// 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

  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic
  // allocation/scheduling.
  //
  // TODO(allenl): Keep a map from outer thread to list of inner threads rather
  // than a single list of threads so aliased nested parallel devices don't

//  outputs - array of TF_Outputs that specify the outputs of the function.
//            If `noutputs` is zero (the function returns no outputs), `outputs`
//            can be null. `outputs` can contain the same tensor more than once.
//  output_names - The names of the function's outputs. `output_names` array
//                 must either have the same length as `outputs`
//                 (i.e. `noutputs`) or be null. In the former case,

// tensors in `retvals`.
//
// 'retvals' must point to a pre-allocated array of TFE_TensorHandle* and
// '*num_retvals' should be set to the size of this array. It is an error if
// the size of 'retvals' is less than the number of outputs. This call sets
// *num_retvals to the number of outputs.
//
// If async execution is enabled, the call may simply enqueue the execution

  /// Note: the `buffer` argument is NOT a null terminated string!
  ///
  /// Plugins:
  ///   * Must set `status` to `TF_OK` if exactly `n` bytes have been read.
  ///   * Must set `status` to `TF_OUT_OF_RANGE` if fewer than `n` bytes have
  ///     been read due to EOF.
  ///   * Must return -1 for any other error and must set `status` to any
  ///     other value to provide more information about the error.

  // Unwatches this tensor on the tape. Mainly used for cleanup when deleting
  // eager tensors.
  void DeleteTrace(const AbstractTensorHandle*);

  // Consumes the internal state of the tape (so cannot be called more than
  // once unless the tape is persistent) and produces the gradient of the target
  // tensors with respect to the source tensors. The output gradients are used

      const std::function<void(BackwardFunction*)>& backward_function_deleter);

  void DeleteTrace(int64_t tensor_id);

  // Consumes the internal state of the tape (so cannot be called more than
  // once) and produces the gradient of the target tensors with respect to the
  // source tensors. The output gradients are used if not empty and not
  // null. The result is populated with one tensor per target element.

TF_CAPI_EXPORT void TF_SetXlaMinClusterSize(int size);

// Gets/Sets TF/XLA flag for whether(true) or not(false) to disable constant
// folding. This is for testing to ensure that XLA is being tested rather than
// Tensorflow's CPU implementation through constant folding.
TF_CAPI_EXPORT unsigned char TF_GetXlaConstantFoldingDisabled();
TF_CAPI_EXPORT void TF_SetXlaConstantFoldingDisabled(
    unsigned char should_enable);