`function_(gradient/value)_(jump/average)` allocates when `u` is a view

[AbdAlazezAhmed]

It seems like getting a view of a view that's not indexed by a range allocate. I ran into this when doing error estimation for AMR, it had a lot of allocations and I didn't suspect this could be the issue back then.

using Ferrite

using BenchmarkTools

grid = generate_grid(Hexahedron, (1,2,1));

ip = Lagrange{RefHexahedron, 1}();

fqr = FacetQuadratureRule{RefHexahedron}(2);

iv = Ferrite.InterfaceValues(fqr, ip);

reinit!(iv, first(InterfaceIterator(grid)));

u = rand(100);

ui = @view u[rand(1:100, 16)];

@benchmark function_gradient_jump($iv, $1, $ui)

BenchmarkTools.Trial: 10000 samples with 987 evaluations per sample.
 Range (min … max):  49.479 ns …  48.324 μs  ┊ GC (min … max):  0.00% … 99.78%
 Time  (median):     63.649 ns               ┊ GC (median):     0.00%
 Time  (mean ± σ):   95.356 ns ± 569.371 ns  ┊ GC (mean ± σ):  17.39% ±  4.06%

  ▅█▇▄▃▃▁▁            ▃▄▄▅▅▄▅▅▄▃▂▁▁▁▁▁                         ▂
  █████████▇▇▅▇██▇▇▆▆▆██████████████████▇█▇▇█▇▆▅▄▆▅▄▄▄▅▄▄▄▅▄▄▅ █
  49.5 ns       Histogram: log(frequency) by time       170 ns <

 Memory estimate: 256 bytes, allocs estimate: 4.

ui = @view u[40:56];

@benchmark function_gradient_jump($iv, $1, $ui)

BenchmarkTools.Trial: 10000 samples with 995 evaluations per sample.
 Range (min … max):  26.543 ns … 547.550 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     28.528 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   28.915 ns ±  10.339 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

   ▁▁  ▂▄▁  ▄▆▂  ▆█▅   ▁                              ▁        ▁
  ▆██▂▃███▃▃███▅████▇▇▆██▆▅▆▆█▆▆▆▆▇▆▅▆▅▆▅▅▄▄▅▅▇█▅▅▃▅▅▇█▄▃▄▅▃▇█ █
  26.5 ns       Histogram: log(frequency) by time      33.9 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

ui = @view u[collect(40:56)];

@benchmark function_gradient_jump($iv, $1, $ui)

BenchmarkTools.Trial: 10000 samples with 986 evaluations per sample.
 Range (min … max):  49.354 ns …  48.101 μs  ┊ GC (min … max):  0.00% … 99.72%
 Time  (median):     91.593 ns               ┊ GC (median):     0.00%
 Time  (mean ± σ):   97.412 ns ± 570.019 ns  ┊ GC (mean ± σ):  16.81% ±  3.94%

  ▆█▇▅▃▃▂▁             ▂▅▄▅▆▅▅▆▆▅▃▂▂▂▂▁▁▁                      ▂
  █████████▇▇▆▇███▇▆▇▇▆█████████████████████▇█▇▇▆▇▆▆▆▅▄▆▅▃▅▃▂▄ █
  49.4 ns       Histogram: log(frequency) by time       163 ns <

 Memory estimate: 256 bytes, allocs estimate: 4.

This might not be an issue if one uses a vector instead of views, I just wanted to document this (and maybe disallow views to be passed to a function that creates views).

[KristofferC]

I'm quite sure this can be fixed.

[termi-official]

Not sure this can be fixed.

julia> @benchmark view($ui, 1:8)
BenchmarkTools.Trial: 10000 samples with 995 evaluations per sample.
 Range (min … max):  30.111 ns …  19.371 μs  ┊ GC (min … max):  0.00% … 99.20%
 Time  (median):     38.207 ns               ┊ GC (median):     0.00%
 Time  (mean ± σ):   84.893 ns ± 481.706 ns  ┊ GC (mean ± σ):  27.57% ±  5.11%

  █▇▇▆▅▄▂▁                 ▁▁▁▁▂▁▁▂▂▂▃▄▄▅▅▅▄▃▂▂▂▂▂▂▂▁▁         ▂
  ██████████▇███▇▇▇▇██████████████████████████████████████▇▆▇▆ █
  30.1 ns       Histogram: log(frequency) by time       132 ns <

 Memory estimate: 128 bytes, allocs estimate: 2.

[termi-official]

Okay, I guess one fix would be to inline the inner call and therefore not create the (allocating) inner view in first place. Not sure this is worth the hassle, as your data vector is now potentially scattered across the memory and thereby potentially introducing cache misses in this hot loop.

[KristofferC]

Without looking at any code, my thought was that there wouldn't be a need to create a view in the first place.

[termi-official]

I marked this as a good first issue for now. As we have pinned down the issue potential contributors can grab this as a refactor to explore the InterfaceValues .