segment_reduce.cu 5.61 KB
Newer Older
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
/*!
 *  Copyright (c) 2020 by Contributors
 * \file array/cuda/segment_reduce.cu
 * \brief Segment reduce C APIs and definitions.
 */
#include <dgl/array.h>
#include <dgl/base_heterograph.h>
#include "./segment_reduce.cuh"
#include "./functor.cuh"
#include "./utils.h"


namespace dgl {

using namespace cuda;

namespace aten {


template <int XPU, typename IdType, int bits>
void SegmentReduce(const std::string& op,
                   NDArray feat,
                   NDArray offsets,
                   NDArray out,
                   NDArray arg) {
  SWITCH_BITS(bits, DType, {
    if (op == "sum") {
      cuda::SegmentReduce<IdType, DType, cuda::reduce::Sum<IdType, DType>>(
          feat, offsets, out, arg);
    } else if (op == "max") {
      cuda::SegmentReduce<IdType, DType, cuda::reduce::Max<IdType, DType>>(
          feat, offsets, out, arg);
    } else if (op == "min") {
      cuda::SegmentReduce<IdType, DType, cuda::reduce::Min<IdType, DType>>(
          feat, offsets, out, arg);
    } else {
      LOG(FATAL) << "Not implemented";
    }
  });
}


template <int XPU, typename IdType, int bits>
void ScatterAdd(NDArray feat,
                NDArray idx,
                NDArray out) {
  SWITCH_BITS(bits, DType, {
    cuda::ScatterAdd<IdType, DType>(feat, idx, out);
  });
}


template <int XPU, typename IdType, int bits>
void UpdateGradMinMax_hetero(const HeteroGraphPtr& g,
                const std::string& op,
                const std::vector<NDArray>& feat,
                const std::vector<NDArray>& idx,
                const std::vector<NDArray>& idx_etype,
                std::vector<NDArray>* out) {
  SWITCH_BITS(bits, DType, {
    cuda::UpdateGradMinMax_hetero<IdType, DType>(g, op, feat, idx, idx_etype, out);
  });
}


template <int XPU, typename IdType, int bits>
void BackwardSegmentCmp(NDArray feat,
                        NDArray arg,
                        NDArray out) {
  SWITCH_BITS(bits, DType, {
    cuda::BackwardSegmentCmp<IdType, DType>(feat, arg, out);
  });
}


template void SegmentReduce<kDLGPU, int32_t, 16>(
    const std::string& op,
    NDArray feat,
    NDArray offsets,
    NDArray out,
    NDArray arg);
template void SegmentReduce<kDLGPU, int64_t, 16>(
    const std::string &op,
    NDArray feat,
    NDArray offsets,
    NDArray out,
    NDArray arg);
template void SegmentReduce<kDLGPU, int32_t, 32>(
    const std::string& op,
    NDArray feat,
    NDArray offsets,
    NDArray out,
    NDArray arg);
template void SegmentReduce<kDLGPU, int64_t, 32>(
    const std::string &op,
    NDArray feat,
    NDArray offsets,
    NDArray out,
    NDArray arg);
template void SegmentReduce<kDLGPU, int32_t, 64>(
    const std::string &op,
    NDArray feat,
    NDArray offsets,
    NDArray out,
    NDArray arg);
template void SegmentReduce<kDLGPU, int64_t, 64>(
    const std::string &op,
    NDArray feat,
    NDArray offsets,
    NDArray out,
    NDArray arg);
template void ScatterAdd<kDLGPU, int32_t, 16>(
    NDArray feat,
    NDArray idx,
    NDArray out);
template void ScatterAdd<kDLGPU, int64_t, 16>(
    NDArray feat,
    NDArray idx,
    NDArray out);
template void ScatterAdd<kDLGPU, int32_t, 32>(
    NDArray feat,
    NDArray idx,
    NDArray out);
template void ScatterAdd<kDLGPU, int64_t, 32>(
    NDArray feat,
    NDArray idx,
    NDArray out);
template void ScatterAdd<kDLGPU, int32_t, 64>(
    NDArray feat,
    NDArray idx,
    NDArray out);
template void ScatterAdd<kDLGPU, int64_t, 64>(
    NDArray feat,
    NDArray idx,
    NDArray out);

template void UpdateGradMinMax_hetero<kDLGPU, int32_t, 16>(
    const HeteroGraphPtr& g, const std::string& op,
    const std::vector<NDArray>& feat, const std::vector<NDArray>& idx,
    const std::vector<NDArray>& idx_etype, std::vector<NDArray>* out);
template void UpdateGradMinMax_hetero<kDLGPU, int64_t, 16>(
    const HeteroGraphPtr& g, const std::string& op,
    const std::vector<NDArray>& feat, const std::vector<NDArray>& idx,
    const std::vector<NDArray>& idx_etype, std::vector<NDArray>* out);
template void UpdateGradMinMax_hetero<kDLGPU, int32_t, 32>(
    const HeteroGraphPtr& g, const std::string& op,
    const std::vector<NDArray>& feat, const std::vector<NDArray>& idx,
    const std::vector<NDArray>& idx_etype, std::vector<NDArray>* out);
template void UpdateGradMinMax_hetero<kDLGPU, int64_t, 32>(
    const HeteroGraphPtr& g, const std::string& op,
    const std::vector<NDArray>& feat, const std::vector<NDArray>& idx,
    const std::vector<NDArray>& idx_etype, std::vector<NDArray>* out);
template void UpdateGradMinMax_hetero<kDLGPU, int32_t, 64>(
    const HeteroGraphPtr& g, const std::string& op,
    const std::vector<NDArray>& feat, const std::vector<NDArray>& idx,
    const std::vector<NDArray>& idx_etype, std::vector<NDArray>* out);
template void UpdateGradMinMax_hetero<kDLGPU, int64_t, 64>(
    const HeteroGraphPtr& g, const std::string& op,
    const std::vector<NDArray>& feat, const std::vector<NDArray>& idx,
    const std::vector<NDArray>& idx_etype, std::vector<NDArray>* out);

template void BackwardSegmentCmp<kDLGPU, int32_t, 16>(
    NDArray feat,
    NDArray arg,
    NDArray out);
template void BackwardSegmentCmp<kDLGPU, int64_t, 16>(
    NDArray feat,
    NDArray arg,
    NDArray out);
template void BackwardSegmentCmp<kDLGPU, int32_t, 32>(
    NDArray feat,
    NDArray arg,
    NDArray out);
template void BackwardSegmentCmp<kDLGPU, int64_t, 32>(
    NDArray feat,
    NDArray arg,
    NDArray out);
template void BackwardSegmentCmp<kDLGPU, int32_t, 64>(
    NDArray feat,
    NDArray arg,
    NDArray out);
template void BackwardSegmentCmp<kDLGPU, int64_t, 64>(
    NDArray feat,
    NDArray arg,
    NDArray out);

}  // namespace aten
}  // namespace dgl