Optimise integer division

[neworderofjamie]

One of the issues with code GeNN generates is that there are a lot of index calculations. A sizeable proportion of these involve dividing by integers which are known at code generation time but not at compile time. This means we miss out on compiler optimisations here which will turn them into essentially fixed point multiplication by the reciprocal. Luckily, the magic required has been encapsulated in libdivide. Benchmarking libdivide's branchless algorithms shows them to result in decent speedup in divide-heavy code.

TODO:

• Branchless version does not support dividing by 1 - probably need to test:
  • If all merged fields are 1 - don't divide at all
  • If no merged fields are 1 - use branchless fast divide
  • If mixed, don't use fast divide
• Make fast dividing backend-specific

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 85.93%. Comparing base (ea3f9ab) to head (b8919d4).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #645      +/-   ##
==========================================
- Coverage   87.83%   85.93%   -1.91%     
==========================================
  Files         104        5      -99     
  Lines       14104     1045   -13059     
==========================================
- Hits        12388      898   -11490     
+ Misses       1716      147    -1569     

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