Fix Base.mapreduce to match the new signature in 0.7/1.0

[ssz66666]

In Julia 0.7/1.0, the signature of Base.mapreduce and its alternative function Base.mapreducedim which allows a dims argument, have changed to use keyword arguments, i.e.
from
mapreduce(f, op[, v0], itr)
mapreducedim(f, op, itr, region[, v0])
to
mapreduce(f, op, itr; dims=:, [init])

It seems that the bot wasn't able to figure out that this is deprecated syntax. The current head is still passing tests since no test case tries to pass a customised initial value to it.

[vchuravy]

Thanks! It would be greatly appreciated if you can add testcases for this as well!

[ssz66666]

Apologies for the messy commit that combines different changes I made. I'll try to make a new branch for each fix next time.
Changes I made (based on master):

1. try to fix Base.show, which includes two variants:

Base.show(io::IO, ::MIME"text/plain", X::GPUArray), which is used by e.g. display to show a human-readable array on REPL.
This calls print_array.

Base.show(io::IO, X::GPUArray), which is used by e.g. print to show a string that is close to its literal value string, like you write [1] to create an Array{Int64,1} with 1 as its sole element.
This calls show_vector for vectors, _show_empty for empty non-vector arrays, and _show_nonempty for non-empty non-vector arrays.

2. Add test cases for:
   showing: I included use of both variants of Base.show as well as the case of showing a 2 by 2 matrix.
   mapreduce: as suggested by @vchuravy , included a test case to check calls with keyword args.

3. Add GPUArrays.allowscalar(false) in test(AT::Type{<:GPUArray})#src/testsuite.jl. This does not seem to be enforced by default and thus allows some scalar index errors to sneak through unit test/CI.

[maleadt]

maybe use Base.parent instead of accessing the field directly

[maleadt]

my reasoning here was that implementations can decide themselves if they want to assert scalar indexing, eg. https://github.com/JuliaGPU/CuArrays.jl/blob/master/test/runtests.jl#L27
but then it probably needs to be added to the JLArray tests.

[ssz66666]

Then maybe moving scalar indexing assertion into runtest.jl of GPUArrays package for JLArray tests, like

@testset "JLArray" begin
    local _prev = GPUArrays._allowscalar[]
    GPUArrays.allowscalar(false)
    GPUArrays.test(JLArray)
    GPUArrays._allowscalar[] = _prev
end

[maleadt]

Something like that, yeah. But we can leave that for a different PR.

[maleadt]

(something like @allowscalar, eg. @disallowscalar would be cleaner then, not having to work with unexported values in GPUArrays)

[ssz66666]

Fixed wrapper field access, will submit a separate PR for the test suite behaviour change.

[maleadt]

Thanks! Too bad we can't run CI on this, guess we'll have to keep a close eye on master after merging PRs like this.

[maleadt]

OK, so this apparently broke CuArrays (as per CI), due to passing a CPU array to the GPU. Did you test with CuArrays.jl? Once JuliaLang/METADATA.jl#17380 is merged CI will report better errors.

EDIT: ah yeah, this is just mapreduce being implemented without kwargs in GPUArrays.

[ssz66666]

Hmm, don't know why the added unit test broke CuArrays. CUDAnative fails to compile the GPU kernel for some reason. Right now it seems that it is the occurrence of anonymous function _addone that is causing the problem. manually running the code in Main outside the testsuite seemed to work fine...

EDIT4: @maleadt I've submitted another PR #151 to attempt to workaround the problem. It seems to me the problem here is CUDAnative cannot handle non-bits types as kernel function parameter, with the exception of a few types, such as CuArray (obviously). A function(closure) capturing a local scope non-bit type variable is itself of non-bits type, thus is not a valid parameter for a GPU kernel.
EDIT3: fix typo
EDIT2: See the code below. Totally confused.

julia> begin
           local a = 1
           local b = Int32
           local c = one(b)
           local d = Ref(1)
           local f(x) = x
           local g(x) = x + a
           local h(x) = x + one(b)
           const j(x) = x + one(b)
           local k(x) = x + c
           local l(x) = x + d[]
           println(isbitstype(typeof(f)))
           println(isbitstype(typeof(g)))
           println(isbitstype(typeof(h)))
           println(isbitstype(typeof(j)))
           println(isbitstype(typeof(k)))
           println(isbitstype(typeof(l)))
       end
true
true
false
true
true
false

EDIT1: not an anonymous function