Mesa 13.0 vs. 17.0 Performance For RADV/RadeonSI: Big Gains For Vulkan, OpenGL Boosts

[ElderSnake]

Deus Ex: MD on my AMD 380X (Mesa 17) will run on High Details no problem except for some areas still give me a 12 FPS kind of performance.

But the weird thing is, hang around in the area long enough and performance starts to really smooth out. Go to a new area, crappy FPS, eventually smooth again.

Could this be related to the lack of disk caching?

[mitch074]

Is that RadeonSI and not AMDGPU? If RadeonSI then that is fantastic!

Padoka PPA around the time 17.0 was forked. That's average fps on a i5 [email protected] GHz, using xfce with compositing disabled. VRAM use is around 6 Gb (it climbs to 8 Gb with Ultra textures, which is the max my card supports). In-game though, walking around Pragues is awful due to strong slowdowns - but it's unrelated to quality settings as even at the lowest settings, frame rate drops something huge.
I would guess that testing anything over Medium or High with these AMD cards slows down to a crawl due to VRAM use: tests done with all settings at Ultra except textures set at Medium would prevent VRAM problems, and really show off Mesa 17's improvements - texture size has little impact on rendering speed.

[cRaZy-bisCuiT]

I have DE :MD and a RX480; I hit 31fps at 1440p on Very high. Windows does 36 with the very same settings.

Are you actually talking about AMDGPU + Mesa or is it AMDGPU-PRO? If that is Mesa it would be awesome! We could even outperform windows then one day if the port and drivers are better optimised under Linux.


About the VRAM: As long as your don't run into into full VRAM there's almost no performance difference between the quality settings of textures.

[mitch074]

I used Oibaf's too, but Oibaf is still on LLVM 3.9 (or was, when I last tried it) and performance under DE:MD was abysmal. I also don't update my packages TOO often, as there are still a few breakdowns sometimes.
I switched to kernel 4.9 recently, but impact on performance was imperceptible.

[darkbasic]

actually it has parameter "rendering resolution" and just scales results to your monitor. but you can run x server in any resolution you like and trick talos

It's not practical for automated testing, that's why Michael just tests at 4K for Talos.

[pal666]

Do you understand that "stable" kernel 4.9.9 does have very little bug fixes.

i know what is 4.9.9 but how is it relevant to your comparison of -rc7 vs -wip?

[pal666]

Distro packagers do not know nothing and does not have resources to test. Take the responsibility of your own computer, when you are using Linux.

i'm not idiot like you and selected distro whose packagers are upstream devs