GLAMOR + RadeonSI 2D Acceleration Is Quite Good For Open-Source AMD 2D Performance

[dungeon]

I wonder why the radeonsi driver has trouble drawing circles as fast as catalyst.

It is not a trouble just a metter of question what would be best default for all (2D, 3D, video) worlds together ,

If you prefer fast 2D, you can turn on ShadowPrimary radeon option and win all these 2D/gtkperf benchmarks at expense of some 3D speed

In amdgpu driver you can even disable glamor so 2D all getting done with the CPU, while 3D/video still accelerated fine, etc..

[smitty3268]

It COULD be the driver, but it could also be GLAMOR since all 2d operations go through glamor. You'd have to debug each individually and see which one is hitting a slow path.

I believe glamor's idea of accelerating circles is to draw a huge number of 1px wide lines, or something really slow like that.

In practice, it doesn't matter, because no actual app is going to draw a circle using XRender.

[MrCooper]

Option "ShadowPrimary"

Michael,

it would be interesting to see a comparison of 2D benchmarks with and without Option "ShadowPrimary"[0] available in current xf86-video-ati Git master. From my experience developing that feature, it can speed up e.g. gtkperf significantly.

[0] For a good user experience, I recommend enabling Option "TearFree" along with Option "ShadowPrimary".

[dungeon]

@MrCooper

Yup gtkperf is very fast with ShadowPrimary (i can beat all these Michael results with Kabini ), of course that comes with expense of minus 20-30% for GL games and lets mention some random browser bencmarks goes much down for example this one:

http://ie.microsoft.com/testdrive/performance/psychedelicbrowsing/Default.html

[bridgman]

i find radeon slower than intel onboard video. simple things like scrolling in chrome are slow.

i'm struggling performance wise with radeon 7750, and i compared to r9 290. and they both seem pretty slow. (they're both gddr5 cards, but r9 290 has way wider data bus)

Shared memory devices (what we call an APU) have advantages on simple benchmarks, since the video memory can be accessed very quickly by the CPU. One of the sad realities of 2D benchmarking is that for most tests the fastest implementation is to have the CPU render into system memory then blit the results to the screen. Probably the same for scrolling in a browser, at least it used to be that way.

[Ericg]

Shared memory devices (what we call an APU) have advantages on simple benchmarks, since the video memory can be accessed very quickly by the CPU. One of the sad realities of 2D benchmarking is that for most tests the fastest implementation is to have the CPU render into system memory then blit the results to the screen. Probably the same for scrolling in a browser, at least it used to be that way.

Other than lower CPU utilization, whats the benefit then of doing it on the GPU? It can't be power efficiency since if the CPU does it faster than the GPU then the CPU wins race-to-idle.

[Ardje]

I would like to see an xterm refresh test.
If you start an xterm and chrome or steam, and you do not have any compositors (why should you...), fglrx is literally slower than my (late) vic-20. So to get it workable, I start compton, then start steam, click to the right item, kill compton, and click.
Radeon otoh runs like a charm. 2d font rendering rocks.
This is on an R9-270.
Any recent amd card is slow. It used to work on the older cards, but HD4890 and R9-270... bah.

[mpw1412]

I'm wondering, weather GTKperf really runs faster on the open driver or does the open driver maybe just lack some feature, which are just skipped then and this might be the reason, why the open driver seems to be so much faster?

As far as I know, this is the case in some 3d benchmarks, where the open driver seems to be faster then the Catalyst driver.

Greetings
MPW

[bridgman]

Other than lower CPU utilization, whats the benefit then of doing it on the GPU? It can't be power efficiency since if the CPU does it faster than the GPU then the CPU wins race-to-idle.

If you're mixing 2D and 3D rendering you end up spending all your time copying data & flushing caches between CPU and GPU, so doing everything on GPU ends up being faster. For 3D rendering the GPU is much faster, obviously. I imagine the ShadowPrimary option mentioned above uses CPU for 2D rendering, but haven't played with it myself.

[MrCooper]

I'm wondering, weather GTKperf really runs faster on the open driver or does the open driver maybe just lack some feature, which are just skipped then and this might be the reason, why the open driver seems to be so much faster?

There's no such thing with gtkperf. If one of the drivers didn't implement some of the functionality, the corresponding rendering would simply be missing, which should be pretty obvious.