// APIs for generically dealing with op attributes (e.g. when forwarding them
// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

// A reference to an op's name -> attribute mapping

void TF_DeleteAbstractTensor(TF_AbstractTensor*);

// TF_OutputList holds the list of TF_AbstractTensor that results from executing
// an operation, or provided to create a function.
// When executing an operation in an eager context, the expected number of
// outputs must be set beforehand with `TF_OutputListSetNumOutputs`.
typedef struct TF_OutputList TF_OutputList;
TF_OutputList* TF_NewOutputList();

// 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
// identify output values when possible and does other aggressive strategies.
// This may cause incorrectness in graph analyses, but is useful for simulation
// or scheduling.
// If include_input_tensor_values is true, the values of constant

class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // (1) If type == TF_ATTR_STRING and is_list == 0
  //     then total_size is the byte size of the string
  //     valued attribute.
  // (2) If type == TF_ATTR_STRING and is_list == 1
  //     then total_size is the cumulative byte size
  //     of all the strings in the list.
  // (3) If type == TF_ATTR_SHAPE and is_list == 0
  //     then total_size is the number of dimensions
  //     of the shape valued attribute, or -1

// Returns a new thread that is running work_func and is identified
// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread
// starts. In this way, data may be passed from the thread back to the caller.
//
// Caller takes ownership of the result and must call TF_JoinThread on it
// eventually.

// 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();

//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"
// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",
//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw

// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update
// the gradients of its inputs. Once we have computed all gradients for a single
// input we can mark this input as done, and this can trigger adding an entry to
// the stack if all outputs of that entry are now done.
//
// When the stack is empty we have gradients for all tensors we're interested
// in.

//
// The API here is subject to changes in the future.
// --------------------------------------------------------------------------

#ifdef __cplusplus
extern "C" {
#endif

// When `enable` is true, set
// tensorflow.ConfigProto.OptimizerOptions.global_jit_level to ON_1, and also
// set XLA flag values to prepare for XLA compilation. Otherwise set
// global_jit_level to OFF.
//