The State of State Trackers In Gallium3D

[phoronix]

Phoronix: The State of State Trackers In Gallium3D

With all of the talk earlier this week regarding the Poulsbo Gallium3D driver and its performance improvements along with the restarted efforts on the Intel 965 Gallium3D driver and then word that Mesa 7.7 may be out by Christmas, it's likely that many are wondering about the current state of the various Gallium3D state trackers that we have been talking about for the past months. Well, here's a few observations on the different state trackers at least where they are at in Git. The most recent Gallium3D state tracker we have talked about is the X.Org/X11 state tracker that really started coming around in September and provides EXA (2D) and X-Video acceleration...

http://www.phoronix.com/vr.php?view=NzY3Mw

[Veerappan]

Thanks for the update. I had been wondering recently what the status of the OpenCL state tracker was. I guess I'm not surprised that it's stalled in favor of getting some of the other work done (xorg state tracker and XvMC state tracker), but it will mean that any OpenCL code I attempt to write will probably end up being on MacOS for now.

Yes, I could use AMD's Stream SDK, and maybe I will for a secondary platform to test on, but I'm hoping to play with GPU-based computing, not CPU.

[leidola]

Yes, I could use AMD's Stream SDK, and maybe I will for a secondary platform to test on, but I'm hoping to play with GPU-based computing, not CPU.

Well, AMD's Stream SDK does support GPU-bases OpenCL -- at least in theory, as it crashes on my Radeon 5750.

[deanjo]

Originally Posted by Veerappan

Yes, I could use AMD's Stream SDK, and maybe I will for a secondary platform to test on, but I'm hoping to play with GPU-based computing, not CPU.

Well there is nvidia that has that available.

In theory, your openCL code should run just fine on any device that has openCL support. Meaning if it runs on the CPU, it should, again in theory, run just fine on a GPU or DSP with openCL support.

[Chad Page]

OpenCL doesn't map terribly well to current GPU's... if you code it wrong at all you'll get far lower GFLOPS than other techniques. radeon 5xxx and moreso Fermi (whenever nVidia can make it actually work!) could help a lot.

I think there's room for something easy to write for that actually runs well on older GPU's

[Veerappan]

Originally Posted by deanjo

Well there is nvidia that has that available.

In theory, your openCL code should run just fine on any device that has openCL support. Meaning if it runs on the CPU, it should, again in theory, run just fine on a GPU or DSP with openCL support.

Yes, in theory, it should work. Anything that I'd be doing in OpenCL at the moment would be toy apps, or proof of concept stuff. At this stage, I'd like that little bit of extra gratification that I'm writing something that runs on the GPU. For now it's toy apps, next spring/fall it will be GPU-computing based classwork in either CUDA or OpenCL.

At the moment, my options for openCL seem to be:
Desktop (Radeon 4770): AMD Stream SDK w/ CPU/GPU openCL in Win7, or CPU/GPU in Linux using fglrx+Stream, or CPU-only using open-source drivers + Stream.
Laptop (GF9400): MacOS w/ GPU openCL. Linux GPU w/ nvidia blob, CPU w/ AMD Stream

It's not that I don't realize that I've got options, just that the one option that I really want is the one that isn't available yet (GPU-based OpenCL in Linux w/ open-source drivers).

[DanL]

Good work, folks! It's really nice to see all the commits on the r300g and softpipe drivers. I hope VIA, nouveau, poulsbo and maybe even SiS users eventually get decent graphics out of this too.