Difference between revisions of "ATLAS"

From CBLFS
Jump to navigationJump to search
m
Line 7: Line 7:
 
| http://www.netlib.org/blas/blas.tgz (F77 source files only, see below)
 
| http://www.netlib.org/blas/blas.tgz (F77 source files only, see below)
 
|}
 
|}
= Introduction to BLAS =
+
= Introduction to BLAS and LAPACK =
  
 
BLAS are '''B'''asic '''L'''inear '''A'''lgebra '''S'''ubprograms written in Fortran77.  There are Level 1 (scalar and vector operations), Level 2 (scalar-matrix operations) and Level 3 (matrix-matrix operations) subprograms.  They are available at [http://www.netlib.org/blas/ as tarballs] of F77 source files and are un-optimized for any particular architecture.  Use the second download source if you want or need these source files.
 
BLAS are '''B'''asic '''L'''inear '''A'''lgebra '''S'''ubprograms written in Fortran77.  There are Level 1 (scalar and vector operations), Level 2 (scalar-matrix operations) and Level 3 (matrix-matrix operations) subprograms.  They are available at [http://www.netlib.org/blas/ as tarballs] of F77 source files and are un-optimized for any particular architecture.  Use the second download source if you want or need these source files.
  
ATLAS (Automatically Tuned Linear Algebra Software) will provide C and FORTRAN interfaces to a BLAS and [[LAPACK]] implementation optimized for a specific machine architecture.  The ideal source of BLAS and LAPACK libraries are those provided by the microprocesor vendor such as AMD's [[ACML]].  However, these are not provided under a license that allows re-distribution without written consent from the vendor.
+
LAPACK is a '''L'''inear '''A'''lgebra '''PACK'''age that provides routines for solving systems of simultaneous linear equations, least-squares solutions of linear systems of equations, eigenvalue problems, and singular value problems.  LAPACK routines are written so that as much as possible of the computation is performed by calls to the BLAS.  The LAPACK libraries produced by ATLAS are not optimal.  Use the [[LAPACK]] package for optimal LAPACK libraries.
 +
 
 +
ATLAS (Automatically Tuned Linear Algebra Software) will provide C and FORTRAN interfaces to a BLAS and LAPACK implementation optimized for a specific machine architecture.  The ideal source of BLAS and LAPACK libraries are those provided by the microprocesor vendor such as AMD's [[ACML]].  However, these are not provided under a license that allows re-distribution without written consent from the vendor.
  
 
== Dependencies ==
 
== Dependencies ==
Line 23: Line 25:
 
Compile the package:
 
Compile the package:
  
    make
+
  make
  
 
The configuration is interactive.  Simply follow the instructions presented on the screen.  If you've installed the FORTRAN provided by gcc-4.0.3 and later use '''gfortran''' when the configuration complains that the F77 version it finds is unsuitable.
 
The configuration is interactive.  Simply follow the instructions presented on the screen.  If you've installed the FORTRAN provided by gcc-4.0.3 and later use '''gfortran''' when the configuration complains that the F77 version it finds is unsuitable.
Line 29: Line 31:
 
Install the package:
 
Install the package:
  
    make install arch=<TARGET_MAKEFILE>
+
  make install arch=<TARGET> &&
 +
  cp -v lib/<TARGET>/*.a /usr/lib
 +
 
 +
If you are building BLAS libraries for more than one TARGET architecture, either place each set of libraries in a subdirectory of /usr/lib or give each set a unique name so as not to overwrite earlier versions.
 +
 
 +
Install the API and other documentation.
  
Where <TARGET_MAKEFILE> will be displayed at the completion of the previous step.
+
  install -v -m755 -d /usr/share/doc/ATLAS-3.6.0 &&
 +
  for doc in doc/<TARGET>/*.{ps,txt}; do
 +
    cp -v $doc /usr/share/doc/ATLAS-3.6.0
 +
  done
 +
 
 +
Where <TARGET> will be displayed at the completion of the 'make' step.
 +
 
 +
The ATLAS compilation and tuning produces several log files in bin/<TARGET>/INSTALL_LOGS.  These may be of interest or helpful if the BLAS libraries are found not to be optimal and you seek help from the ATLAS maintainers.  Install them as follows:
 +
 
 +
  install -v -m755 -d /usr/share/doc/ATLAS-3.6.0/INSTALL_LOGS &&
 +
  for log in bin/<TARGET>/INSTALL_LOGS/*; do
 +
    cp -v $log /usr/share/doc/ATLAS-3.6.0/INSTALL_LOGS
 +
  done
  
 
== Contents ==
 
== Contents ==
 +
 +
Installed Libraries: libatlas.a, libcblas.a, libf77blas.a, liblapack.a, libtstatlas.a

Revision as of 20:27, 13 December 2006

Download Source: http://www.netlib.org/atlas/atlas3.6.0.tgz
Download Source: http://www.netlib.org/blas/blas.tgz (F77 source files only, see below)

Introduction to BLAS and LAPACK

BLAS are Basic Linear Algebra Subprograms written in Fortran77. There are Level 1 (scalar and vector operations), Level 2 (scalar-matrix operations) and Level 3 (matrix-matrix operations) subprograms. They are available at as tarballs of F77 source files and are un-optimized for any particular architecture. Use the second download source if you want or need these source files.

LAPACK is a Linear Algebra PACKage that provides routines for solving systems of simultaneous linear equations, least-squares solutions of linear systems of equations, eigenvalue problems, and singular value problems. LAPACK routines are written so that as much as possible of the computation is performed by calls to the BLAS. The LAPACK libraries produced by ATLAS are not optimal. Use the LAPACK package for optimal LAPACK libraries.

ATLAS (Automatically Tuned Linear Algebra Software) will provide C and FORTRAN interfaces to a BLAS and LAPACK implementation optimized for a specific machine architecture. The ideal source of BLAS and LAPACK libraries are those provided by the microprocesor vendor such as AMD's ACML. However, these are not provided under a license that allows re-distribution without written consent from the vendor.

Dependencies

Optional

  • gcc-3.X.X Benchmarking by the ATLAS maintainers indicates that, with the exception of Core Duo and Pentium M, compiling ATLAS with gcc-3.X produces better x87 code than gcc-4.X.

Non-Multilib or Multilib

Compile the package:

 make

The configuration is interactive. Simply follow the instructions presented on the screen. If you've installed the FORTRAN provided by gcc-4.0.3 and later use gfortran when the configuration complains that the F77 version it finds is unsuitable.

Install the package:

 make install arch=<TARGET> &&
 cp -v lib/<TARGET>/*.a /usr/lib

If you are building BLAS libraries for more than one TARGET architecture, either place each set of libraries in a subdirectory of /usr/lib or give each set a unique name so as not to overwrite earlier versions.

Install the API and other documentation.

 install -v -m755 -d /usr/share/doc/ATLAS-3.6.0 &&
 for doc in doc/<TARGET>/*.{ps,txt}; do
   cp -v $doc /usr/share/doc/ATLAS-3.6.0
 done

Where <TARGET> will be displayed at the completion of the 'make' step.

The ATLAS compilation and tuning produces several log files in bin/<TARGET>/INSTALL_LOGS. These may be of interest or helpful if the BLAS libraries are found not to be optimal and you seek help from the ATLAS maintainers. Install them as follows:

 install -v -m755 -d /usr/share/doc/ATLAS-3.6.0/INSTALL_LOGS &&
 for log in bin/<TARGET>/INSTALL_LOGS/*; do
   cp -v $log /usr/share/doc/ATLAS-3.6.0/INSTALL_LOGS
 done

Contents

Installed Libraries: libatlas.a, libcblas.a, libf77blas.a, liblapack.a, libtstatlas.a