mapreduce: change iteration for compatibility with non-commutative operators.

[maleadt]

Fixes #484

@tkf: do you have an example of an invocation with a scalar element type (to see if the shuffle version also works)?

[codecov]

Codecov Report

Merging #500 into master will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #500   +/-   ##
=======================================
  Coverage   80.54%   80.55%           
=======================================
  Files         166      166           
  Lines        8879     8880    +1     
=======================================
+ Hits         7152     7153    +1     
  Misses       1727     1727           

Impacted Files	Coverage Δ
src/mapreduce.jl	97.82% <100.00%> (+0.04%)	⬆️

---

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 7616664...a2aaabb. Read the comment docs.

[tkf]

Thanks for fixing this!

with a scalar element type

How about this?: reinterpret(UInt32, maximum(CuArray(reinterpret.(Float32, reinterpret(UInt32, NaN32) .+ (UInt32(5):UInt32(100000)))))) == reinterpret(UInt32, NaN32) + 5

It's true if I remove CuArray.

The idea is to use different bit patterns of NaN to distinguish NaNs at different positions.

[maleadt]

That doesn't seem to work:

julia> max(reinterpret(UInt32, NaN32) + 5, reinterpret(UInt32, NaN32) + 6) == reinterpret(UInt32, NaN32) + 5
false

julia> reinterpret(UInt32, maximum(Array(reinterpret.(Float32, reinterpret(UInt32, NaN32) .+ (UInt32(5):UInt32(6)))))) == reinterpret(UInt32, NaN32) + 5
false