new radeon tiling patch

[Roland Scheidegger]

Roland Scheidegger wrote:
> This scheme could potentially have a slight performance impact for 2d 
> use (will do some measurements later today) but I don't expect anything 
> noticeable.
Ok did some testing (x11perf -all -time 1 -repeat 3), with unpatched, 
patched but tiling disabled, and with tiling.
There does not seem to be a lot of difference, usually the results are 
identical. Some results are a bit lower with the new accel code, but 
nothing too dramatic imho. Comparison with tiling enabled or disabled 
doesn't reveal a whole lot of difference neither. As probably expected, 
vertical lines (and related geometric primitives like rectangle 
outlines) get a healthy boost (factor 2 or so faster). Other things 
(like for instance scrolling) are a bit slower with tiling enabled.
All tests done with a desktop resolution of 1280x1024, 85Hz, 32bit,
A64 3200+ (with 32bit SuSE 9.1), Radeon 9000Pro 64MB (rv250).
Full results here if anyone is really interested, use x11perfcomp to 
analyze (btw the version I got has some rounding problem, everything is 
rounded down, which leads to a division by zero somewhere when relative 
output is given):
http://homepage.hispeed.ch/rscheidegger/dri_experimental/x11perfunpatched.gz
http://homepage.hispeed.ch/rscheidegger/dri_experimental/x11perfnontiled.gz
http://homepage.hispeed.ch/rscheidegger/dri_experimental/x11perftiled.gz

Some q3 results as well (these are far more impressive...):
I've done tests with/without pageflip, with/without hyperz, with/without 
color (framebuffer) tiling, and with/without texture tiling (which 
currently leads to errors with small mipmaps, but I'd be surprised if 
that would affect performance). Same test conditions as above, plus q3 
running windowed 1024x768, highest texture detail, trilinear, and in 
general all quality options enabled, executing demo four. 
r_ext_compressed_textures 0 (enabling it would kinda defeat the purpose 
of testing tiled textures...).

pflip,   ctile,   ttile,   hyperz:   124
pflip,   ctile,   ttile,   nohyperz: 101
pflip,   ctile,   nottile, hyperz:   107
pflip,   ctile,   nottile, nohyperz:  88
pflip,   noctile, ttile,   hyperz:    87
pflip,   noctile, ttile,   nohyperz:  74
pflip,   noctile, nottile, hyperz:    77
pflip,   noctile, nottile, nohyperz:  67
nopflip, ctile,   ttile,   hyperz:   107
nopflip, ctile,   ttile,   nohyperz:  89
nopflip, ctile,   nottile, hyperz:    94
nopflip, ctile,   nottile, nohyperz:  79
nopflip, noctile, ttile,   hyperz:    77
nopflip, noctile, ttile,   nohyperz:  67
nopflip, noctile, nottile, hyperz:    70
nopflip, noctile, nottile, nohyperz:  61

So, for Q3 and the rv250, color tiling has the biggest effect on 
performance (30-40%), hyperz is a distant second, and pageflip and 
texture tiling are battling for the 3rd place :-). (note that texture 
tiling will help much less with lower texture detail and/or bilinear 
filtering according to some quick measurements).

Roland