Knet can't convert arrays with more than 15 dimensions

[ilkerkesen]

Knet with 3150e5d, AutoGrad with 6c93428. Knet cannot convert arrays which have more than 15 dimensions. Example code snippet,

julia> a2 = rand(Float32, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1); k2 = KnetArray(a2);
ERROR: MethodError: Cannot `convert` an object of type NTuple{14,Int64} to an object of type Tuple{}
This may have arisen from a call to the constructor Tuple{}(...),
since type constructors fall back to convert methods.
Stacktrace:
 [1] convert(::Type{Tuple{Int64}}, ::NTuple{15,Int64}) at ./essentials.jl:135 (repeats 2 times)
 [2] Type at /KUFS/scratch/ikesen16/.julia/newnode/v0.6/Knet/src/karray.jl:83 [inlined]
 [3] Type at /KUFS/scratch/ikesen16/.julia/newnode/v0.6/Knet/src/karray.jl:104 [inlined]
 [4] Type at /KUFS/scratch/ikesen16/.julia/newnode/v0.6/Knet/src/karray.jl:114 [inlined]
 [5] convert(::Type{Knet.KnetArray{Float32,16}}, ::Array{Float32,16}) at /KUFS/scratch/ikesen16/.julia/newnode/v0.6/Knet/src/karray.jl:141
 [6] Knet.KnetArray(::Array{Float32,16}) at ./sysimg.jl:24

However, we can perform the following,

julia> a3 = rand(Float32, map(x->2, 1:15)...); k3 = KnetArray(a3);

julia> k3 = reshape(k3, 1, size(k3)...);

julia> k3 = reshape(k3, 1, size(k3)..., 1);

julia> ndims(k3)
18

It's a very edge case, though.

[denizyuret]

This seems to effect only julia6, works with julia4 and julia5. I also cannot replicate it using the intermediate steps (karray.jl:114, 104, 83). Seems like a Julia bug, keeping it open for now. Not critical for next release.

[ChrisRackauckas]

This is because there is currently a hard limit on inference for recursion. Splatting only infers for tuples up to size 16.

Xref: JuliaLang/julia#22370 JuliaLang/julia#22545

[denizyuret]

This seems to be resolved in latest Julia, closing.