Intel Gallium3D Graphics Driver Performance

[lapis]

Stephane stopeed the development of driver?

Because using 915g on sauerbraten runs pretty well (including shaders).The i915c runs with lots of bugs and shaders are very slow.

[lapis]

Someone should create a fundraising project to pay a full time developer.

[wirrbeltier]

Hm. Looks like most of the benchmarks were rather CPU bound anyway (largely the same scores even at different resolutions). But still, the gallium driver's results seem to be 'offset', which either means that it produces more CPU churn (by having more overhead) or it utilizes less of the hardware GPU capabilities, therefore using more software fallback... Probably a combination of both.
Are there some charts of CPU utilization during the benchmarks for both classic and gallium drivers? Or perhaps some data about the amounts of memory I/O?

[smitty3268]

Yes, Michael, can you tell us how you built the driver?

So we know if LLVM was enabled or not. I looked over the openbenchmarking site and found a bunch of logs but nothing about how you built the driver.

[FireBurn]

I thought there was an i965g driver too

[agd5f]

I was actually wondering, and couldn't find an answer for this, whether i915g uses llvmpipe for software fallbacks (like the vertex pipeline, since the GMA950 doesn't do vertex processing in HW). If not, why not? It seems this would provide a pretty good performance benefit for machines with decent CPUs (like my tablet that has a Core 2 Duo and a GMA950).

gallium drivers can use llvm for vertex shaders if they are configured to use it. You need to enable llvm in your mesa build to use it.

[TechMage89]

Also, since the tested machine runs an Atom processor, it likely wouldn't derive as much benefit from that as my machine. I would be nice to see a test with a machine that has a higher-end processor. I might just have to test this myself.

[TechMage89]

I was actually wondering, and couldn't find an answer for this, whether i915g uses llvmpipe for software fallbacks (like the vertex pipeline, since the GMA950 doesn't do vertex processing in HW). If not, why not? It seems this would provide a pretty good performance benefit for machines with decent CPUs (like my tablet that has a Core 2 Duo and a GMA950).

[airlied]

I wonder if llvm was used, since that is the major use case for i915g and its major feature difference from i915c.

[ChrisXY]

Actually I made a PKGBUILD for mesa with i915g and uploaded it to the Archlinux AUR:

http://aur.archlinux.org/packages.php?ID=46516

(look at the PKGBUILD for build options)

The next day someone created a thread about how i915 is faster than i915c...

https://bbs.archlinux.org/viewtopic.php?pid=896683

I have not really tested how it really performs but I use it on my eee 1000h now.
Could Phoronix specify why i915g is "not nearly as mature as the Radeon or Nouveau Gallium3D drivers"?
I mean, I can run xonotic on it! The only problem is the performance but you don't expect a netbook to have good performance anyway...