[finsberg]

Summary

Fixes the three test_discrete_gradient[*-False-tetrahedron] failures ("nan location mismatch") on the DOLFINx nightly leg. The bug was in the test, not in scifem.interpolation.interpolation_matrix.

The MatrixCSR branch of the test built one input vector on the column index map of DOLFINx's discrete_gradient, and passed it to the mult of both matrices. MatrixCSR.mult indexes the vector with the matrix's own column numbering, so the vector must be on that matrix's column map. On nightly the two maps differ:

• dolfinx.fem.discrete_gradient uses V's map as is (e.g. 8 ghosts).
• dolfinx.fem.create_matrix, which interpolation_matrix uses, adds ghosts in a different order (9–10 ghosts; at degree 3, 117 vs 81).

So scifem's matrix read past the end of the vector: column index 241 into a vector of length 206 on one rank. Those columns have zero coefficients, so the product was usually right. It became NaN whenever the memory read happened to hold one, which is why the failures were intermittent and only appeared under MPI.

The test now builds u on each matrix's own column map.

Testing

Locally on DOLFINx nightly, against scifem main:

• test_discrete_gradient 5 times each at mpirun -n 3 and -n 4: the old test failed in 9 of 10 runs, the fixed test passed all 10.
• tests/test_interpolation.py passes serially and at 4 ranks.
• pre-commit passes.

🤖 Generated with Claude Code