//            converted to TF_Function.
//  fn_name - the name of the new TF_Function. Should match the operation
//            name (OpDef.name) regexp [A-Z][A-Za-z0-9_.\\-/]*.
//            If `append_hash_to_fn_name` is false, `fn_name` must be distinct
//            from other function and operation names (at least those
//            registered in graphs where this function will be used).

//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//
// All component tensors must have the same dtype. Currently they must also have
// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see

// Create a variable handle with name `variable_name` on a device with name
// `device_name`.
TFE_TensorHandle* CreateVarHandle(TFE_Context* ctx,
                                  const tensorflow::string& device_name,
                                  const tensorflow::string& variable_name);

// Create a variable with value `value` and name `variable_name` on a device
// with name `device_name`.

  // Registers a function with this context, after this the function is
  // available to be called/referenced by its name in this context.
  virtual Status RegisterFunction(AbstractFunction*) = 0;
  // Remove a function. 'func' argument is the name of a previously added
  // FunctionDef. The name is in fdef.signature.name.
  virtual Status RemoveFunction(const string& func) = 0;

 private:
  const AbstractContextKind kind_;
};

// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",
// "description", "label");
// and then incremented that counter for various values of label:
// streamz->GetCell("label-value")->IncrementBy(1);
//
// The test code can then read and test the value of that counter:
//
// auto* reader = TFE_MonitoringNewCounterReader("name");
// test();
// int64_t value = TFE_MonitoringReadCounter1(reader, "label-value");

    TFE_Context* context, const char* device_name,
    const std::array<const char*, num_devices>& underlying_devices,
    TF_Status* status) {
  TFE_CustomDevice device;
  void* device_info;
  tensorflow::parallel_device::AllocateParallelDevice(
      device_name, underlying_devices.data(), underlying_devices.size(),
      &device, &device_info);
  TFE_RegisterCustomDevice(context, device, device_name, device_info, status);
}

// Retrieves a list of directory names on the local machine that may be used for
// temporary storage. You can iterate through the list with TF_StringStreamNext.
// The caller is responsible for freeing the list (see TF_StringStreamDone).
TF_CAPI_EXPORT extern TF_StringStream* TF_GetLocalTempDirectories(void);

// Creates a temporary file name with an extension.

  //
  // Otherwise if all of the tensors have the same shape, returns that via the
  // `shape` output argument. This blocks waiting for async tensors, may return
  // a delayed bad status encountered during async execution, and will return a
  // bad status unless all tensors have the same shape.
  Status Shape(const std::vector<int64_t>** shape) const;

                 const char* name = nullptr,
                 const char* raw_device_name = nullptr);

// Returns a tensor of zeros with the same shape and type as x.
Status ZerosLike(AbstractContext* ctx, AbstractTensorHandle* const x,
                 AbstractTensorHandle** y, const char* name = nullptr,
                 const char* raw_device_name = nullptr);

// Returns the shape of a tensor.

extern "C" {
#endif

// Resets `op_to_reset` with `op_or_function_name` and `raw_device_name`. This
// 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