[Mp4-tech] help

[liu_allen2002 netzero.com]

Hi, all,
In H.264 baseline profile, after 16x16 prediction, the residual will use 4x4 dct, then take every DC coefficient form every 16 4x4 sub block and use hadmand transform. My question is "we should take every DC coeffs after 4x4 dct result be quanlised or before quantized?" I think if we do after quanlized, then the DC matrix coeffs will be smaller after hadmand transform and quantization. Which will make the encoding bit stream shorter. Maybe we should use bigger QP for DC matrix, but how big?
In encoding and decoding the remaining non zero coeff process, if the level[i] value is very big, for example (4000), then how to encoding the number and decode the number?
Thank you for your help.
Allen