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
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
|
/*************************************************************************
* Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#include "TestBedChild.hpp"
#include <thread>
#include <execinfo.h>
#define CHILD_NCCL_CALL(cmd, msg) \
{ \
if (this->verbose) printf("[ NCCL CALL] " #cmd "\n"); \
ncclResult_t status = cmd; \
if (status != ncclSuccess) \
{ \
ERROR("Child process %d fails NCCL call %s with code %d\n", this->childId, msg, status); \
return TEST_FAIL; \
} \
}
#define PIPE_READ(val) \
if (read(childReadFd, &val, sizeof(val)) != sizeof(val)) return TEST_FAIL;
namespace RcclUnitTesting
{
TestBedChild::TestBedChild(int const childId, bool const verbose, int const printValues)
{
this->childId = childId;
this->verbose = verbose;
this->printValues = printValues;
}
int TestBedChild::InitPipes()
{
// Prepare parent->child pipe
int pipefd[2];
if (pipe(pipefd) == -1)
{
ERROR("Unable to create parent->child pipe for child %d\n", this->childId);
return TEST_FAIL;
}
this->childReadFd = pipefd[0];
this->parentWriteFd = pipefd[1];
// Prepare child->parent pipe
this->parentReadFd = -1;
if (pipe(pipefd) == -1)
{
ERROR("Unable to create parent->child pipe for child %d\n", this->childId);
return TEST_FAIL;
}
this->parentReadFd = pipefd[0];
this->childWriteFd = pipefd[1];
return TEST_SUCCESS;
}
void TestBedChild::StartExecutionLoop()
{
// Close unused ends of pipes
close(this->parentWriteFd);
close(this->parentReadFd);
// Wait for commands from parent process
if (verbose) INFO("Child %d enters execution loop\n", this->childId);
int command;
while (read(childReadFd, &command, sizeof(command)) > 0)
{
if (verbose) INFO("Child %d received command [%s]:\n", this->childId, ChildCommandNames[command]);;
ErrCode status = TEST_SUCCESS;
switch(command)
{
case CHILD_GET_UNIQUE_ID : status = GetUniqueId(); break;
case CHILD_INIT_COMMS : status = InitComms(); break;
case CHILD_SET_COLL_ARGS : status = SetCollectiveArgs(); break;
case CHILD_ALLOCATE_MEM : status = AllocateMem(); break;
case CHILD_PREPARE_DATA : status = PrepareData(); break;
case CHILD_EXECUTE_COLL : status = ExecuteCollectives(); break;
case CHILD_VALIDATE_RESULTS: status = ValidateResults(); break;
case CHILD_DEALLOCATE_MEM : status = DeallocateMem(); break;
case CHILD_DESTROY_COMMS : status = DestroyComms(); break;
case CHILD_STOP : status = Stop(); break;
default: exit(0);
}
// Send back acknowledgement to parent
if (status == TEST_FAIL)
ERROR("Child %d failed on command [%s]:\n", this->childId, ChildCommandNames[command]);
if (write(childWriteFd, &status, sizeof(status)) < 0)
{
ERROR("Child %d write to parent failed: %s\n", this->childId, strerror(errno));
break;
}
}
if (verbose) INFO("Child %d exiting execution loop\n", this->childId);
// Close child ends of pipe
close(this->childReadFd);
close(this->childWriteFd);
exit(0);
}
ErrCode TestBedChild::GetUniqueId()
{
if (this->verbose) INFO("Child %d begins GetUniqueId()\n", this->childId);
// Get a unique ID and pass it back to parent process
ncclUniqueId id;
CHILD_NCCL_CALL(ncclGetUniqueId(&id), "ncclGetUniqueId");
write(childWriteFd, &id, sizeof(id));
if (this->verbose) INFO("Child %d finishes GetUniqueId()\n", this->childId);
return TEST_SUCCESS;
}
ErrCode TestBedChild::InitComms()
{
if (this->verbose) INFO("Child %d begins InitComms()\n", this->childId);
// Read values sent by parent [see TestBed::InitComms()]
ncclUniqueId id;
PIPE_READ(id);
PIPE_READ(this->totalRanks);
PIPE_READ(this->rankOffset);
PIPE_READ(this->numCollectivesInGroup);
bool useMultiRankPerGpu;
PIPE_READ(useMultiRankPerGpu);
// Read the GPUs this child uses and prepare storage for collective args / datasets
int numGpus;
PIPE_READ(numGpus);
this->deviceIds.resize(numGpus);
this->streams.resize(numGpus);
this->collArgs.resize(numGpus);
for (int i = 0; i < numGpus; i++)
{
PIPE_READ(this->deviceIds[i]);
this->collArgs[i].clear();
this->collArgs[i].resize(numCollectivesInGroup);
}
// Initialize communicators
comms.clear();
comms.resize(numGpus);
// Initialize within a group call to avoid deadlock when using multiple ranks per child
ErrCode status = TEST_SUCCESS;
CHILD_NCCL_CALL(ncclGroupStart(), "ncclGroupStart");
for (int localRank = 0; localRank < numGpus; ++localRank)
{
int const globalRank = this->rankOffset + localRank;
int const currGpu = this->deviceIds[localRank];
if (hipSetDevice(currGpu) != hipSuccess)
{
ERROR("Rank %d on child %d unable to switch to GPU %d\n", globalRank, this->childId, currGpu);
status = TEST_FAIL;
break;
}
if (hipStreamCreate(&this->streams[localRank]) != hipSuccess)
{
ERROR("Rank %d on child %d unable to create stream for GPU %d\n", globalRank, this->childId, currGpu);
status = TEST_FAIL;
break;
}
if (useMultiRankPerGpu)
{
if (ncclCommInitRankMulti(&this->comms[localRank], this->totalRanks, id, globalRank, globalRank) != ncclSuccess)
{
ERROR("Rank %d on child %d unable to call ncclCommInitRankMulti\n", globalRank, this->childId);
status = TEST_FAIL;
break;
}
}
else
{
if (ncclCommInitRank(&this->comms[localRank], this->totalRanks, id, globalRank) != ncclSuccess)
{
ERROR("Rank %d on child %d unable to call ncclCommInitRank\n", globalRank, this->childId);
status = TEST_FAIL;
break;
}
}
}
if (status == TEST_SUCCESS)
{
CHILD_NCCL_CALL(ncclGroupEnd(), "ncclGroupStart");
}
if (this->verbose) INFO("Child %d finishes InitComms() [%s]\n",
this->childId, status == TEST_SUCCESS ? "SUCCESS" : "FAIL");
return status;
}
ErrCode TestBedChild::SetCollectiveArgs()
{
if (this->verbose) INFO("Child %d begins SetCollectiveArgs()\n", this->childId);
// Read values sent by parent [see TestBed::SetCollectiveArgs()]
int globalRank;
int collId;
ncclFunc_t funcType;
ncclDataType_t dataType;
size_t numInputElements;
size_t numOutputElements;
OptionalColArgs options;
PIPE_READ(globalRank);
PIPE_READ(collId);
PIPE_READ(funcType);
PIPE_READ(dataType);
PIPE_READ(numInputElements);
PIPE_READ(numOutputElements);
PIPE_READ(options);
if (globalRank < this->rankOffset || (this->rankOffset + comms.size() <= globalRank))
{
ERROR("Child %d does not contain rank %d\n", this->childId, globalRank);
return TEST_FAIL;
}
int const localRank = globalRank - rankOffset;
CHECK_HIP(hipSetDevice(this->deviceIds[localRank]));
for (int collIdx = 0; collIdx < collArgs[localRank].size(); ++collIdx)
{
if (collId == -1 || collId == collIdx)
{
CollectiveArgs& collArg = this->collArgs[localRank][collIdx];
CHECK_CALL(collArg.SetArgs(globalRank, this->totalRanks,
this->deviceIds[localRank],
funcType, dataType,
numInputElements, numOutputElements,
options));
if (this->verbose) INFO("Rank %d on child %d sets collective %d [%s]\n",
globalRank, this->childId, collIdx,
collArg.GetDescription().c_str());
// If pre-mult scalars are provided, then create a custom reduction operator
if (options.scalarMode >= 0)
{
CHILD_NCCL_CALL(ncclRedOpCreatePreMulSum(&collArg.options.redOp,
collArg.localScalar.ptr,
dataType,
(ncclScalarResidence_t)options.scalarMode,
this->comms[localRank]),
"ncclRedOpCreatePreMulSum");
if (verbose) INFO("Child %d created custom redop %d for collective %d\n",
this->childId, collArg.options.redOp, collIdx);
}
}
}
if (this->verbose) INFO("Child %d finishes SetCollectiveArgs()\n", this->childId);
return TEST_SUCCESS;
}
ErrCode TestBedChild::AllocateMem()
{
if (this->verbose) INFO("Child %d begins AllocateMem()\n", this->childId);
// Read values sent by parent [see TestBed::AllocateMem()]
int globalRank;
int collId;
bool inPlace;
bool useManagedMem;
PIPE_READ(globalRank);
PIPE_READ(collId);
PIPE_READ(inPlace);
PIPE_READ(useManagedMem);
if (globalRank < this->rankOffset || (this->rankOffset + comms.size() <= globalRank))
{
ERROR("Child %d does not contain rank %d\n", this->childId, globalRank);
return TEST_FAIL;
}
int const localRank = globalRank - rankOffset;
CHECK_HIP(hipSetDevice(this->deviceIds[localRank]));
for (int collIdx = 0; collIdx < collArgs[localRank].size(); ++collIdx)
{
if (collId == -1 || collId == collIdx)
{
CollectiveArgs& collArg = this->collArgs[localRank][collIdx];
CHECK_CALL(collArg.AllocateMem(inPlace, useManagedMem));
if (this->verbose) INFO("Rank %d on child %d allocates memory for collective %d on device %d (%s,%s) Input: %p Output %p\n",
globalRank, this->childId, collIdx, this->deviceIds[localRank],
inPlace ? "in-place" : "out-of-place",
useManagedMem ? "managed" : "unmanaged",
collArg.inputGpu.ptr,
collArg.outputGpu.ptr);
}
}
if (this->verbose) INFO("Child %d finishes AllocateMem()\n", this->childId);
return TEST_SUCCESS;
}
// Fill input memory with pre-known patterned based on rank
ErrCode TestBedChild::PrepareData()
{
if (this->verbose) INFO("Child %d begins PrepareData()\n", this->childId);
// Read values sent by parent [see TestBed::PrepareData()]
int globalRank;
int collId;
CollFuncPtr prepDataFunc;
PIPE_READ(globalRank);
PIPE_READ(collId);
PIPE_READ(prepDataFunc);
if (globalRank < this->rankOffset || (this->rankOffset + comms.size() <= globalRank))
{
ERROR("Child %d does not contain rank %d\n", this->childId, globalRank);
return TEST_FAIL;
}
int const localRank = globalRank - rankOffset;
CHECK_HIP(hipSetDevice(this->deviceIds[localRank]));
for (int collIdx = 0; collIdx < collArgs[localRank].size(); ++collIdx)
{
if (collId == -1 || collId == collIdx)
{
if (this->verbose) INFO("Rank %d on child %d prepares data for collective %d\n",
globalRank, this->childId, collIdx);
CHECK_CALL(this->collArgs[localRank][collIdx].PrepareData(prepDataFunc));
}
}
if (this->verbose) INFO("Child %d finishes PrepareData()\n", this->childId);
return TEST_SUCCESS;
}
ErrCode TestBedChild::ExecuteCollectives()
{
int numRanksToExecute, tempRank;
std::vector<int> ranksToExecute = {};
PIPE_READ(numRanksToExecute);
for (int rank = 0; rank < numRanksToExecute; ++rank){
PIPE_READ(tempRank);
ranksToExecute.push_back(tempRank - this->rankOffset);
}
if (this->verbose) INFO("Child %d begins ExecuteCollectives()\n", this->childId);
// Start group call
CHILD_NCCL_CALL(ncclGroupStart(), "ncclGroupStart");
// Loop over all collectives to be executed in group call
for (int collId = 0; collId < this->numCollectivesInGroup; ++collId)
{
// Loop over all local ranks
for (int localRank = 0; localRank < this->deviceIds.size(); ++localRank)
{
// If ranks to execute is empty, execute all ranks belonging to child
if (!ranksToExecute.empty() && (std::count(ranksToExecute.begin(), ranksToExecute.end(), localRank) == 0)) continue;
CHECK_HIP(hipSetDevice(this->deviceIds[localRank]));
CollectiveArgs const& collArg = this->collArgs[localRank][collId];
if (this->printValues)
{
int const numInputElementsToPrint = (this->printValues < 0 ? collArg.numInputElements : this->printValues);
PtrUnion inputCpu;
size_t const numInputBytes = numInputElementsToPrint * DataTypeToBytes(collArg.dataType);
inputCpu.AllocateCpuMem(numInputBytes);
CHECK_HIP(hipMemcpy(inputCpu.ptr, collArg.inputGpu.ptr, numInputBytes, hipMemcpyDeviceToHost));
printf("[ DEBUG ] Rank %02d Coll %d %-10s: %s\n", collArg.globalRank, collId, "Input",
inputCpu.ToString(collArg.dataType, numInputElementsToPrint).c_str());
inputCpu.FreeCpuMem();
int const numOutputElementsToPrint = (this->printValues < 0 ? collArg.numOutputElements : this->printValues);
size_t const numOutputBytes = numOutputElementsToPrint * DataTypeToBytes(collArg.dataType);
CHECK_HIP(hipMemcpy(collArg.outputCpu.ptr, collArg.outputGpu.ptr, numOutputBytes, hipMemcpyDeviceToHost));
printf("[ DEBUG ] Rank %02d Coll %d %-10s: %s\n", collArg.globalRank, collId, "Pre-Output",
collArg.outputCpu.ToString(collArg.dataType, numOutputElementsToPrint).c_str());
}
switch (collArg.funcType)
{
case ncclCollBroadcast:
CHILD_NCCL_CALL(ncclBroadcast(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numInputElements,
collArg.dataType,
collArg.options.root,
this->comms[localRank],
this->streams[localRank]),
"ncclBroadcast");
break;
case ncclCollReduce:
CHILD_NCCL_CALL(ncclReduce(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numInputElements,
collArg.dataType,
collArg.options.redOp,
collArg.options.root,
this->comms[localRank],
this->streams[localRank]),
"ncclReduce");
break;
case ncclCollAllGather:
CHILD_NCCL_CALL(ncclAllGather(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numInputElements,
collArg.dataType,
this->comms[localRank],
this->streams[localRank]),
"ncclAllGather");
break;
case ncclCollReduceScatter:
CHILD_NCCL_CALL(ncclReduceScatter(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numOutputElements,
collArg.dataType,
collArg.options.redOp,
this->comms[localRank],
this->streams[localRank]),
"ncclReduceScatter");
break;
case ncclCollAllReduce:
CHILD_NCCL_CALL(ncclAllReduce(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numInputElements,
collArg.dataType,
collArg.options.redOp,
this->comms[localRank],
this->streams[localRank]),
"ncclAllReduce");
break;
case ncclCollGather:
CHILD_NCCL_CALL(ncclGather(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numInputElements,
collArg.dataType,
collArg.options.root,
this->comms[localRank],
this->streams[localRank]),
"ncclGather");
break;
case ncclCollScatter:
CHILD_NCCL_CALL(ncclScatter(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numOutputElements,
collArg.dataType,
collArg.options.root,
this->comms[localRank],
this->streams[localRank]),
"ncclScatter");
break;
case ncclCollAllToAll:
CHILD_NCCL_CALL(ncclAllToAll(collArg.inputGpu.ptr,
collArg.outputGpu.ptr,
collArg.numInputElements / collArg.totalRanks,
collArg.dataType,
this->comms[localRank],
this->streams[localRank]),
"ncclAllToAll");
break;
case ncclCollAllToAllv:
CHILD_NCCL_CALL(ncclAllToAllv(collArg.inputGpu.ptr,
collArg.options.sendcounts + (this->rankOffset + localRank)*this->totalRanks,
collArg.options.sdispls + (this->rankOffset + localRank)*this->totalRanks,
collArg.outputGpu.ptr,
collArg.options.recvcounts + (this->rankOffset + localRank)*this->totalRanks,
collArg.options.rdispls + (this->rankOffset + localRank)*this->totalRanks,
collArg.dataType,
this->comms[localRank],
this->streams[localRank]),
"ncclAllToAllv");
break;
case ncclCollSend:
CHILD_NCCL_CALL(ncclSend(collArg.inputGpu.ptr,
collArg.numInputElements,
collArg.dataType,
collArg.options.root,
this->comms[localRank],
this->streams[localRank]),
"ncclSend");
break;
case ncclCollRecv:
CHILD_NCCL_CALL(ncclRecv(collArg.outputGpu.ptr,
collArg.numOutputElements,
collArg.dataType,
collArg.options.root,
this->comms[localRank],
this->streams[localRank]),
"ncclRecv");
break;
default:
ERROR("Unknown func type %d\n", collArg.funcType);
return TEST_FAIL;
}
}
}
// End group call
CHILD_NCCL_CALL(ncclGroupEnd(), "ncclGroupEnd");
// Synchronize
if (this->verbose) INFO("Child %d submits group call. Waiting for completion\n", this->childId);
for (int localRank = 0; localRank < this->streams.size(); ++localRank)
{
CHECK_HIP(hipStreamSynchronize(this->streams[localRank]));
}
if (this->printValues)
{
for (int collId = 0; collId < this->numCollectivesInGroup; ++collId)
for (int localRank = 0; localRank < this->deviceIds.size(); ++localRank)
{
CollectiveArgs const& collArg = this->collArgs[localRank][collId];
int numOutputElementsToPrint = (this->printValues < 0 ? collArg.numOutputElements : this->printValues);
size_t const numOutputBytes = numOutputElementsToPrint * DataTypeToBytes(collArg.dataType);
CHECK_HIP(hipMemcpy(collArg.outputCpu.ptr, collArg.outputGpu.ptr, numOutputBytes, hipMemcpyDeviceToHost));
printf("[ DEBUG ] Rank %02d Coll %d %-10s: %s\n", collArg.globalRank, collId, "Output",
collArg.outputCpu.ToString(collArg.dataType, numOutputElementsToPrint).c_str());
printf("[ DEBUG ] Rank %02d Coll %d %-10s: %s\n", collArg.globalRank, collId, "Expected",
collArg.expected.ToString(collArg.dataType, numOutputElementsToPrint).c_str());
}
}
if (this->verbose) INFO("Child %d finishes ExecuteCollectives()\n", this->childId);
return TEST_SUCCESS;
}
ErrCode TestBedChild::ValidateResults()
{
// Read values sent by parent [see TestBed::ValidateResults()]
int globalRank, collId;
PIPE_READ(globalRank);
PIPE_READ(collId);
if (this->verbose) INFO("Child %d begins ValidateResults()\n", this->childId);
if (globalRank < this->rankOffset || (this->rankOffset + comms.size() <= globalRank))
{
ERROR("Child %d does not contain rank %d\n", this->childId, globalRank);
return TEST_FAIL;
}
int const localRank = globalRank - rankOffset;
CHECK_HIP(hipSetDevice(this->deviceIds[localRank]));
ErrCode status = TEST_SUCCESS;
for (int collIdx = 0; collIdx < collArgs[localRank].size(); ++collIdx)
{
if (collId == -1 || collId == collIdx)
{
if (this->verbose) INFO("Rank %d on child %d validating collective %d results\n",
globalRank, this->childId, collIdx);
if (this->collArgs[localRank][collIdx].ValidateResults() != TEST_SUCCESS)
{
ERROR("Rank %d Collective %d output does not match expected\n", globalRank, collIdx);
status = TEST_FAIL;
}
}
}
if (this->verbose) INFO("Child %d finishes ValidateResults() with status %s\n", this->childId,
status == TEST_SUCCESS ? "SUCCESS" : "FAIL");
return status;
}
ErrCode TestBedChild::DeallocateMem()
{
if (this->verbose) INFO("Child %d begins DeallocateMem\n", this->childId);
// Read values sent by parent [see TestBed::DeallocateMem()]
int globalRank, collId;
PIPE_READ(globalRank);
PIPE_READ(collId);
if (globalRank < this->rankOffset || (this->rankOffset + comms.size() <= globalRank))
{
ERROR("Child %d does not contain rank %d\n", this->childId, globalRank);
return TEST_FAIL;
}
int const localRank = globalRank - rankOffset;
CHECK_HIP(hipSetDevice(this->deviceIds[localRank]));
for (int collIdx = 0; collIdx < collArgs[localRank].size(); ++collIdx)
{
CollectiveArgs& collArg = this->collArgs[localRank][collIdx];
if (collId == -1 || collId == collIdx)
{
if (this->verbose)
{
INFO("Child %d release memory for collective %d (Input: %p Output %p\n",
this->childId, collIdx, collArg.inputGpu.ptr, collArg.outputGpu.ptr);
}
CHECK_CALL(collArg.DeallocateMem());
}
if (collArg.options.scalarMode != -1)
{
CHILD_NCCL_CALL(ncclRedOpDestroy(collArg.options.redOp, this->comms[localRank]),
"ncclRedOpDestroy");
if (verbose) INFO("Child %d destroys custom redop %d for collective %d\n",
this->childId, collArg.options.redOp, collIdx);
}
}
if (this->verbose) INFO("Child %d finishes DeallocateMem\n", this->childId);
return TEST_SUCCESS;
}
ErrCode TestBedChild::DestroyComms()
{
if (this->verbose) INFO("Child %d begins DestroyComms\n", this->childId);
// Release comms
for (int i = 0; i < this->comms.size(); ++i)
{
CHILD_NCCL_CALL(ncclCommDestroy(this->comms[i]), "ncclCommDestroy");
}
for (int i = 0; i < this->streams.size(); ++i)
{
CHECK_HIP(hipStreamDestroy(this->streams[i]));
}
this->comms.clear();
this->streams.clear();
if (this->verbose) INFO("Child %d finishes DestroyComms\n", this->childId);
return TEST_SUCCESS;
}
ErrCode TestBedChild::Stop()
{
return TEST_SUCCESS;
}
}
|
