Bugzilla Bug 201465

mkl-10.0.011 is out (version bump)

Reproducible: Always

Steps to Reproduce:
N/A



I am attaching an ebuild modified from the 9.x series. Note that there seems to
be something wrong with the eselect modules since the MKL libs don't show up
when calling, for example, eselect cblas list.

Created an attachment (id=137864) [details]
mkl-10.0.011.ebuild

... ok, my bad, I had forgotten to call `ebuild mkl-10.0.011.ebuild qmerge`.
So, the ebuild seems to be working correctly, it only needs some cleansing of
commented "patches" and whatnot.

I don't think it's that easy. The docs said they merged all the cluster stuff
into the main package, but that doesn't appear to be reflected in these
changes.

True, I am working on getting the clustering stuff to build too...which also
implies I am working on the latest OpenMPI ebuild...and the installing ifc..
anyways, a cascade of ebuilds. Will post my results when I get it seemingly
working here. For the moment, MKL seems to install correctly as long as the
fftw USE flag is not set.

Please note, that MKL 10.0.1.014 is now the latest released version.

BR,
/Adam

Created an attachment (id=138788) [details]
MKL 10.0.1.014

This ebuild works but there is still some cleaning up to do. The "profile"
concept from 9.x seems to have been dropped since 10.x series introduced a
common interface to all functions (paraphrasing the Readme(s) here).

The ebuild works but is not "clean". What needs to be done:

- Strip out now useless "profile" selection (also verify this)
- Fix up improper use/selection of Fortran compiler.
- Add USE flags for MKL_DOUBLE/MKL_SINGLE selection when compiling interfaces.
-...more...

lib directory is missing:

NOTE the present ebuild is broken, the eselect module is incorrect and the lib
directory structure has changed to a point that the mkl_install_lib() function
needs to be re-written almost entirely.

*** Bug 210718 has been marked as a duplicate of this bug. ***

Hi,

I finally took the time to work on the new mkl. As usual, linking and packaging
changed a lot upstream. The docs are confusing about their new layering system.

Anyway, mkl-10* is now in the science overlay, available with layman. Please
test it and report here. I am hoping to push it in the main tree in a few days
after a bit more polishing. 

Thanks

Hi,

Thanks for taking the time to work on this ebuild. It almost works like a charm
with USE="doc fftw int64 -debug -fortran95 -mpi" on x86_64. A minor remark:

It gives this message when multiple MKL licenses are found. However, after
finishing the emerge, the library works correctly.

>>> Unpacking l_mkl_p_10.0.2.018.tgz to /var/tmp/portage/sci-libs/mkl-10.0.2.018/work
basename: extra operand
`/opt/intel/licenses/1974025126_NCOM_L_MKL__NCD9-8BJGW3CF.lic'
Try `basename --help' for more information.

Furthermore, it would be nice if multiple slots can be used, since there are a
lot of changes in required libraries when linking against 9.1 or 10.0. (just a
reminder for this package)

(
> >>> Unpacking l_mkl_p_10.0.2.018.tgz to /var/tmp/portage/sci-libs/mkl-10.0.2.018/work
> basename: extra operand
> `/opt/intel/licenses/1974025126_NCOM_L_MKL__NCD9-8BJGW3CF.lic'
> Try `basename --help' for more information.

fixed.

> Furthermore, it would be nice if multiple slots can be used, since there are a
> lot of changes in required libraries when linking against 9.1 or 10.0. (just a
> reminder for this package)

Yes I know, I will postpone for the next revision with probably an eselect
framework for all intel packages.

One last thing: mkl linking did not work if I keep LDFLAGS="-Wl,--as-needed" on
packages. Especially with the acx_blas.m4 and acx_lapack.m4 macros that are
used through a lot of packages.
Since we can't patch mkl and can't force unsetting the LDFLAGS, I could not com
e up with a better idea than adding a -WL,--no-as-needed in the pkg-config
files. (I might do the same for acml).

mkl-10.0.2.018 now in cvs. Thanks all for testing.

For the record, mkl fails to build even though I have tried with either
gcc-4.2.2 or gcc-4.3 as active compiler (which implies gfortran-4.2.2 or
gfortran-4.3.0)

The error message:
None of the needed Fortran Compilers (ifc) is installed.
To install one of these, choose one of the following steps:
[1] emerge dev-lang/ifc

Looks like a broken test in the code:

        # blas95 and lapack95 don't compile with gfortran < 4.2
        gcc-version lt 4 2 && FORTRAN="ifc"

that was just fixed already.

Fix confirmed, you're too fast Sébastien ;)