[Mp4-tech] H264 Transform Precision Requirement

[Ravi Bulusu]

Hi All, 
I've asked this question earlier with not much luck. I'm hoping some
expert will answer my question this time. 
I have been trying to understand the data precision requirement for H264
transform. There are a few intermediate stages in the transform
computation in H264. The output of every stage has a constraint that the
range of data should not exceed the values from -2^15 to 2^15 - 1. 
Since the same constraint is applicable for all the four (broadly
classified) stages of transform computation, it does feel like these
constaints are very loosely specified. 
Can someone tell me what would be a "strict" requirement on the input
coefficients to the transform process given the intermediate constraints
that have already been placed by the specification? In other words, if I
mention that a 15 bit precision (i.e, -2^14 to 2^14-1) on the input
coefficients is good enough to meet H264 transform compliance, would
there be a challenging argument (or a proof) to disprove it? 
Another issue I have is with the precision on the output of this
transform. Currently the specification restricts the output data in the
range of [-512, 511]. However, since this is the residual data which is
added to the motion compensated data (which happens to be the 8-bit
unsigned pixel values), does it not suffice for the output of the
transform to be restricted to [-256, 255]? I am unable to understand the
reason behind the extra bit precision requirement at the output. 
thx, 
Ravi
-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/mp4-tech/attachments/20060328/cfc1030c/attachment.html