File size: 10,890 Bytes
68db0a3 | 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 | #!/usr/bin/env python3
"""Conditional h^2 Richardson comparators for the fixed receiver-cost sweep.
The extrapolated responses assume leading h^2 error; they do not certify
continuum error. Candidate waveforms and construction times are unchanged.
"""
import os
for name in ("OPENBLAS_NUM_THREADS", "OMP_NUM_THREADS", "MKL_NUM_THREADS"):
os.environ[name] = "1"
import argparse
import csv
import hashlib
import json
from pathlib import Path
import numpy as np
HERE = Path(__file__).resolve().parent
ROOT = HERE.parent
OUT = HERE / "results"
LABELS = ("grid512", "grid384", "richardson512", "richardson384")
def digest(path):
sha = hashlib.sha256()
with path.open("rb") as stream:
for block in iter(lambda: stream.read(1024 * 1024), b""):
sha.update(block)
return sha.hexdigest()
def relative(value, reference):
return float(np.sqrt(np.sum((value-reference)**2)/np.sum(reference**2)))
def write_csv(path, rows):
with path.open("w", newline="") as stream:
writer = csv.DictWriter(stream, fieldnames=list(rows[0]))
writer.writeheader()
writer.writerows(rows)
def dump(path, value):
path.write_text(json.dumps(value, indent=2, allow_nan=False)+"\n")
def key(row):
return (row["material"], row["grid"], row["method"], row["requested_cfl"])
def winner(rows, label, tolerance, family):
feasible = [r for r in rows if r[label+"_error"] <= tolerance
and r["family"] == family]
return min(feasible, key=lambda r: r["median_seconds"]) if feasible else None
def calculate():
cfg = json.loads((HERE/"config.json").read_text())
inputs = {}
references, differences = {}, []
formula_defects = []
times = np.linspace(0, cfg["time_end"], cfg["samples"])
for material in cfg["materials"]:
raw = {}
for n in (256, 384, 512):
path = ROOT/"propagation"/"results"/f"{material}_n{n}.npz"
with np.load(path) as data:
np.testing.assert_array_equal(data["time"], times)
raw[n] = data["receiver_pressure"].copy()
assert raw[n].shape == (len(times), 4, 3)
inputs[str(path.relative_to(ROOT))] = digest(path)
ext512 = raw[512]+(9/7)*(raw[512]-raw[384])
ext384 = raw[384]+(4/5)*(raw[384]-raw[256])
references[material] = {"grid512": raw[512], "grid384": raw[384],
"richardson512": ext512, "richardson384": ext384}
formula_defects.extend([relative(ext512, (16*raw[512]-9*raw[384])/7),
relative(ext384, (9*raw[384]-4*raw[256])/5)])
differences.append({
"material": material,
"extrapolant_difference_relative_to_richardson512": relative(ext384, ext512),
"grid512_difference_relative_to_richardson512": relative(raw[512], ext512),
"grid384_difference_relative_to_richardson384": relative(raw[384], ext384),
"largest_source_extrapolant_difference": max(relative(ext384[:,:,s], ext512[:,:,s]) for s in range(3)),
"largest_channel_extrapolant_difference": max(relative(ext384[:,r,s], ext512[:,r,s]) for r in range(4) for s in range(3)),
})
records = sorted((OUT/"candidates").glob("*_timing.json"))
prescribed = {(m,n,"full_stormer_verlet",c) for m in cfg["materials"]
for n in cfg["grids"] for c in cfg["explicit_cfl_values"]}
prescribed |= {(m,n,"craig_bampton",None) for m in cfg["materials"] for n in cfg["grids"]}
candidates = []
for path in records:
timing = json.loads(path.read_text())
rawpath = path.with_name(path.name.replace("_timing.json", ".npz"))
assert digest(rawpath) == timing["raw_sha256"]
inputs[str(path.relative_to(ROOT))] = digest(path)
inputs[str(rawpath.relative_to(ROOT))] = timing["raw_sha256"]
with np.load(rawpath) as data:
np.testing.assert_array_equal(data["time"], times)
y = data["receiver_pressure"].copy()
totals = [sum(r.values()) for r in timing["repetitions"]]
assert float(np.median(totals)) == timing["median"]["total_seconds"]
row = {name: timing[name] for name in ("material", "grid", "method", "modes_per_patch", "requested_cfl", "dimension")}
row.update(family="full" if timing["method"].startswith("full_") else "reduced",
median_seconds=float(np.median(totals)),
minimum_seconds=min(totals), maximum_seconds=max(totals))
for label in LABELS:
row[label+"_error"] = relative(y, references[row["material"]][label])
candidates.append(row)
assert {key(row) for row in candidates} == prescribed
assert len(candidates) == len(prescribed)
selections = []
for material in cfg["materials"]:
rows = [r for r in candidates if r["material"] == material]
for tolerance in cfg["receiver_tolerances"]:
primary = {f: winner(rows, "grid512", tolerance, f) for f in ("full", "reduced")}
for label in LABELS:
selected = {f: winner(rows, label, tolerance, f) for f in ("full", "reduced")}
row = {"material": material, "tolerance": tolerance, "comparator": label}
for family, candidate in selected.items():
row[family+"_feasible"] = candidate is not None
for attr in ("grid", "method", "dimension", "requested_cfl", "median_seconds", "minimum_seconds", "maximum_seconds"):
row[family+"_"+attr] = None if candidate is None else candidate[attr]
row[family+"_error"] = None if candidate is None else candidate[label+"_error"]
row[family+"_selection_changed_from_grid512"] = (
(None if candidate is None else key(candidate)) !=
(None if primary[family] is None else key(primary[family])))
samefamily = [r for r in rows if r["family"] == family]
row[family+"_pass_to_fail_from_grid512"] = sum(
r["grid512_error"] <= tolerance < r[label+"_error"] for r in samefamily)
row[family+"_fail_to_pass_from_grid512"] = sum(
r[label+"_error"] <= tolerance < r["grid512_error"] for r in samefamily)
full, reduced = selected["full"], selected["reduced"]
both = full is not None and reduced is not None
row["median_full_over_reduced"] = full["median_seconds"]/reduced["median_seconds"] if both else None
row["fastest_full_over_slowest_reduced"] = full["minimum_seconds"]/reduced["maximum_seconds"] if both else None
row["slowest_full_over_fastest_reduced"] = full["maximum_seconds"]/reduced["minimum_seconds"] if both else None
selections.append(row)
old = list(csv.DictReader((OUT/"receiver_errors.csv").open()))
old_by_key = {(r["material"], int(r["grid"]), r["method"], None if r["requested_cfl"] in ("", "None") else float(r["requested_cfl"])):r for r in old}
replay = max(abs(r[label+"_error"]-float(old_by_key[key(r)][column]))
for r in candidates for label,column in (("grid512","total_receiver_error"), ("grid384","alternative_reference_total_receiver_error")))
assert replay < 1e-13 and max(formula_defects) < 1e-14
inputs[str((OUT/"receiver_errors.csv").relative_to(ROOT))] = digest(OUT/"receiver_errors.csv")
summary = {
"scope": "Conditional Richardson comparators under a leading h^2 waveform-error expansion. Neither extrapolant is a known exact solution or a rigorous error bound.",
"formulas": {"richardson512": "Y512 + (9/7)*(Y512-Y384)", "richardson384": "Y384 + (4/5)*(Y384-Y256)"},
"norm": "Equal-weight Frobenius norm over all 641 times including t=0, four receivers and three sources. Each candidate error uses the corresponding comparator norm.",
"cost": "Unchanged median, minimum and maximum of the three original setup-inclusive stage sums. No construction or timing is repeated.",
"candidate_count": len(candidates), "reference_differences": differences,
"mean_extrapolant_difference": float(np.mean([r["extrapolant_difference_relative_to_richardson512"] for r in differences])),
"selections": selections, "input_sha256": inputs,
"config_sha256": digest(HERE/"config.json"), "analysis_code_sha256": digest(Path(__file__)),
"maximum_previous_error_replay_difference": replay,
"maximum_alternative_formula_difference": max(formula_defects),
}
return candidates, differences, selections, summary
def verify(candidates, differences, selections, summary):
previous = json.loads((OUT/"richardson_summary.json").read_text())
assert previous == summary
maximum = 0.0
for name, rows in (("candidate_errors",candidates), ("reference_differences",differences), ("selections",selections)):
saved = list(csv.DictReader((OUT/f"richardson_{name}.csv").open()))
assert len(saved) == len(rows)
for old, new in zip(saved, rows):
for k,v in new.items():
if isinstance(v, float):
maximum = max(maximum, abs(float(old[k])-v))
else:
assert old[k] == ("" if v is None else str(v))
assert maximum < 1e-13
# The two Richardson weight pairs preserve constants and cancel h^2.
h2_defect = max(abs(16/7/512**2-9/7/384**2), abs(9/5/384**2-4/5/256**2))
assert h2_defect < 1e-20
result = {"status":"pass", "candidate_errors_recomputed":len(candidates)*len(LABELS),
"least_cost_selections_recomputed":len(selections)*2,
"input_hashes_recomputed":len(summary["input_sha256"]),
"maximum_saved_numeric_difference":maximum,
"maximum_previous_error_replay_difference":summary["maximum_previous_error_replay_difference"],
"maximum_alternative_formula_difference":summary["maximum_alternative_formula_difference"],
"leading_h_squared_cancellation_defect":h2_defect,
"no_timing_repeated":True}
dump(OUT/"richardson_verification.json", result)
return result
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--verify", action="store_true", help="Recompute and compare the saved postprocessing outputs")
args = parser.parse_args()
candidates, differences, selections, summary = calculate()
if not args.verify:
for name,rows in (("candidate_errors",candidates), ("reference_differences",differences), ("selections",selections)):
write_csv(OUT/f"richardson_{name}.csv", rows)
dump(OUT/"richardson_summary.json", summary)
print(json.dumps(verify(candidates, differences, selections, summary), indent=2))
if __name__ == "__main__":
main()
|