Skip to content

main train 2026-11-16: RT0 and BDM1 spaces (C8), hex ghost edge row and abort message in the partitioner (C32), three nightly test fixes (C3), index types in the remaining mprans kernels (C33) - #1365

Merged
cekees merged 31 commits into
proteustoolkit:mainfrom
cekees:train/main-2026-11-16
Nov 16, 2026

Conversation

@cekees

@cekees cekees commented Nov 16, 2026

Copy link
Copy Markdown
Contributor

Each change below was reviewed, and passed CI, as its own pull request on cekees/proteus. This branch merges them onto main at 9c40a3b with one merge commit per pull request, in the order listed.

Merge with "Create a merge commit", not squash or rebase, so the commits keep their hashes and the fork's pull requests close themselves.

pull request title commits
cekees#289 C8 round 5: RT0 and BDM1 as FemTools spaces on simplices, checked against the post-processed velocities 2
cekees#290 C32 round 5: macOS abort message, hex ghost edge row, partitioning.h without hdf5.h 4
cekees#291 C3 round 9: partition- and platform-independent tests from the fleet nightly 3
cekees#292 C33 slice 2c part 3: remaining mprans kernel families take proteus_index_t / proteus_csr_index_t 18

Measured, per pull request:

  • make sure tests run with '--isolation' #289 adds proteus/HdivSpaces.py and four classes appended to FemTools.py; no existing line changes and nothing uses the new spaces yet. Against an independent sympy oracle the bases agree to better than 1e-13; the normal jump for generic dofs is 1e-14; interpolation rates are RT0 1.00 and BDM1 1.99 on triangles, 0.99 and 1.97 on tetrahedra. Eight variants of the velocity post-processor convert to the new spaces: the two elements of a face give the same dof to 2e-16, the FemTools field equals the post-processor's to 4e-15, and 2 ranks equal serial dof by dof to 5e-13. Found, not fixed, and pinned by a test: the RT0 velocity of the p1-nc variant is not single valued on interior faces of elements that touch the boundary (67 of 408 faces, up to 5.6e-3 on the 13 x 13 problem). Limits: simplices, affine elements, RT0 and BDM1 only.
  • [WIP] branch with scorec adaptivity #290 changes partitioning.cpp, partitioning.h and tests. A hexahedral subdomain held one or two ghost edge rows (0,0) (4^3 hexahedra, one layer: 2 on 2 ranks, 4 on 4 ranks); now none, and the union of the ranks' edges equals the serial mesh's (new 12-case test; the six stored hex partition digests are re-recorded, the 30 simplex digests are unchanged). partitioning.h no longer includes hdf5.h. abortWith waits 0.2 s between writing its message and MPI_Abort, for the macOS run of 2026-11-15 in which no_rank_holds had empty output; that failure was not reproduced locally (0 of 160), so the next scheduled macOS run is the check. 424 partition tests pass.
  • [WIP] Refactor MeshTools #291 changes four files under test/ only. test_dispatch_run[dam3Bumps]: a mesh node sat exactly on the dam, where the deck's x <= 16 and the spec's heaviside(0) = 1/2 disagree; the dam moves to 16.05 in both and the comparison is still bitwise. test_without_petsc4py[3-dgp1]: the partition was not pinned, and ParMETIS left one of three ranks without elements; pinned, 24 of 24 pass. test_levee_edge_flux_with_implicit_flow_fct is marked xfail (not run) on Linux pip builds: Richards with implicit FCT gives an all-NaN Jacobian there and SuperLU crashes in dgstrf. That is a code defect, reported and not fixed here.
  • [WIP] Refactor Domain #292 changes 35 files, all under proteus/mprans/, and is types only: every remaining mprans kernel family (TADR, VOS3P, RDLS, Kappa, Dissipation, Pres/PresInc/PresInit, AddedMass, RANS2P, RANS2P_IB, RANS3PF, RANS3PSed, SW2D, MoveMesh, CLSVOF, MCorr, MCorr3P, SW2DCV, GN_SW2DCV, with their 2D variants) now takes proteus_index_t / proteus_csr_index_t; flags, tags and face-local numbers stay int. One commit per family. Checked per commit: clang -Wshorten-64-to-32 clean over mprans/*.cpp at index/csr widths 64/32, 64/64 and 32/64; g++ 11.4 -Wall at 32/32, 64/32 and 64/64 with pre-existing warnings only; the family's tests, serial against 2 ranks where a partition test exists; no pinned value moved. The build that ships is still 32-bit, and the full matrix passed on it. Limits: no real 64-bit build or full suite was run; richards/*.h, proteus/ADR.h, experimental/adrh and the meshio calls in Domain.cpp / mesh.cpp do not yet compile at index width 64; MeshSmoothing's MPI buffers stay int; RANS3PSed has little test coverage (26 serial tests).

The four touch disjoint files and merge cleanly. #290's abort-path change is aimed at the no_rank_holds cases that #291 leaves alone. 2.0 only: none of this goes into 1.9.x.

Diff against main: 54 files changed, 5873 insertions(+), 4495 deletions(-).

🤖 Generated with Claude Code

cekees and others added 30 commits November 14, 2026 16:05
…ce dof map, serial = 2 ranks

FemTools had no H(div) space. RT0, BDM1 and BDM2 exist only as
per-element coefficient arrays inside PostProcessingTools, each element
with its own copy of every face unknown. This adds the two lowest-order
spaces as spaces: one unknown per face (RT0) or nd per face (BDM1),
shared by the two elements. Nothing uses them yet; PostProcessingTools
is not touched.

proteus/HdivSpaces.py (new, numpy only):
  rt0    phi_f     = (x - P_f) / (d |K|)
  bdm1   phi_{f,a} = lambda_a (P_a - P_f) / (d |K|),  a a vertex of face f
  The dofs are the normal flux densities |F| (v.n) at the face centroid
  (RT0) or at the face's vertices (BDM1); the bases above are their dual
  bases, and sum_a phi_{f,a} = phi_f. Contravariant Piola map with
  |det J|, so the outward flux through a face is preserved for either
  orientation of the element.

proteus/FemTools.py (appended at the end; no existing line changed):
  HdivOnSimplex               local function space
  HdivInterpolationConditions the sample points
  FaceDOFMap                  face dofs with the global sign and order
  Hdiv_AffineOnSimplex        the space: basis values and divergences on
                              every element, evaluate, divergence,
                              elementFluxes, dofsFromElementValues,
                              interpolate

Sign and order. A dof is a flux density along the face's global normal,
which is fixed by the face's vertices sorted by global vertex number
(the edge turned clockwise in 2D, the cross product in 3D). BDM1's nd
dofs of a face are ordered by the global number of their vertex. Every
rank that holds the face gets the same normal and the same order.
Numbering: dpf * (global face number) + m, owned faces first, no
communication.

Oracle (test/fem_oracle/hdiv_tables.npz, 10 kB, its own generator and
file): a monomial generating set and an exact rational inverse of the
dof matrix in sympy, independent of the closed forms above.

Measured (macOS arm64, MPICH, FI_PROVIDER=tcp, -mat_partitioning_type
current):

  local spaces against the oracle (triangle, tetrahedron; values,
  divergences, normal flux on every face): below 1e-13; duality
  sigma_i(phi_j) = delta_ij 1e-14, on random affine elements of both
  orientations 1e-12

  dof map and space, 7x6-node triangle grid and 4x4x4-node tetrahedron
  grid (and an interval mesh), both families, element vertex lists as
  generated and scrambled per rank; serial and 2 ranks, node and element
  partitions, one layer of overlap:
                                   triangles        tetrahedra
    global dofs rt0 / bdm1         101 / 202        378 / 1134
    normal jump, generic dofs      8.0e-15          1.2e-14
    tangential jump (it is H(div))  >= 10            >= 41
    normal jump without the rule   >= 14            >= 34
    field of the space: values     6.7e-16          8.9e-16
      divergence                   5.1e-15          6.2e-15
      sum of element fluxes
        - |K| div v                8.0e-17          3.5e-17
    element faces whose local
      vertex order differs from
      the global one (2 ranks)     24, 26           180, 260
    2 ranks: the (point, face, value) table of the dofs of one global
    field equals the serial one to 1e-12; ranks that share a dof agree
    on its value to 1e-13 (so on the direction of every normal); owned
    blocks tile the global range

  interpolation of a smooth field, mesh and mesh with twice the
  intervals (L2 error of the field / of its divergence):
    triangles   rt0   8.96e-02 4.48e-02 rate 1.00 | div rate 1.00
                bdm1  2.05e-02 5.17e-03 rate 1.99 | div rate 1.00
    tetrahedra  rt0   1.90e-01 9.55e-02 rate 0.99 | div rate 1.00
                bdm1  8.46e-02 2.16e-02 rate 1.97 | div rate 0.98

Limits: simplices only; RT0 and BDM1 only; affine elements; the space
does not write the archive and no model or post-processor uses it. The
dofs are point values of the normal trace, exact on the space; a
moment-based interpolant of general fields (the commuting one) is not
provided.

Tests run locally: test_hdiv_spaces.py (26), test_hdiv_dofmap.py (8):
34 passed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…he new H(div) spaces, serial = 2 ranks

test/hdiv_vpp_mpi.py and test/test_hdiv_vpp.py. The helper runs
test/post_processing/vpp_partition_mpi.py unchanged (heterogeneous
Poisson, 13 x 13 nodes, 288 triangles, pinned partition), samples the
post-processor's element-wise velocity at the dof points of
Hdiv_AffineOnSimplex on the owned elements and builds the dof vector of
the FemTools space (RT0, or BDM1 for the -bdm variants). Tests only;
PostProcessingTools is not modified.

Measured (macOS arm64, MPICH, FI_PROVIDER=tcp), velocities of size 6:

  serial     the two elements   FemTools field   face fluxes   |K| div v
             of a face give     - post-proc.     - model's     - source
             different dofs     at 5 points/el
  pwl P1     1.7e-16            2.7e-15          3.9e-16       2.7e-14
  pwl P2     1.7e-16            3.6e-15          4.4e-16       1.2e-14
  pwl-bdm P1 2.2e-16            2.7e-15          4.4e-16       2.8e-14
  pwl-bdm P2 1.7e-16            3.6e-15          4.4e-16       1.1e-14
  pwc P1     1.7e-16            3.6e-15          3.9e-16       3.3e-14
  sun-rt0 P1 1.7e-16            2.7e-15          4.4e-16       6.7e-15
  dg P1      1.7e-16            2.7e-15          4.4e-16       1.2e-13
  dg-bdm P1  2.2e-16            2.7e-15          3.9e-16       1.2e-13
  (the last column equals the residual each variant leaves)

  2 ranks: the gathered global dof vector covers every dof (456 for
  RT0, 912 for BDM1); ranks agree where they overlap to 1.7e-14; it
  equals the serial vector dof by dof, matched by point and face:
    pwl P1 overlap 1 / 0       2.3e-13 / 3.2e-13
    pwl-bdm P1 overlap 1 / 0   3.3e-13 / 5.2e-13
    pwl P2 overlap 1           1.7e-13
    dg, dg-bdm P1 (element partition, overlap 1)  2.1e-13, 3.5e-13

So a post-processed velocity of these variants can be carried by one
shared unknown per face (RT0) or two per face (BDM1) without loss.

Found, not fixed (the post-processor's, C10): the RT0 velocity of the
p1-nc variant (Crouzeix-Raviart potential) is single valued in its
normal component only away from the boundary. On every face whose two
elements have no boundary node the two elements agree to round-off; on
interior faces of elements that touch the boundary they do not:
  smooth problem, 13 x 13: 67 of 408 interior faces, up to 5.6e-03
    (25 x 25: 1.5e-03; 49 x 49: 3.9e-04)
  low-K block problem (piecewise constant coefficient, no source):
    40 faces, up to 3.0e-03, all next to the Dirichlet sides
Every element still balances (1.7e-16). The variant is tested here as
it is (test_p1_nc_is_single_valued_only_away_from_the_boundary), so
that a fix shows up as a failure of that test.

Tests run locally: test_hdiv_vpp.py 16 passed in 15 s.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…rom a pair of nodes that is not an edge

The element-partition overlap looped over every pair of an element's nodes
(right for simplices, where every pair is an edge) and, for the neighbour
elements, read nodesEdgeMap_global_new[et] with operator[]. For a hexahedron
28 pairs, 12 of them edges: each of the 16 others was inserted into the map as
edge 0, every later find (the further layers, the owned-node stars) matched it,
and a rank other than 0 added edge 0 to its edges_overlap. Its nodes are not on
that rank, so nodeNumbering_global2subdomain[] inserted 0 for both:
subdomain.edgeNodesArray has the row (0,0) (found by the C8 session,
2026-11-14). find() now, like the other three lookups next to it.

New test_a_hexahedral_subdomain_holds_only_edges_of_the_mesh (12 cases: element
and node partitions, 1 and 2 layers of overlap, 2/3/4 ranks, 5^3 hexahedra,
meshPartition_hex_mpi.py): every row of every rank's edgeNodesArray joins two
nodes one grid step apart, and the ranks' edges together, by coordinates, are
the serial mesh's. -mat_partitioning_type current, 4^3 hexahedra, one layer:

                       serial   2 ranks              4 ranks
  before, bad rows     0        0,1                  0,0,1,1
  before, nEdges       144      144,128              104,144,128,94
  before, union digest 5639728234  fea08c23af (!= serial)  297c0e11a1 (!= serial)
  after,  bad rows     0        0,0                  0,0,0,0
  after,  nEdges       144      144,127              104,144,127,93
  after,  union digest 5639728234  5639728234        5639728234

The node partition (3 ranks, 2 layers, 5^3) was and is clean. The 30 simplex
digests are unchanged: no simplex pair is a non-edge.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…rs uses HDF5)

partitioning.cpp stopped using HDF5 when mappings.h5 went (round 4); its other
includers are cpartitioning.pyx (generated cpartitioning.cpp) and
cartesianPartitioning.cpp, and neither has an H5 call (grep H5|hdf5 over
partitioning.cpp/.pxd, cpartitioning.pyx, cartesianPartitioning.cpp: none).
build_ext --inplace on macOS arm64 builds every extension without it.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The fleet nightly 33659c3 failed test_the_spec_runs_the_deck_bitwise[dam3Bumps-1] and [-2] on all four test nodes
(spark2, ceds-model, chewbacca, darcy): 33 of 81 datasets differ (eta_t0, h_t0 at one node, 0.9375; every later
output follows from it). Cause, measured on spark2 (nightly prefix dev_pip_mumps, wheel 33659c3, one rank): the
unstructured mesh at he=2.0 holds a node at (16.0, 30) exactly, on the dam. The deck writes the jump as x <= 16
(h = 1.875 there), the spec's heaviside(16 - x) is 1/2 there (h = 0.9375); both runs read the same mesh (nodes equal).
The linux-64 conda mesh has no node on x = 16, so CI never saw it. Not a partition effect (one rank fails).

Both sides now put the dam at x = 16.05 (deck option dam_x_location, spec formula), so the comparison no longer
depends on where the mesher places a node. Still bitwise. Before: spark2 dam3Bumps 1 and 2 ranks failed;
after: pass (test_dispatch_run.py 8 passed, spark2, 79 s with test_without_petsc4py).

The deck/spec disagreement at the jump itself (<= versus heaviside(0) = 1/2) is a convention for C42/C5, not changed here.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The fleet nightly 33659c3 failed test_mgm_without_petsc4py_is_the_petsc_path[3-dgp1] on all four test nodes: with
the graph partitioner of the fleet's PETSc (pip, ParMETIS) under MPICH, the 3-rank dgp1 mesh leaves rank 2 with no
elements ("a discontinuous Galerkin space needs every rank to own elements ... rank 2 owns 0"), which
requireEveryRankOwnsElements refuses, correctly. The code is right; the test assumed a partition. test_parallel_space_
multigrid.py already pins `-mat_partitioning_type current` for the same reason (8aaac55); this file's launches did not.

_launch now adds the option to PETSC_OPTIONS for every case (the blocked and unblocked runs compare, so both are
pinned). Before: [3-dgp1] failed on spark2; after: all 24 cases pass on spark2 under the nightly prefix
(test_without_petsc4py.py + test_dispatch_run.py: 32 passed).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…OS no_rank_holds failure)

test_files_lookup_of_a_number_no_rank_holds_ends_the_run[range-entity 37 ...-4]
failed on the macOS Apple Silicon leg of the scheduled dev run of 2026-11-15
(run 43715513563): nonzero exit, stdout+stderr both EMPTY (the assertion
message shows '' after _without_launcher_banners), so the message was not split
or mangled, it never arrived. Round 4's 6ed51be fixed Open MPI's banner inside
the line; macOS runs MPICH, whose hydra launcher kills its proxies on MPI_Abort
and can exit before forwarding what the ranks wrote to their pipes.

It was the abort path that was wrong, not the test: a run ending in an error
must still say why. abortWith now sleeps 0.2 s after the write(2) and before
MPI_Abort (the run is ending; every abort path through abortWith pays 0.2 s).

NOT reproduced locally: macOS arm64, MPICH, FI_PROVIDER=tcp, the failing case
(meshPartition_missing_owner_mpi.py range 37, 4 ranks) 40 runs idle and 120
runs under 8 concurrent copies (32 ranks): 0 empty outputs before the change,
so the cause is inferred from the empty output and the launcher's behaviour,
not demonstrated. After: the 12 no_rank_holds cases pass (8.8 s). The oracle is
the next macOS leg of the scheduled dev run.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…vee test as a known code defect

The fleet nightly 33659c3: test_levee_edge_flux_with_implicit_flow_fct dies with SIGSEGV on spark2 (aarch64) and
ceds-model (x86_64), the Linux pip venvs; it passes on the macOS venvs and linux-64 conda CI. Not a test fault: a probe
on LinearSolvers.LU.prepare (spark2, nightly prefix) shows the Jacobian (n = 13419) finite for four Newton iterations,
then every entry NaN for five prepares, then finite with absmax 1.2e16, then SuperLU dgstrf -> dpivotL segfaults. The
final time of the first long step (0.1, 1, 10 s) changes nothing: ~50 s, crash, every time. Richards' implicit FCT
diverges on these builds and nothing refuses a non-finite matrix before the factorization. Owner: C0 (Richards.h);
reproduction is this test.

The test is xfail(run=False) on a Linux venv without conda-meta, so the nightly reports xfailed instead of failed and
the run does not crash. Before: spark2 1 failed (SIGSEGV); after: test_vdft_compatible.py 20 passed, 1 xfailed, 664 s.
CI (conda, linux-64) still runs the test.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… is gone

test_arrays_identical_to_the_mask_chain_partitioner[hex-element-{brick,scatter}-{3,4,8}]
failed on all four legs of #290 at a424fa8: the reference (recorded from the
2d21116 mask chain) contains the bug fixed in this PR, a non-edge node pair of
a hexahedron entered as edge 0 on every rank > 0 (row (0,0) of edgeNodesArray).
Only the edge arrays change. The 1-rank hex element digests and the other 42
references are unchanged: 89 passed locally (macOS arm64, MPICH, the digest,
graph-replay and empty-rank tests of test_partition_arrays.py).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…index_t

Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…proteus_csr_index_t

Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…teus_csr_index_t

Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype (index arrays and scalars read from ArgumentsDict, element/face/dof locals, numDOFs loop counters,
csr offsets and columns, function parameters incl. pyarray<int>& and the sparse-diffusion rowptr/colind). Flags, tags,
ebN_local*, *_kb* stay int. No .py dtype site needed (these kernels do not read internalNodesArray). Types only.

Checks (group 1 of C33 s2c-p3, all families together):
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors (only the pre-existing
  -Warray-bounds in equivalent_polynomials.h/RDLS/AddedMass).
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: clean at 32/32, 64/32, 64/64.
- macOS arm64 in-place build at 32/32; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_tadr_/test_vos_/
  test_twophaseflow_partition_independence (serial vs 2 ranks, tolerances unchanged), test_rank_interface_flux,
  test_cython_class_attributes, levelset/vortex2D/test_rdls_normal, AddedMass, salinity/test_tadr_edge_flux_recording:
  67 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…r_index_t

Name-class retype as in the group 1 commits, plus second-wave names found here: csr row loops (m = rowptr[I]),
nQuadraturePoints_global / sd_offset (updateSolidParticleTerms), left/right_eN_global and their *_nDOF_trial_element,
ebN_kb(_nSpace) built from a global ebN. Flags (isActiveElement, nSteps, isDOFBoundary_*, boundaryFlags, ...), tags and
local face numbers stay int. RANS2P.py: the sdInfo entries added after OneLevelTransport.__init__ use CSR_INDEX_DTYPE.
Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: only the pre-existing size_type warnings (shape() to int,
  h:181 and the MPI_Allreduce counts) and the -Wwritable-strings at h:2399/1928.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: only the pre-existing -Wwrite-strings, at 32/32, 64/32, 64/64.
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_twophaseflow_partition_independence
  (serial vs 2 ranks), test_rank_interface_flux, test_step_retry, rans2p_ib, solver_tests/test_nse_RANS2P_step,
  TwoPhaseFlow/test_rans2p_preconditioner, test_boundary_integral_gauges, test_hotstart_levelset: 96 passed, 2 skipped;
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype as in the group 1 commits, plus second-wave names found here: csr row loops (m = rowptr[I]),
nQuadraturePoints_global / sd_offset (updateSolidParticleTerms), left/right_eN_global and their *_nDOF_trial_element,
ebN_kb(_nSpace) built from a global ebN. Flags (isActiveElement, nSteps, isDOFBoundary_*, boundaryFlags, ...), tags and
local face numbers stay int. RANS2P_IB.py: same sdInfo dtype fix.
Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: only the pre-existing size_type warnings (shape() to int,
  h:181 and the MPI_Allreduce counts) and the -Wwritable-strings at h:2399/1928.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall: only the pre-existing -Wwrite-strings, at 32/32, 64/32, 64/64.
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_twophaseflow_partition_independence
  (serial vs 2 ranks), test_rank_interface_flux, test_step_retry, rans2p_ib, solver_tests/test_nse_RANS2P_step,
  TwoPhaseFlow/test_rans2p_preconditioner, test_boundary_integral_gauges, test_hotstart_levelset: 96 passed, 2 skipped;
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…csr_index_t

Name-class retype as in the RANS2P commits, plus: GlobPos_*/*_gi (global dof numbers), the *_ith_row_ptr/*_ij/
numDOFs_ith_row/colind_1D csr locals of the 1D-component matrix, surrogate_boundaries and surrogate_boundary_elements
(std::vector<proteus_index_t>), ebN_kb split from ebN_local_kb (a face-local number stays int). RANS3PF.py: rowptr_1D and
colind_1D are CSR_INDEX_DTYPE, as are the sdInfo entries added after OneLevelTransport.__init__. Flags, tags and local
face numbers stay int. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: only the pre-existing size_type and VLA-extension notes.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only the pre-existing -Wreorder/-Wwrite-strings.
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_twophaseflow_partition_independence
  (serial vs 2 ranks), test_cylinder_ibm_parallel, test_rans3pf_boundary_forces, test_hotstart_rans3pf,
  test_av_partition_independence, CLSVOF/with_RANS3PF, test_rank_interface_flux: 33 passed, 2 skipped; no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…eus_csr_index_t

Same name-class retype and second-wave names as the RANS2P/RANS3PF commits (left/right_eN_global blocks split from the
face-local numbers, sd_offset, csr row loops). RANS3PSed.py: the sdInfo entries added after OneLevelTransport.__init__ use
CSR_INDEX_DTYPE. Flags, tags and local face numbers stay int. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: only the pre-existing size_type and VLA-extension notes.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only the pre-existing -Wreorder/-Wwrite-strings.
- macOS arm64 32-bit build; pytest test/RANS3PSed (test_bed_sediment) and test_cython_class_attributes: 26 passed.
  This family has little test coverage (no partition-independence test, serial only); no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype as in the earlier commits. Flags and tags stay int. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors for the C++ kernels.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing -Wreorder/-Wwrite-strings.
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test/SWFlow (10 passed), test_sw2dcv_partition_independence, test_boundaryconditions, test_scatter_dof_array.
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… proteus_index_t

Kernels: name-class retype plus eN_nDOF_trial_element. MeshSmoothing.pyx/.pxd and cMoveMeshMonitor.pyx: the mesh index memoryviews (nodeStarOffsets, nodeStarArray, nodeElementOffsets, nodeElementsArray, elementNodesArray, elementBoundaries/BoundaryNodes/BoundaryElements/NeighborsArray, *_subdomain_owned, *_subdomain2global) and their node/element/variable-number scalars use proteus_index_t from index_type.h. Left int on purpose: flags (nodeMaterialTypes, fixedNodesBoolArray), MPI counts and displacements, and the MPI exchange buffers (result, nodes_2rank_values): changing those needs the MPI datatype changed with them (see handoff). The Cython-generated C++ compiles at index 64 with no errors; its remaining -Wshorten-64-to-32 warnings are Py_ssize_t conversions and local int counters.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors for the C++ kernels.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing -Wreorder/-Wwrite-strings.
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test/MoveMeshMonitor, test_sw2dcv_partition_independence, test_boundaryconditions, ci/test_meshPartitionCharacterisation, test_scatter_dof_array: 432 passed (with the SW2D set).
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype as in the earlier commits (C0 round 11 has landed). Flags and tags stay int. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing warnings (-Wreorder,
  -Wwrite-strings, the multi-line string literal at GN_SW2DCV.h:864).
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test/CLSVOF, test_clsvof_spinup_parallel, test_av_partition_independence, test_twophaseflow_partition_independence, test_step_retry: all pass (48 + 27 across the group 4 runs).
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype as in the earlier commits. isActiveElement and the other flags stay int. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing warnings (-Wreorder,
  -Wwrite-strings, the multi-line string literal at GN_SW2DCV.h:864).
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_mcorr_partition_independence (serial vs 2 ranks), test_twophaseflow_partition_independence: pass.
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype as in the earlier commits. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing warnings (-Wreorder,
  -Wwrite-strings, the multi-line string literal at GN_SW2DCV.h:864).
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_mcorr3p_jacobian, test_twophaseflow_partition_independence: pass.
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Name-class retype plus the helper signatures phiBar, muWallTranspose, wallCoefficients (i dof index; j, ij csr) and the offset_ji loop. SW2DCV.py: rowptr_cMatrix and colind_cMatrix are CSR_INDEX_DTYPE. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing warnings (-Wreorder,
  -Wwrite-strings, the multi-line string literal at GN_SW2DCV.h:864).
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_sw2dcv_partition_independence, test_sw2dcv_bathymetry_partition, test_sw2dcv_bathymetry_forms (serial vs 2 ranks): pass.
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Same as SW2DCV (helper signatures, rowptr_cMatrix and colind_cMatrix dtypes). The relaxation_zone_nodes read is commented out in the header, so nothing to retype there. Types only.

Checks:
- clang++ -Wshorten-64-to-32 at index/csr 64/32, 64/64, 32/64: 0 warnings, 0 errors.
- g++ 11.4.0 on ceds-model -fsyntax-only -Wall at 32/32, 64/32, 64/64: only pre-existing warnings (-Wreorder,
  -Wwrite-strings, the multi-line string literal at GN_SW2DCV.h:864).
- macOS arm64 32-bit build; pytest (FI_PROVIDER=tcp, -mat_partitioning_type current): test_gn_sw2dcv_partition_independence (serial vs 2 ranks): pass.
  no pinned value moved.
- Not run: a 64-bit-index build, the full suite.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… checked against the post-processed velocities

From cekees/c8-r5 at 7a873bb (#289).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…titioning.h without hdf5.h

From cekees/c32-r5 at d0b18c0 (#290).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…m the fleet nightly

From cekees/c3-r9 at 076a0cb (#291).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…e proteus_index_t / proteus_csr_index_t

From cekees/c33-s2c-p3 at 711f7e4 (#292).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@cekees
cekees merged commit 92f24ad into proteustoolkit:main Nov 16, 2026
9 checks passed
@cekees
cekees deleted the train/main-2026-11-16 branch November 16, 2026 10:32
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant