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());

  };
  explicit AbstractOperation(AbstractOperationKind kind) : kind_(kind) {}
  virtual ~AbstractOperation() {}

 public:
  AbstractOperationKind getKind() const { return kind_; }

  // Release any underlying resources, including the interface object.
  //
  // WARNING: The destructor of this class is marked as protected to disallow
  // clients from directly destroying this object since it may manage it's own

TF_CAPI_EXPORT extern const char* TFE_TensorHandleBackingDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Return a pointer to a new TFE_TensorHandle that shares the underlying tensor
// with `h`. On success, `status` is set to OK. On failure, `status` reflects
// the error and a nullptr is returned.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_TensorHandleCopySharingTensor(

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());

  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 =

#include "tensorflow/c/tf_status.h"
#include "tensorflow/core/lib/gtl/cleanup.h"
#include "tensorflow/core/platform/test.h"

namespace {

struct LoggingDevice {
  tensorflow::string device_name;
  tensorflow::string underlying_device;
  // Set to true whenever a TensorHandle is copied onto the device
  bool* arrived_flag;
  // Set to true whenever an operation is executed
  bool* executed_flag;

void RegisterParallelDevice(
    TFE_Context* context, const char* device_name,
    const std::array<const char*, num_devices>& underlying_devices,
    TF_Status* status) {
  TFE_CustomDevice device;
  void* device_info;
  tensorflow::parallel_device::AllocateParallelDevice(
      device_name, underlying_devices.data(), underlying_devices.size(),
      &device, &device_info);
  TFE_RegisterCustomDevice(context, device, device_name, device_info, status);