ENH: Allow specifying reduction axis in nditer/NpyIter

A reduction axis can be a broadcast axis (-1), which does not
exist. Or an existing axis which then must have a length of 1, or
will be allocated with a length of 1.

This allows the correct allocation of reduction outputs with
axes of size 1 (instead of only newaxis/-1), i.e. it allows
supporting the result output with `keepdims=True` in reductions.

It can also be used by generalized ufuncs to specific that one
operand does not need to iterate a core dimensions. In the case
of generalized ufuncs, this is handled like a reduce dimensions,
but that dimension is later removed. This means that the operand
can be specified without read-write.
It mainly is necessary to allow normal broadcasting for the axes
_not_ marked as reduce axis.

In most cases however, an operand should never use broadcasting
when marking axes as reduction.

0 files changed, 0 insertions, 0 deletions