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);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteTensor(t);

  // big-endian platform.
  memcpy(*buffer, &num_strings, sizeof(int32_t));

  // Set offset of strings.
  int32_t start = sizeof(int32_t) * (num_strings + 2);
  for (size_t i = 0; i < offset_.size(); i++) {
    // TODO(b/165919229): This code will need changing if/when we port to a
    // big-endian platform.
    int32_t offset = start + offset_[i];
    memcpy(*buffer + sizeof(int32_t) * (i + 1), &offset, sizeof(int32_t));

    ASSERT_EQ(num_elem_numerical, num_elem_analytical);

    float* dnumerical = new float[num_elem_numerical]{0};
    memcpy(&dnumerical[0], TF_TensorData(numerical_tensor),
           TF_TensorByteSize(numerical_tensor));
    float* danalytical = new float[num_elem_analytical]{0};
    memcpy(&danalytical[0], TF_TensorData(analytical_tensor),
           TF_TensorByteSize(analytical_tensor));

  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.
  int num_elems = TF_TensorElementCount(theta_tensor);
  vector<float> theta_data(num_elems);
  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

  // Initialize space for the numerical gradient.
  vector<float> dtheta_approx(num_elems);

    EXPECT_EQ(stream->stream_id, 14);
    std::memcpy(host_dst, device_src->opaque, size);
  };

  StreamExecutor* executor = GetExecutor(0);
  TF_ASSERT_OK_AND_ASSIGN(auto stream, executor->CreateStream());
  size_t size = sizeof(int);
  int src_data = 34;
  int dst_data = 2;
  DeviceMemoryBase device_src(&src_data, size);
  TF_ASSERT_OK(stream->Memcpy(&dst_data, device_src, size));
  ASSERT_EQ(dst_data, 34);
}

  Status status = XLAShapeToTensorShape(shape_, &tensor_shape);
  if (!status.ok()) {
    done(status);
    return;
  }

  *cpu_tensor = Tensor(dtype, tensor_shape);

  status = stream_->Memcpy(cpu_tensor->data(), device_memory_base_,
                           device_memory_base_.size());
  if (!status.ok()) {
    done(status);
    return;
  }
  status = stream_->RecordEvent(done_event_.get().get());

  xla::Shape shape;
  TF_ASSERT_OK(TensorShapeToXLAShape(DT_FLOAT, TensorShape({2, 2}), &shape));

  // Copy the cpu_tensor to the GPU first before trying to copy it back.
  TF_ASSERT_OK(
      stream->Memcpy(&gpu_dst, origin_cpu_tensor.data(), gpu_dst.size()));
  TF_ASSERT_OK(stream->BlockHostUntilDone());

  TF_ASSERT_OK_AND_ASSIGN(auto se_event, executor->CreateEvent());
  tsl::AsyncValueRef<std::unique_ptr<se::Event>> done_event =

namespace tensorflow {

void XlaHostSendDeviceContext::CopyCPUTensorToDevice(
    const Tensor* cpu_tensor, Device* device, Tensor* device_tensor,
    StatusCallback done, bool sync_dst_compute) const {
  auto status = stream_->Memcpy(device_memory_base_, cpu_tensor->data(),
                                device_memory_base_->size());
  if (!status.ok()) {
    done(status);
    return;
  }
  status = stream_->RecordEvent(done_event_.get().get());

extern "C" {

TF_Buffer* TF_NewBuffer() { return new TF_Buffer{nullptr, 0, nullptr}; }

TF_Buffer* TF_NewBufferFromString(const void* proto, size_t proto_len) {
  void* copy = tensorflow::port::Malloc(proto_len);
  std::memcpy(copy, proto, proto_len);

  TF_Buffer* buf = new TF_Buffer;
  buf->data = copy;
  buf->length = proto_len;
  buf->data_deallocator = [](void* data, size_t length) {
    tensorflow::port::Free(data);
  };

  *len = debug_str.size();
  char* ret = static_cast<char*>(malloc(*len + 1));
  memcpy(ret, debug_str.c_str(), *len + 1);
  return ret;
}

char* TF_FunctionDebugString(TF_Function* func, size_t* len) {
  const auto& debug_str = DebugString(func->record->fdef());
  *len = debug_str.size();
  char* ret = static_cast<char*>(malloc(*len + 1));
  memcpy(ret, debug_str.c_str(), *len + 1);
  return ret;
}