Specialization fixes for mapreducedim.

[maleadt]

Fixes #302

julia> k = KnetArray{Float32}(rand(10,100));

julia> c = CuArray{Float32}(rand(10,100));

julia> @benchmark sum(k)
BenchmarkTools.Trial: 
  memory estimate:  32 bytes
  allocs estimate:  2
  --------------
  minimum time:     8.516 μs (0.00% GC)
  median time:      8.788 μs (0.00% GC)
  mean time:        8.817 μs (0.00% GC)
  maximum time:     23.335 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3

julia> @benchmark sum(c)
BenchmarkTools.Trial: 
  memory estimate:  1.08 KiB
  allocs estimate:  37
  --------------
  minimum time:     12.593 μs (0.00% GC)
  median time:      19.867 μs (0.00% GC)
  mean time:        19.804 μs (0.00% GC)
  maximum time:     325.308 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

julia> @benchmark sum(k,dims=1)
BenchmarkTools.Trial: 
  memory estimate:  288 bytes
  allocs estimate:  11
  --------------
  minimum time:     2.899 μs (0.00% GC)
  median time:      3.016 μs (0.00% GC)
  mean time:        3.072 μs (0.00% GC)
  maximum time:     163.371 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     9

julia> @benchmark sum(c,dims=1)
BenchmarkTools.Trial: 
  memory estimate:  960 bytes
  allocs estimate:  33
  --------------
  minimum time:     3.236 μs (0.00% GC)
  median time:      3.493 μs (0.00% GC)
  mean time:        4.332 μs (4.12% GC)
  maximum time:     5.562 ms (32.07% GC)
  --------------
  samples:          10000
  evals/sample:     8

julia> @benchmark sum(abs2, k)
BenchmarkTools.Trial: 
  memory estimate:  32 bytes
  allocs estimate:  2
  --------------
  minimum time:     8.926 μs (0.00% GC)
  median time:      9.180 μs (0.00% GC)
  mean time:        9.288 μs (0.00% GC)
  maximum time:     128.632 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     3

julia> @benchmark sum(abs2, c)
BenchmarkTools.Trial: 
  memory estimate:  1.08 KiB
  allocs estimate:  37
  --------------
  minimum time:     13.336 μs (0.00% GC)
  median time:      20.882 μs (0.00% GC)
  mean time:        20.743 μs (0.00% GC)
  maximum time:     66.558 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

From x10 to less than 50% overhead.

@denizyuret at this point the largest issues are gone, and it would be good to port over some over the tricks that Knet does. For example, I think the scalar reductions here avoid allocating an output container (I couldn't see a cudaMalloc in the profiler), which might account for the remaining overhead.

[denizyuret]

Looks good. The scalar reduction code is in Knet/deps/cuda20.jl, how can I help with the port?

[codecov]

Codecov Report

Merging #316 into master will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #316      +/-   ##
==========================================
+ Coverage   79.33%   79.34%   +0.01%     
==========================================
  Files         155      155              
  Lines        8902     8900       -2     
==========================================
  Hits         7062     7062              
+ Misses       1840     1838       -2     

Impacted Files	Coverage Δ
src/mapreduce.jl	100.00% <100.00%> (+4.25%)	⬆️

---

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 5dc771d...3107df7. Read the comment docs.