R devel - May 2012 - R Installation Manual - ATLAS BLAS guidance that is not in the current version

Good afternoon.

I am trying to compile a version of Rblas.dll based on ATLAS for the Corei7. I
had remembered that there was mention of which file to adjust and that
"xerbla" needed to be removed from one of the outputs from the last
time I tried a few years ago. The most recent version of the R Installation
manual does not say anything about this. An older version (2.10 I believe) has
the following text:

      Optionally, you can install a version of ATLAS
      (`math-atlas.sourceforge.net'
      (http://math-atlas.sourceforge.net/)) tuned to your system for
      fast linear algebra routines. Pre-built `Rblas.dll' for various
      Pentium and AthlonXP chips are available in the
      `windows/contrib/ATLAS' area on CRAN.  If you are building R from
      source, there are macros `USE_ATLAS' and `ATLAS_PATH' in the file
      `MkRules'.  Set `USE_ATLAS = YES' and `ATLAS_PATH' to where
the
      ATLAS libraries are located.  You will need to make the libraries
      yourself(1): none of the binaries we have seen are compiled for
      the correct compiler.  Since R has its own `xerbla' it is best to
      delete that in ATLAS by
 
           ar d /path/to/libf77blas.a xerbla.o

Would it be possible to restore the above information to the next version of the
manual for future reference, please?
 
Thank you,

Avraham Adler

On 09/05/2012 19:54, Adler, Avraham wrote:
> Good afternoon.
>
> I am trying to compile a version of Rblas.dll based on ATLAS for the
Corei7. I had remembered that there was mention of which file to adjust
and that "xerbla" needed to be removed from one of the outputs from
the
last time I tried a few years ago. The most recent version of the R
Installation manual does not say anything about this. An older version
(2.10 I believe) has the following text:
>
>        Optionally, you can install a version of ATLAS
>        (`math-atlas.sourceforge.net'
>        (http://math-atlas.sourceforge.net/)) tuned to your system for
>        fast linear algebra routines. Pre-built `Rblas.dll' for various
>        Pentium and AthlonXP chips are available in the
>        `windows/contrib/ATLAS' area on CRAN.  If you are building R
from
>        source, there are macros `USE_ATLAS' and `ATLAS_PATH' in the
file
>        `MkRules'.  Set `USE_ATLAS = YES' and `ATLAS_PATH' to
where the
>        ATLAS libraries are located.  You will need to make the libraries
>        yourself(1): none of the binaries we have seen are compiled for
>        the correct compiler.  Since R has its own `xerbla' it is best
to
>        delete that in ATLAS by
>
>             ar d /path/to/libf77blas.a xerbla.o
>
> Would it be possible to restore the above information to the next version
of the manual for future reference, please?
No, as this is no longer supported (and various attempts to make it work 
have failed with incorrect results/segfaults).

We do support the use of the legacy Goto BLAS these days (64-bit only, 
but anyone who cares about performance will be using 64-bit).

As you have found in another thread, avoid over-optimizing code: the 
risk of incorrect answers is serious and the performance gains small.
> Thank you,
>
> Avraham Adler
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

I apologize for the late response, but I missed Professor Ripley's reply.
For what it is worth, I have successfully compiled a 32-bit ATLAS-based BLAS for
Windows for the Core-i7 that passes the make checks and library tools tests
(which may be found at
<http://lib.stat.cmu.edu/R/CRAN/bin/windows/contrib/ATLAS/C2i7/> for those
interested in either testing or using it). Using this BLAS, I find a speedup of
around 500% for matrix multiplications of 1000x1000 dense matrices.
Unfortunately, in the environment in which I operate, I am currently constrained
to a 32-bit operating system. If and when the transition is made to 64bit, I
will gladly use  Professor Nakama's versions, although none of them are
optimized for Core i7 (Sandy Bridge or otherwise) CPUs.

As the necessity to remove the xerbla remains, and implementing it does allow
for the proper compilation of R under Windows, I will just have to archive that
information myself and use it the next time I need to update the BLAS. I am
thankful that the Windows-based MAKEFILE still allows for the compilation of
Rblas.dll linking to an ATLAS-based BLAS, and although I regret that the manuals
will no longer contain that information, I certainly defer to the expertise and
experience of those much more knowledgeable than I in matters programming,
compilation, and R.

Thank you,

Avraham Adler