FEniCS / ufl / 32420098100

80%

push

github

Use the accessed sub-element's degree for indexed mixed-element components (#503)

* Use the accessed sub-element's degree for indexed mixed-element components

SumDegreeEstimator.argument()/.coefficient() return the whole element's
embedded_superdegree, including for a component of a MixedElement
extracted via Indexed. When a mixed space's sub-elements have different
degrees, every component was assigned the mixed element's max degree
instead of its own sub-element's -- inflating the estimated polynomial
degree, and hence the quadrature degree requested, for any form using
an undifferentiated lower-degree component (derivatives already reduce
degree correctly via _reduce_degree, so this only affected undifferentiated
use of a component, e.g. a pressure term in a Taylor-Hood-style form).

Fix indexed() to resolve the correct sub-element when the operand is an
Argument or Coefficient and the multi-index is fully fixed (not free):
flatten the multi-index into the operand's own component space, find
which sub-element that flat component falls in from the mixed element's
sub_elements() and their reference_value_size, and return that
sub-element's degree instead. Falls back to the previous (safe, if
imprecise) whole-element degree whenever a component can't be resolved
this way -- free indices, non-mixed elements, or indexing types other
than Argument/Coefficient.

Updates two pre-existing test assertions in test_degree_estimation.py
that had `# TODO: This should be N, but M is expected behaviour now`
comments documenting exactly this limitation -- both now assert the
corrected value.

For a Taylor-Hood Stokes form (P2 vector / P1 scalar mixed space), this
drops the estimated degree from 3 to 2, and the quadrature rule FFCx
requests from 5 points (with a negative weight) to the minimal 4-point
rule -- confirmed numerically identical (max abs difference 1.5e-15,
floating-point noise) since the true polynomial degree was 2 all along.
Measured 1.23x faster on... (continued)

22 of 22 new or added lines in 1 file covered. (100.0%)

10009 of 12546 relevant lines covered (79.78%)

0.8 hits per line