#include "tensorflow/c/c_api.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"

namespace tensorflow {
namespace parallel_device {

// Allocate a parallel device named `device_name` which forwards operations to
// `underlying_devices`, maintaining "parallel tensors" with components placed
// on each underlying device.
//
// For example if `device_name` is

                                                 const char* data,
                                                 size_t length,
                                                 TF_Status* status);

// Deletes the named file and indicates whether successful in *status.
TF_CAPI_EXPORT extern void TF_DeleteFile(const char* filename,
                                         TF_Status* status);

// If `dx` is nullptr, the implementation will use dx of `OnesLike` for all
// shapes in `y`.
// The partial derivatives are returned in `dy`. `dy` should be allocated to
// size `nx`.
//
// Gradient nodes are automatically named under the "gradients/" prefix. To
// guarantee name uniqueness, subsequent calls to the same graph will
// append an incremental tag to the prefix: "gradients_1/", "gradients_2/", ...

// "/job:localhost/replica:0/task:0/device:CUSTOM:0".
//
// The custom device defines copy operations for moving TensorHandles on and
// off, and an execution operation for named operations. Often execution will
// simply wrap op execution on one or more physical devices.
//
// device_info is an opaque caller-defined type stored with the custom device

                                                   size_t* len);

// On success, dequeues a tensor from a TF-managed FifoQueue given by
// `tensor_id`, associated with `session`. There must be a graph node named
// "fifo_queue_dequeue_<tensor_id>", to be executed by this API call.

// Caller must call TF_DeleteTensor() over the returned tensor. If the queue is
// empty, this call is blocked.
//

template <typename BackwardFunction, typename TapeTensor>
using OpTape =
    std::unordered_map<int64_t, OpTapeEntry<BackwardFunction, TapeTensor>>;

// Operations the tape needs to perform on tensors to do backpropagation. Named
// "vspace" because a subset of these are related to a vector space, such as
// adding gradients, getting zeroes, etc. Currently cannot be implemented
// without using tensorflow python code, hence left unspecified here.
//

// already-finalized operation.
//
// Note that TFE_OpGetFlatInputCount and TFE_OpGetFlatInput operate on a flat
// sequence of inputs, unlike TFE_OpGetInputLength (for getting the length of a
// particular named input list, which may only be part of the op's inputs).
TF_CAPI_EXPORT extern int TFE_OpGetFlatInputCount(const TFE_Op* op,
                                                  TF_Status* status);

  // partitioned tensor are combined into a single entry.
  const TensorSliceReader::VarToDataTypeMap& GetVariableToDataTypeMap() const;

  // Attempts to look up the tensor named "name" and stores the found result in
  // "out_tensor".
  void GetTensor(const string& name,
                 std::unique_ptr<tensorflow::Tensor>* out_tensor,
                 TF_Status* out_status) const;

 private:

typedef struct TF_Filesystem {
  void* plugin_filesystem;
} TF_Filesystem;

typedef struct TF_TransactionToken {
  void* token;
  TF_Filesystem* owner;
} TF_TransactionToken;

// The named union is needed here (as opposed to
// inside the `TF_Filesystem_Option_Value` struct)
// as MSVC does not recognize `typeof`.
typedef union TF_Filesystem_Option_Value_Union {
  int64_t int_val;
  double real_val;