[Nouveau] Status of GF108GLM [NVS 5200M]

[Ilia Mirkin]

On Fri, Apr 10, 2020 at 2:06 PM Jesús J. Guerrero Botella
<jesus.guerrero.botella at gmail.com> wrote:
>
> Hi, Ilia.
>
> Sorry for such a big delay in answering. Real life and that stuff...
>
> I am a newcomer so, please, if I do something wrong regarding my
> quoting style or whatever, just let me know and I'll quickly improve. :)

Perfect so far!

>
> El lun., 30 mar. 2020 a las 13:38, Ilia Mirkin
> (<imirkin at alum.mit.edu>) escribió:
> >
> > Yes, GF108 is Fermi (F = Fermi). Reclocking is currently not available
> > for that generation, unfortunately. You should be able to otherwise
> > use your GPU just fine, but I'm guessing it'll come up in the "07"
> > state when it powers on (in the state as-is it appears powered off,
> > which it will do automatically when not in use), which as you can see
> > is a fraction of the total GPU available perf.
>
> Yes, it comes up in 07 when I launch, let's say glxgears in DRI_PRIME=1.
> It performs similarly to the intel chip, or maybe a bit worse.
> For the rest, yes, it mostly works. But it's no good of course.

Note that glxgears is primarily a measure of PCIe bandwidth in these
setups than actual render performance (basically it's just ferrying
those frames over the PCIe bus as fast as possible to the Intel GPU).
Not sure which intel chip you have ... if it's a SNB, it should be
faster and more featureful (you get GL 4.3 advertised and all the GL
4.5 and most 4.6 exts). If it's IVB or HSW, then it'll be a bit
slower, I expect (esp HSW).


> > There's a very experimental branch that does enable reclocking for
> > Fermi at https://github.com/skeggsb/nouveau/commits/devel-clk .
> > However I believe it was only tested with a single GPU, and my testing
> > with an identical such GPU was negative. On the other hand, you don't
> > have a display hanging off the card, which greatly increases chances
> > of success. Feel free to join #nouveau on irc.freenode.net if you plan
> > on exploring this.
>
> I fetched that and tried to compile it. First failed because of missing envyas,
> which I compiled and installed. Then failed with some undefined symbol kind
> of error which right now I don't have the time to diagnose.

I'm impressed you got this far. You shouldn't need envyas to build, by
the way, only to regenerate after editing fuc source.

So ... this is a funny repository. Basically it's meant to work
alongside the linux kernel tree. It also includes some utilities (like
the ones in bin/*) which essentially run nouveau in userspace. This
can be useful for nouveau development, but most likely not in your
case.

What you want to do is grab kernel v4.16-rc5 (or v4.16 if you're
feeling lucky), build + run that (or alternatively, ask Ben to rebase
the branch on a more recent thing ... or even try that yourself). Then
go into the "drm" directory and run "make" there. That should cause
the nouveau module to get built against your kernel. Then use that
built module for your testing (you could make modules_install from
there, as I recall).

Cheers,

  -ilia