The Current Intel "Iris" Gallium3D OpenGL Performance Against i965 Mesa, Windows 10 OpenGL

[treba]

Thanks for running these, Michael! It's always really nice to have an extra set of data about how we're doing.

I fixed Portal, it's now 3.86% faster than i965 on my laptop. Hadn't looked at that one until I saw this.

Nice

Will we that go into 19.1? I really looking forward to give Iris a try as soon as 19.1 arrives in Fedora

[R41N3R]

Is it possible to activate Iris globally even for a compositor like kwin? I would like to compare the experience on my Intel notebooks.

[treba]

Is it possible to activate Iris globally even for a compositor like kwin? I would like to compare the experience on my Intel notebooks.

I guess putting

MESA_LOADER_DRIVER_OVERRIDE=iris

into your

/etc/environment

should do

[FireBurn]

I guess putting

MESA_LOADER_DRIVER_OVERRIDE=iris

into your

/etc/environment

should do

This patch from Kayden should get it working for you as default - also allows you to not compile i965 or clasic in Mesa https://gitlab.freedesktop.org/kwg/m...dcb631a73d4903

Be warned though vaapi doesn't work quite right yet due to some EGL stuff (my memory if failing me) not being implemented yet

[Eero]

Looking at the bad GpuTest Triangle results...

Phoronix uses 60s run-time for GpuTest tests, and the reported score is simply number of rendered frames, so one can get FPS just by dividing score with 60.0. Which is very low for the Triangle test that on every frame just clears the window and draws single half screen sized triangle. Because shader used in the test is trivial, test is memory bandwidth limited.

I know that:
* Windows doesn't composite fullscreen windows (at least with Intel GPUs)
* On my own machines that are running 18.04 with Unity, they don't get composited either, and I get same performance as on Windows (with Git versions of Mesa & X) with the Fullscreen Triangle test, and in windowed version, Linux is faster
* Ubuntu 19.04 has new enough X server (>= 1.20.1) and Mesa (>= 18.1) that it supports end-to-end render buffer compression out of the box on GEN9

So only conclusion is that Michael's setup composites even fullscreen windows. Composition (more than) doubles the bandwidth required for the Triangle test, which would explain half of the expected performance.

This is going to have impact also in other memory bandwidth limited (= most high FPS) tests. Comparisons against Windows should use desktop environment that disables composition in fullscreen, like Windows does.