If you read different types of manuals how to compile OpenCL software on Linux, then you can get dizzy of all the LD-parameters. Also when installing the SDKs from AMD, Intel and NVIDIA, you get different locations for libraries, header-files, etc. Now GPGPU is old-fashioned and we go for heterogeneous programming, the chances get higher you will have more SDKs on your machine. Also if you want to keep it the way you have, reading this article gives you insight in what the design is after it all. Note that Intel’s drivers don’t give OpenCL support for their GPUs, but CPUs only.
As my mother said when I was young: “actually cleaning up is very simple”. I’m busy creating a PPA for this, but that will take some more time.
First the idea. For developers OpenCL consists of 5 parts:
- GPUs-only: drivers with OpenCL-support
- The OpenCL header-files
- Vendor specific libraries (needed when using -lOpenCL)
- libOpenCL.so -> a special driver
- An installable client driver
Currently GPU-drivers are always OpenCL-capable, so you only need to secure 4 steps. These are discussed below.
Please note that in certain 64-bit distributions there is not lib64, but only ‘lib’ and ‘lib32’. If that is the case for you, you can use the commands that are mentioned with 32-bit.
Header-files
update: A new package “opencl-headers” installs exactly these files for you. Even better: ocl-icd-opencl-dev installs everything for OpenCL 1.2 (problems with OpenCL 2.0 still)
No more export CPPFLAGS=”-I/some_directory/opencl_sdk/include” at last! All SDKs provide the OpenCL 1.1 header-files originated from Khronos (or should).
We only need to put all headers found from the Khronos-webpage in /usr/include/CL/:
cd /usr/include
sudo mkdir CL
cd CL
sudo wget http://www.khronos.org/registry/cl/api/1.2/cl_d3d10.h \
http://www.khronos.org/registry/cl/api/1.2/cl_d3d11.h \
http://www.khronos.org/registry/cl/api/1.2/cl_dx9_media_sharing.h \
http://www.khronos.org/registry/cl/api/1.2/cl_ext.h \
http://www.khronos.org/registry/cl/api/1.2/cl_gl_ext.h \
http://www.khronos.org/registry/cl/api/1.2/cl_gl.h \
http://www.khronos.org/registry/cl/api/1.2/cl.h \
http://www.khronos.org/registry/cl/api/1.2/cl_platform.h \
http://www.khronos.org/registry/cl/api/1.2/opencl.h \
http://www.khronos.org/registry/cl/api/1.2/cl.hpp ;
If you are on mobile, also get EGL:
sudo wget http://www.khronos.org/registry/cl/api/1.2/cl_egl.h
If you want 1.1 headers, do the following:
cd /usr/include
sudo mkdir CL
cd CL
sudo wget http://www.khronos.org/registry/cl/api/1.1/cl_d3d10.h \
http://www.khronos.org/registry/cl/api/1.1/cl_ext.h \
http://www.khronos.org/registry/cl/api/1.1/cl_gl_ext.h \
http://www.khronos.org/registry/cl/api/1.1/cl_gl.h \
http://www.khronos.org/registry/cl/api/1.1/cl.h \
http://www.khronos.org/registry/cl/api/1.1/cl_platform.h \
http://www.khronos.org/registry/cl/api/1.1/opencl.h \
http://www.khronos.org/registry/cl/api/1.1/cl.hpp ;
sudo wget http://www.khronos.org/registry/cl/api/1.1/cl_egl.h
Now you can be sure you have the correct header-files.
Libaries
All vendors have their favourite spot to put their libraries; but actually a “just put your coat where you find a spot” is not the best to do. According to the best answer on stackoverflow, the libraries should be in /usr/local/lib, but since these are shared libraries, Intel has found a good location: /usr/lib/OpenCL/vendors/. There was some discussion about “vendors”, but think of various wrapper-libaries, IBM’s OpenCL Common Runtime, and such. So I agree with their choice.
Intel
update: Intel recently has completely changed the drivers for Linux. They now install in /opt/intel. Best is to copy all files from /opt/intel/opencl-1.2-x.x.xxxxxx to /usr/lib/OpenCL/vendors/intel to keep it orderly. If you choose not to, replace /usr/lib/OpenCL/vendors/intel with /opt/intel/opencl-1.2-x.x.xxxxxx.
The provided rpm can be converted to deb and then works if libnuma1 is installed:
apt-get install libnuma1 alien *.rpm dpkg -i *.deb
Though they’ve put their libraries at a nice spot, they made a little mistake. They put their libOpenCL.so in /usr/lib or /usr/lib64, instead of using a symbolic link. Below I discuss separately all around libOpenCL.so, since this is an important library. You need to copy it to the right directory. For 64 bit:
sudo cp /usr/lib64/libOpenCL.so /usr/lib64/OpenCL/vendors/intel/
For 32 bit systems:
sudo cp /usr/lib/libOpenCL.so /usr/lib/OpenCL/vendors/intel
It is very possible that if you installed another OpenCL SDK later, the library is lost. Not a real problem as explained later, but then you know.
To make the libraries available, I created opencl-vendor-intel.conf in /etc/ld.so.conf.d with the content (64 bit):
[raw]
echo "/usr/lib64/OpenCL/vendors/intel" > /etc/ld.so.conf.d/opencl-vendor-intel.conf
In case you need to have 32-bit libraries too, you can add the location at the end of that file. And for 32 bit systems:
echo "/usr/lib/OpenCL/vendors/intel" > /etc/ld.so.conf.d/opencl-vendor-intel.conf
[/raw]
Then run
ldconfig
to start to using the new LD-location.
AMD
Edit: as suggested by Steffen Moeller in the comments, installing the deb-files in http://wiki.debian.org/ATIStream is easier. Just check if the files are at the right place.
The AMD APP Installer let’s you choose where you want to put the SDK. Just put it somewhere you want the SDK to be. Go to the root of the AMD-APP-SDK and move the lib-directory to /usr/lib(64)/OpenCL/vendors/, for 64 bit systems:
mkdir -p /usr/lib64/OpenCL/vendors/amd/ mv lib/x86_64/* /usr/lib64/OpenCL/vendors/amd/
And for 32 bit systems:
mkdir -p /usr/lib/OpenCL/vendors/amd/ mv lib/x86/* /usr/lib/OpenCL/vendors/amd/
Then we need to add them to ld-config. For 64 bit:
[raw]
echo "/usr/lib64/OpenCL/vendors/amd" > /etc/ld.so.conf.d/opencl-vendor-amd.conf
And for 32 bit systems:
echo "/usr/lib/OpenCL/vendors/amd" > /etc/ld.so.conf.d/opencl-vendor-amd.conf
[/raw]
Then run
ldconfig
NVIDIA
This is somewhat hard. You probably want to use CUDA too, so for that reason we leave the libraries in /usr/local/cuda/lib/ to avoid breaking software. Of course I prefer them to be tidied up under /usr/lib(64)/OpenCL/vendors/ but it is no use to make a symbolic link. Installer can be found here.
Then we need to add them to ld-config, if you haven’t done that. For 64 bit:
[raw]
echo "/usr/local/cuda/lib64" > /etc/ld.so.conf.d/opencl-vendor-nvidia.conf echo "/usr/local/cuda/lib" >> /etc/ld.so.conf.d/opencl-vendor-nvidia.conf
For 32 bit:
echo "/usr/local/cuda/lib" > /etc/ld.so.conf.d/opencl-vendor-nvidia.conf
[/raw]
Then run
ldconfig
LibOpenCL.so
This library handles the selecting of the platforms (the vendors) and providing the correct libraries to the software needing the functionality. It is located under /usr/lib or /usr/lib64. You need to select which vendor you want to use. I personally think this driver should be open sourced and not from a specific vendor. Pick one (first line 64, second 32) out of these 6. But… from my own experience both AMD and Intel give you versions that work best with all 3 platforms, so I suggest you go for one of these.
Khronos open source
Get the “OpenCL 1.2 Installable Client Driver (ICD) Loader” from http://www.khronos.org/registry/cl/ and make the project (needs cmake). In the bin-directory there will be libOpenCL.so.1.2. Remove all files startting with libOpenCL.so* and copy libOpenCL.so.1.2 to /usr/lib/.
sudo ln -s /usr/lib64/libOpenCL.so /usr/lib64/libOpenCL.so.1.2 sudo ln -s /usr/lib/libOpenCL.so /usr/lib/libOpenCL.so.1.2
I use this myself. Will add the binaries later.
AMD
sudo ln -s /usr/lib64/OpenCL/vendors/amd/libOpenCL.so /usr/lib64/libOpenCL.so.1.2 sudo ln -s /usr/lib/OpenCL/vendors/amd/libOpenCL.so /usr/lib/libOpenCL.so.1.2
NVIDIA
Strongly discouraged to use this libOpenCL-library!
sudo ln -s /usr/local/cuda/lib64/libOpenCL.so /usr/lib64/libOpenCL.so.1.1 sudo ln -s /usr/local/cuda/lib/libOpenCL.so /usr/lib/libOpenCL.so.1.1
Intel
sudo ln -s /usr/lib64/OpenCL/vendors/intel/libOpenCL.so /usr/lib64/libOpenCL.so.1.2 sudo ln -s /usr/lib/OpenCL/vendors/intel/libOpenCL.so /usr/lib/libOpenCL.so.1.2
Then we add libOpenCL.so.1, libOpenCL.so.1.0 and libOpenCL.so:
sudo ln -s /usr/lib64/libOpenCL.so.1.2 /usr/lib64/libOpenCL.so.1 sudo ln -s /usr/lib64/libOpenCL.so.1 /usr/lib64/libOpenCL.sosudo ln -s /usr/lib/libOpenCL.so.1.2 /usr/lib/libOpenCL.so.1 sudo ln -s /usr/lib/libOpenCL.so.1 /usr/lib/libOpenCL.so
As libOpenCL.so.1.2 is/should be backwards compatible with libOpenCL.so.1.0 and libOpenCL.so.1.1, you can choose to make those symbolic links too. Only do this when you have wrongly linked software – link to libOpenCL.so.1 in your own software.
Be sure to link to libOpenCL.so.1.1 if you chose to use NVidia’s library.
Installable Client Drivers
important: If you have chosen to leave the files in the original locations and skipped most of this tutorial, be sure to put the whole path and not only the filename in the icd-files.
If you list the platforms available, you actually list the ICDs. If you have written your own compiler, then you can easily add it without interfering with others. Like you can access an Intel CPU via both the AMD-ICD and Intel-ICD.
In /etc/OpenCL/vendor/ all ICDs need to be put. You’ll find them already here, or you have to create them. This is how they are provided now, but I omitted the library-location (which was in nvidia.icd), since it still gives errors if the ldconfig-steps where not done correctly.
[raw]
sudo echo "libatiocl64.so" > /etc/OpenCL/vendors/atiocl64.icd sudo echo "libatiocl32.so" > /etc/OpenCL/vendors/atiocl32.icd sudo echo "libintelocl.so" > /etc/OpenCL/vendors/intelocl.icd sudo echo "libcuda.so" > /etc/OpenCL/vendors/nvidia.icd
[/raw]
You can pick any name for the icd-files. AMD might replace ‘ati’ by ‘amd’ in their libraries, so if it stops working when updating, you know where to look.
Back to programming
When compiling a C or C++ program, you can keep your makefile simple. When the PPA with all this gets available, I’ll let you know via Twitter, Facebook. Tweet or like, if you support this blog!
Hello, many thanks for this neat overview. Pointers I feel to be missing are to
http://wiki.debian.org/ATIStream
and from there to close-to-AMD .deb packages announed on (they work)
http://forums.amd.com/devforum/messageview.cfm?catid=390&threadid=125792&highlight_key=y&keyword1=debian
Hey, what’s the status of your ppa? I thought of packaging the khronos C++ wrapper, but then I saw that you already started.
I can use extra hands and extra ideas! Please contact me directly via http://www.streamhpc.com/about-us/contact/ and I’ll add you to the PPA.
Hi, I’m going this page on how to install OpenCL and i notice this error
For 64bit : sudo cp /usr/lib/libOpenCL.so /usr/lib/OpenCL/vendors/intel/
For 32bit : sudo cp /usr/lib64/libOpenCL.so /usr/lib64/OpenCL/vendors/intel/
should it be lib64 or lib for 32bit?
Thanks for noticing the typo. It’s fixed now.
Hi, thanks for the article. I can install OpenCL fine on Ubuntu 12.10 (I am installing the Intel OpenCL SDK). The code compiles, but it can’t find any platforms or devices at runtime. Have you seen this before and do you perhaps know what could be wrong? Would love to get some code running on Ubuntu, just CPU would be fine for now! 🙂
The latest version of Intel’s SDK is a mess. Probably /etc/OpenCL/vendors/ has no icd-files. You need to do the stuff written under “Installable Client Drivers”, pointing to the location you installed the SDK.
hello, now I meet the same problems as you. my system is Linux Debian and VGA is intel corporation xeon E3-1200 v3/4th Gen Core Processor. Run clinfo , the number of platforms is 0. How did you solve the problem at last, can you help me!!
Hi, I’ve a NVIDIA card and it’s been two days i tried to install OpenCL. It’s just no possible, don’t tell like it is please, it’s just frustrating me so hard. NVIDIA sucks and can go die in hell, I hate it.
Long life to AMD !!!!!
You got quite some frustration, sorry to hear! It is possible to have OpenCL 1.1 on NVIDIA, so try to get some help from the OpenCL forum on Khronos.
NVIDIA’s decision to leave OpenCL in favour of their own proprietary CUDA, is based on a mid-term profit outlook. But please do not hate them for that.
Pingback: OpenCL First Steps | Richi's Blog
Hi,
I am trying to get a AMD and NVidia GPU to run inside the same Ubuntu 14.04 PC for parallel OpenCL development on both. I had achieved this using Ubuntu 12.04 and the information provided here and your other page (https://streamhpc.com/blog/2011-10-12/both-nvidia-gtx-and-amd-radeon-on-linux/).
Unfortunately what I did back then, does not work anymore (*never* change a running system…). My question is, whether you have achieved this, and if so how. Information on this problem is very scarce. The problem is that the libraries and drivers from both clash, when I try to install both at the same time – individually they work fine. And since I am not too familiar with the underlying OpenCL implemenations, I am a little “in over my head”.
Hope someone can perhaps point me in the right direction.
Thanks and best regards!
Sorry for the late reply. We recently did mixed systems again (for porting CUDA to OpenCL), and most problems were with making sure AMD-drivers were fully cleaned – a full purge, manual removal, etc. I’m thinking of releasing a small script to do the full purge. What was your problem?
A year later lol but if any else is trying this amd and nvidia in one system pretty easy install the nvidia then the amd (dispay) I had opencl working on both as well as cuda you just dont use the config that nvidia wants to make but instead use the amdconfig –initial config file this was done with latest drivers for respective clients, so I had opencl working on amd nvidia gpus and amd cpu. Interesting though now I have no amd gpu and 2 nvidia cards and opencl doesnt work after installing amd app sdk for the cpu prolly have to uninstall the nvidia drivers and reinstall them
Hello,
Great tutorial very informative and straightforward/pretty easy to follow. I am trying to actually get OpenCL drivers running on my cubieboard 4 running 32 bit ARM linaro/ubuntu 14.04 “trusty”. It has an Allwinner A80 big.LITTLE octo core processor with an Imagination “Rouge” PowerVR™ G6230 GPU with two shader clusters and 64 ALU cores. It is certified by Khronos for OpenCL 1.2. But i have not been having any luck with getting OpenCL drivers running. I am very knowledgeable when it comes to hardware but sometimes get stuck with the software/programing side of things. Also i’d say i am pretty well versed in linux. Any help would be great… I did get access to the PowerVR SDK but it seems to just be full of examples and tools only. I did not see any device drivers or OpenCL libraries but could just be looking in wrong directory… Thnx in advance
@disqus_cl9Uovo9PE:disqus Did you ever manage to get this working? I’m trying to get OpenCL running on an A80 as well. Thanks!
Are the “vendor specific libraries” mandatory? Are these just the e. g. APP SDK files or are these the .so files, that the .icd files refer to? If the latter case is true (just .icd files), then the naming seems to be a strange. I think the .so files referred to by the .icd files are actually the ICDs. AFAIK you do not even have these .icd files in Windows, but a registry key instead containing the ICD’s DLL file name.
After reading this (http://stackoverflow.com/questions/29146152/do-i-really-need-an-opencl-sdk?lq=1) I assumed, that I only need the ICD loader (libOpenCL.so), the header files and the ICD installed with the GPU driver installation.
The “ICD” on Linux is a text-file with the path+name of the .so library.
To be honest, that does not answer my question 😉
Are the “vendor specific libraries” mandatory? Are these just the e. g. APP SDK files or are these the .so files, that the .icd files refer to?
Yes, like most SDKs, the so-files are libraries for creating the software and not for running the software. The ICD is for runtime. The filenames in the icd-files are in the vendor’s driver-directory – check by opening the icd-files and locate the filenames mentioned there.
Thanks.
Hi, Thanks very much for the guide. I have a question though. When you say: remove all file starting with “libOpenCl.so*” and copy “libOpenCL.so.1.2” … If I remove all files then also libOpenCL.so.1. will be deleted. Also I have another libOpenCL.so.1.2 inside the folder /usr/lib/OpenCL/vendors/intel/opencl-1.2-6.0.0.1049/lib64. What am I supposed to do?
Thanks very much for the answer