Pages

08 February 2013

336. Compiling ATLAS, netblas, lapack and openblas on Arch Linux

Here's another Arch post.

I was a bit surprised to find that there's no ATLAS in the standard Arch repositories (it is in AUR though), so here's how to build some of the more common math libraries for yourself:


ATLAS

pacman -S wget base-devel gcc-fortran cpupower
sudo systemctl enable cpupower

To build ATLAS you should set the governor for your CPU to performance to get the best optimization:

cpupower frequency-set -g performance
sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor

Basically copy the scaling_governor to all cpus (cpu0, cpu1, cpu2 ...) as shown in the last line above. When you set the governor back to e.g. ondemand, follow the same steps.

sudo mkdir /opt/ATLAS
chown ${USER} /opt/ATLAS
mkdir -p ~/tmp/atlas
cd ~/tmp/atlas
wget http://www.netlib.org/lapack/lapack-3.4.2.tgz
wget http://downloads.sourceforge.net/project/math-atlas/Stable/3.10.1/atlas3.10.1.tar.bz2
tar xvf atlas3.10.1.tar.bz2
mkdir build/
cd build/
../ATLAS/./configure --prefix=/opt/ATLAS -Fa alg '-fPIC' --with-netlib-lapack-tarfile=$HOME/tmp/atlas/lapack-3.4.2.tgz --shared
make
make install

Simple as that. You can now change the governor back
cpupower frequency-set -g ondemand
sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor
sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu2/cpufreq/scaling_governor
...


netlib BLAS and lapack
pacman -S wget base-devel gcc-fortran cmake
sudo mkdir /opt/netlib
sudo chown $USER /opt/netlib
mkdir /opt/netlib/blas/lib -p
mkdir -p ~/tmp/blas
cd ~/tmp/blas
wget http://www.netlib.org/blas/blas.tgz
tar xvf blas.tgz
cd BLAS/

Edit make.inc
OPTS = -O3 -shared -m64 -march=native -fPIC
make all
gfortran -shared -Wl,-soname,libnetblas.so -o libblas.so.1.0.1 *.o -lc
ln -s libblas.so.1.0.1 libnetblas.so
cp lib*blas* /opt/netlib/blas/lib
cd ../
wget http://www.netlib.org/lapack/lapack-3.4.2.tgz
tar xvf lapack-3.4.2.tgz
mkdir /opt/netlib/lapack
mkdir build/
cd build/
ccmake ../lapack-3.4.2/ -DCMAKE_INSTALL_PREFIX=/opt/netlib/lapack -DBUILD_SHARED_LIBS=ON -DUSE_OPTIMIZED_BLAS=ON 

Hit c twice to configure, then g to generate.
Edit CMakeCache.txt and add the following lines at the beginning:
########################
# EXTERNAL cache entries
########################
 BLAS_FOUND:STRING=TRUE
 BLAS_GENERIC_FOUND:BOOL=TRUE
 BLAS_GENERIC_blas_LIBRARY:FILEPATH=/opt/netlib/blas/lib/libnetblas.so
 BLAS_LIBRARIES:PATH=/opt/netlib/blas/lib/libnetblas.so

Do
ccmake ../lapack-3.4.2/

again, then hit c once, then g.
Next,
make
make install

Done.

Openblas
Copied from here: http://verahill.blogspot.com.au/2013/02/334-compiling-nwchem-with-openmpi-and.html

Download from http://github.com/xianyi/OpenBLAS/tarball/v0.1.1

pacman -S wget base-devel gcc-fortran
sudo mkdir /opt/openblas
sudo chown $USER /opt/openblas
tar xvf xianyi-OpenBLAS-v0.1.1-0-g5b7f443.tar.gz
cd xianyi-OpenBLAS-e6e87a2/
make all BINARY=64 CC=/usr/bin/gcc FC=/usr/bin/gfortran USE_THREAD=0 INTERFACE64=1 1> make.log 2>make.err
make PREFIX=/opt/openblas install
cp lib*.*  /opt/openblas/lib


4 comments:

  1. Regarding openblas, I followed these instructions but when I type sudo update-alternatives --config libblas.so.3 there is no sign of the new openblas. I.e. my system is still using the reference blas. Any help?

    ReplyDelete
    Replies
    1. Not sure what you mean. If you compile your own version of blas it won't automatically show up under update-alternatives

      Delete
  2. I followed your blas compilation and am having some trouble linking to the libraries. I've done this on Windows, so assume it would be similar in g++ (declare the blas prototype as an external function, call the blas function inside your code, and compile while linking the library). For example:

    1. Matrix multiply prototype - extern void dgemm_(char*, char*, int*, int*,int*, double*, double*, int*, double*, int*, double*, double*, int*);

    2. Function call - dgemm_(&t1, &t2, &m, &n, &n, &fac1, A, &m, B, &m, &fac2, C, &m);

    3. Compile command - g++ -Wall main.cpp /opt/netlib/blas/lib/libnetblas.so -o main

    However, g++ is unable to find the blas library: "undefined reference to 'dgemm_ ....... collect2: error: ld returned 1 exit status"

    Am I linking the blas library correctly or is something else wrong? Thanks in advance!

    ReplyDelete
  3. Is it possible that you could make a post about how to compile this for g++ instead of gcc? I can immediatley see that the .inc file for lapack is set to CC=/usr/bin/gcc. I imagine I would have to change this to the g++ compiler, in addition to a whole slew of other changes.

    Also: your blog is SUPER useful for those of us moving from MATLAB to Linux. Thanks a million!

    ReplyDelete