Just as a preliminary, make sure you’ve got cmake installed (as root) :

yum install cmake

Regular instructions from the Deep Learning site

The standard instructions are :

git clone https://github.com/Theano/libgpuarray.git
cd libgpuarray
mkdir Build
cd Build
## you can pass -DCMAKE_INSTALL_PREFIX=/path/to/somewhere to install to an alternate location
## Use Debug instead of Release if you are investigating a crash
optirun cmake .. -DCMAKE_BUILD_TYPE=Release 
## if the nvidia drivers were installed under bumblebee :
#cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_CUDA_LIBRARY=/lib64/nvidia-bumblebee/libcuda.so -DCUDA_TOOLKIT_INCLUDE=/usr/local/cuda-6.5/include
make
sudo make install
cd ..

Helpful build tips

What isn’t obvious is that (a) the default location (/usr/local/{include,lib}) isn’t much use for Theano without adding a bunch of command-line options, and (b) cmake will require running under optirun in order for it to see the required OpenCL libraries.

If you have mis-steps doing the initial cmake, it helps to clear out the cmake caches using rm CMakeCache.txt.

The key cmake line for building under a bumblebee set-up was :

optirun cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr
# And to include the CUDA install too :
optirun cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr -DCUDA_CUDA_LIBRARY=/lib64/nvidia-bumblebee/libcuda.so -DCUDA_TOOLKIT_INCLUDE=/usr/local/cuda-6.5/include

And finally, install system-wide :

make
sudo make install
[ 50%] Built target gpuarray
[100%] Built target gpuarray-static
Install the project...
-- Install configuration: "Release"
-- Installing: /usr/include/gpuarray/array.h
-- Installing: /usr/include/gpuarray/blas.h
-- Installing: /usr/include/gpuarray/buffer.h
-- Installing: /usr/include/gpuarray/buffer_blas.h
-- Installing: /usr/include/gpuarray/config.h
-- Installing: /usr/include/gpuarray/error.h
-- Installing: /usr/include/gpuarray/extension.h
-- Installing: /usr/include/gpuarray/ext_cuda.h
-- Installing: /usr/include/gpuarray/kernel.h
-- Installing: /usr/include/gpuarray/types.h
-- Installing: /usr/include/gpuarray/util.h
-- Installing: /usr/lib/libgpuarray.so
-- Installing: /usr/lib/libgpuarray-static.a
[root@changi Build]# 

Building PyGPU

NB: If you’re running in a virtualenv, remember to set ‘env’ before the following!

# This must be done after libgpuarray is installed as per instructions above.
cd <main-libgpuarray-directory>
python setup.py build
python setup.py install

# Test it works : 
optirun python -c "import pygpu;pygpu.test()"

Building PyGPU (for developing it)

Apparently, this is just done by repeatedly doing build and install above (there’s no python setup.py develop).

Building PyGPU - but tests fail (after a while)

Apparently, this is to-be-expected behaviour at the moment…

optirun python -c "import pygpu;pygpu.test()"
GpuArrayException: Out of resources
"""opencl0:0"""

But at least it highlights the correct name for the GPU device.

Quick manual test of PyGPU

>>> import pygpu
>>> help(pygpu)
>>> pygpu.init('asdasd')
# FAIL...
>>> pygpu.init('opencl0:0')
# Quiet SUCCESS...


Martin Andrews

{Finance, Software, AI} entrepreneur, living in Singapore with my family.



blog comments powered by Disqus