Hi Roland, Frank, Thanks very much. This was exactly what I was looking for.
Best, Jonah On 15-11-11 04:46 AM, Roland Haas wrote: > Hello all, > > uhh, let me actually read the question. Frank seems to have pointed you > to what I assume (have not checked yet) to be the proper documentation > that shows what can go wrong. Now for how I worked around the issues myself: > > For the symmetric operators branch I use -fp-model precise to the intel > compiler eg on zwicky I use the default zwicky machine defintion file in > simfactory then add "-fp-model precise" to all CFLAGS/CXXFLAGS/F90FLAGS > so that they read: > > C_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise > CXX_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise > F77_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise > F90_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise > > for the gnu compiler on bluewaters I had to also disable some > optimizations (I don't know why): > > C_OPTIMISE_FLAGS = -O3 -fno-expensive-optimizations -fno-ipa-cp-clone > > @Jonah: if you have access to the SXS wiki (say due to Spectre work), > you can find this information on > https://www.black-holes.org/wiki/center_of_mass_conserving_setup > somewhere under the first big table. > > Yours, > Roland > > > On 2015-11-10 22:47, Erik Schnetter wrote: >> Roland, I think you have been using such flags. Which flags exactly did you >> use? >> >> -erik >> >> On Tue, Nov 10, 2015 at 1:45 PM, Frank Loeffler <[email protected]> wrote: >> >>> On Tue, Nov 10, 2015 at 01:38:08PM -0500, Jonah Miller wrote: >>>> I am working on developing a modified stencil for the Einstein Toolkit >>>> and I am discovering that regression tests pass when I use the GNU >>>> compilers but fail when I use the Intel compilers, even with gentle >>>> optimization settings such as -O1. I suspect over-zealous optimization. >>>> Has anyone encountered this sort of problem before? And if so, does >>>> anybody know a good combination of flags that makes the intel compiler >>>> more cautious, without sacrificing too much speed? (For example, >>>> disabling fast math, which seems to be set to active by default.) >>> The default for the Intel compiler is non-exact math. Try some of the >>> options mentioned here: >>> >>> >>> https://software.intel.com/en-us/articles/consistency-of-floating-point-results-using-the-intel-compiler >>> >>> How much that influences effective speed likely depends a lot on your >>> code. >>> >>> Frank >>> >>> >>> _______________________________________________ >>> Users mailing list >>> [email protected] >>> http://lists.einsteintoolkit.org/mailman/listinfo/users >>> >>> >> >> >> >> _______________________________________________ >> Users mailing list >> [email protected] >> http://lists.einsteintoolkit.org/mailman/listinfo/users >> > _______________________________________________ > Users mailing list > [email protected] > http://lists.einsteintoolkit.org/mailman/listinfo/users _______________________________________________ Users mailing list [email protected] http://lists.einsteintoolkit.org/mailman/listinfo/users
