peclet-core 1.3.2__tar.gz → 1.4.0__tar.gz

{peclet_core-1.3.2 → peclet_core-1.4.0}/CITATION.cff

@@ -16,7 +16,7 @@ authors:
     orcid: "https://orcid.org/0000-0001-6099-3583"
     email: e.a.j.f.peters@tue.nl
     affiliation: "Eindhoven University of Technology"
-version: 1.3.2
+version: 1.4.0
 date-released: 2026-10-03
 license: MIT
 doi: 10.5281/zenodo.21132435

{peclet_core-1.3.2 → peclet_core-1.4.0}/CLAUDE.md

@@ -37,7 +37,7 @@ ctest --test-dir build --output-on-failure -LE bench   # 76 ctests (77 with `ben
 export PATH=/usr/local/cuda-13.2/bin:$PATH    # if the Kokkos install targets the CUDA backend
 cmake -S . -B build_kokkos -DPECLET_CORE_ENABLE_KOKKOS=ON \
   -DCMAKE_PREFIX_PATH=../extern/install/nvidia-cuda
-cmake --build build_kokkos -j && ctest --test-dir build_kokkos --output-on-failure -LE bench  # 92 ctests (93 with `bench`): + device halo / geometry / solver / VoF PV fit, np=1,2,4,8
+cmake --build build_kokkos -j && ctest --test-dir build_kokkos --output-on-failure -LE bench  # 96 ctests (97 with `bench`): + device halo / geometry / solver / VoF PV fit / Anderson, np=1,2,4,8
 mpirun -np 4 ./build/benchmarks/bench_halo 48 1 300
 
 # Python modules + their ctests (test_mpi.py np=1,2,4,8; state_hash; ndarray interop). Release is
@@ -199,6 +199,26 @@ Header-only under `include/peclet/core/`:
   `ot_optimizer.hpp` and the AMR package are the consumers. Gate: `tests/test_csr_solver.cpp`
   (Kokkos build) — host row kernel vs device matvec, proper colouring of an asymmetric CSR,
   smoothers, BiCGStab + defect correction.
+  `solver/anderson.hpp` is **`AndersonCore`**, type-II Anderson acceleration of a steady march
+  (design: `../flow/doc/steady_acceleration.md`; register "AndersonCore lives in core from the
+  start"): grid-agnostic over an `AndersonState` (padded state views + roles Velocity / Carried +
+  ghost width; the metric is the velocity only, the revision-0 Pressure role, SDF mask, `cP` and
+  gauge flag were deleted in WO-3b, rev 1's `innerTolerance` in WO-3c), split `prepare` /
+  `complete` around the caller's own step, history `(2m+3)·n_s·8·n_pad` bytes. MPI-free: the
+  collectives are two callables in `AndersonComm`; `solver/anderson_mpi.hpp` builds them from an
+  `MPI_Comm` and is the MPI side. Rev 2 of the design: **no instability guard** — no Ritz
+  estimate, no status "unstable" (`Status` is Active / Disabled); a Ritz value of a non-normal map
+  is no stability test, and stability evidence comes from the caller's plain steps. Pass 2 reduces
+  RR and b only (2·columns + 2 doubles; no reduction reads dG). A restart (ρ > 4·ρ_min at a mixed
+  iterate) **restores the last map output** on every rank (keyed on the broadcast decisions) and
+  commits nothing of the rejected evaluation, so the state never holds a rejected output (review
+  R2). Gates: `tests/test_anderson.cpp`
+  (U1a–c, U2–U6, U4 converges on an unstable map and judges nothing, U4b a stable non-normal map,
+  U10 Carried fields, U11 restarts at a noisy near-neutral Carried mode never leave a rejected
+  output, metric, memory formula) and `tests/test_anderson_mpi.cpp` (U7, np 1/2/4: γ
+  bitwise on all ranks, np = 1 bit-identical to serial; R6: a descriptor invalid on one rank
+  throws on every rank); both print a per-test `digest` of every
+  iterate for bit-identity checks across builds.
 - **`amr/` is gone (2026-09-10)** — the whole AMR tree (the block-local-Morton octree, the distributed
   octree with leaf halos and weighted-ORB rebalancing, the collocated-projection cut-cell
   Navier–Stokes solver, its tests, studies, docs and campaign logs) is the **`peclet-amr`** package,

{peclet_core-1.3.2 → peclet_core-1.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: peclet-core
-Version: 1.3.2
+Version: 1.4.0
 Summary: peclet.core — compatibility shell: peclet.core.geom is now peclet.geom, peclet.core.mpi is now peclet.halo
 Project-URL: Homepage, https://github.com/computational-chemical-engineering/peclet
 Project-URL: Documentation, https://computational-chemical-engineering.github.io/peclet/
@@ -112,7 +112,10 @@ after `cmake --install`, or as a sibling checkout in the suite.
   `face_csr.hpp` (host+device row kernels), `coloring.hpp` (greedy symmetrised graph colouring),
   `csr_operator.hpp` (the device operator, Jacobi and multicolour Gauss–Seidel sweeps),
   `csr_bicgstab.hpp` (preconditioned BiCGStab / defect correction) and `vector_ops.hpp`. Consumed by
-  voro's mesh optimiser and by peclet-amr.
+  voro's mesh optimiser and by peclet-amr. Since 1.4.0 also `anderson.hpp`: **Anderson acceleration
+  of a steady march** (`AndersonCore`, type-II, grid-agnostic, on device; the caller owns the
+  fixed-point map, one time step), MPI-free with its collectives built by `anderson_mpi.hpp`.
+  Built for flow's `march_to_steady`.
 - **Python bindings** (`python/halo_bindings.cpp`) — a host-only **nanobind** module (no Kokkos):
   **`peclet.halo`** exposes the Lagrangian halo (`ParticleMigrator`, `ParticleHalo`: migration /
   ghosts / count-weighted rebalance) to an mpi4py driver. The shared zero-copy Kokkos
@@ -133,7 +136,7 @@ after `cmake --install`, or as a sibling checkout in the suite.
 
 Validated end-to-end by distributed explicit heat-diffusion solvers (plain, and **around an SDF solid
 obstacle**) matching a serial reference cell-for-cell across ranks, and consumed by the validated
-`flow` and `dem` distributed solvers. 72 ctests in the plain host+MPI build, 87 with Kokkos (`np` 1–8),
+`flow` and `dem` distributed solvers. 76 ctests in the plain host+MPI build, 96 with Kokkos (`np` 1–8),
 plus 6 Python ctests in the `python/` build.
 
 ## Build / test / benchmark

{peclet_core-1.3.2 → peclet_core-1.4.0}/README.md

@@ -92,7 +92,10 @@ after `cmake --install`, or as a sibling checkout in the suite.
   `face_csr.hpp` (host+device row kernels), `coloring.hpp` (greedy symmetrised graph colouring),
   `csr_operator.hpp` (the device operator, Jacobi and multicolour Gauss–Seidel sweeps),
   `csr_bicgstab.hpp` (preconditioned BiCGStab / defect correction) and `vector_ops.hpp`. Consumed by
-  voro's mesh optimiser and by peclet-amr.
+  voro's mesh optimiser and by peclet-amr. Since 1.4.0 also `anderson.hpp`: **Anderson acceleration
+  of a steady march** (`AndersonCore`, type-II, grid-agnostic, on device; the caller owns the
+  fixed-point map, one time step), MPI-free with its collectives built by `anderson_mpi.hpp`.
+  Built for flow's `march_to_steady`.
 - **Python bindings** (`python/halo_bindings.cpp`) — a host-only **nanobind** module (no Kokkos):
   **`peclet.halo`** exposes the Lagrangian halo (`ParticleMigrator`, `ParticleHalo`: migration /
   ghosts / count-weighted rebalance) to an mpi4py driver. The shared zero-copy Kokkos
@@ -113,7 +116,7 @@ after `cmake --install`, or as a sibling checkout in the suite.
 
 Validated end-to-end by distributed explicit heat-diffusion solvers (plain, and **around an SDF solid
 obstacle**) matching a serial reference cell-for-cell across ranks, and consumed by the validated
-`flow` and `dem` distributed solvers. 72 ctests in the plain host+MPI build, 87 with Kokkos (`np` 1–8),
+`flow` and `dem` distributed solvers. 76 ctests in the plain host+MPI build, 96 with Kokkos (`np` 1–8),
 plus 6 Python ctests in the `python/` build.
 
 ## Build / test / benchmark

{peclet_core-1.3.2 → peclet_core-1.4.0}/docs/Doxyfile

@@ -5,7 +5,7 @@
 
 DOXYFILE_ENCODING      = UTF-8
 PROJECT_NAME           = "core"
-PROJECT_NUMBER         = 1.3.2
+PROJECT_NUMBER         = 1.4.0
 PROJECT_BRIEF          = "Shared MPI block decomposition, asynchronous ghost-layer exchange, SDF geometry, load balancing (header-only C++20)"
 OUTPUT_DIRECTORY       = docs
 HTML_OUTPUT            = html