File size: 8,049 Bytes
89650c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import sys

import torch.backends.openmp
from torch import autograd
from torch.utils import cpp_extension

module = sys.modules[__name__]


class RSPMMAddMulFunction(autograd.Function):

    @staticmethod
    def forward(ctx, edge_index, edge_type, edge_weight, relation, input):
        node_in, node_out = edge_index
        key = node_in * (node_out.max() + 1) + node_out
        assert (key.diff() >= 0).all(), "Expect sorted `edge_index`"

        if input.device.type == "cuda":
            forward = rspmm.rspmm_add_mul_forward_cuda
        else:
            forward = rspmm.rspmm_add_mul_forward_cpu
        output = forward(edge_index, edge_type, edge_weight, relation, input)
        ctx.save_for_backward(edge_index, edge_type, edge_weight, relation, input, output)
        return output

    @staticmethod
    def backward(ctx, output_grad):
        if output_grad.device.type == "cuda":
            backward = rspmm.rspmm_add_mul_backward_cuda
        else:
            backward = rspmm.rspmm_add_mul_backward_cpu
        weight_grad, relation_grad, input_grad = backward(*ctx.saved_tensors, output_grad)
        return None, None, weight_grad, relation_grad, input_grad


class RSPMMMinMulFunction(autograd.Function):

    @staticmethod
    def forward(ctx, edge_index, edge_type, edge_weight, relation, input):
        node_in, node_out = edge_index
        key = node_in * (node_out.max() + 1) + node_out
        assert (key.diff() >= 0).all(), "Expect sorted `edge_index`"

        if input.device.type == "cuda":
            forward = rspmm.rspmm_min_mul_forward_cuda
        else:
            forward = rspmm.rspmm_min_mul_forward_cpu
        output = forward(edge_index, edge_type, edge_weight, relation, input)
        ctx.save_for_backward(edge_index, edge_type, edge_weight, relation, input, output)
        return output

    @staticmethod
    def backward(ctx, output_grad):
        if output_grad.device.type == "cuda":
            backward = rspmm.rspmm_min_mul_backward_cuda
        else:
            backward = rspmm.rspmm_min_mul_backward_cpu
        weight_grad, relation_grad, input_grad = backward(*ctx.saved_tensors, output_grad)
        return None, None, weight_grad, relation_grad, input_grad


class RSPMMMaxMulFunction(autograd.Function):

    @staticmethod
    def forward(ctx, edge_index, edge_type, edge_weight, relation, input):
        node_in, node_out = edge_index
        key = node_in * (node_out.max() + 1) + node_out
        assert (key.diff() >= 0).all(), "Expect sorted `edge_index`"

        if input.device.type == "cuda":
            forward = rspmm.rspmm_max_mul_forward_cuda
        else:
            forward = rspmm.rspmm_max_mul_forward_cpu
        output = forward(edge_index, edge_type, edge_weight, relation, input)
        ctx.save_for_backward(edge_index, edge_type, edge_weight, relation, input, output)
        return output

    @staticmethod
    def backward(ctx, output_grad):
        if output_grad.device.type == "cuda":
            backward = rspmm.rspmm_max_mul_backward_cuda
        else:
            backward = rspmm.rspmm_max_mul_backward_cpu
        weight_grad, relation_grad, input_grad = backward(*ctx.saved_tensors, output_grad)
        return None, None, weight_grad, relation_grad, input_grad


class RSPMMAddAddFunction(autograd.Function):

    @staticmethod
    def forward(ctx, edge_index, edge_type, edge_weight, relation, input):
        node_in, node_out = edge_index
        key = node_in * (node_out.max() + 1) + node_out
        assert (key.diff() >= 0).all(), "Expect sorted `edge_index`"

        if input.device.type == "cuda":
            forward = rspmm.rspmm_add_add_forward_cuda
        else:
            forward = rspmm.rspmm_add_add_forward_cpu
        output = forward(edge_index, edge_type, edge_weight, relation, input)
        ctx.save_for_backward(edge_index, edge_type, edge_weight, relation, input, output)
        return output

    @staticmethod
    def backward(ctx, output_grad):
        if output_grad.device.type == "cuda":
            backward = rspmm.rspmm_add_add_backward_cuda
        else:
            backward = rspmm.rspmm_add_add_backward_cpu
        weight_grad, relation_grad, input_grad = backward(*ctx.saved_tensors, output_grad)
        return None, None, weight_grad, relation_grad, input_grad


class RSPMMMinAddFunction(autograd.Function):

    @staticmethod
    def forward(ctx, edge_index, edge_type, edge_weight, relation, input):
        node_in, node_out = edge_index
        key = node_in * (node_out.max() + 1) + node_out
        assert (key.diff() >= 0).all(), "Expect sorted `edge_index`"

        if input.device.type == "cuda":
            forward = rspmm.rspmm_min_add_forward_cuda
        else:
            forward = rspmm.rspmm_min_add_forward_cpu
        output = forward(edge_index, edge_type, edge_weight, relation, input)
        ctx.save_for_backward(edge_index, edge_type, edge_weight, relation, input, output)
        return output

    @staticmethod
    def backward(ctx, output_grad):
        if output_grad.device.type == "cuda":
            backward = rspmm.rspmm_min_add_backward_cuda
        else:
            backward = rspmm.rspmm_min_add_backward_cpu
        weight_grad, relation_grad, input_grad = backward(*ctx.saved_tensors, output_grad)
        return None, None, weight_grad, relation_grad, input_grad


class RSPMMMaxAddFunction(autograd.Function):

    @staticmethod
    def forward(ctx, edge_index, edge_type, edge_weight, relation, input):
        node_in, node_out = edge_index
        key = node_in * (node_out.max() + 1) + node_out
        assert (key.diff() >= 0).all(), "Expect sorted `edge_index`"

        if input.device.type == "cuda":
            forward = rspmm.rspmm_max_add_forward_cuda
        else:
            forward = rspmm.rspmm_max_add_forward_cpu
        output = forward(edge_index, edge_type, edge_weight, relation, input)
        ctx.save_for_backward(edge_index, edge_type, edge_weight, relation, input, output)
        return output

    @staticmethod
    def backward(ctx, output_grad):
        if output_grad.device.type == "cuda":
            backward = rspmm.rspmm_max_add_backward_cuda
        else:
            backward = rspmm.rspmm_max_add_backward_cpu
        weight_grad, relation_grad, input_grad = backward(*ctx.saved_tensors, output_grad)
        return None, None, weight_grad, relation_grad, input_grad


def generalized_rspmm(edge_index, edge_type, edge_weight, relation, input, sum="add", mul="mul"):
    name = "RSPMM%s%sFunction" % (sum.capitalize(), mul.capitalize())
    if not hasattr(module, name):
        raise ValueError("No generalized rspmm implementation found for summation `%s` and multiplication `%s`"
                         % (sum, mul))
    Function = getattr(module, name)

    node_in, node_out = edge_index
    key = node_in * (node_out.max() + 1) + node_out
    order = key.argsort()

    return Function.apply(edge_index[:, order], edge_type[order], edge_weight[order], relation, input)


def load_extension(name, sources, extra_cflags=None, extra_cuda_cflags=None, **kwargs):
    if extra_cflags is None:
        extra_cflags = ["-Ofast"]
        if torch.backends.openmp.is_available():
            extra_cflags += ["-fopenmp", "-DAT_PARALLEL_OPENMP"]
        else:
            extra_cflags.append("-DAT_PARALLEL_NATIVE")
    if extra_cuda_cflags is None:
        if torch.cuda.is_available():
            extra_cuda_cflags = ["-O3"]
            extra_cflags.append("-DCUDA_OP")
        else:
            new_sources = []
            for source in sources:
                if not cpp_extension._is_cuda_file(source):
                    new_sources.append(source)
            sources = new_sources

    return cpp_extension.load(name, sources, extra_cflags, extra_cuda_cflags, **kwargs)


print("Load rspmm extension. This may take a while...")
path = os.path.join(os.path.dirname(__file__), "source")
rspmm = load_extension("rspmm", [os.path.join(path, "rspmm.cpp"), os.path.join(path, "rspmm.cu")])