Using BLAS

ACML

The AMD Core Math Library is the preferred BLAS and LAPACK library on the Brazos cluster. We have a separate documentation page here.

OpenBLAS

We have installed OpenBLAS-0.2.11, which is based on GotoBLAS2. This is a multi-architecture library, so one library will run on the various compute nodes of the Brazos cluster.

To use the OpenBLAS library with GNU compilers:

module load gcc
module load openblas

You may choose the single threaded OpenBLAS library, -lopenblas, or the multi-threaded OpenBLAS library (including the system pthreads library), -lopenblasp -lpthread. To link against the OpenBLAS library add the following to your linker command line:

-L$BRAZOS_OPENBLAS_LIB -lopenblas
-or-
-L$BRAZOS_OPENBLAS_LIB -lopenblasp -lpthread

The -lopenblasp OpenBLAS library is multithreaded, but the default environment variable sets OPENBLAS_NUM_THREADS=1. To change this to 8, include the following in your batch script:

export OPENBLAS_NUM_THREADS=8

Be sure to request 8 processes (--ntasks-per-node=8) when you submit your job.

Documentation

General documentation for BLAS is available from Netlib's BLAS site.