ATLAS - shawfdong/hyades GitHub Wiki
ATLAS (Automatically Tuned Linear Algebra Software) is an open source efficient and full implementation of BLAS APIs for C and Fortran 77. It also implements a a few routines from LAPACK. While its performance often trails that of specialized libraries written for one specific hardware platform, e.g., Intel MKL, it is a large improvement over the generic BLAS available at Netlib.
Download the latest stable release of ATLAS and unpack the tarball:
$ cd /scratch $ wget http://downloads.sourceforge.net/project/math-atlas/Stable/3.10.2/atlas3.10.2.tar.bz2 $ tar xfj atlas3.10.2.tar.bz2
Create a build directory:
$ mkdir ATLAS_Linux_SandyBridge $ cd ATLAS_Linux_SandyBridge
According to :
CPU throttling makes pretty much all timings completely random, and so any ATLAS install will be junk. Therefore, before installing ATLAS, turn off CPU throttling.
Let's check the CPU throttling status on Hyades:
$ cat /proc/acpi/processor/CPU0/info processor id: 0 acpi id: 1 bus mastering control: yes power management: no throttling control: no limit interface: noCPU throttling is disabled. Excelente!
Unload the default modules to ensure that GCC will be used to compile ATLAS:
$ module unload intel_mpi $ module unload intel_compilers
Configure ATLAS 3.10.2:
$ /scratch/ATLAS/configure --help $ /scratch/ATLAS/configure \ -Fa alg -fPIC -b 64 -D c \ -DPentiumCPS=2200 \ --prefix=/pfs/sw/serial/gcc/atlas-3.10.2 \ --with-netlib-lapack-tarfile=/scratch/lapack-3.5.0.tgzNote:
- the frequency of the CPUs in Hyades is 2.2 GHz ()
- I use the compiler flag to generate position-independent code (PIC), suitable for dynamic linking[1]. This is necessary, e.g., if you want to build SciPy with ATLAS.
$ make buildwhich will create the following libraries:
- liblapack.a : The serial LAPACK routines provided by ATLAS, including both Fortran and ATLAS's clapack[2] (which is different from either LAPACKE or Netlib CLAPACK) interfaces.
- libcblas.a : The ANSI C interface to the BLAS.
- libf77blas.a : The Fortran77 interface to the BLAS.
- libptlapack.a : The threaded LAPACK routines provided by ATLAS.
- libptcblas.a : The ANSI C interface to the threaded BLAS.
- libptf77blas.a : The Fortran77 interface to the threaded BLAS.
- libatlas.a : The main ATLAS library, providing low-level routines for all interface libs.
$ make check
Sanity-check the threaded libraries:
$ make ptcheck
Check the performance of the libraries:
$ make time
Reference clock rate=3292Mhz, new rate=2200Mhz
Refrenc : % of clock rate achieved by reference install
Present : % of clock rate achieved by present ATLAS install
single precision double precision
******************************** *******************************
real complex real complex
--------------- --------------- --------------- ---------------
Benchmark Refrenc Present Refrenc Present Refrenc Present Refrenc Present
========= ======= ======= ======= ======= ======= ======= ======= =======
kSelMM 1289.9 1521.7 1188.7 1321.7 686.7 815.2 647.4 747.6
kGenMM 198.2 180.7 198.5 161.2 193.9 221.6 196.0 217.3
kMM_NT 193.7 220.5 195.2 223.8 184.2 217.5 188.5 222.6
kMM_TN 198.5 151.0 197.9 227.0 189.8 211.5 189.5 216.9
BIG_MM 1213.8 1386.0 1241.3 1288.0 652.0 736.0 661.4 752.2
kMV_N 224.3 68.1 438.8 511.6 115.9 42.6 205.8 266.0
kMV_T 224.6 89.0 460.3 595.3 123.2 43.8 211.3 81.0
kGER 148.3 75.8 290.2 159.1 73.3 34.7 144.3 70.8
Install the libraries:
$ make install
ATLAS 3.10.2 is installed at .
To compile a Fortran program and link with the serial BLAS library provided by ATLAS, using [3]:
$ gfortran -o blaspgm.x blaspgm.f -L/pfs/sw/serial/gcc/atlas-3.10.2/lib -lf77blas -latlas
To compile a Fortran program and link with the serial BLAS library provided by ATLAS, using the :
$ ifort -o blaspgm.x blaspgm.f -L/pfs/sw/serial/gcc/atlas-3.10.2/lib -lf77blas -latlas -lgfortran
To compile a C program and link with the serial CBLAS library provided by ATLAS, using :
$ gcc -o cblaspgm.x cblaspgm.c -I/pfs/sw/serial/gcc/atlas-3.10.2/include -L/pfs/sw/serial/gcc/atlas-3.10.2/lib -lcblas -latlas
To compile a C program and link with the serial CBLAS library provided by ATLAS, using the :
$ icc -o cblaspgm.x cblaspgm.c -I/pfs/sw/serial/gcc/atlas-3.10.2/include -L/pfs/sw/serial/gcc/atlas-3.10.2/lib -lcblas -latlas
To compile a Fortran program and link with the serial LAPACK library provided by ATLAS, using :
$ gfortran -o lapackpgm.x lapackpgm.f -L/pfs/sw/serial/gcc/atlas-3.10.2/lib -llapack -lf77blas -lcblas -latlas
To compile a Fortran program and link with the serial LAPACK library provided by ATLAS, using the :
$ ifort -o lapackpgm.x lapackpgm.f -L/pfs/sw/serial/gcc/atlas-3.10.2/lib -llapack -lf77blas -lcblas -latlas -lgfortran
Note:
- The full LAPACK library is created by merging ATLAS and Netlib LAPACK, so linking with the LAPACK library requires both C and Fortran77 BLAS interfaces.
- The threaded ATLAS libraries should be avoided in a cluster environment like Hyades; because the maximum number of threads to use is determined at compile time and one can't vary the number of threads ATLAS uses dynamically[4].
- To facilitate the usage of ATLAS, I've created a module to set up its environment. If you load the module, you can use more concise commands to link with the ATLAS libraries. For example, to compile a C program and link with the serial BLAS library provided by ATLAS, you can now use a much shorter command:
$ module load lapack/s_gcc_ATLAS_3.10.2 $ gcc -o cblaspgm.x cblaspgm.c -lcblas -latlas
BLAS and LAPACK libraries are required for building SciPy. Now we can build SciPy with ATLAS[5]:
$ export BLAS=/pfs/sw/serial/gcc/atlas-3.10.2/lib/libf77blas.a $ export LAPACK=/pfs/sw/serial/gcc/atlas-3.10.2/lib/liblapack.a $ export ATLAS=/pfs/sw/serial/gcc/atlas-3.10.2/lib/libatlas.a $ cd scipy-0.14.0 $ python setup.py build $ python setup.py install