R600 Gallium3D Getting Close On OpenGL 3.3 Support

[marek]

No, I didn't run any of those games.

[Kano]

@marek

Did you run Doom 3 BFG or Rage with oss drivers yet? Those games need OpenGL 3.x.

[Kivada]

One would hope though that an exception could be made, just this once if OpenGL3.3 support can be had so that either the coming window is pushed back to include it or a special release that adds only that is put out in the interim for the distros to pick up so we aren't waiting till the 2014 releases before we see it.

[Kayden]

What are the odds of moving Mesa to the kernel's release style? Tag release --> Two week merge window --> 1 RC a week, with say a minimum of 5 RC's, that puts us at 7 weeks. Tag release. Pattern starts again.

I get the benefit of a nice steady 6month release schedule but something like the graphics stack-- like the kernel, I feel like needs to be a little more rapid release. This way if a feature isn't quite ready, its not a big deal to just release a month and a half or 2months later. The piglit tests help to ensure we don't regress and break things so its not even an argument of "Rapid release means rapid breakage!"

Any input from Kayden? Airlied? Marek?

I would be apprehensive about quickening the release process. Doing a release takes a fair amount of work---developers shift focus from adding features or optimizations to stabilization. That may mean fixing the inevitable regressions, or other bugs that have popped up, or just finalizing and polishing things they've worked on. This is a good thing, but it sort of puts forward progress on hold.

In my limited experience with working with the kernel and X server, it seems like the kernel process almost necessitates a person whose entire job is managing releases, merge windows, and so on. This makes sense for a huge project like the kernel, but it seems a high cost for us. A similar issue we have is stable branch maintenance---I know Ian has spent a huge amount of time on that, and could spend even more. (Recently, Andreas stepped up to manage that, which has been extremely helpful. We're very grateful!)

Piglit tests are definitely a huge help in mitigating regressions, but they're by no means foolproof---there are still huge gaps in our test coverage. Also, we don't yet have a system set up where we're running those tests on every platform, so they mostly get run on developers' machines. (We should fix that.) I've definitely fixed a number of 965GM crashers right before release that somehow no one had noticed. And especially these days, breaking people's 3D drivers can be pretty painful. I know distros are really concerned about that (and rightly so).

I do agree with Marek---Mesa master tends to be fairly stable and works most of the time. Breakage does happen, but it's usually caught and corrected fairly quickly. I think for advanced users who enjoy being on the bleeding edge and want the latest and greatest, using Mesa master is a decent option. It's reasonably easy to do these days, and we definitely appreciate the additional testing and feedback. But not everybody wants to be on the bleeding edge.

That said, it's very likely that we'll continue evolving the release process. Maybe we'll move to a shorter release cycle in the future. We'll have to wait and see

[Sidicas]

Okay, hence why I asked for your three's input specifically haha. Just out of curiousity, what is it that you don't like about the kernel's release style? Things seem fairly stable even during RC's (im running 3.8rc3 right now, before I ran rc1)

I would disagree..
Ever since 3.0, from my perspective, the kernel went to hell in a handbasket...

Around 3.0 is where the kernel platform drivers broke my wireless in my 7 year old HP laptop which was working perfectly fine for 6 years until then.. I've had to blacklist the HP platform driver to fix it.
They tell me to test the latest kernels (3.5 head, 3.6 head, etc.), but these kernels lock the system hard on the same laptop, so they won't even boot. The only solution that works perfectly is kernel 3.2 in Debian Wheezy and on top of that, I have to blacklist the HP platform drivers. If I do that, the system is rock solid stable and the wireless works flawlessly. it's also works flawlessly on 2.6* and 2.4* kernels since forever..

Then I bought a new laptop.. Turns out Kernel 3.2 that works perfectly on my other laptop, in Debian Wheezy, randomly locks up hard when running iceweasel/firefox on this new laptop. In fact, it seems to be a common occurance for people with mobile Ivy Bridge processors and intel graphics. I'm so lucky that the 3.5 head and 3.6 head kernels actually boot in this new laptop (although they don't on my other), and not only that, but it fixes the lock ups.. Why haven't the fixes in 3.5/3.6 been backported to 3.2 stable? I have no f-ing clue. Apparently the kernel team is leaving it to the poor folks at Debian to find these crash fixes in the mess of the 3.5/3.6 kernels and backport them in the hopes of making a stable kernel themselves because the kernel team isn't providing a stable 3.X release (yet!).. IMO, it feels like the kernel is a horrible mess right now. I've got 2 laptops that require very specific kernel versions (that are different!!!) to function at all, which means to me that the whole 3.X series is terribly unstable.

[Prescience500]

I'm curious about how Southern Islands is advancing. I'm sure that there's plenty of development going on, but Phoronix has been uncharacteristically silent about it for a long time. r600 seems to be shaping up nicely though. Great work Marek and all the other devs out there working on it!

[marek]

You can hardly say r600 is getting close to OpenGL 3.3 support when a *minor* detail called GLSL 1.5 is missing from 3.2, when GLSL is the most essential part of the specification.

GLSL 1.5 is mostly just a combination of the extensions mentioned in the TODO list below it and most of them are DONE.

[efikkan]

You can hardly say r600 is getting close to OpenGL 3.3 support when a *minor* detail called GLSL 1.5 is missing from 3.2, when GLSL is the most essential part of the specification.

[brent]

So what's the status of geometry shaders support? This is obviously a big and hard feature, but there's already some proof of concept code done, as far as I know. Is anyone working on it?

[marek]

A too-long stabilization period is the main issue. After the merge window, you have to wait 2.5 months before your code is released. If you miss it, you have to wait half a year (and it's rotting in some private branch for 3 months before Linus pulls it), which means waiting at least 3/4 of a year before distributions release it.