Simplify array constructors.

[maleadt]

CuArrays/GPUArrays currently define lots of constructors (and methods of similar and convert) that have diverged from Base and really should not exist. This PR tries to get rid of them, to improve maintainability and code portability.

As part of that work, I propose to move essential functionality such as constructors to the concrete packages (ie. CuArrays). Defining these methods in a generic fashion is cumbersome, introduces loads of ambiguities, and results in applicable but nonsensical method invocations. For example:

julia> convert(GPUArray, cu([1]))
ERROR: MethodError: no method matching GPUArray{Float32,N} where N(::UndefInitializer, ::Tuple{Int64})
Stacktrace:
 [1] similar(::Type{GPUArray{Float32,N} where N}, ::Tuple{Int64}) at ./abstractarray.jl:618
 [2] convert(::Type{GPUArray}, ::CuArray{Float32,1}) at /home/tbesard/Julia/GPUArrays/src/construction.jl:97
 [3] top-level scope at none:0

This because of a generic Base.convert(::Type{<: GPUArray{T1}}, ...) that is only defined to be invoked on concrete subtypes. Base doesn't attempt to share such functionality between Array and AbstractArray.

Obviously massively breaking, but seems to pass tests except for some CUSOLVER failure.

[codecov-io]

Codecov Report

Merging #172 into master will decrease coverage by 0.08%.
The diff coverage is 77.02%.

@@            Coverage Diff             @@
##           master     #172      +/-   ##
==========================================
- Coverage   43.56%   43.47%   -0.09%     
==========================================
  Files          34       34              
  Lines        2128     2123       -5     
==========================================
- Hits          927      923       -4     
+ Misses       1201     1200       -1

Impacted Files	Coverage Δ
src/subarray.jl	64.7% <ø> (ø)	⬆️
src/gpuarray_interface.jl	28.57% <ø> (-6.22%)	⬇️
src/fft/genericfft.jl	56.25% <0%> (ø)	⬆️
src/broadcast.jl	33.33% <0%> (-11.12%)	⬇️
src/dnn/nnlib.jl	58.82% <100%> (ø)	⬆️
src/blas/wrap.jl	52.16% <100%> (ø)	⬆️
src/rand/highlevel.jl	82.75% <100%> (ø)	⬆️
src/solver/highlevel.jl	53.57% <33.33%> (ø)	⬆️
src/fft/CUFFT.jl	52.82% <66.66%> (ø)	⬆️
src/array.jl	49.29% <73.91%> (-1.38%)	⬇️
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update aa73ebe...0fef350. Read the comment docs.

[andreasnoack]

Hm. Master is currently failing a ton of tests for me and I suspect that this PR is involved but it looks like CI went through just fine.

[maleadt]

GPUArrays from master? The current runtests.jl matches branches on CI.

[andreasnoack]

No that was the problem. I'd only checked out CuArrays master.