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
256
257
258
259
260
261
262
263
264
|
# ----------------------------------------------------------------------------
# - Open3D: www.open3d.org -
# ----------------------------------------------------------------------------
# Copyright (c) 2018-2024 www.open3d.org
# SPDX-License-Identifier: MIT
# ----------------------------------------------------------------------------
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = '1' # does not affect results
import argparse
import pickle
from collections import OrderedDict
import numpy as np
import open3d as o3d
from scipy.spatial import cKDTree
import nvidia_smi
import matplotlib.pyplot as plt
from benchmark_utils import measure_memory, print_system_info, print_table_memory
# Define NNS methods
class NNS:
def __init__(self, device, search_type, index_type):
assert index_type in ["int", "long"]
self.device = device
self.search_type = search_type
self.index_type = o3d.core.Int32 if index_type == "int" else o3d.core.Int64
def setup(self, points, queries, radius):
points_dev = points.to(self.device)
queries_dev = queries.to(self.device)
index = o3d.core.nns.NearestNeighborSearch(points_dev, self.index_type)
if self.search_type == "knn":
index.knn_index()
elif self.search_type == "radius":
index.fixed_radius_index(radius)
elif self.search_type == "hybrid":
index.hybrid_index(radius)
else:
raise ValueError(f"{self.search_type} is not supported.")
return index, queries_dev
def search(self, index, queries, search_args):
if self.search_type == "knn":
out = index.knn_search(queries, search_args["knn"])
elif self.search_type == "radius":
out = index.fixed_radius_search(queries, search_args["radius"])
elif self.search_type == "hybrid":
out = index.hybrid_search(queries, search_args["radius"],
search_args["knn"])
else:
raise ValueError(f"{self.search_type} is not supported.")
return out
def __str__(self):
return f"{self.search_type.capitalize()}({self.device}, {self.index_type})_memory"
def compute_avg_radii(points, queries, neighbors):
"""Computes the radii based on the number of neighbors"""
tree = cKDTree(points.numpy())
avg_radii = []
for k in neighbors:
dist, _ = tree.query(queries.numpy(), k=k + 1)
avg_radii.append(np.mean(dist.max(axis=-1)))
return avg_radii
def prepare_benchmark_data():
# setup dataset examples
datasets = OrderedDict()
# random dataset
out_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)),
"testdata")
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
log10_ns = [4, 5, 6, 7]
for log10_n in log10_ns:
print("==================================")
npy_file = os.path.join(out_dir, f"random_1e{log10_n}.npy")
if not os.path.exists(npy_file):
print(f"generating a random dataset, random_1e{log10_n}.npy...")
N = int(np.power(10, log10_n))
points = np.random.randn(N, 3)
np.save(npy_file, points)
print(f"loading the random dataset, random_1e{log10_n}.npy...")
points = queries = o3d.core.Tensor(np.load(npy_file),
dtype=o3d.core.Float32)
queries = queries[::10]
filename = os.path.basename(npy_file)
datasets[filename] = {'points': points, 'queries': queries}
print("")
return datasets
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--search_type",
type=str,
default="knn",
choices=["knn", "radius", "hybrid", "all"])
parser.add_argument("--overwrite", action="store_true")
parser.add_argument("--gpu_idx", type=int, default=3)
args = parser.parse_args()
# devices
nvidia_smi.nvmlInit()
handle = nvidia_smi.nvmlDeviceGetHandleByIndex(args.gpu_idx)
o3d_cpu_dev = o3d.core.Device()
o3d_cuda_dev = o3d.core.Device(o3d.core.Device.CUDA, 0)
# collects runtimes for all examples
results = OrderedDict()
datasets = prepare_benchmark_data()
# prepare methods
if args.search_type == "all":
methods = [
NNS(o3d_cuda_dev, "knn", "int"),
NNS(o3d_cuda_dev, "knn", "long"),
NNS(o3d_cuda_dev, "radius", "int"),
NNS(o3d_cuda_dev, "radius", "long"),
NNS(o3d_cuda_dev, "hybrid", "int"),
NNS(o3d_cuda_dev, "hybrid", "long"),
]
else:
methods = [
NNS(o3d_cuda_dev, args.search_type, "int"),
NNS(o3d_cuda_dev, args.search_type, "long"),
]
neighbors = [int(2**p) for p in range(12)]
# run benchmark
for method in methods:
if not args.overwrite and os.path.exists(f"{method}.pkl"):
print(f"skip {method}")
continue
print(method)
for example_name, example in datasets.items():
points, queries = example['points'], example['queries']
if args.search_type == "knn":
radii = neighbors
else:
radii = compute_avg_radii(points, queries, neighbors)
print(f"{example_name} {points.shape[0]}")
for (knn, radius) in zip(neighbors, radii):
points, queries = example['points'], example['queries']
points = points.contiguous().to(o3d_cuda_dev)
queries = queries.contiguous().to(o3d_cuda_dev)
example_results = {'k': knn, 'num_points': points.shape[0]}
if hasattr(method, "prepare_data"):
points, queries = method.prepare_data(points, queries)
index, queries = method.setup(points, queries, radius)
memory = measure_memory(
lambda: method.search(index, queries,
dict(knn=knn, radius=radius)), handle)
example_results['memory'] = memory
results[
f'{example_name} n={points.shape[0]} k={knn}'] = example_results
del index
del points
del queries
o3d.core.cuda.release_cache()
with open(f"{method}.pkl", 'wb') as f:
pickle.dump(results, f)
results = []
for method in methods:
with open(f"{method}.pkl", "rb") as f:
print(f"{method}.pkl")
data = pickle.load(f)
results.append(data)
print_system_info()
print_table_memory(methods, results)
fig = plt.figure(figsize=(10, 10))
ax1 = fig.add_subplot(2, 2, 1)
ax2 = fig.add_subplot(2, 2, 2)
ax3 = fig.add_subplot(2, 2, 3)
ax4 = fig.add_subplot(2, 2, 4)
dtypes = ["int32", "int64"]
colors = ["b", "r"]
lines = ["^", "o"]
for idx, result in enumerate(results): # int, long
ks = [[], [], [], []] # num_points
ms = [[], [], [], []]
for value in result.values():
if value['num_points'] == int(np.power(10, 4)):
ks[0].append(value['k'])
ms[0].append(value['memory'])
elif value['num_points'] == int(np.power(10, 5)):
ks[1].append(value['k'])
ms[1].append(value['memory'])
elif value['num_points'] == int(np.power(10, 6)):
ks[2].append(value['k'])
ms[2].append(value['memory'])
elif value['num_points'] == int(np.power(10, 7)):
ks[3].append(value['k'])
ms[3].append(value['memory'])
else:
raise ValueError
ax1.plot(ks[0],
ms[0],
marker=lines[idx],
color=colors[idx],
label=dtypes[idx])
ax2.plot(ks[1],
ms[1],
marker=lines[idx],
color=colors[idx],
label=dtypes[idx])
ax3.plot(ks[2],
ms[2],
marker=lines[idx],
color=colors[idx],
label=dtypes[idx])
ax4.plot(ks[3],
ms[3],
marker=lines[idx],
color=colors[idx],
label=dtypes[idx])
ax1.set_title(f"{args.search_type}: N={int(np.power(10, 4))}")
ax1.set_xlabel("K")
ax1.set_ylabel("Memory (GB)")
ax1.legend()
ax2.set_title(f"{args.search_type}: N={int(np.power(10, 5))}")
ax2.set_xlabel("K")
ax2.set_ylabel("Memory (GB)")
ax2.legend()
ax3.set_title(f"{args.search_type}: N={int(np.power(10, 6))}")
ax3.set_xlabel("K")
ax3.set_ylabel("Memory (GB)")
ax3.legend()
ax4.set_title(f"{args.search_type}: N={int(np.power(10, 7))}")
ax4.set_xlabel("K")
ax4.set_ylabel("Memory (GB)")
ax4.legend()
plt.show()
plt.savefig("memory.png")
|
