RadeonSI OpenGL vs. RADV Vulkan Performance For Mad Max

[duby229]

GamingOnLinux is seeing up to 3x performance improvements with Nvidia hardware. It seems to me RADV is pretty much broken at this point and should be avoided unfortuantely.

That's nonsense. It seems that in GPU bound scenario's, like lower end hardware, radv is performing pretty darn well, even outperforming radeonsi. So it seems more like radv needs more optimizations yet to pull the most out of higher end hardware.

[theriddick]

IC non of you have seen the 980ti results, ok well just wait for phoronix to post his. You can yell at NVIDIA all day if you like but their drivers still kill AMD open or closed in almost all scenarios. Saying that's not true is clearly a sign of major denial... I won't be using RADV under Linux any time soon, maybe in a year or two when it works like the closed AMD vulkan driver (which is dramatically faster btw, unless again your in denial mode!)

The truth hurts...

[bridgman]

Are you allowed to say what the hold up is?

No holdup, it's just a big pile of work and "rewriting to eliminate things you can't expose publicly" is not something that can be done in a public tree.

[dungeon]

IC non of you have seen the 980ti results, ok well just wait for phoronix to post his.

Michael does not have AMD GPU hardware to compete with 980Ti or newer/faster from nVidia, his fastest is Fury non X which compete to 980 non Ti

You can yell at NVIDIA all day if you like but their drivers still kill AMD open or closed in almost all scenarios.

And which scenarios are these, usual CPU boundware scenes and draw calls mostly

I hoped to see Ryzen with Fury on AMDGPU-PRO Vulkan, but suddenly Michael announced at same day possible mobo R.I.P: But this is beta anyway newer drivers are expected.

[humbug]

I won't be using RADV under Linux any time soon, maybe in a year or two when it works like the closed AMD vulkan driver (which is dramatically faster btw, unless again your in denial mode!)

Oh ya AMD's official proprietary driver for Vulkan is way faster than RadV. I wish I could use it. The problem I have is that I can't install it on a modern distro like Ubuntu 16.10 or Fedora 25. So then the only option is the community created RadV but that is too slow...

So overall that's kinda disappointing since it's now been over a year since Vulkan was released, the lack of comprehensive support for Vulkan on AMD graphics cards on Linux is a bit sad.

[smitty3268]

We already have a vulkan team and a vulkan driver that supports multiple OSes. If we wanted to support radv, we'd need to hire a bunch of additional developers to support it. That doesn't make much sense.

This reminds me of those "tree in a forest" stories.

If a driver has no users, does it actually exist?

[dungeon]

Oh ya AMD's official proprietary driver for Vulkan is way faster than RadV. I wish I could use it. The problem I have is that I can't install it on a modern distro like Ubuntu 16.10 or Fedora 25. So then the only option is the community created RadV but that is too slow...

So overall that's kinda disappointing since it's now been over a year since Vulkan was released, the lack of comprehensive support for Vulkan on AMD graphics cards on Linux is a bit sad.

It is not an year, but somewhat like 3 and half in mine book, since Mantle abstraction layer means no DiCe for an user





Shit must be green blob to be useful, since people likes clicking on the blobs Since Earth is green, blue the most and red is inside but you don't see it on average as hard to materialise

[agd5f]

You have a team working on your Vulkan driver, but it's in effect unusable. RADV is usable, and you're not currently paying for its support. That makes even less sense.

I admit having it open sourced would be easier, but it's certainly not unusable. It performs well and supports a wide range of asics.

@Marek: I don't understand - RADV is currently shipping with Mesa, so it's part of it, right? Or are you talking about AMDGPU-PRO's Vulkan driver? As for Mesa accepting several drivers, it may be so but more often than not only one is built by default. The one that isn't sooner or later ends up like a slice of pizza forgotten at the back of the fridge.

There are 3 gallium software rasterizers in mesa all of which are actively developed.

[agd5f]

This reminds me of those "tree in a forest" stories.

If a driver has no users, does it actually exist?

In addition to Linux users, it's used by quite a few users on other OSes.

[geearf]

It won't be part of Mesa.

Is it because it's standalone and doesn't share any code / infrastructure with mesa?
If so, in your dev mind, is that more of a pro or a con?