R SIG Debian - Nov 2011 - How to used MKL (not revolution-mkl) with Debian packages

Hi folks,

  if you want to use MKL (the fast BLAS I have tested on my Thinkpad T410)
with the R 2.14.0 built as Debian/Ubuntu packages available on CRAN mirror,
the following tricks may works for without some known side-effects (likes
openmp breaking issues), you may try to build your own libblas.so.3gf.0
with following command:


$ gfortran -L/opt/intel/lib/intel64 -liomp5 -L/opt/intel/mkl/lib/intel64
-lmkl_rt -lmkl_core -lmkl_gf_lp64 -lmkl_gnu_thread -lmkl_gf_ilp64 -lgomp
-lpthread -shared -Wl,-soname,libblas.so.3gf -o libblas.so.3gf.

$ sudo mkdir /usr/local/lib/MKL
$ sudo cp libblas.so.3gf.0 /usr/local/lib/MKL/.
$ cd /usr/local/lib/MKL
$ sudo ln -s libblas.so.3gf.0 libblas.so.3gf
$ sudo -s
$ echo "/usr/local/lib/MKL" > /etc/ld.so.conf.d/MKL.conf
$ exit
$ sudo ldconfig

then test your R with some benchmarks, here I am using
"R-benchmark-25.R",
which is available at http://*r*.research.att.com/benchmarks/.

Note: I assume that your MKL is installed under /opt/intel/.
*

*
-- 
Gong-Yi Liao

Department of Statistics
University of Connecticut
215 Glenbrook Road  U4120
Storrs, CT 06269-4120

	[[alternative HTML version deleted]]

(please ignore previous mail) 

Hi folks,

  if you want to use MKL (the fast BLAS I have tested on my Thinkpad
T410)
with the R 2.14.0 built as Debian/Ubuntu packages available on CRAN
mirror,
the following tricks may work for those who want to use MKL without some
known side-effects (likes
openmp breaking issues), you may try to build your own libblas.so.3gf.0
with following command:


$ gfortran -L/opt/intel/lib/intel64 -liomp5 -L/opt/intel/mkl/lib/intel64
-lmkl_rt -lmkl_core -lmkl_gf_lp64 -lmkl_gnu_thread -lmkl_gf_ilp64 -lgomp
-lpthread -shared -Wl,-soname,libblas.so.3gf -o libblas.so.3gf.

$ sudo mkdir /usr/local/lib/MKL
$ sudo cp libblas.so.3gf.0 /usr/local/lib/MKL/.
$ cd /usr/local/lib/MKL
$ sudo ln -s libblas.so.3gf.0 libblas.so.3gf
$ sudo -s
$ echo "/usr/local/lib/MKL" > /etc/ld.so.conf.d/MKL.conf
$ exit
$ sudo ldconfig

then test your R with some benchmarks, here I am using
"R-benchmark-25.R",
which is available at http://r.research.att.com/benchmarks/.

Note: I assume that your MKL is installed under /opt/intel/.

-- 
Gong-Yi Liao

Department of Statistics
University of Connecticut
215 Glenbrook Road  U4120
Storrs, CT 06269-4120

(please ignore previous mail) 

Hi folks,

  if you want to use MKL (the fast BLAS I have tested on my Thinkpad
T410)
with the R 2.14.0 built as Debian/Ubuntu packages available on CRAN
mirror,
the following tricks may work for those who want to use MKL without some
known side-effects (likes
openmp breaking issues), you may try to build your own libblas.so.3gf.0
with following command:


$ gfortran -L/opt/intel/lib/intel64 -liomp5 -L/opt/intel/mkl/lib/intel64
-lmkl_rt -lmkl_core -lmkl_gf_lp64 -lmkl_gnu_thread -lmkl_gf_ilp64 -lgomp
-lpthread -shared -Wl,-soname,libblas.so.3gf -o libblas.so.3gf.

$ sudo mkdir /usr/local/lib/MKL
$ sudo cp libblas.so.3gf.0 /usr/local/lib/MKL/.
$ cd /usr/local/lib/MKL
$ sudo ln -s libblas.so.3gf.0 libblas.so.3gf
$ sudo -s
$ echo "/usr/local/lib/MKL" > /etc/ld.so.conf.d/MKL.conf
$ exit
$ sudo ldconfig

then test your R with some benchmarks, here I am using
"R-benchmark-25.R",
which is available at http://r.research.att.com/benchmarks/.

Note: I assume that your MKL is installed under /opt/intel/.

-- 
Gong-Yi Liao

Department of Statistics
University of Connecticut
215 Glenbrook Road  U4120
Storrs, CT 06269-4120

Hi folks, 

   after testing with iPython (with pylab), I noticed that linking flag 
"-L/opt/intel/lib/intel64 -ligomp5" can caused many problems, please
remove this part of the linking flag. The correct linking flag should
be:
 
$ gfortran -L/opt/intel/mkl/lib/intel64 -lmkl_rt -lmkl_core
-lmkl_gf_lp64 -lmkl_gnu_thread -lmkl_gf_ilp64 -lgomp -lpthread -shared
-Wl,-soname,libblas.so.3gf -o libblas.so.3gf.0

and 

$ gfortran -L/opt/intel/mkl/lib/intel64 -lmkl_rt -lmkl_core
-lmkl_gf_lp64 -lmkl_gnu_thread -lmkl_gf_ilp64 -lgomp -lpthread -shared
-Wl,-soname,liblapack.so.3gf -o liblapack.so.3gf.0





On Wed, 2011-11-02 at 20:24 -0400, Gong-Yi Liao wrote:
> (please ignore previous mail) 
> 
> Hi folks,
> 
>   if you want to use MKL (the fast BLAS I have tested on my Thinkpad
> T410)
> with the R 2.14.0 built as Debian/Ubuntu packages available on CRAN
> mirror,
> the following tricks may work for those who want to use MKL without some
> known side-effects (likes
> openmp breaking issues), you may try to build your own libblas.so.3gf.0
> with following command:
> 
> 
> $ gfortran -L/opt/intel/lib/intel64 -liomp5 -L/opt/intel/mkl/lib/intel64
> -lmkl_rt -lmkl_core -lmkl_gf_lp64 -lmkl_gnu_thread -lmkl_gf_ilp64 -lgomp
> -lpthread -shared -Wl,-soname,libblas.so.3gf -o libblas.so.3gf.
> 
> $ sudo mkdir /usr/local/lib/MKL
> $ sudo cp libblas.so.3gf.0 /usr/local/lib/MKL/.
> $ cd /usr/local/lib/MKL
> $ sudo ln -s libblas.so.3gf.0 libblas.so.3gf
> $ sudo -s
> $ echo "/usr/local/lib/MKL" > /etc/ld.so.conf.d/MKL.conf
> $ exit
> $ sudo ldconfig
> 
> then test your R with some benchmarks, here I am using
> "R-benchmark-25.R",
> which is available at http://r.research.att.com/benchmarks/.
> 
> Note: I assume that your MKL is installed under /opt/intel/.
>

Hey there,
Thanks for providing this information. I've been trying to link R to MKL 
BLAS for about a week, and this is a huge breakthrough! You're great!

Did you also link libRlapack.so to MKL? If so, did you use the same flags?

Thanks much.