Passing custom struct into kernel function


I’m still a little confused about Adapt.jl. I read several posts in the forum and tried it for my case, but it did not work…:joy:

I have a custom struct like this:

using Adapt
using CUDA

# T is either Float64 or Float32
mutable struct MyStruct{T}
    b::AbstractArray{T, 1}
    c::AbstractArray{T, 2}

struct GPUHelper{T}
    b::CuDeviceVector{T, 1}
    c::CuDeviceMatrix{T, 1}

Adapt.adapt_structure(to, MS::MyStruct) = GPUHelper(
              MS.a ,
    adapt(to, MS.b),
    adapt(to, MS.c)

My aim is to pass data_cpu into GPU (data_gpu) and use it in a kernel function, so I did:

a = 3.8
b = rand(10)
c = rand(10, 2)
data_cpu = MyStruct{Float64}(a, b, c)
data_gpu = MyStruct{Float64}(a, CuArray(b), CuArray(c))

isbits(cudaconvert(data_gpu)) # true

function gpu!(data_gpu)
    # kernel function
    return nothing

and then, I need to copy the data_gpu back to data_cpu by using Adapt.jl.

The above code works, but I have the following questions:

  1. In my case, fields .b and .c are always Vector{T} and Array{Float64, 2}, however, it seems I cannot change their types like this.
  2. As you can see how I get data_gpu. Considering the number of fields, this writing style may not be very pretty. I guess I should be able to implement it with something like data_gpu = adapt(CuArray, data_cpu) or data_cpu = adapt(Array, data_gpu), but unfortunately I get the following error:
ERROR: MethodError: no method matching GPUHelper(::Float64, ::CuArray{Float64, 1, CUDA.Mem.DeviceBuffer}, ::CuArray{Float64, 2, CUDA.Mem.DeviceBuffer})
Closest candidates are:
  GPUHelper(::T, ::CuDeviceVector{T, 1}, ::CuDeviceMatrix{T, 1}) where T at c:\test\case.jl:11
 [1] adapt_structure(to::Type, MS::MyStruct{Float64})
   @ Main c:\test\case.jl:16
 [2] adapt(to::Type, x::MyStruct{Float64})
   @ Adapt C:\Users\test\.julia\packages\Adapt\0zP2x\src\Adapt.jl:40
 [3] top-level scope
   @ REPL[1]:1

Thank you so much~

You can make the types of b and c parametric so that they support both host-side CuArrays and device-side CuDeviceArrays.

struct GPUHelper{T, B<:AbstractVector{T}, C<:AbstractMatrix{T}}