Re: [AD] Color convertors |
[ Thread Index |
Date Index
| More lists.liballeg.org/allegro-developers Archives
]
> Actually the flaw was still there even after your fix. It only showed
> itself in the BeOS port though. The new convertors (which have just
> been committed by Angello) fix this, and also add support for other
> nice features.
You're right, there was still an off-by-one-pixel error after I had fixed
the off-by-one-scanline error. As I am a purist, I will probably touch again
the non-MMX code because your fix breaks the instruction pairing on Pentium
(and the non-MMX code was primarily targeted to it).
Moreover, I see you replaced the 'movl $0, reg' by 'xorl reg, reg'. Why ?
I'm not sure that's the way to go on Pentium.
> There's now a 16 -> 15 and 15 -> 16 convertor, xx ->8 convertors, 8->8
> lut, two additional MMX ones (32->24 and 24->32), 5% speed up on some of
> the convertors, and probably some other stuff I forgot about. The xx ->8
> ones are very slow (recommendations on speed it up are welcome!)
Great work !
However, the 16->15 and 15->16 convertors can't be enabled under Windows for
the time being (see the latest docs for the explanation).
> There are also dummy 15->15, 16->16, 24->24, 32->32 "convertors" which are
> implemented via a super fast memcpy :)
What are they for ? They won't be used under Windows for example.
> The width-must-be-multiple-of-4 is also rendered useless - the convertors
> will happily take anything you care to send to them :)
Great !
> The Windows, X and Qnx (Mac?) drivers will need updating though. The BeOS
> driver was already taken care of by Angello.
I'm going to update the Windows windowed driver.
Btw, could you upload somewhere your big diff and let us know the address ?
---
Eric Botcazou
ebotcazou@xxxxxxxxxx