x11/nvidia-driver-340 and OpenCL

[Scott Bennett]

Jan Beich <jbeich at vfemail.net> wrote:

> Scott Bennett <bennett at sdf.org> writes:
>
> >      Thank you very much for that information, Jan.  I will take a look at
> > that, but if the situation is that chancy and would mean being stuck with
> > 32-bit emulation, it probably wouldn't be my best option after all.
>
> lang/pocl may also be useful, mainly when debugging OpenCL in apps.

     Right now I'm still on 10.3-STABLE, so lang/pocl is not an option.  Soon,
I hope, I will install 11.1-RELEASE, but because it will be a fresh
installation into a ZFS pool, whereas my current 10.3-STABLE is on several UFS2
partitions, I need to find the time to draw up a plan to do it without losing
my existing configuration files and then to get it done.  Then I can install
lang/pocl.
>
> >      Do you know whether a 64-bit driver (for later Nvidia cards, of course)
> > is in the works with up-to-date features?
>
> No. Try pinging upstream on https://devtalk.nvidia.com/default/board/97/

     Thanks for the URL.  Hmm.  There seems to be a problem with it at the
moment, so I'll try it again later.
>
> > Or would I do better by buying a more recent Radeon card?
>
> Did lang/clover not work on Radeon HD 3870 e.g., with radeonkms + r600g?
> If so, try upgrading just the kernel to drm-next branch.

     No, and according to the AMD web sites, the GPUs on those cards are not
capable of it.  Although they support the now abandoned ATI Streams software,
they do not support OpenCL.  The Radeon HD 3870x2 cards were the last to
be released that did not support at least OpenCL 1.0. :-(  IIRC, the GPUs on
these cards are RS780 chips, but I may be misremembering the number because
I haven't looked at those web pages for quite some time.
>
> Personally, I use lang/beignet on Intel Skylake GT2 (iGPU). The performance
> is decent enough to accelerate some jobs e.g., waifu2x-convert-cpp,
> ImageMagick, FFmpeg (a few filters but VAAPI encoding also works). X11 display
> stutters if OpenCL is running on the same GPU, but neither setpriority(2)
> nor rtprio(2) help. Switching X11 display to NVIDIA Kepler GK107 (dGPU)
> eliminated the stutter. However, there's probably a better way to schedule
> jobs on GPU.

     So the problem is not in the feeding of the GPU, but in its internal
scheduling.  I'm not terribly surprised somehow.  The GPU may have to finish
a work item or maybe its entire work queue before it can switch to another
piece of work, e.g., an OpenGL command.  I've long suspected that OpenCL
should probably only be done on GPUs that are not simultaneously supporting
monitors to avoid such problems or contention for graphics device memory.
     I only use one monitor, so I had thought that I could use three of the
four GPUs for OpenCL stuff, but the dratted cards don't do OpenCL, so three of
the four GPUs are of no help to me and just waste electricity.  The second
card prevents use of a slot between the two cards, so removing it would
potentially open both slots, depending upon what I decide to replace it with.
I have an unused USB 3.0 card that would speed up a few things by separating
devices across controllers, or I could buy a SATA 3 card somewhere or possibly
do both.  The first option is the only one that I could do for free at this
point, but that wouldn't get me OpenCL support, although, as you point out,
I should be able to use lang/pocl for development/testing (once I get 11.1
installed), just not for performance.
     Thanks again for the suggestions.


                                  Scott Bennett, Comm. ASMELG, CFIAG
**********************************************************************
* Internet:   bennett at sdf.org   *xor*   bennett at freeshell.org  *
*--------------------------------------------------------------------*
* "A well regulated and disciplined militia, is at all times a good  *
* objection to the introduction of that bane of all free governments *
* -- a standing army."                                               *
*    -- Gov. John Hancock, New York Journal, 28 January 1790         *
**********************************************************************