DevHeads.net

AMD ROCm

Hello,
AMD ROCm - Open Source Platform for HPC and Ultrascale GPU Computing[1]
is packaged by upstream only for RHEL/CentOS and Ubuntu.
Is anybody working on packaging it for Fedora? If not, is anybody
interested in setting up a team to work on it?

Have a nice day

[1]: <a href="https://rocm.github.io/" title="https://rocm.github.io/">https://rocm.github.io/</a>

Comments

Re: AMD ROCm

By Felix Schwarz at 01/21/2019 - 03:39

Am 09.01.19 um 11:50 schrieb Germano Massullo:
Just wanted to mention that Tom Stellard has some open review requests for
rocminfo and hcc:
<a href="https://bugzilla.redhat.com/show_bug.cgi?id=1666836" title="https://bugzilla.redhat.com/show_bug.cgi?id=1666836">https://bugzilla.redhat.com/show_bug.cgi?id=1666836</a>
<a href="https://bugzilla.redhat.com/show_bug.cgi?id=1545479" title="https://bugzilla.redhat.com/show_bug.cgi?id=1545479">https://bugzilla.redhat.com/show_bug.cgi?id=1545479</a>

Reviewers welcome :-)

Btw: I terminated perl license checker script after 8h continuous CPU usage
(only 1 core though) when running "fedora-review" for hcc. Is that a known
problem for bigger software? (hcc bundles llvm).

Felix

PS: If you want to test on F29, please rebuild hsakmt/rocm-runtime from
rawhide before.

Re: AMD ROCm

By Germano Massullo at 01/21/2019 - 13:15

Andreas Schneider [1] that is aware of Fedora community work on packaging ROCm stack, showed me his pending pull request [2] that
"cleans up cmake so that the library can be correctly packaged for distributions. It also cleans up cmake as there are several things which should not be done in cmake."

You may want to give a look to it.

Best regards

[1]: <a href="https://github.com/cryptomilk" title="https://github.com/cryptomilk">https://github.com/cryptomilk</a>
[2]: <a href="https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/pull/25" title="https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/pull/25">https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/pull/25</a>

Re: AMD ROCm

By Ian Kent at 01/21/2019 - 04:11

On Mon, 2019-01-21 at 08:39 +0100, Felix Schwarz wrote:
As much as I would like to help I need to get an additional SSD
for my NUC8i7HNK (Radeon™ RX Vega M GL graphics) for that (I hope
fairly soon).

And I think the Vega needs a later version of ROCm as well.

Re: AMD ROCm

By Felix Schwarz at 01/21/2019 - 07:58

Am 21.01.19 um 09:11 schrieb Ian Kent:
Yes, AFAIK Vega M is not (yet?) supported by amdkfd. Also ROCm is mostly
developed/tested for more powerful hardware so I would not hold my breath for
Vega M support.

Anyway I'm happy to test packages with my Polaris card but I have to admit I
don't feel "knowledgeable" enough to review a more complex package like "hcc".
Probably it would help just doing the normal review process (except actually
running the binary).

Felix

Re: AMD ROCm

By Ian Kent at 01/21/2019 - 10:08

On Mon, 2019-01-21 at 12:58 +0100, Felix Schwarz wrote:
Of course, yes, but I live in hope and I'd like to get familiar
with ROCm!

Either way it's hard to properly test something like this.

I think the focus needs to be primarily on packaging so that we
can help upstream along by simply making it available.

That obviously implies developing a relationship with upstream
which is often lacking in distributions with many packages (for
obvious reasons).

Ian

Re: AMD ROCm

By Allan at 01/11/2019 - 21:58

På Wed, 9 Jan 2019 11:50:00 +0100
Germano Massullo <germano. ... at gmail dot com> skrev:
They have actually made scripts available for Fedora 28 and 29, so
users can compile and use the driver. I did that yesterday, and OpenCL
did work, although a bit slower, than when I previously merged in the
OpenCL drivers from their AMDGPU PRO packages for Centos.

Experimental ROCm:

<a href="https://github.com/RadeonOpenCompute/Experimental_ROC" title="https://github.com/RadeonOpenCompute/Experimental_ROC">https://github.com/RadeonOpenCompute/Experimental_ROC</a>

Allan.

Re: AMD ROCm

By Tom Stellard at 01/09/2019 - 22:28

On 01/09/2019 02:50 AM, Germano Massullo wrote:
hsakmt and the rocm-runtime are currently packaged in Fedora, but aren't
up-to-date with upstream. I've wanted to package hip as well, but I've
been waiting for hip to drop the hcc dependency since it is so hard to package.

I've also done coprs for older versions[1] of ROCm that included a complete
stack up to tensorflow.

-Tom

[1] <a href="https://copr.fedorainfracloud.org/coprs/tstellar/rocm-1.7-userspace" title="https://copr.fedorainfracloud.org/coprs/tstellar/rocm-1.7-userspace">https://copr.fedorainfracloud.org/coprs/tstellar/rocm-1.7-userspace</a>

Re: AMD ROCm

By Felix Schwarz at 01/13/2019 - 07:32

Am 10.01.19 um 03:28 schrieb Tom Stellard:
Do you know when this is supposed to happen? Any public comments from AMD on
that issue?

Besides hcc might come handy in some cases. Would you mind sharing some
information why hcc is hard to package?

(Also I recently read that AMD is interested in shipping the ROCm stack in
more distros as the mainline kernel comes with the necessary bits. Maybe they
would be willing to work with Fedora make hcc easier to package?)

Generally I'd like to see ROCm support in Fedora as I think it really fits
Fedora's mission pretty well: Fedora ships very recent software so that should
be fine and in my experience very few machine learning users looking for
long-term stability. Easy installation and ability to debug issues are more
important.

Also "Machine Learning with Free Software" would make nice Fedora feature name :-)

Felix

Re: AMD ROCm

By Tom Stellard at 01/16/2019 - 13:23

On 01/13/2019 03:32 AM, Felix Schwarz wrote:
I'm not really sure, but I've started to look again at packaging hcc.

The main issues are:

- Lots of hard-coded paths in the source to /opt/hsa
- Bundles its own llvm, clang, and lld.

-Tom

Re: AMD ROCm

By Ian Kent at 01/09/2019 - 20:16

On Wed, 2019-01-09 at 11:50 +0100, Germano Massullo wrote:
Interesting approach to HPC.

Might be a useful way for me to get some exposure to HPC which
I'd like to do.

Re: AMD ROCm

By Matthew Miller at 01/09/2019 - 11:20

On Wed, Jan 09, 2019 at 11:50:00AM +0100, Germano Massullo wrote:
Yes, there's definitely interest -- I actually talked to someone from AMD
about this a few weeks ago.