// A handle to a tensor on a device.
//
// Like a TF_Tensor, a TFE_TensorHandle refers to a tensor with a value, shape,
// type etc. Unlike a TF_Tensor, a TFE_TensorHandle may refer to such tensors
// placed in the memory of different devices or remote address spaces.
typedef struct TFE_TensorHandle TFE_TensorHandle;

TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandle(const TF_Tensor* t,

using tensorflow::string;
using tensorflow::tstring;

TFE_TensorHandle* TestScalarTensorHandle(TFE_Context* ctx, float value) {
  float data[] = {value};
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_FLOAT, nullptr, 0, status);
  memcpy(TF_TensorData(t), &data[0], TF_TensorByteSize(t));
  TFE_TensorHandle* th = TFE_NewTensorHandleFromTensor(ctx, t, status);

}

TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_TensorHandleCopySharingTensor(
    TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid handle");
    return nullptr;
  }

  tensorflow::unwrap(h)->Ref();
  return h;
}

TF_Tensor* TFE_TensorHandleResolve(TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {

std::unique_ptr<ParallelTensor> ParallelDevice::CopyToParallelDevice(
    TFE_Context* context, TFE_TensorHandle* tensor, TF_Status* status) const {
  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);

                        &arrived, &executed, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_TensorHandle* hcpu = TestMatrixTensorHandle(context);
  ASSERT_FALSE(arrived);
  TFE_TensorHandle* hdevice =
      TFE_TensorHandleCopyToDevice(hcpu, context, name, status.get());
  ASSERT_TRUE(arrived);
  ASSERT_FALSE(executed);

//
// Since this function is used to satisfy the TFE_CustomDevice C API,
// device_info is passed in using a C-style generic. It must always be a
// ParallelDevice.
TFE_TensorHandle* CopyToParallelDevice(TFE_Context* context,
                                       TFE_TensorHandle* tensor,
                                       TF_Status* status, void* device_info) {
  TF_SetStatus(
      status, TF_UNIMPLEMENTED,

  const char task2_name[] = "/job:localhost/replica:0/task:2/device:CPU:0";

  // Create one variable per task.
  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.

  {
    // Try to pass a ParallelTensor to TPUReplicatedInput
    std::array<TFE_TensorHandle*, 1> correct_components{value_one.get()};
    TensorHandlePtr combined_value = CreatePerDeviceValues(
        context.get(), correct_components, device_name, status.get());
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

    std::array<TFE_TensorHandle*, 1> incorrect_components{combined_value.get()};

      /*is_async=*/async, /*enable_streaming_enqueue=*/true,
      /*in_flight_nodes_limit=*/0);
  TFE_ContextSetExecutorForThread(ctx, executor);

  TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
  TFE_Op* matmul = MatMulOp(ctx, m, m);
  TFE_TensorHandle* retvals[2] = {nullptr, nullptr};
  int num_retvals = 2;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(1, num_retvals);

  TFE_TensorHandle* (*copy_tensor_to_device)(TFE_Context* context,
                                             TFE_TensorHandle* tensor,
                                             TF_Status* status,
                                             void* device_info);

  // Method to copy a tensor from the custom device to a target device.
  TFE_TensorHandle* (*copy_tensor_from_device)(TFE_Context* context,