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

[bridgman]

Hi Kano;

Is there a real-world use case for switching between fglrx and radeon other than testing ? I understand the desire for fglrx-to-fglrx and radeon-to-radeon (basically "update the driver without rebooting just like Windows") but not sure how to argue for switching between open source and proprietary without reboot.

If your measurement of "betterness" is "programs the same registers the same way" then presumably a reverse-engineered driver is going to win every time. In fairness though, that's not a common way to determine a "good driver".

"It works just like fglrx so it must be the best" ?

[Kano]

Well in theory it is possible to install fglrx in live mode when you add

radeon.modeset=0

and an option to stop X from loading. Then you can install fglrx. But usually thats a bit hard to remember for most people. Use full for example for a games live image which does not require hd install. Why do you think it is needed to do a hd install all the time just for REBOOTING after a driver install, that must be a joke...

[V!NCENT]

@Bridgman,
There is definately a real world use case:
There are millions of people installing Linux by themselves. Even if they bought a pre-loaded Linux computer, they might want to upgrade to newer versions.

Ubuntu, for example, doesn't ship with fglrx by default. So what that means is that users first have to reboot after the install, but then install fglrx and then reboot again. After they found out that for example standby isn't working out of the box and so they switch back. Reboot again.

Then their son wants to play a game. Oops, fglrx again so reboot. But now daddy wants his standby back.

Stupid use case maybe, but these things do happen =x

[nanonyme]

but not sure how to argue for switching between open source and proprietary without reboot

In good old days this usually resulted in a hardlock because registers values have been changed since boot to something unexpected. Dunno whether that's still an issue. Iirc cold booting was even sometimes necessary to clear them up.

[nanonyme]

Dammit, edit timeout.
Still sleepy apparently, didn't notice who that was to.
Anyway, I recall hearing ages ago that to set the registers to the value on boot you'd need to record the values on boot and restore on module unload. It gave me the impression some of the registers might not have guaranteed generic default values.

[Qaridarium]

Precise and Correct is fine.

Precise and Incorrect, not so much

If you drop the precision a bit your chances of being correct go up.

i think the trick is do not try to be right and true is what the people want it to come true.

[rohcQaH]

http://www.overclock.net/linux-unix/...on-driver.html

Note the new instructions starting with 2.6.35. There is also a "mid" profile now IIRC.

thanks.

"dynpm" didn't seem to do anything. Setting the profile to "low" immediately lowered the fan speed, but fglrx is still more silent.
I didn't find a method to output current frequencies, though, so my only measurement is fan noise.

there's a lot of lines during 2d activity (namely xterm):

CS section size missmatch start at (evergreen_accel.c,evergreen_set_default_state,871) 42 vs 45
CS section end at (evergreen_accel.c,evergreen_set_default_state,912)
CS section size missmatch start at (evergreen_accel.c,evergreen_set_default_state,981) 45 vs 46
CS section end at (evergreen_accel.c,evergreen_set_default_state,1023)

These messages can be silenced by changing the numbers in BEGIN_BATCH in the lines indicated, though I'm not sure if that's a proper fix.

glxgears works (~2015 fps), but causes gfx corruption, as if parts were rendered to the wrong parts of the framebuffer, visible as horizontal lines.

Taking a screenshot doesn't quite work. I tried both import and ksnapshot, both result in something like this, which is far more corrupted than what appears onscreen.

~> neverball
neverball: radeon_cs_gem.c:181: cs_gem_write_reloc: Assertion 'boi->space_accounted' failed.
Aborted

ioquake displays it's brown background, then appears to hang.

startkde3 runs (compositing disabled), but there's corruption at some icons, all window borders and the mouse pointer.

Watching a movie (xv) worked halfway through, then X crashed. Last lines of xorg.conf:

vs const buffer map failed
ps const buffer map failed
ps const buffer map failed
ps const buffer map failed
ps const buffer map failed
ps const buffer map failed

Fatal server error:
failed to map pixmap -12

x86_64 smp
HD 5770
X Server 1.7.7
kernel 2.6.35-gentoo-r1
libdrm-2.4.21
xf86-video-ati, current evergreen branch
mesa git from yesterday

no error messages im dmesg.

[rohcQaH]

/edit

of course the question of my first post remains: which of those issues are worth turning into proper bug reports, and what additional information is needed to make those useful?

[trapxvi]

If you open bugs, link them and I'll confirm what I can.

there's a lot of lines during 2d activity (namely xterm):

CS section size missmatch start at (evergreen_accel.c,evergreen_set_default_state,871) 42 vs 45
CS section end at (evergreen_accel.c,evergreen_set_default_state,912)
CS section size missmatch start at (evergreen_accel.c,evergreen_set_default_state,981) 45 vs 46
CS section end at (evergreen_accel.c,evergreen_set_default_state,1023)

I can confirm this on my setup.

glxgears works (~2015 fps), but causes gfx corruption, as if parts were rendered to the wrong parts of the framebuffer, visible as horizontal lines.

Ditto.

Watching a movie (xv) worked halfway through, then X crashed. Last lines of xorg.conf:

vs const buffer map failed
ps const buffer map failed
ps const buffer map failed
ps const buffer map failed
ps const buffer map failed
ps const buffer map failed

Fatal server error:
failed to map pixmap -12

This too. I don't get the buffer map fail messages, but the failure to map pixmap is how X dies after several minutes of use. Mplayer (Xv) or Firefox seem to be surefire ways to kill X within a minute or so, but I've seen this once scrolling in urxvt.

Core 2 Quad Q8300, Running 64 bit kernel and X.
Intel G43 Chipset
Sapphire HD 5770 (0x68B8)
X Server 1.8.1.902
kernel 2.6.35.3
libdrm from this morning's git
xf86-video-ati from this morning's evergreen_accel branch
mesa git from this morning

[rohcQaH]

The "CS section size missmatch"-errors are already fixed in latest git. Apparently it WAS enough to just change the numbers.