// 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>)

  %1 = stablehlo.constant dense<2.000000e+00> : tensor<4xf32>
  %2 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x4xf32>) -> tensor<1x3x3x4xf32>
  %3 = stablehlo.broadcast_in_dim %1, dims = [3] : (tensor<4xf32>) -> tensor<1x3x3x4xf32>
  %4 = stablehlo.add %2, %3 : tensor<1x3x3x4xf32>

        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_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>)

    %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3xf32>} : () -> tensor<2x3xf32>
    %cst_0 = "tf.Const"() {value = dense<4.00000000e-1> : tensor<3xf32>} : () -> tensor<3xf32>
    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<?x2xf32>) -> tensor<?x2xf32>

    // CHECK-DAG: %[[SIGN:.*]] = "tf.Sign"(%[[PROX]]) : (tensor<4xf32>) -> tensor<4xf32>
    // CHECK-DAG: %[[ABS:.*]] = "tf.Abs"(%[[PROX]]) : (tensor<4xf32>) -> tensor<4xf32>
    // CHECK-DAG: %[[SCALED_L1:.*]] = "tf.Mul"(%[[ADAGRAD_LR]], %[[L1]]) : (tensor<4xf32>, tensor<f32>) -> tensor<4xf32>
    // CHECK-DAG: %[[PROX_NEW:.*]] = "tf.Sub"(%[[ABS]], %[[SCALED_L1]]) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

//     : (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>,

  %1 = "tfl.pad"(%arg1, %arg2) : (tensor<2x1x3xf32>, tensor<3x2xi32>) -> tensor<?xf32>
  %2 = "tfl.pad"(%0, %arg2) : (tensor<2x1x3xf32>, tensor<3x2xi32>) -> tensor<?xf32>
  func.return %1, %2 : tensor<?xf32>, tensor<?xf32>

// CHECK: %[[dq:.*]] = "tfl.dequantize"(%arg0)
// CHECK: %[[pad1:.*]] = "tfl.pad"(%arg1, %arg2)
// CHECK: %[[pad2:.*]] = "tfl.pad"(%[[dq]], %arg2)

  %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>

  %8:2 = "tf.TPUReplicatedOutput"(%2#1) : (tensor<?xf32>) -> (tensor<?xf32>, tensor<?xf32>)
  %9 = "tf.opB"(%5#0, %5#1, %7#0, %7#1) : (tensor<?xi32>, tensor<?xi32>, tensor<?xi32>, tensor<?xi32>) -> (tensor<?xi32>)
  %10 = "tf.opC"(%8#0, %8#1) : (tensor<?xf32>, tensor<?xf32>) -> (tensor<?xf32>)