Composite types array, and composite type with array fields

Specific Domains GPU
,
1

Hi,

I am using CUDAnative now and wonder if anyone could point me to a correct direction such that these operations is possible. Is there anyone else found that it will be easier if these 2 functions are supported? I understand that would require some coding in Pointer conversion etc…

  1. composite type array for GPU
struct Type1
    field1::Int32
    field2::Int32
end

array = Vector{Type1}

function kernel_1(a::array)
end
  1. composite with array field
struct Type2
    field1::Vector{Int32}
    field2::Int32
end

function kernel_2(b::Type2)
end
2

Did you try it out yourself? Type1 just works:

julia> struct Type1
           field1::Int32
           field2::Int32
       end

julia> array = [Type1(1,2)]
1-element Array{Type1,1}:
 Type1(1, 2)

julia> gpu_array = adapt(CuArray, array)  # same as `cu` but without the Float32 conversions
1-element CuArray{Type1,1,Nothing}:
 Type1(1, 2)

julia> @cuda ((a)->(@cuprintln(a[1].field1); nothing))(gpu_array)

julia> 1

for Type2 you need to register a GPU-compatible counterpart structs and how to convert to them. But there’s a couple of catches: given how Type2 is currently declared with a hard-coded Vector field, you’d need a Type2GPU field to represent the type with the vector on the GPU, and another for actual use on the GPU. Furthermore, Adapt (the machinery underpinning these conversions) doesn’t change the element type of an array. So it’s probably easier if you use NamedTuple’s instead:

julia> array = [(field1=[1], field2=2)]
1-element Array{NamedTuple{(:field1, :field2),Tuple{Array{Int64,1},Int64}},1}:
 (field1 = [1], field2 = 2)

julia> gpu(x) = adapt(CuArray, x)
gpu (generic function with 1 method)

julia> gpu_array = gpu(gpu.(array))
1-element CuArray{NamedTuple{(:field1, :field2),Tuple{CuArray{Int64,1,Nothing},Int64}},1,Nothing}:
 (field1 = [1], field2 = 2)
1 Like
3

or Type2 you need to register a GPU-compatible counterpart structs

Would it be possible to have a generic counterpart struct for StructArrays.StructArray? That might cover a lot of use cases.

1 Like
4

Or, since StructArray can take any kind of array, would it be enough to make Adapt aware of it somehow?

CC @piever

1 Like
5

There is a PR for that, but it seems a bit too complex. In principle I’d be happy to make StructArrays fully GPU compliant though.

2 Likes
6

Hi,

Yes thanks for your comments it really works. And I just realized if everything in type2 is tuple instead of vector then everything that I wanted will just work, due to the static allocation.

Thanks