Running Ivy Bridge HD Graphics With Linux 3.16 + Mesa 10.3

[phoronix]

Phoronix: Running Ivy Bridge HD Graphics With Linux 3.16 + Mesa 10.3

Given our recent updated Sandy Bridge benchmarks on Linux, for those with Ivy Bridge processors curious how the HD Graphics are handling the latest Mesa and kernel, I have some updated benchmarks for you this Sunday...

http://www.phoronix.com/vr.php?view=MTc3MTg

[_deepfire]

Pretty please..

The quadruple increase in the real-world benchmarks (Unigine Tropics and Sanctuary) "aren't exciting enough"?

I would say -- to the contrary!

It would be interesting to see the real scope of the phenomenon -- it could well be possible that it affects other modern GPU loads as well..
It might be that a change in the set of supported OpenGL extensions triggered a change in the rendering path, and so it might be confined to Unigine specifically. But it could also be that a certain bottleneck was killed, and it has a wider scope. It could be a mix of the two. The possibilities are numerous..

Yes, you said "aren't exciting enough to warrant a multi-page featured article on Phoronix", but this is what I'm questioning.

Such an enormous improvement in a _real world load_ is worth an investigation!

[mendieta]

The quadruple increase in the real-world benchmarks (Unigine Tropics and Sanctuary) "aren't exciting enough"?

I would say -- to the contrary!

It would be interesting to see the real scope of the phenomenon -- it could well be possible that it affects other modern GPU loads as well..
It might be that a change in the set of supported OpenGL extensions triggered a change in the rendering path, and so it might be confined to Unigine specifically. But it could also be that a certain bottleneck was killed, and it has a wider scope. It could be a mix of the two. The possibilities are numerous..

Yeah, I thought the same. A lot of the time, these amazing improvements have to do with a slow feature actually not supported in the "fast" software stack (say, in the fast test, antialising is not supported, so that test runs much faster). This is really where benchmarking graphics becomes tricky. It would be nice if a test that requires a certain feature set, either goes through or fails (as opposed to to running faster). But, as you point ouy, this _could_ be real.

[mendieta]

I just tested on my Haswell i5 system (I run something very similar, Oibaf ppa and kernel 3.16.1). I get in one of the unigine tests a white screen and about 100 fps (it is normally about 15 fps).

[chrisf]

The 4x increases are unfortunately the result of the app being broken.

Sanctuary and Tropics both use extensions in their shaders without bothering to enable them [IIRC, ARB_instanced_arrays, which provides gl_InstanceID]. Mesa has an existing workaround for this, where for the unigine demos we enable *all* extensions in their shaders, but in "warning mode".

Unfortunately, now that mesa supports ARB_gpu_shader5, that is getting enabled as well, which turns "sample" (which is also used by these shaders as an identifier) into a keyword, causing them to fail to compile.

So we're in the unpleasant situation of having to workaround the workaround for their buggy shaders


Would have been nice to get an actual 4x boost in these demos, but not this time...

[mar04]

Wouldn't it be easier to ask Unigine for a proper fix?

[Kayden]

Wouldn't it be easier to ask Unigine for a proper fix?

Feel free to ask them. I don't think they're interested in updating their older demos (Tropics and Sanctuary). I've personally had no luck.

[dungeon]

@_deepfire

When i see "improvments" in very large margin like this, i first think rendering must be broken there . And also here:

http://www.phoronix.com/scan.php?pag...tem&px=MTc3MTM
http://www.phoronix.com/scan.php?pag...e_linux1&num=2

It sounds much more believeable, but i also think render is broken in OpenArena 0.8.5/UrbanTerror benchmarks too . So someone need to investigate those cases too.

But in one thing might be you are right, it is broken or something really extraordinary happen .