Module lane_kernels 

Elementwise lane kernels over indexed sources.

Replaces &[T] with an IndexedSource trait: each lane read is unsafe fn get_unchecked(i) -> Item, independent across iterations. For &[T] this inlines to the same indexed load as the slice kernel; for LaneZip(&[A], &[B]) it gives two independent indexed reads per lane — both shapes the auto-vectorizer handles.

The module is split into:

• source — the IndexedSource trait, LaneZip, and read-only adapters.
• sink — the IndexedSink trait and ReinterpretSink.
• map_into — out-of-place kernels via IndexedSourceExt (writes into a caller-provided &mut [MaybeUninit<R>]).
• map_in_place — in-place kernels via IndexedSinkExt (writes back through the sink itself).

The kernels never allocate. Both kernel families handle a mask with a non-byte-aligned offset and with a logical len shorter than the underlying byte buffer, via BitBuffer::chunks.

Re-exports

pub use map_in_place::IndexedSinkExt;

pub use map_into::IndexedSourceExt;

pub use sink::IndexedSink;

pub use sink::ReinterpretSink;

pub use source::IndexedSource;

pub use source::LaneZip;

Modules

map_in_place

In-place lane kernels: read from an IndexedSink and write back through the same sink (no separate output buffer).

map_into

Out-of-place lane kernels: read from an IndexedSource and write into a caller-provided &mut [MaybeUninit<R>].

sink

Writable lane sink — the IndexedSink trait and the ReinterpretSink adapter for in-place type-punning kernels.

source

Read-only lane source — the IndexedSource trait and the LaneZip adapter.