NVIDIA's Proposal For A New API Better Than GBM Has Already Made Some Progress

[Delgarde]

Unbelievable. They prefer start a new API instead of supporting GMB. The whole point of people not wanting to use EGLStreams was to not throw away all the work already done. Now Nvidia want people to do even more work by implementing a new API, instead of fixing/implementing what GBM lacks.

Think about it - why would NVidia take this path, flatly rejecting the community choice of GBM, but being open to coming developing a shared new API? It's obvious that GBM just wasn't an option for them - that they simply couldn't implement it for their driver. And that's not unlikely, because while GBM is a workable solution for any driver based on the Mesa/DRM stack, it wasn't designed to be shared with other stacks, like whatever NV use internally.

It's looking pretty clear that collaborating on a new shared API for this is the only way we're going to see progress on this. NV won't accept GBM, probably because they can't implement it. The open guys won't accept EGLStreams, for what I assume to be valid reasons. And since pretending NV don't exist probably isn't a realistic option, that new API is probably the only way forward.

[Delgarde]

The GBM people were always prepared to improve it. Nobody claimed GBM was complete, sufficient, or a final API/ABI.

I don't think it's a question of improvement - I suspect that the unstated reason for all this fuss is that the GBM design is in some way fundamentally incompatible with the NVidia driver architecture... it was designed by the Mesa/DRM guys, so it likely has a lot of assumptions that work for them, but aren't valid for NVidia.

If so, no amount of improvement will suffice - the only option is the one NVidia seem to be taking, pushing for collaboration on a new API that works for both parties.


For the life of me, I don't know why NVidia don't just come out and confirm this, though. It's the only rational explanation for the path they're taking, but for all that they're wanting to be open and collaborative on solving this impasse, they're remarkably silent about the motivation.

[boxie]

It is good to see that the impasse has a possible path to resolution that people are willing to work on.

From the presentation there were a bunch of desirable features that were not implemented in any of the solutions, so I hope that this new idea can solve the problems.

Though, it does seem like https://xkcd.com/927/ might be getting a work out

[SaucyJack]

I don't think it's a question of improvement - I suspect that the unstated reason for all this fuss is that the GBM design is in some way fundamentally incompatible with the NVidia driver architecture... it was designed by the Mesa/DRM guys, so it likely has a lot of assumptions that work for them, but aren't valid for NVidia.

If so, no amount of improvement will suffice - the only option is the one NVidia seem to be taking, pushing for collaboration on a new API that works for both parties.


For the life of me, I don't know why NVidia don't just come out and confirm this, though. It's the only rational explanation for the path they're taking, but for all that they're wanting to be open and collaborative on solving this impasse, they're remarkably silent about the motivation.

If AMD handled gbm fine then nvidia has no excuse.

[SaucyJack]

Considering that NVIDIA makes the best GPUs on the planet (they are the Intel to ATi's AMD) it makes sense to listen to what NVIDIA says. Love them or hate them they have more market share, more money, more experience and drivers which run on more platforms than AMDs.

I actually felt a little sick after reading that much bullshit.

[smitty3268]

Well, that also kind of proves there are indeed things to improve wrt GBM and Nvidia wasn't just making up excuses. At least not 100%.

That's long been known and agreed. Intel has had people working on extending GBM to deal with those problems for a while now, and they were the ones that kept asking NVidia why they had to use EGLStreams rather than just the extensions and fixes they were making for GBM.

[robclark]

why does he call gbm an Unix Device Memory API ? gbm already has a name

The scope is (probably) bigger, covering what formats can be shared between multiple devices, incl multiple gpu's but also various embedded/android type use-cases such as separate gpu+display (or camera/vidc/etc), and who can allocate, etc. Basically what gralloc does but done properly.

[computerquip]

If AMD handled gbm fine then nvidia has no excuse.

That's kind of a silly thing to say given any two architectures from both companies are different.

[microcode]

Seems like they could just start by implementing GBM, then work with the community on defining the next generation. Instead we'll have a situation where their driver will work with no existing software for the next year or two.

[misp]

Nvidia: So, let's change the world because the world is wrong. We're just a bunch of standup guys and you are all dumb.


I don't like this approach to anything. There are guys who can make it work with GBM for free, but nvidia doesn't want to help them as well. Who knows what the fuck they have in those blobs when they are fighting so hard to keep it in-house.

Hopefully they won't screw amd's efforts who played along with the community. But, you know, nice guys usually get screwed. And it would be nice if the community backed up somebody that played nice with them (Some political messages are needed here).

People are moving to Linux because of security issue and I don't mind saying nv to FO. If anyone is concerned about privacy and security, they can change vendors and get something that works on Linux. Maybe it increases amd's market share and brings better competition.

In the end: they have the money to fund development of differently programmed drivers, why should community rewrite everything just to accommodate those assholes?