Rangsiman Ketkaew

Compiling powerful linear algebra libraries on Linux

Posted by Rangsiman Ketkaew on Jul 29, 2019

Table of Content

Table of Content
Compiling Linear Algebra Libraries
Linear algebra software
Caveat
BLAS
LAPACK
CBLAS
ScaLAPACK
ATLAS
Error while compiling ATLAS
MKL

Compiling Linear Algebra Libraries

This post will show how to compile and install linear algebra libraries such as BLAS and LAPACK on Linux-based operating system step-by-step. These linear algebra mathematical libraries have been generously used in computational science and development for compiling a mid- and top-level scientific programming such as Fortran, C, and C++ programming languages.

Linear algebra software

Basic Linear Algebra Subprogram (BLAS)
Linear Algebra Package (LAPACK)
C/C++ interface to BLAS (CBLAS)
Scalable Linear Algebra PACKage (ScaLAPACK)
Automatically Tuned Linear Algebra Soft (ATLAS)
Intel Math Kernel Library (MKL)

The latest version of each library can be obtained from Netlib Repository: http://www.netlib.org/liblist.html, except ATLAS, which its latest version is available at https://sourceforge.net/projects/math-atlas/files/Stable/ as well as Intel MKL that available only at Intel software suite website.

Caveat

I use GNU Compiler and OpenMPI of writing this post.
I install all libraries in my home directory (/home/rangsiman/).

BLAS

Download a tarball of BLAS release from http://www.netlib.org/blas/
Uncompress the tarball and change the directory

tar -xzvf blas-3.10.0.tgz
cd BLAS-3.10.0

Compile

make
mv blas_LINUX.a libblas.a

LAPACK

Download a tarball of LAPACK release from this website http://www.netlib.org/lapack/.
Uncompress the tarball and change directory

tar-xzvf v3.10.0.tar.gz
cd lapack-3.10.0

Create make.inc by copying from makefile.inc.example, which is a standard one

cp make.inc.example make.inc

Compile and Install using Makefile

make

There should be three libraries in LAPACK directory

liblapack.a
librefblas.a
libtmglib.a

Note: As a low-level BLAS software is part of LAPACK, I could install only LAPACK. No need to install BLAS separately:

# in LAPACK directory, make lapack with blas
make blaslib
make
# create a symbolic link of librefblas.a as libblas.a library
ln -s librefblas.a libblas.a

CBLAS

A libblas.a library is required for compiling CBLAS

Download the tar.gz. file of CBLAS from BLAS forum on Netlib repository to your Linux http://www.netlib.org/blas/index.html#_cblas.
Unpack the tar file

tar -xzvf cblas.tgz

Move to CBLAS directory

cd CBLAS

Edit PATH of libblas.a and name of CBLAS library at line 26 and 27 in Makefile.in file

vi Makefile.in
BLLIB = /home/rangsiman/lapack-3.8.0/libblas.a
CBLIB = ../lib/libcblas.a

Note that a libblas.a static library can be obtained from LAPACK directory, where you installed in previous section.

Install CBLAS using make

make

There should be libcblas.a in CBLAS/lib sub-directory.

ScaLAPACK

Download a tarball of ScaLAPACK from Netlib repository to your Linux http://www.netlib.org/scalapack/#_software
Unpack the tarball

tar -xzvf scalapack-2.0.2.tgz

Move to ScaLAPACK directory

cd scalapack-2.0.2

Copy SLmake.inc.example to SLmake.inc

cp SLmake.inc.example SLmake.inc

Install ScaLAPACK using make

make

This can take several minutes.

There should be libscalapack.a in ScaLAPACK directory.

ATLAS

I am going to install ATLAS under my home directory as a normal user.

Uncompress a tarball

mv tlas3.8.3.tar.gz $HOME/
cd $HOME
tar -xzvf atlas3.8.3.tar.gz

Create new directory (either inside or outside ATLAS directory), for example, ATLAS_BUILD

mkdir $HOME/ATLAS_BUILD

Move to ATLAS_BUILD directory and run configure script (from ATLAS directory) to set up compilation configuration.

cd $HOME/ATLAS_BUILD
../ATLAS/configure

Compile ATLAS

make

You may run compilation using parallel making, says use 8 processors

make -j 8

This step can take several minutes/hours (depending on your computer performance).

Then check library before install

make check
make ptcheck
make time

If having many error occur while checking program, you might ignore those error and proceed next step.

Install software

make install

When the installation is finished (with or without error), please check if CBLAS, LAPACK, and ATLAS are installed at $HOME/ATLAS_INSTALL/lib/.

Error while compiling ATLAS

There is two error that most users frequently face when install ATLAS using make install command.

Cannon create /usr/local/atlas The default setting of target directory of compiled ATLAS library is /usr/local/atlas. If you are not root or sudo, you cannot install anything on global shared space. Thus you should tell make to install ATLAS under your home directory. Refer to a Makefile file at line 3, for example, change

DESTDIR=/usr/local/atlas

DESTDIR=/home/rangsiman/ATLAS_INSTALL

Cannot find some libraries

cp: cannot stat ‘/home/rangsiman/ATLAS_BUILD/lib/libsatlas.dylib’: No such file or directory
make[1]: [install_lib] Error 1 (ignored)
cp /home/rangsiman/ATLAS_BUILD/lib/libtatlas.dylib /home/rangsiman/ATLAS_INSTALL/lib/.
cp: cannot stat ‘/home/rangsiman/ATLAS_BUILD/lib/libtatlas.dylib’: No such file or directory
make[1]: [install_lib] Error 1 (ignored)
cp /home/rangsiman/ATLAS_BUILD/lib/libsatlas.dll /home/rangsiman/ATLAS_INSTALL/lib/.
cp: cannot stat ‘/home/rangsiman/ATLAS_BUILD/lib/libsatlas.dll’: No such file or directory
make[1]: [install_lib] Error 1 (ignored)
cp /home/rangsiman/ATLAS_BUILD/lib/libtatlas.dll /home/rangsiman/ATLAS_INSTALL/lib/.
cp: cannot stat ‘/home/rangsiman/ATLAS_BUILD/lib/libtatlas.dll’: No such file or directory
make[1]: [install_lib] Error 1 (ignored)
cp /home/rangsiman/ATLAS_BUILD/lib/libsatlas.so /home/rangsiman/ATLAS_INSTALL/lib/.
cp: cannot stat ‘/home/rangsiman/ATLAS_BUILD/lib/libsatlas.so’: No such file or directory
make[1]: [install_lib] Error 1 (ignored)
cp /home/rangsiman/ATLAS_BUILD/lib/libtatlas.so /home/rangsiman/ATLAS_INSTALL/lib/.
cp: cannot stat ‘/home/rangsiman/ATLAS_BUILD/lib/libtatlas.so’: No such file or directory
make[1]: [install_lib] Error 1 (ignored)
make[1]: Leaving directory `/home/u7/rangsiman/ATLAS_BUILD'

These error messages can be ignored as long as the static libraries of linear algebra software are build successfully. All library files should be stored at

/home/rangsiman/ATLAS_INSTALL/lib/

MKL

Intel MKL provides a suite of ready-to-use math libraries including optimized linear algebra and Fast Fourier Transforms (FFT). Qualified student or educator and open-source software developer can obtain Intel MKL free of charge from https://software.intel.com/en-us/mkl.

The feature of MKL includes