#include "tensorflow/c/conversion_macros.h"
#include "tensorflow/c/eager/immediate_execution_context.h"

// Wraps a pointer to a context implementation.
//
// WARNING: Since the underlying object could be ref-counted a user of this
// interface cannot destruct the underlying context object. Instead, call
// TFE_DeleteContext who calls Release() on the context pointer and deletes
// the TFE_Context structure.
typedef struct TFE_Context TFE_Context;

#include "tensorflow/c/conversion_macros.h"
#include "tensorflow/c/eager/immediate_execution_operation.h"

// Wraps a pointer to an operation implementation.
//
// WARNING: Since the underlying object could be ref-counted a user of this
// interface cannot destruct the underlying operation object. Instead, call
// TFE_DeleteOp who calls Release() on the operation pointer and deletes
// the TFE_Op structure.
typedef struct TFE_Op TFE_Op;

#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"

// Wraps a pointer to a tensor handle implementation.
//
// WARNING: Since the underlying object could be ref-counted a user of this
// interface cannot destruct the underlying handle object. Instead, call
// TFE_DeleteTensorHandle who calls Release() on the handle pointer and deletes
// the TFE_TensorHandle structure.

  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};
  RegisterParallelDevice(context.get(), device_name, underlying_devices,
                         status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

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
//   "/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"}

                                            TF_Status* status) const;

  // The number of devices operations run on.
  size_t num_underlying_devices() const { return underlying_devices_.size(); }

  // The devices operations run on.
  const std::vector<std::string>& underlying_devices() const {
    return underlying_devices_;
  }

  // Takes a description of a single operation being executed on the

  std::vector<TensorHandlePtr> components;
  components.reserve(underlying_devices_.size());
  for (const std::string& underlying_device_name : underlying_devices_) {
    TFE_TensorHandle* t = TFE_TensorHandleCopyToDevice(
        tensor, context, underlying_device_name.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    components.emplace_back(t);
  }

  const char* second_device_name =
      "/job:localhost/replica:0/task:0/device:CPU:1";
  std::array<const char*, 2> underlying_devices{first_device_name,
                                                second_device_name};
  RegisterParallelDevice(context.get(), device_name, underlying_devices,
                         status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  /// reads from the underlying filesystem.
  const size_t block_size_;
  /// The maximum number of bytes (sum of block sizes) allowed in the LRU cache.
  const size_t max_bytes_;
  /// The maximum staleness of any block in the LRU cache, in seconds.
  const uint64_t max_staleness_;
  /// The callback to read a block from the underlying filesystem.
  const BlockFetcher block_fetcher_;

  std::vector<std::string> underlying_devices_vector;
  underlying_devices_vector.reserve(num_underlying_devices);
  for (int device_index = 0; device_index < num_underlying_devices;
       ++device_index) {
    underlying_devices_vector.push_back(underlying_devices[device_index]);
  }
  std::unique_ptr<ParallelDevice> parallel_device(
      new ParallelDevice(underlying_devices_vector));
  *device_info =