[XviD-devel] swo more speedups possible

Christoph Lampert chl at math.uni-bonn.de
Sat Jan 3 12:50:27 CET 2004


Hi,

I'd say it's a good idea, I'm sure there is something to gain from
getting rid of redundant stuff in high quality modes. 

But please after 1.0. I know, it's not "big" changes, but still
it changes crucial parts of the code, and it's not a bugfix or anything.

gruel 

P.S. Have there been any complaints about beta3? Any problems left?

On Sat, 3 Jan 2004, Radek Czyz wrote:

> There are two places in the code where we can add easy (and not very 
> big) speedups. I never mentioned that before because I assumed this is 
> after-1.0 thing, but then again this is really easy to add, and the risk 
> of breaking something isn't huge (but not smaller than usual).
> 
> Both have to do with the fact that VHQ is performing the same 
> calculations as MBTransquant.
> 
> 1. after VHQ, we know the coded-block-pattern for all inter-coded 
> macroblocks. We can skip fdct and quantization for the blocks which will 
> not be coded. This works provided that:
>    - we're talking about chroma blocks, or
>    - we're talking about luma blocks in halfpel mode
>    - and Trellis will not re-encode some blocks which would otherwise be 
> not-coded (TRUE for our trellis).
> 
> The first two conditions come from the fact that VHQ, as all ME, uses an 
> approximation of qpel filter. As a result, CBP is not always identical, 
> with a differece of a few blocks per frame.
> Perhaps it would even be better to copy ME's CBP than calculate a new 
> one... <- just a thought. Or, we can make it benefit hpel/chroma alone.
> 
> 2. if VHQ decided on INTRA coding, it has already performed AC/DC 
> prediction, and it's in HQ_ACDC_PRED mode. We can copy the 
> predicted/not-predicted flag from VHQ rather than re-calculate it.
> 
> What do you think? CBP is already remembered for p-frames (I used it as 
> ASSERT-like debug tool for VHQ), we just have to initialize it (as 0xff 
> perhaps) in bframes and add some MBTransquant code, done.
> AC/DC prediction is eaven easier.
> 
> Comments?
> Radek
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel
> 



More information about the XviD-devel mailing list