convert(Vector{Int}, o) is much slower than copy(PyVector{Int}(o))

[stevengj]

As discussed on julia-users, the two pathways to converting a Python list to a Vector{Int} seem mysteriously different in performance. For example:

iv = PyVector(rand(Int16, 10^6))
@time copy!(Array(Int, length(iv)), iv);
@time convert(Vector{Int}, iv.o);

yields 3-5× slower performance for the second version on my machine.

[cortner]

I just tested this on Mac OS X, Julia 0.4-rc2, because into which I though was related to this. The timings are now

0.198072 seconds (1.00 M allocations: 22.888 MB, 27.51% gc time)
0.198673 seconds (1.00 M allocations: 22.888 MB, 27.98% gc time)
0.400316 seconds (3.97 M allocations: 68.180 MB, 27.95% gc time)
0.339683 seconds (3.97 M allocations: 68.180 MB, 22.49% gc time)

[rofinn]

FWIW, this is still a problem.

julia> @benchmark copy!(Array(Int, length($iv)), iv)
BenchmarkTools.Trial:
  memory estimate:  22.89 MiB
  allocs estimate:  1000003
  --------------
  minimum time:     95.546 ms (0.00% GC)
  median time:      104.093 ms (0.00% GC)
  mean time:        210.753 ms (30.82% GC)
  maximum time:     1.425 s (53.43% GC)
  --------------
  samples:          24
  evals/sample:     1

julia> @benchmark convert(Vector{Int}, $(iv.o))
BenchmarkTools.Trial:
  memory estimate:  52.93 MiB
  allocs estimate:  2968713
  --------------
  minimum time:     391.859 ms (14.19% GC)
  median time:      436.203 ms (18.82% GC)
  mean time:        441.488 ms (19.38% GC)
  maximum time:     502.386 ms (22.78% GC)
  --------------
  samples:          12
  evals/sample:     1

[rofinn]

Okay, I've narrowed it down to the py2array call in convert.

julia> @benchmark PyCall.py2array(Int, $A, $(iv.o), 1, 1)
BenchmarkTools.Trial:
  memory estimate:  45.30 MiB
  allocs estimate:  2968480
  --------------
  minimum time:     404.982 ms (17.03% GC)
  median time:      410.736 ms (16.95% GC)
  mean time:        410.792 ms (17.07% GC)
  maximum time:     421.355 ms (16.31% GC)
  --------------
  samples:          13
  evals/sample:     1