I’m running some stellar atmosphere code on a Mac. When the code is compiled with the pgf95 compiler using the flags -mp -Mlarge_arrays -O2 -fast -Mconcur
it takes about 3 hours to run.

When I compile the same code using gfortran and the switches
-O2 -fopenmp -m64
it takes about 5 hours to run.

In both cases I’m running nominally on 4 cores, although the gfortran compilation may not
be fully utilizing them based on what I see using “top”.

I’d appreciate any suggestions. The “-fast” in pgf95 seems to do a lot of good things, but I can’t seem to identify anything analogous in gfortran. Thanks for any advice.

Philip Massey <massey@lowell.edu> schrieb:
> I’m running some stellar atmosphere code on a Mac.

Which type? Apple Silicon or x86? (I'm not sure that Apple
Silicon is officially supported yet).

Which version of gfortran?

> When the code is compiled with the pgf95 compiler using the flags -mp -Mlarge_arrays -O2 -fast -Mconcur
> it takes about 3 hours to run.
>
> When I compile the same code using gfortran and the switches
> -O2 -fopenmp -m64
> it takes about 5 hours to run.

A significant difference, but this is not unheard of.

A few suggestions:

Upgrade to a newer version, if possible (the lastest release
gfortran 13.2).

For time-critical code use at least -O3 -march=native -mtune=native

If you call a lot of procedures from other places, try -flto.
This will expose a lot of optmiization opportunities. It will also
make your compilations much longer, but for your calculation times,
that may still be a win.

You could also try out -ffast-math, but if your code requires any
guarantees on numerical rearrangement or accuracy, then this option
could impact it negatively. You could also try out the indiviual
gcc flags controlling numerics to see if they make a difference.

-Ofast also includes -ffast-math, plus it does some other things;
it can also increase your stack usage, and you could get seemingly
random crashes. This is a known problem from measly maximum
allocation for openmp threads on MacOS.

There are also options for less accurate calculation of square
roots, but if these help (and their large inaccuracy does not
disturb you) this may be worth looking into.

> In both cases I’m running nominally on 4 cores, although the gfortran compilation may not
> be fully utilizing them based on what I see using “top”.

Can't speak to that.

> I’d appreciate any suggestions. The “-fast” in pgf95 seems to do a lot of good things, but I can’t seem to identify anything analogous in gfortran. Thanks for any advice.

Well, -ffast-math does some things that are rightly considered
bad in some use cases, and are considered good (increases speed
and does not negatively impact correctness) in others. I would
not recommend using it without some thought.

On Tuesday, October 24, 2023 at 10:08:45 AM UTC-7, Thomas Koenig wrote:
> Philip Massey <mas...@lowell.edu> schrieb:
> > I’m running some stellar atmosphere code on a Mac.
> Which type? Apple Silicon or x86? (I'm not sure that Apple
> Silicon is officially supported yet).

I'm currently running on an intel machine but hope to implement it on an M2 when gfortran for apple silicon is a little
more mature.
>
> Which version of gfortran?
> > When the code is compiled with the pgf95 compiler using the flags -mp -Mlarge_arrays -O2 -fast -Mconcur
> > it takes about 3 hours to run.
> >
> > When I compile the same code using gfortran and the switches
> > -O2 -fopenmp -m64
> > it takes about 5 hours to run.
> A significant difference, but this is not unheard of.
>
> A few suggestions:
>
> Upgrade to a newer version, if possible (the lastest release
> gfortran 13.2).

I'm using 13.2.
>
> For time-critical code use at least -O3 -march=native -mtune=native
>
> If you call a lot of procedures from other places, try -flto.
> This will expose a lot of optmiization opportunities. It will also
> make your compilations much longer, but for your calculation times,
> that may still be a win.
>
> You could also try out -ffast-math, but if your code requires any
> guarantees on numerical rearrangement or accuracy, then this option
> could impact it negatively. You could also try out the indiviual
> gcc flags controlling numerics to see if they make a difference.
>
> -Ofast also includes -ffast-math, plus it does some other things;
> it can also increase your stack usage, and you could get seemingly
> random crashes. This is a known problem from measly maximum
> allocation for openmp threads on MacOS.
>
> There are also options for less accurate calculation of square
> roots, but if these help (and their large inaccuracy does not
> disturb you) this may be worth looking into.
> > In both cases I’m running nominally on 4 cores, although the gfortran compilation may not
> > be fully utilizing them based on what I see using “top”..
> Can't speak to that.
> > I’d appreciate any suggestions. The “-fast” in pgf95 seems to do a lot of good things, but I can’t seem to identify anything analogous in gfortran. Thanks for any advice.
> Well, -ffast-math does some things that are rightly considered
> bad in some use cases, and are considered good (increases speed
> and does not negatively impact correctness) in others. I would
> not recommend using it without some thought.

These are all very useful, valuable suggestions, and exactly what I was after. Thanks!