Mesa 19.1-RC5 Is Out With A Handful Of RADV & Intel/Iris Changes

[phoronix]

Phoronix: Mesa 19.1-RC5 Is Out With A Handful Of RADV & Intel/Iris Changes

Mesa 19.1 is in overtime and today marks the fifth weekly release candidate as the developers try addressing the last two blocker bugs to get out this quarterly feature release...

http://www.phoronix.com/scan.php?page=news_item&px=Mesa-19.1-RC5-Released

[kripteks]

With my vega 56:
19.0.3-1, 19.0.5: glxgears 24k fps
19.1.0-rc5, 19.2-devel: glxgears 16k fps
with unigine heaven 4 no problem equal fps.
(i use on debian testing/experimental: git mesa, cd mesa, mkdir build, cd build, meson --prefix=/usr --libdir=/usr/lib/x86_64-linux-gnu/ --libexecdir=/usr/lib/x86_64-linux-gnu/ .., ninja install)


I know glxgears is not a gpu benchmark:
7850: 28 fps (oc 1.1 ghz)
7950: 26 fps (oc 1.175 ghz)
vega 56: 23-24 fps (stock 1.59 ghz)
In unigine heaven vega 56 > 7950 > 7850, but why fps down when gpu power is up with glxgears i dont understand.

[smitty3268]

With my vega 56:
19.0.3-1, 19.0.5: glxgears 24k fps
19.1.0-rc5, 19.2-devel: glxgears 16k fps
with unigine heaven 4 no problem equal fps.
(i use on debian testing/experimental: git mesa, cd mesa, mkdir build, cd build, meson --prefix=/usr --libdir=/usr/lib/x86_64-linux-gnu/ --libexecdir=/usr/lib/x86_64-linux-gnu/ .., ninja install)


I know glxgears is not a gpu benchmark:
7850: 28 fps (oc 1.1 ghz)
7950: 26 fps (oc 1.175 ghz)
vega 56: 23-24 fps (stock 1.59 ghz)
In unigine heaven vega 56 > 7950 > 7850, but why fps down when gpu power is up with glxgears i dont understand.

Wasn't there a change a while back to help out somebody with an external GPU over thunderbolt connection? As I recall, it changed things so that there were 2 kernel calls where there used to be one, and it was expected to hurt glxgears performance for that reason.

Ultimately kernel API needed to be changed to allow 1 call again, but AMD doesn't seem to be in a hurry to do it, and it doesn't really affect anything except glxgears.

[nuetzel]

Wasn't there a change a while back to help out somebody with an external GPU over thunderbolt connection? As I recall, it changed things so that there were 2 kernel calls where there used to be one, and it was expected to hurt glxgears performance for that reason.

Ultimately kernel API needed to be changed to allow 1 call again, but AMD doesn't seem to be in a hurry to do it, and it doesn't really affect anything except glxgears.

The other way around:

We need a new call with two command buffers at the same time.
I reported the slowdown to Marek.

Was:
Re: [Mesa-dev] [PATCH 0/4] RadeonSI: Upload constants to VRAM via SDMA
[-]>> Am 26.02.2019 07:36, schrieb Marek Olšák:
>>> We need to extend the CS ioctl to allow submitting 2 command
>> buffers
>>> at the same time.

https://lists.freedesktop.org/archives/mesa-dev/2019-February/215816.html

It came with the 'primitive culling stuff'.

https://lists.freedesktop.org/archives/mesa-dev/2019-February/215810.html

[-]
Trivial benchmarks such as glxgears can expect 20% decrease
in performance due to the added cost of the SDMA CS ioctl that wasn't
there before.
[-]

Greetings,
Dieter