// done in parallel.
  //
  // This graph is:
  // (Const0, Const0) -> MatMul0
  // (Const1, Const1) -> MatMul1
  // (MatMul0, MatMul1) -> MatMulCombined
  //
  // Device0: [Const0, Const0, MatMul0]
  // Device1: [Const1, Const1, MatMul1, MatMulCombined]
  //
  // Cluster0: [Const0, Const0, MatMul0]
  // Cluster1: [Const1, Const1, MatMul1]
  // Cluster2: [MatMulCombined]

  %3 = stablehlo.add %2, %1 : tensor<1x3xf32>
  func.return %3: tensor<1x3xf32>
}
// CHECK: %[[CONST_0:.*]] = stablehlo.constant dense<2.000000e+00>
// CHECK: %[[CONST_1:.*]] = stablehlo.constant dense<2.000000e+00>
// CHECK: %[[XLA_CALL_MODULE:.*]] = "tf.XlaCallModule"(%arg0, %[[CONST_0]], %[[CONST_1]])
// CHECK: return %[[XLA_CALL_MODULE:.*]] : tensor<1x3xf32>
// CHECK: }

    //        |         |
    //      Const_0    Const_1
    //
    const float const0_val[] = {1.0, 2.0, 3.0, 4.0};
    const float const1_val[] = {1.0, 0.0, 0.0, 1.0};
    TF_Operation* const0 = FloatConst2x2(graph_, s_, const0_val, "Const_0");
    TF_Operation* const1 = FloatConst2x2(graph_, s_, const1_val, "Const_1");
    TF_Operation* matmul = MatMul(graph_, s_, const0, const1, "MatMul");

// CHECK: %[[CONST_0:.+]] = stablehlo.constant() <{value = dense<{{.*}}> : tensor<2x3xi8>}> : () -> tensor<2x3x!quant.uniform<i8<-127:127>:f32:1, {{.*}}>
// CHECK: %[[UNIFORM_QUANTIZE_0:.+]] = stablehlo.uniform_quantize %[[ARG_0]] : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, {{.*}}>>

  Output id1 = ops::Identity(root.WithOpName("id1"), sw.output_true);

  Output const0 = ops::Const(root.WithOpName("const0"), 1);
  Output const1 = ops::Const(root.WithOpName("const1"), 2);

  Output add = ops::Add(root.WithOpName("add"), const0, const1);

  root.graph()->AddControlEdge(id0.node(), const0.node());
  root.graph()->AddControlEdge(id1.node(), const1.node());

    %15 = "tf.Cast"(%14) {Truncate = false} : (tensor<1x3x2x2xi32>) -> tensor<1x3x2x2xi8>
    return %15 : tensor<1x3x2x2xi8>
  }

// CHECK-LABEL: func @conv_with_bias_and_relu
// CHECK-DAG: %[[CONST_0:.*]] = "tf.Const"() <{value = dense<[1, 2]> : tensor<2xi32>}> : () -> tensor<2xi32>
// CHECK-DAG: %[[CONST_1:.*]] = "tf.Const"() <{value = dense<1> : tensor<2xi32>}> : () -> tensor<2xi32>

      {});
}

TEST_F(CApiFunctionTest, UsingOneOutputOfSplit) {
  /*
   *                      feed
   *                       |
   *             +---------+---+
   *             | const0  |   |
   *             |    |    |   |
   *             |    v    /   |
   *             |    split    |
   *             |   |  |  |   |
   *             |   v  |  v   |
   *             |      |      |

variables, for the buffer and current size. Each push will be turned into
```mlir
  %old_val = "tf.ReadVariableOp"(%buffer)
  %old_size = "tf.ReadVariableOp"(%size)
  %offsets = "tf.ConcatV2"(%old_size, %other_dims_0s, %const0)
  %new_val = "tf.XlaDynamicUpdateSlice"(%old_val, %push_val, %offsets)
  "tf.AssignVariableOp"(%buffer, %new_val)
  %new_size = "tf.AddV2"(%old_size, %const1)
  "tf.AssignVariableOp"(%size, %new_size)
```