Xarray optimizations

Brings the baseline runtime for the most simple (single link) 365-days simulation from 850 ms to 180 ms (no unit conversion) / from 870 ms to 430 ms (with unit conversion).

Push&pull with numpy data goes from 2 ms to 1.4 ms, for xarray data from 2 ms to 0.37 ms.

• Avoid conversion in to_units() if not reqiured
  • No-op from 400us to 16us
• Assign time in assign_time() inplace if axis exists
  • From 800us to 25us
  • From 37% of simple run to 3%
  • Breaks indexing with time values via xdata.loc[t]
    • Will leave as is for now, as we don't use indexing anywhere, and make an issue for it.
• Add a simplified implementation for to_datetime, without all the overhead
  • From 260us to 25us
  • From 20% of simple run to 3%

assigned to @mlange

changed target branch from main to test-outputs-copy

changed the description

removed performance label

marked this merge request as draft

added 1 commit

• 3d707e3a - assign time inplace if axis exists: 800us -> 25us

Compare with previous version

changed the description

added 1 commit

• 0cffe81e - simplify and optimize units extraction tools

Compare with previous version

changed the description

added 1 commit

• d089edd4 - use a faster custom implementation for to_datetime

Compare with previous version

mentioned in issue #88 (closed)

added 1 commit

• 6e6ae766 - add benchmarks for units comparison functions

Compare with previous version

added 1 commit

• fa56edfa - slightly optimize tool function `equivalent_units`

Compare with previous version

changed the description

added 1 commit

• 63d9b3a9 - add profiling and benchmarks for runs with numpy arrays

Compare with previous version