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
|
#include "coll_custom_huge_all_reduce_mesh_executor.h"
namespace hccl
{
CollCustomHugeAllReduceMeshExecutor::CollCustomHugeAllReduceMeshExecutor(const HcclDispatcher dispatcher,
std::unique_ptr<TopoMatcher> &topoMatcher)
: CollCommExecutor(dispatcher, topoMatcher)
{
}
// Calculate the amount of scratch memory to request
HcclResult CollCustomHugeAllReduceMeshExecutor::CalcScratchMemSize(u64 &scratchMemSize)
{
// We don't need to use Scratch memory
scratchMemSize = 0U;
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][CalcScratchMemSize] scratchMemSize: %u", scratchMemSize);
return HCCL_SUCCESS;
}
// Calculate the number of streams to be requested
HcclResult CollCustomHugeAllReduceMeshExecutor::CalcStreamNum(u32 &streamNum)
{
u32 totalStreamNum = topoAttr_.deviceNumPerAggregation;
streamNum = totalStreamNum - 1U;
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][CalcStreamNum] streamNum: %u", streamNum);
return HCCL_SUCCESS;
}
// Calculate the number of Notify to be requested
HcclResult CollCustomHugeAllReduceMeshExecutor::CalcNotifyNum(u32 streamNum, u32 ¬ifyNum)
{
// ReducesScatter needs 2 Notify, AllGather needs 2 Notify
notifyNum = 4U * streamNum;
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][CalcNotifyNum] notifyNum: %u", notifyNum);
return HCCL_SUCCESS;
}
// Set up the level-0 mesh topology required for the AllReduce operation
HcclResult CollCustomHugeAllReduceMeshExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport> &opTransport)
{
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][CalcCommInfo]");
// Define the source and destination memory types for communication
TransportMemType inputType = TransportMemType::CCL_INPUT;
TransportMemType outputType = TransportMemType::CCL_OUTPUT;
// Construct a mesh topology for level 0
CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
// Compute and populate the transport plan for level-0 communication domain
CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
return HCCL_SUCCESS;
}
// Calculate the number of iterations for loop processing
u64 CollCustomHugeAllReduceMeshExecutor::CalcLoopMaxCount(const u64 cclBuffSize, const u32 unitSize)
{
u64 maxCountPerLoop = cclBuffSize / unitSize;
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][CalcLoopMaxCount] maxCountPerLoop: %u", maxCountPerLoop);
return maxCountPerLoop;
}
// Entry point for orchestrating the AllReduce algorithm execution
HcclResult CollCustomHugeAllReduceMeshExecutor::Orchestrate(OpParam ¶m, AlgResourceResponse &algRes)
{
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][Orchestrate] count: %u", param.DataDes.count);
tag_ = param.tag;
algResResp_ = &algRes;
// Cast and validate pointers to the user-provided input and output memory buffers
u8 *userInputPtr = static_cast<u8 *>(param.inputPtr);
u8 *userOutputPtr = static_cast<u8 *>(param.outputPtr);
CHK_PTR_NULL(userInputPtr);
CHK_PTR_NULL(userOutputPtr);
// Determine the number of ranks in the communication domain
CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
u32 rankSize = level0CommInfo.localRankSize;
// Determine the size of a single data unit (e.g., 4 bytes for fp32)
u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
// Calculate max number of elements per iteration based on available CCL_Out memory
u64 maxCountPerLoop = CalcLoopMaxCount(algRes.cclOutputMem.size(), unitSize) * (rankSize - BROKEN_RANK_COUNT);
// Loop to process data in chunks if the total data count exceeds maxCountPerLoop
// countLeft: Remaining number of data elements to process
// curCount: Number of data elements processed in the current loop iteration
// inputOffset: Current byte offset in user's input memory buffer
// outputOffset: Current byte offset in user's output memory buffer
for (u64 countLeft = param.DataDes.count, curCount = 0, inputOffset = 0, outputOffset = 0; countLeft > 0;)
{
// Determine the number of data elements for the current iteration
curCount = (countLeft > maxCountPerLoop) ? maxCountPerLoop : countLeft;
// Determine the total data size for the current iteration (unit: bytes)
u64 curSize = curCount * unitSize;
// Construct memory information for the current loop iteration
ExecMem execMem;
execMem.count = curCount; // Number of data elements to process in this loop
execMem.inputPtr = userInputPtr + inputOffset; // Pointer to the current offset of user's input memory buffer
execMem.outputPtr = userOutputPtr + outputOffset; // Pointer to the current offset of user's output memory buffer
execMem.inputMem = algRes.cclInputMem; // Local CCL_Input memory
execMem.outputMem = algRes.cclOutputMem; // Local CCL_Output memory
execMem.scratchMem = algRes.scratchMem; // Local Scratch memory
// Execute the kernel logic for the current data chunk
CHK_RET(KernelRun(param, execMem));
// Update offsets for the next loop iteration
countLeft -= curCount;
inputOffset += curSize;
outputOffset += curSize;
}
return HCCL_SUCCESS;
}
// Process data for a single iteration of the AllReduce algorithm execution
HcclResult CollCustomHugeAllReduceMeshExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
{
// Get sub-communication domain information for level 0
CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
u32 rankSize = level0CommInfo.localRankSize;
u32 rankId = level0CommInfo.localRank;
u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
// Cut the chunk into multiple blocks
CHK_RET(PrepareBlock(execMem.count, unitSize, rankSize - BROKEN_RANK_COUNT));
// Perform ReduceScatter
if (IsHealthyRank(rankId))
{
CHK_RET(HealthyPerformReduceScatter(param, execMem));
}
else
{
CHK_RET(UnhealthyPerformReduceScatter(param, execMem));
}
// Perform AllGather
if (IsHealthyRank(rankId))
{
CHK_RET(HealthyPerformAllGather(param, execMem));
}
else
{
CHK_RET(UnhealthyPerformAllGather(param, execMem));
}
HCCL_WARNING("[HCCLContest][CollCustomHugeAllReduceMeshExecutor][KernelRun] localRank: %u, localRankSize: %u",
level0CommInfo.localRank, level0CommInfo.localRankSize);
return HCCL_SUCCESS;
}
// Perform ReduceScatter
// col0 col1 col2 | col0 col1 col2
// R0: [ 2, 2, 2 ] | R0: [ 8, 2, 2 ]
// R1: [ 2, 2, 2 ] | R1: [ 2, 8, 2 ]
// R2: [ 2, 2, 2 ] | R2: [ 2, 2, 8 ]
// ------------------- --------------------
// R3: [ 2, 2, 2 ] | R3: [ 2, 2, 2 ]
HcclResult CollCustomHugeAllReduceMeshExecutor::HealthyPerformReduceScatter(const OpParam ¶m, ExecMem &execMem)
{
// Retrieve the sub-communication domain information and master stream
CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
// Dertermine the information for the topology and the data
u32 rankSize = level0CommInfo.localRankSize;
u32 rankId = level0CommInfo.localRank;
u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
// For healthy ranks, copy data block from user input buffer to CCL_Out
{
u32 blockIdx = GetResponsibleBlockIdx(rankId);
u64 blockSize = GetBlockSize(blockIdx);
u64 blockOffset = GetBlockOffset(blockIdx);
if (blockSize != 0)
{
DeviceMem src = DeviceMem::create(static_cast<char *>(execMem.inputPtr) + blockOffset, blockSize);
DeviceMem dst = DeviceMem::create(execMem.outputMem.ptr(), blockSize);
CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, masterStream));
}
}
// Post notify signals from master stream to all slave streams
CHK_RET(MainPostToSlaves(param, execMem));
// Slave streams wait for master's notification
CHK_RET(SlavesWaitForMain(param, execMem));
// Case A: The healthy rank notifies the unhealthy rank that it can reduce data
// Case B: Normal communication between healthy ranks
for (u32 round = 1; round < rankSize; round++)
{
// Get the slave stream assigned for communication with dstRank
u32 dstRank = (round + rankId) % rankSize;
Stream &subStream = algResResp_->slaveStreams[round - 1];
// Case A: The healthy rank notifies the unhealthy rank that it can reduce data
if (!IsHealthyRank(dstRank))
{
// Notify remote rank data has been ready
CHK_RET(level0CommInfo.links[dstRank]->TxAck(subStream));
// Wait for the remote rank to notify that data transfer and operations have completed
CHK_RET(level0CommInfo.links[dstRank]->RxDataSignal(subStream));
continue;
}
// Case B: Normal communication between healthy ranks
// Dertermine the information for the data
u32 blockIdx = GetResponsibleBlockIdx(dstRank);
u64 blockSize = GetBlockSize(blockIdx);
u64 blockOffset = GetBlockOffset(blockIdx);
// Notify remote rank data has been ready
CHK_RET(level0CommInfo.links[dstRank]->TxAck(subStream));
// Wait for notification from the remote rank
CHK_RET(level0CommInfo.links[dstRank]->RxAck(subStream));
// Source address on localRank user input (local)
void *srcLocalMemPtr = static_cast<char *>(execMem.inputPtr) + blockOffset;
DeviceMem srcLocal = DeviceMem::create(static_cast<char *>(srcLocalMemPtr), blockSize);
// Destination address on dstRank CCL_Out (remote)
void *dstRemoteMemPtr = nullptr;
CHK_RET(level0CommInfo.links[dstRank]->GetRemoteMem(UserMemType::OUTPUT_MEM, &dstRemoteMemPtr));
DeviceMem dstRemote = DeviceMem::create(static_cast<char *>(dstRemoteMemPtr), blockSize);
// Perform HcclD2DMemcpyAsync: Copy and reduce data from local user input to remote CCL_Out
CHK_RET(HcclReduceAsync(dispatcher_, static_cast<void *>(srcLocal.ptr()), blockSize / unitSize,
param.DataDes.dataType, param.reduceType, subStream,
static_cast<void *>(dstRemote.ptr()),
level0CommInfo.links[dstRank]->GetRemoteRank(),
level0CommInfo.links[dstRank]->GetLinkType(),
INLINE_REDUCE_BIT));
// Notify the remote rank that data transfer and operations have completed
CHK_RET(level0CommInfo.links[dstRank]->TxDataSignal(subStream));
// Wait for data transfer and operations has been completed
CHK_RET(level0CommInfo.links[dstRank]->RxDataSignal(subStream));
}
// Slave streams notify the master stream that their tasks are completed
CHK_RET(SlavesPostToMain(param, execMem));
// The master stream waits for all slave streams to complete their tasks
CHK_RET(MainWaitForSlaves(param, execMem));
return HCCL_SUCCESS;
}
HcclResult CollCustomHugeAllReduceMeshExecutor::UnhealthyPerformReduceScatter(const OpParam ¶m, ExecMem &execMem)
{
// Retrieve the sub-communication domain information and the master stream
CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
// Dertermine the information for the topology and the data
u32 rankSize = level0CommInfo.localRankSize;
u32 rankId = level0CommInfo.localRank;
u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
// Post notify signals from master stream to all slave streams
CHK_RET(MainPostToSlaves(param, execMem));
// Slave streams wait for master's notification
CHK_RET(SlavesWaitForMain(param, execMem));
// Case: Unhealthy ranks reduce data block to healthy ranks
for (u32 round = 1; round < rankSize; round++)
{
// Get the slave stream assigned for communication with dstRank
u32 dstRank = (round + rankId) % rankSize;
Stream &subStream = algResResp_->slaveStreams[round - 1];
// Skipping unhealthy ranks
if (!IsHealthyRank(dstRank))
{
CHK_RET(SlaveExecEmptyTask(param, execMem, round - 1));
continue;
}
// Dertermine the information for the data
u32 blockIdx = GetResponsibleBlockIdx(dstRank);
u64 blockSize = GetBlockSize(blockIdx);
u64 blockOffset = GetBlockOffset(blockIdx);
// Wait for notification from the remote rank
CHK_RET(level0CommInfo.links[dstRank]->RxAck(subStream));
// Source address on localRank user input (local)
void *srcLocalMemPtr = static_cast<char *>(execMem.inputPtr) + blockOffset;
DeviceMem srcLocal = DeviceMem::create(static_cast<char *>(srcLocalMemPtr), blockSize);
// Destination address on dstRank CCL_Out (remote)
void *dstRemoteMemPtr = nullptr;
CHK_RET(level0CommInfo.links[dstRank]->GetRemoteMem(UserMemType::OUTPUT_MEM, &dstRemoteMemPtr));
DeviceMem dstRemote = DeviceMem::create(static_cast<char *>(dstRemoteMemPtr), blockSize);
// Perform HcclD2DMemcpyAsync: Copy and reduce data from local user input to remote CCL_Out
CHK_RET(HcclReduceAsync(dispatcher_, static_cast<void *>(srcLocal.ptr()), blockSize / unitSize,
param.DataDes.dataType, param.reduceType, subStream,
static_cast<void *>(dstRemote.ptr()),
level0CommInfo.links[dstRank]->GetRemoteRank(),
level0CommInfo.links[dstRank]->GetLinkType(),
INLINE_REDUCE_BIT));
// Notify the remote rank that data transfer and operations have completed
CHK_RET(level0CommInfo.links[dstRank]->TxDataSignal(subStream));
}
// Slave streams notify the master stream that their tasks are completed
CHK_RET(SlavesPostToMain(param, execMem));
// The master stream waits for all slave streams to complete their tasks
CHK_RET(MainWaitForSlaves(param, execMem));
return HCCL_SUCCESS;
}
// Perform AllGather
// col0 col1 col2 | col0 col1 col2
// R0: [ 8, 2, 2 ] | R0: [ 8, 8, 8 ]
// R1: [ 2, 8, 2 ] | R1: [ 8, 8, 8 ]
// R2: [ 2, 2, 8 ] | R2: [ 8, 8, 8 ]
// ------------------- --------------------
// R3: [ 2, 2, 2 ] | R3: [ 8, 8, 8 ]
HcclResult CollCustomHugeAllReduceMeshExecutor::HealthyPerformAllGather(const OpParam ¶m, ExecMem &execMem)
{
// Retrieve the sub-communication domain information and the master stream
CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
// Dertermine the information for the topology and the data
u32 rankSize = level0CommInfo.localRankSize;
u32 rankId = level0CommInfo.localRank;
u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
// Post notify signals from master stream to all slave streams
CHK_RET(MainPostToSlaves(param, execMem));
// Slave streams wait for master's notification
CHK_RET(SlavesWaitForMain(param, execMem));
// For healthy ranks, copy data block from CCL_Out to user output buffer
{
u32 blockIdx = GetResponsibleBlockIdx(rankId);
u64 blockSize = GetBlockSize(blockIdx);
u64 blockOffset = GetBlockOffset(blockIdx);
if (blockSize != 0)
{
DeviceMem src = DeviceMem::create(execMem.outputMem.ptr(), blockSize);
DeviceMem dst = DeviceMem::create(static_cast<char *>(execMem.outputPtr) + blockOffset, blockSize);
CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, masterStream));
}
}
// Case A: The healthy rank notifies the unhealthy rank that it can copy data
// Case B: Normal communication between healthy ranks
for (u32 round = 1; round < rankSize; round++)
{
// Get the slave stream assigned for communication with dstRank
u32 dstRank = (rankId - round + rankSize) % rankSize;
Stream &subStream = algResResp_->slaveStreams[round - 1];
// Case A: The healthy rank notifies the unhealthy rank that it can copy data
if (!IsHealthyRank(dstRank))
{
// Notify remote rank data has been ready
CHK_RET(level0CommInfo.links[dstRank]->TxAck(subStream));
// Wait for the remote rank to notify that data transfer and operations have completed
CHK_RET(level0CommInfo.links[dstRank]->RxDataSignal(subStream));
continue;
}
// Case B: Normal communication between healthy ranks
// Dertermine the information for the data
u32 blockIdx = GetResponsibleBlockIdx(dstRank);
u64 blockSize = GetBlockSize(blockIdx);
u64 blockOffset = GetBlockOffset(blockIdx);
// Notify remote rank data has been ready
CHK_RET(level0CommInfo.links[dstRank]->TxAck(subStream));
// Wait for notification from the remote rank
CHK_RET(level0CommInfo.links[dstRank]->RxAck(subStream));
// Source address on dstRank CCL_Out (remote)
void *srcRemoteMemPtr = nullptr;
CHK_RET(level0CommInfo.links[dstRank]->GetRemoteMem(UserMemType::OUTPUT_MEM, &srcRemoteMemPtr));
DeviceMem srcRemote = DeviceMem::create(static_cast<char *>(srcRemoteMemPtr), blockSize);
// Destination address on localRank user-output (local)
void *dstLocalMemPtr = static_cast<char *>(execMem.outputPtr) + blockOffset;
DeviceMem dstLocal = DeviceMem::create(static_cast<char *>(dstLocalMemPtr), blockSize);
// Perform HcclD2DMemcpyAsync
CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstLocal, srcRemote, subStream,
level0CommInfo.links[dstRank]->GetRemoteRank(),
level0CommInfo.links[dstRank]->GetLinkType()));
// Notify the remote rank that data transfer and operations have completed
CHK_RET(level0CommInfo.links[dstRank]->TxDataSignal(subStream));
// Wait for data transfer and operations has been completed
CHK_RET(level0CommInfo.links[dstRank]->RxDataSignal(subStream));
}
// Slave streams notify the master stream that their tasks are completed
CHK_RET(SlavesPostToMain(param, execMem));
// The master stream waits for all slave streams to complete their tasks
CHK_RET(MainWaitForSlaves(param, execMem));
return HCCL_SUCCESS;
}
HcclResult CollCustomHugeAllReduceMeshExecutor::UnhealthyPerformAllGather(const OpParam ¶m, ExecMem &execMem)
{
// Retrieve the sub-communication domain information and the master stream
CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
// Dertermine the information for the topology and the data
u32 rankSize = level0CommInfo.localRankSize;
u32 rankId = level0CommInfo.localRank;
u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
// Post notify signals from master stream to all slave streams
CHK_RET(MainPostToSlaves(param, execMem));
// Slave streams wait for master's notification
CHK_RET(SlavesWaitForMain(param, execMem));
// Case: Unhealthy ranks copy data block from healthy ranks
for (u32 round = 1; round < rankSize; round++)
{
// Get the slave stream assigned for communication with dstRank
u32 dstRank = (rankId - round + rankSize) % rankSize;
Stream &subStream = algResResp_->slaveStreams[round - 1];
// Skipping unhealthy ranks
if (!IsHealthyRank(dstRank))
{
CHK_RET(SlaveExecEmptyTask(param, execMem, round - 1));
continue;
}
// Dertermine the information for the data
u32 blockIdx = GetResponsibleBlockIdx(dstRank);
u64 blockSize = GetBlockSize(blockIdx);
u64 blockOffset = GetBlockOffset(blockIdx);
// Wait for notification from the remote rank
CHK_RET(level0CommInfo.links[dstRank]->RxAck(subStream));
// Source address on dstRank CCL_Out (remote)
void *srcRemoteMemPtr = nullptr;
CHK_RET(level0CommInfo.links[dstRank]->GetRemoteMem(UserMemType::OUTPUT_MEM, &srcRemoteMemPtr));
DeviceMem srcRemote = DeviceMem::create(static_cast<char *>(srcRemoteMemPtr), blockSize);
// Destination address on localRank user-output (local)
void *dstLocalMemPtr = static_cast<char *>(execMem.outputPtr) + blockOffset;
DeviceMem dstLocal = DeviceMem::create(static_cast<char *>(dstLocalMemPtr), blockSize);
// Perform HcclD2DMemcpyAsync
CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstLocal, srcRemote, subStream,
level0CommInfo.links[dstRank]->GetRemoteRank(),
level0CommInfo.links[dstRank]->GetLinkType()));
// Notify the remote rank that data transfer and operations have completed
CHK_RET(level0CommInfo.links[dstRank]->TxDataSignal(subStream));
}
// Slave streams notify the master stream that their tasks are completed
CHK_RET(SlavesPostToMain(param, execMem));
// The master stream waits for all slave streams to complete their tasks
CHK_RET(MainWaitForSlaves(param, execMem));
return HCCL_SUCCESS;
}
// Determine if a rank is healthy
bool CollCustomHugeAllReduceMeshExecutor::IsHealthyRank(u32 rankId)
{
return rankId != BROKEN_RANK_X && rankId != BROKEN_RANK_Y;
}
// Cut the chunk into multiple parts and ensure that the bytes are aligned
HcclResult CollCustomHugeAllReduceMeshExecutor::PrepareBlock(u64 dataCount, u32 unitSize, u32 blockCount)
{
Block temp;
u64 totalSize = dataCount * unitSize;
dataBlock.clear();
dataBlock.reserve(blockCount);
if (blockCount == 0)
{
return HCCL_E_PARA;
}
u64 sizePerBlock = (totalSize + blockCount - 1) / blockCount;
sizePerBlock = RoundUpWithDivisor(sizePerBlock, HCCL_MIN_SLICE_ALIGN_910B);
u64 residueSize = totalSize;
u32 i = 0;
while (residueSize > 0)
{
u64 blockSize = sizePerBlock < residueSize ? sizePerBlock : residueSize;
temp.size = blockSize;
temp.offset = totalSize - residueSize;
i++;
if (blockSize <= 0)
{
return HCCL_E_PARA;
}
residueSize -= blockSize;
dataBlock.push_back(temp);
}
while (i < blockCount)
{
temp.size = 0;
temp.offset = totalSize;
i++;
dataBlock.push_back(temp);
}
return HCCL_SUCCESS;
}
// Get the size of the specific block
u64 CollCustomHugeAllReduceMeshExecutor::GetBlockSize(u32 blockIdx)
{
return dataBlock[blockIdx].size;
}
// Get the offset of the specific block
u64 CollCustomHugeAllReduceMeshExecutor::GetBlockOffset(u32 blockIdx)
{
return dataBlock[blockIdx].offset;
}
// Get the responsible block index of the specific rank
u32 CollCustomHugeAllReduceMeshExecutor::GetResponsibleBlockIdx(u32 rankId)
{
return rankId - 2;
}
// Post notify signals from master stream to all slave streams
HcclResult CollCustomHugeAllReduceMeshExecutor::MainPostToSlaves(const OpParam ¶m, ExecMem &execMem)
{
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
for (u32 signalIndex = 0; signalIndex < algResResp_->slaveStreams.size(); signalIndex++)
{
CHK_RET(LocalNotify::Post(masterStream, dispatcher_,
algResResp_->notifiesAux[signalIndex], PROF_STAGE_1));
}
return HCCL_SUCCESS;
}
// Slave streams wait for master's notification
HcclResult CollCustomHugeAllReduceMeshExecutor::SlavesWaitForMain(const OpParam ¶m, ExecMem &execMem)
{
for (u32 streamIndex = 0; streamIndex < algResResp_->slaveStreams.size(); streamIndex++)
{
CHK_RET(LocalNotify::Wait(algResResp_->slaveStreams[streamIndex], dispatcher_,
algResResp_->notifiesAux[streamIndex], PROF_STAGE_1));
}
return HCCL_SUCCESS;
}
// Slave streams notify the master stream that their tasks are completed
HcclResult CollCustomHugeAllReduceMeshExecutor::SlavesPostToMain(const OpParam ¶m, ExecMem &execMem)
{
for (u32 streamIndex = 0; streamIndex < algResResp_->slaveStreams.size(); streamIndex++)
{
CHK_RET(LocalNotify::Post(algResResp_->slaveStreams[streamIndex], dispatcher_,
algResResp_->notifiesMain[streamIndex], PROF_STAGE_1));
}
return HCCL_SUCCESS;
}
// The master stream waits for all slave streams to complete their tasks
HcclResult CollCustomHugeAllReduceMeshExecutor::MainWaitForSlaves(const OpParam ¶m, ExecMem &execMem)
{
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
for (u32 signalIndex = 0; signalIndex < algResResp_->slaveStreams.size(); signalIndex++)
{
CHK_RET(LocalNotify::Wait(masterStream, dispatcher_,
algResResp_->notifiesMain[signalIndex], PROF_STAGE_1));
}
return HCCL_SUCCESS;
}
// The master stream executes an empty task to ensure synchronization
HcclResult CollCustomHugeAllReduceMeshExecutor::MainExecEmptyTask(const OpParam ¶m, ExecMem &execMem)
{
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
DeviceMem srcTmp = DeviceMem::create(execMem.inputPtr, 0);
DeviceMem dstTmp = DeviceMem::create(execMem.outputPtr, 0);
CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstTmp, srcTmp, masterStream));
return HCCL_SUCCESS;
}
// The slave stream execute an empty task to ensure synchronization
HcclResult CollCustomHugeAllReduceMeshExecutor::SlaveExecEmptyTask(const OpParam ¶m, ExecMem &execMem, u32 streamIndex)
{
hccl::Stream &masterStream = const_cast<hccl::Stream &>(param.stream);
DeviceMem srcTmp = DeviceMem::create(execMem.inputPtr, 0);
DeviceMem dstTmp = DeviceMem::create(execMem.outputPtr, 0);
CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstTmp, srcTmp, algResResp_->slaveStreams[streamIndex]));
return HCCL_SUCCESS;
}
REGISTER_EXEC("CustomHugeAllReduceMeshExecutor", CustomHugeAllReduceMesh, CollCustomHugeAllReduceMeshExecutor);
} // namespace hccl
|
