Store intra4x4_pred_mode per row only. about 5 cpu cycles slower in the local code but should be overall faster due to reduced cache use. (my sample though has too few intra4x4 blocks for this to be meassureable easily either way)