auto x = ops::Placeholder(root.WithOpName("x"), DT_FLOAT);
  auto add1 = ops::Add(root.WithOpName("add1"), x, x);
  add1.node()->AddAttr("_cluster", "cluster1");
  auto add2 = ops::Add(root.WithOpName("add2"), add1, add1);
  add2.node()->AddAttr("_cluster", "cluster2");
  auto out = ops::Mul(root.WithOpName("mul"), add1, add2);

  Graph graph_before_encapsulation(OpRegistry::Global());

  %4 = "tf.Log"(%3) {device = "/job:localhost/replica:0/task:0/device:GPU:0"}: (tensor<4x4xf32>) -> tensor<4x4xf32>

  // CHECK: %[[ADD1:.*]] = "tf.AddV2"
  // CHECK: %[[LOG1:.*]] = "tf.Log"(%[[ADD1]])
  %5 = "tf.AddV2"(%4, %1): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>
  %6 = "tf.Log"(%5): (tensor<4x4xf32>) -> tensor<4x4xf32>

            %island1:3 = tf_executor.island {
              %add1 = "tf.Add"(%arg0, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
              %add2 = "tf.Add"(%add1, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
              %res = "tf.Print"(%add2) { message = "add result" } : (tensor<*xi32>) -> (tensor<*xi32>)
              tf_executor.yield %add1, %add2 : tensor<*xi32>, tensor<*xi32>
            }

  }, {
    ^bb0(%barg0: tensor<*xf32>, %barg1: tensor<i32>):
      %add0 = "tf.Add"(%barg0, %barg0) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
      %add1 = "tf.Add"(%barg1, %barg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
      "tf.Yield"(%add0, %add1) : (tensor<*xf32>, tensor<i32>) -> ()
  }) { is_stateless = false } : (tensor<*xf32>, tensor<i32>) -> (tensor<*xf32>, tensor<i32>)

  func.return %0#0 : tensor<*xf32>
}

openAVX2Tail128LoopB:
	ADDQ     $16, itr2
	chachaQR_AVX2(AA1, BB1, CC1, DD1, TT0)
	VPALIGNR $4, BB1, BB1, BB1
	VPALIGNR $8, CC1, CC1, CC1
	VPALIGNR $12, DD1, DD1, DD1
	chachaQR_AVX2(AA1, BB1, CC1, DD1, TT0)
	VPALIGNR $12, BB1, BB1, BB1
	VPALIGNR $8, CC1, CC1, CC1
	VPALIGNR $4, DD1, DD1, DD1
	CMPQ     itr2, itr1
	JB       openAVX2Tail128LoopA
	CMPQ     itr2, $160
	JNE      openAVX2Tail128LoopB