AMD ROCm 1.9 Available With Vega 20 Support Plus Upstream Kernel Compatibility

[phoronix]

Phoronix: AMD ROCm 1.9 Available WIth Vega 20 Support & Upstream Kernel Compatibility

For months we have been looking forward to ROCm 1.9 as the latest feature update to the Radeon Open Compute stack while on Friday that big release finally took place. This ROCm update for GPU compute purposes has a lot of new features...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-ROCm-1.9-Released

[ms178]

Great news, even though I don't own a compatible card yet (waiting impatiently for Navi to arrive). I am following the HSA development for years and still hope to reap the benefits out of it as a user some day in the future. Congrats to Gregory Stoner and the involved teams! I very much like his way of interacting with the community on Github, being as transparent as possible about development. By the way, the HSA Foundation was meant to release an update to its specification this year based on the cooperation with their Chinese chapter. As there is little to no information flow in this regard, it would be nice to know in which direction they are heading and how it fits into the bigger picture, e.g. in terms of standardization efforts to make heterogenious computing finally a reality as well on the desktop.

[Aeder]

It seems the rocm docker hasn't been updated yet.

[tildearrow]

I'm late but unimportant typo:

Phoronix: AMD ROCm 1.9 Available WIth Vega 20 Support & Upstream Kernel Compatibility

[boxie]

right, let's give this thing ago - it's about time I gave the vega10+ryzen1700 combo a good rocm workout

[Soul_keeper]

So what exactely is rocm, besides an ubuntu package ubuntu users download ?
Is this supposed to eventually be part of mesa or the kernel ?
All I really know is that with the latest kernel and mesa alone, there is no opencl support ... If I run an opencl program the system just hardlocks.

[boxie]

So what exactely is rocm, besides an ubuntu package ubuntu users download ?
Is this supposed to eventually be part of mesa or the kernel ?
All I really know is that with the latest kernel and mesa alone, there is no opencl support ... If I run an opencl program the system just hardlocks.

so - ROC = Radeon Open Compute

it is AMD's Open Source GPU compute software

one of the things it provides is OpenCL

[Happy Heyoka]

Ok, so I'm on 18.04 and have a 4.18 kernel installed... should I skip the dkms and do the udev step or do I need to regress to an earlier kernel or??
Thinking about this "Some ROCm features are not available in the upstream KFD" and more specifically "Interoperability between graphics and compute" which sounds like it's hiding a whole day of debugging right there

[TemplarGR]

I am a little confused after reading their github page. When they refer to GFX8 cards they never mention Tonga, only Fiji and Polaris. Vega, GFX9, got the ability to run without pcie atomics, that is, on pcie 2.0 cpus. This ability is not supported on GFX8 cards (yet?). Experimental support for older GFX7 cards does not support or require pcie atomics at all and can work on pcie 2.0 cpus (will they support it in the future?). Where does that leave poor Tonga, the red haired step child of AMD? LOL.

All i want to know is that if at some time in the future my FX processor + R9 380 combo will be able to run ROCm.

[Shevchen]

Its nice to see progress on this somewhat... internal project of AMD. But I have to be honest here: A friend of mine does GPU computing and wanted to give it a go. He is using Arch Linux at work (which he had to fight for but got it eventually) and I send him my RX580 to "test it out". His workplace has 4 Titans for AI computing that is just "plug and play" for the most part. (Of course he has to code the scripts - but thats about it)

There is nothing on the aur that is work-capable, the only thing he found was some kind of fuzzy Ubuntu package without any kind of documentation. After hours of fiddling with this stuff he gave up on it. Sorry - but this isn't going to fly.

Release the thing is a usable manner. As plugin, as lib, as whatever you want but make it work. There is no problem if people have to compile it, but give the package out with a documentation on how to inject ROCm between the CUDA part and the AMD GPU. Thats all.

There are people out there who want to try this stuff out. Don't let it die because of poor support.