I'm writing this because I believe there is general interest in this subject. I'm not an expert, so errors and mistakes in this post are highly likely. Corrections and improvements are gratefully accepted. Credits for finding these simple but less than obvious procedures go to wizards dipak and Luke Iwanski.
Although Open CL has been supported for quite a while by AMD GPUs, a problem appeared with the introduction of Ubuntu 16.04. If I understand right, prior to Ubuntu 16.04, the Linux kernel supported AMD graphics cards, but then the card support was transferred to the external amdgpu open-source driver. Unfortunately, amdgpu does not support Open CL. Although Open CL is supported by the proprietary driver amdgpu-pro, this driver only supports newer cards, here meaning GPUs implementing the GCN-1.2 architecture or newer.
GCN-1.0 ("old", ~2011): HD 7750, HD 7770, R7 250X, HD 7850, HD 7870, R9 270X, R9 270, R7 370, HD 7950, R9 280, HD 7970, R9 280X;
GCN-1.1 ("old"): HD 7790, R7 260, R7 260X, R7 360, R9 290, R9 390, R9 290X, R9 390X;
GCN-1.2 ("newer", ~2014): R9 285, R9 380, R9 380X...
A quick glance at the web gives the impression that it is difficult to run Open CL on AMD GPUs, but this isn't true. It is actually quite easy, at least for newer GPUs. Here is how it can be done, starting from an up-to-date vanilla Ubuntu 16.04 LTS installation:
sudo apt install opencl-headers
g++ MatrixTranspose.cpp -L/opt/amdgpu-pro/lib/x86_64-linux-gnu/ -lOpenCL
WARNING to relatively inexperienced people like myself: The operations below could trash the system. Make sure there is a backup. Beware that dual-booted Ubuntu 16.04 and 14.04 can sometimes break each other, even if you have separate boot and root partitions!
Running Open CL on old cards is a little bit trickier than on newer cards. For old cards, downgrading to Ubuntu 14.04 is necessary, the kernel of which must be further downgraded to 3.19 (details below). I assume (and have tested the procedure below on) a Ubuntu 14.04 LTS server clean installation. For simplicity, here are the steps in the form of a bash script:
case $1 in
1) # Install Linux kernel 3.19
sudo apt-get install linux-image-3.19.0-79-generic linux-image-extra-3.19.0-79-generic linux-headers-3.19.0-79-generic ;;
2) # Check current kernel
uname -r ;;
3) # Remove it
sudo apt-get remove linux-image-$(uname -r) ;;
4) # Update boot
sudo update-grub ;;
5) # Reboot
sudo reboot ;;
6) # Check that kernel 3.19.0-79-generic is loaded
uname -r ;;
7) # Install build-essential and Open CL 1.2 headers
sudo apt-get install build-essential ocl-icd-opencl-dev opencl-headers ;;
8) # Download AMD drivers
9) # Install the downloaded drivers
sudo dpkg -i fglrx-core_15.302-0ubuntu1_amd64_ub_14.01.deb fglrx_15.302-0ubuntu1_amd64_ub_14.01.deb ;;
10) # Reboot
sudo reboot ;;
11) # Check GPU
12) # Choose MatrixTranspose directory and compile
g++ MatrixTranspose.cpp -L/usr/lib/x86_64-linux-gnu/ -lOpenCL ;;
A few comments: Use by
bash script.sh <step number>
etc. Don't forget to watch the output messages, in particular at step 3. It can happen that another kernel is automatically installed upon removal. Make certain that the Linux kernel is not newer than 3.19 in step 6!
In order to compile the MatrixTranspose program without error messages, a minor change in CLUtil.hpp is necessary. The lines reading
need to be commented out.