Intel ( i845G ) profiling

Simon Thum simon.thum at gmx.de
Fri Mar 14 05:14:26 PDT 2008


Carsten Haitzler (The Rasterman) wrote:
> On Thu, 13 Mar 2008 22:50:16 +0100 Simon Thum <simon.thum at gmx.de> babbled:
> 
>>> ten i can get onto my next favorite game and bitching about the quality of
>>> the rendering ( e.g. filtered downscaling in xrender :):) )
>> I hadn't assumed this runs through the 3D pipeline at all. Also, XRender 
>> accel is a domain of the driver, so things may vary. Does it?
> 
> on the r200 it does go through the 3d pipe. it doesn't much matter what pipe it
> ends up going through, but as long as it is accelerated and can draw with
> decent quality... that's what we want (and accelerated well, not in such a way
> that it's slower than software). in this day and age such accel likely will go
> through the 3d pipe using fbo's ad rect textures etc. and thus the whole
> texture filterign talk.
Thanks for explaining me. I've got the r350 in my laptop. Just out of 
curiosity: I got it that 3D functions get 'swithed off' as a whole,
thus reducing power consumption for 'pure' 2D. Does that mean my 
render-composited
glyphs make the laptop draw more power? (ignoring any freetype overhead 
of course)

> in the end we can talk about all sorts of funky down or up filtering, BUT what
> we need is some form of decent filtering at all. full box filters would be
> nice. but ansiotropic with a high level would be a compromise an d give good
> quality and be very fast as its a core texturing feature in most gpu's - no
> need for fragment shaders etc. of course in future other exotic filters can be
> discussed but just some decent filter - just by enabling ansiotropic filtering
> at a high level (4, 8 or 16 or more) would be dead-simple to enable and work
> well, if you are going via the 3d pipe for this.
As long as it gets any better I'm fine with it. It wasn't my intention 
to tell people what to do, just to give the chance to avoid a far too 
common error. My personal favourite (ramblings follow):

http://lear.inrialpes.fr/people/triggs/pubs/Triggs-iccv01-subpix.pdf

That's a prime example of 'downscaling quality sucks, it must be the 
filter function, here's the never-optimal PRF as proof'. Not that I had 
heard of anyone performing scaling in frequency domain, but the solution 
suddenly has to be there? IMO, that's fixing a screwed implementation 
with another screwed assumption.

So to say the least, you're in best company. (Not meant) And yes, for 
the r200, it probably won't get us far to talk about funky filtering.

Regards,

Simon



More information about the xorg mailing list