replace_storage ignores type parameters

[cuihantao]

Hi,

I have the MWE below to replace the storage with CuArray data types. It looks like replace_storage recognizes CuArray out of CuArray{Float32} but ignores the Float32 part. Did I miss something? Thanks for your help!

using CUDA
using StructArrays

foo = StructArray(a=[1.0 for i = 1:10])

bar1 = replace_storage(CuArray, foo)  # defaults to CuArray{Float64}

@show typeof(bar1)

bar2 = replace_storage(CuArray{Float32}, foo) # still yields CuArray{Float64}

bar3 = convert(CuArray{Float32}, foo.a)  # works fine

[piever]

The docstring of replace_storage mentions that the eltype of the output is expected to be the same. The reason is that for complex types (not NamedTuple types but custom structs), it is a bit difficult to figure out what the output eltype should be. Basically, one would need a method createtype(T, type_1, type_2, ..., type_n) that computes the updated eltype based on column types rather than column values (this could be added to the interface).

If you have a StructArray of NamedTuple, you could instead use

julia> foo = StructArray(a=[1.0 for i = 1:3]);

julia> StructArray(map(Array{Float32}, StructArrays.components(foo)))
3-element StructArray(::Vector{Float32}) with eltype NamedTuple{(:a,), Tuple{Float32}}:
 (a = 1.0,)
 (a = 1.0,)
 (a = 1.0,)

Hope this helps!

[cuihantao]

Thank you. I do have a StructArray of NamedTuple. This is very helpful.