-
Notifications
You must be signed in to change notification settings - Fork 25
/
Copy pathutilities.py
166 lines (127 loc) · 4.19 KB
/
utilities.py
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
import time
import math
import collections
import importlib
import os
import numpy as np
class Timer:
def __init__(self):
self.elapsed = float("nan")
def __enter__(self):
self._start = time.perf_counter()
return self
def __exit__(self, type, value, traceback):
if value is None:
self.elapsed = time.perf_counter() - self._start
def estimate_repetitions(func, args=(), target_time=10, powers_of=10):
# call function once for warm-up
func(*args)
# some backends need an extra nudge (looking at you, PyTorch)
func(*args)
# call again and measure time
with Timer() as t:
func(*args)
time_per_rep = t.elapsed
exponent = math.log(target_time / time_per_rep, powers_of)
num_reps = int(powers_of ** round(exponent))
return max(powers_of, num_reps)
def compute_statistics(timings, burnin=1):
stats = []
for (backend, size), t in timings.items():
t = t[burnin:]
repetitions = len(t)
if repetitions:
mean = np.mean(t)
stdev = np.std(t)
percentiles = np.percentile(t, [0, 25, 50, 75, 100])
else:
mean = stdev = float("nan")
percentiles = [float("nan")] * 5
stats.append(
(size, backend, repetitions, mean, stdev, *percentiles, float("nan"))
)
stats = np.array(
stats,
dtype=[
("size", "i8"),
("backend", object),
("calls", "i8"),
("mean", "f4"),
("stdev", "f4"),
("min", "f4"),
("25%", "f4"),
("median", "f4"),
("75%", "f4"),
("max", "f4"),
("Δ", "f4"),
],
)
# add deltas
sizes = np.unique(stats["size"])
for s in sizes:
mask = stats["size"] == s
# measure relative to NumPy if present, otherwise worst backend
if "numpy" in stats["backend"][mask]:
reference_time = stats["mean"][mask & (stats["backend"] == "numpy")]
else:
reference_time = np.nanmax(stats["mean"][mask])
stats["Δ"][mask] = reference_time / stats["mean"][mask]
return stats
def format_output(stats, benchmark_title, device="cpu"):
stats = np.sort(stats, axis=0, order=["size", "mean", "max", "median"])
header = stats.dtype.names
col_widths = collections.defaultdict(lambda: 8)
col_widths.update(size=12, backend=10)
def format_col(col_name, value, is_time=False):
col_width = col_widths[col_name]
if np.issubdtype(type(value), np.integer):
typecode = ","
else:
typecode = ".3f"
if is_time:
format_string = f"{{value:>{col_width}{typecode}}}"
else:
format_string = f"{{value:<{col_width}}}"
return format_string.format(value=value)
out = [
"",
benchmark_title,
"=" * len(benchmark_title),
f"Running on {device.upper()}",
"",
" ".join(format_col(s, s) for s in header),
]
out.append("-" * len(out[-1]))
current_size = None
for row in stats:
# print empty line on size change
size = row[0]
if current_size is not None and size != current_size:
out.append("")
current_size = size
out.append(
" ".join(
format_col(n, s, not isinstance(s, str)) for n, s in zip(header, row)
)
)
out.extend(
[
"",
"(time in wall seconds, less is better)",
]
)
return "\n".join(out)
def get_benchmark_module(file_path):
base_path = os.path.dirname(os.path.abspath(__file__))
module_path = os.path.relpath(file_path, base_path)
import_path = ".".join(os.path.split(module_path))
bm_module = importlib.import_module(import_path)
return bm_module, import_path
def check_consistency(res1, res2):
if isinstance(res1, (tuple, list)):
if not len(res1) == len(res2):
return False
return all(check_consistency(r1, r2) for r1, r2 in zip(res1, res2))
assert isinstance(res1, np.ndarray)
assert isinstance(res2, np.ndarray)
return np.allclose(res1, res2)