Fixes for PR717#1
Draft
hughcars wants to merge 43 commits into
Draft
Conversation
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
3 times, most recently
from
fe5c897 to
e6f2324
Compare
hughcars
changed the title
Add WavePortOperator::AddBoundaryMassBdrCoefficients and GetWavePortKn
helpers, plus SpaceOperator::GetWavePortBoundaryMassMatrix, exposing
the per-port μ⁻¹ boundary mass M^(p) and the scalar kₙ,p(ω) separately.
The existing AddExtraSystemBdrCoefficients now calls the new helper
internally, so behaviour and assembled operators are unchanged.
These helpers are foundational plumbing for upcoming changes to the
RomOperator: storing per-port projected M^(p)_r once and assembling
the wave-port contribution online as Σ_p kₙ,p(ω)·M^(p)_r recovers an
O(n²) online phase for adaptive sweeps with wave ports, and unblocks
circuit synthesis with wave ports.
Adds a unit test that compares Im{A2(ω)·v} to Σ_p kₙ,p(ω)·M^(p)·v on
the existing CPW wave-port example at three frequencies in the
configured sweep band, in serial and under MPI.
(cherry picked from commit 9248d8f)
The reduced-order driven sweep currently re-assembles the full HDM
A2(ω) and projects it onto the basis at every online frequency, which
defeats the asymptotic cost model of the reduced model: each online
solve was paying for one HDM-scale boundary form assembly plus
dim_V HDM-size mat-vecs.
Wave ports are the dominant A2 contributor in practice. Their
contribution factors as A_wp(ω) = i·Σ_p kₙ,p(ω)·M^(p)_{μ⁻¹}, with
kₙ,p the only ω-dependent factor (a scalar from the per-port
cross-section EVP). Store the per-port HDM operators in RomOperator,
project them once when the basis grows, and apply the wave-port term
online as Σ_p kₙ,p(ω)·M^(p)_r. Online cost per ω drops from
HDM-scale to O(n²).
To keep correctness for any non-wave-port A2 contributors
(second-order farfield, surface conductivity), add an
include_wave_ports flag to GetExtraSystemMatrix and call the
"excluding wave ports" path on the slow fallback. The default
behaviour is unchanged for all existing call sites.
Validated against the parent branch on cpw_wave_adaptive: port-S
agrees to ~1e-11 relative, domain-E to ~1e-13.
(cherry picked from commit 9938828)
(cherry picked from commit b5cdbd08bcf90ea3ed24c7c4da6c6e4eca35fab5)
…tion
Two post-rebase cleanups:
- test-waveportoperator.cpp now loads the cpw config and mesh from
PALACE_TEST_DATA_DIR/regression/input/cpw (staged via the regression
fixtures that symlink to examples/cpw and are dereferenced on install)
instead of a dedicated PALACE_SOURCE_EXAMPLES_DIR pointing at the source
tree. Drop the now-unused PALACE_SOURCE_EXAMPLES_DIR compile definition.
- Make the port index p a consistent subscript in the comments: the
per-port boundary mass is now written M_{μ⁻¹,p} (and M_{r,p}, M_p) to match
kₙ,p, rather than mixing a superscript M^(p) with the subscript kₙ,p.
Put both subscripts in one group so the port index p is a true subscript
(matching M_{μ⁻¹,p}) rather than a baseline ",p" after the Unicode kₙ. Also
render the standalone constant as k_n for a consistent brace-subscript style.
Comment-only.
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
4 times, most recently
from
f90412f to
851d8b0
Compare
Split wave port ExtraSystemMatrix term
Remove unusual superbuild combinations.
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
from
851d8b0 to
598ae7f
Compare
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
2 times, most recently
from
a3795eb to
60f5261
Compare
Bind untrusted github.* and inputs.* context values to env variables and reference them as shell variables in run: blocks, instead of interpolating them directly into the script text. This closes the remaining instances of the script-injection class flagged by AppSec (Talos) on main: - docs.yml: "Decide whether to build" (github.ref, github.event_name, steps.filter.outputs.test), "Build and deploy" (inputs.deploy_tag), and "Verify stable docs deployment" (inputs.deploy_tag || github.ref_name) - ghcr-cleanup.yml: both version-count steps (github.repository_owner) Completes the hardening started in #787 (which covered build-and-test-linux, long-tests, and long-test-status-manager). Follows GitHub's security-hardening guidance for untrusted input.
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
2 times, most recently
from
1cd11c9 to
79043ce
Compare
* Fix for Spack PR 52627 spack/spack#52627 changes how env is resolved. So, here we add PWD to search for the Dockerfile template in the current directory as well * Pin Spack to the 1.2 release series instead of develop Pin the Spack tooling version used by CI to the 1.2 release series (releases/v1.2, currently 1.2.1) instead of the moving develop branch, so CI builds are reproducible: - setup-runner: default spack-version and all callers (palace-ci, build-container, run-regression-tests) now check out releases/v1.2 - container spack.yaml: images.spack pinned to the prebuilt spack/ubuntu-noble:1.2 Docker Hub image - docs.yml: setup-spack@v3 gets spack_ref=releases/v1.2 and packages_ref=releases/v2026.06 (spack 1.2 uses the separate spack-packages repo); the old 'ref' key was a no-op for @V3 Binary-cache channel names (binaries.spack.io/develop and the palace-develop-testing GHCR cache) are left unchanged: they are cache identifiers, not tooling versions, and still resolve under Spack 1.2. * Add spack-develop lookahead job Expose a spack-version input on the palace-ci action (defaulting to the pinned releases/v1.2) and add a lookahead job that builds and tests with Spack develop. This gives an early signal when upstream Spack or spack-packages changes break Palace, without affecting the pinned CI.
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
2 times, most recently
from
5c76bcc to
5f205a9
Compare
Previously, MeasureSParameter() skipped S-parameter post-processing when a Floquet port coexisted with lumped or wave ports, but InitializePortS() did not replicate that condition, so port-S.csv was still written with the raw port-mode projections. The written values were wrong in two ways: - For lumped/wave-driven excitations, the incident subtraction (-1 on the diagonal) was never applied, so diagonal entries were 1 + S_ii instead of S_ii. - Floquet ports normalize the incident wave to unit time-averaged power (P_avg = 1), while lumped/wave ports normalize to unit peak power (integral of E x H* . n dS = 1, so P_avg = 1/2). Cross-type observations were therefore off by sqrt(2) in amplitude (3.01 dB in power). Support the mixed configuration by applying a sqrt(2) power normalization bridge to cross-type S-parameters: divide lumped/wave observations by sqrt(2) when a Floquet port drives, and multiply Floquet observations by sqrt(2) when a lumped/wave port drives. This restores S-matrix unitarity (power balance) and complex reciprocity for the mixed case, verified on Floquet + lumped and Floquet + wave port test structures (sum |S|^2 = 1 to better than 0.01% for lossless structures, S_ij = S_ji exactly). Floquet-only and lumped/wave-only configurations are unaffected: the bridge factor applies only to cross-type observations.
Two new regression cases that validate the sqrt(2) power normalization bridge for cross-type S-parameter observations: - floquet_lumped: periodic unit cell (1x1x4 cm) with a Floquet TM port and a lumped resistive port (R = 377 Ohm). Verifies S-matrix unitarity (sum |S|^2 = 1.000) and reciprocity for the lumped/Floquet pairing. - floquet_wave: open-ended waveguide array unit cell (2.5x2.5 cm periodic cell, 2x1 cm rectangular waveguide feed). TE10 mode propagates 8-11 GHz, single-mode on both port types. Verifies S-matrix unitarity and reciprocity for the wave/Floquet pairing. Both cases include mesh generation scripts (Julia + Gmsh), Palace config, and reference CSV data. Custom checks use the existing TestFloquetSParams helper for port-floquet-S.csv (magnitude-only, NaN-tolerant).
The CI arm64 runners use 64 MPI ranks. With only 128 elements (4x4x8), some ranks get empty subdomains (0 vertices), causing a fatal abort in the lumped port boundary setup. Increase to 432 elements (6x6x12) so every rank has at least 6 elements even at 64 ranks.
These small meshes (432 and ~600 elements) don't partition cleanly at high MPI rank counts (the CI arm64 runners use 64 ranks). With so few elements, METIS can produce subdomains that have no faces on a port boundary, causing the port operator setup to abort. The tests validate S-parameter power normalization correctness, not parallel scaling — the existing dielectric_grating_uniform test already covers Floquet ports in parallel.
This reverts commit aa68591.
…ning The regression meshes were too small for the CI arm64 runners' 64 MPI ranks: below ~7 elements per rank, METIS produces degenerate partitions (subdomains with 0 vertices) which crash in parallel mesh/boundary setup (segfault in the solve, or "non-periodic face cannot have multiple boundary elements" in GetFaceToBdrElementMap). The failure is independent of port type — it reproduces with plain PEC boundaries on the same mesh — and disappears at ~2000+ elements. Increase floquet_lumped to 8x8x20 (2112 elements) and floquet_wave to a finer aperture discretization (3076 elements). Both now pass at 48, 56, 60, and 64 ranks. Serial runtimes remain modest (7 s and 24 s). Power balance on the regenerated references holds to 1e-6.
Fix S-parameters when Floquet ports are used with lumped or wave ports
For pull_request events from forks, GITHUB_TOKEN is read-only regardless of the workflow's declared permissions, so the peaceiris/actions-gh-pages push to gh-pages fails with 403 and aborts the job before the regression test steps run (observed on PR #770, the coverage matrix row). Guard the deploy and the summary-link steps with the same same-repo condition already used by docs.yml for its GHCR cache push. The coverage report artifact upload is left unguarded since it works with a read-only token, so fork PRs still get a downloadable coverage report.
Skip gh-pages coverage deploy for fork PRs
Add a SurfaceRationalImpedanceOperator implementing a Robin boundary whose per-square surface impedance is a user-specified rational function of frequency, Zs(s) = N(s)/D(s) with s = iω. Numerator and denominator polynomial coefficients (zeros and poles) are given as lists in the new boundaries["RationalImpedance"] config block, allowing any passive lumped-network response. Being a general function of frequency, the boundary contributes iω/Zs(iω) to the frequency-dependent system matrix A2(ω) (like SurfaceConductivityOperator), and is available for the frequency domain driven problem type. Coefficients are nondimensionalized as a_k/(Z0*tc^k) and b_k/tc^k. Wiring spans configfile (RationalImpedanceData and parser), iodata (nondimensionalization, problem-type warnings, nonlinear-eigenmode sparsity-pattern handling), spaceoperator (assembly, auxiliary-space marker, multiple-BC check), the JSON schema, and the docs. Add a verification example (examples/rational_impedance) comparing, on a parallel-plate TEM line, a parallel-RLC sheet realized as a RationalImpedance against the same RLC realized as a LumpedPort; the two S11 responses agree to solver tolerance.
…undary
Validate user-specified rational surface impedances and warn on physically
questionable inputs, following the Palace Mpi::Warning convention:
- Warn when the numerator/denominator degree difference exceeds one, since a
positive-real (passive) impedance has |deg(N) - deg(D)| <= 1.
- Warn, once per boundary, when Re{Zs(iw)} < 0 at an evaluated frequency, with
a relative tolerance so lossless reactive terminations do not trigger it.
Real-valued coefficients (enforced by the JSON schema and parser) guarantee a
Hermitian Zs(iw), and hence a real time-domain response.
Add the RationalImpedance boundary attributes to the aggregated boundary attribute list built in BoundaryData. Without this, external rational-impedance boundaries raised a spurious "no associated boundary condition, PMC assumed" warning, and internal rational-impedance sheets were skipped by the interface-element (mesh cracking) logic, so the boundary term was not applied on them.
…iven-only restriction
Co-authored-by: Layla Ghaffari <49916147+laylagi@users.noreply.github.com>
…nings, regression cleanup
Add rational surface impedance boundary condition
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
from
5f205a9 to
fd5c2cb
Compare
cuDSS is a high-performance CUDA library for Direct Sparse Solvers. This commit enables the use of cuDSS in Palace by: - Wrapping mfem::CuDSSSolver in a new palace::CuDSSSolver; - Enabling the use of CuDSSSolver for KSP and WavePort; and - Adding cuDSS configuration options to CMake, for both Palace and MFEM.
Maintainer feedback on the cuDSS support: - Rebase cleanup: in cudss.cpp swap the transitive utils/iodata.hpp include for utils/labels.hpp, the header that actually defines the MatrixSymmetry / SymbolicFactorization types the file uses (include-what-you-use); drop the dead IoData forward-decl in cudss.hpp and the unused cudss.hpp include in waveportoperator.cpp; and document the intentionally-unused reorder argument in the CuDSSSolver constructor. - Consolidate the "not built with cuDSS" guard to a single config-time check in iodata.cpp (mirroring the SuperLU/STRUMPACK/MUMPS guards), and drop the redundant #else MFEM_ABORT from the ksp.cpp cuDSS case so it matches the empty-guarded MUMPS pattern. - Bump the configuration schema $id to 1-2-0 for the new "cuDSS" Solver.Linear.Type enum value: a new allowed enum value is a backward- compatible SchemaVer REVISION (required by scripts/check-schema-version). Tag the CHANGELOG entry with "SchemaVer 1-2-0" per the bump policy; the notes.md release table is intentionally left untouched (updated only at release time). - Document cuDSS in the install guide (build flag + spack example) and add unit tests: every Solver.Linear.Type value (incl. "cuDSS") validates against the schema, and Type="cuDSS" parses to LinearSolver::CUDSS when built with cuDSS. - Enable cuDSS in the GPU CI: build with +cuda+cudss and add a cuDSS entry to the GPU regression-test solver matrix.
hughcars
force-pushed
the
hughcars/pr717-cudss
branch
from
fd5c2cb to
204ff5b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes for PR717: