    def __call__(self, inputs, state, scope=None):
        num_proj = self._num_units if self._num_proj is None else self._num_proj

        if self._state_is_tuple:
            (c_prev,m_prev) = state
        else:
            c_prev = array_ops.slice(state, [0, 0], [-1, self._num_units])
            m_prev = array_ops.slice(state, [0, self._num_units], [-1, num_proj])

        dtype = inputs.dtype
        input_size = inputs.get_shape().with_rank(2)[1]
        if input_size.value is None:
            raise ValueError("Could not infer input size from inputs.get_shape()[-1]")
        with vs.variable_scope(scope or type(self).__name__,
                               initializer=self._initializer):
            concat_w = tf.nn.rnn_cell._get_concat_variable(
                "W", [input_size.value + num_proj, 3 * self._num_units],
                dtype, self._num_unit_shards)

            b = vs.get_variable(
                "B", shape=[3 * self._num_units],
                initializer=init_ops.zeros_initializer, dtype=dtype)

            cell_inputs = array_ops.concat(1,[inputs, m_prev])
            ltm_matrix = nn_ops.bias_add(math_ops.matmul(cell_inputs, concat_w), b)
            i,j,o = array_ops.split(1,3,ltm_matrix) # i,j,o: [1,num_units]
            c = c_prev + sigmoid(i)*self._activation(j)
            if self._cell_clip is not None:
                c = clip_ops.clip_by_value(c, -self._cell_clip, self._cell_clip)
            m = sigmoid(o) * self._activation(c)
            if self._num_proj is not None:
                concat_w_proj = tf.nn.rnn_cell._get_concat_variable(
                                "W_P", [self._num_units, self._num_proj],
                                dtype, self._num_proj_shards)
                m = math_ops.matmul(m, concat_w_proj)
                if self._proj_clip is not None:
                    m = clip_ops.clip_by_value(m, -self._proj_clip, self._proj_clip)
        new_state = (tf.nn.rnn_cell.LSTMStateTuple(c,m) if self._state_is_tuple
                     else array_ops.concat(1,[c,m]))
        return m, new_state

 def call(self, inputs, state):
   """Most basic RNN: output = new_state = act(W * input + U * state + B)."""
   inputs = self._tflite_wrapper.add_input(
       inputs, tag="input", name="input", aggregate="stack", index_override=0)
   state = self._tflite_wrapper.add_input(
       state,
       tag="hidden_state",
       name="hidden_state",
       aggregate="first",
       index_override=4)
   weights = array_ops.transpose(
       array_ops.concat([self._input_weights, self._recurrent_weights], 1))
   gate_inputs = math_ops.matmul(array_ops.concat([inputs, state], 1), weights)
   gate_inputs = nn_ops.bias_add(gate_inputs, self._bias)
   output = self._activation(gate_inputs)
   output = self._tflite_wrapper.add_output(
       output,
       tag="output",
       name="output",
       index_override=1,
       aggregate="stack")
   return output, output

def _SimulateFusedConv2dBiasActivationInt8(conv_input_scale, conv_input, kernel,
                                           padding, strides, side_input_scale,
                                           side_input, biases, apply_relu):
  """Simulates the int8 fused 2-D convolution op using separate float ops.

    The arguments and return values have the same format, meanings and
    restrictions as the actual op.
  Args:
    conv_input_scale: A scalar 'float'.
    conv_input: A `Tensor` of type `qint8` in NCHW_VECT_C layout.
    kernel: A `Tensor` of type `qint8` in OIHW_VECT_I layout.
    padding: A `string` from: `"SAME", "VALID"`.
    strides: A list of `ints`.
    side_input_scale: A scalar 'float'.
    side_input: A `Tensor` of type `qint8` in NCHW_VECT_C layout.
    biases: A `Tensor` of type `float32` in NCHW layout.
    apply_relu: A boolean to specify whether to apply "Relu" activation function
      that clips outputs to the range [0, 127], or "None" activation that clips
      to the range [-128, 127].
  Returns:
    A `Tensor` of type `qint8` in NCHW_VECT_C layout.
  """
  conv_result = nn_ops.conv2d(
      _NchwVectCToNchw(gen_array_ops.dequantize(conv_input, -128, 127)),
      _OihwVectIToHwio(gen_array_ops.dequantize(kernel, -128, 127)),
      strides=strides,
      padding=padding,
      data_format="NCHW") * conv_input_scale

  conv_and_side_inputs = conv_result + side_input_scale * _NchwVectCToNchw(
      gen_array_ops.dequantize(side_input, -128, 127))

  output = nn_ops.bias_add(conv_and_side_inputs, biases, data_format="NCHW")
  if apply_relu:
    output = nn_ops.relu(output)

  result, _, _ = gen_array_ops.quantize_v2(
      _NchwToNchwVectC(output), -128, 127, dtypes.qint8)
  return result

  def call(self, inputs, state):
    """Run one time step of the IndRNN.

    Calculates the output and new hidden state using the IndRNN equation

      `output = new_state = act(W * input + u (*) state + b)`

    where `*` is the matrix multiplication and `(*)` is the Hadamard product.

    Args:
      inputs: Tensor, 2-D tensor of shape `[batch, num_units]`.
      state: Tensor, 2-D tensor of shape `[batch, num_units]` containing the
        previous hidden state.

    Returns:
      A tuple containing the output and new hidden state. Both are the same
        2-D tensor of shape `[batch, num_units]`.
    """
    gate_inputs = math_ops.matmul(inputs, self._input_kernel)
    recurrent_update = math_ops.multiply(state, self._recurrent_kernel)
    gate_inputs = math_ops.add(gate_inputs, recurrent_update)
    gate_inputs = nn_ops.bias_add(gate_inputs, self._bias)
    output = self._activation(gate_inputs)
    return output, output

def _test_fully_connected(tensor_in_sizes, filter_in_sizes, bias_in_size=None):
    """ One iteration of fully connected """

    total_size_1 = 1
    total_size_2 = 1
    for s in tensor_in_sizes:
        total_size_1 *= s
    for s in filter_in_sizes:
        total_size_2 *= s
    # Initializes the input tensor with array containing incrementing
    # numbers from 1.
    data_array = [f * 1.0 for f in range(1, total_size_1 + 1)]
    filter_array = [f * 1.0 for f in range(1, total_size_2 + 1)]
    assert int(total_size_1 / tensor_in_sizes[0]) == filter_in_sizes[0], \
        "input size and filter size are mismatched"

    with tf.Graph().as_default():
        in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype='float32')
        in_filter = constant_op.constant(filter_array, shape=filter_in_sizes, dtype='float32')

        # reshape N H W C into N H*W*C
        in_data_reshape = array_ops.reshape(in_data, [tensor_in_sizes[0], -1])

        out = math_ops.mat_mul(in_data_reshape, in_filter)

        # if we have bias
        if bias_in_size:
            assert bias_in_size[0] == filter_in_sizes[1], "bias and filter size are mismatched"
            bias_array = [f * 1.0 for f in range(1, bias_in_size[0] + 1)]
            in_bias = constant_op.constant(bias_array, shape=bias_in_size, dtype='float32')
            out = nn_ops.bias_add(out, in_bias)

        tflite_data_array = np.reshape(data_array, tensor_in_sizes).astype('float32')
        tvm_data_array = np.transpose(tflite_data_array, axes=(0, 3, 1, 2))
        compare_tflite_with_tvm(tflite_data_array, tvm_data_array,
                                'Placeholder:0', [in_data], [out])