*
     *    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("data_format", &data_format));

    // Grad for A
    grad_inputs[0] = upstream_grad;

                         TF_Status* status) {
  status->status = absl::OkStatus();
  for (int i = 0; i < noutputs; ++i) {
    c_outputs[i] = nullptr;
  }
}

// TF_TensorToTensorV1 decodes a string serialization to DT_RESOURCE.
// In the TFv1 convention, TF_Tensor can hold a string serialization of
// DT_RESOURCE. The string serialization is converted back to a

    r = JUnitXml.fromfile(f)
  except Exception as e:  # pylint: disable=broad-except
    print("Ignoring this XML parse failure in {}: ".format(f), str(e))

  source_file = re.search(
      r"/(bazel_pip|tensorflow)/.*", f.decode("utf-8")
  ).group(0)
  for testsuite in r:
    testsuite._elem.set("source_file", source_file)
    # Remove empty testcases
    for p in testsuite._elem.xpath(".//testcase"):

    per_device_output_tensors.push_back(device_thread->Join(status));
    // We will run every Join even if there are bad statuses in case the user
    // wants to recover and continue running ops on the parallel device (which
    // would otherwise deadlock).
    if (TF_GetCode(status) != TF_OK &&
        (first_bad_status == nullptr

      output = subprocess.check_output(cmd, stderr=stderr)
    except subprocess.CalledProcessError as e:
      output = e.output
  else:
    output = subprocess.check_output(cmd, stderr=stderr)
  return output.decode('UTF-8').strip()


def cygpath(path):
  """Convert path from posix to windows."""
  return os.path.abspath(path).replace('\\', '/')


def get_python_path(environ_cp, python_bin_path):

  // Not owned; provides operations on Tensors which are currently only
  // available in language bindings (e.g. Python).
  const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace_;

  // Decides if tangents are vectorized or not
  bool use_batch_;

  struct AccumulatorCallState {
    AccumulatorCallState(
        GradientTape<Gradient, BackwardFunction, TapeTensor>* backward_tape,
        bool accumulating)

-#  endif
-# endif
-
-/* GCC < 3.0 optimizes memset(s, 0, n) but not bzero(s, n).
-   The optimization is broken before EGCS 1.1.
-   GCC 3.0+ has __builtin_bzero as well, but at least till GCC 3.4
-   if it decides to call the library function, it calls memset
-   and not bzero.  */
-# if __GNUC_PREREQ (2, 91)
-#  define __bzero(s, n) __builtin_memset (s, '\0', n)
-# endif
-

                                           const char* path,
                                           TF_TransactionToken** token,
                                           TF_Status* status);

  /// Decodes transaction token in `token` to human readable format for
  /// debugging.
  ///
  /// A new `char*` buffer must be allocated by this method. Core TensorFlow

        `tf.debugging.is_inf`, `tf.debugging.is_nan`.
    *   New endpoints in `tf.dtypes` namespace: `tf.dtypes.as_string`.
    *   New endpoints in `tf.io` namespace: `tf.io.decode_base64`,
        `tf.io.decode_compressed`, `tf.io.decode_json_example`,
        `tf.io.decode_raw`, `tf.io.encode_base64`, `tf.io.matching_files`,
        `tf.io.parse_tensor`, `tf.io.read_file,`tf.io.write_file`.