Mesa Developers Discussing Again Whether To Fork Or Drop Non-Gallium3D Drivers

[TemplarGR]

In short - modern APUs can deliver the GPU performance from GPUs that are 5+ years old and are cheap as hell. So if performance is an issue, but budget is a limiting factor: There are solutions.

LOL, in your dreams, perhaps. Even the best APU in stores today can barely catch up to a 100$ dGPU (for example, AMD R7 260x) from 2015 in performance. In 2014-15 we had stuff like 7970 and Fury, and no APU will touch those in some time to come... The main issue is VRAM bandwidth, and until this gets solved by some L4 cache or HBM solution, APUs will be left to play simple/old games and display GUIs, that's it.

As for old hardware, the issue is not that the hardware is old, the issue is that Intel were stubborn for years and refused to switch to the gallium infastructure for their drivers until recently. AMD has a gallium driver for even the R300 architecture, which is approaching 2 decades of existence...

I disagree with creating a MESA LTS version. For many reasons. One is that LTS usually means "only bugfixes allowed" but there could be feature work that could apply to Classic drivers in the future, so why call it LTS? And another reason is that we don't need another kernel situation where we have 5-6 different MESA LTS versions, LOL.

I think what MESA should create is a MESA CLASSIC version, and put all the classic MESA drivers there. People with classic MESA drivers will be installing the Classic package, and people with modern gallium drivers will be installing the normal MESA package. Seems clear to me.

[-MacNuke-]

Bug fixes in shared code? It *will* bit-rot out of tree. How many out-of-tree Mesa driver drivers are maintained? Overhead and additional system complexity from GLVND?

Since Intel switched to Gallium it also bit-rots in mainline. Let's face it: Pretty much nobody cares for them right now and fewer will care for them later no matter how long they stay in the main tree. It's not like that the mesa devs are checking every chipset they support before every release.

I think splitting the trees is a benefit to both sides. Modern drivers can develop without being hold of by decades old hardware and decades old hardware have a place to get some visibility.

And since OpenGL is pretty much dead at this point the Gallium/OpenGL part of Mesa becomes even more irrelevant in the future. It will all be "legacy software support" at some point. And even that could be mapped to Vulkan (i.e. Zink)

[kpedersen]

I think it boils down to the fact that Linux is not just a toy for kids and gamers. A high number of machines with this kind of hardware are still used within the industry. Even commercial operating systems like Windows provide a driver for them out of the box.

In the ideal world, they *should* be dropped but then replaced with a wonderful Gallium3D based replacement

I the Linux community really struggling to find man power to maintain them? Or is Intel being wasteful? This could be an example of relying too much on a company when it comes to digital / hardware preservation.

[Shevchen]

Sorry, but as you allude to this is nonsense. "Performance" is arguably even *more* important on older hardware, especially mobile devices like laptops because without acceleration, they're pretty much useless. There is no "modern tech race", either the hardware performs adequately for the given task, or it doesn't. If you want to run AAA games titles, you need a capable system, there's no other way around it, dropping support for last years hardware isn't going to turn today's budget systems magically into PS5s.

You have missed my point. Old hardware that runs today on todays tasks doesn't need any more improvement, as it is already doing what it is supposed to do. As it has been mentioned already, there have been only minor changes over the past years and Intels IGP are generally not suited for "gaming".

And we do have a tech race - there is a reason why ACO is pushed so hard.

What if you have *no* budget for a new system? Why would you want to replace a 5+ year old system with a new one with equivalent performance anyway?

Then you ride on the legacy drivers.

Intel dropping support for Beignet is a real PITA*, I have an high end mobile IVB GT2 which despite claims to the contrary achieves pretty reasonable OpenCL performance, certainly better than not using it! But Intel decided to only support GEN8+ with the new driver. Same situation with the new Gallium3D driver, I'm sure few would have issue if they supported all of their OpenGL4 capable hardware, but Ivybridge/Haswel are slightly different to later revised designs while nearly as capable.

I do agree on the point that Ivy Bridge isn't "old" - but the IGP in those processors was and is weak to begin with. Gaming on those chips in a case for "low spec gamer" wich in itself is already APITA. You can get pretty cheap GPUs that have a heap load of more processing power (RX580) which will carry a long time.

The "no budget" point is a problem, but I also don't demand raytracing if I only have 50 bucks in my hand. (broadly speaking)

There is an extension of your suggestion - that would sort of make more sense; that's to drop all support for Intel CPUs with critical security flaws which require kernel mitigation. That really would allow a lot of code to be dropped from the kernel and the possibility of higher performance...

* I've not been able to build Beignet for a while (can't figure out why) and just to rub in salt I've not been able to get the legacy Fermi NVIDIA driver to work recently so I've lost all OpenCL support on this system!

You don't drop it, you have your legacy code that runs as fast as always, but developers don't have to waste their time on those things anymore, while they try to tweak the new cards for their optimal performance. Hardware and software advances over time and at a certain point, the overhead to support legacy hardware becomes so big, that the benefits don't outweight the invested time and effort.

I think what MESA should create is a MESA CLASSIC version, and put all the classic MESA drivers there. People with classic MESA drivers will be installing the Classic package, and people with modern gallium drivers will be installing the normal MESA package. Seems clear to me.

I think that is reasonable. As for the APUs: The 3400G is strong enough to give you ~30-40FPS on 1080p. Not great, but usable. Tweaking the memory can give you a huge boost on those APUs. But we are already talking about overclocking here, which few people do.

[Raka555]

(And in Intels case: Who plays games on an IGP?)

To your surprise, Intel has 10% of the graphics market according to steam: https://store.steampowered.com/hwsur...lcome-to-Steam
BTW AMD only has 15% ...

[yoshi314]

LOL, in your dreams, perhaps. Even the best APU in stores today can barely catch up to a 100$ dGPU (for example, AMD R7 260x) from 2015 in performance. In 2014-15 we had stuff like 7970 and Fury, and no APU will touch those in some time to come... The main issue is VRAM bandwidth, and until this gets solved by some L4 cache or HBM solution, APUs will be left to play simple/old games and display GUIs, that's it..

i always thought APU used system ram , which would avoid the bandwidth issues. i guess i was wrong there.

[kpedersen]

To your surprise, Intel has 10% of the graphics market according to steam: https://store.steampowered.com/hwsur...lcome-to-Steam
BTW AMD only has 15% ...

Not to mention, these days the GPU isn't just about games. You can barely run a Wayland compositor or web browser effectively without a GPU driver stack. Which I agree is daft. But hey, developers are daft!

[Raka555]

Not to mention, these days the GPU isn't just about games. You can barely run a Wayland compositor or web browser effectively without a GPU driver stack. Which I agree is daft. But hey, developers are daft!

I agree, but the developers do as they are told/asked.
It is the industry and the perception of what "modern software" are, that are daft.

[RussianNeuroMancer]

If anyone need example of how much Gen.7 is broken on mesa mainline - just take a look at this screenshots.

[ix900]

I assume it wouldn't be an argument if Mesa was designed in a way that makes this point void. That there would then be the real problem, Mesa itself. They obviously bungled it so that its tangled and causing conflict.

Mesa devs should work to fix it, not have debates if it should be fixed.