Mesa 20.1-dev RADV vs. RADV+ACO vs. AMDVLK vs. AMDGPU-PRO Vulkan Radeon Linux Gaming Performance

[phoronix]

Phoronix: Mesa 20.1-dev RADV vs. RADV+ACO vs. AMDVLK vs. AMDGPU-PRO Vulkan Radeon Linux Gaming Performance

Here is an up-to-date look at how the very latest Mesa 20.1 Git performance is for the Radeon "RADV" Vulkan driver both out-of-the-box and when enabling the Valve-backed ACO compiler back-end alternative to AMDGPU LLVM. Plus there are benchmarks of the latest AMDVLK open-source AMD Vulkan driver and also when using AMDGPU-PRO's Vulkan packages that still rely upon AMD's proprietary shader compiler.

http://www.phoronix.com/vr.php?view=28951

[oibaf]

Note that mesa in the PPA currently builds against llvm9 rather then llvm10, since the with the latter mesa won't build.

[mphuZ]

Jesus Christ! Great job, AMD!
AMDVLK is now better than AMDGPU-PRO! I think we are closer to the day when AMD will switch to the open Vulkan stack.

It's also worth noting that AMDVLK is quietly competing with RADV, and even the ACO doesn 't help. Working with DXVK is also improving - this is very pleasing!

[cute2dgirl]

Unfortunate timing. There was a Vega-specific bug with tessellation in ACO which would cause hangs, but that got fixed a few hours ago.

[towo2099]

Note that mesa in the PPA currently builds against llvm9 rather then llvm10, since the with the latter mesa won't build.

Disable build of opencl and mesa 201-dev builds against llvm-10.

towo:Defiant> inxi -G
Graphics: Device-1: Advanced Micro Devices [AMD/ATI] Ellesmere [Radeon RX 470/480/570/570X/580/580X/590] driver: amdgpu
v: kernel
Display: x11 server: X.Org 1.20.7 driver: amdgpu,ati unloaded: fbdev,modesetting,vesa resolution: 1920x1080~60Hz
OpenGL: renderer: Radeon RX 580 Series (POLARIS10 DRM 3.36.0 5.6.0-rc6-siduction-amd64 LLVM 10.0.0)
v: 4.6 Mesa 20.1.0-devel

[qarium]

yes amd and the FLOSS community really did the job.

now the biggest problem is the legacy stuff of outdated CAD software using non-standard OpenGL
if we can manage to solve this problem then AMD can just turn of the plugs from the old closed source driver.

this PRO/FireGL customers according to Bridgman do cross-subsidizatizing the linux/FLOSS development.
But if money isn't the problem of this Rich customers why they do not use the money to fix their software?
or just write profiles for mesa to run the Legacy apps.

however i think geting rid of the Legacy closed source driver should result in good PR dor AMD.

[Venemo]

Sorry guys, the Vega problem was my fault. I don't have a Vega myself so didn't have the opportunity to test ACO tessellation there... Now that the problem has been fixed, it would be nice if Michael could repeat the tests with those fixes included.

[microcode]

Sorry guys, the Vega problem was my fault. I don't have a Vega myself so didn't have the opportunity to test ACO tessellation there... Now that the problem has been fixed, it would be nice if Michael could repeat the tests with those fixes included.

I think I have an extra (Vega 64) I could send you when Corona-chan moves on.

[tildearrow]

Typo:

this appears to be a regression in the AMD stack with the AMDGPU-PRO 19.50 perfoming okay and it being an older revision.

[smitty3268]

Hmm,

The -pro drivers really don't have any performance advantage anymore, over either the amdvlk or radv drivers.

The amdvlk performance looks pretty good, but the strange brigade results shows AMD isn't taking it seriously as a gaming driver yet, just a place to dump open sourced code. That behavior is most likely due to a poor implementation of vulkan's mailbox presentation mode - which is integrated with X11, and therefore likely written for scratch in this driver and not copied from something else.

And the Vega 56 issue in ACO is rather unfortunate for this set of tests, though it sounds like it's already been fixed up.