[MPlayer-dev-eng] [PATCH] bugfix for black bar besides gaussed vobsubs

[D Richard Felker III]

On Sat, Jan 08, 2005 at 03:33:37PM +0100, Reimar Döffinger wrote:
> Hi,
> On Sat, Jan 08, 2005 at 02:52:05PM +0100, Csillag Kristof wrote:
> > 2005-01-08, szo keltezéssel 11.35-kor Reimar Döffinger ezt írta:
> > > How about memset(d2, 255, ds*dh); before scaling? Probably not much
> > > slower (well, when I look at that code I think speed wasn't really much
> > > considered anyway) and easier to understand.
> > > Think that should work, too...
> > Now that you mention it, the following solution is faster and alsa more
> > simple (both to do, and to understand). (See attached patch!)
> > 
> > Here is how it works now:
> > 
> > 1. It gets the alpha values straight:
> > 	- 0 means totally transparent
> > 	- 1 means almost totally transparent
> > 	- 255 means not transparent at all
> > 	(Until so far, we set the values to the
> > 	_opposite_ values before scaling.)
> > 
> > 2. Scales up the stuff.
> > 	(The padded parts are 0 - that's totally transparent,
> > 				 no more problem)
> > 
> > 3. Reverts the alpha values to the "normal" way:
> > 	- 0 still means totally transparent
> > 	- 1 is not transparent at all
> > 	- 255 is almost totally transparent
> > 
> > 	(May I ask, why do we use this interpretation? It's a pain...)
> 
> Agree, I dislike it and it should be changed. I think it's because it
> was like that in the first OSD implementation (and is used when you
> define FAST_OSD). Not 100% sure though.

Definitely should not be changed. It's written this way for a reason,
to optimize the alpha blender, which is called on every frame. OSD is
already slow enough..

> Wouldn't 256 - s2[i] (which is by the way equivalent -s2[i]) make more
> sense than (s2[i] - 1) ^ 255 ??

I dunno, if you have a way that works it's ok, but my code is correct
and not slow anyway.

Rich