[atlas-devel] 3.9.42

[Clint W. <wh...@cs...>]

Guys,

I have released 3.9.42.  It has a number of bugfixes, as detailed below.

I am mainly working on stabilizing things for 3.10, so the main new 
development I'm working on is improving ATLAS's ability to auto-test and
time.  On the testing front, I have added a directory EXtest/, where
I'm going to put in a series of one-off testers that are written to
look for particular bugs.  As time goes on, this directory should fill
up with testers written by myself and ATLAS users that test corner cases
that the standard testers aren't getting at.  Right now, it just has
some memory alignment testers for the L2BLAS.  If you have tests to
contribute, please send them in, including simple things like scripts
that call the normal ATLAS testers in ways that have found bugs in the past..

The main new thing is the results/ directory.  I have added the capability
for ATLAS to automatically time ATLAS versus the system BLAS (or a previous
release of ATLAS, to catch performance regressions/improvements), and to
generate bar charts with the comparison using ploticus:
   http://ploticus.sourceforge.net/doc/welcome.html

You can also easily extend this framework for arbitrary benchmarking/charting
using the simple tvec tools I have provided in this directory.  A rough
overview of these new tools and charting capabilities is provided in a
chapter of ATLAS/doc/atlas_devel.pdf.

As usual, you can use the ATLAS framework to measure other BLAS/LAPACK.
I hope that if you develop cool measures and charting capabilities with
these tools, you'll share them!

The bad news is that this makes the comparison very easy, and right now
both ACML and MKL *crush* me on parallel factorization performance.
Part of this is that they have better GEMMs, part of it is that they
use persistent worker threads, but there is clearly more than this
to it.  Their numbers jump around a lot, but at their best they can 
pretty much double my parallel performance (on a platform where they
are only 25% faster in serial).  Unless I find a problem in ATLAS
artificially depressing my current performance, it will undoubtedly
be a while before I can close such an impressive gap.

Let me know if you find the new charting capability helpful.

Cheers,
Clint

ATLAS 3.9.42 released 06/22/11, changes from 3.9.41:
   * Added ability to autobuild performance charts in results/
   * Added EXtest/ and all-aligment testing for GER and GEMV
   * Fixed bug in BETA=0 case of ATL_cgemvN_8x4_sse3.c
   * Added results/ directory that can autobuild performance charts
   * numerous fixes to qrtest and some fixes for the QR fact routines
   * Added missing $(F77SYSLIB) in Make.lib's dylib and ptdylib targets
   * Added chapter in atlas_install explaining how to use mmflagsearch
   * Fixed uninitialized memory read caused by copying data I don't reference
     in parallel GEMM.
   * Fixed unitialized memory read in gemvT
   * Changed extendedmodel=2, model=5 from Corei2 to Corei1 in archinfo_x86


**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************