lazymatrix 0.6.0

Lazy normalized design matrices: apply column centering/scaling without materializing (X - 1cᵀ)S⁻¹, preserving sparsity.
Documentation 
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

#!/usr/bin/env python3
"""Join warmed allocation counts and Criterion estimates for overwrite comparisons."""

import argparse
import csv
import json
import math
import re
import sys
from pathlib import Path


def expected_names():
    names = set()
    dense = [
        "faer_dense",
        "nalgebra_dense",
        "ndarray",
        "zarrs_[256, 16]",
        "zarrs_[1024, 32]",
    ]
    sparse = ["faer_csc", "nalgebra_csc", "sprs_csc"]
    for shape in ["2000x32", "10000x128"]:
        for backends, densities in [(dense, ["1"]), (sparse, ["0.01", "0.1"])]:
            for backend in backends:
                for density in densities:
                    for normalization in ["raw", "centered", "standardized"]:
                        for intercept in ["", "/intercept"]:
                            case = f"{backend}/{shape}/density_{density}/{normalization}{intercept}"
                            for path in ["overwrite", "workspace"]:
                                for direction in ["forward", "transpose"]:
                                    names.add(
                                        f"overwrite_products/{case}/{direction}/{path}"
                                    )
                                names.add(f"overwrite_iterations/{case}/{path}")
    return names


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "log", type=Path, help="Combined stdout and stderr from the full run"
    )
    parser.add_argument("--criterion-dir", type=Path, default=Path("target/criterion"))
    parser.add_argument("--baseline", default="overwrite-comparison")
    args = parser.parse_args()

    counts = {}
    pattern = re.compile(
        r"^allocations (overwrite_(?:products|iterations)/.+): "
        r"(\d+) calls, (\d+) bytes per (\d+) operations$"
    )
    for line in args.log.read_text().splitlines():
        if match := pattern.fullmatch(line):
            name, calls, requested_bytes, batch_size = match.groups()
            if name in counts:
                raise ValueError(f"Duplicate allocation sample: {name}")
            counts[name] = tuple(map(int, (calls, requested_bytes, batch_size)))

    expected = expected_names()
    if counts.keys() != expected:
        raise ValueError(
            f"Allocation coverage differs: {len(expected - counts.keys())} missing, "
            f"{len(counts.keys() - expected)} unexpected"
        )

    estimates = {}
    for metadata_path in args.criterion_dir.rglob(f"{args.baseline}/benchmark.json"):
        metadata = json.loads(metadata_path.read_text())
        name = metadata["full_id"]
        if name in expected:
            if name in estimates:
                raise ValueError(f"Duplicate Criterion estimate: {name}")
            estimate = json.loads(metadata_path.with_name("estimates.json").read_text())
            estimates[name] = estimate["mean"]
    if estimates.keys() != expected:
        raise ValueError(
            f"Missing Criterion estimates: {sorted(expected - estimates.keys())}"
        )

    writer = csv.writer(sys.stdout, lineterminator="\n")
    writer.writerow(
        [
            "backend",
            "rows",
            "predictors",
            "density",
            "normalization",
            "intercept",
            "operation",
            "path",
            "batch_size",
            "allocations",
            "requested_bytes",
            "mean_batch_ns",
            "lower_batch_ns",
            "upper_batch_ns",
            "operations_per_second",
        ]
    )
    for name in sorted(expected):
        parts = name.split("/")
        prefix, backend, shape, density, normalization = parts[:5]
        intercept = "intercept" in parts[5:]
        operation = parts[-2] if prefix == "overwrite_products" else "iteration"
        path = parts[-1]
        calls, requested_bytes, batch_size = counts[name]
        estimate = estimates[name]
        mean = estimate["point_estimate"]
        interval = estimate["confidence_interval"]
        lower, upper = interval["lower_bound"], interval["upper_bound"]
        if batch_size != 10 or not all(
            math.isfinite(v) and v > 0 for v in (mean, lower, upper)
        ):
            raise ValueError(f"Invalid measurement: {name}")
        rows, predictors = map(int, shape.split("x"))
        writer.writerow(
            [
                backend,
                rows,
                predictors,
                density.removeprefix("density_"),
                normalization,
                str(intercept).lower(),
                operation,
                path,
                batch_size,
                calls,
                requested_bytes,
                f"{mean:.3f}",
                f"{lower:.3f}",
                f"{upper:.3f}",
                f"{batch_size * 1e9 / mean:.3f}",
            ]
        )


if __name__ == "__main__":
    main()