Open-Source 2D, 3D For ATI Radeon HD 5000 Series GPUs

[admiral0]

No 2d or 3d for me..

I have an acer laptop with switchable ironlake/radeon hd 5470 graphics. I have compiled code from evergreen_accel branch, but it hangs when starting X.

Anyone has this issue?

[bridgman]

[*]geartrain didn't crash the system, but it took a trip to the 60s: screenshot. Those colours change, synchronized to the turning of the gears (normal- and colour-buffers mixed up or something? Then again, why don't we see non-saturated colors?). Those red stripes weren't visible, they turned up sometime during making and copy/pasting the screenshot around.

Whoa, that's a funky looking geartrain...

[agd5f]

Not all power tables specify lower memory clocks (especially on desktop cards and in multi-head power states).

[rohcQaH]

thanks, just updated mesa, let's see..

• glxgears: I don't see corruption.
• neverball: colours still wrong -> screenshot. Even though glxgears behaved this time, neverball seemed to corrupt others (see the red stripes in the window border for example)
• geartrain didn't crash the system, but it took a trip to the 60s: screenshot. Those colours change, synchronized to the turning of the gears (normal- and colour-buffers mixed up or something? Then again, why don't we see non-saturated colors?). Those red stripes weren't visible, they turned up sometime during making and copy/pasting the screenshot around.
• When I tried to do stuff with geartrain running, X crashed. The backtrace lacks debugging symbols, I'll see if I can fix that. Maybe xorg doesn't like splitdebug? Hmm..
• kde4: quickly got a segfault similar to the above

Those red stripe-corruption only appears when 3d apps are running. Using a pure 2d desktop (or using xv) seems to work quite well so far.


My second screen turning black: happened again, reproducible. Switch to a VT, back to X, second screen is off. As said, a workaround is to re-issue
xrandr --output DVI-1 --off; xrandr --output DVI-1 --auto --right-of DVI-0
Now where would I report that bug? kernel/KMS? Xorg? xf86-radeon? Or is it known already?


Power Management:

 ~> echo "low" > /sys/class/drm/card0/device/power_profile
 ~> cat /sys/kernel/debug/dri/0/radeon_pm_info 
default engine clock: 850000 kHz
current engine clock: 399990 kHz
default memory clock: 1200000 kHz
current memory clock: 1200000 kHz
voltage: 950 mV

doesn't look like it's working. I'm using a dual-head system, but even with one monitor switched off (xrandr --output DVI-1 --off), the memory clocks won't go down.

[agd5f]

I just pushed a bunch of evergreen patches to r600c in mesa. They should fix at least some of the issues you are seeing. Also, changing the memory clocks on evergreen works just fine; same as other asics. See this page for more on the pm options:

http://wiki.x.org/wiki/RadeonFeature

[rohcQaH]

New summary (compare to my last post):

2D:

• xv works again in latest git (-> bug)
• haven't seen any flaws in 2D operations since applying the patch from this bug. RenderAccel is on and works.

Will try to use the oss drivers to get some real work done, let's see how that turns out.

3D, mesa classic:

• neverball starts now, but some of the colours are slightly wrong. Geometry appears correct.
• glxgears still overwrites parts of the VRAM it shouldn't touch
• still got
  

  radeon 0000:02:00.0: forbidden register 0x00009508 at 10
  [drm:radeon_cs_ioctl] *ERROR* Invalid command stream !

  same register as reported before, different app this time, more recent kernel (2.6.36-rc3). I'll try to figure out if I can consistently reproduce it and file a bug.
• haven't tried anything else after the crash

3D, gallium3d:

~> glxinfo | grep render
direct rendering: Yes
OpenGL renderer string: Gallium 0.4 on EVERGREEN

• glxgears runs without any kind of corruption, but after a while it stalls with this being spammed through my dmesg a few hundred times:
  

  [drm:radeon_cs_ioctl] *ERROR* Failed to parse relocation -22!

• neverball: different problems than r600c: both colours and geometry are correct, but the menu text is invisible. It's seems slower than on r600c. Stalls after a few seconds with the above dmesg-spam.
• geartrain runs, but it appears that either the viewport isn't set correctly or some transformations failed; only parts of the scene are visible zoomed-in.
• ioquake3 starts, but glyphs in the menu are broken, textures appear to be missing. Didn't start a game though.

somewhere along the way my second monitor turned itself off, despite xrandr claiming it was running. Not sure what happened there, but it was fixed by xrandr --output DVI-1 --off; xrandr --output DVI-1 --auto --right-of DVI-0

I wish there was memory downclocking, because the card runs a bit louder than on fglrx.

[Wingfeather]

Excellent stuff! I get no crashes now.

New summary:
In KDE, using openGL compositing with RenderAccel disabled, I get almost no corruption at all! Prefectly usable. Though, most of the icons in the taskbar and on the desktop are corrupted.

E17 still won't start up at all with its compositing module (either with or without RenderAccel).

Xv is very slow (but works). When displaying a video scaled up to full-screen, there are a lot of blocking-type artefacts that don't seem to be there when displaying at the native size (though I can't be sure) and motion is choppy. But it works, which is a huge step forward.

glxgears still introduces the horizontal lines (across the background/wallpaper, in my case) mentioned by others. In KDE, dragging windows around sees the frame-rate go pretty low (lower than when software does it) and, as ever, resizing a window takes *forever*. Probably cranks out a new frame every 6-7 seconds or so.

Some niggles, to be sure, but a huge improvement overall! Thanks to all the devs for their fantastic work.

[V!NCENT]

Awesome, no more GPU lockups with that patch. Xv is also working again and it seems the low power profile is enough to render videos at 720p.

Time to put my HD5770 back in or not? (and learn doing lego yeah-yeah-yeah )

[monraaf]

Awesome, no more GPU lockups with that patch. Xv is also working again and it seems the low power profile is enough to render videos at 720p.

[pingufunkybeat]

w00t w00t! Great job!