Bikeshed datum and array domain definition

[bernhardmgruber]

Defining the datum domain in LLAMA is quite verbose and feels alien. We will eventually get around this with proper reflection in C++, but this will still take time. Until then I would like to make the user experience as little painful and as understandable as possible. Currently, a datum domain could be defined like this:

namespace tag {
    struct Pos {};
    struct X {};
    struct Y {};
    struct Z {};
    struct Momentum {};
    struct Weight {};
    struct Flags {};
}

using XYZ = llama::DS<
    llama::DE<tag::X, double>,
    llama::DE<tag::Y, double>,
    llama::DE<tag::Z, double>
>;

using Particle = llama::DS<
    llama::DE<tag::Pos, XYZ>,
    llama::DE<tag::Weight, float>,
    llama::DE<tag::Momentum, XYZ>,
    llama::DE<tag::Flags, bool[4]> // == llama::DE<tag::Flags, llama::DA<bool, 4>>
>;

We already provide the aliases llama::DS, llama::DE and llama::DA for llama::DatumStruct, llama::DatumElement and llama::DatumArray to take out some of the verbosity. For static arrays I even implemented a shorthand notation, e.g. bool[4]. I still don't like the the abbreviations and considered renaming them at multiple occations:

using XYZ = llama::Struct<
    llama::Member<tag::X, double>,
    llama::Member<tag::Y, double>,
    llama::Member<tag::Z, double>
>;

using Particle = llama::Struct<
    llama::Member<tag::Pos, XYZ>,
    llama::Member<tag::Weight, float>,
    llama::Member<tag::Momentum, XYZ>,
    llama::Member<tag::Flags, bool[4]>
>;

Instead of Member I also considered DataMember, Element or Attribute.

The array domain is later defined using a type holding the dimensions and an instance holding the runtime sizes:

using ArrayDomain = llama::ArrayDomain<2>;
ArrayDomain arrayDomain{16, 16};

I am kind of happy with the way this works, but I am unsure if ArrayDomain is the right name. I considered IndexSpace at some point, because it defines what indices are allowed to access a llama::View. However, we can also see the tag types of the datum domain as (compile time) indices, so the array domain does not solely define the index space.

I am happy for any feedback on the current choice of names! Thank you!

[bernhardmgruber]

In a recent discussion with @jblomer we found Record and Field to be good substitutes for DatumStruct and DatumElement.

[bernhardmgruber]

This idea has been implemented in the meantime with #194.