FEniCS / ffcx
89%

DEFAULT BRANCH: main
Repo Added
Files 60
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

LAST BUILD ON BRANCH main
branch: main
CHANGE BRANCH
x

coverage: 88.921% (+0.2%) from 88.75%
41411093991

push

github

web-flow 
Enable sum factorisation by default and improve O2 kernel generation (#906)

* Fix sum factorisation with several or one point quadrature rules

- 1D tensor factor tables were named FE_TF0, FE_TF1, ... from zero for
  each quadrature rule, so the tables of integrals with different
  quadrature degrees clashed and one rule read the other's tables, giving
  a wrong element tensor. Include the quadrature rule in their name.
- A piecewise table with tensor factors, such as a coefficient on a one
  point rule, is accessed outside the quadrature loop, where the
  quadrature index can't be split by direction, which failed an
  assertion. Its factors are constant over the points, so read point 0.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Make quadrature weights tensor products

* Add opt-in sum-factorised contraction for tensor product cell integrals

With `sum_factorization_contraction`, cell integrals on tensor product
rules cache the integrand on the whole rule and contract one direction at
a time, instead of updating every element tensor entry at every quadrature
point: O(N^2 Q^2 + N^4 Q) rather than O(Q^2 N^4) in 2D, and O(N^6 Q)
rather than O(Q^3 N^6) in 3D.

- The 1D quadrature weights are folded into the first argument's 1D
  tables at code generation time.
- Terms with the same dof loops share one loop nest; stage arrays and
  contributions shared by several terms are computed once, and each
  element tensor entry is updated once.
- Integrals whose buffers exceed a 256 KiB stack budget, or with an
  argument that doesn't factorise, keep the fused loop.
- Without the option, kernels use the fused loop with the full weights
  table, as before 59e849d6.
- Tensor factor tables are indexed by their own single permutation and
  entity.
- Unread element tables are pruned when the contraction is used.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Plan sum-factorised contractions once and declare only ... (continued)

425 of 426 new or added lines in 11 files covered. (99.77%)

13 existing lines in 4 files now uncovered.

5209 of 5858 relevant lines covered (88.92%)

0.89 hits per line

Relevant lines Covered
5858 RELEVANT LINES 5209 COVERED LINES
0.89 HITS PER LINE
Source Files on main

Recent builds

Builds Branch Commit Type Ran Committer Via Coverage
41411093991 main Enable sum factorisation by default and improve O2 kernel generation (#906) * Fix sum factorisation with several or one point quadrature rules - 1D tensor factor tables were named FE_TF0, FE_TF1, ... from zero for each quadrature rule, so the ... push web-flow github
88.92
40291595106 main Enable sum factorisation by default and improve O2 kernel generation (#906) * Fix sum factorisation with several or one point quadrature rules - 1D tensor factor tables were named FE_TF0, FE_TF1, ... from zero for each quadrature rule, so the ... push web-flow github
88.92
40228706075 main Enable sum factorisation by default and improve O2 kernel generation (#906) * Fix sum factorisation with several or one point quadrature rules - 1D tensor factor tables were named FE_TF0, FE_TF1, ... from zero for each quadrature rule, so the ... push web-flow github
88.92
40210089223 main Sum-factorise coefficient evaluation in sum-factorised contractions (#904) * Sum-factorise coefficient evaluation in sum-factorised contractions With `sum_factorization_contraction`, the integrand was cached on the whole tensor product rule, but... push web-flow github
88.75
40136004794 main Add documentation version switcher (#931) * Add documentation version switcher Enable the pydata-sphinx-theme version dropdown. The version list is generated by FEniCS/docs from the published documentation, and CI sets DOCS_VERSION to the branch... push web-flow github
88.1
40087880283 main Include the Basix and UFL versions in the JIT signature (#926) Generated kernels embed Basix's tabulated values and depend on UFL's form processing, but the cache key only had the FFCx version, so a Basix or UFL patch release could reuse stale ke... push web-flow github
88.1
40086869165 main Add a minimal CMake + C example (#918) * Add a minimal CMake + C example A small, self-contained example of calling an FFCx-generated kernel from C, with no other FEniCS component involved. CMake compiles poisson.py with FFCx at build time, and ... push web-flow github
88.1
39383040288 main Generate the numba test module once per module (#930) Follow-up to #911, which fixed the xdist race by generating into a pytest temporary directory but left the fixture function-scoped, so the Poisson module was regenerated for each of the four p... push web-flow github
88.1
39221885436 main Fix test_ds_prism entity_index dtype on big-endian platforms (#925) The kernel reads entity_index as int*, but the test passed int64 arrays, so on big-endian machines (s390x) the kernel read 0 instead of 1. Fixes #815 Co-authored-by: Claude Son... push web-flow github
88.1
39217432823 main Fix test_ds_prism entity_index dtype on big-endian platforms (#925) The kernel reads entity_index as int*, but the test passed int64 arrays, so on big-endian machines (s390x) the kernel read 0 instead of 1. Fixes #815 Co-authored-by: Claude Son... push web-flow github
88.1
See All Builds (3299)