FYI what I"m seeing right now is gears running properly on rv770 and mostly drawing black on an rv620 although dragging the window around makes the gears appear intermittently. Based on limited testing it is probably safe to say "seems to work on 7xx, colours aren't right on 6xx yet".
The same applies to a number of other tests, eg the redbook demos.
Does it matter which variant of RV770 you use?
According to wikipedia my Hd4870 would be an RV770XT, and I'd like to help testing the driver when I can. I'd help develop but I simply don't have the gfx card background for it.
And maybe you can also point me to the correct repository for this?
Based on the small amount of testing we have done so far (ie a couple of rv770s and an rv730 aka HD4670) your card should be fine.
If you go to my blog post and scroll down to the end I listed the build instructions I have been using. If you want to pull updates to drm and mesa afterwards you can replace the first three commands with :
Alex just merged the last few weeks of changes from master into r6xx-rewrite (in preparation for merging r6xx-rewrite back into master) and we're seeing some regressions, so you might want to hold off on pulling source for a bit. Only 587 files changed, including a lot of the common files used by 6xx/7xx, what could possibly go wrong ?
Monraaf, sounds like what I'm seeing on rv620, which makes sense because the 3D core is similar (actually hd3200 is like rv610). It's good to see you getting some performance benefit from your IGP
All the devs noticed that the sun seemed to shine a bit more brightly today
For the short term, I guess the tasks are :
(a) continue to fix bugs,
(b) get back to 1500fps from 15fps by accelerating the back-to-front buffer copy,
(c) get the mesa code ready to merge into master
The driver actually draws decently fast, it's just the buffer copy that slows everything down. The GPU can copy VRAM-to-VRAM at 2,000 MB/s or better, but the CPU can only copy at around 5 MB/s when it has to read from uncached VRAM. The CPU is probably reading one 32-bit word at a time, and taking maybe 800nS to go out to the GPU over PCIE, have the GPU read a word from VRAM, and return the data to the CPU. The glxgears window is ~300x300 pixels, or 360K bytes - 5MB/sec over 360KB/frame gives about 15 fps.
Just curious. Can you measure these speeds/bottle necks when the driver is running?
Originally Posted by bridgman
Running fullscreen (say 1680x1050) the copy really does start to give us "Frames per Day" performance
I don't hope AMD's markedtings department is getting ideas now
Funny, if I play a video with mplayer (Xv) and use glxgears at the same time and both the video and the gears window are visible then the gears are shown in the color that they are supposed to be, but as soon as I pause the video the gears turn black & white again.