File size: 18,569 Bytes
39ea985 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 | """Pass B1: molecular-orbital coefficients from the gbw files into the p2 store.
For every p2 shard the worker reads each calculation's gbw (staged by Globus at
<gbw_root>/<dataset>/<calc_id>.gbw.zstd0), takes the FULL coefficient matrix of both spin
channels, checks that the gbw and the Pass A text describe the same calculation, and appends the
result to the shard group. Zarr lets arrays be added to an existing group, so p2 keeps its layout
and row order; every new array is row-aligned with the shard's existing calculations.
Added arrays:
cmo_a, cmo_a_offsets MO coefficients, MO-major: calculation i is
cmo_a[o[i]:o[i+1]].reshape(nbas, nbas) = C^T, i.e. row k is MO k
over the AOs (ORCA AO order). Occupied orbitals are the first
nocc rows, so C_occ is a contiguous prefix. float32 by default.
cmo_b, cmo_b_offsets the same for the beta channel (empty for RHF)
gbw_eps_a/b (+_offsets) fp64 orbital energies of all nbas orbitals from the gbw
gbw_occ_a/b (+_offsets) fp64 occupations of all nbas orbitals from the gbw
mo_i int columns: nocc_a, nocc_b, gbw_nbas, gbw_nop
mo_f8 diagnostics: occ_sum_a/b; eps_err_a/b = max |gbw - printed| over
the printed orbital energies; fc_offdiag_a/b = largest
off-diagonal of C_occ^T F C_occ (Eh); fc_diag_err_a/b =
max |diag(C_occ^T F C_occ) - eps_occ| (Eh); fc_bound_a/b = what
the 5e-7 Eh print rounding of F can do to that block
mo_flags gbw_found, nbas_match, nspin_match, nelec_match, spin_match,
eps_checked, eps_match, occ_contiguous, fock_checked, fock_match,
mo_ok
mo_ok is the conjunction of every check that could be run. fock_match is only meaningful when
fock_checked is set (a Fock matrix was printed and nbas agreed); it asks that the printed Fock
matrix be diagonal in the gbw's occupied orbitals to 1e-2 Eh. A mismatched pair is off by
0.1 to 1 Eh, so the flag is a pairing test, not a convergence test. Typical values are ~1e-6 Eh,
but in near-linearly-dependent bases (smallest overlap eigenvalue ~1e-6, not removed by ORCA)
the 5e-7 Eh print rounding of F is amplified by the large MO coefficients into ~1e-3 Eh on the
occupied block even though F_print agrees with C^-T diag(e) C^-1 to 1e-5 in the AO basis. The
fc_bound columns give that rounding amplification per calculation (5e-7 x max_i ||C_i||_1^2);
compare fc_diag_err against it before reading a large value as an inconsistency.
Usage:
python pass_b1.py --store $PSCRATCH/omol_100k --gbw-root $PSCRATCH/gbw_100k --workers 48
"""
from __future__ import annotations
import argparse, glob, json, os, sys, time, traceback
import multiprocessing as mp
import numpy as np
import zarr
# One thread per worker process: zarr's default pool (os.cpu_count() threads) times 64 to 96
# forked workers thrashed a 128-core node, cutting throughput several-fold.
zarr.config.set({"threading.max_workers": 1, "async.concurrency": 2})
try:
import numcodecs.blosc
numcodecs.blosc.set_nthreads(1)
numcodecs.blosc.use_threads = False
except Exception:
pass
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from gbw_reader import read_gbw
from omol_store import put_array, inflate_fock
MO_I = ("nocc_a", "nocc_b", "gbw_nbas", "gbw_nop")
MO_F8 = ("occ_sum_a", "occ_sum_b", "eps_err_a", "eps_err_b",
"fc_offdiag_a", "fc_offdiag_b", "fc_diag_err_a", "fc_diag_err_b",
"fc_bound_a", "fc_bound_b")
MO_FLAGS = ("gbw_found", "nbas_match", "nspin_match", "nelec_match", "spin_match",
"eps_checked", "eps_match", "occ_contiguous", "fock_checked", "fock_match", "mo_ok")
INFORMATIONAL = ("eps_checked", "fock_checked", "mo_ok")
EPS_TOL = 5e-6 # printed to 6 decimals; rounding alone gives 5e-7
FOCK_TOL = 1e-2 # Eh; see the docstring on why the occupied block can be off by ~1e-3
CMO_CHUNK = 1 << 22
OLD_ARRAYS = ("cocc_a", "cocc_a_offsets", "cocc_b", "cocc_b_offsets", "cocc_occ_a",
"cocc_occ_a_offsets", "cocc_occ_b", "cocc_occ_b_offsets", "cocc_i", "cocc_f8",
"cocc_flags")
def p2_shards(store):
out = []
for depth in ("*", "*/*/*"):
out += glob.glob(os.path.join(store, "p2", depth, "*.zarr"))
return sorted(set(out))
def fock_check(F, C, eps):
"""Largest off-diagonal of C^T F C and largest |diag - eps|, in Eh."""
M = C.T @ (F @ C)
d = np.diag(M).copy()
np.fill_diagonal(M, 0.0)
return float(np.abs(M).max()) if M.size else 0.0, float(np.abs(d - eps).max()) if d.size else 0.0
def process_shard(args):
sp, gbw_root, force, dtype = args
t0 = time.time()
try:
g = zarr.open_group(sp, mode="r+")
if g.attrs.get("has_mo") and g.attrs.get("mo_dtype") == dtype and not force:
return sp, "skip", {}, [], 0.0, ""
n = int(g.attrs["n_calc"])
ids = g.attrs["calc_id"]
ds = g.attrs["dataset"]
si = np.asarray(g["scalar_i"])
ic = {k: j for j, k in enumerate(g.attrs["scalar_i_cols"])}
fl = np.asarray(g["flags"]).astype(bool)
fc = {k: j for j, k in enumerate(g.attrs["flag_cols"])}
eps = {s: np.asarray(g[f"eps_{s}"]) for s in "ab"}
eoff = {s: np.asarray(g[f"eps_{s}_offsets"]) for s in "ab"}
foff = {s: np.asarray(g[f"fock_{s}_offsets"]) for s in "ab"}
nbas_all = si[:, ic["nbas"]].astype("i8")
nsp_all = np.where(fl[:, fc["is_uhf"]], 2, 1)
# Preallocate the coefficient buffers from the text's nbas (verified against the gbw
# below; a mismatch leaves that calculation's slot empty and flagged).
cap = {s: np.zeros(n + 1, dtype="i8") for s in "ab"}
cap["a"][1:] = np.cumsum(nbas_all ** 2)
cap["b"][1:] = np.cumsum(np.where(nsp_all == 2, nbas_all ** 2, 0))
buf = {s: np.zeros(int(cap[s][-1]), dtype=dtype) for s in "ab"}
filled = {s: np.zeros(n, dtype=bool) for s in "ab"}
geps = {s: [] for s in "ab"}
gocc = {s: [] for s in "ab"}
mi = np.zeros((n, len(MO_I)), dtype="i8")
mf = np.full((n, len(MO_F8)), np.nan)
mfl = np.zeros((n, len(MO_FLAGS)), dtype="i1")
failures = []
for i in range(n):
nbas = int(nbas_all[i])
nelec = int(si[i, ic["nelec"]])
mult = int(si[i, ic["mult"]])
is_uhf = bool(fl[i, fc["is_uhf"]])
has_fock = bool(fl[i, fc["has_fock"]])
flags = dict.fromkeys(MO_FLAGS, False)
path = os.path.join(gbw_root, ds, ids[i] + ".gbw.zstd0")
gb = None
if os.path.exists(path):
try:
gb = read_gbw(path)
flags["gbw_found"] = True
except Exception as e:
failures.append((ids[i], f"gbw read error: {type(e).__name__}: {str(e)[:120]}"))
else:
failures.append((ids[i], "gbw missing"))
if gb is None:
for s in "ab":
geps[s].append(np.zeros(0)); gocc[s].append(np.zeros(0))
mfl[i] = [flags[k] for k in MO_FLAGS]
continue
dim, nop = gb["nbas"], gb["nop"]
mi[i, MO_I.index("gbw_nbas")] = dim
mi[i, MO_I.index("gbw_nop")] = nop
flags["nbas_match"] = dim == nbas
flags["nspin_match"] = nop == (2 if is_uhf else 1)
occ_sum = {}
contiguous = True
eps_ok = True
eps_checked = False
fock_checked = fock_ok = True
for k, s in enumerate("ab"):
if k >= nop:
geps[s].append(np.zeros(0)); gocc[s].append(np.zeros(0))
occ_sum[s] = 0.0
continue
op = gb["ops"][k]
occ_v, en_v, C = op["occ"], op["energies"], op["C"]
mask = occ_v > 0
nocc = int(mask.sum())
idx = np.flatnonzero(mask)
if nocc and (idx[-1] != nocc - 1):
contiguous = False
if dim == nbas and (s == "a" or is_uhf):
# MO-major (C^T) so the occupied block is a contiguous prefix
buf[s][cap[s][i]:cap[s][i + 1]] = C.T.ravel().astype(dtype, copy=False)
filled[s][i] = True
geps[s].append(en_v.copy())
gocc[s].append(occ_v.copy())
mi[i, MO_I.index(f"nocc_{s}")] = nocc
occ_sum[s] = float(occ_v.sum())
mf[i, MO_F8.index(f"occ_sum_{s}")] = occ_sum[s]
# printed orbital energies (6 decimals) against the gbw. A reduced print level
# (the no-Fock metal_organics inputs) lists only the first few hundred orbitals,
# so compare over the printed prefix; energies of removed orbitals print as 0.
e_txt = eps[s][eoff[s][i]:eoff[s][i + 1]]
if 0 < len(e_txt) <= dim:
eps_checked = True
m = e_txt != 0.0
err = float(np.abs(en_v[:len(e_txt)][m] - e_txt[m]).max()) if m.any() else 0.0
mf[i, MO_F8.index(f"eps_err_{s}")] = err
if err > EPS_TOL:
eps_ok = False
elif len(e_txt) > dim:
eps_checked = True
eps_ok = False
# printed Fock diagonal in the gbw occupied orbitals
if has_fock and dim == nbas and nocc:
f0, f1 = int(foff[s][i]), int(foff[s][i + 1])
if f1 > f0:
Co = np.ascontiguousarray(C[:, mask])
F = inflate_fock(np.asarray(g[f"fock_{s}"][f0:f1]), nbas)
od, de = fock_check(F, Co, en_v[mask])
mf[i, MO_F8.index(f"fc_offdiag_{s}")] = od
mf[i, MO_F8.index(f"fc_diag_err_{s}")] = de
mf[i, MO_F8.index(f"fc_bound_{s}")] = 5e-7 * float(
(np.abs(Co).sum(axis=0) ** 2).max())
if od > FOCK_TOL or de > FOCK_TOL:
fock_ok = False
del F, Co
else:
fock_checked = False
else:
fock_checked = False
tot = occ_sum["a"] + occ_sum["b"]
flags["nelec_match"] = abs(tot - nelec) < 1e-6
if nop == 2:
flags["spin_match"] = abs((occ_sum["a"] - occ_sum["b"]) - (mult - 1)) < 1e-6
else:
flags["spin_match"] = mult == 1 and abs(occ_sum["a"] - nelec) < 1e-6
flags["eps_checked"] = eps_checked
flags["eps_match"] = eps_ok
flags["occ_contiguous"] = contiguous
flags["fock_checked"] = fock_checked
flags["fock_match"] = fock_ok
flags["mo_ok"] = all(flags[k] for k in MO_FLAGS if k not in INFORMATIONAL)
mfl[i] = [flags[k] for k in MO_FLAGS]
if not flags["mo_ok"]:
bad = [k for k in MO_FLAGS if k not in INFORMATIONAL and not flags[k]]
failures.append((ids[i], "failed: " + ",".join(bad)))
def ragged(parts, dt):
offs = np.zeros(len(parts) + 1, dtype="i8")
offs[1:] = np.cumsum([len(p) for p in parts])
flat = np.concatenate(parts).astype(dt) if parts else np.zeros(0, dt)
return flat, offs
for name in OLD_ARRAYS:
if name in g:
del g[name]
for s in "ab":
# offsets follow the filled slots; an unfilled slot (gbw missing or nbas mismatch)
# gets a zero-length block so the array stays dense
lens = np.where(filled[s], np.diff(cap[s]), 0)
offs = np.zeros(n + 1, dtype="i8")
offs[1:] = np.cumsum(lens)
if filled[s].all():
flat = buf[s]
else:
flat = np.concatenate([buf[s][cap[s][i]:cap[s][i + 1]] for i in range(n) if filled[s][i]]
) if filled[s].any() else np.zeros(0, dtype)
put_array(g, f"cmo_{s}", flat, codec=None,
chunks=(max(1, min(len(flat), CMO_CHUNK)),), overwrite=True)
put_array(g, f"cmo_{s}_offsets", offs, overwrite=True)
flat, offs = ragged(geps[s], "f8")
put_array(g, f"gbw_eps_{s}", flat, overwrite=True)
put_array(g, f"gbw_eps_{s}_offsets", offs, overwrite=True)
flat, offs = ragged(gocc[s], "f8")
put_array(g, f"gbw_occ_{s}", flat, overwrite=True)
put_array(g, f"gbw_occ_{s}_offsets", offs, overwrite=True)
put_array(g, "mo_i", mi, overwrite=True)
put_array(g, "mo_f8", mf, overwrite=True)
put_array(g, "mo_flags", mfl, overwrite=True)
for k in ("has_cocc", "cocc_schema", "cocc_i_cols", "cocc_f8_cols", "cocc_flag_cols",
"cocc_layout", "cocc_gbw_root"):
g.attrs.pop(k, None)
g.attrs.update({
"has_mo": True,
"mo_schema": "omol_elec/pass_b1/2",
"mo_dtype": dtype,
"mo_i_cols": list(MO_I),
"mo_f8_cols": list(MO_F8),
"mo_flag_cols": list(MO_FLAGS),
"mo_layout": "cmo_x[o[i]:o[i+1]].reshape(nbas, nbas) = C^T (row k = MO k over AOs, "
"ORCA AO order); occupied MOs are the first nocc_x rows",
"mo_gbw_root": gbw_root,
})
stats = {
"n": n,
"found": int(mfl[:, MO_FLAGS.index("gbw_found")].sum()),
"ok": int(mfl[:, MO_FLAGS.index("mo_ok")].sum()),
"fock_checked": int(mfl[:, MO_FLAGS.index("fock_checked")].sum()),
"bytes_cmo": int(sum(int(g[f"cmo_{s}"].shape[0]) for s in "ab") * np.dtype(dtype).itemsize),
"eps_err_max": float(np.nanmax(mf[:, [2, 3]])) if np.isfinite(mf[:, [2, 3]]).any() else float("nan"),
"fc_offdiag_max": float(np.nanmax(mf[:, [4, 5]])) if np.isfinite(mf[:, [4, 5]]).any() else float("nan"),
"fc_offdiag_p50": float(np.nanmedian(mf[:, [4, 5]])) if np.isfinite(mf[:, [4, 5]]).any() else float("nan"),
}
for k in MO_FLAGS:
stats["fail_" + k] = int(n - mfl[:, MO_FLAGS.index(k)].sum())
return sp, "ok", stats, failures, time.time() - t0, ""
except Exception:
return sp, "fail", {}, [], time.time() - t0, traceback.format_exc(limit=4)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--store", required=True)
ap.add_argument("--gbw-root", required=True)
ap.add_argument("--workers", type=int, default=48)
ap.add_argument("--dtype", default="f4", choices=("f4", "f8"))
ap.add_argument("--limit", type=int, default=0)
ap.add_argument("--force", action="store_true", help="recompute shards that already have MOs")
args = ap.parse_args()
shards = p2_shards(args.store)
if args.limit:
shards = shards[:args.limit]
print(f"pass B1 (full C, {args.dtype}): {len(shards)} p2 shards in {args.store}, "
f"gbw from {args.gbw_root}", flush=True)
jobs = [(s, args.gbw_root, args.force, args.dtype) for s in shards]
t0 = time.time()
tot = {}
all_fail, shard_fail = [], []
n_skip = 0
eps_max = fc_max = 0.0
fc_p50 = []
with mp.Pool(min(args.workers, len(jobs))) as pool:
for k, (sp, status, st, failures, dt, err) in enumerate(pool.imap_unordered(process_shard, jobs), 1):
if status == "ok":
for key, v in st.items():
if key.startswith(("n", "found", "ok", "fock_checked", "bytes", "fail_")):
tot[key] = tot.get(key, 0) + v
if st["eps_err_max"] == st["eps_err_max"]:
eps_max = max(eps_max, st["eps_err_max"])
if st["fc_offdiag_max"] == st["fc_offdiag_max"]:
fc_max = max(fc_max, st["fc_offdiag_max"])
if st["fc_offdiag_p50"] == st["fc_offdiag_p50"]:
fc_p50.append(st["fc_offdiag_p50"])
all_fail.extend((sp, cid, why) for cid, why in failures)
elif status == "skip":
n_skip += 1
else:
shard_fail.append((sp, err))
print(f"SHARD FAIL {sp}\n{err}", flush=True)
if k % 50 == 0 or k == len(jobs):
el = time.time() - t0
print(f" {k}/{len(jobs)} calcs {tot.get('n',0):,} found {tot.get('found',0):,} "
f"ok {tot.get('ok',0):,} {tot.get('bytes_cmo',0)/1e9:.0f} GB {el/60:.1f} min "
f"eta {el/k*(len(jobs)-k)/60:.1f} min", flush=True)
dt = time.time() - t0
print(f"\nB1 done in {dt/60:.1f} min: shards ok {len(jobs)-n_skip-len(shard_fail)}, "
f"skipped {n_skip}, failed {len(shard_fail)}")
print(f"calculations {tot.get('n',0):,}: gbw found {tot.get('found',0):,}, "
f"mo_ok {tot.get('ok',0):,}, fock checked {tot.get('fock_checked',0):,}")
for k in MO_FLAGS:
print(f" not {k:15s}: {tot.get('fail_'+k, 0):,}")
print(f"C bytes {tot.get('bytes_cmo',0)/1e12:.3f} TB "
f"({tot.get('bytes_cmo',0)/max(tot.get('n',1),1)/1e6:.2f} MB/calc, {args.dtype})")
print(f"max eps error (gbw vs printed) {eps_max:.2e} Eh; "
f"C_occ^T F C_occ off-diagonal: median-of-shard-medians "
f"{np.median(fc_p50) if fc_p50 else float('nan'):.2e}, max {fc_max:.2e} Eh")
out = os.path.join(args.store, "b1_failures.tsv")
with open(out, "w") as fh:
fh.write("shard\tcalc_id\treason\n")
for sp, cid, why in all_fail:
fh.write(f"{os.path.relpath(sp, args.store)}\t{cid}\t{why}\n")
print(f"{len(all_fail)} per-calculation problems written to {out}")
with open(os.path.join(args.store, "b1_summary.json"), "w") as fh:
json.dump({"totals": tot, "dtype": args.dtype, "eps_err_max": eps_max,
"fc_offdiag_max": fc_max, "wall_s": dt,
"shard_failures": [s for s, _ in shard_fail]}, fh, indent=1)
if __name__ == "__main__":
mp.set_start_method("fork", force=True)
main()
|