From: Felix <fx...@gm...> - 2005-02-08 21:37:39
|
Am Dienstag, den 08.02.2005, 20:38 +0100 schrieb Roland Scheidegger: > Ok, I was FINALLY able to come up with something for texture tiling=20 > which seems to work - this was very very annoying, it _almost_ worked=20 > literally within minutes, but I needed a lot of time until it finally=20 > did really work. > I needed to convert back the drivers to use a multi-byte upload scheme=20 > again (they did this years ago until it was abandoned for a simpler=20 > method), since if the blitter auto-tiles the textures on upload color=20 > format and texture pitch need to match the exact texture properties (at=20 > least I was unable to make it work otherwise). > Also, there are tons of special cases for small textures in the drm,=20 > which I don't like but couldn't avoid. > I could not make it work for texture rectangles, I think at least on=20 > r200 microtiling for that case should work, but it didn't. Well I don't=20 > think it's important. > It is only tested with "normal" 16-bit and 32-bit RGB/RGBA textures, in=20 > theory all other formats should probably work too I hope. I lack a good=20 > test, tests/manytex and redbook/mipmap are a bit too simple (and you=20 > can't see if the tiling is actually correct...). 8-bit formats should=20 > work too, if I got the math correctly... FWIW, text in foobillard uses an 8bit texture format, at least with the Savage driver, in case someone wants to test it. I don't have my Radeon installed right now to test it myself. > Also, there seem to be some minor differences between r100 and r200=20 > chips as far as macro tiling is concerned (didn't have much time to test=20 > on r100 though), and for the heck of it I couldn't figure out what that=20 > second micro-tile bit is good for on r100. >=20 > Quake3 got about a 15% boost on a 9000pro, and 11% on a 7200 sdr, if the=20 > highest texture setting/trilinear/32bit was used. Well, compressed=20 > textures are still faster :-). Cool stuff. I tested HyperZ and color tiling for the first time last weekend on my Radeon 7500. I found the speedup pretty impressive. A shame that this potential had been wasted for such a long time. >=20 > For the drm, I've also included Andreas Stenglein's cube map patch for=20 > the r100 (since this needs a version bump as well). > I've only attached the -core version, I actually plan to do a non-core=20 > version too, but I've noticed the non-core version does no longer seem=20 > to get all fixes the core version does. So am I the only one who still=20 > checks things in for that version? In that case I'll immediately stop=20 > touching it... >=20 > Roland --=20 | Felix K=FChling <fx...@gm...> http://fxk.de.vu | | PGP Fingerprint: 6A3C 9566 5B30 DDED 73C3 B152 151C 5CC1 D888 E595 | |