Reduction functions with pre-allocated output storage.

[lindahua]

Inplace reduction functions (including sum!, prod!, maximum!, minimum!, all!, and any!) have been in the Base for a while (without being exported). This PR exports them with a modified interface that is safe & flexible.

The interface is defined as:

sum!(r, A[; init=true])

The function reduces A to r along dimensions determined by the size of r and A. This behavior has been in the Base before this PR. What I modify is to add an init keyword argument, such that when init is true (default), it initializes r properly before doing the reduction, and when init is false, it simply accumulates A to r.

A few downstream packages rely on these functions, so that they can do reduction along certain dimensions without allocating new arrays.

[StefanKarpinski]

I have to confess I'm a bit confused about what these functions do. Can you add some documentation and/or just post it here? In-place cumsum and cumprod make sense to me, but I'm not sure this does.

[mbauman]

It's not really an in-place reduction - I had that thought, too. They just allow you to provide a pre-allocated output buffer, r, to store the output of the reduction of A.

[timholy]

Thanks for doing this. Your implementation looks fine to me.

We'll need documentation too. Right now I'm too busy with other things to contribute towards that, but if this isn't done in a week or two I can help out.

[lindahua]

Sorry, this is not in-place. They just allow you to write the results to a pre-allocated storage.

For example,

x = rand(4, 3);
r = zeros(1, 3);
sum!(r, x);    # write sum(x, 1) to a pre-allocated r

[lindahua]

I have changed to title to express the meaning more accurately.

[StefanKarpinski]

Oh, right – yes, that makes perfect sense. Thanks for doing this.

[lindahua]

I have added the documentation & news entry. Should be ready to merge when it passes travis.