ifort 2013 -fp-model changes from 2011

ifort 2013 -fp-model changes from 2011

Portrait de Bassem G.

I have my fortran code solving finite elemnt problems for CFD applications. I'm using ifort since 2010. I was frustrated today to find that test cases I used to solve with no problem are actually not working any more. I spent more than 10 hours today trying to find out why. ifort 2011 works fine, it is 2013 which I have just installed on one of my machines. I had to specifically include -fp-model source in the compiler options in order to get ifort 2013 to work. I also included -fimf-arch-consistency=true . I'm wondering what is the detault -fp-model in 2013 and why you have changed it from that of 2011. I'm also curuis about how much would speciifing both option affact the performance

Thanks!

Bassem Girgis

3 posts / 0 nouveau(x)
Dernière contribution
Reportez-vous à notre Notice d'optimisation pour plus d'informations sur les choix et l'optimisation des performances dans les produits logiciels Intel.
Portrait de Steve Lionel (Intel)

The default did not change - it is still "fast". What is more likely is that your code is sensitive to optimization choices and that a new or different optimization caused results to change.

Do you know that you need -fimf-arch-consistency? You might if you run the program across different processor generations or brands.

Steve
Portrait de TimP (Intel)

-fp-model source will prevent vectorization of sum reductions. It also sets -prec-div -prec-sqrt -noftz. Any of those you could set back to your desired setting; their effect on performance varies with CPU model as well as with your application. If your reason for setting -fp-model source is to get standard treatment of parentheses, -assume protect_parens will do that without disabling other optimizations.
-fimf-arch-consistency would most likely affect performance of math functions. We saw big improvements in ifort 13.0.1 where it no longer reduces performance significantly.

Connectez-vous pour laisser un commentaire.