Typedef sequant::PeakSink

• Defined in File eval.hpp

Typedef Documentation

using sequant::PeakSink = std::atomic<double>*

Opt-in sink for the batched evaluator’s per-batch scratch high-watermarks. The batched replay runs each aux/mu~ batch in a separate scratch CacheManager whose transients never enter the outer cache, so the outer cache’s working_set_hwmark() misses the batched-inner peak. When a non-null PeakSink is threaded through make_batched_custom_evaluator (and its nested re-instantiations), each scratch’s high-watermark folds (max) into this one global accumulator, yielding the true batched-replay peak. A null sink (the default) folds nothing and leaves each scratch’s mark local.