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
|
#include "graclus_cuda.h"
#include <ATen/cuda/CUDAContext.h>
#include "utils.cuh"
#define THREADS 1024
#define BLOCKS(N) (N + THREADS - 1) / THREADS
#define BLUE_P 0.53406
__device__ bool done_d;
__global__ void init_done_kernel() { done_d = true; }
__global__ void colorize_kernel(int64_t *out, const float *bernoulli,
int64_t numel) {
const int64_t thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
if (thread_idx < numel) {
if (out[thread_idx] < 0) {
out[thread_idx] = (int64_t)bernoulli[thread_idx] - 2;
done_d = false;
}
}
}
bool colorize(torch::Tensor out) {
auto stream = at::cuda::getCurrentCUDAStream();
init_done_kernel<<<1, 1, 0, stream>>>();
auto numel = out.size(0);
auto props = torch::full(numel, BLUE_P, out.options().dtype(torch::kFloat));
auto bernoulli = props.bernoulli();
colorize_kernel<<<BLOCKS(numel), THREADS, 0, stream>>>(
out.data_ptr<int64_t>(), bernoulli.data_ptr<float>(), numel);
bool done_h;
cudaMemcpyFromSymbol(&done_h, done_d, sizeof(done_h), 0,
cudaMemcpyDeviceToHost);
return done_h;
}
__global__ void propose_kernel(int64_t *out, int64_t *proposal,
const int64_t *rowptr, const int64_t *col,
int64_t numel) {
const int64_t thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
if (thread_idx < numel) {
if (out[thread_idx] != -1)
return; // Only vist blue nodes.
bool has_unmatched_neighbor = false;
for (int64_t i = rowptr[thread_idx]; i < rowptr[thread_idx + 1]; i++) {
auto v = col[i];
if (out[v] < 0)
has_unmatched_neighbor = true; // Unmatched neighbor found.
if (out[v] == -2) {
proposal[thread_idx] = v; // Propose to first red neighbor.
break;
}
}
if (!has_unmatched_neighbor)
out[thread_idx] = thread_idx;
}
}
template <typename scalar_t>
__global__ void weighted_propose_kernel(int64_t *out, int64_t *proposal,
const int64_t *rowptr,
const int64_t *col,
const scalar_t *weight, int64_t numel) {
const int64_t thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
if (thread_idx < numel) {
if (out[thread_idx] != -1)
return; // Only vist blue nodes.
bool has_unmatched_neighbor = false;
int64_t v_max = -1;
scalar_t w_max = 0;
for (int64_t i = rowptr[thread_idx]; i < rowptr[thread_idx + 1]; i++) {
auto v = col[i];
if (out[v] < 0)
has_unmatched_neighbor = true; // Unmatched neighbor found.
// Find maximum weighted red neighbor.
if (out[v] == -2 && weight[i] >= w_max) {
v_max = v;
w_max = weight[i];
}
}
proposal[thread_idx] = v_max; // Propose.
if (!has_unmatched_neighbor)
out[thread_idx] = thread_idx;
}
}
void propose(torch::Tensor out, torch::Tensor proposal, torch::Tensor rowptr,
torch::Tensor col,
torch::optional<torch::Tensor> optional_weight) {
auto stream = at::cuda::getCurrentCUDAStream();
if (!optional_weight.has_value()) {
propose_kernel<<<BLOCKS(out.numel()), THREADS, 0, stream>>>(
out.data_ptr<int64_t>(), proposal.data_ptr<int64_t>(),
rowptr.data_ptr<int64_t>(), col.data_ptr<int64_t>(), out.numel());
} else {
auto weight = optional_weight.value();
auto scalar_type = weight.scalar_type();
AT_DISPATCH_ALL_TYPES_AND(at::ScalarType::Half, scalar_type, "_", [&] {
weighted_propose_kernel<scalar_t>
<<<BLOCKS(out.numel()), THREADS, 0, stream>>>(
out.data_ptr<int64_t>(), proposal.data_ptr<int64_t>(),
rowptr.data_ptr<int64_t>(), col.data_ptr<int64_t>(),
weight.data_ptr<scalar_t>(), out.numel());
});
}
}
__global__ void respond_kernel(int64_t *out, const int64_t *proposal,
const int64_t *rowptr, const int64_t *col,
int64_t numel) {
const int64_t thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
if (thread_idx < numel) {
if (out[thread_idx] != -2)
return; // Only vist red nodes.
bool has_unmatched_neighbor = false;
for (int64_t i = rowptr[thread_idx]; i < rowptr[thread_idx + 1]; i++) {
auto v = col[i];
if (out[v] < 0)
has_unmatched_neighbor = true; // Unmatched neighbor found.
if (out[v] == -1 && proposal[v] == thread_idx) {
// Match first blue neighbhor v which proposed to u.
out[thread_idx] = min(thread_idx, v);
out[v] = min(thread_idx, v);
break;
}
}
if (!has_unmatched_neighbor)
out[thread_idx] = thread_idx;
}
}
template <typename scalar_t>
__global__ void weighted_respond_kernel(int64_t *out, const int64_t *proposal,
const int64_t *rowptr,
const int64_t *col,
const scalar_t *weight, int64_t numel) {
const int64_t thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
if (thread_idx < numel) {
if (out[thread_idx] != -2)
return; // Only vist red nodes.
bool has_unmatched_neighbor = false;
int64_t v_max = -1;
scalar_t w_max = 0;
for (int64_t i = rowptr[thread_idx]; i < rowptr[thread_idx + 1]; i++) {
auto v = col[i];
if (out[v] < 0)
has_unmatched_neighbor = true; // Unmatched neighbor found.
if (out[v] == -1 && proposal[v] == thread_idx && weight[i] >= w_max) {
// Find maximum weighted blue neighbhor v which proposed to u.
v_max = v;
w_max = weight[i];
}
}
if (v_max >= 0) {
out[thread_idx] = min(thread_idx, v_max); // Match neighbors.
out[v_max] = min(thread_idx, v_max);
}
if (!has_unmatched_neighbor)
out[thread_idx] = thread_idx;
}
}
void respond(torch::Tensor out, torch::Tensor proposal, torch::Tensor rowptr,
torch::Tensor col,
torch::optional<torch::Tensor> optional_weight) {
auto stream = at::cuda::getCurrentCUDAStream();
if (!optional_weight.has_value()) {
respond_kernel<<<BLOCKS(out.numel()), THREADS, 0, stream>>>(
out.data_ptr<int64_t>(), proposal.data_ptr<int64_t>(),
rowptr.data_ptr<int64_t>(), col.data_ptr<int64_t>(), out.numel());
} else {
auto weight = optional_weight.value();
auto scalar_type = weight.scalar_type();
AT_DISPATCH_ALL_TYPES_AND(at::ScalarType::Half, scalar_type, "_", [&] {
weighted_respond_kernel<scalar_t>
<<<BLOCKS(out.numel()), THREADS, 0, stream>>>(
out.data_ptr<int64_t>(), proposal.data_ptr<int64_t>(),
rowptr.data_ptr<int64_t>(), col.data_ptr<int64_t>(),
weight.data_ptr<scalar_t>(), out.numel());
});
}
}
torch::Tensor graclus_cuda(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_weight) {
CHECK_CUDA(rowptr);
CHECK_CUDA(col);
CHECK_INPUT(rowptr.dim() == 1 && col.dim() == 1);
if (optional_weight.has_value()) {
CHECK_CUDA(optional_weight.value());
CHECK_INPUT(optional_weight.value().dim() == 1);
CHECK_INPUT(optional_weight.value().numel() == col.numel());
}
cudaSetDevice(rowptr.get_device());
int64_t num_nodes = rowptr.numel() - 1;
auto out = torch::full(num_nodes, -1, rowptr.options());
auto proposal = torch::full(num_nodes, -1, rowptr.options());
while (!colorize(out)) {
propose(out, proposal, rowptr, col, optional_weight);
respond(out, proposal, rowptr, col, optional_weight);
}
return out;
}
|
