cumsum(DA; dims) throws a ConcurrencyViolationError #770

MWE

julia> using Dagger

julia> DA = distribute(rand(4, 4), Blocks(2, 2));

julia> cumsum(DA, dims=1)
ERROR: ConcurrencyViolationError("Cannot resolve aliasing for object of type Any (element of Matrix{Any}), execution may become sequential")

julia> cumsum(distribute(rand(4), Blocks(2)))
4-element DVector{Any} with 2 partitions of size 2:
 0.0206831
 0.437184
 1.14558
 1.77878

cumprod and accumulate behave the same. The vector case returns the right values, but with an Any eltype.

Cause

There is no accumulate method for a DArray, so Base's generic _accumulate! runs element by element. Base allocates the output as a DMatrix{Any}. Reading A[i, I] with I::CartesianIndex reaches getindex(A::DArray, idx...), which returns a 0-dimensional DArray and not a scalar:

julia> DA[1, 2]
0.5772936968348066

julia> DA[1, CartesianIndex(2)]
0-dimensional DArray{Float64, 0} with 1 partitions of size 1:
0.577294

Writing that into the output is then a copyto! between two DArrays under spawn_datadeps, which cannot work out the aliasing of a Matrix{Any}.

I can write a partition wise cumsum/cumprod (accumulate inside each partition, then carry the last slice forward along dims) if that is something you want.

Dagger master (02381759, 0.22.6)
Julia 1.12.7, Linux x86_64