Error mapreducing over a 0 dimensional array

[GiggleLiu]

MWE,

julia> using CUDA

julia> sum(CUDA.CuArray(fill(randn(), ())))
ERROR: MethodError: no method matching ndims(::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{0}, Nothing, typeof(identity), Tuple{CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}}})
Closest candidates are:
  ndims(::CUDA.CUSPARSE.CuSparseMatrix) at /home/leo/.julia/packages/CUDA/9T5Sq/lib/cusparse/array.jl:166
  ndims(::AbstractFFTs.Plan) at /home/leo/.julia/packages/AbstractFFTs/JebmH/src/definitions.jl:15
  ndims(::Base.Iterators.ProductIterator) at iterators.jl:983
  ...
Stacktrace:
  [1] check_reducedims(R::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}, A::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{0}, Nothing, typeof(identity), Tuple{CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}}})
    @ Base ./reducedim.jl:209
  [2] mapreducedim!(f::typeof(identity), op::typeof(Base.add_sum), R::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}, A::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{0}, Nothing, typeof(identity), Tuple{CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}}}; init::Float64)
    @ CUDA ~/.julia/packages/CUDA/9T5Sq/src/mapreduce.jl:168
  [3] _mapreduce(f::typeof(identity), op::typeof(Base.add_sum), As::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}; dims::Colon, init::Nothing)
    @ GPUArrays ~/.julia/packages/GPUArrays/UBzTm/src/host/mapreduce.jl:62
  [4] mapreduce(::Function, ::Function, ::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}; dims::Function, init::Nothing)
    @ GPUArrays ~/.julia/packages/GPUArrays/UBzTm/src/host/mapreduce.jl:28
  [5] mapreduce(::Function, ::Function, ::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer})
    @ GPUArrays ~/.julia/packages/GPUArrays/UBzTm/src/host/mapreduce.jl:28
  [6] _sum(f::Function, a::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}, ::Colon; kw::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ Base ./reducedim.jl:894
  [7] _sum(f::Function, a::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}, ::Colon)
    @ Base ./reducedim.jl:894
  [8] _sum(a::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}, ::Colon; kw::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ Base ./reducedim.jl:893
  [9] _sum(a::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}, ::Colon)
    @ Base ./reducedim.jl:893
 [10] #sum#722
    @ ./reducedim.jl:889 [inlined]
 [11] sum(a::CuArray{Float64, 0, CUDA.Mem.DeviceBuffer})
    @ Base ./reducedim.jl:889
 [12] top-level scope
    @ REPL[78]:1

CUDA version 3.4.2, Julia version 1.7.0-beta3, OS: Ubuntu 20.04

I guess it is better to fix the method defined in: CUDA ~/.julia/packages/CUDA/9T5Sq/src/mapreduce.jl:168 ?

Edit: defining the following function would fix the problem

Base.ndims(::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{0}, Nothing, typeof(identity), Tuple{CuArray{Float64, 0, CUDA.Mem.DeviceBuffer}}}) = 0

Where are other related functions defined?

[maleadt]

Sorry, forgot about this issue.

Many of these functions live in GPUArrays, so let's close this in favor of JuliaGPU/GPUArrays.jl#362.
There's a PR there too, JuliaGPU/GPUArrays.jl#363, but that doesn't define ndims.