The Embedded Linux GPU Mess & How It Can Be Fixed

[bridgman]

The other point Dave made was that the performance requirements (and hence driver complexity) for these small-footprint embedded systems seems to be lower than the norm for PCs, presumably because of the smaller screen size.

One thing that surprised me was how many of these systems have been using pure software rendering, with hardware acceleration only showing up very recently.

[bridgman]

Somedev, are you talking about open source drivers (where I agree) or proprietary drivers (where I disagree strongly) ?

If you look at the size of the proprietary driver teams for major HW vendors you'll find at least one order of magnitude more developers than can be "bought up" trained and ready to go. Not sure I understand your point about "developments that a very technical person could not independently produce with sufficient expertise" -- what matters is what you have today, not what you *could* have in the future.

Dave's point is that the dynamics may be different for embedded GPUs, which are typically below the lowest end discrete GPUs in terms of performance and features. That's an interesting point and worth looking at.

[somedev]

You would be surprised how little of the magic voodoo is hardware specific. Quite a bit of the proprietary driver techniques would be portable to competitors hardware, and even more so to a new competitor who independently designed hardware that was architecturally similar to yours.



If all of the hardware vendors contribute fairly to the common development effort, that's true. The problem is that it also effectively funnels benefits from vendors who do contribute to vendors who do not (ie the ones who aren't spending much on R&D in the first place). I suspect all the vendors agree that allowing open source developers to program their hardware is a Good Thing - the challenge is making that possible without giving up too much of your hardware advantage. That said, it can be done (and IMO should be done); it's just not as easy as some of the claims suggest.



Strictly speaking I don't think many hardware mfgs would agree with that. If you can give your hardware a competitive advantage via investing in software developoment it's not clear how removing that advantage and commoditizing the hardware benefits the hardware vendor.

Again, I'm not saying that vendors shouldn't support open source driver development, just that many of the seemingly obvious arguments don't actually hold up when you look at them more closely. I still think opening up hardware is a Good Thing, but if you want to understand why you haven't seen the wholesale rush to open source driver support that those arguments predict it's good to try to see things from the hardware vendor's POV as well.

Bottom line is that you need arguments which make sense to a hardware vendor. Making sense to everyone else is nice but doesn't get the hardware opened up. Dave's post seemed like a good step towards refining those arguments.

I find this line of reasoning quite suspect. R&D on the software side has in my observations followed merely this form: buy up software designers who already had the knowledge and/or interesting designs going for them before a penny was spent, and now they just get to apply that to one company or another's pet product. In terms of the rise of virtual machines - this happened with Sun/Java and Microsoft/CLR/etc. - and GPUs would appear to be going that way.

I would be quite surprised if R&D funding for software design under the driver stack has seriously resulted in developments that a very technical person could not independently produce with sufficient expertise. I suspect that most of what is considered proprietary magic voodoo has really been implemented independently by most serious graphics vendors and that it's nothing more than a strawman to keep drivers closed. Sounds good to the layman and the media, and that's what counts?

Especially as the hardware becomes increasingly generic, the technology issue is funneled back into the usual compilation and memory management strategies for which there is already plenty of publicly available literature on effective techniques.

[Henri]

Because Linux isn't the only operating system to use them - the BSD users would probably be a little upset if somebody attempted to change the licence.

Sure. (Although really, X11 and BSD allow you to do that, it seems to me that if you're upset when people actually use that ability you should have been using GPL/LGPL or similar in the first place. As a Wine developer I may be a bit biased here though.)

However, my point is more that if AMD, Intel, or other hardware vendors are concerned about competitors reusing driver code, I would have expected r600 for example to be released under the LGPL. In the broader view you'd also expect the hardware vendors and distributions (RedHat, really) to push for (or at least be in favor of) LGPL Mesa / Gallium3D, but as far as I know that discussion has never come up, which suggests to me that people are mostly happy with the way things are. I could see how in the past Tungsten Graphics could have made the argument that they depend on the ability to create closed drivers based on Mesa, but I don't think the same argument would work so well for VMware.

[archibald]

In that context, it has always surprised me a bit that both Mesa itself and most (all?) of the drivers are X11 licensed rather than something like LGPL.

Because Linux isn't the only operating system to use them - the BSD users would probably be a little upset if somebody attempted to change the licence.

[Svartalf]

In that context, it has always surprised me a bit that both Mesa itself and most (all?) of the drivers are X11 licensed rather than something like LGPL.

Well... X11's not the same thing as the 3D stuff is, to start out with. Couple that with inertia on licensing (I know there wasn't any impetus to change it with Utah-GLX when it was the thing and I was one of the contributors to it...) and it kind of makes some sense that it's still MIT/X11.

[bridgman]

...in the embedded space, Windows is to Linux, what Linux is to Windows in desktops, but they seem to be trying to enforce the same mindset.

Well, you know how much everyone hates having to come up with a new mindset all by themselves

[airlied]

It's a bit different in the CPU world because the API standard is still the instruction set, not a much higher level interface like DX or OpenGL. In the CPU world locking up the instruction set still makes a big difference. You may be right about hardware representing a bigger chunk of the secret sauce in the embedded space - I don't think that's the case in the PC market to the same extent though.

The PC market isn't an indicator is this situation, the first point I made is that the PC market is all driven by Windows development, in the embedded space, Windows is to Linux, what Linux is to Windows in desktops, but they seem to be trying to enforce the same mindset.

Dave.

[Henri]

You would be surprised how little of the magic voodoo is hardware specific. Quite a bit of the proprietary driver techniques would be portable to competitors hardware, and even more so to a new competitor who independently designed hardware that was architecturally similar to yours.

In that context, it has always surprised me a bit that both Mesa itself and most (all?) of the drivers are X11 licensed rather than something like LGPL.

[bridgman]

It's a bit different in the CPU world because the API standard is still the instruction set, not a much higher level interface like DX or OpenGL. In the CPU world locking up the instruction set still makes a big difference. You may be right about hardware representing a bigger chunk of the secret sauce in the embedded space - I don't think that's the case in the PC market to the same extent though.