const char* local_device_name) {
  TF_Status* status = TF_NewStatus();
  TFE_TensorHandle* h0_task0 = TestMatrixTensorHandle(ctx);

  TFE_Op* matmul = MatMulOp(ctx, h0_task0, h0_task0);
  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);

    TFE_Op* matmul = MatMulOp(ctx, hcpu, hgpu);
    if (cpu_op) {
      string cpu_device_name;
      ASSERT_TRUE(GetDeviceName(ctx, &cpu_device_name, "CPU"));
      TFE_OpSetDevice(matmul, cpu_device_name.c_str(), status.get());
    } else {
      TFE_OpSetDevice(matmul, gpu_device_name.c_str(), status.get());
    }
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

    TFE_TensorHandle* retvals[1];

  TFE_OpSetAttrShape(op, "shape", {}, 0, status);
  TFE_OpSetAttrString(op, "container", "localhost", 0);
  TFE_OpSetAttrString(op, "shared_name", "", 0);
  if (!device_name.empty()) {
    TFE_OpSetDevice(op, device_name.c_str(), status);
  }
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  TFE_Execute(op, &var_handle, &num_retvals, status);

    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetAttrInt(op.get(), "num_replicas", 1);
    TFE_OpAddInput(op.get(), value_one.get(), status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetDevice(op.get(), device_name, status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;

    TFE_TensorHandle* result_handles;
    int num_retvals = 1;

  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_OpAddInput(func, packed_handle, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  if (remote) {
    TFE_OpSetDevice(func, task1_name, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  }

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;

TF_CAPI_EXPORT extern TFE_Context* TFE_OpGetContext(const TFE_Op* op,
                                                    TF_Status* status);

TF_CAPI_EXPORT extern void TFE_OpSetDevice(TFE_Op* op, const char* device_name,
                                           TF_Status* status);
// The returned string remains valid throughout the lifetime of 'op'.

  if (op_ == nullptr) {
    TFE_ContextSetExecutorForThread(context, executor_.get());
    op_.reset(TFE_NewOp(context, operation_name, status));
    if (TF_GetCode(status) != TF_OK) return;
    TFE_OpSetDevice(op_.get(), device_.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return;
  } else {
    TFE_OpReset(op_.get(), operation_name, device_.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return;
  }

    // device names repeatedly.
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> const_op(
        TFE_NewOp(context, "Const", status), TFE_DeleteOp);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetDevice(const_op.get(), underlying_devices_[device_index].c_str(),
                    status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetAttrTensor(const_op.get(), "value", tensor.get(), status);

        hcpu, ctx, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;

// creating a new op every time. If `raw_device_name` is `NULL` or empty, it
// does not set the device name. If it's not `NULL`, then it attempts to parse
// and set the device name. It's effectively `TFE_OpSetDevice`, but it is faster
// than separately calling it because if the existing op has the same
// `raw_device_name`, it skips parsing and just leave as it is.
TF_CAPI_EXPORT extern void TFE_OpReset(TFE_Op* op_to_reset,