Error transforming a reshaped 0-dimentional GPU array to a CPU array

[GiggleLiu]

julia> b = CuArray(zeros())
0-dimensional CuArray{Float64,0}:
0.0

julia> reshape(b, ()) |> Array
ERROR: ArgumentError: reducing over an empty collection is not allowed
Stacktrace:
 [1] _empty_reduce_error() at ./reduce.jl:299
 [2] reduce_empty(::typeof(Base.add_sum), ::Type{Union{}}) at ./reduce.jl:319
 [3] reduce_empty(::Base.BottomRF{typeof(Base.add_sum)}, ::Type{Union{}}) at ./reduce.jl:328
 [4] reduce_empty_iter at ./reduce.jl:355 [inlined]
 [5] reduce_empty_iter at ./reduce.jl:354 [inlined]
 [6] foldl_impl(::Base.BottomRF{typeof(Base.add_sum)}, ::NamedTuple{(),Tuple{}}, ::Tuple{}) at ./reduce.jl:49
 [7] mapfoldl_impl(::typeof(identity), ::typeof(Base.add_sum), ::NamedTuple{(),Tuple{}}, ::Tuple{}) at ./reduce.jl:44
 [8] mapfoldl(::Function, ::Function, ::Tuple{}; kw::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at ./reduce.jl:160
 [9] mapfoldl at ./reduce.jl:160 [inlined]
 [10] #mapreduce#208 at ./reduce.jl:287 [inlined]
 [11] mapreduce(::Function, ::Function, ::Tuple{}) at ./reduce.jl:287
 [12] sum(::Function, ::Tuple{}) at ./reduce.jl:494
 [13] sum(::Tuple{}) at ./reduce.jl:511
 [14] _memory_offset(::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}, ::Int64) at ./abstractarray.jl:1016
 [15] pointer at /home/ubuntu/.julia/packages/CUDA/gKMm0/src/array.jl:168 [inlined]
 [16] unsafe_copyto!(::Array{Float64,0}, ::Int64, ::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}, ::Int64, ::Int64) at /home/ubuntu/.julia/packages/CUDA/gKMm0/src/array.jl:288
 [17] copyto!(::Array{Float64,0}, ::Int64, ::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}, ::Int64, ::Int64) at /home/ubuntu/.julia/packages/CUDA/gKMm0/src/array.jl:257
 [18] copyto!(::Array{Float64,0}, ::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}) at /home/ubuntu/.julia/packages/CUDA/gKMm0/src/array.jl:261
 [19] copyto_axcheck!(::Array{Float64,0}, ::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}) at ./abstractarray.jl:946
 [20] Array{Float64,0}(::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}) at ./array.jl:562
 [21] Array(::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}) at ./boot.jl:430
 [22] |>(::Base.ReshapedArray{Float64,0,CuArray{Float64,0},Tuple{}}, ::Type{T} where T) at ./operators.jl:834
 [23] top-level scope at REPL[16]:1

I can fix it by dispatching the 0-dimentional reshaped array function, but I think you may know how to fix this issue better.

[maleadt]

Issue in Base:

julia> a = zeros()
0-dimensional Array{Float64,0}:
0.0

julia> b = Base.ReshapedArray(a, (), ())
0-dimensional reshape(::Array{Float64,0}, ) with eltype Float64:
0.0

julia> pointer(b, 1)
ERROR: ArgumentError: reducing over an empty collection is not allowed
Stacktrace:
 [1] _empty_reduce_error() at ./reduce.jl:299
 [2] reduce_empty(::typeof(Base.add_sum), ::Type{Union{}}) at ./reduce.jl:319
 [3] reduce_empty(::Base.BottomRF{typeof(Base.add_sum)}, ::Type{Union{}}) at ./reduce.jl:328
 [4] reduce_empty_iter at ./reduce.jl:355 [inlined]
 [5] reduce_empty_iter at ./reduce.jl:354 [inlined]
 [6] foldl_impl(::Base.BottomRF{typeof(Base.add_sum)}, ::NamedTuple{(),Tuple{}}, ::Tuple{}) at ./reduce.jl:49
 [7] mapfoldl_impl(::typeof(identity), ::typeof(Base.add_sum), ::NamedTuple{(),Tuple{}}, ::Tuple{}) at ./reduce.jl:44
 [8] mapfoldl(::Function, ::Function, ::Tuple{}; kw::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at ./reduce.jl:160
 [9] mapfoldl at ./reduce.jl:160 [inlined]
 [10] #mapreduce#208 at ./reduce.jl:287 [inlined]
 [11] mapreduce(::Function, ::Function, ::Tuple{}) at ./reduce.jl:287
 [12] sum(::Function, ::Tuple{}) at ./reduce.jl:494
 [13] sum(::Tuple{}) at ./reduce.jl:511
 [14] _memory_offset(::Base.ReshapedArray{Float64,0,Array{Float64,0},Tuple{}}, ::Int64) at ./abstractarray.jl:1016
 [15] pointer(::Base.ReshapedArray{Float64,0,Array{Float64,0},Tuple{}}, ::Int64) at ./abstractarray.jl:1009
 [16] top-level scope at REPL[15]:1

I'm considering reverting the SubArray/ReshapedArray/ReinterpretArray change, as it has caused several issues.

[GiggleLiu]

I'm considering reverting the SubArray/ReshapedArray/ReinterpretArray change, as it has caused several issues.

Great, helps a lot to OMEinsum.jl .

[maleadt]

Great, helps a lot to OMEinsum.jl .

Could you elaborate? Can't you run into these wrappers with Base arrays too? Array is special-cased to return a regular Array again, but with other inputs you'd still see SubArray/ReinterpretArray/ReshapeArray (and the DenseCuArray etc unions should take care of those).

[GiggleLiu]

Currently, It is mainly about dispatch, output type stability et. al.
My main complaint is the CuArray reshape. Reshape an Array does not give you a ReshapedArray but reshaping a CuArray does. ReshapedArray makes the type unstable in some places. The rest is actually fine.

I have fixed the issue of dispatch in this PR
https://github.com/under-Peter/OMEinsum.jl/pull/100/files

but not yet tested the performance, I will return to this issue after benchmarking (when I have time).

[maleadt]

I'm considering reverting the SubArray/ReshapedArray/ReinterpretArray change, as it has caused several issues.

#498, feedback appreciated.