return nullptr;
}

}  // end namespace tracing
}  // end namespace tensorflow

// =============================================================================
// Public C API entry points
//
// These are only the generic entry points for the C API. This file does not
// have any visibility into the graph/eager implementation and is only providing
// C bindings to the abstract classes defined in the

// DT_RESOURCE. The string serialization is converted back to a
// ResourceHandle during Session run where the TF_Tensor is converted to a
// Tensor.
// TFv2 does not depend on this conversion. There is no matching
// TF_TensorFromTensorV1 because the conversion to string is performed by the
// python side of Session.
static Status TF_TensorToTensorV1(const TF_Tensor* src, Tensor* dst) {

  if (enable) {
    optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::ON_1);

    // These XLA flags are needed to trigger XLA properly from C (more generally
    // non-Python) clients. If this API is called again with `enable` set to
    // false, it is safe to keep these flag values as is.
    tensorflow::MarkForCompilationPassFlags* flags =
        tensorflow::GetMarkForCompilationPassFlags();

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK);
  // Can't take the shape of a parallel tensor with varying numbers of axes, but
  // running operations on them is OK.
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> size_op(

  TF_AddInput(desc, {zero, 0});
  TF_AddInput(desc, {input, 0});
  TF_SetAttrInt(desc, "num_split", 3);
  TF_SetAttrType(desc, "T", TF_INT32);
  // Set device to CPU since there is no version of split for int32 on GPU
  // TODO(iga): Convert all these helpers and tests to use floats because
  // they are usually available on GPUs. After doing this, remove TF_SetDevice
  // call in c_api_function_test.cc
  TF_SetDevice(desc, "/cpu:0");

  }
  for (int i = ndim - 2; i >= 0; --i) {
    (*stride_arr)[i] = (*shape_arr)[i + 1] * (*stride_arr)[i + 1];
  }

  dlm_tensor->dl_tensor.shape = shape_arr->data();
  // There are two ways to represent compact row-major data
  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.

  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.
// These APIs are mainly to facilitate testing and are subject to change.
namespace internal {
Status Reset(AbstractOperation* op_, const char* op,
             const char* raw_device_name, ForwardOperation* forward_op_) {

// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either
// need to return something here or address these use-cases one by one.
TFE_TensorHandle* CopyTensorFromParallelDevice(TFE_Context* context,

using ExecutorPtr = std::unique_ptr<TFE_Executor, ExecutorDeleter>;

}  // namespace

// Allows a single op at a time to be launched without blocking.
//
// DeviceThread itself is thread-safe, in that StartExecute will block if there
// is a pending execution. Since StartExecute is equivalent to grabbing a lock,
// multiple DeviceThreads should always be accessed in the same order to avoid
// deadlocks.
class DeviceThread {
 public:

#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
#include "tensorflow/c/tf_status_internal.h"
#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.