dtype control for corrcoef

[orenbenkiki]

Feature

I am computing a cross-correlation matrix of 150K * 150K matrix of float32.

Numpy insists on making the results dtype be float64. This causes the amount of memory to be "too much".

I would debate that default, but changing it would probably be too disruptive (not backward compatible).

However, how about a dtype parameter for corrcoef?

In general I would have liked to see a dtype parameter "everywhere" to override the float64 default...

[lschwetlick]

Looks like corrcoef relies on cov in which deliberately casts to float64.

I made an experimental PR (#17456) for this, and it seems to work without much of a hitch.

[This is my first PR to numpy so I'm happy to get any feedback you might have!]

[orenbenkiki]

Wow, thanks!