// 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>)

func.func @BiasAddGrad_unranked(%arg0: tensor<*xf32>) -> tensor<?xf32> {
  // CHECK: tf.BiasAddGrad
  %0 = "tf.BiasAddGrad"(%arg0) {data_format = "NCHW"} : (tensor<*xf32>) -> tensor<?xf32>
  func.return %0 : tensor<?xf32>
}

// CHECK-LABEL: func @rsqrt_grad
// CHECK-SAME: (%[[ARG0:.*]]: tensor<2xf32>, %[[ARG1:.*]]: tensor<2xf32>)
func.func @rsqrt_grad(%arg0: tensor<2xf32>, %arg1: tensor<2xf32>) -> tensor<2xf32> {

        tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,

  %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>

    %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>
  }

    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<?x3xf32>) -> tensor<?x3xf32>
    return %2 : tensor<?x3xf32>
  }
// CHECK: func.func private @quantize_dot_general_with_bias_dynamic_fn(%[[ARG_0:.+]]: tensor<?x2xf32>) -> tensor<?x3xf32> attributes {tf._original_func_name = "main_0"}

// 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>)

//     : (tensor<1x2xf32>) -> tensor<2xf32>
//   %items1:4 = "tf.Unpack"(%[[INP1]]) {axis = 0 : i64}
//     : (tensor<4x2xf32>) -> (tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
//     tensor<2xf32>)
//   %axis = "tf.Const"() {value = dense<0> : tensor<i64>}
//   %0 = "tf.Pack"(items1#3, items1#2, items1#1, items1#0, %items0, %axis)
//     : (tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,

  %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>

    %0 = "tf.Div"(%arg1, %cst_0) : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
    %1 = "tf.AddV2"(%0, %cst_1) : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
    %2 = "tf.Floor"(%1) : (tensor<2x2xf32>) -> tensor<2x2xf32>
    %3 = "tf.ClipByValue"(%2, %cst_5, %cst_6) : (tensor<2x2xf32>, tensor<f32>, tensor<f32>) -> tensor<2x2xf32>
    %4 = "tf.Cast"(%3) {Truncate = false} : (tensor<2x2xf32>) -> tensor<2x2xi8>