: (tensor<?x3x2x1xi8>, tensor<2x1x1x1xi8>) -> tensor<?x2x2x1xi32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<?x2x2x1xi32>, tensor<?x2x2x1xi32>) -> tensor<?x2x2x1xi32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<?x2x2x1xi32>, tensor<1xi32>) ->tensor<?x2x2x1xi32>
  %3 = chlo.broadcast_add %2, %bias : (
      tensor<?x2x2x1xi32>, tensor<1xi32>) ->tensor<?x2x2x1xi32>
  return %3 : tensor<?x2x2x1xi32>
}

// -----

  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<[1, 3]> : tensor<2xi64>}> : (tensor<2x3xf32>) -> tensor<1x2x2x3xf32>
  %1 = mhlo.multiply %0, %cst1 : tensor<1x2x2x3xf32>
  // CHECK:      return %[[RES]] : tensor<1x2x2x3xf32>
  func.return %1 : tensor<1x2x2x3xf32>
}

// CHECK-LABEL: @foldBroadcastInDimBeforeMulOp_bcast_dim_1D_int

// CHECK-LABEL: testSpaceToDepth
func.func @testSpaceToDepthF32(%arg0: tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32> {
  // CHECK: %[[ARG:.*]]: tensor<1x2x2x1xf32>
  // CHECK: "tfl.space_to_depth"(%[[ARG]]) <{block_size = 2 : i32}> : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  %0 = "tfl.space_to_depth"(%arg0) {block_size = 2: i32} : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  func.return %0 : tensor<1x1x1x4xf32>
}

// -----

// CHECK{LITERAL}: %cst_1 = arith.constant dense<[[[[1.000000e+00], [1.250000e+00]]]]> : tensor<1x1x2x1xf32>
// CHECK: %3 = tfl.mul %2, %cst_1 {fused_activation_function = "NONE"} : tensor<1x1x2x1xf32>
// CHECK: %cst_2 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK: %4 = "tfl.transpose"(%3, %cst_2) : (tensor<1x1x2x1xf32>, tensor<4xi32>) -> tensor<1x1x1x2xf32>

  %8 = "tf.FloorMod"(%7, %4) {device = ""} : (tensor<1x1x2x1xi32>, tensor<i32>) -> tensor<1x1x2x1xi32>
  %9 = "tf.FloorDiv"(%arg1, %4) {device = ""} : (tensor<1x1x2x1xi32>, tensor<i32>) -> tensor<1x1x2x1xi32>
  %10 = "tf.Pack"(%2, %9, %8, %2) {axis = 0 : i64, device = ""} : (tensor<1x1x2x1xi32>, tensor<1x1x2x1xi32>, tensor<1x1x2x1xi32>, tensor<1x1x2x1xi32>) -> tensor<4x1x1x2x1xi32>

}

func.func @cast(%arg0: tensor<1x2x2x5xi32>) -> tensor<1x2x2x5xf32> {
  %0 = "tf.Cast"(%arg0) : (tensor<1x2x2x5xi32>) -> tensor<1x2x2x5xf32>
  func.return %0 : tensor<1x2x2x5xf32>

  // CHECK-LABEL: cast
  // CHECK: "tfl.cast"(%arg0) : (tensor<1x2x2x5xi32>) -> tensor<1x2x2x5xf32>
}

func.func @castFloat32ToI16(%arg0: tensor<1x2x2x5xf32>) -> tensor<1x2x2x5xi16> {

      -> tensor<1x2x2x2xf32>
    %3 = "quantfork.stats"(%arg2) {layerStats = dense<[7.05456924, 7.11401462]> : tensor<2xf32>} : (tensor<2xf32>) -> tensor<2xf32>
    %4 = "quantfork.stats"(%2) {layerStats = dense<[-1.36523, 3.57373]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %5 = "chlo.broadcast_add"(%4, %3) : (tensor<1x2x2x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>

func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexDepthToSpace", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
// CHECK: return %[[CUSTOM_0]] : tensor<1x2x2x1xf32>
}

    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    // CHECK-SKIP: tf.PartitionedCall
    // CHECK-NOSKIP: call @some_other_func
    // CHECK-NOSKIP-NOT: tf.PartitionedCall
    func.return %1: tensor<1x2x2x3xf32>
  }

  // CHECK-SKIP: func.func private @some_func
  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> attributes {tf._noinline = true} {

    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    func.return %0: tensor<1x2x2x3xf32>
  }

  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> {
    return %arg0 : tensor<1x2x2x3xf32>
  }
}

// CHECK: module
// CHECK-NOT: tf.PartitionedCall
// CHECK-NOT: some_func