On 09/09/2016 07:37 PM, Dominik 'Rathann' Mierzejewski wrote:
On Friday, 09 September 2016 at 18:22, Orion Poplawski wrote:
On 09/09/2016 08:24 AM, Jerry James wrote:
There is no soname bump in this release, but one section of the API
changed in a backwards-incompatible way.
If they broke ABI, why wasn't the soname bumped?
If the suitesparse maintainer does not object, I would also like to
fix something I noticed in the build logs. GCC complains about
unrecognized pragmas.
- #pragma ivdep: I propose to change all instances of this to #pragma GCC ivdep.
- #pragma novector: there is no GCC equivalent, so nothing can be done here.
- #pragma omp ...: I propose to build CHOLMOD with -fopenmp so these
will be defined and used.
I'm concerned about this last change - if I understand it correctly everything
that link to CHOLMOD will now need to use -fopenmp as well. I'm not
necessarily opposed to this, but it does have larger ramifications. I know in
various places libraries will provide both serial and openmp versions. I
wonder if it's time for Fedora to work out a scheme for this, or perhaps
simply embrace the multi-core age and accept openmp versions as standard.
I'd be wary against making it default. Thread-safety is still not
universal and sometimes multi-threading makes things slower.
Regards,
Dominik
Many codes can use multithreading support, though one sometimes gets
conflicts. It would be beneficial to have two builds - single threaded
and multithreaded, perhaps with single threaded as default. People can
then choose appropriate package. Multithreaded can lead to improvements,
but is problem dependent. Are there standard flags for multithreaded builds?
--
devel mailing list
devel@lists.fedoraproject.org
https://lists.fedoraproject.org/admin/lists/devel@lists.fedoraproject.org
