pub fn reduce_axis<T: Copy + Send + Sync, Op: ElementOp<T>, M, R, U>(
src: &StridedView<'_, T, Op>,
axis: usize,
map_fn: M,
reduce_fn: R,
init: U,
exec: Exec<'_>,
) -> Result<StridedArray<U>>Expand description
Reduce along a single axis, returning a new StridedArray.
This allocates a column-major output and delegates to
reduce_axis_into.
Every output folds init once and then its reduced elements. When the
reduced axis has unit stride those elements are folded with the lane
scheme of reduce, so reduce_fn must be associative and commutative.
When the leading kept axis has unit stride, a block of adjacent outputs
is accumulated one contiguous source column at a time, which keeps the
left to right order per output. Outputs are split across the budget of
exec once the number of source elements read exceeds the threading
threshold; each output is computed by one task, so the result does not
depend on the budget.
§Errors
Returns StridedError::InvalidAxis when axis >= src.ndim(), and
StridedError::OffsetOverflow when the source element count cannot be
represented in usize. The destination layout is built here, so the
destination-shape errors of reduce_axis_into cannot occur.
§Panics
Panics when the kept source extents cannot be described in isize, which
an empty source extent hides from the element-count check: reducing axis 0
of a [0, usize::MAX] source builds the output strides with
col_major_strides, which panics instead of reporting an error.