File size: 5,254 Bytes
ff2eba6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Benchmark small FP32 Cholesky against preallocated PyTorch POTRF."""

from __future__ import annotations

import argparse
import importlib
import math
import statistics
import sys
from pathlib import Path

import torch

TESTS = Path(__file__).resolve().parents[1] / "tests"
sys.path.insert(0, str(TESTS))
from _source_loader import load_source_ops  # noqa: E402


def load_installed_ops(artifact: str | None):
    if artifact:
        sys.path.insert(0, artifact)
    try:
        return importlib.import_module("small_matrix_cholesky")
    finally:
        if artifact:
            sys.path.remove(artifact)


def make_spd(batch: int, n: int) -> torch.Tensor:
    generator = torch.Generator(device="cuda").manual_seed(41000 + n + batch)
    x = torch.randn(
        batch,
        n,
        n,
        device="cuda",
        dtype=torch.float32,
        generator=generator,
    ) / n**0.5
    return (
        x @ x.transpose(-1, -2)
        + 0.5 * torch.eye(n, device="cuda", dtype=torch.float32)
    ).contiguous()


def median_ms(fn, warmup: int, iterations: int) -> float:
    for _ in range(warmup):
        fn()
    torch.cuda.synchronize()
    samples: list[float] = []
    for _ in range(iterations):
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        start.record()
        fn()
        end.record()
        end.synchronize()
        samples.append(start.elapsed_time(end))
    return statistics.median(samples)


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--backend", choices=["source", "installed"], default="source"
    )
    parser.add_argument("--artifact", default=None)
    parser.add_argument("--registration-include", default=None)
    parser.add_argument("--warmup", type=int, default=10)
    parser.add_argument("--iterations", type=int, default=50)
    args = parser.parse_args()

    if not torch.cuda.is_available():
        raise RuntimeError("CUDA is required")
    ops = (
        load_source_ops(args.registration_include)
        if args.backend == "source"
        else load_installed_ops(args.artifact)
    )

    shapes = [(4096, 32), (1024, 64), (256, 128)]
    candidate_times: list[float] = []
    baseline_times: list[float] = []
    compiled_times: list[float] = []

    def pytorch_reference(
        input: torch.Tensor,
        output: torch.Tensor,
        info: torch.Tensor,
    ) -> torch.Tensor:
        torch.linalg.cholesky_ex(
            input,
            check_errors=False,
            out=(output, info),
        )
        return output

    compiled_reference = torch.compile(pytorch_reference, fullgraph=True)
    print(
        "batch,n,candidate_ms,pytorch_eager_ms,pytorch_compile_ms,"
        "speedup_eager,speedup_compile,candidate_tflops,io_gbps"
    )
    for batch, n in shapes:
        input = make_spd(batch, n)
        candidate_output = torch.empty_like(input)
        baseline_output = torch.empty_like(input)
        info = torch.empty(batch, device="cuda", dtype=torch.int32)

        def candidate() -> None:
            ops.cholesky_small_fp32(input, out=candidate_output)

        def baseline() -> None:
            pytorch_reference(input, baseline_output, info)

        def baseline_compiled() -> None:
            compiled_reference(input, baseline_output, info)

        candidate()
        baseline()
        baseline_compiled()
        torch.cuda.synchronize()
        torch.testing.assert_close(
            candidate_output,
            baseline_output,
            rtol=5e-4,
            atol=2e-4,
        )

        candidate_ms = median_ms(candidate, args.warmup, args.iterations)
        baseline_ms = median_ms(baseline, args.warmup, args.iterations)
        compiled_ms = median_ms(
            baseline_compiled, args.warmup, args.iterations
        )
        flops = batch * n**3 / 3.0
        tflops = flops / (candidate_ms * 1e-3) / 1e12
        io_bytes = 2 * batch * n * n * 4
        io_gbps = io_bytes / (candidate_ms * 1e-3) / 1e9
        candidate_times.append(candidate_ms)
        baseline_times.append(baseline_ms)
        compiled_times.append(compiled_ms)
        print(
            f"{batch},{n},{candidate_ms:.6f},{baseline_ms:.6f},"
            f"{compiled_ms:.6f},{baseline_ms / candidate_ms:.3f},"
            f"{compiled_ms / candidate_ms:.3f},{tflops:.3f},{io_gbps:.3f}"
        )

    candidate_geomean = math.exp(
        sum(math.log(value) for value in candidate_times)
        / len(candidate_times)
    )
    baseline_geomean = math.exp(
        sum(math.log(value) for value in baseline_times)
        / len(baseline_times)
    )
    compiled_geomean = math.exp(
        sum(math.log(value) for value in compiled_times)
        / len(compiled_times)
    )
    print(f"candidate_geomean_ms={candidate_geomean:.6f}")
    print(f"pytorch_geomean_ms={baseline_geomean:.6f}")
    print(f"pytorch_compile_geomean_ms={compiled_geomean:.6f}")
    print(
        f"geomean_speedup_eager={baseline_geomean / candidate_geomean:.3f}"
    )
    print(
        f"geomean_speedup_compile={compiled_geomean / candidate_geomean:.3f}"
    )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())