}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

    %26 = "tf.Cast"(%25) {Truncate = false} : (tensor<2xi64>) -> tensor<2xf32>
    %27 = "tf.Equal"(%14, %26) {incompatible_shape_error = true} : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>
    %28 = "tf.Cast"(%27) {Truncate = false} : (tensor<2xi1>) -> tensor<2xf32>
    %29 = "tf.Sum"(%28, %6) {keep_dims = false} : (tensor<2xf32>, tensor<1xi32>) -> tensor<f32>

func.func @add_dense_dense_float_mixfng_1_n() -> tensor<2x2xf32> {
  %cst_0 = arith.constant dense<[[1.5, -2.5]]> : tensor<1x2xf32>
  %cst_1 = arith.constant dense<[[-3.], [4.]]> : tensor<2x1xf32>

  %0 = "tfl.add"(%cst_0, %cst_1) {fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<2x1xf32>) -> tensor<2x2xf32>

  func.return %0 : tensor<2x2xf32>

func.func @my_add_n_single_input(%arg0: tensor<2x3xf32>) -> tensor<2x3xf32> {
  %0 = "tf.MyAddN"(%arg0) {N=1:i32} : (tensor<2x3xf32>) -> tensor<2x3xf32>
  func.return %0 : tensor<2x3xf32>

// CHECK-NEXT: return %arg0 : tensor<2x3xf32>
}

// CHECK-LABEL: my_add_n_multiple_inputs
func.func @my_add_n_multiple_inputs(%arg0: tensor<2x3xf32>, %arg1: tensor<2x3xf32>, %arg2: tensor<2x3xf32>) -> tensor<2x3xf32> {

      %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
      return %0 : tensor<2x2xf32>
    }
  }
)mlir";

constexpr absl::string_view kModuleHybridQuantized = R"mlir(
  module {

}

// CHECK-LABEL: prepareAdd
func.func @prepareAdd(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {
  %cst = arith.constant dense<[[0.0, 1.0], [2.0, 255.0]]> : tensor<2x2xf32>
  %add = "tfl.add"(%arg0, %cst) {fused_activation_function="NONE"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
  func.return %add : tensor<2x2xf32>

          %outputs_9, %control_10 = tf_executor.island wraps "tf.XlaSharding"(%outputs_7) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01", sharding = "\08\03\1A\02\02\01\22\02\00\01", unspecified_dims = []} : (tensor<2x2xf32>) -> tensor<2x2xf32>
          tf_executor.fetch %outputs_9 : tensor<2x2xf32>
        }
        return %0 : tensor<2x2xf32>
      }
    }
  )";

    %3 = "tfl.reshape"(%2, %cst_0) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>
    %4 = "tfl.reshape"(%3, %cst_1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
    return %4 : tensor<2x1xf32>
  }
```

#### Compute Costs Pass
In the compute cost pass, we will essentially compute the cost of each op within

    %0 = "tfl.batch_matmul"(%arg0, %cst_0) {adj_x = false, adj_y = false} : (tensor<1x256xf32>, tensor<256x256xi8>) -> tensor<1x256xf32>
    %1 = "tfl.batch_matmul"(%0, %cst_1) {adj_x = false, adj_y = false} : (tensor<1x256xf32>, tensor<256x256x!quant.uniform<i8:f32, 1.000000e+00>>) -> tensor<1x256xf32>
    %2:2 = "tfl.while"(%cst_2, %1) ({
    ^bb0(%arg1: tensor<i32>,  %arg2: tensor<1x256xf32>):
      %cst_3 = arith.constant dense<10> : tensor<i32>

    %identity_out1, %control_identity1 = tf_executor.island wraps "tf.Identity"(%partitioned_out#0) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
    %identity_out2, %control_identity2 = tf_executor.island wraps "tf.Identity"(%partitioned_out#1) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>

    tf_executor.fetch
  }
  func.return
}