keithp at keithp.com
Mon Apr 18 16:51:55 PDT 2005
On Mon, 2005-04-18 at 19:40 -0400, Zack Rusin wrote:
> On Monday 18 April 2005 17:51, you wrote:
> > scanlines don't deal with filters and transforms well at all; I'd
> > like to see this code use square patches (8x8 or so) which seems like
> > a good fit for both MMX and transforms.
> I think i mentioned on irc at some point that I will try this approach
> but I'm wondering what makes you think that they don't deal with
> filters or transforms well.
General convolutions require data from several adjacent scanlines, which
means lots of data if you have long scanlines. Regular patches means
you've got fixed memory needs for arbitrary image sizes.
Same for transforms, except worse when rotating as you'll use very
little information from each scanline.
Other systems (gimp) use patches instead of scanlines.
Also, with fixed size patches, you'll eliminate lots of variable values
in loops, enabling (I hope) nifty optimization opportunities with
completely fixed computations.
But, I have to say I haven't implemented any of this, so I may well be
speaking out of my ear.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 189 bytes
Desc: This is a digitally signed message part
More information about the xorg