A frog around the world

In the numerical computing world, the cornerstones libraries are BLAS and LAPACK. They have been used in most of the numerical software for decades (like Scilab, R, numpy, OpenOffice with calc, etc).

During that time, many implementations appeared to improve the performances taking advantages of clusters, multicore, SEE{1,2,3,4}, various levels of cache...
Between the reference BLAS (refblas) to an optimized one like ATLAS or MKL (Math Kernel Library by Intel - non-free), it is not rare to have a 15 factor.

In Debian, we use by default the reference implementation of BLAS (168 reverse dependencies) and LAPACK (178 reverse dependencies). If the results are usually bad, they are pretty easy to use. What is hard to use, is switch between highly optimized libraries.
For now, the main one in the archive is ATLAS. ATLAS build process will launch many computations to know what will work best on the architecture. Results are usually excellent.

1) Upload of a refactoring of the ATLAS package.
I have been working on this for a while and after 19 uploads into Debian Experimental and I am happy (and kind of relief) to upload into debian unstable the release 3.8.3 of ATLAS.

The new key elements in this release are:

• Package of the release 3.8.3 ... Long overdue
• Much more packages for recent architectures (sse3, core2sse3, etc)
• A simplified maintenance
• Easy to build a custom package: fakeroot debian/rules custom
• Easy upgrade to version 3.9.X when it is stable
• 12 bugs closed in Debian (including 4 RCs)
• 6 bugs closed in Launchpad.
• MMX optimized package removed

Note that, as before, all prebuilt binaries of ATLAS will be always slower than if you built them on the target architecture (but using Debian binary packages will save a few kilograms of Uranium).

And one of most important feature is the capability to switch to any ATLAS implementation.

2) Switch between the different implementation
The problem in Debian (and Ubuntu) was that it was hard to switch between the ref BLAS/LAPACK and the optimized libraries. The user has to play with the LD_LIBRARY_PATH to use the various optimized packages and since there is no convention between the various distribution, the upstream developer has to develop crappy tricks to handle such things.

It is why I implemented the following proposal: Handle different versions of BLAS and LAPACK.

The main idea is to use the update-alternatives system to allow a quick and easy switch. For example:

# update-alternatives --config libblas.so.3gf 
There are 3 choices for the alternative libblas.so.3gf (providing /usr/lib/libblas.so.3gf).

  Selection    Path                                           Priority   Status
------------------------------------------------------------
* 0            /usr/lib/atlas-core2sse3/atlas/libblas.so.3gf   55        auto mode
  1            /usr/lib/atlas-base/atlas/libblas.so.3gf        35        manual mode
  2            /usr/lib/atlas-core2sse3/atlas/libblas.so.3gf   55        manual mode
  3            /usr/lib/libblas/libblas.so.3gf                 10        manual mode

# update-alternatives --config liblapack.so.3gf
There are 3 choices for the alternative liblapack.so.3gf (providing /usr/lib/liblapack.so.3gf).

  Selection    Path                                             Priority   Status
------------------------------------------------------------
* 0            /usr/lib/atlas-core2sse3/atlas/liblapack.so.3gf   55        auto mode
  1            /usr/lib/atlas-base/atlas/liblapack.so.3gf        35        manual mode
  2            /usr/lib/atlas-core2sse3/atlas/liblapack.so.3gf   55        manual mode
  3            /usr/lib/lapack/liblapack.so.3gf                  10        manual mode

Thanks to this, it is just trivial to switch from one to the other...

Conclusion:
I just pushed the changes into Debian unstable for blas, lapack and atlas.
I have been testing a lot these deep modifications and I fixed all the problems that I found. However, in case I missed something, please report a bug...