Performance: getindex(a, i::Array{Int})

[denizyuret]

This is important when computing loss:

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

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

julia> i = sort(rand(1:1000,100))

julia> @benchmark k[i]
@benchmark k[i]
BenchmarkTools.Trial: 
  memory estimate:  1.48 KiB
  allocs estimate:  23
  --------------
  minimum time:     26.711 μs (0.00% GC)
  median time:      29.097 μs (0.00% GC)
  mean time:        30.774 μs (0.00% GC)
  maximum time:     787.018 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

julia> @benchmark c[i]
@benchmark c[i]
BenchmarkTools.Trial: 
  memory estimate:  2.48 KiB
  allocs estimate:  84
  --------------
  minimum time:     44.229 μs (0.00% GC)
  median time:      48.130 μs (0.00% GC)
  mean time:        48.999 μs (0.00% GC)
  maximum time:     833.070 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

[maleadt]

Looks fixed with the recent round of improvements:

Knet.jl
BenchmarkTools.Trial: 
  memory estimate:  1.50 KiB
  allocs estimate:  23
  --------------
  minimum time:     11.236 μs (0.00% GC)
  median time:      12.286 μs (0.00% GC)
  mean time:        15.525 μs (3.11% GC)
  maximum time:     14.404 ms (33.53% GC)
  --------------
  samples:          10000
  evals/sample:     1

CUDA.jl
BenchmarkTools.Trial: 
  memory estimate:  1.06 KiB
  allocs estimate:  46
  --------------
  minimum time:     10.209 μs (0.00% GC)
  median time:      11.297 μs (0.00% GC)
  mean time:        12.564 μs (0.00% GC)
  maximum time:     1.114 ms (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1