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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
|
import torch
import jax
import einx
import timeit
import types
import einops
import random
import argparse
import math
import jax.numpy as jnp
import numpy as np
from functools import partial
from collections import defaultdict
parser = argparse.ArgumentParser()
parser.add_argument("--n", type=int, default=1000)
args = parser.parse_args()
k = 1
n = args.n // k
rows = []
envs = [
types.SimpleNamespace(
name="numpy",
backend=einx.backend.get("numpy"),
jit=lambda x: x,
block_until_ready=lambda x: x,
to_numpy=np.asarray,
ones=lambda shape, dtype: np.ones(shape, dtype=dtype),
transpose=np.transpose,
mean=np.mean,
einsum=np.einsum,
),
types.SimpleNamespace(
name="torch-eager",
backend=einx.backend.get("torch"),
jit=lambda x: x,
block_until_ready=lambda x: torch.cuda.synchronize(),
to_numpy=lambda x: np.asarray(x.cpu()),
ones=lambda shape, dtype: torch.ones(shape, dtype=vars(torch)[dtype]).cuda(),
transpose=torch.permute,
mean=torch.mean,
einsum=torch.einsum,
),
types.SimpleNamespace(
name="torch-compile",
backend=einx.backend.get("torch"),
jit=torch.compile,
block_until_ready=lambda x: torch.cuda.synchronize(),
to_numpy=lambda x: np.asarray(x.cpu()),
ones=lambda shape, dtype: torch.ones(shape, dtype=vars(torch)[dtype]).cuda(),
transpose=torch.permute,
mean=torch.mean,
einsum=torch.einsum,
),
types.SimpleNamespace(
name="jax-jit",
backend=einx.backend.get("jax"),
jit=jax.jit,
block_until_ready=lambda x: x.block_until_ready(),
to_numpy=lambda x: np.asarray(x),
ones=lambda shape, dtype: jnp.ones(shape, dtype=dtype),
transpose=jnp.transpose,
mean=jnp.mean,
einsum=jnp.einsum,
),
]
for env in envs:
experiments = []
f = 4 if env.name == "numpy" else 1
x = env.ones((16 // f, 512 // f, 512 // f, 64 // f), "float32")
x2 = env.ones((16 // f, 256 // f, 256 // f, 64 // f), "float32")
y = env.ones((512 // f, 512 // f), "float32")
z1 = env.ones((64 // f,), "float32")
w = env.ones((64 // f, 128 // f), "float32")
def benchmark_einx(x):
return einx.rearrange("b h w c -> b c h w", x, backend=env.backend)
def benchmark_einops(x):
return einops.rearrange(x, "b h w c -> b c h w")
def benchmark_idx(x):
return env.transpose(x, (0, 3, 1, 2))
experiments.append(("rearrange", (benchmark_einx, benchmark_einops, benchmark_idx), (x,), 5.0))
def benchmark_einx(x):
return einx.mean("b [s...] c", x)
def benchmark_einops(x):
return einops.reduce(x, "b h w c -> b c", reduction="mean")
def benchmark_idx(x):
return env.mean(x, axis=(1, 2))
experiments.append((
"spatial_mean",
(benchmark_einx, benchmark_einops, benchmark_idx),
(x,),
5.0,
))
def benchmark_einx(x):
return einx.mean("b s... [c]", x)
def benchmark_einops(x):
return einops.reduce(x, "b h w c -> b h w", reduction="mean")
def benchmark_idx(x):
return env.mean(x, axis=3)
experiments.append((
"channel_mean",
(benchmark_einx, benchmark_einops, benchmark_idx),
(x,),
5.0,
))
def benchmark_einx(x, y):
return einx.add("b [s...] c", x, y)
def benchmark_idx(x, y):
return x + y[None, ..., None]
experiments.append(("spatial_add", (benchmark_einx, None, benchmark_idx), (x, y), 5.0))
def benchmark_einx(x, y):
return einx.add("b s... [c]", x, y)
def benchmark_idx(x, y):
return x + y
experiments.append(("channel_add", (benchmark_einx, None, benchmark_idx), (x, z1), 5.0))
def benchmark_einx(x, w):
return einx.dot("b... [c1->c2]", x, w)
def benchmark_einops(x, w):
return einops.einsum(x, w, "... c1, c1 c2 -> ... c2")
def benchmark_idx(x, w):
return env.einsum("b h w c, c d -> b h w d", x, w)
experiments.append(("einsum", (benchmark_einx, benchmark_einops, benchmark_idx), (x2, w), 5.0))
for name, methods, inputs, mul in experiments:
name = env.name + " " + name
print(name)
# Assert correctness
results = []
for method in methods:
if method is not None:
results.append(method(*inputs))
results = [env.to_numpy(r) for r in results]
for r2 in results[1:]:
assert np.allclose(results[0], r2)
# Initialization
for _ in range(5):
for method in methods:
if method is not None:
env.block_until_ready(method(*inputs))
methods = [env.jit(m) if m is not None else None for m in methods]
for _ in range(5):
for method in methods:
if method is not None:
env.block_until_ready(method(*inputs))
# Benchmark
times = defaultdict(list)
order = "random"
if order == "random":
methods2 = list(methods)
for _ in range(max(1, int(n * mul))):
random.shuffle(methods2)
for method in methods2:
if method is not None:
times[method.__name__].append(
timeit.repeat(
lambda: env.block_until_ready(method(*inputs)), repeat=1, number=k
)[0]
/ k
)
elif order == "sequential":
for method in methods:
if method is not None:
for _ in range(max(1, int(n * mul))):
times[method.__name__].append(
timeit.repeat(
lambda: env.block_until_ready(method(*inputs)), repeat=1, number=k
)[0]
/ k
)
else:
raise AssertionError()
# Store and print results
for key in list(times.keys()):
p = int(len(times[key]) * 0.2)
times[key] = sorted(times[key])[p:-p]
for method in methods:
if method is not None:
print(
f"{method.__name__:>25}: {1000.0 * np.mean(times[method.__name__]):0.6f} "
f"+- {1000.0 * np.std(times[method.__name__]):0.6f}"
)
rows.append((name, times))
print()
# Print markup table
import tabulate
table = []
def tostr(times):
if len(times) == 0 or times is None:
return ""
m = f"{np.mean(times):0.3f}"
s = f"{np.std(times):0.3f}"
return f"{m:>7} +- {s:>7}"
for name, times in rows:
times = {k: np.asarray(v) for k, v in times.items()}
table.append([
name,
1000000.0 * (np.mean(times["benchmark_einx"]) - np.mean(times["benchmark_idx"])),
1000000.0 * (np.mean(times["benchmark_einops"]) - np.mean(times["benchmark_idx"]))
if "benchmark_einops" in times
else "",
tostr(1000.0 * times["benchmark_einx"]),
tostr(1000.0 * times["benchmark_einops"]) if "benchmark_einops" in times else "",
tostr(1000.0 * times["benchmark_idx"]),
])
print(
tabulate.tabulate(
table,
headers=[
"Method",
"einx overhead (us)",
"einops overhead (us)",
"einx (ms)",
"einops (ms)",
"index-based (ms)",
],
tablefmt="github",
)
)
|
