Repository navigation
Vectorized getindex of wrapped array falls back to scalar indexing #509
Description
I wasn't sure how much this issue overlaps with existing ones, so apologies if this is a duplicate.
I constantly get into the aforementioned scalar indexing issue when working with transpose.
I opened a new PR (JuliaLang/julia#52626) on base Julia intended to fix the problem, so I wanted to raise awareness and get feedback from the experts here as well.
copying the MWE here to demonstrate the error (which is avoided after the PR changes):
julia> using LinearAlgebra, JLArrays
julia> JLArrays.allowscalar(false)
julia> x = rand(Float32, 2,2) |> jl ; y = transpose(x);
julia> y[:,1]
ERROR: Scalar indexing is disallowed.
Invocation of getindex resulted in scalar indexing of a GPU array.
This is typically caused by calling an iterating implementation of a method.
Such implementations *do not* execute on the GPU, but very slowly on the CPU,
and therefore should be avoided.
If you want to allow scalar iteration, use `allowscalar` or `@allowscalar`
to enable scalar iteration globally or for the operations in question.
Stacktrace:
[1] error(s::String)
@ Base ./error.jl:35
[2] errorscalar(op::String)
@ GPUArraysCore ~/.julia/packages/GPUArraysCore/GMsgk/src/GPUArraysCore.jl:155
[3] _assertscalar(op::String, behavior::GPUArraysCore.ScalarIndexing)
@ GPUArraysCore ~/.julia/packages/GPUArraysCore/GMsgk/src/GPUArraysCore.jl:128
[4] assertscalar(op::String)
@ GPUArraysCore ~/.julia/packages/GPUArraysCore/GMsgk/src/GPUArraysCore.jl:116
[5] getindex
@ ~/.julia/packages/GPUArrays/EoKy0/src/host/indexing.jl:48 [inlined]
[6] scalar_getindex
@ ~/.julia/packages/GPUArrays/EoKy0/src/host/indexing.jl:34 [inlined]
[7] _getindex
@ ~/.julia/packages/GPUArrays/EoKy0/src/host/indexing.jl:17 [inlined]
[8] getindex
@ ~/.julia/packages/GPUArrays/EoKy0/src/host/indexing.jl:15 [inlined]
[9] getindex
@ ~/julia/usr/share/julia/stdlib/v1.11/LinearAlgebra/src/adjtrans.jl:334 [inlined]
[10] macro expansion
@ ./multidimensional.jl:921 [inlined]
[11] macro expansion
@ ./cartesian.jl:64 [inlined]
[12] _unsafe_getindex!
@ ./multidimensional.jl:916 [inlined]
[13] _unsafe_getindex(::IndexCartesian, ::Transpose{Float32, JLArray{Float32, 2}}, ::Base.Slice{Base.OneTo{Int64}}, ::Int64)
@ Base ./multidimensional.jl:907
[14] _getindex
@ Base ./multidimensional.jl:893 [inlined]
[15] getindex(::Transpose{Float32, JLArray{Float32, 2}}, ::Function, ::Int64)
@ Base ./abstractarray.jl:1314
[16] top-level scope
@ REPL[10]:1In particular, I want to hear if there's something I'm missing like some memory contiguous requirements etc.
Cheers
Activity
[-]scalar indexing triggetring on Transpose objects[/-][+]scalar indexing triggering on Transpose objects[/+]maleadt commentedon Jan 2, 2024
Not really; support for wrapped arrays in Julia just... isn't great. We have functionality for scalar and vectorized indexing implemented here, https://github.com/JuliaGPU/GPUArrays.jl/blob/master/src/host/indexing.jl, where the
vectorized_getindexfunctionality in principle supports any kind of array-like GPU input (i.e. not just a JLArray, but also its transpose). However, to make Julia pick up those definitions, we'd need to extend thegetindexdefinition to cover any wrappedAbstractGPUArrays, which the type system just isn't great at. We have an Adapt.jl-basedAnyGPUArraytype, but that's a horrible union which significantly slows down method deserialization, only covers one level of wrapping, and introduces lots of ambiguities. As such, I'm wary to just extend all our functions from::AbstractGPUArrayto::AnyGPUArray. This probably needs a more fundamental fix, maybe something like JuliaLang/julia#51910.[-]scalar indexing triggering on Transpose objects[/-][+]Vectorized getindex of wrapped array falls back to scalar indexing[/+]nomadbl commentedon Jan 2, 2024
Ok thanks for your response.
For now it looks like the PR I linked will at least alleviate some issues but a more general fix would be nice at some point. Perhaps it is also a matter of time and accumulated experience that's needed...
maleadt commentedon Oct 19, 2026
Fixed by #512 (GPUArrays v10.1);
transpose(x)[:,1]now uses the vectorized indexing kernel instead of scalar iteration.