Benchmarking Valve's RADV+ACO Yields Fastest Open-Source Radeon Vulkan Driver

[aufkrawall]

Do you mean the former Unity guy mentioned over here a not so long ago? AMD's compiler team does consist of more than one person though. They will hire a new one, life goes on.

Sure, but the first amdvlk-open release was at the end of 2017, and now we are in mid 2019 and its shader compile time is still utterly slow even vs. mainline-llvm.
Furthermore, vsync is still broken, FreeSync missing, still more visual artifacts in DXVK than radv, likely still slower CPU performance.. When will it reach a mature state?

[ernstp]

There are some interesting differences in minimum frame rates, but no clear pattern I could see in terms of which was better.

Some nice gains in Tomb Raider! But I guess that's just a bonus for now and the main focus is shader compile times...?

[Etherman]

It seems that you, unfortunately, forgot to compare the performance of ACO with Windows 10. But anyway, thanks for your work!
I'm interested in comparing with Windows 10 because today W10 is the benchmark for performance. But there is another reason and it is that I saw a few screenshots where the use of ACO led to the fact that the performance on GNU/Linux with Wine+DXVK in the game Nier:Automata turned out more than on Windows 10. But these are just screenshots, but I would like to see the full quality performance comparison, such as you, Michael, always do. Thank you very much!

38fps - GNU/Linux, Wine+DXVK, Mesa RADV+LLVM: https://i.imgur.com/2RTDKr7.png


55.7fps - GNU/Linux, Wine+DXVK, Mesa RADV+ACO: https://i.imgur.com/ZTeN1wN.png


49fps - Windows 10 (native): https://i.imgur.com/7Yw3fcG.png

Hate to tell you. There's a burning stick.

[dpanter]

Unfortunately, these screenshots were not made by me, but I only found it on the Web as it are, but I don’t know anything about hardware specifications, graphics settings and everything else. Therefore, I ask Michael to make a full comparison of the performance of ACO vs. Windows 10.

I thought the images looked familiar.
Source is a post on GamingOnLinux by YoRHa-2B made in this article comments.

https://imgur.com/a/DBpo43T

YoRHa-2B's listed PC specs on GoL: Ryzen 7 2700X, 16GB RAM, RX480, Arch, KDE Plasma.
It's very likely that this machine was the one used in the tests. No idea what the game settings were beyond the noted 1440p rendering resolution.

[ZFKerr]

I thought the images looked familiar.
Source is a post on GamingOnLinux by YoRHa-2B made in this article comments.

https://imgur.com/a/DBpo43T

YoRHa-2B's listed PC specs on GoL: Ryzen 7 2700X, 16GB RAM, RX480, Arch, KDE Plasma.
It's very likely that this machine was the one used in the tests. No idea what the game settings were beyond the noted 1440p rendering resolution.

If I'm not mistaken, he is Philip Rebohle, the author of DXVK. Thank you!

[ms178]

Sure, but the first amdvlk-open release was at the end of 2017, and now we are in mid 2019 and its shader compile time is still utterly slow even vs. mainline-llvm.
Furthermore, vsync is still broken, FreeSync missing, still more visual artifacts in DXVK than radv, likely still slower CPU performance.. When will it reach a mature state?

Aye, the state of amdvlk seems to be less than desireable (I haven't used it yet myself). It seems that its development is not a top priority for them and from a strategic perspective it might be better to focus development on RADV for Vulkan and RadeonSI for OpenGL and help to improve ACO further? I remember the heated discussions from years back around this topic and wasn't fond of the idea of two coexisting AMD Vulkan drivers myself. But I see a dilemma here, they need to support their own Vulkan driver for other platforms anyway, so giving up on the Linux side of amdvlk wouldn't even save them much, I guess. Maybe ACO could be also used within amdvlk? But I am not familiar with any of the internals, but at least it would mean some ressource sharing between more drivers. But ACO would need to be very flexible and to be cross-platform, I don't know if that is on the Agenda here.

[pal666]

AMD should make themselves honest that the open amdvlk strategy didn't work out.

both radeonsi and radv benefited from open amdvlk strategy. you should educate yourself before posting fantasies

[pal666]

BNieuwenhuizen samuel Does ACO generate different binaries than current LLVM solution for the same shader?

well, identical binaries will have identical speed

[pal666]

You accidentally prove yourself wrong: When amdvlk isn't faster, there is no need for a notoriously immature driver outside of mesa.

for you. so you are not making driver outside of mesa. now move on

[pal666]

today W10 is the benchmark for performance.
...
Wine+DXVK

wine + dxvk vs w10 is benchmark of two different apps. you need linux vulkan vs w10 vulkan. i don't know such games, maybe some demo