the LAPACKers can certainly discuss this problem next time there is a
meeting. Just to answer quickly, the BLAS directory in LAPACK is not
advertised nor meant to be the whole BLAS. It covers only the BLAS
routines needed by LAPACK. This was a choice made at a point in time. The
compromise is still up to date:
- either you only provide the BLAS you need and then people complains that
you are not providing the full BLAS implementation,
- or you provide the full BLAS for sake of completeness, and then you have
as much people complaining for dead codes.
The initial goal of the BLAS in LAPACK was to satisfied LAPACK needs.
Now it is true that the BLAS provided by LAPACK is almost complete and
whether or not to include the whole BLAS is argueed regularly.
I am sure the LAPACKers can discuss this item once more time :). We let
you know the answer of the group.
Thanks for rising the issue,
Just another quick point: I hope you are aware of the sacrifice in term of
performance you are making by using reference BLAS for improving
portability, ease of installation, etc. This is really not a trivial
choice you are making. Ideally, you should link with any optimized BLAS
implementaion beneath. I do not know exactly what is the situation with
the Goto BLAS, ATLAS BLAS, ESSL, ACML, MKL BLAS, but that would be much
nicer if there is some kind of way for you to link against them instead of
our not optimized reference implementation.
Just raising the issue, not providing solution though.
On Sat, 22 Jul 2006, Van Dam, HJJ \(Huub\) wrote:
I have recently downloaded and installed LAPACK to be used in the
context of a Python installation with the NumPy and SciPy extensions.
Due to various problems with the installation of Python and its
extensions I decided to compromise on performance and compile everything
from source so that I would have the best control over how things were
compiled to avoid getting bitten by issues such as 32-bit vs. 64-bit
integers, no-underscores through to double underscores on subroutine
names, etc. So I compiled and used the BLAS reference implementation
that is supplied with LAPACK. However when I came to link SciPy I
discovered there were a few BLAS routines missing, they are:
These are available from the BLAS library provided by netlib
http://www.netlib.org/blas . Could you add these to the LAPACK BLAS
tar-ball, please? The Python setup I mentioned is rather popular where I
work therefore I am sure we'll hit this problem again, so if you could
fix it it would surely help.
Huub van Dam email: h.j.j.vandam@Domain.Removed
CCLRC Daresbury Laboratory phone: +44-1925-603933
Daresbury, Warrington fax: +44-1925-603634
Lapack mailing list