Or perhaps add a "noprec" option that does NOT set any precision flags at all, leaving it up to the easyconfig to specify?
On 11/19/19 3:17 PM, Åke Sandgren wrote: > Yeah, I think we have to add an option for this in > toolchains/compiler/inteliccifort.py > > Kenneth, any thoughts on that? > > On 11/19/19 3:12 PM, Loris Bennett wrote: >> Hi Bart, >> >> Thanks for the information. I'll try 'precise'. However, the developer >> of the program I need to compile (Siesta) explicitly says that '-no-ftz' >> should be used, so it certainly would be handy to be able to set that >> option. >> >> Cheers, >> >> Loris >> >> Bart Oldeman <[email protected]> writes: >> >>> Hello Loris, >>> >>> unfortunately it's not so straightforward to easily override a few flags to >>> my knowledge (maybe somebody else knows?) >>> >>> but you can try >>> toolchainopts = {'usempi': True, 'precise': True} >>> which sets -fp-model precise, which may be what you want anyway if you care >>> about denormals. >>> >>> you can also use 'strict': True instead if you want even more FP safety. >>> >>> Cheers, >>> Bart >>> >>> On Tue, 19 Nov 2019 at 08:03, Loris Bennett <[email protected]> >>> wrote: >>> >>> Hi, >>> >>> I want to use the Intel compiler option '-no-ftz' and have set >>> >>> toolchainopts = {'usempi': True, 'optarch': 'm64 -xHost -prec-div >>> -prec-sqrt -qopt-prefetch -no-ftz'} >>> >>> However, in the log I see >>> >>> CFLAGS set to -O2 -m64 -xHost -prec-div -prec-sqrt -qopt-prefetch >>> -no-ftz -ftz -fp-speculation=safe -fp-model source >>> >>> Is this correct? Will the earlier option really win? My expectation >>> would be that the later option would take precedence? >>> >>> Cheers, >>> >>> Loris >>> >>> -- >>> Dr. Loris Bennett (Mr.) >>> ZEDAT, Freie Universität Berlin Email [email protected] > -- Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden Internet: [email protected] Phone: +46 90 7866134 Fax: +46 90-580 14 Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se
