First attempt at embarrasingly parallel execution to compute head grids.

[dbrakenhoff]

Early result suggests ~10x speedup on my machine.

Todo

• Implement to_numba_tuple() for all element families intended for Numba path.
  Add fast potinf and potential kernels for remaining element classes. Missing:
  • CircAreaSink
  • LeakyWall
  • LeakyWallString
  • WellStringBase
• Support aquifer-region aware evaluation in Numba backend. Reason: model currently
  selects aquifer by location; backend now uses a single aqtuple and ignores aq_id.
• Preserve Model feature parity in Numba headgrid path. Must include: layers
  selection behavior, steady contribution, and output shape conventions.
• Add Numba-optimized discharge-vector path: disvec and velocity grids still route
  through Python loops. Write fast parallel implementations for disvec_inf, disvec
  functions.
• Testing
  • Add correctness tests comparing serial vs Numba-parallel across element mixes.
  • Add regression tests for edge cases that often break Numba kernels. Examples:
    points at well radius, near singular line endpoints, zero/very small times, many
    intervals, mixed g/v/z elements.
• Documentation
  • Add help for contributors describing how to add new elements to fast numba code.

[dbrakenhoff]

Parallel example (based on example from Discussion #115)

parallel_example.zip

Note that first run of numba code triggers the compilation step, which means it runs in about ~15s on my machine. The next run takes about ~3s. Normal timflow (using parallel=True, referring to multithreading) is about ~30s.

EDIT: example won't run until #139 is merged into dev and subsequently this branch.

[eriktoller]

Cool work and awesome to see such speed-up already!

Some thoughts:

• It could be a good idea to skip fastmath=True, at least when still developing the code
• it would be interesting to look at the difference if you change results_arr=bessellsv2(...) to bessellsv2(restuls_arr,...) and reuse a work array rather than creating a new one per linesink per point (300 linesink for 100 by 100 grid will be a lot of time for memory allocation).
• it would also be cool if you could time the pre-processing, computations and post-processing of the numba approach to see where the majority of the effort goes into.

I will follow the progress with great interest, great work Davíd 🎉