limitations under the License.
==============================================================================*/

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

#include <string>

#include "include/dlpack/dlpack.h"  // from @dlpack
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"

  // custom device.
  //
  // Many devices will want to simply return an "unimplemented" status
  // here. This is the default behavior if `pack` is null when passed to
  // TFE_RegisterCustomDevice.
  TFE_TensorHandle* (*pack)(TFE_Context* context, TFE_TensorHandle** handles,
                            int num_handles, TF_Status* s,
                            void* device_info) = nullptr;

        "//tensorflow/core/common_runtime/eager:tensor_handle",
        "@dlpack",
    ],
    alwayslink = 1,
)

tf_cuda_cc_test(
    name = "dlpack_test",
    size = "small",
    srcs = [
        "dlpack_test.cc",
    ],
    args = [],
    tags = [],
    deps = [
        ":c_api",
        ":dlpack",
        "//tensorflow/core:test",
        "//tensorflow/core:test_main",

        TFE_DeleteTensorHandle(outputs[i]);
      }
    }
    return status.status;
  }

  tensorflow::Status Pack(absl::Span<ImmediateExecutionTensorHandle*> handles,
                          ImmediateExecutionTensorHandle** result) override {
    TF_Status status;
    *result = tensorflow::unwrap(device_.pack(context_,
                                              tensorflow::wrap(handles.data()),

      // One of the inputs we're trying to pack is on a custom device. We'll let
      // the first custom device we see handle all of the packing.
      auto* custom_device_handle =
          tensorflow::down_cast<tensorflow::CustomDeviceTensorHandle*>(
              unwrapped_handle);
      tensorflow::ImmediateExecutionTensorHandle* result;
      status->status = custom_device_handle->device()->Pack(

  if (!status->status.ok()) return;

  // Initialize a input_tensor vector with `nullptr` values.
  std::vector<const Tensor*> input_tensors_vector(num_inputs, nullptr);
  // A vector to keep track of newly created `tf::Tensor` objects.
  std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {

    status->status = cc_status;
    return;
  }
  if (var_info->var_info.var() == nullptr) {
    cc_status = absl::InvalidArgumentError(
        "VariableInfo does not track a resource variable.");
    status->status = cc_status;
    return;
  }

  cc_status = cc_ctx->allocate_temp(var_info->var_info.var()->tensor()->dtype(),

// A ParallelDevice on its own is not registered with a TFE_Context, and so has
// no device name (e.g. for `tf.device`). `NamedParallelDevice` associates a
// name with it, which lets us pack its `ParallelTensor`s into TFE_TensorHandles
// placed on the parallel device.
class NamedParallelDevice {
 public:
  NamedParallelDevice(const std::string& name,

      Multiply(context.get(), second_combined_value.get(),
               second_negative_one.get(), status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Un-pack the parallel tensor to verify that the operation was
  // successful. The resulting structure should be:
  //   second_device{first_device{1. * 3., 2. * 3.}, 3. * 3.}.
  std::array<TensorHandlePtr, 2> second_components;

                   &FloatDeallocator, nullptr),
      TF_DeleteTensor);
  return TensorHandlePtr(TFE_NewTensorHandle(tensor.get(), status));
}

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,