SoA Benchmarks

• We have set up a benchmark repository: https://github.com/cern-nextgen/wp1.7-soa-benchmark
• David solved the performance problem from last week. Now the GCC code is even faster than the baseline ?!

Simplify the SoA Code

Most type conversions are of a very specific form. Now they are very commmplicated:

auto id = [](auto& member, std::size_t) -> decltype(auto) { return member; };
return helper::apply_to_members<M, value_type&, proxy_type<reference, S>>(data[i], id);

Both the library code and the user code can be made much simpler by providing user defined conversion operators. 

• We get more implicit conversions
• More temmplates can be deduced, i.e. don't need to be specified

The folloeing code is just an example explaining the principle.

#include <span>
#include <vector>

template <class T>
using value = T;

template <class T>
using reference = T&;

template <template <class> class F>
struct S {
    template <template <class> class F_out>
    operator S<F_out>() { return {x, y, z}; }
    F<int> x, y, z;
};

int main() {
    S<std::vector> v {{1, 2, 3}, {4, 5, 6}, {7, 8, 9}};
    S<std::span> s = v;
    
    S<value> a {1, 2, 3};
    S<reference> s = a;

    return 0;
}