Give access to non-allocating combinations iterator

[mschauer]

This gives access to the non-allocating Combinations internals such that e.g. a copy of the changing state can be made with SVector.

Old:

julia> @btime collect(combinations(1:50,5))
  883.211 ms (8475043 allocations: 695.09 MiB)

julia> function f()
           i = 0
           for c in combinations(1:50, 5)
               i = i + c[1]
           end
           i
       end
f (generic function with 1 method)

julia> @btime f()
  1.079 s (8475041 allocations: 678.92 MiB)
18009460

New:

julia> @btime collect(combinations(1:50,5))
  303.784 ms (6356283 allocations: 404.12 MiB)

julia> @btime f()
  246.004 ms (6356281 allocations: 387.96 MiB)
18009460

julia> function g()
           i = 0
           for c in Combinatorics.Combinations(50, 5)
               i = i + c[1]
           end
           i
       end
g (generic function with 1 method)

julia> @btime g()
  71.276 ms (2118761 allocations: 64.66 MiB)
18009460

julia> @btime map(SVector{5,Int},Combinatorics.Combinations(50,5))
  100.676 ms (2118764 allocations: 145.48 MiB)

Actually, I do not see why it should still allocate in each iteration in g, maybe you see something or one should sprinkle some inbounds.

[Moelf]

why does it still allocate 😭 in each iteration

[mschauer]

Does it still?

[Moelf]

julia> function f(N)
           for c in Combinations(N, 4)
           end
       end
f (generic function with 1 method)

julia> @benchmark f(10)
BenchmarkTools.Trial: 10000 samples with 9 evaluations.
 Range (min … max):  2.305 μs … 142.252 μs  ┊ GC (min … max): 0.00% … 96.89%
 Time  (median):     2.609 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   2.851 μs ±   4.948 μs  ┊ GC (mean ± σ):  6.16% ±  3.49%

  ▃▅▅▃▁▁▅▇█▇▅▂                            ▁▂▂▂                ▂
  █████████████▇██▇██▇▆▆▅▅▁▄▄▃▃▃▄▃▁▄▃▁▄▆▇██████▆▅▁▅▄▅▅▄▄▁▅▆▇█ █
  2.31 μs      Histogram: log(frequency) by time      4.43 μs <

 Memory estimate: 6.66 KiB, allocs estimate: 211.

julia> @benchmark f(20)
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):  48.622 μs …  1.510 ms  ┊ GC (min … max): 0.00% … 95.13%
 Time  (median):     53.622 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   58.645 μs ± 82.108 μs  ┊ GC (mean ± σ):  7.96% ±  5.47%

  ▄▆▄▂▁ ▅█▇▅▃▁▁                                               ▂
  ██████████████▇▇▇▇▆▆▅▄▃▁▁▁▁▁▁▁▁▁▁▁▁▃▃▁▁▁▁▃▄▁▁▃▃▅▆█▇▇▅▄▃▅▇██ █
  48.6 μs      Histogram: log(frequency) by time      86.9 μs <

 Memory estimate: 151.50 KiB, allocs estimate: 4846.

notice the capital Combination, I use it combined with @view in reality

[mschauer]

The iterator state is a vector or nothing, I don't know if small unions with vectors can be stack allocated.

[Moelf]

isn't the problem this line always heap allocate?

comb = [c.a[si] for si in s]

[mschauer]

Can you check if #125 helps for your use case?

[Moelf]

yep, it does help with the above toy example.