Installing libclc on Ubuntu 14.04 LTS

libclc is needed for using Clang with OpenCL. See here for more details. Unfortunately it requires LLVM to be 3.7 or higher, while Ubuntu 14.04 only has up to 3.6.

Make sure you don't have LLVM installed before proceeding.To get around this you first have to get LLVM 3.7 by adding this line:

deb http://llvm.org/apt/precise/ llvm-toolchain-precise-3.7 main

 to this file:

/etc/apt/sources.list.d/llvm.list

I used nano. The original instructions using output redirection didn't work for me. You can find the address for other versions of Ubuntu or LLVM here. Now install llvm-3.7 and clang-3.7

sudo apt-get install llvm-3.7 clang-3.7

It will complain that it can't verify the source, but we added the source so go ahead and ignore it. Next get libclc if you don't already have it:

git clone http://llvm.org/git/libclc.git

Go through the readme and use llvm-config-3.7 instead of llvm-config. You may get an error like so:

/usr/bin/ld: cannot find -ledit

In this case just install libedit:

 sudo apt-get install libedit-dev

 However I still got many warnings of the following type:

WARNING: Linking two modules of different data layouts: 'amdgcn--/lib/image/write_image_impl.ll.tahiti.bc' is '' whereas 'llvm-link' is 'e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64'

A quick google search returned nothing. Because the data layout of the first module is empty I'm going to proceed with the assumption that I can ignore this.

Issues when installing GPU Ocelot

Although I encountered all of these problems, most of them are not specific to GPU Ocelot

1. 'PTXLexer' is not a member of 'parser'
As per here, just edit ocelot/.release_build/ptxgrammar.hpp (make sure to use sudo) and comment out line 352.

 2. typedef 'GlobalMap' locally defined but not used
Turn off Werror by using "--no_werr" when running ./build.py

3. /usr/local/lib/libocelot.so: undefined reference to `tigetnum'
As per here, edit  SConscript (after running sudo ./build.py --install) on line 138 to add "-ltinfo" to ocelot_libs before it's used in OcelotConfig

4. /usr/bin/ld: cannot find -lboost_system-mt
According to this, -mt has been removed from Boost. As suggested I made the necessary symbolic links like so:

sudo ln -s /usr/lib/x86_64-linux-gnu/libboost_system.so /usr/lib/x86_64-linux-gnu
/libboost_system-mt.so

How to install OpenCL for Nvidia GPUs on Ubuntu

These instructions are for installing the Ubuntu packages that contain the necessary drivers and libraries for Nvidia and OpenCL. I prefer doing it this way over the ones provided by Nvidia because it makes installation and/or cleanup a lot easier. However, you won't end up with the latest driver and toolkit. For reference I did this with Ubuntu server release14.04 LTS and an Nvidia GTX 780.

Step 1: Get the Nvidia driver

Use the following command to have the packaged driver for your GPU listed. Use the non-updates versions.

sudo ubuntu-drivers devices

Then install the driver:

sudo apt-get install nvidia-<version here>

 It should look like this for example. Reference

sudo apt-get install nvidia-331

Step 2: Get more Nvidia driver packages

Not quite sure what the following packages are for,  but you need them for this to work. Reference

sudo apt-get install nvidia-<version here>-uvm nvidia-opencl-dev nvidia-modprobe

Step 3: Get the toolkit

Easiest step here. I know it says cuda, but it includes OpenCL libraries too:

sudo apt-get install nvidia-cuda-toolkit

getPlatformIDs returns -1001

For me this was resolved by applying step 2, but I saw people getting this error for other reasons. See here and here

Next steps

Run nvidia-smi to confirm that the driver was installed correctly. Next build and run the deviceQuery sample to ensure you can build and run OpenCL programs successfully.