[Dune-devel] Explicit template instantiation

[Simon Praetorius]

Hi Carsten

Am 16.02.21 um 20:46 schrieb Carsten Gräser:
>
>> Recently, I have experimented with explicit template instantiation of
>> some grid types leading to a reasonable speedup of compile processes.
>> In a few applications I have compile times for a single target > 5min
>> on a decent machine and was wondering how to to reduce that time.
>>
>> In summary: explicit template instantiation is a way to enforce an
>> instantiation of a template in a translation unit and to prevent from
>> implicit instantiations of that same template.
> this sounds like a good idea. While link time optimization
> may help to do inlineing, unrolling, ... across translation
> until boundaries, this potential runtime penalty could be
> a valid reason against this.
>
> Maybe this could be done in an opt-in fashion:
>
> * Do not use `extern template` in the header.
> * Do some documented explicit instantiation in the library.
> * Users may (but don't have to) prohibit implicit instantiation
>   by using `extern template ...` in their application after
>   including the header. For sure this can only be done for
>   those well-documented instantiations contained in the lib.
> * If the user does not use `extern template ...` she will
>   get implicit instantiation as usual.

I don't know what is the best way to go. But I'm afraid that not writing 
the "extern termplate" in the headers would be more irritating than 
helping. We would just gain longer compile times for the libraries - not 
the opposite. And then the explicit instantiation (with prevented 
implicit instantiation) would mostly not be used. But it is something 
very positive for the user. I think just "well documentation" is not 
enough.

However, I'm aware of the potential problems e.g. that we limit the 
inline-abilty of some functions by explicit instantiation. Whether this 
comes of a performance penalty I don't want to judge without 
benchmarking. Initial tests showed that there is not necessary any 
performance difference. This clearly depends on the application and on 
the template that is instantiated.

An alternative to let the user manually write `extern template class 
...` with the correct template parameters that are already instantiated 
would be a simple compile flag, something along the line:

in the header:

#if ENABLE_TEMPLATE_INSTATIATION
extern template class YaspGrid<2,EquidistantCoordinate<double,2>>;
extern template class YaspGrid<3,EquidistantCoordinate<double,3>>;
#endif

in the source:

#if ENABLE_TEMPLATE_INSTATIATION
template class YaspGrid<2,EquidistantCoordinate<double,2>>;
template class YaspGrid<3,EquidistantCoordinate<double,3>>;
#endif

Or even conditionally add the source to the library target:

in the CMakeLists.txt

option(ENABLE_TEMPLATE_INSTATIATION "Enable explicit template 
instantiation" ON/OFF)
if(ENABLE_TEMPLATE_INSTATIATION)
   target_sources(dunegrid PRIVATE yaspgrid.cc)
endif()

with a corresponding entry in the config.h file

#cmakedefine ENABLE_TEMPLATE_INSTATIATION 1

Whether it is turned ON or OFF by default could be discussed, though.

Best,
Simon


>
> _______________________________________________
> Dune-devel mailing list
> Dune-devel at lists.dune-project.org
> https://lists.dune-project.org/mailman/listinfo/dune-devel

-- 
Dr. Simon Praetorius
Technische Universität Dresden
Institute of Scientific Computing
phone: +49 351 463-34432
mail: simon.praetorius at tu-dresden.de
web: https://tu-dresden.de/mn/math/wir/das-institut/beschaeftigte/simon-praetorius