// Use other output
  %3:6 = "tf.FusedBatchNormV3"( %2#0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", U = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<167x64xf32>
  %1 = stablehlo.constant dense<2.000000e+00> : tensor<64xf32>
  %2 = stablehlo.dot_general %arg0, %0, contracting_dims = [2] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  %3 = stablehlo.broadcast_in_dim %1, dims = [2] : (tensor<64xf32>) -> tensor<1x1x64xf32>
  %4 = stablehlo.add %2, %3 : tensor<1x1x64xf32>
  func.return %4: tensor<1x1x64xf32>
}

  %outputs_50 =  "tf.Mul"(%outputs_30, %outputs_48) {device = ""} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
  %outputs_52 =  "tf.Reshape"(%outputs_50, %outputs_0) {device = ""} : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>
  %outputs_54 =  "tf.MatMul"(%outputs_40, %outputs_52) {device = "", transpose_a = false, transpose_b = false} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

  %4 = "tfl.minimum"(%0, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>
  %5 = "tfl.minimum"(%1, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>
  %6 = "tfl.minimum"(%2, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>
  %7 = "tfl.minimum"(%3, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>

}

// CHECK-LABEL: func @add_f32
func.func @add_f32(%arg0: tensor<3xf32>, %arg1: tensor<3xf32>) -> tensor<3xf32> {
  // CHECK-DAG: [[ADD:%.+]] = "tf.AddV2"(%arg0, %arg1) : (tensor<3xf32>, tensor<3xf32>) -> tensor<3xf32>
  %add = "tf.Add"(%arg0, %arg1) : (tensor<3xf32>, tensor<3xf32>) -> tensor<3xf32>
  func.return %add : tensor<3xf32>
}

// CHECK-LABEL: func @BiasAddGrad_NHWC

// CHECK-SAME: (%[[ARG_0:.*]]: tensor<!tf_type.resource<tensor<32xf32>>>, %[[ARG_1:.*]]: tensor<!tf_type.resource<tensor<32xf32>>>, %[[ARG_2:.*]]: tensor<!tf_type.resource<tensor<32xf32>>>, %[[ARG_3:.*]]: tensor<!tf_type.resource<tensor<32xf32>>>, %[[ARG_4:.*]]: tensor<!tf_type.resource<tensor<32xf32>>>)

// CHECK:           %[[VAL_4:.*]]:2 = call @func_0_GPU_FLOAT(%[[VAL_0]], %[[VAL_1]], %[[VAL_2]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>, tensor<1xf32>)

      %0 = tf_executor.graph {
        %read0, %read0_control = tf_executor.island wraps "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf_type.resource<tensor<32xf32>>>) -> tensor<32xf32>
        %assign0_control = tf_executor.island(%read0_control) wraps "tf.AssignVariableOp"(%arg0, %arg1) : (tensor<*x!tf_type.resource<tensor<32xf32>>>, tensor<32xf32>) -> ()

        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>) -> tensor<*xf32>
  %24 = "quantfork.stats"(%23) {layerStats = dense<[-1.0, 2.0]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>

  // CHECK: %[[VAL:.*]] = "tf.Const"() <{value = dense<[1.000000e+00, 2.000000e+00, 3.000000e+00]> : tensor<3xf32>}> : () -> tensor<3xf32>
  %value = "tf.Const"() {value = dense<[1.0, 2.0, 3.0]> : tensor<3xf32>} : () -> tensor<3xf32>
  // CHECK: %[[READ_VAR:.*]] = "tf.ReadVariableOp"(%[[VAR]])
  // CHECK: %[[UPDATE_SLICE:.*]] = "tf.Reshape"(%[[VAL]]
  // CHECK-SAME: -> tensor<1x3xf32>