Radeon's AMDVLK Vulkan Driver Picks Up A Warhammer II Optimization

[abott]

AMDVLK is a shit driver. skeevy420 You can almost be assured RADV takes AMDVLK out back and curb stomps it to death. One year ago it was barely even usable. Today, it's fantastic. And faster. With more general support

[duby229]

Yet Google picked it for Stadia over radv for some reason. I'd prefer them to use radv instead.

I'm just learning about stadia. Seems it will only be available to chromebook users, so... Not much use to almost anybody....

But good for AMD, bad for Googles customers.

[humbug]

I'm just learning about stadia. Seems it will only be available to chromebook users, so... Not much use to almost anybody....

Stadia will not be limited to chromebook users. That makes no sense, it's running off the cloud.

[duby229]

Stadia will not be limited to chromebook users. That makes no sense, it's running off the cloud.

In which case that just makes amdvlk a worse choice....

[humbug]

Yet Google picked it for Stadia over radv for some reason. I'd prefer them to use radv instead.

Google revealed what drivers they are gonna use? Do you have a source.

[skeevy420]

AMDVLK is a shit driver. skeevy420 You can almost be assured RADV takes AMDVLK out back and curb stomps it to death. One year ago it was barely even usable. Today, it's fantastic. And faster. With more general support

Did AMDVLK come to your house and pee in your breakfast cereal? Geez.

Last night with F1 2017 I had a better FPS average with RADV (19.0.1), 52/45, and a better frame time averages with AMDVLK (latest git), 16ms/20ms. I used the High preset with SSRT shadows on, 1080p, TAA, 16x AF, and the default benchmark track. RADV went second. I'm wondering if running each benchmark twice would have resulted in them being more equal or showing one as better than the other due to shader cache or whatnot. AMDVLK from AMDGPU-Pro ddn't work due to missing some Vulkan features.

Like I said, with the games I play I get mixed results. Some games are equal and some games prefer one over the other.

[abott]

If you don't see how AMDVLK holla everything back, you're actually retarded.

[aufkrawall]

In my tests is renders faster.

But it loses more performance with dxgi.maxFrameLatency = 1 in DXVK than radv and still has corruption issues in lots of games.
It's also excessively slow in mpv and many bugreports don't gain any traction at all. In most cases, either use radv or proprietary amdvlk. Open amdvlk still hardly makes any sense to use. Oh, and compiling shaders on CPU with it is even slower than radv. Just wtf?

[skeevy420]

AMDVLK is garbage. Test it with Shadow of the Tomb Raider Trial. The rendering is many times slower than with RADV and has grey lines through Laras face. The game runs in a window only.

Don't have that Tomb Raider yet.

With F1 2017, however, after running multiple tests yesterday, RADV clearly came out ahead after multiple benchmarks and enabling some AMD_DEBUG features (sisched,nir,forcedma,unsafemath,checkir). They were both competitive with each other, within 3-5 fps on the minimum and average, high was the same (capped at 60) and within 1-2ms on the frame times.

[skeevy420]

But it loses more performance with dxgi.maxFrameLatency = 1 in DXVK than radv and still has corruption issues in lots of games.
It's also excessively slow in mpv and many bugreports don't gain any traction at all. In most cases, either use radv or proprietary amdvlk. Open amdvlk still hardly makes any sense to use. Oh, and compiling shaders on CPU with it is even slower than radv. Just wtf?

Proprietary AMDVLK doesn't work with F1 2017 due to missing Vulkan features.

RADV crashes the vulkaninfo tool while neither AMD variant does.

All I'm pointing out is that they all have their issues.

Personally I just try all three, run a few benchmarks with each (multiple times in a row now), and go with whichever works the best on the 2nd+ benchmarks...and when the both run at 60fps smooth I'll just go with the tried and true "feels" benchmark.