Performance improvements for sparse triu!, tril!, dropzeros!, droptol!

[KristofferC]

This changes the anonymous functions used in the calls to fkeep! made by droptol!, dropzeros!, triu! and tril! to instead use functors.

Performance differences below are with this benchmark:

A = sprand(10^5, 10^5, 0.001)
B = copy(A)

import Base.droptol!
import Base.dropzeros!

B = copy(A);
@time droptol!(B, 0.1);

B = copy(A);
@time dropzeros!(B);

B = copy(A);
@time triu!(B);

B = copy(A);
@time tril!(B);

Master

 605.263 milliseconds (49914 k allocations: 762 MB, 7.72% gc time)
 422.887 milliseconds (29908 k allocations: 456 MB, 8.26% gc time)
 376.937 milliseconds (29908 k allocations: 456 MB, 9.14% gc time)
 374.099 milliseconds (29908 k allocations: 456 MB, 9.38% gc time)

PR

  32.674 milliseconds (4 allocations: 144 bytes)
  31.266 milliseconds (4 allocations: 144 bytes)
  26.735 milliseconds (38 allocations: 2416 bytes)
  27.053 milliseconds (38 allocations: 2416 bytes)

[ssfrr]

Wow, super improvement. Is there any documentation on this technique, perhaps it could live in the Performance Tips section of the docs? I found a couple discussions of similar things in the julia-users list and I see some places where it's used in base but I'm having trouble piecing together exactly what's going on and how/why it's different from just defining the functions as functions.

[KristofferC]

The README in https://github.com/lindahua/NumericFuns.jl has a little info about functors and why they sometimes (currently) are needed in Julia.

[dmbates]

Great idea. When I wrote that section of csparse.jl I was imitating the C code and wasn't aware of Functors. Obviously this is a good use of Functors.

[ViralBShah]

The sparse function already does the same too. Great idea. Can merge once appveyor completes.

[KristofferC]

@ViralBShah Funny coincidence, I made the issue on the previously used anonymous functions in sparse #10694.

At that point I didn't know enough Julia to fix it myself though.

[KristofferC]

Tests passing so should be fine to merge.