Extending ndarray for binary arrays

[AdamWRichardson]

Hi everyone,

I want to create a new crate which extends on the great ndarray work but is specific for binary matrices. My two new types I'd like to create are:

• Bitpacked binary array - where the underlying data is a bitvec or bitarray (from this crate)
• Sparse binary array - which could have many different implementations

I was hoping to get some pointers on the best way to achieve this. Thanks!

[akern40]

Hello! On the topic of bitpacked arrays, this has been brought up in #1258 but wasn't implemented. Is this the same idea - essentially a BitArrD? On the topic of sparse representations, we currently don't really have a way to switch out the underlying memory representation under ndarray, nor a set of traits that could be implemented for anything that is like an ndarray. Both of these things are of interest to me, however.

[AdamWRichardson]

Thanks for the reply, good spot with that issue although a shame about the proprietary implementation. Is there already a proposal for a trait based memory API, or a discussion about how to implement it? I think that would be incredibly useful to have not just for my use cases but also to enable testing and bench-marking of different implementations (for example when std simd is stabilised).

[akern40]

There isn't an official RFC - you can check out the design-doc branch of akern40/ndarray-design, where I have a document that lays out one potential design for a new base type that allows for customized backends (among other things). I'm frankly unsatisfied with that particular design, but there are some difficult pros and cons to weigh between a trait-based vs type-based design for switchable backends.