Planned Vega support in Linux

Alex Deucher alexdeucher at gmail.com
Wed Aug 9 04:07:40 UTC 2017


On Tue, Aug 8, 2017 at 11:46 PM, David Niklas <doark at mail.com> wrote:
> BTW: I'm on list, no need to CC.
>
> On Tue, 8 Aug 2017 20:30:02 +0000
> "Deucher, Alexander" <Alexander.Deucher at amd.com> wrote:
>> > From: David Niklas [mailto:doark at mail.com]
>> > Hello,
>> > I'm not seeking any secret info.
>> > I'm a Gentoo Linux user who wants to run some massively parallel
>> > experiments.
>> > I'm unwilling to wait and see what kind of support Linux gets because
>> > the annoying bitcoin miners tend to cause the price to go through the
>> > roof quickly.
>> > My current card is an AMD SI HD7780 by MSI.
>> > I know that currently some form of headless support already exists,
>> > but then I'd have to use crossfire (which does not work with the
>> > opensource driver AFAIK).
>> > I filed a ticket at AMD on 05/07/17 but AMD has not gotten back to me
>> > on this matter and like I said, if I wait then I risk having to buy
>> > something way out of my price range.
>> > --And I really have tried to wait till the last minute--
>> > Anything official or unofficial will do.
>> > I am planning to use the OpenCL language to do my MP experiments.
>>
>> I'm not quite sure what you are asking.  We released initial open
>> source support for vega10 months ago and it continues to evolve as new
>> SKUs are launched.
>
> I used AMD pro stack for sometime before the AMDGPU driver evolved to
> something useful. Believe me, I do well to ask[1].
>
>> It is expected that you can get a complete working
>> stack in open source at launch time for whatever sku you are interested
>> in.  Not all of the patches are upstream, but they are all public.
>
> Is the opensource driver you're thinking of at the url that harry gave
> me?
> What it the state? (beta, alpha, etc.).
> About how complete is it with respect to the other opensource GCN drivers?

It's the same driver.  However, kernel and other open source
components do not always align to product cycles so support may not be
upstream in the various open source projects yet.  The vega10 code
that is upstream doesn't not have full functionality for all vega10
skus or features because older kernels were released before all the
products were finalized.  The amd-staging-4.12 branch Harry pointed
out has the latest code which provides full functionality for the
asic.

>
>> OpenCL support is provided by the ROCm stack which you can get from the
>> gpuopen site.
>
> "We do not support ROCm with PCIe Gen 2 enabled CPUs such as the AMD
> Opteron, Phenom, Phenom II, Athlon, Athlon X2, Athlon II and Older Intel
> Xeon and Intel Core Architecture and Pentium CPUs."
>
> Ouch!
> Mine is a Phenom II 6 core 1090T.
> I can't go spending a ton on a new CPU, MB, RAM, and GPU (and I'd need a
> less hacked case too :)
> Granted I could get an i3, cheap the MB and RAM, but I've worked so hard
> for a great platform... 32GB RAM, 6 core Phenom II, HD 7780 GPU, RAID 5
> on 2TB HDs, Bluray drive, Mechanical keyboard, HP laser printer,
> 4800x4800dpi HP scanner...
> AMD is normally so good at compatibility too.
> Ok, I'm over it, what is doable at this point?

PCIe gen 3 is required for pcie atomics.  I'm not a ROCm expert so I'm
not sure if that is a hard requirement or only if you need platform
atomics.

Alex

>
>> Headless boards are supported just fine and have been
>> for years.  If you prefer packaged drivers, you can download the ROCm
>> stack from gpuopen or the pro stack from amd.com.  All stacks build on
>> the same open source kernel driver and other components.
>>
>> Alex
>>
>
> Good to know!
>
>
>
> [1] No offence, but AMD's gpu-pro would not allow the system to suspend
> (sleep), also I'd often get a black and white set of vertical lines
> through my display and the kernel would totally freeze (no sysreq). The
> worst part is that it would happen totally randomly (really).
>
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>


More information about the dri-devel mailing list