Skip to content

Unify ionic-step header keyword, cube channel suffixes, and Vxc(R) sparse output filename in IO outputs - #8293

Open
mohanchen wants to merge 126 commits into
deepmodeling:developfrom
mohanchen:2026-11-05-a
Open

mohanchen wants to merge 126 commits into
deepmodeling:developfrom
mohanchen:2026-11-05-a

Conversation

@mohanchen

Copy link
Copy Markdown
Collaborator

Summary

This PR consolidates a set of IO-output naming clean-ups that have grown
inconsistent across different writers. Three coordinated changes:

1. Unify ionic-step header keyword to Ionic_Step N

Output files written during geometry relaxations / MD previously used a
mix of header markers — RELAX STEP N, STEP: N, IONIC STEP N,
# istep: — across different writers. All text-based per-step outputs
now use the single form # Ionic_Step N (one-based), matching the
existing STRU-frame convention and the project-wide rule recorded in
AGENTS.md.

Files updated:

  • source_io/module_hs/* (HS/H(R)/S(R) dense + sparse writers)

  • source_io/module_dos/write_pdos_text.cpp

  • source_io/module_wfc/* (WFC / band header lines)

  • Related integration fixtures under tests/02_NAO_Gamma/**/*.ref

2. Disambiguate Libxc cube channel suffixes from real spin indices

write_libxc_r.cpp previously tagged every cube channel with _sN,
which collided with the real-spin convention used by rho / vrho
(s1 = spin-up, s2 = spin-down). Tensor quantities (sigma,
vsigma, v2rho2, v2rhosigma, v2sigma2, v3rho3, … v4sigma4)
now carry explicit Libxc channel-index tuples (_00, _01, _11, …)
and scalar quantities (exc, amag) drop the redundant _s1 suffix.

category before after
real spin rho_s1.cube, vrho_s2.cube unchanged (_sN is meaningful)
scalar exc_s1.cube, amag_s1.cube exc.cube, amag.cube
3-ch tensor sigma_s1/s2/s3.cube sigma_00/01/11.cube
6-ch tensor v2rhosigma_s1..s6.cube v2rhosigma_00_0 / 00_1 / … .cube
higher-order v3sigma3_s1..s10.cube v3sigma3_00_00_00 / … .cube

A small set of helper lambdas generates the suffix lists, so the
Libxc channel ordering stays consistent with libxc_tools.cpp
((00), (01), (11) for 3-channel tensors, symmetric-combination order
for higher ranks).

3. Rename Vxc(R) sparse CSR fixture to canonical form

source_io/module_hs/vxc_op_r.cpp was writing the Vxc(R) sparse matrix
as Vxc_R_spin0.ref in test fixtures. The file is now produced under
the canonical LCAO-matrix naming vxcrs{spin}_nao.csr, aligned with
hsr_writer / dsr_writer conventions. The fixture under
tests/02_NAO_Gamma/scf_out_hxc/ is renamed and its header assertion
updated accordingly.

Verification

  • make -j$(nproc) : full build passes, zero errors

  • source/source_io/module_chgpot/write_libxc_r.cpp compiles cleanly under __LIBXC

  • Integration references under tests/02_NAO_Gamma/, tests/03_NAO_multik/,
    and tests/08_RI/ have been updated to match the new header strings

Migration note for downstream users

Any post-processing scripts that parse the per-step header from
HS/H(R)/DOS/PDOS/WFC output files should look for Ionic_Step N
instead of STEP: N, RELAX STEP N, IONIC STEP N, or istep:.
Cube-filename consumers should switch from the _sN suffix pattern to
the new explicit index tuple for tensor quantities; real-spin files
(rho_s1, vrho_s2) are unaffected.

abacus_fixer added 12 commits November 5, 2026 15:08
…out_app_flag=2

The band output did not honor out_app_flag=2: the filename never received
a g{step} suffix and always wrote into global_out_dir, so multi-step runs
(MD/relax) would overwrite bands1.txt / bands2.txt and only keep the last
step.

Changes:
- Add global_band_dir to System_para, mapped to OUT/<suffix>/BANDS/ in mode 2
  and falling back to global_out_dir otherwise; includes MPI broadcast.
- write_bands: use global_band_dir as the output directory, add g{istep+1}
  suffix when out_app_flag==2, and lazily ensure the BANDS/ directory exists.
- Tests: sync io_sys_var_test directory assertions and test_write_bands
  TestParameters with the new global_band_dir.

Verified: MODULE_IO_write_bands (10/10), MODULE_IO_system_variable_test pass;
esolver builds cleanly; governance check has no new blockers.
Replace the legacy "RELAX STEP" keyword and pre-existing inconsistent
"Ionic Step" (with space) occurrences with "Ionic_Step" (underscore) to
match the established convention already used by write_cube.cpp and
related integration-test fixtures.

Production code:
- source_relax/relax_stru_io.cpp: STRU file provenance header
- source_io/module_output/print_info.cpp: stdout/ofs_running run logs
- source_io/module_hs/vxc_op_tools.cpp: eband decomposition energy header
- source_pw/module_pwdft/dftu_base_io.cpp: DFTU occ_mat provenance header
- source_io/module_dm/write_dmk.cpp: DM(k) frame header (historical)
- source_io/module_dm/write_dmr.cpp: DM(R) CSR frame header (historical)
- source_io/module_hs/hsr_writer.cpp: H(R)/S(R) CSR frame header (historical)
- source_io/module_mulliken/output_mulliken.cpp: Mulliken output frame header (historical)

Tests:
- source_relax/unittests/test_relax_stru_io.cpp
- source_io/unittests/test_print_info.cpp
- source_io/module_dm/test/write_dmk_test.cpp
- source_io/module_dm/test/test_write_dmr.cpp
- source_io/module_hs/unittests/test_hsr_writer.cpp

Comments / doc strings:
- source_base/module_out/csr_reader.h: CSR format example
- source_io/module_parameter/read_inp_out.cpp: out_stru parameter description

Verification:
- make -j 30 -> 100%
- ctest -R "(RELAX_relax_stru_io|IO_print_info|IO_vxc_op_tools|PW_dftu_base|
  IO_mulliken|IO_write_dmk|IO_write_dmr|IO_hsr_writer)" -> 10/10 PASSED
- grep -rn "RELAX STEP" / "Ionic Step " -> zero matches in source tree
Production code:
- source_io/module_wf/write_wfc_nao.cpp: wf{ki}s{spin}_nao.txt provenance header
- source_io/module_energy/write_bands.cpp: band eigenvalue provenance header

Tests:
- source_io/unittests/test_write_wfc_nao.cpp
- source_io/unittests/test_read_wfc_nao.cpp
- source_io/unittests/test_write_bands.cpp

Verification:
- make -j 30 -> 100%
- ctest -R "(IO_write_wfc_nao|IO_read_wfc_nao|IO_write_bands|
  RELAX_relax_stru_io|IO_print_info)" -> 8/8 PASSED
- grep -rn "RELAX STEP\|Ionic Step \|IONIC STEP" -> zero matches
Three source output files wrote "# ionic step" (or "X # ionic step") in
their per-ionic-step header comments, while STRU and cube outputs already
use "Ionic_Step". Rename to the unified keyword across:
- cal_dos.cpp      -> "X    # Ionic_Step" (doss*.txt / pdoss*.txt)
- write_eig_occ.cpp -> "X     # Ionic_Step" (eig_occ.txt)
- hs_dense_io.cpp  -> "# Ionic_Step X"  (HS dense matrix)

Sync test assertions and INPUT parameter description accordingly.
Also refines the INPUT file under examples/17_relax/04_relax_with_output_lcao
with end-of-line filename comments for each enabled out_* parameter.
The prior commit 4038796 unified "# ionic step" to "# Ionic_Step" in
hs_dense_io.cpp but missed test_hs_dense_io.cpp, causing TextHeaderContains
StepAndDimension and TextHeaderStepTwo to fail. Sync the EXPECT_NE find()
strings and a nearby comment.

Verification: MODULE_IO_hs_dense_io_test -> 11/11 PASSED.
Three writers produced per-frame headers as "STEP: N" or "IONIC_STEP: N":
- pos_op_csr.cpp       (rr_nao.csr, rx_nao.csr, ...)
- hs_sparse_io.cpp     (hrs*.csr, srg*.csr, trs*.csr, ...)
- dhs_sparse_writer.cpp (dhrx*.csr, dsrx*.csr, ...)
- overlap.cpp          (S_async(R) overlap)

Rename to "Ionic_Step N" (keyword-first, no colon) to match STRU cube,
doss, eig_occ, and HS dense-matrix headers unified in the previous commit.

Sync test assertions, fixture SR.csr, and comments accordingly.
PDOS text output (write_pdos_text.cpp) used "# istep:" as its per-step
header keyword while all other outputs had already been unified to
"Ionic_Step". This was the last remaining non-Ionic_Step keyword in
production output.

Production code:
- source_io/module_dos/write_pdos_text.cpp

Tests:
- source_io/module_dos/unittests/test_write_pdos_text.cpp (5 assertions)

Docs:
- docs/advanced/elec_properties/dos.md (PDOS format example and description)

Verification:
- make -j 30 -> 100%
- MODULE_DOS_write_pdos_text -> 5/5 PASSED
hs_dense_io.cpp writes "# Ionic_Step N" since the keyword unification, but
the hk/sk/vxc/dH reference files still contained "# ionic step". CompareFile.py
compares comment lines token by token, so those matrix comparisons would fail.
Update 31 references under tests/; value data is unchanged.
out_mat_r writes a single rr_nao.txt (or MAT/rrg{step}_nao.txt when
out_app_flag = 2) rather than separate rxrs/ryrs/rzrs CSR files. Update
the INPUT description, the position-matrix guide, and the example INPUT
comment, and align the documented frame header with the emitted
Ionic_Step keyword.
Tensor quantities (sigma, vsigma, v2rho2, v2rhosigma, v2sigma2, v3rho3,
v3rho2sigma, v3rhosigma2, v3sigma3, v4rho4, v4rho3sigma, v4rho2sigma2,
v4rhosigma3, v4sigma4) previously used _s1/_s2/_s3 suffixes that collide
with the real spin-index convention used by rho and vrho.  Replace them
with explicit Libxc-channel index tuples (_00, _01, _11 for 3-channel
tensors, etc.).  Scalar quantities (exc, amag) no longer carry a
redundant _s1 suffix.  Rho and vrho keep _sN since those map to actual
spin indices.  A set of small helper lambdas generates the suffix lists
so the channel ordering stays consistent with Libxc throughout.
Follow the repository NAO sparse-matrix naming convention: one-based
spin index, "_nao.csr" suffix, and the g{step} suffix plus append
behavior driven by out_app_flag in mode 2. Pass the ionic step and
out_app_flag explicitly into write_Vxc_R instead of hard-coding
istep = -1. Update the out_mat_xc2 help text, the example INPUT
comment, and the scf_out_hxc reference fixture (including its
Ionic_Step header).
@mohanchen mohanchen added Input&Output Suitable for coders without knowing too many DFT details Refactor Refactor ABACUS codes The Absolute Zero Reduce the "entropy" of the code to 0 Solve before 3.11 labels Nov 5, 2026

@ABACUS-PR-BOT ABACUS-PR-BOT left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The output changes introduce mismatched derivative labels, a CSR step-reader regression, stale Vxc records on reruns, and unmigrated integration headers. The global-dependency budget and three non-blocking documentation/test items also need attention. Static source and API-contract review of 63431111; no builds or tests were run.

Comment thread source/source_io/module_hs/vxc_op_r.cpp Resolved
abacus_fixer added 3 commits November 6, 2026 14:44
The earlier commit 91465ba updated 31 .txt.ref files under tests/ for
the Ionic_Step unification, but left .csr.ref, rr_nao.txt.ref,
mulliken.txt.ref and cube.ref fixtures still carrying the legacy
"--- Ionic Step N ---", "STEP: N" or "IONIC_STEP: N" headers.
CompareFile.py does a token-by-token compare and those mismatches cause
the integrating tests (md_out_hk_syns, scf_out_hk_spin2, scf_out_mul,
scf_out_mul_spin2, ...) to fail. Update 41 remaining reference files
under tests/02_NAO_Gamma/ and tests/03_NAO_multik/ to the new unified
"Ionic_Step N" keyword. Data values are untouched.
The eig_occ.txt per-ionic-step header was unified from "X     # ionic
step" to "X     # Ionic_Step", but the ASE reader still counted frames
with endswith('# ionic step'), yielding nframe = 0 and failing the
assertion len(ik) == nframe * nspin * nk in read_band_from_eig_occ.

Match the header with a case-insensitive regex that accepts both the
old and the new spelling, and sync the nspin4-gamma-eigocc fixture to
the current format.
The out_band BANDS/ change read PARAM.globalv.global_band_dir inside
write_bands, which pushed the PR over the global dependency budget
(added=4, removed=3, net_delta=1) and blocked CI.

Resolve the directory once in ESolver_KS::after_scf and pass it as an
explicit argument, following the existing write_eig_file /
output_mat_sparse convention. This keeps the band writer free of PARAM
and brings the PR-wide delta to zero.

Verified: MODULE_IO_write_bands 2/2 passed; abacus_basic_para target
builds; governance diff counting gives net_delta=0.

@ABACUS-PR-BOT ABACUS-PR-BOT left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The reader now accepts both step-header spellings. The suggestion below covers the legacy format. Static incremental review; builds and tests were not run.

abacus_fixer added 8 commits November 7, 2026 11:33
The non-MPI-specific tests in test_hsr_writer.cpp performed file I/O
on all MPI ranks simultaneously, causing file corruption and race
conditions when the test binary was launched with mpirun -np 2.

- For the 7 tests that call the non-collective write_hcontainer_csr /
  write_hcontainer_csr_binary directly, add an early return on
  non-zero ranks so only rank 0 writes/reads/removes the files.
- For WriteHsrTextCsrCarriesPerSpinFermiFromEfermi, keep the collective
  write_hsr call (which internally uses gatherParallels) on all ranks,
  but guard the file read-back, assertions, and cleanup to rank 0 only.

Verified: single-process (21 passed, 1 skipped) and mpirun -np 2
(22 passed on both ranks), plus ctest serial+parallel both pass.
The MO-basis Vxc matrix written by out_mat_xc reused the "vxc" property,
which resolved to the same vxck{ik}s{spin}_nao.txt filename as the
NAO-basis matrix from out_mat_vxck; with out_app_flag 0/1 the later
writer silently overwrote the former. Give the MO-basis matrix its own
"vxcmo" property and pass the real ionic step and out_app_flag into
write_Vxc instead of locally shadowed constants, which restores the
one-based Ionic_Step header and the mode-2 g{step} suffix. Update the
two integration fixtures and props_mat.sh accordingly.
write_Vxc_R is called from ctrl_runner_lcao, which runs once at the end
of the whole calculation, so out_app_flag=2 only ever produced a single
vxcrs{spin}g{N}_nao.csr file. Drop the per-step suffix and always
truncate, matching the documented vxcrs{spin}_nao.csr name, and remove
the now unused out_app_flag parameter.
write_h_term looped both spin channels while write_hk_common enumerated
only the spin-up k indices, so Vnl(k)/Vl(k)/Vh(k) were written twice
into s1 files, and Vxc(k), which is spin dependent, placed both channels
in the same s1 file. Enumerate the local k points of each spin through
K_Vectors::isk as write_hsk does, pass nspin=1 to filename_output for
spin-independent terms so they carry no spin suffix, and apply the same
enumeration to the EXX writer and to T(k).
Emit the column header only once per file and stamp each frame with
"# Ionic_Step <1-based step>" so the angular-momentum matrices match the
step labelling of the other matrix outputs. Appending frames no longer
repeats the full header.
Mode 0 previously kept every written frame and mode 2 ignored the step
number. Open the file per the selected mode: overwrite for mode 0, append
after the first step for mode 1, and write mullikeng{step}.txt in mode 2.
write_eig_file always appended after the first step and ignored the ionic
output frequency, so mode 0 kept every frame and mode 2 had no step suffix.
Gate the write by out_freq_ion and open the file per out_app_flag: overwrite
in mode 0, append after the first step in mode 1, and write eig_occg{step}.txt
in mode 2.
out_band was dumped on every ionic step regardless of out_freq_ion, so a
run with out_freq_ion = 2 produced five band frames while the other
step-scoped outputs produced three. Gate write_bands with the same
frequency used by eig_occ.txt; a non-positive frequency still writes
every step.
abacus_fixer added 11 commits November 18, 2026 09:07
output_SR wrote a compact "Matrix Dimension" header while write_hsr wrote
the self-describing CSR header for the same sr_nao.csr file name. Pass the
unit cell explicitly and enable csr_format so both paths emit one layout;
the binary record path is unchanged.
Every R-space writer assembled its own "<prefix>[s<spin>][g<step>]_nao"
name, which let the S(R) path drift: output_SR ignored out_app_flag=2 and
always wrote sr_nao.csr into the run directory with step 0.

Add make_sparse_R_filename in the shared detail header and route H(R),
S(R), DM(R), dH/dR, r(R), L(R), T(R) and Vxc(R) through it. output_SR now
takes the ionic step and writes into global_matrix_dir with the mode-2
step suffix; existing names are unchanged.
save_sparse now writes only R blocks with elements above the threshold,
but the out_mat_t and out_mat_r descriptions (and the r(R) format note)
still promised one block per candidate R vector. Describe the header
count and block list as the non-empty R set actually written.
prepare_exx_nscf assumed a flat directory with fixed file names, so an
SCF run with out_app_flag=2 (eig_occg{step}.txt in the output root and
wavefunctions under PSI/) could not feed the hybrid NSCF. Resolve the
latest written ionic/electronic step, preferring the fixed name so that
modes 0 and 1 keep their existing paths.
State that source files need no renaming: fixed names win, and otherwise
the latest g{step}/e{iter} snapshot is read from the output root and its
PSI/ subdirectory.
Remove the RT-TDDFT-only out_freq_td branches from every writer; in RT-TDDFT
istep is already the cumulative electronic step, so the same frequency works.
Keep out_freq_td as a deprecated input alias that is forwarded to out_freq_ion
when the canonical key is not set, and ignore it outside RT-TDDFT.
The CSR angular-momentum writer only fills the scalar real-harmonic block, so
a spinor basis would produce a silently incomplete L(R). Fail in ReadInput
instead of aborting after SCF, document the reason, and keep the writer guard
as a defensive check.
Move the duplicated CSR description banner to the lower ModuleIO layer so
Output_HContainer and the R-space sparse writers emit byte-identical text
without a source_hamilt -> source_io dependency.
The text dH/dR and r(R) writers must keep a zero block for every R counted
in their shared header, but a globally empty component needs no MPI
collective. Write the zero block directly on the owning rank so empty
components no longer pay one allreduce per row, while output stays
byte-identical.
The local istep_mat=-1 bypassed the zero-based to one-based conversion
in save_mat, producing "# Ionic_Step 0" in vxc matrix headers. Use the
istep parameter directly so the header prints step 1 as intended.
…at_tk

The dense text body writer asserts a positive precision, but the three
INPUT keywords that feed it accepted 0 or negative values and tripped the
assertion during text output. Validate the precision during input parsing
in prepare_input, where the error string is broadcast to every rank before
aborting, and document the positive-integer requirement for each keyword.
Cover the boundary with input-error cases for out_hsk/out_dmk/out_mat_tk.

@ABACUS-PR-BOT ABACUS-PR-BOT left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At cdf5bcb0, get_s output, RT-TDDFT cadence/input validation, and restart-source consistency need the corrections below and in the existing threads.

Static incremental review since d46e01e6 against fcd57f6d; no builds or runtime tests were run, and CI was not verified. The merge-tree check against f05ff9f2 reports a conflict in source/source_md/unittests/CMakeLists.txt; preserve both sets of test registrations when resolving it.

Comment thread source/source_esolver/esolver_gets.cpp Resolved
Comment thread source/source_lcao/lcao_set.cpp Resolved
abacus_fixer added 16 commits November 19, 2026 08:21
… 2026-11-05-a

Resolve add/add conflict in source/source_md/unittests/CMakeLists.txt
by keeping both sides' tests: upstream's migrated MD test targets
(lj/snap/dp/nep/fire/verlet/nhc/msst/lgv) and the local
MODULE_MD_final_step target for run_md loop predicates.
…ectory

The write_Vxc_LIP caller passes global_mat_dir, so with basis_type =
lcao_in_pw the fixed-name matrices and eband_vxc*.txt files go to the
MAT directory (OUT.${suffix}/MAT/ when out_app_flag = 2), not the root
output directory. Clarify the mode-2 behavior and give concrete file
names for Gamma-only and multi-k cases.
…test

The TextAppendTwoStepsAndFirstStepTruncates test previously deleted the
file before writing, so it never exercised replacing a stale record.
It also only checked step headers, not the matrix values.

Seed a stale frame first, then write a fresh-run step 0 (truncating)
and step 1 (appending). Assert exactly two frames remain, their upper-
triangle payloads match the new matrices, and no stale values survive.
With calculation=get_s and out_app_flag=2, output_SR opens a .csr file
under global_matrix_dir (MAT/) before any ensure_directory call ran, so
a fresh output directory made the open fail. Create the directory once
before output_SR and drop the now-redundant calls in the out_mat_r and
out_mat_ds branches.
read_dmk only checked nkstot and the k-point index, so a reordered or
shifted k-mesh of the same size was silently accepted; the init_chg=dm
fallback then folded the stored DM(k) with the current KPT phases and
produced a wrong density. The reader now parses header[6] (kvec_d) and
rejects the file unless the direct coordinates match the running mesh
modulo a reciprocal lattice vector (integer shifts leave e^{ik·R}
unchanged). A same-count mismatch test covers swapped, shifted and
wrapped meshes.
Rank 0 now probes the DMR CSR files and broadcasts the result to all
ranks, ensuring every rank selects the same communication path (DMR
distribution vs DM(k) broadcast). Previously each rank decided
independently, which could deadlock when file visibility differed
across ranks.
Hardcoded /tmp paths ignore the configured scratch directory, and two
failed writes could previously compare as equal empty outputs. Build the
paths from testing::TempDir() per rule 17 and assert both output and
readback streams open before comparing contents.
Rule 5 prohibits adding default arguments to existing interfaces. Drop
the default "pw" from write_input/reject_input in the input-domain test
fixture and provide an explicit one-argument overload that forwards
"pw", so the 24 existing PW call sites stay unchanged while the LCAO
case keeps its basis explicit.

Verified: OMP_NUM_THREADS=1 ./MODULE_IO_input_domain_reader, 19/19 pass.
Change the dense dH/dk filename to put the 1-based k-point index right
after the prefix and use _ia for the atom index: dhk1x_ia1_nao.txt
replaces dhkx_iat1_ik0_nao.txt. The _nao suffix is kept. Update the
unit test and the scf_out_dh / scf_out_dh_t reference files and
result.ref keys to match.
Check strtod range errors and finite components so malformed restart
matrices cannot propagate infinity or NaN into the density.
The DM(k) text files produced by write_dmk carry D(i,j) = sum_n w_n
conj(c_n,i) c_n,j because dm_from_psi builds DM = conj(P) (the physical
one-body density matrix P with the complex conjugate on the first band
index). This convention was only implicit in a consumer-side note at
source/source_io/module_mulliken/output_mulliken.cpp ~L37 (added in PR
deepmodeling#8250), so restart readers and external post-processing tools could
silently mix conj-first and conj-last interpretations.

- Emit a '#' comment line inside the 11-line header of every DM(k) file
  stating the exact element convention. Placing it inside the header block
  lets read_dmk's read_data_line skip it transparently; put it after the
  UnitCell block would break the matrix-body stream extractor which does
  not skip comment lines.
- Document the same convention in write_dmk's Doxygen @details, including
  the relation to the physical density matrix sum_n w_n c_n,i conj(c_n,j).
- Add a unit test HeaderStatesConjFirstConvention that asserts the header
  comment is present and that a non-symmetric complex probe round-trips
  through write_dmk/read_dmk without losing off-diagonal imaginary signs.
…format upgrade

The hs_sparse_io writer was upgraded in c1ff9a6 ("align tr_nao.csr with
the self-describing CSR layout", 2026-11-09) to emit the header block with
#-comment annotation, lattice info, and CSR format banner. The commit updated
two integration-test refs but did not touch get_s/sr_nao.csr.ref, which was
still in the legacy "Matrix Dimension of S(R)" / "Matrix number of S(R)"
format. A follow-up update at 1500cec ("Align Vxc(R) CSR output with the
H(R) CSR layout") then changed the R count of the two tr_nao.csr.ref files
to values that do not match the case INPUT (125 and 9 vs the actual 9 and 1,
respectively). Both gaps are fixed here by running the three cases under the
current writer and capturing what it emits:

- get_s/sr_nao.csr.ref: legacy -> new self-describing header, R count 3
  (c1ff9a6 did not update this file at all).
- nscf_out_hsr_tr_rr/tr_nao.csr.ref: R count 125 -> 9 (matches c1ff9a6's
  correct value; 1500cec had changed it by hand).
- scf_out_dh_t/tr_nao.csr.ref: R count 27 -> 1 (the case INPUT emits one R
  block; previous ref was R=27 from c1ff9a6, already a case/INPUT mismatch).

Also remove the untracked KPT file left behind under tests/02_NAO_Gamma/ by a
local run. No INPUT changes are present in this commit — the refs now reflect
what the cases actually compute.
…rdinates as DM

UcellIO::read_ucell used to be a void line skipper: when the per-type
atom-count line in a DMK file was malformed (non-numeric, empty), the
std::istringstream loop yielded total_natom == 0, so the function only
consumed the "Direct" marker and left the coordinate rows on the stream.
A subsequent small-matrix body read (e.g. nlocal = 2 gamma) would then
happily consume those coordinate numbers and hand back a bogus density
matrix.  The init_chg = dm restart path trusts read_dmk's return value,
so a silently wrong density would seed the next SCF.

- Change UcellIO::read_ucell signature from void to bool.
- Every getline now checks the stream state; the atom-count line must
  yield at least one non-negative integer summing to a positive total,
  otherwise the block is indeterminate and we bail out.
- read_dmk handles the new return value: failure sets read_success,
  prints a diagnostic, and breaks out of the rank0 body so the existing
  bcast_bool cooperative-failure path propagates the error.
- Add a handwritten unit test ReadDMKRejectsMalformedAtomCount that
  rewrites the per-type count line to a non-numeric token and asserts
  read_dmk returns false without touching the caller buffer.

Test: DMKTest.* 24/24 pass (1 new, 23 existing no regression).
… 2026-11-05-a

Conflicts resolved:
- source/source_io/CMakeLists.txt: merged both sides' object lists, keeping
  parse_args.cpp + upstream's new parse_command_line.cpp + local libxc_r_suffix.cpp
- source/source_md/unittests/test_run_md.cpp: kept upstream's consolidate
  (deepmodeling#8322) RunMDTest/RunMDCadenceTest suite and appended local RunMdFinalStepTest
The old description compressed scalar, spin-resolved and tensor outputs
into one optional-suffix pattern, so users could not derive file names
such as xc_v2rhosigma_0_00.cube. List the three channel forms explicitly
and align the examples with libxc_r_channel_suffixes.

@ABACUS-PR-BOT ABACUS-PR-BOT left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The merged MD test target is missing required link dependencies. Changed DMK findings and source-level fixes are updated in the original threads. This review used static source and reference-data analysis; no builds, tests or MPI/CI checks were run.

Comment on lines +211 to +213
addtest(
TARGET MODULE_MD_final_step
SOURCES test_run_md.cpp)

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P1] After the merge, test_run_md.cpp also constructs cell objects and calls the non-inline Run_MD::prepare_mdcell, but this target still compiles it alone. AddTest supplies only GTest and the numerical runtime here, leaving the required ABACUS implementations unlinked. Please remove this redundant target in favor of MODULE_MD_run, or give the header-only predicate tests a separate source.

Comment on lines +778 to +779
MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
MPI_Comm_size(MPI_COMM_WORLD, &world_size);

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Non-blocking: please route the new rank/size, barrier and reduction calls through the project parallel helpers, and encapsulate the split/free lifecycle or document a narrow exception. The added barrier in test_hs_dense_io.cpp has the same rule-9 issue; existing MPI calls are outside this request.

}
fk += "_nao.txt";
const bool append_fk = (out_app_flag == 1) && (params.istep > 0);
std::string fk = kprefix + std::to_string(params.kv->ik2iktot[ik] + 1) + tag;

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Non-blocking: please update out_mat_dh help and write_dh.h to match this filename change. The writer now produces names such as dhk1z_ia2_nao.txt, while the help still promises dhk[xyz]_iat{I} with an optional _ik suffix, so users following it will look for nonexistent files.

@AsTonyshment AsTonyshment left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Still a few minor issues to address before merging.

{
const std::string filename = ModuleIO::filename_output(readin_dir, "wf", "pw", iq,
source_points->ik2iktot, inp.nspin, nks, 2, false, false, -1);
const std::string filename = resolve_wfc(iq, source_points->ik2iktot, nks);

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this has been pointed out by the PR bot earlier. WFC and occupation files independently select their latest snapshots, potentially mixing different ionic/electronic steps and producing an incorrect EXX operator. We should somehow ensure all source files belong to the same snapshot.

Comment on lines +313 to +315
// A per-step file (mode 2) must describe the step its name
// announces, otherwise the restart would silently mix steps.
if (out_app_flag == 2 && frame_step >= 0 && frame_step != istep + 1)

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Mode 0 does not check consistency of Ionic_Step across k/spin files. An interrupted write can therefore mix density matrices from different steps, producing an incorrect restart density. Maybe we should validate their common step ID.

// final step is independent of SCF convergence (an unconverged final
// step is still useful), and it is not duplicated when it happens to
// fall on the cadence because this function writes at most once.
if (istep % inp.out_freq_ion == 0 || last_ion_step)

@AsTonyshment AsTonyshment Nov 19, 2026 •

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this has also been pointed out by the PR bot earlier. In RT-TDDFT, last_ion_step remains true for every electronic substep of the final ionic step, bypassing out_freq_ion. Maybe restrict the final-frame override to the last electronic substep?

For example, with md_nstep=1, estep_per_md=5, and out_freq_ion=4, all 6 electronic steps are written instead of only steps 1, 5, and 6 (one-based).

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Input&Output Suitable for coders without knowing too many DFT details Refactor Refactor ABACUS codes Solve before 3.11 The Absolute Zero Reduce the "entropy" of the code to 0

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants