[on hold] Draft/demo: Use Boost.PFR to reflect record dimension

[bernhardmgruber]

Do not merge this PR! For now this is just a demo.

Using Boost.PFR we can reflect existing structures and explores the possibilities:

• Create LLAMA's datum domain without the need of these unwieldy typelists, just by reflection an existing struct.
• This existing struct is also used to represent data on the stack, so llama::One<T> should also no longer be needed.
• Use your overloaded operators on your struct instead of the ones LLAMA provides.
• Navigate on your data structure using view(x, y)->pos.x instead of view(x, y)(tag::Pos{}, tag::X{}) during reading.

Datum domain definition before:

using Particle = llama::DS<
    llama::DE<tag::Pos, llama::DS<
        llama::DE<tag::X, FP>,
        llama::DE<tag::Y, FP>,
        llama::DE<tag::Z, FP>
    >>,
    llama::DE<tag::Vel, llama::DS<
        llama::DE<tag::X, FP>,
        llama::DE<tag::Y, FP>,
        llama::DE<tag::Z, FP>
    >>,
    llama::DE<tag::Mass, FP>
>;

After:

struct Vec
    FP x;
    FP y;
    FP z;
    // ... operators
};

struct Particle {
    Vec pos;
    Vec vel;
    FP mass;
};

Compare the change of the n-body pPInteraction function from:

void pPInteraction(VirtualParticleI&& pi, VirtualParticleJ pj) {
    auto dist = pi(tag::Pos{}) - pj(tag::Pos{});
    dist *= dist;
    const FP distSqr = EPS2 + dist(tag::X{}) + dist(tag::Y{}) + dist(tag::Z{});
    const FP distSixth = distSqr * distSqr * distSqr;
    const FP invDistCube = 1.0f / std::sqrt(distSixth);
    const FP sts = pj(tag::Mass{}) * invDistCube * TIMESTEP;
    pi(tag::Vel{}) += dist * sts;
}

to:

void pPInteraction(VirtualParticleI&& pi, VirtualParticleJ pj) {
    Particle dist = pi->pos - pj->pos; // your struct, your operators
    dist *= dist;
    const FP distSqr = EPS2 + dist.x + dist.y + dist.z;
    const FP distSixth = distSqr * distSqr * distSqr;
    const FP invDistCube = 1.0f / std::sqrt(distSixth);
    const FP sts = pj->mass * invDistCube * TIMESTEP;
    pi(1_DC) = pi->vel + dist * sts; // 1_DC == llama::DatumCoord<1>
}

Also consider the particle creation loop changed from:

Particle p = particles(i);
p(tag::Pos{}, tag::X{}) = dist(engine);
p(tag::Pos{}, tag::Y{}) = dist(engine);
p(tag::Pos{}, tag::Z{}) = dist(engine);
p(tag::Vel{}, tag::X{}) = dist(engine) / FP(10);
p(tag::Vel{}, tag::Y{}) = dist(engine) / FP(10);
p(tag::Vel{}, tag::Z{}) = dist(engine) / FP(10);
p(tag::Mass{}) = dist(engine) / FP(100);

to:

Particle p;
p.pos.x = dist(engine);
p.pos.y = dist(engine);
p.pos.z = dist(engine);
p.vel.x = dist(engine) / FP(10);
p.vel.y = dist(engine) / FP(10);
p.vel.z = dist(engine) / FP(10);
p.mass = dist(engine) / FP(100);
particles(i) = p;

This prototype has a drawback, because it invokes the mapping function unnecessarily. During each datum access view(x, y)->, a full datum is fetched which is used for navigation. The compiler discards the unused loads and computations in the mapping function. But this still leads to wrong results if the mapping function is stateful.