sum_axis_f32

void sum_axis_f32(const float * in, float * out, int outer, int reduce_dim, int inner)

source edit

Reduces a single axis by summation into a packed (outer, inner) output.

For each (o, i) pair the kernel walks the reduce dimension contiguously when inner == 1 and dispatches to Accelerate's vDSP_sve for a vectorised compensated-sum implementation; in the strided case it falls back to the scalar axis_reduce template with identity 0.

Math

$\text{out}[o, i] = \sum_{r=0}^{\text{reduce\_dim}-1} \text{in}[o, r, i].$

See Also

max_axis_f32, min_axis_f32, prod_axis_f32 : Companion reductions sharing the same (outer, reduce_dim, inner) layout convention.

Parameters

inconst float*

Source buffer laid out as (outer, reduce_dim, inner) in row-major order.

outfloat*

Destination buffer of shape (outer, inner); written densely.

outerstd::size_t

Product of all dimensions strictly before the reduce axis. May be 0.

reduce_dimstd::size_t

Size of the axis being summed. A value of 0 yields 0 per cell (the additive identity).

innerstd::size_t

Product of all dimensions strictly after the reduce axis.

Notes

The contiguous fast path uses vDSP_sve which performs a vectorised single-pass summation on Apple Silicon NEON. Numerical results may differ from the scalar loop by a few ULP for very long reductions but are typically stabler.