This guide will show you how to set up your fresh Ubuntu 20.4 LTS OS to be ready to run TensorFlow projects, using ROCm to take advantage of the power of your RX580 graphics card (or any gfx803) in a tested, easy and fast way. The majority of the content is a resume of the official guide by AMD and the unofficial guide by Mathieu Poliquin. I highly recommend to check those links, in order to understand what you are doing and why.

In this guide we stick to ROCm 3.5.1, and there are two reasons for this:

1. There are some bugs in posterior versions which haven't been fixed at all (bugs: 1, 2)
2. AMD dropped official support for the RX580 (3)

So... lets get started!

You can skip this step if you haven't installed ROCm yet

It's preferrable to do a fresh Ubuntu reinstall instead of removing ROCm, but if for some reason you still need to remove ROCm then follow the steps bellow:

1. Run sudo apt autoremove rocm-dkms
2. Make sure that all packages are removed under /opt/rocm-x
3. Remove any other additional packages you installed explicitly with sudo apt remove --purge. To find relevant packages, you can use sudo dpkg -l | grep -E 'hsa|hip|llvm|rocm|rock'.
4. Reboot the system

ROCm requieres the kernel version 5.4 (you can check your currently running kernel version with uname -r).

1. Install the requiered kernel: sudo apt install --install-recommends linux-generic
2. Reboot your computer. In the GRUB menu choose "Additional options for Ubuntu" and select "Boot with kernel 5.4.0-x-generic"
3. Remove the other kernels: sudo apt remove --purge linux-generic-hwe-20.04 linux-oem-20.04 linux-hwe-* linux-oem-* linux-modules-5.1* linux-modules-5.8.0-* linux-modules-5.6.0-*

• Check your progress: You should see something like the following as the output of uname -r:

5.4.0-x-generic

sudo apt update
sudo apt dist-upgrade
sudo apt install libnuma-dev
sudo reboot

Add the repo and install rocm-dkms:

wget -q -O - https://repo.radeon.com/rocm/rocm.gpg.key | sudo apt-key add -
echo 'deb [arch=amd64] http://repo.radeon.com/rocm/apt/3.5.1/ xenial main' | sudo tee /etc/apt/sources.list.d/rocm.list
sudo apt update
sudo apt install rocm-dkms && sudo reboot

Set user permissions to access video card features:

sudo usermod -a -G video $LOGNAME
sudo usermod -a -G render $LOGNAME
sudo reboot

Add rocm to PATH:

echo 'export PATH=$PATH:/opt/rocm/bin:/opt/rocm/profiler/bin:/opt/rocm/opencl/bin' | sudo tee -a /etc/profile.d/rocm.sh
sudo ldconfig
sudo reboot

• Check your progress: At this point you should be able to enter rocminfo into a terminal without getting any error. You should also see your video card name in the command output (something like this):

Name: gfx803
Uuid: GPU-XX
Marketing Name: Ellesmere [Radeon RX 470/480/570/570X/580/580X/590]

You are half the way now!

sudo apt install python3 python3-pip
sudo apt install rocm-libs miopen-hip
pip3 install -Iv tensorflow-rocm==2.2.0
sudo apt install rccl
sudo apt install libtinfo5
sudo reboot

• Check your progress: At this point you should be able to import Tensorflow in Python, make a simple operation, and exit without any error. Try the following in a Python interactive session:

import tensorflow as tf
tf.add(2,5)
exit()

In case everything is OK, you should see something like this:

<tf.Tensor: shape=(), dtype=int32, numpy=7>

Otherwise, you may get an error similar to this:

Could not load dynamic library 'libhip_hcc.so'; dlerror: libhip_hcc.so: cannot open shared object file: No such file or directory

In that case, here is a workaroud extracted from this link. Don't forget make the change permanent if it works for you:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/hip/lib
sudo ldconfig

Congrats! Your machine is now ready to use tensorflow-rocm! You should still consider testing it with something more complete, like a benchmark.

sudo apt install git
git clone https://github.com/tensorflow/benchmarks
cd ./benchmarks/scripts/tf_cnn_benchmarks
python3 tf_cnn_benchmarks.py --num_gpus=1 --batch_size=32 --model=resnet50

Expect it to take some time (5-10 minutes). If you see something like the output bellow (numbers may vary a lot, someone got 36.56) then go and get yourself a taco 🌮, you did it!

total images/sec: 87.92

Have fun! :D

HIP allows developers to convert CUDA code to portable C++. The same source code can be compiled to run on NVIDIA or AMD GPUs. HIP code can be developed either on AMD ROCm platform using HIP-Clang compiler, or a CUDA platform with NVCC installed. HIP-Clang is the compiler for compiling HIP programs on AMD platform (Clang as front end, LLVM as its back end).

If you want to compile aplications/samples using HIP or if you are having issues with Clang (not being installed, or being incorrectly installed), then check this link for detailed instructions on installing/building HIP-Clang.