Building An AMD HIP Stack From Upstream Open-Source Code

[phoronix]

Phoronix: Building An AMD HIP Stack From Upstream Open-Source Code

While AMD ships pre-built ROCm/HIP stacks for the major enterprise Linux distributions, if you are using not one of them or just want to be advantageous and compile your own stack for building HIP programs for running on AMD GPUs, one of the AMD Linux developers has written a how-to guide...

https://www.phoronix.com/news/Building-Upstream-HIP-Stack

[cend]

Technical instructions are always welcome! How to compile this so I can train models using unsupported contemporary GPUs?

[zcansi]

> if you are using not one of them or just want to be advantageous

Assuming that was supposed to be "adventurous", although I do want to be advantageous...

[Jabberwocky]

Hat's off to everyone involved!

Now let's push to remove step 4. I'm joking (also not joking). Seriously, I'm always the one bitching and moaning about this so I really appreciate the direction that this is going.

It's going to take me a really long time to test all of it. I'm very excited.

Technical instructions are always welcome! How to compile this so I can train models using unsupported contemporary GPUs?

For starters you can read the article ;-)

[Linuxhippy]

... using the libhsa-runtime64.so library as the lowest-level host-side library of the ROCm stack

I have to admit, I've completly lost track regarding all the attemps of AMD to re-gain a place in the gpu compute space.
So many buzzwords and "products" which were introduced only to disappear after short while with too little to late investment.

But I still remember HSA, which was about queues and shared virtual memory. So HSA is not dead after all?
At least the last news on the HSA foundation website is from 4 years ago: https://hsafoundation.com/category/news/

[agd5f]

But I still remember HSA, which was about queues and shared virtual memory. So HSA is not dead after all?
At least the last news on the HSA foundation website is from 4 years ago: https://hsafoundation.com/category/news/

HSA ~= ROCm. HSA started out as a vendor independent compute infrastructure, but only AMD used it, so eventually it just became ROCm.

[Snaipersky]

Love seeing this kind of thing Michael! Wouldn't complain if you included more instructional posts like this one.

[ktb123]

Is there a reason why most distributions (like ubuntu 23.10) do not build and distribute rocm on their own but instead hope that amd provides the libraries for them?

[cgmb]

Is there a reason why most distributions (like ubuntu 23.10) do not build and distribute rocm on their own but instead hope that amd provides the libraries for them?

Ubuntu 23.10 has HIP in its packages. In fact, it has a version of rocBLAS that has been enabled on far more architectures than the AMD-provided rocBLAS. You should be able to use basic tools like llama-cpp on any discrete AMD GPU from the Vega, RDNA 1, RDNA 2, CDNA 1 or CDNA 2 architectures. The only notable omission is RDNA 3 support, which I'm working hard to try to get packaged before the Ubuntu 24.04 cutoff.

Here's an example of how to build llama-cpp from source with ROCm support on Ubuntu 23.10: https://gist.github.com/cgmb/be113c0...37aa33c3e4ea33

[sophisticles]

As far as I'm concerned the fact that this guide is necessary is one of the glaring shortcomings of Linux use in general.

Linux proponents are always saying how it's open source and you can do anything you want with it, but the reality is that this is not true.

Look at the AMD method for getting ROCm working:

https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html

As long as you are using Ubuntu, Red Hat Enterprise, or SUSE Enterprise Server, it's nice and easy, less than 2 minutes and a quick reboot.

What happens if you are using Solus, OpenMandriva, Manjaro, Slackware, Void or Gentoo?

Sucks to be you, you need to follow the tutorial, which frankly looks like it's a pain in the ass and will probably take you at least an hour, assuming you don't screw something up.

Same thing happens with Intel, it took me a while to figure out how to get Open CL to work with the iGPU on my Ice Lake:

https://www.intel.com/content/www/us/en/developer/articles/tool/opencl-drivers.html#proc-graph-section

https://github.com/intel/compute-runtime

I know people are going to hate this but NVIDIA is much more Linux friendly in this regard:

https://developer.nvidia.com/opencl

You install the official drivers, reboot and done.

And they have a slew of code samples to get you started.

AMD and Intel can really learn a lesson from NVIDIA in this regard.