Module pool 

Rayon-backed parallel for: par_for(total, grain, |off, cnt| …).

Replaces the old per-worker Condvar pool with Rayon’s work-stealing scheduler. Same (offset, count) chunk API so all existing call sites (BLAS tiling, SDPA, LayerNorm, …) pick up Rayon without changes.

Constants

ELEMENTWISE_PAR_FLOOR

Minimum elements that make threading an element-wise kernel worthwhile. Below this the rayon hand-off costs more than the work it saves. Kept in one place so conv/transpose/pool/relu/region kernels share a single, tunable cutover instead of sprinkling magic constants at each call site.

Functions

chunk_floor

par_for chunk floor for a flat range of total elements: aim for one chunk per worker, never below the amortization floor.

num_threads

Total Rayon worker count (configured from [RuntimeConfig::pool_workers]).

outer_chunk

par_for chunk floor for an outer loop of units independent items (e.g. (n,c) planes): one chunk per worker, at least one unit.

par_for

par_range

Parallel for: split total items across threads. f(off, cnt) is called once per chunk with disjoint regions.

should_parallelize

Should an op over work elements run in parallel? True only with >1 worker and enough work to keep them busy past the amortization floor. Use this at every data-parallel kernel call site instead of a hand-picked threshold.