[XviD-devel] Changes to get_pmv2

[Christoph Lampert]

On Mon, 20 Sep 2004, Tom Jacobs wrote:

> hi
> 
> i am still threading mostion estimation and am looking at ways of gettin
> around the problem of each blocks predictions being based on its
> neighbours.
> ideally we would make the prediction from the pixel above, below, left and
> right. because of the scan method used in ME we have changed it to left top
> and top right.
> you there be a great quality decrease if it was changed to top left, top,
> top right? for my work this would make it possible for a whole row to be
> processed in parallel. 

Hi,

the decrease in quality (or rather the raise in bitrate) wouldn't be big, 
but, if you write the resulting MVs to the bitstream like that, it would 
violate the MPEG-4 standard, the resulting bitstream wouldn't be playable 
by MPEG-4 codecs anymore. 
What you can do is do ME with the "wrong" predictors, but then afterwards 
use the right predictors to write the bitsream. This will slow the 
process down a little, and cause another raise in bitrate, but unless you 
target very low bitrates, I guess it would still be acceptable. 

However, more important, the gain in speed isn't that big by doing only ME
in parallel. I tried a while ago (when machine were much slower),
splitting the image vertically in two halfs. But the overhead of creating
the threads, doing ME and rejoining the results was bigger than expected.
The speedup was less than 10% for 2 CPUs. 

gruel