// 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);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  // TODO(gjn): Add support for waiting on async local mirrors
  if (!remote && !async && !remote_func_outputs) {

    int num_retvals = 1;
    TFE_Execute(matmul, &retvals[0], &num_retvals, status.get());
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

    // The CPU handle should have been copied and have a mirror on the GPU
    ASSERT_TRUE(cpu_arg->HasLocalMirror(gpu_device));

    TFE_DeleteOp(matmul);
    TFE_DeleteTensorHandle(retvals[0]);
    TFE_DeleteTensorHandle(hgpu);
  }

  TF_DeleteTensor(t);