Inplace multiplication of sub-matrices without allocations

General Usage Performance
, ,
9

Out of curiosity, why do we use view(A, :, :, dim) instead of View(A)[:, :, dim]? It seems like much of the motivation for the @view macro is to allow the begin and end keywords which are valid in [] square brackets, but if we simply had an indexable type we get those back.

Playing around, it seems like this should be doable:

struct View{A<:AbstractArray} a::A end
Base.getindex(v::View, args...) = view(v.a, args...)
Base.axes(v::View, args...) = axes(v.a, args...)
View(A)[:, :, end] # this just works

Applying it here:

julia> function foo6!(A,B,C,dim)
           mul!(View(C)[:,:,dim], View(A)[:,:,dim], B)
           nothing
       end
foo6! (generic function with 1 method)

julia> @btime foo6!(A,B,C,1)
  330.808 ns (0 allocations: 0 bytes)
2 Likes