use abstract array where possible

[kleinschmidt]

This widens type signatures in a bunch of places where Matrix is hardcoded to be AbstractMatrix, and adds type parameters to the Normal[Inverse]Wishart structs that record the types of the mean and covariance matrix fields.

I did this primarily to support StaticArrays, which gives me large speed ups when doing lots of posterior calculations (as in my particle filter code). The support isn't perfect, because in a number of places we fall back to Distributions.jl PDMats.jl, both of which hard-code something more specific than AbstractMatrix in a number of places. But it does get the posterior updates correct.

[mschauer]

What is the use of the conversions/casts, convert(AbstractMatrix, ...) and AbstractMatrix(...)?

[kleinschmidt]

That's how you extract the full matrix from the cholesky factorization since the `full` function was removed. They should all be converts not casts so I guess I missed some... AbstractMatrix is so that static arrays don't get converted back to vanilla arrays.

…

On Jan 15, 2019, at 08:00, Moritz Schauer ***@***.***> wrote: What is the use of the conversions/casts, convert(AbstractMatrix, ...) and AbstractMatrix(...)? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[mschauer]

Ok, makes sense, just check the issue with convert/constructor calls.

[mschauer]

Ahem, actually, I should know... JuliaLang/julia#24595 (comment)

[mschauer]

Merging shortly.

[kleinschmidt]

If you're okay with it go ahead. I lost track of this and can't remember if you still thought changes were necessary. Feel free to update as you see fit!