cu(::SVector) gives SVector, cu(::MVector) gives CuArray

[islent]

Here's a MWE:

using CUDA, StaticArrays
N = 10
SV = @SVector rand(N);
SA = @SMatrix rand(N,N);
MV = @MVector rand(N);
MA = @MMatrix rand(N,N);

typeof(cu(SV))
typeof(cu(SA))
typeof(cu(MV))
typeof(cu(MA))

Results:

julia> typeof(cu(SV))
SVector{10, Float64} (alias for SArray{Tuple{10}, Float64, 1, 10})

julia> typeof(cu(SA))
SMatrix{10, 10, Float64, 100} (alias for SArray{Tuple{10, 10}, Float64, 2, 100})

julia> typeof(cu(MV))
CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}

julia> typeof(cu(MA))
CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}

Environments:

(@v1.7) pkg> st CUDA
      Status `E:\.julia\environments\v1.7\Project.toml`
  [052768ef] CUDA v3.5.0

(@v1.7) pkg> st StaticArrays
      Status `E:\.julia\environments\v1.7\Project.toml`
  [90137ffa] StaticArrays v1.2.13

julia> versioninfo()
Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Xeon(R) W-10885M CPU @ 2.40GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)
Environment:
  JULIA_DEPOT_PATH = E:\.julia
  JULIA_NUM_THREADS = 1
  JULIA_EDITOR = code

Is this a bug or feature?

[maleadt]

A feature; SVector is a bitstype so can be used on the GPU directly. MArray is mutable, so you can't pass it to a kernel, and the fallback definition treating it as an array (so yielding a CuArray) kicks in.

[maleadt]

Although there's something to be said that cu(SVector) should also return a CuArray. But then cu((1,)) will also result in a CuArray, which may be a fairly breaking change. I'm not sure it's worth doing so.