NVIDIA RTX Remix 0.4 Released With Updated DXVK, Performance Improvements & Fixes

[MrCooper]

If there was a performance hit with implicit sync, either radeonsi & RADV couldn't compete with the nvidia driver now, or they'd be expected to consistently beat it once they use explicit sync.

[user1]

The latter doesn't require explicit sync in the Wayland protocol, see Ensuring steady frame rates with GPU-intensive clients (which wasn't covered by Phoronix for some reason) on how I implemented this in mutter!1880 which landed in mutter 44.

Interesting. When I used Gnome 44, I did notice that my cursor didn't stutter anymore while I played 30> fps YouTube videos. But I thought this was a different issue.
Will this work still provide benefit even when Mutter and the rest of the graphics stack will fully support explicit sync?

[mdedetrich]

If there was a performance hit with implicit sync, either radeonsi & RADV couldn't compete with the nvidia driver now, or they'd be expected to consistently beat it once they use explicit sync.

This was already explained numerous times before, the radeonsi/RADV drivers were built with implicit sync in mind and they never removed that capability which is why they are not effeted. Where as the NVidia driver was rebuilt from scratch 15 years ago completely based on explicit sync.

The NVidia developers said this on mesa gitlab

[smitty3268]

Dude get the god damn hint

There is no f**ken way that NVidia putting in hacks to their driver just to pass some checkbox would have "helped" the Wayland transition (and helped here is in quotation marks because its not really a help, its just making people feel good). Instead of people complaining about x not working in some way, they would instead complain about x not working in another way.

I disagree.

All of the relevant people (and yes that means other GPU manufacturers, not just NVidia) disagree with you

Really? Can you show some references where someone from Intel and AMD publicly state that failing to support implicit sync on linux is the correct thing to do? Because I've seen the opposite, and you had an AMD employee in this very comment section saying so.

and if you actually cared about the Wayland transition rather than crying "Waaah, why isn't Wayland working with NVidia right now via some hacks"

I'm not the one that's getting super upset and crying about it. I'm just stating that Nvidia had options and chose not to do them. I'm allowed to dislike that choice and argue that it's poor for the consumer, no matter how good it was for nvidia's bottom line.

And this is all on the presumption that those hacks which you are asking for NVidia to do would have worked properly

Yes, it is. I do believe nvidia is competent at writing drivers, not incompetent, which means they could have made it work if they wanted to.

[mdedetrich]

I disagree.

Disagree about what? Its NVidia and their driver and in generally its a really stupid idea to put hacks into something as critical/core as a driver, especially if the only motivator is to placate some early adaptors of Wayland given that you can always just use X11/Xorg (no one is forcing you to use Wayland).

Really? Can you show some references where someone from Intel and AMD publicly state that failing to support implicit sync on linux is the correct thing to do? Because I've seen the opposite, and you had an AMD employee in this very comment section saying so.

This was quoted so many times before. I can't be assed to trawl through previous Phoronix discussions, but you can read https://www.collabora.com/news-and-b...-gap-on-linux/ . Also just so you know, the author of that article is the main heavy pusher of explicit sync and he doesn't even work for NVidia, he works at Intel and used to work at AMD (in both cases on the graphics stack).

You can just search the guys name on mesa's gitlab, he is actually the strongest pusher of explicit sync and he has nothing to do with NVidia aside from working with NVidia dev's on the graphics stack.

I'm not the one that's getting super upset and crying about it. I'm just stating that Nvidia had options and chose not to do them. I'm allowed to dislike that choice and argue that it's poor for the consumer, no matter how good it was for nvidia's bottom line.

Yes, it is. I do believe nvidia is competent at writing drivers, not incompetent, which means they could have made it work if they wanted to.

Right, and NVidia said its not possible to competently implement implicit sync in their driver without an excessive amount of effort, why don't just take that for what it is and leave it? I mean it really is that that, if there was such a simple solution to this NVidia would have done it since there is no reason otherwise.

Sometimes these things happen, not every problem has the solution you want

[smitty3268]

Disagree about what?

With what you said in the bit i directly quoted, about how nvidia didn't hurt the wayland transition and that providing implicit sync support would have been just as bad as not. I quoted you for a reason there.

Its NVidia and their driver and in generally its a really stupid idea to put hacks into something as critical/core as a driver, especially if the only motivator is to placate some early adaptors of Wayland given that you can always just use X11/Xorg (no one is forcing you to use Wayland).

I disagree. Graphics drivers are full of hacks, many of them explicitly to support the latest games coming out that are completely broken, and nobody bats an eye about that. It's what drivers do. Do you truly believe that nvidia's driver doesn't have hacks in it already for weird corner cases of OpenGL just so they can "check a box" and pass the conformance tests? Beyond that, monitors famously are almost never to spec and so drivers are full of workarounds to make them work. Same for pretty much every other hardware driver in the kernel that touches anything complex.

This was quoted so many times before. I can't be assed to trawl through previous Phoronix discussions, but you can read https://www.collabora.com/news-and-b...-gap-on-linux/ .

Thanks. I've read that article before, actually, and I just read it again. They do a good job of explaining why explicit sync is the way forward.

However, and this is important: I never said otherwise

Seriously, this really makes me think you aren't even reading what I'm typing, because I saw zero places in that article where he said nvidia shouldn't provide temporary support for implicit sync while the underlying driver infrastructure on linux is overhauled. In fact, a great deal of it was about how to seamlessly update drivers to the new model while keeping the old one working, for a smooth transition. Precisely what nvidia is NOT doing, and what I'm talking about.

If you think I am somehow against moving to explicit sync, you aren't paying attention to what I'm saying. Now, if I somehow missed something in particular in that article where they said nvidia shouldn't provide backwards compatibility please let me know and I'll apologize.

Also just so you know, the author of that article is the main heavy pusher of explicit sync and he doesn't even work for NVidia, he works at Intel and used to work at AMD (in both cases on the graphics stack).

You can just search the guys name on mesa's gitlab, he is actually the strongest pusher of explicit sync and he has nothing to do with NVidia aside from working with NVidia dev's on the graphics stack.

I'm well aware of who Faith is, and their history within the Mesa project. They've given several great xorg dev conference presentations i've watched online over the years. Since you respect their opinion so much, check out the comment section on that article, where they say:

Faith Ekstrand:
Mar 27, 2023 at 07:37 PM
That's a very long-term project. We're multiple years out from having a design that's ready to land upstream. Also, it's not really fair to say that that's the blocker. NVIDIA could implement proper implicit sync in their driver stack if they chose to. They've chosen not to.

Pretty much the same thing I just said here.

Right, and NVidia said its not possible to competently implement implicit sync in their driver without an excessive amount of effort, why don't just take that for what it is and leave it? I mean it really is that that, if there was such a simple solution to this NVidia would have done it since there is no reason otherwise.

Have you never had any experience working for a corporation that does ROI to determine whether they do something or not? That's corp-speak for "We could do this but it isn't worth the effort so we aren't going to bother".

I have no doubt it would take more effort than Nvidia wants to expend. Because they want to expend 0 effort. Anything non-zero is by definition "excessive". If you can give me a number of man-hours estimate, I'd take that more seriously, but obviously nvidia won't release that kind of info publicly.

You can choose to believe that their definition of excessive lines up with whatever you want to believe, but personally I think it's below my definition. At least for a quick, hacky version (that would have some performance implications that they don't like). A fully optimized version probably would be excessive, I'll grant that much, and I suspect that's what they are really getting at in their response. Note how they said "competent" - I think that probably means virtually zero performance loss, which means deep integration and likely excessive work.

[MrCooper]

Interesting. When I used Gnome 44, I did notice that my cursor didn't stutter anymore while I played 30> fps YouTube videos. But I thought this was a different issue.

It can certainly help for that with mutter 44. To be fair though, the KMS thread which my colleague Jonas Ådahl landed in mutter 45 allows for even smoother cursor movement, and this work wasn't required for that.

Will this work still provide benefit even when Mutter and the rest of the graphics stack will fully support explicit sync?

Yes, in fact, the transaction mechanism is required for supporting the WIP explicit sync v2 Wayland protocol both correctly and with good performance at the same time. Explicit or implicit sync in the protocol are just slightly different ways to determine when a transaction is ready to be used for an output frame, that's kind of a detail compared to the transaction mechanism itself though.

One could actually say that this is a kind of explicit sync internally in mutter, just that it also works with implicit sync in the Wayland protocol.

I can't be assed to trawl through previous Phoronix discussions, but you can read https://www.collabora.com/news-and-b...-gap-on-linux/ . Also just so you know, the author of that article is the main heavy pusher of explicit sync and he doesn't even work for NVidia, he works at Intel and used to work at AMD (in both cases on the graphics stack).

Faith is currently with Collabora (where the blog was posted, but it seems you couldn't be bothered to actually read it anyway), AFAIK she never worked for AMD though (I did 2011-2019).

[mdedetrich]

With what you said in the bit i directly quoted, about how nvidia didn't hurt the wayland transition and that providing implicit sync support would have been just as bad as not. I quoted you for a reason there.

They delayed a worse version of Wayland in order to make a more correct/better version of Wayland, take that for what you will. For technical reasons its actually a good thing that this was delayed, I have already explained why.

I disagree. Graphics drivers are full of hacks, many of them explicitly to support the latest games coming out that are completely broken, and nobody bats an eye about that. It's what drivers do. Do you truly believe that nvidia's driver doesn't have hacks in it already for weird corner cases of OpenGL just so they can "check a box" and pass the conformance tests? Beyond that, monitors famously are almost never to spec and so drivers are full of workarounds to make them work. Same for pretty much every other hardware driver in the kernel that touches anything complex.

Except in this case you would have completely tanked performance and then people like you would be crying about how shitty NVidia is. Thats what the result of NVidia's hack would be, it would involve stupid things like having to manually check the graphics buffer pipeline to insert fences in order for implict sync to work, tanking performance by 30% or more.

Seriously, this really makes me think you aren't even reading what I'm typing, because I saw zero places in that article where he said nvidia shouldn't provide temporary support for implicit sync while the underlying driver infrastructure on linux is overhauled. In fact, a great deal of it was about how to seamlessly update drivers to the new model while keeping the old one working, for a smooth transition. Precisely what nvidia is NOT doing, and what I'm talking about.

You are asking to provide evidence of something that is not possible because the only people that can provide evidence of what NVidia is saying is NVidia themselves because its a closed source driver. No one else can back that up because in order to do so you would have to see the source code.

So you either take what NVidia says at good faith or just drop the argument

I have no doubt it would take more effort than Nvidia wants to expend. Because they want to expend 0 effort. Anything non-zero is by definition "excessive". If you can give me a number of man-hours estimate, I'd take that more seriously, but obviously nvidia won't release that kind of info publicly.

And this is where you are completely wrong because NVidia currently helping with explicit sync is not taking 0 effort. You seem to be under some delusion that NVidia took the easy way out that has zero effort, thats completely wrong. NVidia dev's are constantly participating on the merge request discussions and providing feedback about what makes sense and what doesn't and they are also testing changes with an internal version of their own driver to actually verify that the explicit sync changes are working correctly. You can see this yourself from the discussions on mesa's gitlab.

Im sorry to break it to you thats not zero effort. What NVidia definitely doesn't want to do is to waste their effort on a giant substandard hack just to people please early Wayland adaptors. You still seem to have issues accepting that.

[smitty3268]

...

We're starting to just repeat ourselves at this point, and since it's clear neither of us is going to change our opinions I'm going to let this die here.