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);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> matmul(

  parallel_device.Execute(context.get(), handle_inputs, "ReadVariableOp",
                          TFE_OpGetAttrs(read_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),

    auto remote_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(h1_task2));
    // The input handles should never change since they have been mirrored.
    ASSERT_FALSE(remote_arg->HasLocalMirror(nullptr));
  }

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

        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(hcpu));
    auto gpu_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(hgpu));
    auto gpu_device = gpu_arg->device();
    ASSERT_FALSE(cpu_arg->HasLocalMirror(gpu_device));

    TFE_Op* matmul = MatMulOp(ctx, hcpu, hgpu);
    if (cpu_op) {
      string cpu_device_name;
      ASSERT_TRUE(GetDeviceName(ctx, &cpu_device_name, "CPU"));

  // Test not found errors in TF_Operation*() query functions.
  EXPECT_EQ(-1, TF_OperationOutputListLength(feed, "bogus", s));
  EXPECT_EQ(TF_INVALID_ARGUMENT, TF_GetCode(s));

  ASSERT_FALSE(GetAttrValue(feed, "missing", &attr_value, s));
  EXPECT_EQ(string("Operation 'feed' has no attr named 'missing'."),
            string(TF_Message(s)));

  // Make a constant oper with the scalar "3".