ESMF needs -fp-model=precise

[anton-seaice]

Per COSIMA/access-om3#40 (comment) the ESMF dependency needs to be built with -fp-model=precise to ensure BFB reproducibility.

In the COSIMA spack we had ESMF+debug set, which may have included this flag.

We don't want to always build with debug and +debug has been removed in the config under test. We need to set -fp-model=precise separately for production build. Is this only an option for intel-compilers?

I think we need to update the ESMF package definition to include the fp-model=precise flag?

I am not totally on top of the details, we might need @micaeljtoliveira 's input to be sure of the details.

Ping @CodeGat

[CodeGat]

This might also need @harshula s eyes. Will look into it.

[harshula]

Hi @micaeljtoliveira ,

The access-om3-nuopc SPD currently has: depends_on("parallelio fflags='-qno-opt-dynamic-align -convert big_endian -assume byterecl -ftz -traceback -assume realloc_lhs -fp-model source' cflags='-qno-opt-dynamic-align -fp-model precise -std=gnu99'"). i.e. fflags with fp-model set to source and cflags with fp-model set to precise.

Do you want depends_on("esmf@8.3.0:") to include cflags='-fp-model precise' and fflags='-fp-model precise' ?

[micaeljtoliveira]

In the COSIMA spack we had ESMF+debug set, which may have included this flag.

The debug variant will add the -debug and -g compiler flags, which, by default, will disable optimizations (one has to explicitly add the optimization flags to prevent this)). So in practice, as a side-effect, it does makes the build bit-wise reproducible.

Do you want depends_on("esmf@8.3.0:") to include cflags='-fp-model precise' and fflags='-fp-model precise' ?

If you want to enforce bit-wise reproducibility, yes.

[harshula]

Hi @anton-seaice , Can you please test the spack-packages development branch? It contains: ACCESS-NRI/spack-packages@main...development