Template Function sequant::opt::detail::pareto_insert_ceiling¶

Defined in File cost_model.hpp

Function Documentation¶

template<typename FP> void sequant::opt::detail::pareto_insert_ceiling(container::vector<FP> &f, FP p, bool use_nsl)¶

Slice-count-aware Pareto insert for the perf-first batched frontier.

With use_nsl == false this behaves as pareto_insert does (plain (peak, flops) domination; FP::nsl is not consulted). With use_nsl == true (perf-first + a finite peak_threshold) the cumulative sliced-mode count FP::nsl becomes a third Pareto objective, so a point dominates only when it is no worse in peak, flops, and slice count. Because contracted slicing is flops-neutral, the unsliced realization (higher peak, nsl == 0) and a sliced one (lower peak, nsl > 0) are then Pareto- incomparable: both survive on the frontier, all the way to the root, so an ancestor that needs the lower-peak (sliced) subtree still has it and the root selection still has the unsliced one to pick when it fits the budget. select_root then chooses the least-sliced feasible schedule — no free slicing below the ceiling. (A per-flops peak/nsl trade-off keeps at most one point per distinct slice count, so the frontier stays bounded: slicing more modes monotonically lowers peak.) The peak-first path passes false and is unchanged.