Talk:GPGPU

From ArchWiki
Jump to: navigation, search

video hardware decoding

Is it worth mentioning video hardware decoding with cuda (through mpv) ? Ivanoff (talk) 07:21, 21 May 2017 (UTC)

I'd say no, that kind of stuff belongs to Hardware video acceleration, though we might want to add that to Related articles kralyk 08:31, 21 May 2017 (UTC)
what about the List of OpenCL and CUDA accelerated software section Ivanoff (talk) 09:12, 21 May 2017 (UTC)
That wouldn't be right, video decoding acceleration is not built on top of OpenCL or CUDA, it is a native feature of the GPU. I have added a link to related articles, that should do it. | kralyk 10:05, 21 May 2017 (UTC)
Understood. I mainly talked abount the option of mpv with the option --hwdec=cuda (Any platform CUDA is available) , which uses cuda directly instead of vdpau. I admit, slightly off topic. Just saying. Regards Ivanoff (talk) 12:08, 21 May 2017 (UTC)
Ah, so it is actually CUDA-based, that's a different story. Sorry, I didn't understand you right. Okay, in that case I suppose it could go the List of OpenCL and CUDA accelerated software | kralyk 08:37, 22 May 2017 (UTC)

Missing compute-runtime ("NEO") for intel drivers

As per this Phoronix news and the note at Beignet upstream linked in the beignet, the Beignet is deprecated for Broadwell+ in favor of the new "NEO" OpenCL driver. I was planning to add this driver to the wiki, but I couldn't find the package for it neither in the repository or AUR. With that in mind, the question is: should the entry be added without appropriate package until someone make one, or it shouldn't be added at all until the package arrives? The code for the new driver is at GitHub's intel/compute-runtime and the project file is at 01.org's compute-runtime, but I'm not sure I'm up for task to maintain an AUR package for it plus I'd want to confirm if there's no package for the new NEO driver, or I just can't find one. Faalagorn / 15:02, 30 April 2018 (UTC)

EDIT: I have added the link to project's github for now. In case an AUR or official package exist, please add them there Faalagorn / 18:12, 2 May 2018 (UTC).

Adding missing and explaining AMD OpenCL implementations confusion?

I've been trying to get grasp of OpenCL implementations in AMD cards, and so far I've been confused. From what I understand, there are at least three implementations of open sourced OpenCL – ROCm for the Polaris, Vega and upcoming cards, "PAL" for all (GCN?) cards that is slightly slower and Mesa's GalliumCompute/"Clover" (stuck at OpenCL 1.2 and generally worse performing) and at least one proprietary implementations, "legacy" for older GPUS (including TeraScale), that is the main reason to use AMD Catalyst on older cards, as it provides OpenCL 2.0.

Now, AMDGPU-PRO (lately renamed to Radeon Software for Linux) lists three switches: --opencl=legacy, --opencl=pal and --opencl=rocm (in addition to combo of --opencl=legacy,pal). However, both ROCm and PAL were opensourced, although I'm not sure if PAL's opensource OpenCL exist.

Then there's the dubbed old "Orca" [1] [2] OpenCL implementation, is it the same as legacy?

Finally, the wiki lists three (four including CPU runtime, which I wasn't aware of) ways to install AMD drivers: opencl-amdAUR: proprietary standalone runtime for AMDGPU, amdgpu-pro-openclAUR: proprietary runtime for AMDGPU PRO (PAL? ROCm?), opencl-catalystAUR: AMD proprietary runtime, deprecated in favor of AMDGPU (Legacy/"Orca"?) – it all dubbes them "proprietary", but is any of them related to ROCm or PAL that were opensourced down the road? I can't tell, as I wasn't following AMD that closely or using OpenCL at the time they open sourced ROCm library, plus like I told I'm not sure I'm not

Finally, the article doesn't mention the open source implementation of ROCm at all.

It would be nice to clear some confusion, especially for AMD cards! Faalagorn / 18:06, 11 May 2018 (UTC)

EDIT: Seems that while the most PAL code is opensourced, the OpenCL-to-PAL portion is not, so PAL OpenCL indeed remains proprietary for now: [3]. Faalagorn / 20:03, 11 May 2018 (UTC)