NVIDIA Maxwell Support Patches For The Nouveau X.Org Driver

[bridgman]

According to this, by 2008 you weren't: http://www.xbitlabs.com/news/graphic...vidia_JPR.html
But of course, next you're going to tell me you were doing open source drivers at ATI since you were a kid.

Sure, but the decision was made in early 2007, not 2008.

I was no longer a kid by the time ATI was founded.

[bug77]

Sure, but the decision was made in early 2007, not 2008.

I was no longer a kid by the time ATI was founded.

Right. And 9 years later you have a driver that works on a couple generations of cards. If you don't need OpenCL. Or Vulkan. Amazing job you did right there.
By comparison, since 2007 we have had the rise of the iPhone and Qualcomm/Snapdragon (which ATI helped build). And Android. And in all this time, you almost wrote a driver. Kudos.

Btw, according to that graph, you lost your leadership sometime in 2006. So you clearly didn't make the decision to switch when thing were going well. That's all I'm saying.

[bridgman]

Right. And 9 years later you have a driver that works on a couple generations of cards.

r300/400, r5xx, r6xx, r7xx/Evergreen, NI, SI, CI, VI/Polaris, and some newer ones (each generation bringing significantly different HW)

In my country a couple is almost always less than 8. Where do you live and how many are in a couple there ?

Btw, according to that graph, you lost your leadership sometime in 2006. So you clearly didn't make the decision to switch when thing were going well. That's all I'm saying.

Your theory seems to assume that we all had time machines to skip ahead over the ~8 month lag between the start of a quarter and the time consolidated industry reports become available for the period up to the end of the quarter. I'm not saying it's impossible but I submit to you that if we had had a time machine we would have been much more profitable.

The decision to restart open source driver development was initially driven and funded from the CPU side of the business, based on requirements and feedback from their existing customer base.

In a business where 99% of revenues came from Windows sales the first reaction to slipping Windows GPU market share would not have been "hey let's take on more cost & risk and pull a bunch of our top technical people off Windows to do open source graphics drivers for Linux".

I'm not saying that your theory is dumb, or that it does not reflect a commonly seen pattern, just that it does not apply in this case. Besides, the usual pattern is to throw existing code over the wall, RIF all the people who worked on it, and hope someone outside does something with the code - not write new code and keep hiring more new people to work on it. Or are you saying we did that wrong too ?

[bug77]

See, this is where we disagree. You did put out 8 drivers, but you haven't actually finished any of them: https://www.x.org/wiki/RadeonFeature/
As for "each generation bringing significantly different HW", I don't follow hardware developments that closely, but Wikipedia seems to suggest only minor additions with each version.

It's a shame, too. If only you'd have offered AMDGPU PRO in a timely manner, I would have given AMD another chance this year. Then again, even with this new driver, we get the same sucky performance of fglrx

[bridgman]

See, this is where we disagree. You did put out 8 drivers, but you haven't actually finished any of them: https://www.x.org/wiki/RadeonFeature/

Huh ? I didn't say 8 drivers, I said driver support for 8 generations.

Maybe we are looking at things differently, but that page looks pretty green to me. The main exceptions are OpenCL, where we recently obtained approval to open source the current closed driver rather than continuing to develop a separate open driver, and some of the workstation functionality which needs to stay on the closed source GL driver because Mesa does not support compatibility profiles.

The actual driver components change when the hardware changes are too big to be able to cover with runtime logic in a single driver, eg there were three different 3D HWL drivers (r300, r600, radeonsi) required to cover those generations.

[bug77]

We definitely are looking at things differently, otherwise we would not be having this conversation.
There are other things missing from that matrix. Go back two years on that page and you'll see a lot more was missing. So the fact remains that until very recently, the open source drivers have been offering only basic support. To you, these are all non-issues, to me they're missing functionality (and that matrix speaks nothing about the level of performance offered). You say "some of the workstation functionality which needs to stay on the closed source", but there's stuff like Crossfire and a few others that are not workstation only.
I understand you defending your brain-child, but at the same time, let me remind you that a program that almost works is just like a plane that almost flies. That is the only point I'm trying to make. I'm not going to switch because you hit some internal milestones. I'm going to switch when you offer me the whole package. Ok, I would have disregarded Crossfire, because I don't go there anyway, but support for older cards is very important.
You may pat yourself on the back for reaching this point after 9 years, but I've been enjoying full support from Nvidia during all this time. You have more work to do if you want to win users over. And I know, you don't want that, it's all about workstations...

[bridgman]

We definitely are looking at things differently, otherwise we would not be having this conversation. There are other things missing from that matrix. Go back two years on that page and you'll see a lot more was missing. So the fact remains that until very recently, the open source drivers have been offering only basic support.

Sure, but we were talking about today not two years ago. If you had said "two years ago" my response would have been different - I would not have agreed with "only basic support" but I would have agreed with the gist of what you were saying.

To you, these are all non-issues, to me they're missing functionality

What specific missing functionality are you talking about. OpenCL in the all-open stack ?

How could you possibly think I regard missing functionality as a non-issue ? I don't think I have ever said anything that could be twisted around to that conclusion unless you start rearranging my words into new sentences.

(and that matrix speaks nothing about the level of performance offered).

Then please don't use the matrix as your only reference. I don't necessarily disagree with what you are thinking, just what you are typing.

You say "some of the workstation functionality which needs to stay on the closed source", but there's stuff like Crossfire and a few others that are not workstation only.

Crossfire/SLI has never really worked for Linux gaming other than the odd special case, for the simple reason that game developers don't have the time/resources to implement support. It has been years since something like Crossfire/SLI could "just work" without game support.

What that leaves is workstation, which generally uses older GL versions and (more importantly) no difficult-to-split effects like motion blur... but GPUs are sufficiently fast these days that workstation doesn't care much about it either.

I understand you defending your brain-child, but at the same time, let me remind you that a program that almost works is just like a plane that almost flies. That is the only point I'm trying to make. I'm not going to switch because you hit some internal milestones. I'm going to switch when you offer me the whole package. Ok, I would have disregarded Crossfire, because I don't go there anyway, but support for older cards is very important.

I'm having a tough time understanding what you mean by "the whole package". I suspect you are thinking about OpenGL performance, but you aren't actually saying it so it's tough for me to respond the way you expect.

You may pat yourself on the back for reaching this point after 9 years, but I've been enjoying full support from Nvidia during all this time. You have more work to do if you want to win users over. And I know, you don't want that, it's all about workstations...

Again, please be specific. I don't necessarily disagree with you but it's tough when I have to guess what you are thinking.

If you are saying "it took a long time to build an open source driver stack that is essentially caught up with the closed-source driver" then yeah, that's true, but I don't think anyone (other than random posters on the internet) ever expected anything different. People are running the closed-source OpenCL driver with the all-open stack today, although it requires some tweaking for now - what else do you feel is missing ?

[Kano]

The main thing I see is that if you stay with a distribution like Debian stable for a release cycle then you would need officially supported backports including LLVM. Without LLVM compiling mesa is simple, but mesa users with AMD can mainly only use rolling releases or maybe Ubuntu with 2 releases every year. Usually updating binary parts without hard depends is easier. But certainly the out of the box experience is much better now with AMD GFX now compared to NVIDIA without binary drivers. It is certainly still a question if you should go for full open stack or use for games optimized binary drivers. AMDGPU-PRO shows however no huge improvements overy current mesa code. I think AMD could improve more in that area but most likely you have to wait for game devs to use Vulkan for higher speed.

[bug77]

bridgman I fell we've drifted way off the subject. My initial assertion was that companies don't simply embrace open source because open source is pretty. They have to see some gain in order to make this change. My theory is ATI/AMD has only switched when they lost market share to Nvidia. You say the effort (which was made public some time after Nvidia became the #1 player) was started before that happened, but you can't prove that, because you're not going to start publishing internal documents here just to prove me wrong.
The corollary to the above would be that people shouldn't simply expect Nvidia to start rewriting a perfectly good driver because a bunch of people are complaining on Phoronix. Whether they'll be forced to go open source because of Wayland or some other technology we don't know about yet, whether they'll loose market share and won't be able to maintain a closed driver or whether they'll never have an open source driver, we simply don't know. But what we know is today they're not motivated to change.

And btw, I do think ATI/AMD saw moving to open source as a cost cutting solution as hordes of developers would take over once they released the documentation. When that didn't pan out, they had to shift programmers (I was about to say "resources", but God I hate it when programmers are referred to as such) around and that's why it took them so long to offer a decent driver.

[bridgman]

I fell we've drifted way off the subject. My initial assertion was that companies don't simply embrace open source because open source is pretty. They have to see some gain in order to make this change.

My theory is ATI/AMD has only switched when they lost market share to Nvidia. You say the effort (which was made public some time after Nvidia became the #1 player) was started before that happened, but you can't prove that, because you're not going to start publishing internal documents here just to prove me wrong.

Correct. What I can tell you is that I was part of the team that did the initial planning, and at no time did that come up as a reason.

There were some outside complaints about not having proper driver support for r5xx and r600 (so those outside voices were thinking the open source driver might fill a gap if there was never going to be 5xx/6xx Linux Catalyst support), but the reality was that we were in the final stages of a major architectural change to Linux Catalyst that shared a lot more code with Windows and moved to an all-new OpenGL driver.

The new Linux Catalyst driver was released around the same time as the first open source driver code; a few days ahead IIRC.

Remember that the idea of rebuilding the proprietary driver on top of open source code came much later... at the start the open source driver was a second code base independent of the (ongoing) Linux Catalyst driver. We started talking about it casually in maybe 2008 or 2009 (when the first customer asked for KMS support in Linux Catalyst, probably), but it didn't become a real plan until much later.

The corollary to the above would be that people shouldn't simply expect Nvidia to start rewriting a perfectly good driver because a bunch of people are complaining on Phoronix. Whether they'll be forced to go open source because of Wayland or some other technology we don't know about yet, whether they'll loose market share and won't be able to maintain a closed driver or whether they'll never have an open source driver, we simply don't know. But what we know is today they're not motivated to change.

I agree with you completely there. People still think that we re-started open source driver support because of petitions or angry mobs, but the reality was that a significant percentage of AMD's server CPU sales were used with Linux (at least 35%, possibly higher), and some of those were in graphics-related areas like compute farms for movie and special-effects rendering, so when big customers in that area started saying "you really need open source GPU drivers" the sales teams listened because those customers represented a big chunk of our business.

And btw, I do think ATI/AMD saw moving to open source as a cost cutting solution as hordes of developers would take over once they released the documentation. When that didn't pan out, they had to shift programmers (I was about to say "resources", but God I hate it when programmers are referred to as such) around and that's why it took them so long to offer a decent driver.

Nope, you still seem to be working on the assumption that the open source stack was supposed to replace the Linux Catalyst driver. That was never the plan, then or now. The open source driver has definitely helped on the consumer side, but that was never the main (workstation) focus for Linux Catalyst.

There were a lot of internet posts about "10,000 developers writing a replacement for fglrx in a few months" after we announced our plans but that never seemed credible (even if you ratcheted it down by a factor of 100) and was never a factor in our plans. Remember that we were working directly with community developers during the planning effort, so there was none of the false optimism you are describing.

Right from the start the open source driver effort was an additional cost on top of what we were spending already, and the benefits were primarily expected to come in the form of (a) happy customers buying our high end server CPUs (that didn't go so well) and a few additional GPU sales (albeit with no way to track them back to driver decisions). I presented a plan to Dirk, Phil Hester (CTO) and a few others then asked for some budget on top of existing GPU funding; Dirk asked one of the VPs on the CPU side to find us the money and approved the plan.

The funny thing is that the biggest benefit from the open source graphics effort actually turned out to be in the embedded graphics area, which was not an area we had thought about much during the initial planning.