[XviD-devel] iDCT coefficients different...

Christoph Lampert xvid-devel@xvid.org
Thu, 25 Jul 2002 14:56:27 +0200 (CEST)


Hi,

> I'm really surprised that a +- 1.0 difference is ALLOWED, because of the
> effects visible here. 

Could can hardly allow an error of _less_ than 1 for integer values, can
you? Then you would have to simply disallow _any_ errors which is somehow 
ridiculous, too, because rouding is very arbitrary business. 

Btw. I was wrong in two places: 

1) "For any pixel location, the peak error shall not exceed 2 in
magnitude"... so even 2 would be allowed! 

2) "Every macroblock is required to be refreshed before it is coded 132
times as predictive macroblocks"... so it's 132, not 180. Btw. XviD does
not care about this rules, at the moment, since normally a "keyframe
interval" of not too much more is given anyway (often around 10 seconds). 

3) "Where arithmetic precision is not specified, such as the calculation
fo the IDCT, the precision shall be sufficient so that significant errors
do not occur in the final integer values." ... although I don't know how
this should any programmer working on IDCT.

Is also states that 3) is meant additionally to 1), so you may say that
MMX quant is _not_ allowed by the standard, if you consider your problems
as "significant errors". 

Still, we won't remove MMX ;-) If this really is a problem, we could add a
"high precision" flag that disables MMX (i)DCT. Speed will of course drop
dramatically.

> (BTW, source material is a very still scene, not the typical "MTV" action.
> This
> is the reason that we reach the I-frame threshold of 180 frames even the
> same
> scene is visible)

Well, without the material, there's not very else much we can do about
it...

Christoph 

P.S. Isibaar, if you remember the "degrading image quality in fullpel
mode" problem we had a while ago. Did we do tests without MMX then? 
I don't remember... 

-- 
Christoph H. Lampert chl@math.uni-bonn.de | Diese Signature wurde maschi-     
Beringstr. 6, Raum 14 Tel. (0228) 73-2948 | nell erstellt und bedarf
Sprechstunden: keine, aber meistens da    | keiner Unterschrift. AZ 27B-6