-
Notifications
You must be signed in to change notification settings - Fork 804
/
enqueue.cc
2428 lines (2237 loc) · 104 KB
/
enqueue.cc
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
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#include "enqueue.h"
#include "argcheck.h"
#include "coll_net.h"
#include "gdrwrap.h"
#include "bootstrap.h"
#include "channel.h"
#include "cudawrap.h"
#include "profiler.h"
#include "transport.h"
#include <cstring> // std::memcpy
#include <cinttypes> // PRIx64
NCCL_PARAM(L1SharedMemoryCarveout, "L1_SHARED_MEMORY_CARVEOUT", 0);
// Returns maximum kernel stack size of all CUDA kernels
ncclResult_t ncclInitKernelsForDevice(int cudaArch, size_t* maxStackSize) {
ncclResult_t result = ncclSuccess;
if (maxStackSize) *maxStackSize = 0;
int carveout = ncclParamL1SharedMemoryCarveout();
for (int k=0; k < ncclDevKernelCount; k++) {
void* fn = ncclDevKernelList[k];
if (fn == nullptr) continue;
if (maxStackSize) {
cudaFuncAttributes attr = {0};
CUDACHECKGOTO(cudaFuncGetAttributes(&attr, fn), result, ignore0);
if (attr.localSizeBytes > *maxStackSize) *maxStackSize = attr.localSizeBytes;
ignore0:;
}
if (carveout) {
CUDACHECKGOTO(cudaFuncSetAttribute(fn,
cudaFuncAttributePreferredSharedMemoryCarveout, carveout),
result, ignore1);
ignore1:;
}
if (ncclShmemDynamicSize(cudaArch) != 0) {
CUDACHECKGOTO(cudaFuncSetAttribute(fn,
cudaFuncAttributeMaxDynamicSharedMemorySize, ncclShmemDynamicSize(cudaArch)),
result, next_kernel);
}
next_kernel:;
}
return result;
}
////////////////////////////////////////////////////////////////////////////////
// Data movement metrics.
static inline int ncclFuncTrafficPerByte(ncclFunc_t func, int nRanks) {
switch (func) {
case ncclFuncAllReduce: return 2;
case ncclFuncAllGather: return nRanks;
case ncclFuncReduceScatter: return nRanks;
default: return 1;
}
}
static inline size_t ncclFuncSendCount(ncclFunc_t func, int nRanks, size_t count) {
return func == ncclFuncReduceScatter ? nRanks*count : count;
}
static inline size_t ncclFuncRecvCount(ncclFunc_t func, int nRanks, size_t count) {
return func == ncclFuncAllGather ? nRanks*count : count;
}
static inline size_t ncclFuncMaxSendRecvCount(ncclFunc_t func, int nRanks, size_t count) {
return func == ncclFuncAllGather || func == ncclFuncReduceScatter ? nRanks*count : count;
}
/*****************************************************************************/
/* Launch system : synchronization and CUDA kernel launch */
/*****************************************************************************/
static ncclResult_t addProxyOpIfNeeded(struct ncclComm* comm, struct ncclKernelPlan* plan, struct ncclProxyOp* op) {
bool needed = true;
NCCLCHECK(ncclProxySaveOp(comm, op, &needed));
if (needed) {
struct ncclProxyOp* q = ncclMemoryPoolAlloc<struct ncclProxyOp>(&comm->memPool_ncclProxyOp, &comm->memPermanent);
*q = *op; // C++ struct assignment
ncclIntruQueueEnqueue(&comm->planner.wipPlan.channels[op->channelId].proxyOpQueue, q);
}
return ncclSuccess;
}
static void addWorkBatchToPlan(
struct ncclComm* comm, struct ncclKernelPlan* plan, int channelId,
enum ncclDevWorkType workType, int devFuncId, uint32_t workOffset,
int p2pRound = -1
) {
ncclKernelPlanner::WipPlan::Channel* chan = &comm->planner.wipPlan.channels[channelId];
size_t workSize = ncclDevWorkSize(workType);
// Conditions causing us to create a new blank batch.
bool newBatch = (chan->workBatchQueue.tail == nullptr);
struct ncclDevWorkBatch* batch = nullptr;
if (!newBatch) {
batch = &chan->workBatchQueue.tail->batch;
// All of the conditions that prevent us from appending to current batch.
newBatch |= batch->workType != (uint8_t)workType;
newBatch |= batch->funcId != devFuncId;
// The following ensure the device can handle a batch this large. They have to
// account for all extension batches being fused together which is why
// wipBatch.workBytes and wipBatch.nP2ps aren't reset to 0 for a new extension
// batch further down.
newBatch |= NCCL_MAX_DEV_WORK_BATCH_BYTES < chan->wipBatch.workBytes + workSize;
if (workType == ncclDevWorkTypeP2p) {
newBatch |= chan->wipBatch.nP2ps == NCCL_MAX_DEV_WORK_P2P_PER_BATCH;
for (int i=0; i < chan->wipBatch.nP2ps; i++) {
newBatch |= p2pRound == chan->wipBatch.p2pRounds[i];
}
}
}
// Conditions causing us to create an extension batch (prev->nextExtends=1)
uint32_t offset = newBatch ? 0 : (workOffset - batch->offsetBase);
bool extendBatch = 63*workSize < offset;
extendBatch |= 0 != offset%workSize;
if (newBatch || extendBatch) {
if (!newBatch) batch->nextExtends = extendBatch; // Extending the previous batch.
struct ncclWorkBatchList* batchNode = ncclMemoryStackAlloc<ncclWorkBatchList>(&comm->memScoped);
// Coverity thinks that ncclIntruQueueEnqueue will access chan->workBatchQueue->tail, which might
// be NULL. But that code is guarded by chan->workBatchQueue->head not being NULL, in which
// case tail won't be NULL either.
// coverity[var_deref_model:FALSE]
ncclIntruQueueEnqueue(&chan->workBatchQueue, batchNode);
batch = &batchNode->batch;
batch->nextExtends = 0;
batch->workType = (uint32_t)workType;
batch->funcId = devFuncId;
batch->offsetBase = workOffset;
batch->offsetBitset = 0;
offset = 0;
if (newBatch) {
// Since extension batches are fused together on the device, and these values
// account for constraints on the fused batch, we only reset the values on
// a new batch
chan->wipBatch.workBytes = 0;
chan->wipBatch.nP2ps = 0;
// We don't count extension batches since this is used to derive a proxyOpCount,
// and we wan't all ops which are fused together to have the same value.
chan->nWorkBatchesP2p += (workType == ncclDevWorkTypeP2p ? 1 : 0);
}
plan->nWorkBatches += 1;
}
batch->offsetBitset |= 1ull<<(offset/workSize);
chan->wipBatch.workBytes += workSize;
if (workType == ncclDevWorkTypeP2p) {
// We need to ensure that a single batch doesn't have multiple p2p's
// of the same round since they would use the same connections.
chan->wipBatch.p2pRounds[chan->wipBatch.nP2ps++] = p2pRound;
}
}
static void finishPlan(struct ncclComm* comm, struct ncclKernelPlan* plan) {
ncclKernelPlanner::WipPlan::Channel* wipChannels = comm->planner.wipPlan.channels;
size_t workBytes = plan->workBytes;
size_t batchBytes = plan->nWorkBatches*sizeof(struct ncclDevWorkBatch);
plan->threadPerBlock = std::max(plan->threadPerBlock, NCCL_MIN_NTHREADS);
// If we can fit everything into the kernel args we do so.
if (sizeof(ncclDevKernelArgs) + batchBytes + workBytes <= comm->workArgsBytes) {
plan->workStorageType = ncclDevWorkStorageTypeArgs;
}
plan->kernelArgsSize = sizeof(struct ncclDevKernelArgs) + batchBytes;
plan->kernelArgsSize += (plan->workStorageType == ncclDevWorkStorageTypeArgs) ? workBytes : 0;
plan->kernelArgsSize = alignUp(plan->kernelArgsSize, 16);
plan->kernelArgs = (struct ncclDevKernelArgs*)ncclMemoryStackAlloc(&comm->memScoped, plan->kernelArgsSize, /*align=*/16);
plan->kernelArgs->comm = comm->devComm;
plan->kernelArgs->channelMask = plan->channelMask;
plan->kernelArgs->workStorageType = plan->workStorageType;
// Put batches into the kernel arguments. The first batch for each channel
// must be located at batchZero[blockIdx.x]. To achieve this we round robin
// over the channels in ascending order until they're exhausted.
uint64_t hasBatchMask = plan->channelMask;
struct ncclDevWorkBatch* batchPrev[MAXCHANNELS] = {}; // {0...}
struct ncclDevWorkBatch* batchZero = (struct ncclDevWorkBatch*)(plan->kernelArgs+1);
int batchIx = 0;
while (hasBatchMask != 0) {
uint64_t tmpMask = hasBatchMask; // channels with a batch for this round.
do {
int c = popFirstOneBit(&tmpMask);
if (!ncclIntruQueueEmpty(&wipChannels[c].workBatchQueue)) {
struct ncclWorkBatchList* batchNode = ncclIntruQueueDequeue(&wipChannels[c].workBatchQueue);
if (batchPrev[c] != nullptr) {
batchPrev[c]->nextJump = int(&batchZero[batchIx] - batchPrev[c]);
}
batchPrev[c] = &batchZero[batchIx];
batchZero[batchIx++] = batchNode->batch;
}
if (ncclIntruQueueEmpty(&wipChannels[c].workBatchQueue)) {
hasBatchMask ^= 1ull<<c;
}
} while (tmpMask != 0);
}
// Merge-sort per-channel proxy-op lists by opCount when merging them into plan->proxyOpQueue
// Phase 1: scan first op of each channel, store opCount in headIds[c].
uint64_t headIds[MAXCHANNELS];
int nHeads = 0;
int channelUbound = 0;
for (int c=0; c < MAXCHANNELS; c++) {
struct ncclProxyOp* op = ncclIntruQueueHead(&wipChannels[c].proxyOpQueue);
headIds[c] = op ? op->opCount : uint64_t(-1);
if (op) nHeads += 1;
if (op) plan->hasProxyOps = true;
if (op) channelUbound = c+1;
}
// Phase 2: Dequeue from planner->channels[c], enqueue in merged order to plan
while (nHeads != 0) {
int c = -1;
uint64_t minId = uint64_t(-1);
// Find channel with least proxy-op id. We store the heads[c]->opCount in
// headIds[c] to remove indirect loads from this loop.
for (int c1=0; c1 < channelUbound; c1++) {
uint64_t id = headIds[c1];
id = (id>>1 | id<<63); // Move tag bit to order collectives before p2p's
if (id < minId) { c = c1; minId = id; }
}
struct ncclProxyOp* op = ncclIntruQueueDequeue(&wipChannels[c].proxyOpQueue);
struct ncclProxyOp* opNext = ncclIntruQueueHead(&wipChannels[c].proxyOpQueue);
headIds[c] = opNext ? opNext->opCount : uint64_t(-1);
nHeads -= opNext ? 0 : 1;
ncclIntruQueueEnqueue(&plan->proxyOpQueue, op);
}
}
int64_t ncclParamLocalRegister();
NCCL_PARAM(GraphRegister, "GRAPH_REGISTER", 1);
struct ncclIpcCleanupCallback {
struct ncclCommCallback base;
void* ptr;
};
static ncclResult_t cleanupIpc(struct ncclComm* comm, struct ncclCommCallback* cb) {
struct ncclIpcCleanupCallback* me = (struct ncclIpcCleanupCallback*)cb;
CUDACHECKIGNORE(cudaIpcCloseMemHandle(me->ptr));
free(me);
return ncclSuccess;
}
static ncclResult_t registerCheckP2PConnection(struct ncclComm* comm, struct ncclConnector* conn, struct ncclTopoGraph* graph, int peer, bool* needReg) {
if (conn->connected) {
if (conn->conn.flags & (NCCL_IPC_READ | NCCL_IPC_WRITE | NCCL_DIRECT_READ | NCCL_DIRECT_WRITE)) {
*needReg = true;
} else {
// network connection
*needReg = false;
}
} else {
struct ncclPeerInfo* peerInfo = &comm->peerInfo[peer];
struct ncclPeerInfo* myInfo = &comm->peerInfo[comm->rank];
int canConnect = 0;
NCCLCHECK(ncclTransports[0]->canConnect(&canConnect, comm, graph, myInfo, peerInfo));
if (canConnect) {
*needReg = true;
} else {
*needReg = false;
}
}
return ncclSuccess;
}
static ncclResult_t registerCollBuffers(
struct ncclComm* comm, struct ncclTaskColl* info,
void* outRegBufSend[NCCL_MAX_LOCAL_RANKS],
void* outRegBufRecv[NCCL_MAX_LOCAL_RANKS],
struct ncclIntruQueue<struct ncclCommCallback, &ncclCommCallback::next>* cleanupQueue,
bool* regNeedConnect
) {
ncclResult_t result = ncclSuccess;
info->regBufType = NCCL_REGULAR_BUFFER;
*regNeedConnect = true;
if (!(ncclParamLocalRegister() || (comm->planner.persistent && ncclParamGraphRegister()))) goto exit;
#if CUDART_VERSION >= 11030
if (info->algorithm == NCCL_ALGO_NVLS || info->algorithm == NCCL_ALGO_NVLS_TREE) {
if (!comm->nvlsRegSupport || info->opDev.op == ncclDevPreMulSum) goto exit;
bool regBufUsed = false;
const void *sendbuff = info->sendbuff;
void *recvbuff = info->recvbuff;
if (info->func == ncclFuncAllGather) sendbuff = NULL;
if (info->func == ncclFuncReduceScatter) recvbuff = NULL;
size_t elementSize = ncclTypeSize(info->datatype);
size_t sendbuffSize = elementSize*ncclFuncSendCount(info->func, comm->nRanks, info->count);
size_t recvbuffSize = elementSize*ncclFuncRecvCount(info->func, comm->nRanks, info->count);
/* first try local registration. */
if (ncclParamLocalRegister()) {
ncclNvlsLocalRegisterBuffer(comm, sendbuff, recvbuff, sendbuffSize, recvbuffSize, ®BufUsed, outRegBufSend, outRegBufRecv);
}
if (regBufUsed == false && comm->planner.persistent && ncclParamGraphRegister()) {
ncclNvlsGraphRegisterBuffer(comm, sendbuff, recvbuff, sendbuffSize, recvbuffSize, ®BufUsed, outRegBufSend, outRegBufRecv, cleanupQueue, &info->nCleanupQueueElts);
}
if (regBufUsed) {
*regNeedConnect = false;
/* tweak NVLS channels usage; for registered NVLS buffer, we only need 4/5 channels to
* saturate bandwidth. */
if (comm->nNodes == 1) {
if (info->func == ncclFuncReduceScatter)
info->nMaxChannels = std::max(comm->config.minCTAs, std::min(comm->config.maxCTAs, 5));
else
info->nMaxChannels = std::max(comm->config.minCTAs, std::min(comm->config.maxCTAs, 4));
} else {
info->nMaxChannels = std::max(comm->config.minCTAs, std::min(comm->config.maxCTAs, 6));
}
info->regBufType = NCCL_NVLS_REG_BUFFER;
}
} else if ((info->algorithm == NCCL_ALGO_COLLNET_DIRECT || info->algorithm == NCCL_ALGO_COLLNET_CHAIN) && comm->collNetRegSupport && info->opDev.op != ncclDevPreMulSum && info->opDev.op != ncclDevSumPostDiv) {
size_t elementSize = ncclTypeSize(info->datatype);
size_t sendbuffSize = elementSize*ncclFuncSendCount(info->func, comm->nRanks, info->count);
size_t recvbuffSize = elementSize*ncclFuncRecvCount(info->func, comm->nRanks, info->count);
int sendRegBufFlag = 0;
int recvRegBufFlag = 0;
void *sendHandle, *recvHandle;
if (ncclParamLocalRegister()) {
ncclCollnetLocalRegisterBuffer(comm, info->sendbuff, sendbuffSize, collNetSend, &sendRegBufFlag, &sendHandle);
info->sendMhandle = sendHandle;
if (sendRegBufFlag) {
ncclCollnetLocalRegisterBuffer(comm, info->recvbuff, recvbuffSize, collNetRecv, &recvRegBufFlag, &recvHandle);
info->recvMhandle = recvHandle;
}
}
if ((sendRegBufFlag == 0 || recvRegBufFlag == 0) && comm->planner.persistent && ncclParamGraphRegister()) {
if (!sendRegBufFlag) {
ncclCollnetGraphRegisterBuffer(comm, info->sendbuff, sendbuffSize, collNetSend, &sendRegBufFlag, &sendHandle, cleanupQueue, &info->nCleanupQueueElts);
info->sendMhandle = sendHandle;
}
if (sendRegBufFlag && !recvRegBufFlag) {
ncclCollnetGraphRegisterBuffer(comm, info->recvbuff, recvbuffSize, collNetRecv, &recvRegBufFlag, &recvHandle, cleanupQueue, &info->nCleanupQueueElts);
info->recvMhandle = recvHandle;
}
}
if (sendRegBufFlag && recvRegBufFlag) {
info->nMaxChannels = 1;
info->regBufType = NCCL_COLLNET_REG_BUFFER;
if (sendRegBufFlag == 1 && recvRegBufFlag == 1) {
INFO(NCCL_REG, "rank %d successfully registered collNet sendbuff %p (handle %p), sendbuff size %ld, recvbuff %p (handle %p), recvbuff size %ld", comm->rank, info->sendbuff, sendHandle, sendbuffSize, info->recvbuff, recvHandle, recvbuffSize);
}
}
} else if (comm->intraNodeP2pSupport && info->protocol == NCCL_PROTO_SIMPLE) {
// IPC buffer registration
if (info->func == ncclFuncReduceScatter) goto exit;
if (info->algorithm == NCCL_ALGO_RING && ((info->func == ncclFuncAllReduce && info->sendbuff == info->recvbuff) || info->func == ncclFuncReduce)) goto exit;
if ((info->algorithm == NCCL_ALGO_TREE || info->algorithm == NCCL_ALGO_COLLNET_CHAIN) && info->sendbuff == info->recvbuff) goto exit;
if (info->func == ncclFuncAllGather && info->algorithm == NCCL_ALGO_PAT) goto exit;
int peerRanks[NCCL_MAX_LOCAL_RANKS];
int nPeers = 0;
size_t elementSize = ncclTypeSize(info->datatype);
size_t sendbuffSize = elementSize*ncclFuncSendCount(info->func, comm->nRanks, info->count);
size_t recvbuffSize = elementSize*ncclFuncRecvCount(info->func, comm->nRanks, info->count);
int regBufFlag = 0;
memset(peerRanks, 0xff, sizeof(int) * NCCL_MAX_LOCAL_RANKS);
if (info->algorithm == NCCL_ALGO_COLLNET_DIRECT) {
struct ncclChannel* channel = comm->channels;
for (int r = 0; r < NCCL_MAX_DIRECT_ARITY; ++r) {
for (int updown = 0; updown < 2; ++updown) {
int peer;
if (updown == 0)
peer = channel->collnetDirect.up[r];
else
peer = channel->collnetDirect.down[r];
if (peer != -1) {
struct ncclConnector* peerConn = &channel->peers[peer]->recv[0];
bool needReg = false;
NCCLCHECK(registerCheckP2PConnection(comm, peerConn, &comm->graphs[NCCL_ALGO_COLLNET_DIRECT], peer, &needReg));
if (needReg) {
bool found = false;
for (int p = 0; p < nPeers; ++p) {
if (peerRanks[p] == peer) {
found = true;
break;
}
}
if (!found) peerRanks[nPeers++] = peer;
}
}
}
}
if (nPeers > 0) {
if (ncclParamLocalRegister())
ncclIpcLocalRegisterBuffer(comm, info->sendbuff, sendbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->sendbuffOffset, &info->sendbuffRmtAddrs);
if (!regBufFlag && comm->planner.persistent && ncclParamGraphRegister()) {
ncclIpcGraphRegisterBuffer(comm, info->sendbuff, sendbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->sendbuffOffset, &info->sendbuffRmtAddrs, cleanupQueue, &info->nCleanupQueueElts);
}
if (regBufFlag) {
if (ncclParamLocalRegister())
ncclIpcLocalRegisterBuffer(comm, info->recvbuff, recvbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->recvbuffOffset, &info->recvbuffRmtAddrs);
if (!regBufFlag && comm->planner.persistent && ncclParamGraphRegister()) {
ncclIpcGraphRegisterBuffer(comm, info->recvbuff, recvbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->recvbuffOffset, &info->recvbuffRmtAddrs, cleanupQueue, &info->nCleanupQueueElts);
}
}
}
if (regBufFlag) {
info->regBufType = NCCL_IPC_REG_BUFFER;
}
} else if (info->algorithm == NCCL_ALGO_RING) {
struct ncclReg* recvRegRecord;
NCCLCHECK(ncclRegFind(comm, info->recvbuff, recvbuffSize, &recvRegRecord));
if (recvRegRecord == NULL) goto exit;
for (int c = 0; c < comm->nChannels; ++c) {
struct ncclChannel* channel = comm->channels + c;
for (int r = 0; r < 2; ++r) {
bool needReg = false;
int peer;
struct ncclConnector* peerConn;
// P2P transport
if (r == 0)
peer = channel->ring.prev;
else
peer = channel->ring.next;
peerConn = &channel->peers[peer]->recv[0];
NCCLCHECK(registerCheckP2PConnection(comm, peerConn, &comm->graphs[NCCL_ALGO_RING], peer, &needReg));
if (needReg) {
bool found = false;
for (int p = 0; p < nPeers; ++p) {
if (peerRanks[p] == peer) {
found = true;
break;
}
}
if (!found) peerRanks[nPeers++] = peer;
}
}
}
if (nPeers > 0) {
if (ncclParamLocalRegister()) {
ncclIpcLocalRegisterBuffer(comm, info->recvbuff, recvbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->recvbuffOffset, &info->recvbuffRmtAddrs);
}
if (!regBufFlag && comm->planner.persistent && ncclParamGraphRegister()) {
ncclIpcGraphRegisterBuffer(comm, info->recvbuff, recvbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->recvbuffOffset, &info->recvbuffRmtAddrs, cleanupQueue, &info->nCleanupQueueElts);
}
}
if (regBufFlag) {
info->regBufType = NCCL_IPC_REG_BUFFER;
}
} else if (info->algorithm == NCCL_ALGO_TREE || info->algorithm == NCCL_ALGO_COLLNET_CHAIN) {
struct ncclReg* recvRegRecord;
NCCLCHECK(ncclRegFind(comm, info->recvbuff, recvbuffSize, &recvRegRecord));
if (recvRegRecord == NULL) goto exit;
for (int c = 0; c < comm->nChannels; ++c) {
struct ncclChannel* channel = comm->channels + c;
struct ncclTree* tree = NULL;
int peers[NCCL_MAX_TREE_ARITY + 1];
if (info->algorithm == NCCL_ALGO_TREE)
tree = &channel->tree;
else
tree = &channel->collnetChain;
for (int p = 0; p < NCCL_MAX_TREE_ARITY; ++p) peers[p] = tree->down[p];
peers[NCCL_MAX_TREE_ARITY] = tree->up;
for (int p = 0; p < NCCL_MAX_TREE_ARITY + 1; ++p) {
int peer = peers[p];
bool peerNeedReg = false;
struct ncclConnector* recvConn = NULL;
// P2P transport
if (peer == -1 || peer == comm->nRanks) continue;
recvConn = &channel->peers[peer]->recv[0];
NCCLCHECK(registerCheckP2PConnection(comm, recvConn, &comm->graphs[info->algorithm], peer, &peerNeedReg));
if (peerNeedReg) {
bool found = false;
for (int pindex = 0; pindex < nPeers; ++pindex) {
if (peerRanks[pindex] == peer) {
found = true;
break;
}
}
if (!found) peerRanks[nPeers++] = peer;
}
}
}
if (nPeers > 0) {
if (ncclParamLocalRegister()) {
ncclIpcLocalRegisterBuffer(comm, info->recvbuff, recvbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->recvbuffOffset, &info->recvbuffRmtAddrs);
}
if (!regBufFlag && comm->planner.persistent && ncclParamGraphRegister()) {
ncclIpcGraphRegisterBuffer(comm, info->recvbuff, recvbuffSize, peerRanks, nPeers, NCCL_IPC_COLLECTIVE, ®BufFlag, &info->recvbuffOffset, &info->recvbuffRmtAddrs, cleanupQueue, &info->nCleanupQueueElts);
}
}
if (regBufFlag) {
info->regBufType = NCCL_IPC_REG_BUFFER;
}
}
if (info->regBufType == NCCL_IPC_REG_BUFFER && comm->nNodes == 1 && 16 < info->nMaxChannels && info->nMaxChannels <= 24) {
info->nMaxChannels = 16;
}
}
exit:
#endif
return result;
}
static ncclResult_t registerP2pBuffer(struct ncclComm* comm, void* userbuff, int peerRank, size_t size, int* regFlag, void** regAddr, struct ncclIntruQueue<struct ncclCommCallback, &ncclCommCallback::next>* cleanupQueue) {
ncclResult_t ret = ncclSuccess;
uintptr_t offset = 0;
uintptr_t* peerRmtAddrs = NULL;
*regFlag = 0;
if (ncclParamLocalRegister()) {
ncclIpcLocalRegisterBuffer(comm, userbuff, size, &peerRank, 1, NCCL_IPC_SENDRECV, regFlag, &offset, &peerRmtAddrs);
}
if (*regFlag == 0 && comm->planner.persistent && ncclParamGraphRegister()) {
ncclIpcGraphRegisterBuffer(comm, userbuff, size, &peerRank, 1, NCCL_IPC_SENDRECV, regFlag, &offset, &peerRmtAddrs, reinterpret_cast<void*>(cleanupQueue), NULL);
}
if (*regFlag)
*regAddr = (void*)((uintptr_t)peerRmtAddrs + offset);
return ret;
}
static ncclResult_t getCollNetSupport(struct ncclComm* comm, struct ncclTaskColl* task, int* collNetSupport);
static ncclResult_t getAlgoInfo(
struct ncclComm* comm, struct ncclTaskColl* task,
int collNetSupport, int nvlsSupport, int numPipeOps, ncclSimInfo_t* simInfo = NULL
);
static ncclResult_t calcCollChunking(
struct ncclComm* comm, struct ncclTaskColl* task, int nChannels, size_t nBytes,
/*outputs*/uint32_t* outChunkSize, uint32_t* outDirectFlags, struct ncclProxyOp* proxyOp
);
struct ncclKernelPlanBudget {
ssize_t inArgsBytes; // Space available within kernel args struct
ssize_t outArgsBytes; // Space available outside of args struct (fifo or persistent buf)
};
static bool testBudget(
struct ncclKernelPlanBudget* budget, int nWorkBatches, ssize_t workBytes
) {
ssize_t batchBytes = nWorkBatches*sizeof(struct ncclDevWorkBatch);
bool ok = false;
ok |= (batchBytes + workBytes <= budget->inArgsBytes);
ok |= (batchBytes <= budget->inArgsBytes) && (workBytes <= budget->outArgsBytes);
return ok;
}
// Called once per ncclGroup to organize the user submitted tasks in
// comm->planner so that they can be peeled off into plans.
ncclResult_t ncclPrepareTasks(struct ncclComm* comm, bool* algoNeedConnect, bool* needConnect, ncclSimInfo_t* simInfo) {
struct ncclKernelPlanner* planner = &comm->planner;
// Tasks from the sorter come out ordered size descending.
struct ncclTaskColl* task = ncclTaskCollSorterDequeueAll(&planner->collSorter);
// Tasks are assembled by (fn,op,ty) size ascending.
struct ncclTaskColl* tasksByFnOpTy[ncclNumFuncs*ncclNumDevRedOps*ncclNumTypes];
memset(tasksByFnOpTy, 0, sizeof(tasksByFnOpTy));
int fnOpTyIndices[ncclNumFuncs*ncclNumDevRedOps*ncclNumTypes];
int fnOpTyCount = 0;
// Walk the size sorted tasks, binning them by (fn,op,ty).
while (task != nullptr) {
struct ncclTaskColl* next = task->next;
int index = ((int)task->func*ncclNumDevRedOps + (int)task->opDev.op)*ncclNumTypes + (int)task->datatype;
// Add to set of (fn,op,ty) indices on first occurrence
if (tasksByFnOpTy[index] == nullptr) fnOpTyIndices[fnOpTyCount++] = index;
// Add to LIFO for this (fn,op,ty)
task->next = tasksByFnOpTy[index];
tasksByFnOpTy[index] = task;
// Next task
task = next;
}
// Walk (fn,op,ty) bins, compute algo and proto etc. Then bin them by their
// scheduling constraints (collnet x nvls).
struct ncclIntruQueue<struct ncclTaskColl, &ncclTaskColl::next> collBins[2][2] = {};
for (int cursor=0; cursor < fnOpTyCount; cursor++) {
struct ncclTaskColl* aggBeg = tasksByFnOpTy[fnOpTyIndices[cursor]];
int collNetSupport = 0;
NCCLCHECK(getCollNetSupport(comm, aggBeg, &collNetSupport));
int nvlsSupport = comm->nvlsSupport && (ncclNvlsSupported(aggBeg->opDev.op, aggBeg->datatype) || aggBeg->func == ncclFuncAllGather);
// Crudely estimate number of tasks per channel. This is using the wrong number
// of channels for NVLS algos, but knowing the algo requires having this value,
// so either be crude our iterate until fixed point, we chose the former.
int nTasksPerChannel = divUp(comm->planner.nTasksColl, comm->nChannels);
do {
struct ncclTaskColl* aggEnd = aggBeg->next;
struct ncclTaskColl agg = *aggBeg;
// We aggregate operations that are within 4X size of each other.
while (aggEnd != nullptr && aggEnd->trafficBytes < 4*aggBeg->trafficBytes) {
agg.count += aggEnd->count;
agg.trafficBytes += aggEnd->trafficBytes;
aggEnd = aggEnd->next;
}
NCCLCHECK(getAlgoInfo(comm, &agg, collNetSupport, nvlsSupport, nTasksPerChannel, simInfo));
agg.devFuncId = ncclDevFuncId(agg.func, agg.opDev.op, agg.datatype, agg.algorithm, agg.protocol);
int isCollnet=0, isNvls=0;
switch (agg.algorithm) {
case NCCL_ALGO_NVLS:
case NCCL_ALGO_NVLS_TREE:
isNvls = 1;
isCollnet = agg.algorithm == NCCL_ALGO_NVLS && comm->nNodes > 1;
break;
case NCCL_ALGO_COLLNET_CHAIN:
case NCCL_ALGO_COLLNET_DIRECT:
isCollnet = 1;
break;
}
// Update the aggregated tasks with the computed values.
do {
struct ncclTaskColl* next = aggBeg->next;
aggBeg->algorithm = agg.algorithm;
aggBeg->protocol = agg.protocol;
aggBeg->nMaxChannels = agg.nMaxChannels;
aggBeg->nWarps = agg.nWarps;
aggBeg->devFuncId = agg.devFuncId;
aggBeg->isCollnet = isCollnet;
aggBeg->isNvls = isNvls;
ncclIntruQueueEnqueue(&collBins[isCollnet][isNvls], aggBeg);
aggBeg = next;
} while (aggBeg != aggEnd);
} while (aggBeg != nullptr);
}
// Concatenate `collBins[*][*]` together into final list `planner->collTaskQueue`.
// Collnet is the outer dimension since that affects how we divide over the
// channels.
for (int isCollnet=0; isCollnet <= 1; isCollnet++) {
for (int isNvls=0; isNvls <= 1; isNvls++) {
ncclIntruQueueTransfer(&planner->collTaskQueue, &collBins[isCollnet][isNvls]);
}
}
// Walk tasks again to:
// 1. Possibly register buffers.
// 2. Build ncclDevWorkColl structs.
// 3. Bin the work structs according to the number of valid channels they
// may be assigned to {collnet, nvls, standard}
task = ncclIntruQueueHead(&planner->collTaskQueue);
while (task != nullptr) {
// Build a ncclDevWorkColl[Reg?] struct for each task.
void* regBufSend[NCCL_MAX_LOCAL_RANKS];
void* regBufRecv[NCCL_MAX_LOCAL_RANKS];
bool regNeedConnect = true;
registerCollBuffers(comm, task, regBufSend, regBufRecv, &planner->collCleanupQueue, ®NeedConnect);
if (comm->runtimeConn && comm->initAlgoChannels[task->algorithm] == false) {
if (task->algorithm == NCCL_ALGO_NVLS_TREE && comm->initAlgoChannels[NCCL_ALGO_NVLS] == false && regNeedConnect == true) {
comm->initAlgoChannels[NCCL_ALGO_NVLS] = true;
algoNeedConnect[NCCL_ALGO_NVLS] = true;
}
if (task->algorithm != NCCL_ALGO_NVLS || regNeedConnect == true) {
comm->initAlgoChannels[task->algorithm] = true;
algoNeedConnect[task->algorithm] = true;
*needConnect = true;
}
}
struct ncclDevWorkColl devWork = {};
devWork.sendbuff = (void*)task->sendbuff;
devWork.recvbuff = (void*)task->recvbuff;
devWork.sendbuffOffset = task->sendbuffOffset;
devWork.recvbuffOffset = task->recvbuffOffset;
devWork.sendbuffRmtAddrs = task->sendbuffRmtAddrs;
devWork.recvbuffRmtAddrs = task->recvbuffRmtAddrs;
devWork.root = task->root;
devWork.nWarps = task->nWarps;
devWork.redOpArg = task->opDev.scalarArg;
devWork.redOpArgIsPtr = task->opDev.scalarArgIsPtr;
devWork.oneNode = (comm->nNodes == 1);
devWork.regUsed = task->regBufType;
struct ncclWorkList* workNode;
switch (task->regBufType) {
case NCCL_REGULAR_BUFFER:
case NCCL_IPC_REG_BUFFER:
case NCCL_COLLNET_REG_BUFFER:
{ workNode = ncclMemoryStackAllocInlineArray<ncclWorkList, ncclDevWorkColl>(&comm->memScoped, 1);
workNode->workType = ncclDevWorkTypeColl;
workNode->size = sizeof(struct ncclDevWorkColl);
memcpy((void*)(workNode+1), (void*)&devWork, workNode->size);
} break;
case NCCL_NVLS_REG_BUFFER:
{ struct ncclDevWorkCollReg workReg = {};
workReg.coll = devWork; // C++ struct assignment
/* NVLS only has one send and recv buffer registered */
workReg.dnInputs[0] = regBufSend[0];
workReg.dnOutputs[0] = regBufRecv[0];
workNode = ncclMemoryStackAllocInlineArray<ncclWorkList, ncclDevWorkCollReg>(&comm->memScoped, 1);
workNode->workType = ncclDevWorkTypeCollReg;
workNode->size = sizeof(struct ncclDevWorkCollReg);
memcpy((void*)(workNode+1), (void*)&workReg, workNode->size);
} break;
default:
/* impossible value */
WARN("Invalid regBufType %d", task->regBufType);
return ncclInvalidArgument;
}
ncclIntruQueueEnqueue(&planner->collWorkQueue, workNode);
task = task->next;
}
return ncclSuccess;
}
static ncclResult_t scheduleCollTasksToPlan(
struct ncclComm* comm, struct ncclKernelPlan* plan, struct ncclKernelPlanBudget* budget
) {
struct ncclKernelPlanner* planner = &comm->planner;
// Estimate number of tasks that will fit in this plan.
int nPlanColls = 0;
size_t trafficBytes[2*2] = {0, 0, 0, 0}; // [collnet][nvls]
int nChannels[2*2] = {0, 0, 0, 0}; // [collnet][nvls]
int const nMaxChannels[2*2] = {comm->nChannels, comm->nvlsChannels, // [collnet][nvls]
comm->nChannels, comm->nvlsChannels};
constexpr size_t MinTrafficPerChannel = 16 << 10; // 16K traffic as minimal
do {
size_t workBytes = 0;
struct ncclTaskColl* task = ncclIntruQueueHead(&planner->collTaskQueue);
struct ncclWorkList* workNode = ncclIntruQueueHead(&planner->collWorkQueue);
while (task != nullptr) {
int nBatches = divUp(nPlanColls, 4); // Rough guess: 4 colls per batch.
if (!testBudget(budget, nBatches, workBytes + workNode->size)) goto plan_full;
nPlanColls += 1;
workBytes += workNode->size;
int kind = 2*task->isCollnet + task->isNvls;
trafficBytes[kind] += std::max(MinTrafficPerChannel, task->trafficBytes);
nChannels[kind] += task->nMaxChannels;
nChannels[kind] = std::min(nChannels[kind], nMaxChannels[kind]);
task = task->next;
workNode = workNode->next;
}
plan_full:;
} while (0);
int kindPrev = -1;
size_t trafficPerChannel = 0;
int channelId = 0;
size_t currentTraffic = 0;
while (nPlanColls!=0 && !ncclIntruQueueEmpty(&planner->collTaskQueue)) {
struct ncclTaskColl* task = ncclIntruQueueHead(&planner->collTaskQueue);
struct ncclWorkList* workNode = ncclIntruQueueHead(&planner->collWorkQueue);
struct ncclDevWorkColl* devWork = (struct ncclDevWorkColl*)(workNode+1);
size_t elementSize = ncclTypeSize(task->datatype);
int kind = 2*task->isCollnet + task->isNvls;
if (kind != kindPrev) {
trafficPerChannel = std::max<size_t>(MinTrafficPerChannel, trafficBytes[kind]/nChannels[kind]);
kindPrev = kind;
channelId = 0;
currentTraffic = 0;
}
if (task->isCollnet) {
int nChannels = task->nMaxChannels;
// Ensure room for worst case of one new batch per channel
if (!testBudget(budget, plan->nWorkBatches + nChannels, plan->workBytes + workNode->size)) {
return ncclSuccess;
}
size_t globalBytesPerElement = elementSize*ncclFuncMaxSendRecvCount(task->func, comm->nRanks, 1);
struct ncclProxyOp proxyOp;
uint32_t chunkSize, directFlags=0;
NCCLCHECK(calcCollChunking(comm, task, nChannels, globalBytesPerElement*task->count, &chunkSize, &directFlags, &proxyOp));
devWork->channelLo = 0;
devWork->channelHi = nChannels-1;
devWork->collnet.count = task->count;
devWork->collnet.chunkCount = chunkSize/ncclTypeSize(task->datatype);
devWork->direct = directFlags;
uint64_t proxyOpId = uint64_t(plan->collOpCount++)<<1 | 0;
for (int c=devWork->channelLo; c <= (int)devWork->channelHi; c++) {
proxyOp.channelId = c;
proxyOp.opCount = proxyOpId;
proxyOp.task.coll = task;
proxyOp.rank = comm->rank;
addWorkBatchToPlan(comm, plan, c, workNode->workType, task->devFuncId, plan->workBytes);
NCCLCHECK(addProxyOpIfNeeded(comm, plan, &proxyOp));
}
} else { // not task->isCollnet
int trafficPerByte = ncclFuncTrafficPerByte(task->func, comm->nRanks);
size_t cellSize = divUp(divUp(MinTrafficPerChannel, (size_t)trafficPerByte), 16) * 16;
int elementsPerCell = cellSize/elementSize;
size_t cells = divUp(task->count*elementSize, cellSize);
size_t trafficPerElement = elementSize*trafficPerByte;
size_t trafficPerCell = cellSize*trafficPerByte;
size_t cellsPerChannel = std::min(cells, divUp(trafficPerChannel, trafficPerCell));
size_t cellsLo;
if (channelId+1 == nMaxChannels[kind]) { // On last channel everything goes to "lo"
cellsLo = cells;
} else {
cellsLo = std::min(cells, divUp((trafficPerChannel-currentTraffic),trafficPerCell));
}
int nMidChannels = (cells-cellsLo)/cellsPerChannel;
size_t cellsHi = (cells-cellsLo)%cellsPerChannel;
int nChannels = (cellsLo!=0 ? 1 : 0) + nMidChannels + (cellsHi!=0 ? 1 : 0);
if (nMaxChannels[kind] < channelId + nChannels) { // Overflowed available channels
nMidChannels = nMaxChannels[kind] - channelId - 2;
cellsPerChannel = (cells-cellsLo)/(nMidChannels+1);
cellsHi = cellsPerChannel + (cells-cellsLo)%(nMidChannels+1);
}
if (cellsHi == 0 && nMidChannels != 0) {
cellsHi = cellsPerChannel;
nMidChannels -= 1;
}
if (cellsLo == 0) { // Least channel skipped. Make the next channel the new least.
channelId += 1;
if (nMidChannels == 0) { cellsLo = cellsHi; cellsHi = 0; }
else { cellsLo = cellsPerChannel; nMidChannels -= 1; }
}
size_t countMid = nMidChannels!=0 ? cellsPerChannel*elementsPerCell : 0;
size_t countLo = cellsLo*elementsPerCell;
size_t countHi = cellsHi*elementsPerCell;
(countHi != 0 ? countHi : countLo) -= cells*elementsPerCell - task->count;
nChannels = (countLo!=0 ? 1 : 0) + nMidChannels + (cellsHi!=0 ? 1 : 0);
// Ensure room for worst case of one new batch per channel
if (!testBudget(budget, plan->nWorkBatches + nChannels, plan->workBytes + workNode->size)) {
return ncclSuccess;
}
devWork->channelLo = channelId;
devWork->channelHi = channelId + nChannels-1;
devWork->cbd.countLo = countLo;
devWork->cbd.countMid = countMid;
devWork->cbd.countHi = countHi;
// calcCollChunking() uses global bytes instead of traffic which differs
// in that allreduce isn't multiplied by 2.
size_t globalBytesPerElement = elementSize*ncclFuncMaxSendRecvCount(task->func, comm->nRanks, 1);
struct ncclProxyOp proxyOpLo, proxyOpMid, proxyOpHi;
uint32_t chunkSize, directFlags=0;
size_t grainSize = ncclProtoGrainSize(task->protocol);
if (countLo != 0) {
NCCLCHECK(calcCollChunking(comm, task, /*nChannels=*/1, globalBytesPerElement*countLo, &chunkSize, &directFlags, &proxyOpLo));
devWork->cbd.chunkGrainsLo = chunkSize/grainSize;
}
if (countHi != 0) {
NCCLCHECK(calcCollChunking(comm, task, /*nChannels=*/1, globalBytesPerElement*countHi, &chunkSize, &directFlags, &proxyOpHi));
devWork->cbd.chunkGrainsHi = chunkSize/grainSize;
}
if (nMidChannels != 0) {
NCCLCHECK(calcCollChunking(comm, task, /*nChannels=*/1, globalBytesPerElement*countMid, &chunkSize, &directFlags, &proxyOpMid));
devWork->cbd.chunkGrainsMid = chunkSize/grainSize;
}
devWork->direct = directFlags;
// Update the current channel and vacant traffic budget.
if (countHi != 0) {
channelId += nChannels-1;
currentTraffic = cellsHi*elementsPerCell*trafficPerElement;
} else if (nMidChannels != 0) {
channelId += nChannels;
currentTraffic = 0;
} else {
currentTraffic += cellsLo*elementsPerCell*trafficPerElement;
}
if (currentTraffic >= trafficPerChannel && channelId+1 != nMaxChannels[kind]) {
channelId += 1;
currentTraffic = 0;
}
uint64_t proxyOpId = uint64_t(plan->collOpCount++)<<1 | 0;
for (int c=devWork->channelLo; c <= (int)devWork->channelHi; c++) {
struct ncclProxyOp* proxyOp;
if (c == (int)devWork->channelLo) {
proxyOp = &proxyOpLo;
} else if (c == (int)devWork->channelHi) {
proxyOp = &proxyOpHi;
} else {
proxyOp = &proxyOpMid;
}
proxyOp->channelId = c;
proxyOp->opCount = proxyOpId;
proxyOp->task.coll = task;
proxyOp->rank = comm->rank;
addWorkBatchToPlan(comm, plan, c, workNode->workType, task->devFuncId, plan->workBytes);
// Coverity reports "proxyOp->connection" as being possibly uninitialized. It's hard to
// determine if that's actually true but it's also not clear if that would be an issue.
// coverity[uninit_use_in_call:FALSE]
NCCLCHECK(addProxyOpIfNeeded(comm, plan, proxyOp));
}
}
plan->channelMask |= (2ull<<devWork->channelHi) - (1ull<<devWork->channelLo);
plan->threadPerBlock = std::max(plan->threadPerBlock, task->nWarps*WARP_SIZE);
if (!plan->kernelSpecialized) {
plan->kernelFn = ncclDevKernelForFunc[task->devFuncId];
plan->kernelSpecialized = ncclDevKernelForFuncIsSpecialized[task->devFuncId];
}
if (comm->rank == 0) {
if (task->isCollnet) {
TRACE(NCCL_COLL, "Collective %s(%s, %s, %s, %s) count=%ld devFuncId=%d channel{Lo..Hi}={%d..%d} count=%ld chunkCount=%d",
ncclFuncToString(task->func), ncclDevRedOpToString(task->opDev.op),
ncclDatatypeToString(task->datatype), ncclAlgoToString(task->algorithm),
ncclProtoToString(task->protocol),
(long)task->count, task->devFuncId, devWork->channelLo, devWork->channelHi,
(long)devWork->collnet.count, devWork->collnet.chunkCount);
} else {
TRACE(NCCL_COLL, "Collective %s(%s, %s, %s, %s) count=%ld devFuncId=%d channel{Lo..Hi}={%d..%d} count{Lo,Mid,Hi}={%ld,%ld,%ld} chunkBytes{Lo,Mid,Hi}={%d,%d,%d}",
ncclFuncToString(task->func), ncclDevRedOpToString(task->opDev.op),
ncclDatatypeToString(task->datatype), ncclAlgoToString(task->algorithm),
ncclProtoToString(task->protocol),
(long)task->count, task->devFuncId, devWork->channelLo, devWork->channelHi,
(long)devWork->cbd.countLo, (long)devWork->cbd.countMid, (long)devWork->cbd.countHi,
int(devWork->cbd.chunkGrainsLo*ncclProtoGrainSize(task->protocol)),
int(devWork->cbd.chunkGrainsMid*ncclProtoGrainSize(task->protocol)),
int(devWork->cbd.chunkGrainsHi*ncclProtoGrainSize(task->protocol)));
}
}
for (int i=0; i < task->nCleanupQueueElts; i++) {
ncclIntruQueueEnqueue(&plan->cleanupQueue, ncclIntruQueueDequeue(&planner->collCleanupQueue));
}
ncclIntruQueueDequeue(&planner->collTaskQueue);
ncclIntruQueueDequeue(&planner->collWorkQueue);
nPlanColls -= 1;
planner->nTasksColl -= 1;
ncclIntruQueueEnqueue(&plan->collTaskQueue, task);
ncclIntruQueueEnqueue(&plan->workQueue, workNode);
plan->workBytes += workNode->size;
}
return ncclSuccess;
}
NCCL_PARAM(P2pLLThreshold, "P2P_LL_THRESHOLD", 16384);
NCCL_PARAM(ChunkSize, "CHUNK_SIZE", 0);
// Put p2p op in plan assuming there is sizeof(ncclDevWorkBatch) in batch budget
// and sizeof(ncclDevWorkP2p) in work budget. "sendRank" and "recvRank" must
// match the corresponding values for this round of the p2p schedule (no -1's).
// No-op's are encoded with a -1 size.
static ncclResult_t addP2pToPlan(
struct ncclComm* comm, struct ncclKernelPlan* plan,
int nChannelsMin, int nChannelsMax, int p2pRound,
int sendRank, void* sendAddr, ssize_t sendBytes,
int recvRank, void* recvAddr, ssize_t recvBytes,
struct ncclTaskP2p** p2pTasks
) {
constexpr int connIndex = 1;
bool selfSend = (sendRank == comm->rank);
// recv: dir=0, send: dir=1
void* addrs[2] = {recvAddr, sendAddr};
ssize_t bytes[2] = {recvBytes, sendBytes};
bool protoLL[2] = {!selfSend, !selfSend};
bool network[2] = {false, false};
bool proxySameProcess[2] = {true, true};
uint8_t base = ncclP2pChannelBaseForRound(comm, p2pRound);
if (!selfSend) {
for (int part=0; part < nChannelsMax; part++) {
int channelId = ncclP2pChannelForPart(comm->p2pnChannels, base, part);
struct ncclChannelPeer** channelPeers = comm->channels[channelId].peers;
for (int dir=0; dir <= 1; dir++) {
int peerRank = dir ? sendRank : recvRank;
struct ncclConnector* conn = dir ? &channelPeers[peerRank]->send[connIndex]
: &channelPeers[peerRank]->recv[connIndex];
protoLL[dir] &= conn->conn.buffs[NCCL_PROTO_LL] != nullptr;
network[dir] |= conn->transportComm == (dir ? &netTransport.send : &netTransport.recv);
proxySameProcess[dir] &= conn->proxyConn.sameProcess;
}
}
}
ssize_t thresholdLL = nChannelsMax*ncclParamP2pLLThreshold();
ssize_t paramChunkSize = ncclParamChunkSize();
// Arrays indexed by dir where recv=0, send=1:
int nChannels[2];
int protocol[2];
int stepSize[2];
int chunkSize[2];
int chunkDataSize[2];
int chunkDataSize_u32fp8[2];
bool registered[2] = {false, false};
bool ipcRegistered[2] = {false, false};
for (int dir=0; dir < 2; dir++) { // 0=recv, 1=send
if (bytes[dir] != -1) protoLL[dir] &= bytes[dir] <= thresholdLL;
protocol[dir] = protoLL[dir] ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
stepSize[dir] = comm->buffSizes[protocol[dir]]/NCCL_STEPS;
if (protocol[dir] == NCCL_PROTO_SIMPLE) stepSize[dir] = comm->p2pChunkSize;
chunkSize[dir] = stepSize[dir];
if (paramChunkSize != 0) {
chunkSize[dir] = paramChunkSize;
} else if (network[dir]) {
// Tune chunk size for the network
if (protocol[dir] == NCCL_PROTO_SIMPLE && bytes[dir] < stepSize[dir]) chunkSize[dir] /= 4;
else if (bytes[dir] < 8*stepSize[dir]) chunkSize[dir] /= 2;
}