AMD Athlon's R3 Graphics: RadeonSI Gallium3D vs. Catalyst

[dungeon]

Why not Mesa devel?

Because if i tested that yesterday, then you will probably ask why not with Marek's commits of today .

I am interested into looking if i can reproduce hyperz artifacts and lockups... but i can't found that yet .

[przemoli]

It is one frame worse in openarena . This is with Debian Sid, kernel 3.13, mesa 10.1, etc. hyperz enabled = 32 fps .

http://openbenchmarking.org/result/1404151-PL-OPENARENA67

With 3.15-rcs i have 31 fps, but missed to upload it



Hyperz is always good to have, even in 1280x1024 it goes from 25 fps to 32 fps .

Why not Mesa devel?

[dungeon]

It would be interesting to see how 3.15 compares as well since that contains a number of memory management optimizations.

It is one frame worse in openarena . This is with Debian Sid, kernel 3.13, mesa 10.1, etc. hyperz enabled = 32 fps .

http://openbenchmarking.org/result/1404151-PL-OPENARENA67

With 3.15-rcs i have 31 fps, but missed to upload it

Isn't 1080p high enough resolution where hyperz we would make a difference?

Hyperz is always good to have, even in 1280x1024 it goes from 25 fps to 32 fps .

[smitty3268]

Isn't 1080p high enough resolution where hyperz we would make a difference?

Also, does anyone know what the gpumark tests triangles and pixbuf piano actually do?
The former sounds like it is just creating a bunch of triangles, which, if done in the worst way, would seem to be a good indicator of how fast each driver can issue instructions. That might provide some kind of useful info into driver efficiency?

Dungeon showed massive performance boosts for hyperz at 1080p resolution. It makes a big difference on bandwidth limited platforms like this. Even at the lower resolutions i think it was 5-10%, which is a decent chunk of some of these results. Some of the memory optimizations in 3.15 and that curaga looked at could make a big difference on those tests that show 3 fps instead of 5 fps.

I believe triangle ends up the same as glxgears, which last i heard is basically limited by the latency rate of X commands in the DRI2 drivers, because it has a synchronous X command that blocks everything until it returns each time. DRI3 makes the call async which should speed it way up, comparable to the proprietary drivers. But i haven't actually seen any test results to confirm that.

[curaga]

The triangle test is basically glxgears, it measures swap speed/cpu overhead/memcpy speed.

[liam]

Isn't 1080p high enough resolution where hyperz we would make a difference?

Also, does anyone know what the gpumark tests triangles and pixbuf piano actually do?
The former sounds like it is just creating a bunch of triangles, which, if done in the worst way, would seem to be a good indicator of how fast each driver can issue instructions. That might provide some kind of useful info into driver efficiency?

[agd5f]

It would be interesting to see how 3.15 compares as well since that contains a number of memory management optimizations.

[AJSB]

Nice results, can't wait to see dGPU results in special with GTX750(Ti) with NVIDIA blob

[mannerov]

Do you have a clue why you had a DRI3 XCB error?

DRI3 support isn't enabled for these cards both in Mesa and in the radeon DDX, so that's strange.


I've made myself some benchmarks with my radeonsi card (hd7730m), and for unigine heave 4.0,
I got last year a score of 234 (9.3 fps) with Catalyst.
Now with recent Mesa and kernel 3.13, I get a score of 200, which is very close.

[ua=42]

Summary.
Open drivers when running latest kernel mesa is 2-3 times faster than Ubuntu 14.04 stock.
The latest open driver stack performed within 20% to 80% of Catalyst on 3D benchmarks.
There were performance gains both from a newer kernel and from using the latest mesa.
Micheal is planning to run some 2d tests on the same hardware.

Opinion:
As always, I'm impressed by the rate that the open drivers progress at, but it is obvious that for 3D tasks, this hardware's driver needs some more work.