JuliaGPU / CUDA.jl

CUDA programming in Julia.
https://juliagpu.org/cuda/
Other
1.16k stars 206 forks source link

Adapt + strictly-typed fields can trigger confusing errors #2384

Open maleadt opened 1 month ago

maleadt commented 1 month ago

Say, a user wants to pass a struct with CuArrays to the GPU:

julia> struct Foo
       x::CuVector{Float32}
       end

julia> @cuda Returns(nothing)(Foo(CUDA.rand(1)))
ERROR: GPU compilation of MethodInstance for (::Returns{Nothing})(::Foo) failed
KernelError: passing and using non-bitstype argument

We may tell him to define an Adapt rule to convert the CuArray to CuDeviceArray. However, because the field isn't parametric, that triggers a very confusing error:

julia> Adapt.@adapt_structure Foo

julia> @cuda Returns(nothing)(Foo(CUDA.rand(1)))
ERROR: This function is not intended for use on the CPU
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] arrayref(A::CuDeviceVector{Float32, 1}, index::Int64)
    @ CUDA ~/Julia/pkg/CUDA/src/device/utils.jl:29
  [3] getindex
    @ ~/Julia/pkg/CUDA/src/device/array.jl:164 [inlined]
  [4] copyto_unaliased!
    @ ./abstractarray.jl:1088 [inlined]
  [5] copyto!
    @ ./abstractarray.jl:1068 [inlined]
  [6] copyto_axcheck!
    @ ./abstractarray.jl:1177 [inlined]
  [7] Array
    @ ./array.jl:673 [inlined]
  [8] Array
    @ ./boot.jl:501 [inlined]
  [9] convert
    @ ./array.jl:665 [inlined]
 [10] CuArray
    @ ~/Julia/pkg/CUDA/src/array.jl:406 [inlined]
 [11] CuArray
    @ ~/Julia/pkg/CUDA/src/array.jl:410 [inlined]
 [12] convert(::Type{CuArray{Float32, 1}}, a::CuDeviceVector{Float32, 1})
    @ GPUArrays ~/Julia/pkg/GPUArrays/src/host/construction.jl:4
 [13] Foo
    @ ./REPL[5]:2 [inlined]
 [14] adapt_structure(to::CUDA.KernelAdaptor, obj::Foo)
    @ Main ~/Julia/pkg/Adapt/src/macro.jl:11
 [15] adapt
    @ ~/Julia/pkg/Adapt/src/Adapt.jl:40 [inlined]
 [16] cudaconvert
    @ ~/Julia/pkg/CUDA/src/compiler/execution.jl:198 [inlined]
 [17] map(f::typeof(cudaconvert), t::Tuple{Foo})
    @ Base ./tuple.jl:291
 [18] top-level scope
    @ ~/Julia/pkg/CUDA/src/compiler/execution.jl:110

I'm not sure it makes sense to try and convert fields that aren't parametric (because it would just convert the values back again). So either we fix this in Adapt, or we provide some convert implementation that errors early and makes the issue more clear.

FWIW, to fix the actual issue the field needs to be parametric:

julia> struct Bar{T<:AbstractArray}
       x::T
       end

julia> Adapt.@adapt_structure Bar

julia> @cuda Returns(nothing)(Bar(CUDA.rand(1)))