%0:4 = "tfl.split"(%arg0, %arg1) {num_splits = 4 : i32, tac.device = "CPU"} : (tensor<i32>, tensor<1x8x8x1024xf32>) -> (tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>)
  func.return %0#0, %0#1, %0#3 : tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>
}