[SciPy-User] RZ Factorization

[Sturla Molden]

Thomas Unterthiner <thomas_unterthiner at web.de> wrote:

> I just went to see for myself: Netlib's LAPACKE does indeed do that even 
> in cases when it isn't needed (e.g. in *potrf). Is there a reason why 
> this is done, or could other implementations (e.g. MKL) potentially have 
> a more intelligent solution?

The Netlib LAPACKE code is written by Intel, and probably similar to the
LAPACKE code in MKL.

Another PITA is that LAPACKE does not cover all of LAPACK. 

Also there is no agreement on a standard C interface for LAPACK. Intel has
LAPACKE, AMD has its own wrappers, Apple has CLAPACK, and Netlib has
CLAPACK and LAPACKE (depending on LAPACK version). I think SciPy's decision
to stick with Fortran LAPACK is the better option here, even though there
is a f2c/g77 vs. gfortran/ifort ABI problem. Fortran 2003 was supposed to
remove this ABI problem, but it does not work here as f2c and g77 cannot
compile Fortran 2003 code. 

A while ago I thought I should just ditch Fortran and use C. But now the C
standards comittee has decided to screw over all numerical C programmers,
and make the numerical capabilites added to C99 optional. I am getting
depressed... :(

Sturla