//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation
//   function.  This deallocation function will point to client code
//   for tensors populated by the client.  So the client can do things
//   like shadowing a numpy array.

// 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
// which is passed to the functions referenced in the TFE_CustomDevice struct
// `device` (execute, delete_device, etc.). It can for example contain the
// names of wrapped devices.
//

// function and deleted (as the backprop code creates lots of gradients the user
// is not interested in).
//
// BackwardFunction needs to be a closure which stores intermediate activations
// from the forward computation and calls a vector-jacobian product function
// (also known as adjoint function) to compute, given downstream gradients,
// upstream gradients.
//

// `GradientTape::Watch`.) The op_id is simply a unique index assigned to each
// op executed under the tape. A separate map (`tensorflow::eager::OpTape`)
// maintains the map from `op_id` to a `OpTapeEntry` which stores the `op_type`,
// inputs and outputs and the gradient function These data structures combined
// allow us to trace the data dependencies between operations and hence compute
// gradients.
//

  // 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: