Experiment different Loop unrolling profiles on different arm microarchitectures (falkor, vulcan, a72) and find the optimal amount of loop unrolling based on experiments.

Based on current experiments A72 on Marvell seems to give better nos from switching to Dual Loop (less aggressive prefetching) from Quad Loop (more aggressive prefetching). The reason is pfrm (prefetch instruction) is seen as a hotspot in A72 on Marvell. The same is not the case for Falkor CPU.

Experiments to be done and different architecture specific profiles to be used,