max_axis_f32

void max_axis_f32(const float * in, float * out, int outer, int reduce_dim, int inner)

source edit

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

The kernel always uses the generic axis_reduce template — Accelerate does not expose a stride-aware vDSP equivalent for max with the same semantics, so vectorisation is left to the compiler. Initial accumulator is -inf which means an empty reduce dimension yields -inf per cell.

Math

$\text{out}[o, i] = \max_{r} \text{in}[o, r, i].$

Parameters

inconst float*

Source buffer of shape (outer, reduce_dim, inner).

outfloat*

Destination buffer of shape (outer, inner).

outer, reduce_dim, innerstd::size_t

Layout extents; see sum_axis_f32.

Notes

NaN propagation follows the C++ a > b semantics: a NaN on either side of the comparison loses, so a NaN element will not necessarily survive the reduction. Downstream code that requires NaN-poisoning must guard explicitly.