#ifndef TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_
#define TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.

==============================================================================*/

#ifndef TENSORFLOW_C_C_API_MACROS_H_
#define TENSORFLOW_C_C_API_MACROS_H_

#ifdef SWIG
#define TF_CAPI_EXPORT
#else
#if defined(_WIN32)
#ifdef TF_COMPILE_LIBRARY
#define TF_CAPI_EXPORT __declspec(dllexport)
#else
#define TF_CAPI_EXPORT __declspec(dllimport)
#endif  // TF_COMPILE_LIBRARY
#else
#ifdef TF_CAPI_WEAK

  /// Look up the entry with key `key` and copy it to `value` if found. If not
  /// found, call `compute_func`. If `compute_func` set `status` to `TF_OK`,
  /// store a copy of the output parameter in the cache, and another copy in
  /// `value`.
  void LookupOrCompute(const std::string& key, T* value,
                       const ComputeFunc& compute_func, TF_Status* status) {
    if (max_age_ == 0) {
      return compute_func(key, value, status);

// is for performance optimization by reusing an exiting unused op rather than
// creating a new op every time. If `raw_device_name` is `NULL` or empty, it
// does not set the device name. If it's not `NULL`, then it attempts to parse
// and set the device name. It's effectively `TFE_OpSetDevice`, but it is faster
// than separately calling it because if the existing op has the same
// `raw_device_name`, it skips parsing and just leave as it is.

    TF_GraphProperties* graph_properties);

// Infer tensor shapes through abstract interpretation.
// If assume_valid_feeds is true, it can help infer shapes in the fanout of fed
// nodes. This may cause incorrectness in graph analyses, but is useful for
// simulation or scheduling.
// If aggressive_shape_inference is true, nodes are executed on the host to

                          ImmediateExecutionTensorHandle* input) = 0;

  virtual ImmediateExecutionContext* GetContext() const = 0;

  // Following two methods are used to support custom device.
  // Return true if the inputs contain custom device tensor handle. It means
  // that the argument need to be handled by a custom device.
  virtual bool HasCustomDeviceInput() const = 0;

  virtual const tensorflow::OpDef* OpDef() const = 0;

namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and
 * parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *

// Append a tensor at the end of the output list, growing its size by one.
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status*);

// TF_ExecuteOperation will, if in eager mode, execute, if in graph mode, maybe
// capture some inputs and then add a node in the graph. The output tensors are
// returned through the provided TF_OutputList.
// Any active tape will observe the effects of this execution.

  auto it = tensor_usage_.find(tensor_id);
  if (it == tensor_usage_.end()) {
    return;
  }
  it->second--;
  if (it->second != 0) {
    return;
  }
  tensor_usage_.erase(it);
  auto tensor_op_it = tensor_tape_.find(tensor_id);
  if (tensor_op_it == tensor_tape_.end()) {
    return;
  }
  const int64_t op_id = tensor_op_it->second;
  if (op_id == -1) {
    // Do not delete watched tensors.

// run on device_type.
TF_CAPI_EXPORT extern void TF_AttrBuilderCheckCanRunOnDevice(
    TF_AttrBuilder* builder, const char* device_type, TF_Status* status);

// For argument number input_index, fetch the corresponding number_attr that
// needs to be updated with the argument length of the input list.
// Returns nullptr if there is any problem like op_name is not found, or the