Repository navigation
test(dfpt): test for DFPT phonon frequency output - #8308
Zanthoxylum wants to merge 19 commits into
Conversation
Add the first PW-DFPT integration test and the props_dfpt.sh collector that brings phonon frequencies into the result.ref comparison: - props_dfpt.sh parses the per-q 'DFPT phonon frequencies' and the Gamma 'DFPT LO-TO corrected frequencies' blocks from the test log, emitting dfpt_freq_q<i>_m<j> / dfpt_loto_q<i>_m<j> keys; the mode index keeps its zero-based printing so keys stay stable across cell sizes. Guarded by esolver_type == dfpt so non-DFPT cases are untouched. - catch_properties.sh sources the module and calls run_dfpt_props after run_basic_props (the documented extension point). - 110_PW_DFPT_Si_X: 2-atom Si, ecutwfc 20 / ecutrho 80, fixed occupations, Gamma-centered 4x4x4, X-point q (0, 0.5, 0.5); the reference pins etot and the six X-point frequencies (133.44x2, 407.6x2, 459.8x2 cm^-1 at CHECKACCURACY=2). Verified end-to-end: fresh run, result.ref generated through catch_properties.sh, regenerated result.out compared key-by-key at the CHECKACCURACY=2 rounding (8/8 keys identical).
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
Two gaps currently weaken the new DFPT regression test: the normal suite does not select the case, and missing phonon output can pass comparison. The inline comments identify both and two nonblocking reproducibility details.
Static source/configuration review at 7e4a3d31e9e00b1d415092b178135a78ae684b00; whitespace, governance and shell syntax checks passed. No ABACUS/runtime or numerical checks were run, and CI diagnostics were unavailable.
…x README Review follow-up on all four comments: - CASES_CPU.txt: register 110_PW_DFPT_Si_X so the normal CTest suite actually selects the case (the reference was never checked before). - props_dfpt.sh: fail the collection (return 1) when a DFPT log contains no dfpt_freq_ key at all (truncated log, crashed solver, format drift); catch_properties.sh propagates the failure with exit 1, so the case can no longer pass on etot alone. The develop sync had also dropped the props_dfpt wiring from catch_properties.sh, making the collector a silent no-op; re-wired together with the propagation. - README: fix the lattice constant (2 * 2.715 * 1.8897261255 = 10.2612 Bohr = 5.4300 Angstrom, was misstated as 10.2) and name the pseudo file; document the intended tolerance: a per-case threshold file (1e-4) covers the cm^-1 keys — same-binary reruns reproduce the six printed decimals exactly, so the margin only absorbs last-digit rounding across toolchains (the Autotest.sh default 1e-7 would demand bit-identical output for ~500 cm^-1 values). Verified: fresh Si_qmid log collects 12 keys (rc=0), an emptied log fails with rc=1 through catch_properties.sh, bash -n clean.
…ly serial The CI Test job runs Autotest.sh with -n 4; the DFPT esolver is currently a serial implementation (MPI domain decomposition is planned future work), so multiple ranks crash in the response-solve FFT gather-scatter (MPI_Alltoallv with rank-inconsistent stick maps, reproduced locally at -n 2 and -n 4). Mirror the of_ml_gene_data special case: cases with esolver_type dfpt run with a single MPI rank. README documents the limitation and the future plan.
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
The new launcher handles the checked-in case, but valid uppercase INPUT keywords bypass its one-rank restriction. The partial-spectrum issue remains in its existing thread.
Static source review at fdc78aadff4840906bad66864174c73ee07400ac; whitespace, governance and shell syntax checks passed. No ABACUS runtime or numerical checks were run; CI diagnostics were unavailable.
…case-insensitive match 1. Key-scoped threshold (1e-4 only for dfpt_* keys, etot keeps 1e-7) 2. Complete-spectrum guard (dense 0..N mode indices + cross-ref vs result.ref) 3. Case-insensitive esolver_type match (-qi) 4. README: per-key threshold + serial note + LibXC build provenance 5. Remove case-wide threshold file
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
Two collector gaps remain: initial/direct reference generation can accept partial spectra, and valid uppercase INPUT keywords can skip phonon comparison. A nonblocking inline suggestion covers malformed frequency values.
Static source review at f4fd288d32f1db5dbd0ff22cfb1888824a2ee657; whitespace, governance and three shell syntax checks passed. No ABACUS runtime or numerical checks were run; CI diagnostics were unavailable.
…okup, numeric guard - Rename 110_PW_DFPT_Si_X to dfpt_x (maintainer: follow naming convention) - Shorten README to 2 lines (maintainer request) - props_common.sh: case-insensitive get_input_key_value (ESOLVER_TYPE DFPT no longer silently skips collection) - props_dfpt.sh: reject nonnumeric frequency tokens
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
The keyword lookup and malformed-frequency rejection are source-fixed at 39ba0c1; the separate first-reference completeness gap remains open in its existing discussion. The new feedback concerns diagnostic text, reference-build metadata and guidance for running the case individually.
Static source review of 39ba0c1 against 48e2494; exact-revision whitespace and LF checks passed. No collector or ABACUS runtime checks ran, and CI diagnostics were unavailable.
Maintainer feedback: 24 min is 288x over CI budget. Reduce to Gamma-only k, Gamma-only q, 8/32 Ry, nbands 4, no E-field. Runtime ~4.3s.
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
The new atom counter can reject valid mixed-species DFPT spectra, and first-time reference generation still accepts incomplete spectra. Both need collector corrections.
Static source review of c33ee19 against 173620a; exact-revision whitespace and LF checks passed. No collector or ABACUS runtime checks ran, and CI diagnostics were unavailable.
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
The per-block checks fix the previously reported mode-0, empty-block and short LO-TO cases for a valid atom count. Two remaining atom-count cases are detailed in the existing thread.
Static review of 34bb3982 against 173620a1; exact-revision whitespace and LF checks passed. No collector or ABACUS runtime checks ran; CI diagnostics were unavailable.
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
The two reported atom-counter examples are source-corrected, and frequency-guard tests are now registered. Valid STRU whitespace still leaves a completeness gap, detailed in the existing thread.
Static source review at 069ef527 against 173620a1. No collector tests, ABACUS runs or builds were executed; CI diagnostics were unavailable.
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
Reviewed 069ef527 → 240ead3: the new STRU fixtures and zero-count guard address the earlier examples, but the remaining atom-count cases are detailed in the existing F4 thread. Static source review only; governance and git diff --check passed, while tests, builds and CI were not run.
Summary
Add the first PW-DFPT integration test (smoke test) and the
props_dfpt.shproperty collector that brings phonon frequenciesinto the standard
result.refcomparison.Test case
dfpt_si_g: 2-atom Si, Gamma-only k-mesh, Gamma-only q,8/32 Ry, nbands 4, no E-field — runs in ~4 seconds following the
suite's conventions (1-2 atoms, Gamma k, sub-second per case).
props_dfpt.sh: parses the per-qDFPT phonon frequenciesblocksfrom the test log, emitting
dfpt_freq_q<i>_m<j>/dfpt_loto_q<i>_m<j>key-value lines. Guarded by
esolver_type == dfpt(case-insensitive),with completeness checks (dense mode indices, cross-reference against
result.ref, nonnumeric token rejection).
Autotest.shapplies adedicated 1e-4 threshold to the frequency keys while keeping the 1e-7
default for etot.
Autotest.sh: cases withesolver_type dfptrun with a singleMPI rank (the DFPT esolver is currently a serial implementation; MPI
domain decomposition is planned future work).
Verification
catch_properties.sh→ result.ref → regeneratedresult.out → 8/8 keys identical
bash -nclean on all modified shell scriptsThe full precision validation (20/80 Ry, 4×4×4 k, Γ+X q, E-field,
LO-TO) is maintained in a separate external campaign, not in CI.