SciML / SciML/ComponentArrays.jl
Scalar indexing on GPU when computing `Zygote.gradient` of `dot(x::CA, x::CA)`
Open
Nobody has claimed this yet.
- Dominant language
- Julia
- Stars
- 375
- Forks
- 42
- Avg merge
- 7h 25m
- Merged PRs (30d)
- 17
Description
julia> Zygote.gradient(x -> dot(x, x), CUDA.ones(4)) # works
(Float32[2.0, 2.0, 2.0, 2.0],)
julia> Zygote.gradient(x -> dot(x, x), (;x=ones(4)) |> ComponentArray |> cu)
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 are only permitted from the REPL for prototyping purposes.
If you did intend to index this array, annotate the caller with @allowscalar.
Stacktrace:
[1] error(s::String)
@ Base ./error.jl:35
[2] assertscalar(op::String)
@ GPUArraysCore ~/.julia/packages/GPUArraysCore/uOYfN/src/GPUArraysCore.jl:103
[3] getindex
@ ~/.julia/packages/GPUArrays/dAUOE/src/host/indexing.jl:48 [inlined]
[4] getindex
@ ~/.julia/dev/ComponentArrays/src/array_interface.jl:94 [inlined]
[5] getindex
@ ~/.julia/packages/OffsetArrays/0MOrf/src/OffsetArrays.jl:436 [inlined]
[6] _broadcast_getindex
@ ./broadcast.jl:675 [inlined]
[7] _getindex
@ ./broadcast.jl:705 [inlined]
[8] _broadcast_getindex
@ ./broadcast.jl:681 [inlined]
[9] getindex
@ ./broadcast.jl:636 [inlined]
[10] macro expansion
@ ./broadcast.jl:1004 [inlined]
[11] macro expansion
@ ./simdloop.jl:77 [inlined]
[12] copyto!
@ ./broadcast.jl:1003 [inlined]
[13] copyto!
@ ./broadcast.jl:956 [inlined]
[14] copy
@ ./broadcast.jl:928 [inlined]
[15] materialize
@ ./broadcast.jl:903 [inlined]
[16] accum(x::OffsetArrays.OffsetVector{…}, ys::OffsetArrays.OffsetVector{…})
@ Zygote ~/.julia/packages/Zygote/YYT6v/src/lib/lib.jl:25
[17] accum
@ ComponentArraysZygoteExt ~/.julia/dev/ComponentArrays/ext/ComponentArraysZygoteExt.jl:10 [inlined]
[18] (::Zygote.var"#75#76"{Zygote.Pullback{Tuple{var"#64#65", ComponentVector{…}}, Tuple{Zygote.ZBack{…}}}})(Δ::Float32)
@ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface.jl:45
[19] gradient(f::Function, args::ComponentVector{Float32, CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}, Tuple{Axis{…}}})
@ Zygote ~/.julia/packages/Zygote/YYT6v/src/compiler/interface.jl:97
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Reproduce the gradient example and start with ext/ComponentArraysZygoteExt.jl, then inspect src/array_interface.jl and the stack-traced indexing paths. Determine why the ComponentArrays/Zygote gradient path performs scalar indexing on the GPU. Done means the shown gradient computation succeeds without the scalar-indexing error.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- julia
- Domain
- data
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100