File size: 6,791 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
"""Derive the occupied-only variant of a full-C store.

For every p2 shard the group directory is copied file for file (no decode or re-encode of the
Fock and table arrays), except the `cmo_*` arrays, which are rewritten with only the occupied
rows of each calculation's C^T block. Both variants share one schema: `mo_content` in the group
attributes is "full" or "occupied", and `cmo_x[o[i]:o[i+1]].reshape(-1, nbas)` is C^T restricted to
the stored orbitals either way. p1 and the root files are copied as they are.

    python derive_occ_store.py --src $PSCRATCH/omol_100k --dst $PSCRATCH/omol_100k_occ --workers 64
"""
from __future__ import annotations
import argparse, glob, os, shutil, sys, time, traceback
import multiprocessing as mp
import numpy as np
import zarr

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 omol_store import put_array

CMO_CHUNK = 1 << 22


def p2_shards(store):
    out = []
    for depth in ("*", "*/*/*"):
        out += glob.glob(os.path.join(store, "p2", depth, "*.zarr"))
    return sorted(set(out))


def derive(args):
    src, dst = args
    t0 = time.time()
    try:
        if os.path.isdir(dst):
            if zarr.open_group(dst, mode="r").attrs.get("occ_derived_verified"):
                return dst, "skip", 0, 0.0, ""
            shutil.rmtree(dst)
        gs = zarr.open_group(src, mode="r")
        if gs.attrs.get("mo_content", "full") != "full":
            raise RuntimeError("source is not a full-C store")
        skip = {n for n in gs.array_keys() if n.startswith("cmo_")}
        shutil.copytree(src, dst, ignore=lambda d, names: [n for n in names if d == src and n in skip])
        # the copied root zarr.json may carry consolidated metadata that still lists the excluded
        # arrays; drop it and re-consolidate once the new arrays are in place
        import json
        meta_path = os.path.join(dst, "zarr.json")
        meta = json.load(open(meta_path))
        if "consolidated_metadata" in meta:
            meta["consolidated_metadata"] = None
            json.dump(meta, open(meta_path, "w"))
        gd = zarr.open_group(dst, mode="r+", use_consolidated=False)
        n = int(gs.attrs["n_calc"])
        nbas = np.asarray(gs["scalar_i"])[:, list(gs.attrs["scalar_i_cols"]).index("nbas")].astype("i8")
        mi = np.asarray(gs["mo_i"])
        nocc = {s: mi[:, list(gs.attrs["mo_i_cols"]).index(f"nocc_{s}")].astype("i8") for s in "ab"}
        written = 0
        for s in "ab":
            zsrc = gs[f"cmo_{s}"]
            off_src = np.asarray(gs[f"cmo_{s}_offsets"])
            lens = np.where(np.diff(off_src) > 0, nocc[s] * nbas, 0)
            offs = np.zeros(n + 1, dtype="i8")
            offs[1:] = np.cumsum(lens)
            buf = np.empty(int(offs[-1]), dtype=zsrc.dtype)
            for i in range(n):
                if lens[i]:
                    a = int(off_src[i])
                    buf[offs[i]:offs[i + 1]] = zsrc[a:a + int(lens[i])]
            put_array(gd, f"cmo_{s}", buf, codec=None,
                      chunks=(max(1, min(len(buf), CMO_CHUNK)),), overwrite=True)
            put_array(gd, f"cmo_{s}_offsets", offs, overwrite=True)
            # verify: every stored block equals the source prefix
            back = np.asarray(gd[f"cmo_{s}"][...])
            if not np.array_equal(back, buf):
                raise RuntimeError(f"read-back mismatch cmo_{s}")
            written += buf.nbytes
        # verify the copied arrays too (cheap ones fully, Fock by a spot check of the first calc)
        for name in gs.array_keys():
            if name in skip:
                continue
            if name.startswith("fock_") and not name.endswith("_offsets"):
                k = min(int(gs[name].shape[0]), 200_000)
                if not np.array_equal(np.asarray(gs[name][:k]), np.asarray(gd[name][:k])):
                    raise RuntimeError(f"copy mismatch {name}")
            else:
                a, b = np.asarray(gs[name][...]), np.asarray(gd[name][...])
                ok = np.array_equal(a, b, equal_nan=True) if a.dtype.kind == "f" else np.array_equal(a, b)
                if not ok:
                    raise RuntimeError(f"copy mismatch {name}")
        gd.attrs.update({
            "mo_content": "occupied",
            "mo_layout": "cmo_x[o[i]:o[i+1]].reshape(nocc_x, nbas) = C_occ^T (row k = occupied MO k "
                         "over AOs, ORCA AO order); derived from the full-C store",
            "occ_derived_verified": True,
        })
        if gs.metadata.consolidated_metadata is not None:
            zarr.consolidate_metadata(gd.store)
        return dst, "ok", written, time.time() - t0, ""
    except Exception:
        return dst, "fail", 0, time.time() - t0, traceback.format_exc(limit=4)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--src", required=True)
    ap.add_argument("--dst", required=True)
    ap.add_argument("--workers", type=int, default=64)
    args = ap.parse_args()
    src, dst = os.path.abspath(args.src), os.path.abspath(args.dst)
    os.makedirs(dst, exist_ok=True)
    # p1 and root files verbatim
    if not os.path.isdir(os.path.join(dst, "p1")):
        shutil.copytree(os.path.join(src, "p1"), os.path.join(dst, "p1"))
    for name in os.listdir(src):
        p = os.path.join(src, name)
        if os.path.isfile(p):
            shutil.copy2(p, dst)
        elif name == "code":
            shutil.copytree(p, os.path.join(dst, "code"), dirs_exist_ok=True)
    shards = p2_shards(src)
    jobs = [(s, os.path.join(dst, os.path.relpath(s, src))) for s in shards]
    print(f"derive occupied-only store: {len(jobs)} p2 shards {src} -> {dst}", flush=True)
    t0 = time.time()
    n_ok = n_skip = n_fail = 0
    written = 0
    with mp.Pool(min(args.workers, len(jobs))) as pool:
        for k, (path, status, w, dt, err) in enumerate(pool.imap_unordered(derive, jobs), 1):
            if status == "ok":
                n_ok += 1; written += w
            elif status == "skip":
                n_skip += 1
            else:
                n_fail += 1
                print(f"FAIL {path}\n{err}", flush=True)
            if k % 100 == 0 or k == len(jobs):
                print(f"  {k}/{len(jobs)} ok={n_ok} skip={n_skip} fail={n_fail} "
                      f"cocc {written/1e9:.1f} GB  {(time.time()-t0)/60:.1f} min", flush=True)
    print(f"\ndone: ok {n_ok}, skipped {n_skip}, failed {n_fail}, occupied C {written/1e9:.1f} GB, "
          f"wall {(time.time()-t0)/60:.1f} min")


if __name__ == "__main__":
    mp.set_start_method("fork", force=True)
    main()