Jan Blechta <[email protected]> writes: > It is OpenBLAS. > > Citing from https://github.com/xianyi/OpenBLAS/wiki/faq#multi-threaded > > If your application is already multi-threaded, it will conflict with > OpenBLAS multi-threading. > > We must be cautious of this. (Are PETSc devs? CCing Jed Brown.)
I consider threaded BLAS to always be a mistake for Level 1 BLAS (the part that is used for sparse linear algebra). In general, I think it's better to deal with threads at a higher level. But PETSc can't fail when a user configures with threaded BLAS (and we can't even portably determine that it's threaded). BLAS with threads is a disaster and vendors still pretend like it's acceptable to make threading decisions at link time. The right fix is at the interface level, but nobody will do that. I suggest turning off threads always (threads are often a performance mess even when not over-subscribed).
signature.asc
Description: PGP signature
_______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
