Do away with task reweighting
Task reweighting is slow, serial, and pretty much useless for multistep.
I therefore suggest the following:
- Use only approximate weights for each task type, depending on the particle counts and pair geometry.
- Only have to reweight after a rebuild, since the task costs are constant.
- Get rid of the task
tic
/toc
counters, i.e. wrap them in an#ifdef
so that we can still have them when we want to generate task plots. - Eventually build a more elaborate model for estimating task costs as a bi-quadratic of the particle counts and geometry.