SIMD integration

[bernhardmgruber]

LLAMA's goal is to provide an open set of memory layouts without user program change. Some of these memory layouts may result in automatic vectorization in modern compilers. But fundamentally, vectorization of a user program is independent of the memory layout used, as only load and store operations are affected.

We would like to more clearly define the interface between LLAMA objects such as View, VirtualRecord and mappings, and SIMD vectors. This could result in explicit APIs to move data between SIMD vectors and LLAMA objects, independently of the chosen memory layout, generating appropriate (aligned) load, store, gather or scatter instructions.

One way this could work is by providing an additional LLAMA construct to One<RecordDomain>, e.g. called Some<RecordDomain>, which contains some instances of the RecordDomain. Some would contain SIMD vectors of the record domain. A LLAMA view would need an additional mechanism to fetch a Some instead of a One. Depending on the mapping, this fetch could be an aligned vector load, an unaligned vector load or a gather. Stores would work similarly.