Intel ( i845G ) profiling

[Carsten Haitzler (The Rasterman)]

On Thu, 06 Mar 2008 19:39:51 +0100 Simon Thum <simon.thum at gmx.de> babbled:

> > anyway - i'm glad its useful. keep up the benchmarking and lets hope not too
> > long from now most chipsets and drivers are running smoothing with xrender.
> If one could make or point me to some short description on how to 
> measure, I'd glady devote some of my spare time. I still use some 
> howebrew XAA hack to evade the worst cases, even though EXA is faster in 
> general.

1. oprofile
2. run a test suite eg:
expedite -e x11 -a
expedite -e xr -a
expedite -e gl -a
(take a guess which rendering methods they use) :)
expedite can be gotten from enlgihtenment cvs (http://www.enlightenment.org).
it requires evas (same source).

:)

> > ten i can get onto my next favorite game and bitching about the quality of
> > the rendering ( e.g. filtered downscaling in xrender :):) )
> You've accidentially played my favourite game :) Thus, feel free to 
> ignore my ramblings.
> Filtered downscaling: I think it's broken (usually). People tend to 
> attribute quality shortcomings to the poor frequency response of the box 
> filter, but I think that in the vast majority of cases the true source 
> is the broken way of dealing with it. The supersampling/box filter/about 
> most stuff has to be performed in an intensity-linear space to yield a 
> correct result. See
> http://en.wikipedia.org/wiki/Gamma_correction
> http://www.w3.org/Graphics/Color/sRGB (at end of part 1)
> 
> IOW, supersampling or filtering without prelinearization is plain wrong. 
> This stems from the fact that addition of gammacoded values has no 
> pysical counterpart, strictly speaking, it is not defined. This won't 
> cause the computer to blow up of course, but output quality suffers. 
> Prefiltering successfully hides this in many cases, but makes matters 
> worse for others. Including glyphs, IMO.
> 
> If you knew that, as your name implies, I'm sorry for bugging you and 
> being a nuisance in general. But please, at least give it a shot before 
> implementing the 1000th well-meant workaround.

aah - yes. i do know of the theory that using gamma correction in conjunction
with down-filtering (box) can help look nicer, but frankly i'd be happy with
box filtering. right now the filtering you get is just "nearest" (point
sampling) or at best linear interpolation (so down to 1/2 size is ok, but below
is as awful as nearest/sampling).

in the end we can theorise on filtering, but we want it FAST. that means taking
whatever filters hardware happens to commonly implement. in this case
ansiotropic filtering would be a big improvement compared to "none worth
talking about" right now. crank it up to 8 or more levels and we may be talking
vaguely acceptable.

of course a pure box filter would be even better than this... and i suspect
some hardware may support that. beyond that you might not find much hardware
support and if its back to software in xrender with what i consider "phd
research level filter experiments" then i dont think this is something xorg
developers should be worrying about. if you want such a filter - it will end up
in software anyway and so you may as well implement it yourself and pre-scale in
software.

note - i am skipping fragment shaders here. i consider them even more esoteric
in the scheme of hardware and support and am not really thinking about them for
a scale filter. in theory it could be done, but given where we stand, and that
we'd probably want good quality - but we can argue on details of filters
forever, just ansiotropic would be a huge win.

> >> If someone wants me to profile something different, please feel free to
> >> ask :)
> As I wrote you just before I read this, I've got a very similar R280. FF 
> 2/3 (tabs) and thunderbird (scrolling the inbox) are unuseable to me w/EXA.
> Maybe you can profile this?
> 
> Thanks a lot,
> 
> Simon
> 


-- 
------------- Codito, ergo sum - "I code, therefore I am" --------------
The Rasterman (Carsten Haitzler)    raster at rasterman.com