The ROCm Enablement Tool Makes It Easier To Setup AMD's Open-Source Compute Stack

[gsedej]

It's one of the reasons I ended up running Tensorflow on a CPU instead. The second issue is that they haven't even upstreamed their patches to tensorflow so you have to use their fork that, at least back then, was outdated, and didn't play nice with newer libraries.

What issues did you have?
I was trying different TensorFlow projects and it seemed it worked. Running TF on CPU is really slow, at least for networks that I use.

[Marc Driftmeyer]

The project is just too large and fragmented. There is too much redundancy. Various projects that are already installed on people's systems are avoided in favor of internal copies.
And good luck trying to find which svn address or version needs to be used between the various subprojects. ROCm needs to be a single unified project with an actual build system instead of dozens of shell scripts. And please follow the gnu/linux philosophy, upstream the stuff to llvm or any other project instead of duplicating projects. When every distro can create packages easily (ie: installing to bin, lib, include dirs without hacking the thing to work), then maybe it'll be a respectable project.

Agreed, and they should start by making all their Clang/LLVM bits upstreamed into LLVM Project Proper. I sure as hell am not gutting my Debian system for it.

[mannerov]

I agree with all the comments. Even installing an Ubuntu partition and testing ROCm on it, it was not easy, and in the end somehow it didn't work (tensorflow tests failed).

[vegabook]

Honestly I have tensorflow 1.13 working very well on Ubuntu 19.04 and it's just a few "apt"s away after plugging in the ppi. Then it's pip install tensorflow-rocm. And 280+ on Resnet50 with batch size 128, versus 7 on 2700x. Well above 2080 and getting close to 2080ti, before we even starting talking about fp64.

I don't know what all the moaning is about.

https://rocm.github.io/tensorflow.html

https://pypi.org/project/tensorflow-rocm/

[Aeder]

What issues did you have?
I was trying different TensorFlow projects and it seemed it worked. Running TF on CPU is really slow, at least for networks that I use.

it wouldn't work with librosa/keras/pandas because of conflicts, i had to use outdated versions of everything. But I checked now and it seems to be at the same version as upstream stable.

The other problem was installing rocm in any distro not called ubuntu/centos/rhel with ease. I could install the rpms in Fedora but then they would just segfault when running samples.

This was a while back and I haven't tried again since Rocm 1.8

[gsedej]

Honestly I have tensorflow 1.13 working very well on Ubuntu 19.04 and it's just a few "apt"s away after plugging in the ppi. Then it's pip install tensorflow-rocm. And 280+ on Resnet50 with batch size 128, versus 7 on 2700x. Well above 2080 and getting close to 2080ti, before we even starting talking about fp64.

I don't know what all the moaning is about.

https://rocm.github.io/tensorflow.html

https://pypi.org/project/tensorflow-rocm/

Nice results. What card did you test?
Btw, my old tests of caffe-cpu: Ryzen 1600x vs i7 6700 were much worse for ryzen. IIRC cpu implementation of MATH libraries (BLAS) was very bad at the time. Don't know how is situation now.

Also different tasks (different networks) perform different. My tests of RX 480 vs RTX 2080 showed that radeon was generally a bit slower. In some cases much slower (10 times) and in some cases just 10%.

I would like to test on some HBM card, but I lack of it...

[microcode]

Cool, this shouldn't be too tough to turn into a PKGBUILD.

[karolherbst]

this RET thing is a joke right?

just some golden highlights from the README:

"Note: it is required to start with a clean system"
"Note: DO NOT update and upgrade your system"

Instead of making it less painful to package and build ROCm, some joke bandage is wrapped around it, which isn't supposed to work unless you do a clean ubuntu installation? Seriously?

At least it's not an official AMD project, that would have been very bad for AMD. Also from a look at the sources, it really seems to be a more or less dangerous tool as it tries to manage installed kernels and all funky stuff you don't want to have to be done by some random tool :/

[Raven3x7]

Cool, this shouldn't be too tough to turn into a PKGBUILD.

There were a couple of attempts to do that. It seems there are some issues with it.

[microcode]

There were a couple of attempts to do that. It seems there are some issues with it.

Yeah, I noticed. At least the whole process is here though, which is a big step up.