TF_Status* status = TF_NewStatus();
  TF_NewWritableFile(full_path, &h, status);
  if (TF_GetCode(status) != TF_OK) {
    fprintf(stderr, "TF_NewWritableFile failed: %s\n", TF_Message(status));
    return 1;
  }
  fprintf(stderr, "wrote %s\n", full_path);
  free(full_path);
  TF_CloseWritableFile(h, status);
  if (TF_GetCode(status) != TF_OK) {
    fprintf(stderr, "TF_CloseWritableFile failed: %s\n", TF_Message(status));

  // tasks to initialize the eager context and TF server for multi-client
  // execution.
  virtual Status EnableCollectiveOps(const ServerDef& server_def) = 0;

  // Check if the remote task is alive.
  virtual Status CheckRemoteAlive(const std::string& remote_task_name,
                                  bool* is_alive) = 0;

  // Get pointer to the coordination service agent instance.

typedef struct TF_AttrMetadata {
  // A boolean: 1 if the attribute value is a list, 0 otherwise.
  unsigned char is_list;

  // Length of the list if is_list is true. Undefined otherwise.
  int64_t list_size;

  // Type of elements of the list if is_list != 0.
  // Type of the single value stored in the attribute if is_list == 0.
  TF_AttrType type;

  // Total size the attribute value.

TFE_TensorHandle* TestAxisTensorHandle(TFE_Context* ctx);

// Return an op taking minimum of `input` long `axis` dimension.
TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis);

// If there is a device of type `device_type`, returns true
// and sets 'device_name' accordingly.
// `device_type` must be either "GPU" or "TPU".
bool GetDeviceName(TFE_Context* ctx, tensorflow::string* device_name,

  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.

{
	int i, max;

	max = *(int*)p;
	for(i=0; i<max; i++)
		goDummy();
	return NULL;
}

int
callGoInCThread(int max)
{
	pthread_t thread;

	if (pthread_create(&thread, NULL, goDummyCallbackThread, (void*)(&max)) != 0)
		return -1;
	if (pthread_join(thread, NULL) != 0)
		return -1;

	return max;

// 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.

// Set the shapes and types of the output's handle.
//
// The lengths of the arrays pointed to by `shapes`, `ranks`, and `types` must
// all be equal to `num_shapes_and_types`. If `ranks[i] != -1`, (i.e., if the
// rank is known), then it must be equal to the length of `shapes[i]`; if
// `ranks[i] == 1`, then `shapes[i]` may be nullptr.
//
// TODO(akshayka): Implement a corresponding getter method.

                                   absl::Span<AbstractTensorHandle*>)>;

// Runs `model` maybe wrapped in a function call op. This can be thought as
// being equivalent to the following python code.
//
// if use_function:
//   outputs = tf.function(model)(inputs)
// else:
//   outputs = model(inputs)
Status RunModel(Model model, AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,

{
	int i, max;
	
	max = *(int*)p;
	for(i=0; i<max; i++)
		Add(i);
	return 0;
}

void
doAdd(int max, int nthread)
{
	enum { MaxThread = 20 };
	int i;
	uintptr_t thread_id[MaxThread];
	
	if(nthread > MaxThread)
		nthread = MaxThread;
	for(i=0; i<nthread; i++)
		thread_id[i] = _beginthreadex(0, 0, addThread, &max, 0, 0);
	for(i=0; i<nthread; i++) {
		WaitForSingleObject((HANDLE)thread_id[i], INFINITE);