Support isbits Union Arrays

[benchislett]

Describe the bug
In attempting to move an array of multiple different kinds of structs, I discovered some strange trunctation where they are all forced into the type of one of them.

I initially wanted to use an array of abstract types, but that seemed to work even worse due to their not having a definite size.

To Reproduce
The Minimal Working Example (MWE) for this bug:

julia> using CuArrays

julia> struct Circle
         radius::Int32
       end

julia> struct Square
         width::Int32
         height::Int32
       end

julia> shapes = Array{Union{Circle, Square}}([Circle(4), Square(5, 5)])
2-element Array{Union{Circle, Square},1}:
 Circle(4)   
 Square(5, 5)

julia> cu(shapes)
2-element CuArray{Union{Circle, Square},1,Nothing}:
 Circle(4)
 Circle(5)

Expected behavior
One would think that the structs should come out as they went in.

Build log

julia> Pkg.build()
  Building SpecialFunctions → `~/.julia/packages/SpecialFunctions/ne2iw/deps/build.log`
  Building Conda ───────────→ `~/.julia/packages/Conda/kLXeC/deps/build.log`
  Building FFTW ────────────→ `~/.julia/packages/FFTW/loJ3F/deps/build.log`
  Building ImageMagick ─────→ `~/.julia/packages/ImageMagick/vMfoS/deps/build.log`
  Building GR ──────────────→ `~/.julia/packages/GR/oiZD3/deps/build.log`
  Building FFMPEG ──────────→ `~/.julia/packages/FFMPEG/guN1x/deps/build.log`
  Building Plots ───────────→ `~/.julia/packages/Plots/Ih71u/deps/build.log`
  Building ZipFile ─────────→ `~/.julia/packages/ZipFile/oD4uG/deps/build.log`
  Building NNlib ───────────→ `~/.julia/packages/NNlib/udnNA/deps/build.log`
false

Environment details (please complete this section)
Details on Julia:

Julia Version 1.2.0
Commit c6da87ff4b (2019-08-20 00:03 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

Julia packages:

• CuArrays.jl:

CUDA: toolkit and driver version

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Sun_Jul_28_19:07:16_PDT_2019
Cuda compilation tools, release 10.1, V10.1.243

CUDA Version 10.1.243

EDIT: so we basically need https://docs.julialang.org/en/v1/devdocs/isbitsunionarrays/

[maleadt]

This one isn't trivial, and needs support across the stack:

• add maxsizeandoffsetfields toCuArray` to support inline storage of Union bitstypes: https://docs.julialang.org/en/v1/devdocs/isbitsunionarrays/#isbits-Union-Arrays-1 we can just store the selector bytes past the data like Base does with the current information we have
• copy selector bytes when doing transfers between cpu and gpu (or between gpu arrays): https://github.com/JuliaLang/julia/blob/18783434e9c0e48152b99d5e3717deb8711f492b/base/array.jl#L276-L280
• do the above two steps for CuDeviceArray too
• read and interpret the union selector byte in CUDAnative's arrayref counterpart (currently unsafe_load, but might need additional abstraction level): https://github.com/JuliaLang/julia/blob/18783434e9c0e48152b99d5e3717deb8711f492b/src/codegen.cpp#L2695-L2718
• properly handle all this when creating derived CuArrays (view, reshape)

[maleadt]

In the mean time, JuliaGPU/CuArrays.jl#552 will make this code error.

[tkf]

FYI, I started working on supporting Union eltype for GPU arrays without explicitly coding for GPU arrays: JuliaFolds/UnionArrays.jl#2

[maleadt]

Interesting! I hadn't considered doing this in Julia (well, in regular user code, without extending the arrayref intrinsic).

[maleadt]

By @ViralBShah:

I tried to see what it might take to run DataFrames.jl on GPUs as discussed in JuliaData/DataFrames.jl#2713

Working with the neuro dataset from RDatasets, I could not convert a vector with missing data into a CuArray:

julia> cu(neuro.V1)
ERROR: CuArray does not yet support union bits types
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:33
  [2] CuArray{Union{Missing, Float64}, 1}(#unused#::UndefInitializer, dims::Tuple{Int64})
    @ CUDA ~/.julia/packages/CUDA/Zmd60/src/array.jl:18
  [3] (CuArray{Union{Missing, Float64}, N} where N)(#unused#::UndefInitializer, dims::Tuple{Int64})