// tensors/ops/etc. and usable in APIs like OpSetDevice/ResetOp/etc.

#define TFE_CUSTOM_DEVICE_VERSION 4

// Struct to be filled in. Functions are required except where indicated.
typedef struct TFE_CustomDevice {
  int version = TFE_CUSTOM_DEVICE_VERSION;
  // Method to copy a tensor to the custom device.
  TFE_TensorHandle* (*copy_tensor_to_device)(TFE_Context* context,

typedef struct TF_AbstractFunction TF_AbstractFunction;

// This allows the client to swap the implementation of the tracing engine.
// Any future call to TF_CreateFunction will use the implementation defined
// here.
void TF_SetTracingImplementation(const char* name, TF_Status*);

// Creates a new TensorFlow function. A Function is an execution context, and as
// such it can trace operations through TF_ExecuteOperation. After completing

TF_CAPI_EXPORT extern void TF_AddControlInput(TF_OperationDescription* desc,
                                              TF_Operation* input);

// Request that `desc` be co-located on the device where `op`
// is placed.
//
// Use of this is discouraged since the implementation of device placement is
// subject to change. Primarily intended for internal libraries

  // used to cancel all other operations if any fails.
  //
  // Set step_id to configure the step id used for rendezvous creation. step id
  // of value -1 is reserved for global rendezvous and should not be set here.
  void StartExecute(TFE_Context* context,
                    const std::vector<ParallelTensor*>& inputs,
                    const char* operation_name, const TFE_OpAttrs* attributes,

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,

TF_CAPI_EXPORT unsigned char TF_GetXlaConstantFoldingDisabled();
TF_CAPI_EXPORT void TF_SetXlaConstantFoldingDisabled(
    unsigned char should_enable);

// Create a serialized tensorflow.ConfigProto proto, where:
//
// a) ConfigProto.optimizer_options.global_jit_level is set to ON_1 if
// `enable_xla_compilation` is non-zero, and OFF otherwise.
// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.

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

  // Returns true if `Accumulate` is active somewhere above on the stack and
  // there isn't an intervening PushState. This is useful for ordering
  // ForwardAccumulators, where more deeply nested accumulators should not see
  // computations from less deeply nested accumulators.
  bool BusyAccumulating() const { return call_state_.top().accumulating; }

//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.
// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

TF_CAPI_EXPORT extern TFE_Op* TFE_NewOp(TFE_Context* ctx,
                                        const char* op_or_function_name,

// Macro to verify that the field `struct_size` of STRUCT_OBJ is initialized.
// `struct_size` is used for struct member compatibility check between core TF
// and plug-ins with the same C API minor version. More info here:
// https://github.com/tensorflow/community/blob/master/rfcs/20200612-stream-executor-c-api/C_API_versioning_strategy.md
#define TF_VALIDATE_STRUCT_SIZE(STRUCT_NAME, STRUCT_OBJ, SIZE_VALUE_NAME) \

/// data in the structures here to prevent cases when padding/packing enabled by
/// different compiler options breaks compatibility. For more details about how
/// this is used, please consult next sections. Here we just wrap these tables
/// in lint warnings so that changes here cause changes to the versioning data
/// as well. Here is a short summary of what changes are allowed: