auto arg0 = ops::_Arg(scope.WithOpName("a_0_arg"), DT_INT32, 0);
  auto arg1 = ops::_Arg(scope.WithOpName("b_0_arg"), DT_FLOAT, 1);
  auto arg2 = ops::_Arg(scope.WithOpName("c_0_arg"), DT_INT32, 2);
  auto arg3 = ops::_Arg(scope.WithOpName("d_0_arg"), DT_FLOAT, 3);

  auto arg4 = ops::_Arg(scope.WithOpName("u_0_arg"), DT_RESOURCE, 4);
  auto arg5 = ops::_Arg(scope.WithOpName("v_0_arg"), DT_RESOURCE, 5);

      {
          {{"G"}, "BinaryTest", {"e_0_arg", "f_0_arg"}},
          {{"I"},
           "BinaryTest",
           {"f_0_arg", "outside_compilation_O1_host_compute:outputs:0"}},
          {{"outside_compilation_O1_host_compute"},
           "XlaHostCompute",
           {"d_0_arg", "G:o:0"},
           {{"Tinputs", absl::Span<const DataType>({DT_FLOAT, DT_FLOAT})},

      auto in_rtt =
          llvm::dyn_cast<RankedTensorType>(cst_arg->getResult(0).getType());

      auto inverse_type =
          RankedTensorType::get(PermuteShape(in_rtt.getShape(), inverse_perm),
                                in_rtt.getElementType());

      tposed_const = rewriter.create<TFL::TransposeOp>(
          cst_arg->getLoc(), inverse_type, cst_arg->getResult(0),
          inverse_perm_op);
    }

    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

    auto cpu_arg =
        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) {