Drawing Open-Source Triangles On The ATI RV770

[agd5f]

RS780 is now working, I've just pushed the fixes.

[mshopf]

RS780 is now working, I've just pushed the fixes.

For those who want to test - the fix is in radeonhd (RS780 wasn't detected as an PCIE device), so you have to fetch the git head.

[bugmenot]

Great!
I tested it and were able to draw triangles

But: I cannot do anything else, iirc everything with the CP fails (invalid argument and so on are still there, see last log)

And if I try to use exa it seems to work somehow. I can log in into gnome, but if I type something in somewhere, the letters are not drawn fine. And everywhere is a bit graphics corruption from time to time. Also scrolling is very slow, much slower than with the default, but faster than "noaccel". And if I move a window then both cores are used at 100% and if I move on, the computer crashes, I have to turn it off. I cannot enable/disable num lock any more and moving the mouse is working, but mouse pointer flickers somehow.

What shall I do?

Triangles are great anyway, though

[mshopf]

Great!
But: I cannot do anything else, iirc everything with the CP fails (invalid argument and so on are still there, see last log)

Not exactly. Drawing triangles already uses the CP

The EXA behavior is to be expected. It's not working completely yet.

[bridgman]

Yep. Right now the EXA code is moving one line at a time and (I believe) flushing the GPU pipe and caches after every line when copying overlapping areas, which is really slow (dragging and scrolling both involve copying overlapping src/dst areas). As bugmenot said, it's faster than no acceleration but not as fast as shadowfb. That shouldn't be hard to optimize, but right now we're trying to make sure all the programming information is available and solid before getting into driver optimization.

I'm not sure if we were still expecting corruption, will check with agd5f re: whether corruption reports are useful.

[agd5f]

There is corruption with the EXA composite (which is currently enabled) and occasionally with overlapping blits (I think due to cache flushing issues). At the moment if you want relatively fast and corruption free, disable EXA composite (Option "NoEXAComposite" IIRC) and enable a composite manager (e.g., xcompmgr -a).

Also as to Xv, the current shader is for nv12 formats, other planar formats should be easy (just changes to the swizzles of the incoming components in the texture setup). Currently, you only get a gradient rather than an image.

[bugmenot]

Hello,
I tried that, but the computer crashed again. It also crashes, if I only restart the xserver. Have always to take the pc the power away, I'm afraid this isn't good for my computer, if I do this too often.

Any way I can help to debug?
Thanks!

[monraaf]

I also get some ugly fonts and the server also locks up after a short time on my 780G with exa-acceleration. If I run X with just an xterm an no desktop and window manager it's more stable and even mplayer with Xv output does work! Xaa acceleration seems stable, but I get motion sick scrolling a web page and the Xv port is gone.

Looking at the Xorg log file from a locked up X session with exa-acceleration it gets spammed with a lot of:

[mi] EQ overflowing. The server is probably stuck in an infinite loop.
[mi] mieqEnequeue: out-of-order valuator event; dropping.

[agd5f]

Xaa acceleration seems stable

There is no XAA acceleration. It's falling back to shadowfb or none.

[monraaf]

There is no XAA acceleration. It's falling back to shadowfb or none.

Ok, It's still documented in the radeonhd(4) man page as an option however but maybe that's for the older cards.