Faster ExclusiveTopology construction

[KnutAM]

Had this as notes lingering around after seeing #617, nothing urgent or major.
Timings should be verified by someone since this is so heavy on GC it seems to be quite unreliable to time it and profiling changes from run to run...

Somewhat representative btime runs give

grid = generate_grid(Hexahedron, (40, 40, 40)) # more stable timings than 3x80
@btime ExclusiveTopology($grid);
528.067 ms (3582741 allocations: 378.69 MiB) # PR
580.268 ms (4094719 allocations: 464.89 MiB) # master

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.84%. Comparing base (1fcdef1) to head (b09ac4b).
Report is 2 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #927      +/-   ##
==========================================
- Coverage   93.86%   93.84%   -0.02%     
==========================================
  Files          36       36              
  Lines        5313     5296      -17     
==========================================
- Hits         4987     4970      -17     
  Misses        326      326              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[KristofferC]

Is there an advantage to making this ordered so the order in the cell_neighbor_table[cell_id] will not be "random"? Maybe it doesn't matter.

[termi-official]

Probably this does not matter, because we do not enforce any ordering and locality properties on our grids in general.

[termi-official]

I cannot see any significant difference between the PR and master unfortunately.

Master 966.810 ms (4094719 allocations: 464.89 MiB)
PR 956.878 ms (3582741 allocations: 378.69 MiB)

So a about 1% difference on my machine. We need to think about how to increase data locality and reduce allocations here if we want to see a significant speedup.

[KnutAM]

Thanks for checking it!