Add Hcat * Vcat special case

[dlfivefifty]

Starts to address #48

@antoine-levitt This should give an idea on what to do. At this point just overloading

materialize!(::MatMulMatAdd{<:VcatLayout,<:HcatLayout})

should suffice.

[antoine-levitt]

Thanks! I was just looking at this.

[antoine-levitt]

I assume we also need something to support products of cats and regular matrices? Or is that magically covered?

[dlfivefifty]

It will fall back to the slow default, but something like the following should be fast for Hcat * Strided (e.g. Hcat * Vector):

function materialize!(M::MatMulVecAdd{<:HcatLayout})
    α,A,B,β,C =  M.α,M.A,M.B,M.β,M.C
    T = eltype(C)
    _fill_lmul!(β,C) # this is temporary until strong β = false is supported
    ind = 0
    for a in A.arrays
        jr = ind+1:ind+size(a,2)
        b = view(B, jr, :) 
        materialize!(MulAdd(α,a,b,one(T),C))
        ind = last(jr)
    end
    C
 end

[antoine-levitt]

OK, thanks! Will try to do the 3^2 combinations with Hcat,Vcat and dense.

[dlfivefifty]

I think you only need 4, not 3^2: Hcat * Vcat, Vcat * Hcat, Vcat * any, Hcat * any. This is because I don’t think Vcat * Vcat or Hcat * Hcat are that useful....

(Ok each will need to be done three times, once for matrix * matrix, matrix * vector and vector * matrix...)