Template Function sequant::detail::make_batched_scratch

• Defined in File eval.hpp

Function Documentation

template<typename TreeNode, bool FHC, typename Members>
BatchedScratch<TreeNode, FHC> sequant::detail::make_batched_scratch(Members const &members, CacheManager<TreeNode, FHC> const &real)

Builds the scratch CacheManager for one batched replay pass over members (each a subtree root paired with its batch axis).

Walks every member subtree with the same pruned counting walk as cache_manager() (descend on first visit of a canonical-equal node, so counts match access counts under caching &#8212; and also on a re-encounter whose slicing signature differs from the first visit’s, so that descendants’ signatures under an inconsistently-sliced occurrence are recorded rather than hidden by the prune) and registers every internal subnode that repeats AND has a consistent slicing signature &#8212; the position of the containing member’s batch axis in the subnode’s canon_indices(), or its absence &#8212; across all occurrences. Signature consistency is what makes a scratch hit exact across members: canonical equality maps the index at canonical position p to the index at position p, so equal signatures plus equal realized element ranges (guaranteed by the caller’s grouping) imply identical slices. Inconsistently-sliced subnodes are not registered and so are evaluated per occurrence, unshared. Count inexactness arising from the pruned walk is benign: an undercount makes evaluate() recompute a drained entry, an overcount keeps an entry until the per-batch reset().

Subnodes whose signature is consistently ‘absent’ (no leaf below carries the axis &#8212; the axis is contracted at the member’s root, so a subtree containing an axis-carrying leaf carries the axis free in its canon_indices()) have batch-invariant full values; those that are alive persistent entries of real are returned as seeds, and the caller copies their values into the scratch before the batch loop.