Generally the performance looks very decent considering this is open source and coded with relatively few resources. This gives me a lot of hope for future versions. I would consider 50% performance of the proprietary driver a huge success, especially since it will just work out of the box and solve for example the tearing issue. And I guess getting to ~20% as now is probably much much closer to 50% than it looks
Yeah, which is the reason I believe the Linux driver model isn't up to par with proprietary operating systems. Binary blobs on Linux need to replace big chunks of X.Org to work properly and the kernel interface sucks.
Originally Posted by 89c51
There's also a "conspiracy theory" around this, saying that this is intended by the developers so that big Linux users pay for support from the likes of Red Hat and such.
Enabling color tiling and pageflipping, and disabling compiz would at least double the open source results.
Originally Posted by bongmaster2
Still, the interesting thing about the results is that the driver is heavily CPU-bound, which means once these bottlenecks are found and fixed, the performance should increase rapidly.
All r600 and r700 GPUs have basically the same performance, although some of them are 5x faster than others.
No, actually, they support the vast majority of 3d related features, and are missing a minority.
Originally Posted by Mr James
The main problem with the open drivers (other than GL3+ support) is performance.
It would make sense to run a window manager which switches compositing off for full-screen apps.
Originally Posted by BlackStar
It's 2011, after all. If compiz can't cut it, use something more sane.
Basic income could be a solution to this problem...
Originally Posted by 89c51
Not true, SM3.0 is a strict subset of GLSL1.2 with ARB_shader_texture_lod, both of which can be supported by a GL1 driver.
Originally Posted by evolution
Based on my findings, the low performance can be attributed to these facts:
- We don't use hardware to its full potential (e.g. HyperZ is either not implemented (r600g) or not enabled by default (r300g)). Page flipping is about to emerge in the next kernel. Tiling is not implemented in r600g to my knowledge (and according to the feature matrix). I don't think this point matters much, because the CPU is what limits performance the most.
- There is a command stream checker in the kernel which gives away a little of speed in exchange for security. I'd personally like to see that dropped, but Linus wouldn't accept it.
- The Mesa state tracker and Mesa core slow things down. Some state tracker optimizations are being worked on and that should make the performance a little better. The whole thing could be made faster if the classic driver interface was dropped and both the core and the state tracker were merged together and simplified. This won't happen until Intel jump over to Gallium. Intel are still doing the majority of work on Mesa, so we can't just ignore them.
- Uploading user buffers (like vertices not being in VBOs) is embarrassingly slow. If you wonder why Lightsmark is such a huge failure with open source drivers, then this is the reason.
Ranting about missing GL3 without mentioning WHY GL3 cannot be supported gets old. There are patent issues preventing us from implementing GL3 fully, so it's out of the question now.
It would be very useful if we had comparisons like this article every month or more often, so that we don't live in illusions.
Things havenīt changed much in the last 4 years, performance wise. Gallium3D only boosted performance a bit in certain cases. Knowing about the situation with the open drivers, I donīt really understand why all the complaints about performance. What has changed in the last 3-4 years is the amount of support from the driver. Itīs simply amazing how far it has come and this is what really should be praised above all else.
Originally Posted by marek
marek since you are a dev and probably know more on the subject do you see Intel dropping classic and going Gallium and if when whats the timeframe.
i remember reading that they were not keen on making the jump