Repository navigation
Conversation
…utput
- Add out_freq_ion/out_freq_elec INPUT parameters to write numbered
occupation-matrix snapshots dm_onsiteg{istep+1}.txt, with one section
per recorded electronic step containing energy, magnetism and status
- Rebuild the SOC occupation matrix in spinor space from the four Pauli
blocks and diagonalize it with zheev; the previous code read the charge
block four times and reported identical eigenvalues and magnetism
- Write dm_onsite.txt in spin-basis layout for nspin=4 so that it matches
read_occup_m() on init_chg=file restart
- Add unit tests for the base IO helpers and update INPUT documentation
mohanchen
added
DFT+U
added 8 commits
October 5, 2026 21:23
….cpp The anonymous-namespace helper extract_soc_pauli_blocks() lives at file scope, so the enum and its enumerators from namespace DFTU_BASE require explicit qualification. Verified: make module_pwdft -j4 builds successfully.
write_occup_m() now prefixes every atom block with "<Element> <index>"
and appends the Cartesian position, so dm_onsite.txt and the
dm_onsiteg{#}.txt snapshots share the identical header line, e.g.
Fe 1 Atom=1 L=2 0.00000000 0.00000000 0.00000000
read_occup_m() skips tokens up to "Atom=", accepts both the compact
"L=2" and the legacy split "L= 2", and treats "ORBITAL=" as optional
(zeta defaults to 0), so legacy dm_onsite.txt files still parse.
Verified: make module_pwdft -j4 builds successfully.
…site.txt with g files
- Move dm_onsite.txt writing from iter_init to iter_finish so each
electronic-step snapshot records the actual charge-density residual
drho alongside the configured threshold scf_thr.
- Share one section writer between dm_onsite.txt and dm_onsiteg{#}.txt:
provenance header (version, timestamp, RELAX STEP), then
"# Electronic step N / # scf_thr / # drho" plus the compact matrix body.
- Snapshot " N/A" instead of a silent zero matrix at PW istep 0 / iter 1
when no occupation matrix exists yet (loaded-from-file runs keep the
real matrix via is_occmat_ready); LCAO always writes the real matrix.
- read_occup_m skips the new comment lines and the N/A token, so
restart parsing is unaffected.
- Replace the scf_dftu_spin2/spin4 integration cases with
relax_dftu_spin2/spin4 (relax_nmax 5, scf_nmax 5) and register them
in CASES_CPU/GPU.txt; add AppendSnapshotNAPlaceholderAndReady unit
test for the N/A placeholder behavior.
Verification: incremental build clean; ctest
MODULE_PW_dftu_base_test + MODULE_PW_test_dftu_pw_tools 2/2 pass;
relax_dftu_spin2 run exit=0, dm_onsite.txt identical to the last
dm_onsiteg5.txt section; drho sequence varies per step as expected;
nspin4 SOC run + restart (init_chg=file) exit=0; governance check
exit=0 with global-dependency budget net_delta=0.
…matrix output
out_occ_mat (Boolean, default 1) decouples the DFT+U occupation-matrix
files (dm_onsite.txt and dm_onsiteg{#}.txt) from out_chg: users no longer
need to pay the charge-density cube IO to obtain occupation matrices.
The switch is carried by OccmatOutputCfg::out_occ_mat and gates all
writers in one place each: write_latest_occmat() (dm_onsite.txt),
prepare_ion_step_file() via output()/iter_init_dftu_pw() (numbered-file
header), and append_ion_step_snapshot() (numbered-file sections). The
bool out_chg parameters of output(), iter_init_dftu_pw() and
finish_dftu_lcao() served only occupation-matrix output and are removed.
Side fix: with out_chg=0 and out_freq_ion>0, append_ion_step_snapshot()
previously created headerless orphan files; the new gate prevents that.
Global dependency budget is migration-neutral (net_delta=0).
…{#}.txt
dm_onsite.txt and dm_onsiteg{#}.txt are renamed to occ_mat.txt and
occ_matg{#}.txt; gen_ion_step_dm_onsite_filename() is renamed to
gen_ion_step_occ_mat_filename() accordingly. read_occup_m() error
messages now report the actual probed filename.
init_chg=file first probes occ_mat.txt and falls back to the legacy
dm_onsite.txt, so output directories written by older versions can
still be used for restarts. dm_onsite_ini.txt (occ_mat_ctrl input)
keeps its name.
The omc parameter note now points to occ_mat.txt and out_occ_mat=1
(the default) as the way to produce it.
Verification: MODULE_PW_dftu_base_test and MODULE_DFTU_base_io pass;
PW DFT+U SCF writes occ_mat.txt + occ_matg{1,3,5}.txt; NSCF with
init_chg=file reads occ_mat.txt and, after renaming it to
dm_onsite.txt, produces the identical total energy via the fallback;
with both files absent the run quits with a clear error.
band.md and dos.md now reference occ_mat.txt in the NSCF workflow; run_scf_nscf.sh copies occ_mat.txt (the new output name) from the SCF output; the dm_onsite.txt mentions in tests/17_DS_DFTU/README.md are updated accordingly. Verification: bash -n run_scf_nscf.sh passes; no non-legacy dm_onsite references remain in docs/advanced/elec_properties or tests/17_DS_DFTU.
- Add init_occ_mat (Integer, 0/1/2) as the primary occupation-matrix initialization mode parameter. - Keep omc as a legacy alias; when both are set, init_occ_mat takes precedence (same pattern as out_mat_hs / out_mat_hs2). - Rename occ_mat_ctrl member/getter to init_occ_mat across dftu_base, esolver call sites, and tests. - Regenerate docs/parameters.yaml and input-main.md.
- init_occ_mat 1/2 now searches read_file_dir for dm_onsite_ini.txt, then falls back to occ_mat.txt, then the legacy dm_onsite.txt, so a previously output occ_mat.txt can be reused without renaming. - The init_chg=file branch reuses the same helper. - Move find_first_existing_file into DFTU_BASE namespace and add a unit test for candidate priority. - Update init_occ_mat description and regenerate docs.
- PW DFT+U: when init_occ_mat=2, rebuild pot_onsite and E_U from the
fixed file-loaded occupation matrix at every electronic step
(including the first); the matrix itself is still neither
reprojected from psi nor mixed. Previously the whole cal_occ_pw
was skipped, leaving V_U and E_plusU at zero for the entire run.
- Remove the per-type U listing and per-atom occupation matrices
from the running log; the same data is available through the
out_occ_mat-gated occ_mat.txt / occ_matg{#}.txt files and the log
block floods the output for large cells.
- Print the full energy table at every electronic step for DFT+U
calculations so E_plusU can be monitored.
- Convert relax_dftu_spin2 into an init_occ_mat=2 regression case
(KPT 2x1x1, tighter SCF) with an occ_mat.txt fixture.
Add AGENTS.md rule 17: unit tests must not create or delete files or
directories via std::system("mkdir/rm"), std::remove, or
std::filesystem::remove_all; use committed fixtures or
testing::TempDir() instead.
Fix dftu_base_test.cpp to comply:
- Replace std::system("mkdir -p/rm -rf") calls in
FindFirstExistingFileTest and InitBaseReadsOccMatFileOnlyOnce with
testing::TempDir(); use unique fixture names so leftover files from
a previous run do not affect the "no candidate" case.
- Simulate "file no longer present" on the second ionic step by passing
a non-existent readin dir instead of deleting the file.
- Add a custom main() that calls MPI_Init/Finalize so the binary runs
under MPI builds where local_occup_bcast() broadcasts on COMM_WORLD.
Rule 17 now forbids only shell-based file/dir IO
(std::system("mkdir/rm")) in unit tests. Direct library calls
(std::remove, std::filesystem::remove_all) on test-written files
are acceptable since they do not go through a shell and cannot
delete directory trees, so there is no injection or runaway-delete
risk.
…density-deviation log
… in locals
Pass nspin as an explicit parameter to iter_init_dftu_pw instead of reading
PARAM.inp.nspin inside the function body. In both esolver iter_finish
functions, read PARAM.globalv.{global_out_dir,npol,gamma_only_local} once
into locals so the repeated DFT+U call sites do not each re-enter the
global dependency surface. PR net_delta drops from +5 (BLOCK) to -4 (WARN).
write_latest_occmat and append_ion_step_snapshot only checked out_occ_mat (default true), not dft_plus_u. Non-DFT+U LCAO/PW runs dereferenced an empty l_channel vector in has_l_channel() during iter_finish, segfaulting at GE1 with a nil address. Add dft_plus_u to OccmatOutputCfg and early-out when dft_plus_u <= 0, honouring the documented contract that out_occ_mat only takes effect for DFT+U calculations. Add DftPlusUDisabledWritesNothing regression test using a default-constructed Plus_U_Base (empty l_channel). Verified: abacuslite scf.py converges (Si LCAO, E=-210.6232 eV); MODULE_PW_dftu_base_test 11 existing tests pass; new regression test compiled locally (user runs final ctest confirmation).
…1/3) Introduce a dormant scf_rerun_ member in ESolver_KS and a rerun hook at the bottom of ESolver_KS::runner's SCF loop. When a subclass sets scf_rerun_, the loop clears the flag and restarts from iter=1. This replaces the historical iter=0 restart-signal trick that crashed DFT+U occupation-matrix writers validating iter>=1. Behavior is unchanged in this step: scf_rerun_ is never set yet, so the hook is dormant and EXX continues to rely on ++iter to turn iter=0 back into iter=1. Build + full link pass in build_max_para_test.
EXX no longer mutates the caller's iter. Change Exx_HelperBase and Exx_Helper signatures: iter_finish and exx_after_converge now take iter by value. Drop the "iter = 0" write in exx_after_converge that used to trigger an SCF rerun via ++iter in the for-loop. The SCF rerun is now requested explicitly: ESolver_KS_PW::iter_finish captures conv_esolver before the EXX call and, when SCF had converged but EXX overrode conv_esolver to false, sets ESolver_KS::scf_rerun_. ESolver_KS::runner (updated in step 1) then restarts the loop at iter=1. This keeps iter strictly 1-based for every downstream reader (check_deltaspin_oscillation, write_latest_occmat, append_ion_step_snapshot, ctrl_iter_pw) and removes the crash in append_ion_step_snapshot's iter>=1 validation. Build + full link pass.
… 3/3) Move the dft_plus_u/out_occ_mat guard to the top of append_ion_step_snapshot, before the nspin/istep/iter/frequency validation. When DFT+U is off this function is a no-op, so an invalid iter (defense in depth against any future upstream bug that might reintroduce an iter=0 sentinel) no longer crashes the iter>=1 validation. Mirrors the guard-first pattern already used by write_latest_occmat. Validation still runs in full when DFT+U is on (the only case where this function does real work), so the existing contract is preserved. MODULE_PW_dftu_base_test passes; full abacus build + link pass.
This branch has not been deployed
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
1 participant
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.
Fix #8084
Reminder
AGENTS.mdanddocs/developers_guide/agent_governance.md.source/changes.Linked Issue
Fix #
Unit Tests and/or Case Tests for my changes
What's changed?
Governance Notes