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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
|
#include <unistd.h>
#include <c10/util/irange.h>
#include <torch/csrc/distributed/c10d/ProcessGroupMPI.hpp>
#include <cstdlib>
#include <iostream>
#include <sstream>
#include <string>
#include <thread>
#define STR_HELPER(x) #x
#define STR(x) STR_HELPER(x)
// Wait for work to complete
std::vector<std::vector<at::Tensor>> waitWork(
c10::intrusive_ptr<::c10d::ProcessGroupMPI> pg,
std::vector<c10::intrusive_ptr<c10d::Work>> works) {
std::vector<std::vector<at::Tensor>> outputTensors;
for (auto& work : works) {
try {
work->wait();
} catch (const std::exception& ex) {
std::cerr << "Exception received: " << ex.what() << std::endl;
pg->abort();
}
outputTensors.emplace_back(work->result());
}
return outputTensors;
}
// Wait using Futures
std::vector<std::vector<at::Tensor>> waitFuture(
c10::intrusive_ptr<::c10d::ProcessGroupMPI> pg,
std::vector<c10::intrusive_ptr<c10d::Work>> works) {
std::vector<std::vector<at::Tensor>> outputTensors;
for (auto& work : works) {
auto fut = work->getFuture();
try {
fut->wait();
} catch (const std::exception& ex) {
std::cerr << "Exception received: " << ex.what() << std::endl;
pg->abort();
}
auto result = fut->value();
if (result.isNone()) {
outputTensors.emplace_back();
} else if (result.isTensorList()) {
outputTensors.emplace_back(result.toTensorVector());
} else {
TORCH_CHECK(false, "future result should be tensor list or none");
}
}
return outputTensors;
}
void testAllreduce(int iter = 1000) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
// Generate inputs
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
for (const auto i : c10::irange(iter)) {
auto tensor = at::ones({16, 16}) * i;
std::vector<at::Tensor> tensors = {tensor};
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->allreduce(tensors);
works.push_back(std::move(work));
}
auto outputTensors = waitFuture(pg, works);
// Get the world size
auto worldSize = pg->getSize();
// Verify outputs
for (const auto i : c10::irange(iter)) {
const auto expected = worldSize * i;
auto data = outputTensors[i][0].data_ptr<float>();
for (auto j = 0; j < outputTensors[i][0].numel(); ++j) {
if (data[j] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
void testBroadcast(int iter = 10000) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
for (const auto i : c10::irange(iter)) {
auto tensors = std::vector<at::Tensor>();
if (pg->getRank() == 0) {
auto tensor = at::ones({16, 16}) * i;
tensors = std::vector<at::Tensor>({tensor});
} else {
auto tensor = at::zeros({16, 16});
tensors = std::vector<at::Tensor>({tensor});
}
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->broadcast(tensors);
works.push_back(std::move(work));
}
auto outputTensors = waitFuture(pg, works);
// Verify outputs
for (const auto i : c10::irange(iter)) {
const auto expected = i;
auto data = outputTensors[i][0].data_ptr<float>();
for (auto j = 0; j < outputTensors[i][0].numel(); ++j) {
if (data[j] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
void testReduce(int iter = 10000) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
for (const auto i : c10::irange(iter)) {
auto tensor = at::ones({16, 16}) * i;
auto tensors = std::vector<at::Tensor>({tensor});
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->reduce(tensors);
works.push_back(std::move(work));
}
auto outputTensors = waitFuture(pg, works);
// Get the world size
auto worldSize = pg->getSize();
if (pg->getRank() == 0) {
// Verify outputs
for (const auto i : c10::irange(iter)) {
const auto expected = worldSize * i;
auto data = outputTensors[i][0].data_ptr<float>();
for (auto j = 0; j < outputTensors[i][0].numel(); ++j) {
if (data[j] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
}
void testAllgather(int iter = 10000) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
// Get the world size
auto worldSize = pg->getSize();
auto rank = pg->getRank();
// Generate inputs
for (const auto i : c10::irange(iter)) {
auto tensor = at::ones({16, 16}) * i * rank;
auto tensors = std::vector<at::Tensor>({tensor});
auto outputs = std::vector<std::vector<at::Tensor>>(1);
outputs[0].resize(worldSize);
for (const auto j : c10::irange(worldSize)) {
outputs[0][j] = at::zeros({16, 16});
}
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->allgather(outputs, tensors);
works.push_back(std::move(work));
}
auto outputTensors = waitFuture(pg, works);
// Verify outputs
for (const auto i : c10::irange(iter)) {
for (const auto j : c10::irange(worldSize)) {
const auto expected = i * j;
auto data = outputTensors[i][j].data_ptr<float>();
for (auto k = 0; k < outputTensors[i][j].numel(); ++k) {
if (data[k] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
}
void testGather(int iter = 10000) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
// Get the world size
auto worldSize = pg->getSize();
auto rank = pg->getRank();
// Generate inputs
for (const auto i : c10::irange(iter)) {
auto tensor = at::ones({16, 16}) * i * rank;
auto tensors = std::vector<at::Tensor>({tensor});
auto outputs = std::vector<std::vector<at::Tensor>>(0);
if (rank == 0) {
outputs = std::vector<std::vector<at::Tensor>>(1);
outputs[0].resize(worldSize);
for (const auto j : c10::irange(worldSize)) {
outputs[0][j] = at::zeros({16, 16});
}
}
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->gather(outputs, tensors);
works.push_back(std::move(work));
}
auto outputTensors = waitFuture(pg, works);
// Verify outputs
if (rank == 0) {
for (const auto i : c10::irange(iter)) {
for (const auto j : c10::irange(worldSize)) {
const auto expected = i * j;
auto data = outputTensors[i][j].data_ptr<float>();
for (auto k = 0; k < outputTensors[i][j].numel(); ++k) {
if (data[k] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
} else {
for (const auto i : c10::irange(iter)) {
if (outputTensors[i].size() != 0) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
void testScatter(int iter = 1) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
// Get the world size
auto worldSize = pg->getSize();
auto rank = pg->getRank();
// Generate inputs
for (const auto i : c10::irange(iter)) {
auto tensor = at::zeros({16, 16});
auto tensors = std::vector<at::Tensor>({tensor});
auto inputs = std::vector<std::vector<at::Tensor>>(0);
if (rank == 0) {
inputs = std::vector<std::vector<at::Tensor>>(1);
inputs[0].resize(worldSize);
for (const auto j : c10::irange(worldSize)) {
inputs[0][j] = at::ones({16, 16}) * i * j;
}
}
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->scatter(tensors, inputs);
works.push_back(std::move(work));
}
auto outputTensors = waitFuture(pg, works);
// Verify outputs
for (const auto i : c10::irange(iter)) {
for (const auto j : c10::irange(worldSize)) {
const auto expected = i * j;
auto data = outputTensors[i][0].data_ptr<float>();
for (auto k = 0; k < outputTensors[i][0].numel(); ++k) {
if (data[k] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
}
void testSendRecv(bool recvAnysource, int iter = 10000) {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
// Generate inputs
std::vector<c10::intrusive_ptr<::c10d::Work>> works;
// pg->send does not keep sent tensors alive, so we need to.
std::vector<std::vector<at::Tensor>> sendTensors(iter);
auto rank = pg->getRank();
for (const auto i : c10::irange(iter)) {
if (rank == 0) {
auto tensor = at::ones({16, 16}) * i;
sendTensors[i] = std::vector<at::Tensor>({tensor});
// Queue the work.
c10::intrusive_ptr<::c10d::Work> work = pg->send(sendTensors[i], 1, 0);
works.push_back(std::move(work));
} else {
auto tensor = at::zeros({16, 16});
auto recvTensors = std::vector<at::Tensor>({tensor});
// Queue the work.
if (!recvAnysource) {
c10::intrusive_ptr<::c10d::Work> work = pg->recv(recvTensors, 0, 0);
works.push_back(std::move(work));
} else {
c10::intrusive_ptr<::c10d::Work> work =
pg->recvAnysource(recvTensors, 0);
works.push_back(std::move(work));
}
}
}
auto outputTensors = waitWork(pg, works);
if (rank == 0) {
return;
}
std::vector<int> srcRanks;
if (recvAnysource) {
for (const auto& work : works) {
srcRanks.push_back(work->sourceRank());
}
}
// Verify outputs
for (const auto i : c10::irange(iter)) {
if (recvAnysource && srcRanks[i] != 0) {
TORCH_CHECK(false, "src rank is wrong for recvAnysource");
}
const auto expected = i;
auto data = outputTensors[i][0].data_ptr<float>();
for (auto j = 0; j < outputTensors[i][0].numel(); ++j) {
if (data[j] != expected) {
TORCH_CHECK(false, "BOOM!");
}
}
}
}
void testBackendName() {
auto pg = c10d::ProcessGroupMPI::createProcessGroupMPI();
if (pg->getBackendName() != std::string(c10d::MPI_BACKEND_NAME)) {
TORCH_CHECK(false, "BOOM!");
}
}
int main(int argc, char** argv) {
#ifdef MPIEXEC
// If we are within an openmpi mpirun, then skip the exec
if (!std::getenv("OMPI_COMM_WORLD_SIZE")) {
std::cout << "Execute mpiexec from: " << STR(MPIEXEC) << std::endl;
execl(STR(MPIEXEC), "-np 2", argv[0], (char*)nullptr);
}
testAllreduce();
testBroadcast();
testReduce();
testAllgather();
testGather();
testScatter();
testSendRecv(false);
testSendRecv(true);
testBackendName();
std::cout << "Test successful" << std::endl;
#else
std::cout << "MPI executable not found, skipping test" << std::endl;
#endif
return EXIT_SUCCESS;
}
|
