HARDWARE: Open-Source-Friendly Graphics Cards -- Viable?
I'm posting from home, so this won't look right. Sorry.
Anyhow, Andre Eisenbach said this:
Quoted 1 line
Quoted 1 line
Quoted 1 line
If the graphics card mostly supports 2D initially, it's really not much better then just about any off the shelf graphics card with VESA drivers. As in, the hardware doesn't need to be open for just that. Most (all?) the frustration in Linux graphics card land comes from unsupported/closed 3D drivers.
<<<
I have tried using cards with VESA drivers before, and I found it to be very painful. Certainly, you can turn off certain features and get a reasonably useful UI experience, but dragging windows around with "show window contents while moving" enabled is painfully slow, even with AGP 4x. Just imagine doing it over PCI.
When it comes to desktop applications, the FIRST thing you need is good 2D acceleration. In fact, that's really the ONLY thing. OpenOffice does not need to use OpenGL. GNOME doesn't need to use OpenGL. In fact, for the most part, they don't bother. There are some instances where they use OpenGL, but most of what a workstation user does fits squarely within all the functionality supplied by Xlib, which is entirely 2D.
Ok, so with limited 3D support, it's almost not worth trying to play Doom II (let alone III), but that's never affected me. On Linux, I use nedit, Mozilla, GNOME, KDE... ALL 2D apps. I use Linux as a development platform for chip logic, X server modules, and web sites. I also do a fair amount of tinkering with CPU-intensive things like genetic algorithms. In fact, the ONLY time I have EVER played with 3D on Linux was when I fiddled with some of the screensavers.
Ok, so I'm really limited in my use. But what about what a secretary would use? GNOME, OpenOffice, Evolution, Mozilla. All 100% 2D apps. How about a sysadmin? Well, he wants something simple in his server that lets him run his Red Hat configuration tools. What's a system integrator want? Something that won't result in any tech-support calls.
Nevertheless, I do think 3D is very important. MacOS has moved to pure 3D, and Longhorn's Aero Glass thingy is all 3D too. Plus there's that Sun thing. With Linux, we're kinda constrained by the fact that core X11 protocol is strictly 2D, but soon, GNOME and KDE will surely find a way around that too. I know we could not sell a graphics device which did not have ANY 3D support. But keep in mind that the more sophistocated the 3D support, the larger an FPGA you'll need. The prices of FPGA's go up exponentially with die area.
I've been given a very limited budget here for development. (Well, I haven't been given a budget yet--I've just been told that we're not going to spend $100K to do an ASIC for something this speculative.) I'm further constrained by the impact of FPGA chip cost on the end product.
Here's an off-the-cuff guess as to the parts cost for one board (I'm sure I have most of the prices wrong):
- FPGA $30
- PCB $5
- DAC $10 - DVI transmitter $10 - RAM $20 - Assembly $??
- Development cost $??
- Profit $??
The parts alone are $75, and I've left plenty out. If the board is profitable at $100... who will buy it?
I'll do whatever anyone wants, as long as it fits into these constraints!!
One idea I have is to merge the 2D and 3D pipelines into one. This way, we can get better 3D functionality, and 2D comes in for free. The problem is that 2D performance would be a LOT slower in this case. But forget I said that, because it's absolutely pointless to talk implementation details at this point.
The whole issue comes down to this: This is technically feasible. Should we do it?
The open source community often complains about hardware vendors being too tight-lipped with their IP. Here, we have a golden opportunity to get what we want, both in terms of features and disclosure. How do we figure out how to not miss that opportunity?
In case you're wondering why I'm writing so much about this... it's because I REALLY REALLY REALLY want to do this. As a geek who enjoys designing stuff, this is an exciting idea to me. I'm also a free software zealot, but I often feel like a leech because I haven't given anything back (well, there's GTerm, but who cares.). I just don't know how to justify the cost of this project to my employer.
__________________________________
Do you Yahoo!?
Take Yahoo! Mail with you! Get it on your mobile phone.
http://mobile.yahoo.com/maildemo
On 2004.10.21, Timothy Miller wrote:
...
Quoted 9 lines
When it comes to desktop applications, the FIRST thing you need is good 2D acceleration. In fact, that's really the ONLY thing. OpenOffice does not need to use OpenGL. GNOME doesn't need to use OpenGL. In fact, for the most part, they don't bother. There are some instances where they use OpenGL, but most of what a workstation user does fits squarely within all the functionality supplied by Xlib, which is entirely 2D.
Have you looked at xorg-x11 recently ? IE, the Composite, Damage and Render extensions ?
OSX uses OpenGL because it is the API they have to access things like alpha blending, image scaling, and so on, so they can do those nice effects of transparencies, shadows, genie's and so on. At least until Panther. For me, it looks like the new Tiger implementation (CoreImage) is their own implementation of the OpenGL pixel pipeline, talking directly to drivers instead of using OpenGL as intermediary.
Probably desktop systems would not need the T&L part of 3D, but be sure they will need at least managing windows at different depths, blending them, anti-aliasing them an so on.
So, as I see it, for an appealing 2D card, you need to program a 2 1/2 graphics engine, with really _fast_ alpha blending and antialiasing. You can only kill the matrix part. I do not know if you will be able to get rid completely of floating point, for those alpha mixes and assorted candy...
--
J.A. Magallon <jamagallon()able!es> \ Software is like sex:
werewolf!able!es \ It's better when it's free
Mandrakelinux release 10.1 (Community) for i586
Linux 2.6.9-rc4-mm1 (gcc 3.4.1 (Mandrakelinux 10.1 3.4.1-4mdk)) #4
Quoted 35 lines · --- "J.A. Magallon
On 2004.10.21, Timothy Miller wrote:
...Quoted 2 lines
When it comes to desktop applications, the FIRST thing you need is
good
Quoted 1 line
2D acceleration. In fact, that's really the ONLY thing.
OpenOffice
Quoted 2 lines
does not need to use OpenGL. GNOME doesn't need to use OpenGL. In fact, for the most part, they don't bother. There are some
instances
Quoted 1 line
where they use OpenGL, but most of what a workstation user does
fits
Quoted 3 lines
squarely within all the functionality supplied by Xlib, which is entirely 2D.
Have you looked at xorg-x11 recently ? IE, the Composite, Damage and Render extensions ?
OSX uses OpenGL because it is the API they have to access things like alpha blending, image scaling, and so on, so they can do those nice effects of transparencies, shadows, genie's and so on. At least until Panther. For me, it looks like the new Tiger implementation (CoreImage) is
their own implementation of the OpenGL pixel pipeline, talking directly
to drivers instead of using OpenGL as intermediary.Probably desktop systems would not need the T&L part of 3D, but be sure
they will need at least managing windows at different depths, blending them,
anti-aliasing them an so on.
Quoted 8 lines
So, as I see it, for an appealing 2D card, you need to program a 2 1/2
graphics engine, with really _fast_ alpha blending and antialiasing. You can only kill the matrix part. I do not know if you will be able to
get rid completely of floating point, for those alpha mixes and assorted
candy...
Alright. Excellent points. If I don't have to do any scaling or rotation, alpha-blending won't be that big of a deal. I already would have to read the destination of applying a raster-op or a planemask, so alpha-blend would become just another part of the merge at the end of the pipeline.
Of course, with only 8 significant bits, you could get noticable cumulative round-off error.
As you start to add 3D features to the 2D pipeline, the point in keeping them separate diminishes.
Oh, and there's antialiasing... now, sometimes, it just LOOKS like antialiasing, like with the font server used by X11. When it renders characters, it produces an 8-bit grayscale bitmap which is pre-antialiased, which you then have to color while rendering. If you want to do other things, then you get back into scaling, which is just a special case of texture-mapping.
_______________________________
Do you Yahoo!?
Declare Yourself - Register online to vote today!
http://vote.yahoo.com
Quoted 5 lines · J.A. Magallon
So, as I see it, for an appealing 2D card, you need to program a 2 1/2 graphics engine, with really _fast_ alpha blending and antialiasing. You can only kill the matrix part. I do not know if you will be able to get rid completely of floating point, for those alpha mixes and assorted candy...
Alpha blending (and Porter-Duff operations in general) can easily be done in integer. You do want to read Rasterman's excellent comments in the Imlib2 sources, though.
Been there, done that... (for a STB without FPU and without hardware blending)
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
Quoted 18 lines · Timothy Miller
I'm posting from home, so this won't look right. Sorry.
Anyhow, Andre Eisenbach said this:
Quoted 1 line
Quoted 1 line
Quoted 1 line
If the graphics card mostly supports 2D initially, it's really not much better then just about any off the shelf graphics card with VESA drivers. As in, the hardware doesn't need to be open for just that. Most (all?) the frustration in Linux graphics card land comes from unsupported/closed 3D drivers.
<<<I have tried using cards with VESA drivers before, and I found it to be very painful. Certainly, you can turn off certain features and get a reasonably useful UI experience, but dragging windows around with "show window contents while moving" enabled is painfully slow, even with AGP 4x. Just imagine doing it over PCI.
Quoted 7 lines
When it comes to desktop applications, the FIRST thing you need is good 2D acceleration. In fact, that's really the ONLY thing. OpenOffice does not need to use OpenGL. GNOME doesn't need to use OpenGL. In fact, for the most part, they don't bother. There are some instances where they use OpenGL, but most of what a workstation user does fits squarely within all the functionality supplied by Xlib, which is entirely 2D.
[snip]
My opinion, for what its worth:
Do 3D first and only. 2D is a subset of 3D. Implement as much of OpenGL as you can in hardware and software can emulate any 2D interface desired.
I agree that existing graphics cards do 2D just fine. I can get a ATI card for $20 that does all the 2D I need. But 2D isn't enough for me. I spend $400 on one Nvidia card. Maybe I'm not the average, common user, but users like me have the highest profit margin. :-)
I'm a pragmatic user. I'd like full-featured Open Source drivers for my Nvidia card but I use the binary because they work really well and for me, (excellent_performance - closed_drivers) > (crappy_performance + open_drivers).
If it can be done well enough to run Doom 3 in 640x480 at 20 fps for less than $500, I'll buy one. That's the performance level where I'd consider sacrificing 60 fps for the open drivers.
Of course, in 5 years I'll expect 120 fps so its definitely a moving target.
Signature · 1 line
Zan Lynx <[email protected]>
Quoted 28 lines · Zan Lynx
Quoted 18 lines · Timothy Miller
I'm posting from home, so this won't look right. Sorry.
Anyhow, Andre Eisenbach said this:
If the graphics card mostly supports 2D initially, it's really not much better then just about any off the shelf graphics card with VESA drivers. As in, the hardware doesn't need to be open for just that. Most (all?) the frustration in Linux graphics card land comes from unsupported/closed 3D drivers.
<<<I have tried using cards with VESA drivers before, and I found it to be very painful. Certainly, you can turn off certain features and get a reasonably useful UI experience, but dragging windows around with "show window contents while moving" enabled is painfully slow, even with AGP 4x. Just imagine doing it over PCI.
Quoted 7 lines
When it comes to desktop applications, the FIRST thing you need is good 2D acceleration. In fact, that's really the ONLY thing. OpenOffice does not need to use OpenGL. GNOME doesn't need to use OpenGL. In fact, for the most part, they don't bother. There are some instances where they use OpenGL, but most of what a workstation user does fits squarely within all the functionality supplied by Xlib, which is entirely 2D.
Quoted 7 lines
[snip]
My opinion, for what its worth:
Do 3D first and only. 2D is a subset of 3D. Implement as much of OpenGL as you can in hardware and software can emulate any 2D interface desired.
If that's what's important, then fine. Just keep in mind the resulting performance hit for many 2D operations.
A 2D engine is simple, and it's very parallelizable. For instance, the logic to process 8 pixels at a time isn't much more than what's required to process 1 pixel at a time. Since everything has a fixed orientation, you can make lots of simplifying assumptions. With 3D, you can't parallelize things like texture-mapping in the same way, because although the destination pixels are fixed in orientation, the source pixels are not. So while you can have one 2D pipeline that processes 8 pixels, the 3D equivalent would be 8 separate 3D pipelines. In other words, 2D scales better than 3D.
Quoted 5 lines
I agree that existing graphics cards do 2D just fine. I can get a ATI card for $20 that does all the 2D I need. But 2D isn't enough for me. I spend $400 on one Nvidia card. Maybe I'm not the average, common user, but users like me have the highest profit margin. :-)
I don't think we can produce a $20 card, at least not as a first-run.
To get our volumes up, we would TRY to sell this card into the Windows market. Yeah, I know... PAIN. But even so, Windows is a different world where ideals like open source don't matter, which means the benefits of this product mostly disappear.
As I understand it, one of the reasons nVidia doesn't release specs is because they don't own all of their IP, so they're contractually disallowed from releasing certain information. And that's just fine. The problem is that if you want to save money and buy a faster, cheaper nVidia card, you're stuck with closed-source drivers. If you run into a kernel bug, LKML members are going to be reluctant to help you with your tainted kernel.
I don't know anything about ATI's position, but since I don't want to deal with closed-source drivers (and the fact that ATI's drivers don't play nice with fbconsole), I'm stuck with a Radeon 9000. (I do believe the 9200 also works with the open source drivers.) I'm also stuck with 3D acceleration which is not nearly what I'd get from ATI-supplied drivers.
I REALLY LIKE ATI and nVidia cards.... for WINDOWS. For Linux, I think I want something else.
Hmmm... so it seems that if you want open source drivers, you're going to have to live with slow 3D no matter WHAT you do. :)
Quoted 4 lines
I'm a pragmatic user. I'd like full-featured Open Source drivers for my Nvidia card but I use the binary because they work really well and for me, (
excellent_performance-closed_drivers) > (crappy_performance+open_drivers).
I haven't upgraded my kernel on this RH7.2 box here in AGES, because every time I do, I have to try to remember how to rebuild the nVidia driver. Eventually, I just gave up on it. Of course, that leaves me without all sorts of enhancements and bug fixes. Some time soon, I'm going to switch over to Gentoo, but to do that, I've procured a Radeon 9200 so I don't have to worry about closed-source drivers.
Quoted 3 lines
If it can be done well enough to run Doom 3 in 640x480 at 20 fps for less than $500, I'll buy one. That's the performance level where I'd consider sacrificing 60 fps for the open drivers.
Well, that might be doable at $500. I don't know. I haven't started thinking about some of the more high-end stuff like programmable virtex shading, bump mapping, or applying multiple textures in a single pass.
Quoted 2 lines
Of course, in 5 years I'll expect 120 fps so its definitely a moving target.
Heh.
Oh, hey... another advantage of doing the open approach is that if we decide to completely redesign the register set, it isn't a problem, because there won't be any guessing about what's changed and what hasn't.