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
|
from itertools import product
import pytest
import torch
import torch_geometric.typing
from torch_geometric.profile import benchmark
from torch_geometric.testing import withCUDA, withDevice, withPackage
from torch_geometric.utils import group_argsort, group_cat, scatter
from torch_geometric.utils._scatter import scatter_argmax
def test_scatter_validate():
src = torch.randn(100, 32)
index = torch.randint(0, 10, (100, ), dtype=torch.long)
with pytest.raises(ValueError, match="must be one-dimensional"):
scatter(src, index.view(-1, 1))
with pytest.raises(ValueError, match="must lay between 0 and 1"):
scatter(src, index, dim=2)
with pytest.raises(ValueError, match="invalid `reduce` argument 'std'"):
scatter(src, index, reduce='std')
@withDevice
@withPackage('torch_scatter')
@pytest.mark.parametrize('reduce', ['sum', 'add', 'mean', 'min', 'max'])
def test_scatter(reduce, device):
import torch_scatter
src = torch.randn(100, 16, device=device)
index = torch.randint(0, 8, (100, ), device=device)
if device.type == 'mps' and reduce in ['min', 'max']:
with pytest.raises(NotImplementedError, match="for the MPS device"):
scatter(src, index, dim=0, reduce=reduce)
return
out1 = scatter(src, index, dim=0, reduce=reduce)
out2 = torch_scatter.scatter(src, index, dim=0, reduce=reduce)
assert out1.device == device
assert torch.allclose(out1, out2, atol=1e-6)
jit = torch.jit.script(scatter)
out3 = jit(src, index, dim=0, reduce=reduce)
assert torch.allclose(out1, out3, atol=1e-6)
src = torch.randn(8, 100, 16, device=device)
out1 = scatter(src, index, dim=1, reduce=reduce)
out2 = torch_scatter.scatter(src, index, dim=1, reduce=reduce)
assert out1.device == device
assert torch.allclose(out1, out2, atol=1e-6)
@withDevice
@pytest.mark.parametrize('reduce', ['sum', 'add', 'mean', 'min', 'max'])
def test_scatter_backward(reduce, device):
src = torch.randn(8, 100, 16, device=device, requires_grad=True)
index = torch.randint(0, 8, (100, ), device=device)
if device.type == 'mps' and reduce in ['min', 'max']:
with pytest.raises(NotImplementedError, match="for the MPS device"):
scatter(src, index, dim=1, reduce=reduce)
return
out = scatter(src, index, dim=1, reduce=reduce)
assert src.grad is None
out.mean().backward()
assert src.grad is not None
@withDevice
def test_scatter_any(device):
src = torch.randn(6, 4, device=device)
index = torch.tensor([0, 0, 1, 1, 2, 2], device=device)
out = scatter(src, index, dim=0, reduce='any')
for i in range(3):
for j in range(4):
assert float(out[i, j]) in src[2 * i:2 * i + 2, j].tolist()
@withDevice
@pytest.mark.parametrize('num_groups', [4])
@pytest.mark.parametrize('descending', [False, True])
def test_group_argsort(num_groups, descending, device):
src = torch.randn(20, device=device)
index = torch.randint(0, num_groups, (20, ), device=device)
out = group_argsort(src, index, 0, num_groups, descending=descending)
expected = torch.empty_like(index)
for i in range(num_groups):
mask = index == i
tmp = src[mask].argsort(descending=descending)
perm = torch.empty_like(tmp)
perm[tmp] = torch.arange(tmp.numel(), device=device)
expected[mask] = perm
assert torch.equal(out, expected)
empty_tensor = torch.tensor([], device=device)
out = group_argsort(empty_tensor, empty_tensor)
assert out.numel() == 0
@withCUDA
def test_scatter_argmax(device):
src = torch.arange(5, device=device)
index = torch.tensor([2, 2, 0, 0, 3], device=device)
old_state = torch_geometric.typing.WITH_TORCH_SCATTER
torch_geometric.typing.WITH_TORCH_SCATTER = False
argmax = scatter_argmax(src, index, dim_size=6)
torch_geometric.typing.WITH_TORCH_SCATTER = old_state
assert argmax.tolist() == [3, 5, 1, 4, 5, 5]
@withDevice
def test_group_cat(device):
x1 = torch.randn(4, 4, device=device)
x2 = torch.randn(2, 4, device=device)
index1 = torch.tensor([0, 0, 1, 2], device=device)
index2 = torch.tensor([0, 2], device=device)
expected = torch.cat([x1[:2], x2[:1], x1[2:4], x2[1:]], dim=0)
out, index = group_cat(
[x1, x2],
[index1, index2],
dim=0,
return_index=True,
)
assert torch.equal(out, expected)
assert index.tolist() == [0, 0, 0, 1, 2, 2]
if __name__ == '__main__':
# Insights on GPU:
# ================
# * "sum": Prefer `scatter_add_` implementation
# * "mean": Prefer manual implementation via `scatter_add_` + `count`
# * "min"/"max":
# * Prefer `scatter_reduce_` implementation without gradients
# * Prefer `torch_sparse` implementation with gradients
# * "mul": Prefer `torch_sparse` implementation
#
# Insights on CPU:
# ================
# * "sum": Prefer `scatter_add_` implementation
# * "mean": Prefer manual implementation via `scatter_add_` + `count`
# * "min"/"max": Prefer `scatter_reduce_` implementation
# * "mul" (probably not worth branching for this):
# * Prefer `scatter_reduce_` implementation without gradients
# * Prefer `torch_sparse` implementation with gradients
import argparse
from torch_geometric.typing import WITH_TORCH_SCATTER, torch_scatter
parser = argparse.ArgumentParser()
parser.add_argument('--device', type=str, default='cuda')
parser.add_argument('--backward', action='store_true')
parser.add_argument('--aggr', type=str, default='all')
args = parser.parse_args()
num_nodes_list = [4_000, 8_000, 16_000, 32_000, 64_000]
if args.aggr == 'all':
aggrs = ['sum', 'mean', 'min', 'max', 'mul']
else:
aggrs = args.aggr.split(',')
def pytorch_scatter(x, index, dim_size, reduce):
if reduce == 'min' or reduce == 'max':
reduce = f'a{aggr}' # `amin` or `amax`
elif reduce == 'mul':
reduce = 'prod'
out = x.new_zeros(dim_size, x.size(-1))
include_self = reduce in ['sum', 'mean']
index = index.view(-1, 1).expand(-1, x.size(-1))
out.scatter_reduce_(0, index, x, reduce, include_self=include_self)
return out
def pytorch_index_add(x, index, dim_size, reduce):
if reduce != 'sum':
raise NotImplementedError
out = x.new_zeros(dim_size, x.size(-1))
out.index_add_(0, index, x)
return out
def own_scatter(x, index, dim_size, reduce):
return torch_scatter.scatter(x, index, dim=0, dim_size=num_nodes,
reduce=reduce)
def optimized_scatter(x, index, dim_size, reduce):
return scatter(x, index, dim=0, dim_size=dim_size, reduce=reduce)
for aggr, num_nodes in product(aggrs, num_nodes_list):
num_edges = num_nodes * 50
print(f'aggr: {aggr}, #nodes: {num_nodes}, #edges: {num_edges}')
x = torch.randn(num_edges, 64, device=args.device)
index = torch.randint(num_nodes, (num_edges, ), device=args.device)
funcs = [pytorch_scatter]
func_names = ['PyTorch scatter_reduce']
if aggr == 'sum':
funcs.append(pytorch_index_add)
func_names.append('PyTorch index_add')
if WITH_TORCH_SCATTER:
funcs.append(own_scatter)
func_names.append('torch_scatter')
funcs.append(optimized_scatter)
func_names.append('Optimized PyG Scatter')
benchmark(
funcs=funcs,
func_names=func_names,
args=(x, index, num_nodes, aggr),
num_steps=100 if args.device == 'cpu' else 1000,
num_warmups=50 if args.device == 'cpu' else 500,
backward=args.backward,
)
|
