Fused type required for floating point Cython operations

[stefanv]

Since we're now (soon ;) letting through float32 arrays in img_as_float, we need to add a fused dtype to all Cython code that consumes doubles to also accept single precision floats.

[grlee77]

Agreed, although some cases may need to use double precision internally for accuracy. An example is the change from float32 to float64 when using the fast integral-image representation for non-local means. A good image demonstrating the problem with float32 is here: #2878 (comment)

[jni]

@grlee77 for those functions we should be using img_as_float64 instead of img_as_float.

[rfezzani]

A big 👍 for this!
The np_floats fused type is already defined in skimage._shared.fused_numerics. Generalizing its use in Cython operations may in fact save memory use and computation time.
It also prevent some hidden data casting from float32 to double, like for example in _slic_cython where the input image is silently converted to double...

[rfezzani]

Oups, I just realized that this issue is more than one year old!!! 😆

[stefanv]

I think we still need someone to do a proper review of all changes required; but since this issue is so vague I'm going to close it.

[rfezzani]

I started this review yesterday when I noticed the silent cast to double in _slic_cython. That's how I found this discussion :-)
I will start by reporting the _slic_cython issue and refer to this one. I will try to investigate more on this.