Index mapping is expensive

[icui]

The cost of accessing the indices through saved index mapping, e.g.

const int iglob = field.assembly_index_mapping(field.index_mapping(ispec, iz, ix), static_cast<int>(MediumTag));

seems very expensive. In specfem2d all the indices are computed on the fly. A toy benchmark shows that if we could make that happen there could be ~20% speedup, which accounts for the current performance difference between specfem2d and specfem++.

Acoustic:
0.594864s (remove iglob and index mapping)
0.628938s (remove iglob mapping)
0.778316s (current)

Elastic isotropic:
0.886881s (remove iglob and index mapping)
0.934282s (remove iglob mapping)
1.10545s (current)

Elastic anisotropic:
1.08351s (remove iglob and index mapping)
1.13898s (remove iglob mapping)
1.30267s (current)

[Rohit-Kakodkar]

How would we track assembly of the domain in that case? I.e. when we do atomic add at the end of stiffness routine, multiple (ispec, iz, ix) can relate to the same iglob. The mapping is not a simple one i.e. some linear combination of (ispec, iz, ix) when the domain is not regular.