TF_Tensor* orig = TFE_TensorHandleResolve(n, status);
  void* orig_ptr = TF_TensorData(orig);
  TF_DeleteTensor(orig);

  TFE_Op* add_op = AddOp(ctx, n, m);
  std::string cpu_device_name;
  ASSERT_TRUE(GetDeviceName(ctx, &cpu_device_name, "CPU"));
  TFE_OpSetDevice(add_op, cpu_device_name.c_str(), status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  if (forward_input) {
    TFE_DeleteTensorHandle(n);

  TFE_DeleteOp(op);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  CHECK_EQ(0, num_retvals);

  TF_DeleteStatus(status);

  return var_handle;
}

TFE_Op* AddOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "AddV2", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

TF_Buffer* TF_GetAllOpList() {
  std::vector<tensorflow::OpDef> op_defs;
  tensorflow::OpRegistry::Global()->GetRegisteredOps(&op_defs);
  tensorflow::OpList op_list;
  for (const auto& op : op_defs) {
    *(op_list.add_op()) = op;
  }
  TF_Buffer* ret = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(op_list, ret));
  return ret;
}

// --------------------------------------------------------------------------

<pre class="ebnf">
Expression = UnaryExpr | Expression binary_op Expression .
UnaryExpr  = PrimaryExpr | unary_op UnaryExpr .

binary_op  = "||" | "&amp;&amp;" | rel_op | add_op | mul_op .
rel_op     = "==" | "!=" | "&lt;" | "&lt;=" | ">" | ">=" .
add_op     = "+" | "-" | "|" | "^" .
mul_op     = "*" | "/" | "%" | "&lt;&lt;" | "&gt;&gt;" | "&amp;" | "&amp;^" .

unary_op   = "+" | "-" | "!" | "^" | "*" | "&amp;" | "&lt;-" .

    *   Added reference implementation for 16-bit int unquantized `add`.
    *   Added reference implementation for 16-bit int and 32-bit unsigned int unquantized `mul`.
    *   `add_op` supports broadcasting up to 6 dimensions.
    *   Added 16-bit support for `top_k`.

*   `tf.function`