Validate preconditioner evidence before claiming speedups
This commit is contained in:
251
experiments/check_preconditioner_benchmark.py
Normal file
251
experiments/check_preconditioner_benchmark.py
Normal file
@@ -0,0 +1,251 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Audit frozen preconditioner artifacts and compare equal-accuracy solve costs.
|
||||
|
||||
A completed experiment is valid even when neither inverse reaches the requested
|
||||
linear tolerance. Those observations must survive the audit, without acquiring
|
||||
an invented time-to-convergence speedup.
|
||||
"""
|
||||
import argparse
|
||||
import csv
|
||||
import hashlib
|
||||
import json
|
||||
import math
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
REQUIRED_ARTIFACTS = {
|
||||
'metadata.txt', 'layout.csv', 'preconditioner_summary.csv', 'preconditioner_actions.csv',
|
||||
'preconditioner_blocks.csv', 'preconditioner_correction_blocks.csv', 'preconditioner_history.csv',
|
||||
'preconditioner_invariants.csv', 'preconditioner_reference.csv', 'preconditioner_scope.txt',
|
||||
'benchmark_physical_base.txt', 'benchmark_normalized_base.txt', 'benchmark_normalized_rhs.txt',
|
||||
'benchmark_state_factors.txt', 'benchmark_residual_factors.txt',
|
||||
'benchmark_invariant_direction.txt', 'benchmark_invariant_action.txt',
|
||||
'context_reference_normalized_correction.txt',
|
||||
} | {role + suffix for role in ('baseline', 'candidate') for suffix in (
|
||||
'_normalized_correction.txt', '_physical_correction.txt',
|
||||
'_normalized_linear_defect.txt', '_physical_linear_defect.txt')}
|
||||
|
||||
|
||||
def digest(path):
|
||||
return hashlib.sha256(path.read_bytes()).hexdigest()
|
||||
|
||||
|
||||
def hash_matches(path, expected):
|
||||
return path.is_file() and digest(path) == expected
|
||||
|
||||
|
||||
def rows(path):
|
||||
with path.open() as stream:
|
||||
return list(csv.DictReader(stream))
|
||||
|
||||
|
||||
def vector(path):
|
||||
values = path.read_text().split()
|
||||
if not values:
|
||||
raise ValueError(f'Empty vector: {path}')
|
||||
count = int(values[0])
|
||||
result = [float(v) for v in values[1:]]
|
||||
if count <= 0 or count != len(result) or not all(math.isfinite(v) for v in result):
|
||||
raise ValueError(f'Invalid finite vector: {path}')
|
||||
return result
|
||||
|
||||
|
||||
def norm(values):
|
||||
return math.sqrt(math.fsum(v * v for v in values))
|
||||
|
||||
|
||||
def close(left, right):
|
||||
return math.isfinite(left) and math.isfinite(right) and math.isclose(left, right, rel_tol=2e-12, abs_tol=1e-28)
|
||||
|
||||
|
||||
def ratio(numerator, denominator):
|
||||
# Exact residual zero is a valid outcome. Do not invent a finite ratio or
|
||||
# serialize infinity when the comparison denominator is exactly zero.
|
||||
return numerator / denominator if denominator > 0 else None
|
||||
|
||||
|
||||
def invalidate_comparison(result):
|
||||
comparison = result['comparison']
|
||||
comparison['comparison_valid'] = False
|
||||
comparison['outcome'] = 'invalid_artifacts'
|
||||
for key in ('candidate_over_baseline_defect', 'candidate_over_baseline_measured_cost', 'equal_gate_speedup'):
|
||||
comparison[key] = None
|
||||
|
||||
|
||||
def invalid_case(error):
|
||||
result = {'passed': False, 'checks': [{'check': 'readable_complete_artifacts', 'passed': False,
|
||||
'error': str(error)}], 'comparison': {}}
|
||||
invalidate_comparison(result)
|
||||
return result
|
||||
|
||||
|
||||
def inspect_case(folder):
|
||||
try:
|
||||
return inspect_case_artifacts(folder)
|
||||
except (OSError, ValueError, KeyError, TypeError, OverflowError, ZeroDivisionError) as error:
|
||||
return invalid_case(error)
|
||||
|
||||
|
||||
def inspect_case_artifacts(folder):
|
||||
checks = []
|
||||
def check(name, result):
|
||||
checks.append({'check': name, 'passed': bool(result)})
|
||||
summary = rows(folder / 'preconditioner_summary.csv')
|
||||
check('exactly_one_baseline_and_candidate', len(summary) == 2 and {r['role'] for r in summary} == {'baseline', 'candidate'})
|
||||
by_role = {r['role']: r for r in summary}
|
||||
rhs = vector(folder / 'benchmark_normalized_rhs.txt')
|
||||
rhs_norm = norm(rhs)
|
||||
check('finite_nonzero_rhs', math.isfinite(rhs_norm) and rhs_norm > 0)
|
||||
layout = rows(folder / 'layout.csv')
|
||||
residual_blocks = [r for r in layout if r['kind'] == 'residual']
|
||||
covered = []
|
||||
for row in residual_blocks:
|
||||
covered.extend(range(int(row['offset']), int(row['offset']) + int(row['size'])))
|
||||
check('residual_layout_covers_vector_once', covered == list(range(len(rhs))))
|
||||
invariants = rows(folder / 'preconditioner_invariants.csv')
|
||||
required = {'accepted_physical_bitwise_identical', 'accepted_normalized_bitwise_identical',
|
||||
'accepted_residual_bitwise_identical', 'correction_bitwise_identical',
|
||||
'trial_state_bitwise_identical', 'trial_residual_bitwise_identical',
|
||||
'candidate_physical_bitwise_identical', 'state_factors_bitwise_identical',
|
||||
'residual_factors_bitwise_identical', 'preparation_generation_unchanged',
|
||||
'normalized_operator_current', 'jacobian_action_bitwise_identical'}
|
||||
for role in ('context_reference', 'baseline', 'candidate'):
|
||||
role_rows = [r for r in invariants if r['role'] == role]
|
||||
metrics = {r['metric']: float(r['value']) for r in role_rows}
|
||||
check(role + '_unique_invariants', len(role_rows) == len(metrics))
|
||||
check(role + '_all_invariants_recorded', required <= metrics.keys())
|
||||
check(role + '_all_invariants_pass', all(metrics.get(name) == 1.0 for name in required))
|
||||
check(role + '_exact_jacobian_unchanged', metrics.get('jacobian_action_difference_norm') == 0.0)
|
||||
action_rows = rows(folder / 'preconditioner_actions.csv')
|
||||
check('action_checks_present', len(action_rows) == 2 and
|
||||
{row['role'] for row in action_rows} == {'baseline', 'candidate'})
|
||||
for row in action_rows:
|
||||
expected = by_role[row['role']]['variant'] == 'independent'
|
||||
check(row['role'] + '_action_variant_matches', row['variant'] == by_role[row['role']]['variant'])
|
||||
check(row['role'] + '_equivalence_requirement_recorded',
|
||||
row['context_equivalence_required'] == ('1' if expected else '0'))
|
||||
if expected:
|
||||
difference = float(row['context_action_relative_difference'])
|
||||
check(row['role'] + '_fresh_inverse_matches_default',
|
||||
math.isfinite(difference) and 0 <= difference <= 256 * 2.220446049250313e-16)
|
||||
check('baseline_is_independent', by_role['baseline']['variant'] == 'independent')
|
||||
block_rows = rows(folder / 'preconditioner_blocks.csv')
|
||||
for role, row in by_role.items():
|
||||
defect = vector(folder / (role + '_normalized_linear_defect.txt'))
|
||||
solution = vector(folder / (role + '_normalized_correction.txt'))
|
||||
true_norm = norm(defect)
|
||||
check(role + '_compatible_vector_sizes', len(defect) == len(rhs) == len(solution))
|
||||
check(role + '_rhs_norm', close(float(row['rhs_norm']), rhs_norm))
|
||||
check(role + '_defect_norm', close(float(row['true_residual_norm']), true_norm))
|
||||
check(role + '_relative_defect', close(float(row['true_relative']), true_norm / rhs_norm))
|
||||
check(role + '_solution_norm', close(float(row['normalized_correction_norm']), norm(solution)))
|
||||
target = float(row['relative_tolerance']) * rhs_norm
|
||||
check(role + '_valid_positive_tolerance', math.isfinite(target) and target > 0)
|
||||
check(role + '_unchanged_target', close(float(row['absolute_target']), target))
|
||||
check(role + '_valid_gate_flag', row['true_gate_pass'] in ('0', '1'))
|
||||
check(role + '_honest_gate', (row['true_gate_pass'] == '1') == (true_norm <= target))
|
||||
check(role + '_cold_start', close(float(row['reported_initial_norm']), rhs_norm))
|
||||
check(role + '_iteration_count', int(row['iterations']) == int(row['preconditioner_applications'])
|
||||
and 0 < int(row['iterations']) <= int(row['maximum_iterations']))
|
||||
timing_names = ['construction_seconds', 'normalization_binding_seconds', 'krylov_setup_seconds', 'krylov_seconds']
|
||||
check(role + '_valid_timing', all(math.isfinite(float(row[k])) and float(row[k]) >= 0 for k in timing_names))
|
||||
check(role + '_complete_setup_solve_cost', close(float(row['setup_plus_solve_seconds']), sum(float(row[k]) for k in timing_names)))
|
||||
extra_timings = ['true_check_seconds', 'jacobian_seconds', 'preconditioner_seconds']
|
||||
check(role + '_valid_verification_application_timing',
|
||||
all(math.isfinite(float(row[k])) and float(row[k]) >= 0 for k in extra_timings))
|
||||
full_cost = float(row['setup_solve_and_true_check_seconds'])
|
||||
check(role + '_complete_verified_solve_cost', full_cost > 0 and
|
||||
close(full_cost, float(row['setup_plus_solve_seconds']) + float(row['true_check_seconds'])))
|
||||
check(role + '_one_independent_true_check', int(row['true_check_jacobian_applications']) == 1)
|
||||
check(role + '_nonoverlapping_application_timers',
|
||||
float(row['jacobian_seconds']) + float(row['preconditioner_seconds']) <= float(row['krylov_seconds']) + 1e-8)
|
||||
role_blocks = [r for r in block_rows if r['role'] == role]
|
||||
recorded = {r['block']: r for r in role_blocks}
|
||||
check(role + '_unique_residual_blocks', len(role_blocks) == len(recorded))
|
||||
check(role + '_all_residual_blocks_present', set(recorded) == {b['block'] for b in residual_blocks})
|
||||
for block in residual_blocks:
|
||||
start, count = int(block['offset']), int(block['size'])
|
||||
measured = recorded[block['block']]
|
||||
check(role + '_' + block['block'] + '_defect_norm',
|
||||
close(float(measured['normalized_defect_norm']), norm(defect[start:start + count])))
|
||||
check(role + '_' + block['block'] + '_rhs_norm',
|
||||
close(float(measured['normalized_rhs_norm']), norm(rhs[start:start + count])))
|
||||
baseline, candidate = by_role['baseline'], by_role['candidate']
|
||||
for key in ('restart', 'maximum_iterations', 'relative_tolerance', 'rhs_norm', 'absolute_target'):
|
||||
check('matched_' + key, baseline[key] == candidate[key])
|
||||
both_pass = baseline['true_gate_pass'] == candidate['true_gate_pass'] == '1'
|
||||
outcome = 'both_pass' if both_pass else ('candidate_only_pass' if candidate['true_gate_pass'] == '1'
|
||||
else 'baseline_only_pass' if baseline['true_gate_pass'] == '1' else 'neither_passes')
|
||||
comparison = {'outcome': outcome, 'recorded_gate_outcome': outcome, 'comparison_valid': True,
|
||||
'candidate_over_baseline_defect': ratio(float(candidate['true_relative']), float(baseline['true_relative'])),
|
||||
'candidate_over_baseline_measured_cost': ratio(float(candidate['setup_solve_and_true_check_seconds']), float(baseline['setup_solve_and_true_check_seconds'])),
|
||||
'equal_gate_speedup': float(baseline['setup_solve_and_true_check_seconds']) / float(candidate['setup_solve_and_true_check_seconds']) if both_pass else None,
|
||||
'baseline': baseline, 'candidate': candidate}
|
||||
result = {'passed': all(c['passed'] for c in checks), 'checks': checks, 'comparison': comparison}
|
||||
if not result['passed']:
|
||||
invalidate_comparison(result)
|
||||
return result
|
||||
|
||||
|
||||
def inspect_campaign(campaign):
|
||||
provenance = json.loads((campaign / 'provenance.json').read_text())
|
||||
checks = [{'check': 'immutable_executable_hash', 'passed': hash_matches(campaign / 'plateau_diagnostic', provenance['binary_sha256'])},
|
||||
{'check': 'mesh_hash', 'passed': hash_matches(Path(provenance['mesh']), provenance['mesh_sha256'])},
|
||||
{'check': 'nonempty_unique_cases', 'passed': bool(provenance['cases']) and
|
||||
len(provenance['cases']) == len({case['id'] for case in provenance['cases']})}]
|
||||
results = {}
|
||||
for case in provenance['cases']:
|
||||
name = case['id']
|
||||
record = json.loads((campaign / (name + '_run.json')).read_text())
|
||||
checks.append({'check': name + '_completed', 'passed': bool(record.get('finished_utc')) and
|
||||
record.get('exit_code') == 0 and not record.get('timed_out', False)})
|
||||
checks.append({'check': name + '_required_outputs_hashed',
|
||||
'passed': REQUIRED_ARTIFACTS <= record.get('output_hashes', {}).keys()})
|
||||
arguments = case.get('arguments', [])
|
||||
expected_inputs = set()
|
||||
for index, argument in enumerate(arguments):
|
||||
if argument in ('--probe-state', '--probe-direction', '--replay-candidate'):
|
||||
path = Path(arguments[index + 1])
|
||||
if not path.is_absolute(): path = Path(provenance['source']) / path
|
||||
expected_inputs.add(str(path.resolve()))
|
||||
checks.append({'check': name + '_required_inputs_hashed',
|
||||
'passed': expected_inputs <= record.get('input_state_hashes', {}).keys()})
|
||||
for path, expected in record.get('input_state_hashes', {}).items():
|
||||
checks.append({'check': name + '_input_hash_' + path, 'passed': hash_matches(Path(path), expected)})
|
||||
for path, expected in record.get('output_hashes', {}).items():
|
||||
checks.append({'check': name + '_output_hash_' + path, 'passed': hash_matches(campaign / name / path, expected)})
|
||||
results[name] = inspect_case(campaign / name)
|
||||
if '--preconditioner-variant' in arguments:
|
||||
expected_variant = arguments[arguments.index('--preconditioner-variant') + 1]
|
||||
candidate = results[name]['comparison'].get('candidate', {})
|
||||
checks.append({'check': name + '_requested_variant', 'passed': candidate.get('variant') == expected_variant})
|
||||
passed = all(c['passed'] for c in checks) and all(r['passed'] for r in results.values())
|
||||
if not all(c['passed'] for c in checks):
|
||||
# A valid-looking CSV cannot substantiate a speedup when its producing
|
||||
# executable or input/output provenance failed verification.
|
||||
for result in results.values():
|
||||
result['passed'] = False
|
||||
invalidate_comparison(result)
|
||||
assessment = {'passed': passed, 'source_commit': provenance['commit'], 'provenance_checks': checks, 'cases': results,
|
||||
'scope': 'Frozen linear-solve comparison only. A failed baseline has no finite equal-gate speedup.'}
|
||||
return assessment
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description=__doc__)
|
||||
parser.add_argument('campaign', type=Path)
|
||||
parser.add_argument('--output', type=Path)
|
||||
args = parser.parse_args()
|
||||
campaign = args.campaign.resolve()
|
||||
assessment = inspect_campaign(campaign)
|
||||
path = args.output or campaign / 'preconditioner_assessment.json'
|
||||
path.write_text(json.dumps(assessment, indent=2, allow_nan=False) + '\n')
|
||||
for name, result in assessment['cases'].items():
|
||||
c = result['comparison']
|
||||
print(name, 'PASS' if result['passed'] else 'FAIL', c['outcome'],
|
||||
'defect_ratio=', c['candidate_over_baseline_defect'], 'equal_gate_speedup=', c['equal_gate_speedup'])
|
||||
return 0 if assessment['passed'] else 1
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
raise SystemExit(main())
|
||||
Reference in New Issue
Block a user