Line data Source code
1 : /**
2 : * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3 : * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 : * CANN Open Software License Agreement Version 2.0 (the "License").
5 : * Please refer to the License for details. You may not use this file except in compliance with the License.
6 : * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7 : * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8 : * See LICENSE in the root of the software repository for the full text of the License.
9 : */
10 :
11 : #include <cmath>
12 : #include "alg_template_register.h"
13 : #include "reduce_scatter_plant_local_reduce.h"
14 :
15 : namespace hccl {
16 : constexpr u32 DEVICE_EIGHT = 8;
17 : constexpr u32 FACTOR_NUM_TWO = 2;
18 0 : ReduceScatterPlantLocalReduce::ReduceScatterPlantLocalReduce(const HcclDispatcher dispatcher)
19 0 : : AlgTemplateBase(dispatcher)
20 0 : {}
21 :
22 0 : ReduceScatterPlantLocalReduce::~ReduceScatterPlantLocalReduce()
23 0 : {}
24 :
25 0 : HcclResult ReduceScatterPlantLocalReduce::Prepare(void *inputMemPtr, DeviceMem &cclInMem, DeviceMem &outputMem,
26 : const Stream &stream, std::vector<Stream> &subStreams, std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
27 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, GroupSlicesInfo &grouSlicesInfo,
28 : const HcclReduceOp reductionOp, u32 all2allOffset, const HcclDataType dataType, bool isNeedSpaceBorrow,
29 : bool reverseMemUsage, bool isA3CrossNode)
30 : {
31 0 : inputMemPtr_ = inputMemPtr; // UserInPtr，All2All使用
32 0 : inputMem_ = cclInMem; // 空拷贝 & 存放最后一块数据（Allreduce非整除场景）
33 0 : outputMem_ = outputMem; // 单算子CclOut 图模式Scrach/UserOut，LocalReduce使用
34 0 : stream_ = stream;
35 0 : subStreams_ = subStreams;
36 0 : meshSignalPtr_ = &meshSignal;
37 0 : meshSignalAuxPtr_ = &meshSignalAux;
38 0 : groupSlicesInfo_ = std::move(grouSlicesInfo);
39 0 : reductionOp_ = reductionOp;
40 0 : all2allOffset_ = all2allOffset;
41 0 : dataType_ = dataType;
42 0 : isNeedSpaceBorrow_ = isNeedSpaceBorrow;
43 0 : isA3CrossNode_ = isA3CrossNode;
44 0 : if (reverseMemUsage) {
45 : // 交换两块buffer的用途，in buffer作为输出buffer
46 0 : HCCL_INFO("[%s] reverse memory usage.", __func__);
47 0 : std::swap(scratchMemType_, outputMemType_);
48 0 : std::swap(inputMem_, outputMem_);
49 : }
50 0 : return HCCL_SUCCESS;
51 : }
52 :
53 0 : HcclResult ReduceScatterPlantLocalReduce::MainRecordSub(Stream &mainStream, u32 firstSubStreamIndex,
54 : u32 totalTask)
55 : {
56 0 : for (u32 streamIndex = firstSubStreamIndex; streamIndex < totalTask; streamIndex++) {
57 0 : CHK_RET(LocalNotify::Post(mainStream, dispatcher_, (*meshSignalAuxPtr_)[streamIndex], profilerInput_.stage));
58 : }
59 0 : return HCCL_SUCCESS;
60 : }
61 :
62 0 : HcclResult ReduceScatterPlantLocalReduce::SubWaitMain(u32 firstSubStreamIndex, u32 totalTask)
63 : {
64 0 : for (u32 streamIndex = firstSubStreamIndex; streamIndex < totalTask; streamIndex++) {
65 0 : CHK_RET(LocalNotify::Wait(subStreams_[streamIndex], dispatcher_,
66 : (*meshSignalAuxPtr_)[streamIndex], profilerInput_.stage));
67 : }
68 0 : return HCCL_SUCCESS;
69 : }
70 :
71 0 : HcclResult ReduceScatterPlantLocalReduce::MainWaitSub(Stream &mainStream, u32 firstSubStreamIndex, u32 totalTask)
72 : {
73 0 : for (u32 streamIndex = firstSubStreamIndex; streamIndex < totalTask; streamIndex++) {
74 0 : CHK_RET(LocalNotify::Wait(mainStream, dispatcher_, (*meshSignalPtr_)[streamIndex], profilerInput_.stage));
75 : }
76 0 : return HCCL_SUCCESS;
77 : }
78 :
79 0 : HcclResult ReduceScatterPlantLocalReduce::SubRecordMain(u32 firstSubStreamIndex, u32 totalTask)
80 : {
81 0 : for (u32 streamIndex = firstSubStreamIndex; streamIndex < totalTask; streamIndex++) {
82 0 : CHK_RET(LocalNotify::Post(subStreams_[streamIndex], dispatcher_, (*meshSignalPtr_)[streamIndex],
83 : profilerInput_.stage));
84 : }
85 0 : return HCCL_SUCCESS;
86 : }
87 :
88 0 : HcclResult ReduceScatterPlantLocalReduce::MainRecordLocalReduceWait(u32 lRMainStreamIndex)
89 : {
90 0 : CHK_RET(LocalNotify::Post(stream_, dispatcher_, (*meshSignalAuxPtr_)[lRMainStreamIndex], profilerInput_.stage));
91 0 : CHK_RET(LocalNotify::Wait(subStreams_[lRMainStreamIndex], dispatcher_, (*meshSignalAuxPtr_)[lRMainStreamIndex],
92 : profilerInput_.stage));
93 0 : return HCCL_SUCCESS;
94 : }
95 :
96 0 : u32 ReduceScatterPlantLocalReduce::CalcOutputIndex(const u32 round)
97 : {
98 0 : return (all2allOffset_ + round + localRank_) % rankSize_;
99 : }
100 :
101 0 : bool ReduceScatterPlantLocalReduce::isLastGroup(const u32 groupId)
102 : {
103 0 : return groupId == groupSlicesInfo_.size() - 1;
104 : }
105 :
106 0 : bool ReduceScatterPlantLocalReduce::isLastRank(const u32 rankId)
107 : {
108 0 : return rankId == rankSize_ - 1;
109 : }
110 :
111 0 : bool ReduceScatterPlantLocalReduce::isLastBlockData(const u32 outputIndex)
112 : {
113 0 : return outputIndex == rankSize_ - 1;
114 : }
115 :
116 0 : HcclResult ReduceScatterPlantLocalReduce::RunAsync(const u32 rank, const u32 rankSize, const std::vector<LINK> &links)
117 : {
118 0 : HCCL_INFO("ReduceScatterPlantLocalReduce run: rank[%u] ranksize[%u] inputMem[%p] outputMem[%p].",
119 : rank, rankSize, inputMem_.ptr(), outputMem_.ptr());
120 0 : CHK_SMART_PTR_NULL(dispatcher_);
121 0 : CHK_PTR_NULL(stream_.ptr());
122 0 : CHK_PRT_RET(links.size() < rankSize, HCCL_ERROR("[%s]rank[%u] linksize[%llu] is less than rankSize[%u]",
123 : __func__, rank, links.size(), rankSize), HCCL_E_INTERNAL);
124 :
125 0 : rankSize_ = rankSize;
126 0 : localRank_ = rank;
127 :
128 : // All2All主流（主流）通知LocalReduce主流开始准备执行,
129 : // All2All需要rankSize条流，其中主流完成LocalCopy&第一个A2A任务，因此主从同步需要rankSize-2个任务。lRMainStreamId_需要-2
130 0 : all2allSubStreamNum_ = isA3CrossNode_ ? std::min(rankSize, DEVICE_EIGHT) - 1 : rankSize - 2;
131 0 : lRMainStreamId_ = all2allSubStreamNum_;
132 :
133 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, stream_, dispatcher_));
134 0 : CHK_RET(MainRecordLocalReduceWait(lRMainStreamId_));
135 : // 额外一次LocalReduce主流通知All2All主流准备好接受信息（通知第一次执行完的All2AllWait）
136 0 : CHK_RET(LocalNotify::Post(subStreams_[lRMainStreamId_], dispatcher_, (*meshSignalPtr_)[lRMainStreamId_],
137 : profilerInput_.stage));
138 :
139 0 : HcclResult ret = HCCL_SUCCESS;
140 0 : for (u32 groupId = 0; groupId < groupSlicesInfo_.size(); groupId++) {
141 0 : const MemBlockInfo& memBlockInfo = groupSlicesInfo_[groupId];
142 0 : if (isA3CrossNode_) {
143 0 : ret = RunGroupAlltoAll(links, groupId, memBlockInfo);
144 : } else {
145 0 : ret = RunAlltoAll(links, groupId, memBlockInfo);
146 : }
147 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[%s]RunAlltoAll or RunGroupAlltoAll failed, localRank[%u], groupId[%u]",
148 : __func__, localRank_, groupId), ret);
149 :
150 0 : CHK_RET(LocalNotify::Wait(stream_, dispatcher_, (*meshSignalPtr_)[lRMainStreamId_], profilerInput_.stage));
151 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, stream_, dispatcher_));
152 0 : CHK_RET(MainRecordLocalReduceWait(lRMainStreamId_));
153 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, subStreams_[lRMainStreamId_], dispatcher_));
154 :
155 0 : ret = RunLocalReduce(groupId, memBlockInfo);
156 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[%s]LocalReduce failed, localRank[%u], groupId[%u]",
157 : __func__, localRank_, groupId), ret);
158 :
159 : // LocalReduce主流通知All2All主流执行完成，可以下发下一次LocalReduce操作
160 0 : CHK_RET(LocalNotify::Post(subStreams_[lRMainStreamId_], dispatcher_, (*meshSignalPtr_)[lRMainStreamId_],
161 : profilerInput_.stage));
162 : }
163 :
164 : // All2All主流等待最后一次LocalReduce执行完成
165 0 : CHK_RET(LocalNotify::Wait(stream_, dispatcher_, (*meshSignalPtr_)[lRMainStreamId_], profilerInput_.stage));
166 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, stream_, dispatcher_));
167 0 : HCCL_INFO("ReduceScatterPlantLocalReduce finished: localRank[%u] ranksize[%u]", localRank_, rankSize_);
168 0 : return HCCL_SUCCESS;
169 : }
170 :
171 0 : HcclResult ReduceScatterPlantLocalReduce::LocalCopy(u32 groupId, const MemBlockInfo& memBlockInfo)
172 : {
173 0 : u64 sliceSize = memBlockInfo.size[localRank_];
174 0 : if (sliceSize == 0) {
175 0 : return HCCL_SUCCESS;
176 : }
177 :
178 0 : DeviceMem src = DeviceMem::create(static_cast<u8 *>(inputMemPtr_) +
179 0 : memBlockInfo.userInputOffsets[localRank_], sliceSize);
180 :
181 : // 当非最后一组最后一卡且outputIndex是最后一块时，Copy至CclIn/UserIn预留位
182 0 : DeviceMem dst;
183 0 : u32 outputIndex = CalcOutputIndex(localRank_);
184 0 : if (isNeedSpaceBorrow_ && isLastBlockData(outputIndex) && !(isLastRank(localRank_) && isLastGroup(groupId))) {
185 0 : dst = inputMem_.range(memBlockInfo.outputOffsets[localRank_], sliceSize);
186 : } else {
187 0 : dst = outputMem_.range(memBlockInfo.outputOffsets[outputIndex], sliceSize);
188 : }
189 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, stream_));
190 0 : return HCCL_SUCCESS;
191 0 : }
192 :
193 0 : HcclResult ReduceScatterPlantLocalReduce::RunAlltoAll(const std::vector<LINK> &links, u32 groupId,
194 : const MemBlockInfo& memBlockInfo)
195 : {
196 : // 本卡优先拷贝同号位数据
197 0 : CHK_RET(LocalCopy(groupId, memBlockInfo));
198 : // 主流通知从流可以开始接受数据
199 0 : u32 all2allfirstSubStreamId = 0;
200 0 : CHK_RET(MainRecordSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
201 0 : CHK_RET(SubWaitMain(all2allfirstSubStreamId, all2allSubStreamNum_));
202 :
203 : // 开始数据拷贝
204 0 : u32 streamIndex = 0;
205 0 : for (u32 round = 0; round < rankSize_; round++) {
206 0 : if (round == localRank_) {
207 0 : continue;
208 : }
209 0 : Stream &subStream = (streamIndex == 0) ? stream_ : subStreams_[streamIndex - 1];
210 0 : CHK_SMART_PTR_NULL(links[round]);
211 0 : CHK_RET(links[round]->TxAck(subStream));
212 0 : CHK_RET(links[round]->RxAck(subStream));
213 0 : streamIndex++;
214 : }
215 :
216 0 : CHK_RET(SubRecordMain(all2allfirstSubStreamId, all2allSubStreamNum_));
217 0 : CHK_RET(MainWaitSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
218 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, subStreams_[lRMainStreamId_], dispatcher_));
219 :
220 0 : CHK_RET(MainRecordSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
221 0 : CHK_RET(SubWaitMain(all2allfirstSubStreamId, all2allSubStreamNum_));
222 0 : streamIndex = 0;
223 0 : for (u32 round = 0; round < rankSize_; round++) {
224 0 : if (round == localRank_) {
225 0 : continue;
226 : }
227 0 : Stream &subStream = (streamIndex == 0) ? stream_ : subStreams_[streamIndex - 1];
228 0 : CHK_SMART_PTR_NULL(links[round]);
229 :
230 0 : u64 sliceSize = memBlockInfo.size[round];
231 0 : if (sliceSize != 0) {
232 0 : u64 userMemInOffset = memBlockInfo.userInputOffsets[round];
233 0 : DeviceMem src = DeviceMem::create(static_cast<u8 *>(inputMemPtr_) + userMemInOffset, sliceSize);
234 0 : u32 outputIndex = CalcOutputIndex(round);
235 0 : u64 dstOffset = 0;
236 0 : void *remMemPtr = nullptr;
237 0 : if (isNeedSpaceBorrow_ && isLastBlockData(outputIndex) && !(isLastRank(round) && isLastGroup(groupId))) {
238 0 : CHK_RET(links[round]->GetRemoteMem(scratchMemType_, &remMemPtr));
239 0 : dstOffset = memBlockInfo.outputOffsets[round];
240 : } else {
241 0 : CHK_RET(links[round]->GetRemoteMem(outputMemType_, &remMemPtr));
242 0 : dstOffset = memBlockInfo.outputOffsets[outputIndex];
243 : }
244 0 : DeviceMem dst = DeviceMem::create(static_cast<u8 *>(remMemPtr) + dstOffset, sliceSize);
245 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, subStream, links[round]->GetRemoteRank(),
246 : links[round]->GetLinkType()));
247 0 : }
248 0 : CHK_RET(links[round]->TxDataSignal(subStream));
249 0 : CHK_RET(links[round]->RxDataSignal(subStream));
250 0 : streamIndex++;
251 : }
252 :
253 : // 从流通知主流完成拷贝
254 0 : CHK_RET(SubRecordMain(all2allfirstSubStreamId, all2allSubStreamNum_));
255 0 : CHK_RET(MainWaitSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
256 0 : return HCCL_SUCCESS;
257 : }
258 :
259 0 : HcclResult ReduceScatterPlantLocalReduce::RunGroupAlltoAll(const std::vector<LINK> &links, u32 groupId,
260 : const MemBlockInfo& memBlockInfo)
261 : {
262 0 : constexpr u32 numInGroup = DEVICE_EIGHT;
263 0 : u32 numOfGroups = (rankSize_ + numInGroup - 1) / numInGroup;
264 :
265 : // 本卡优先拷贝同号位数据
266 0 : CHK_RET(LocalCopy(groupId, memBlockInfo));
267 :
268 0 : for (u32 idGroup = 0; idGroup < numOfGroups; ++idGroup) {
269 : // 主流通知从流可以开始接受数据
270 0 : u32 all2allfirstSubStreamId = 0;
271 0 : CHK_RET(MainRecordSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
272 0 : CHK_RET(SubWaitMain(all2allfirstSubStreamId, all2allSubStreamNum_));
273 :
274 : // 开始数据拷贝
275 0 : u32 streamIndex = 0;
276 0 : for (u32 cnt = 0, round = idGroup * numInGroup; round < rankSize_ && cnt < numInGroup; ++round, ++cnt) {
277 0 : if (round == 0) {
278 0 : continue;
279 : }
280 0 : u32 sendRank = (localRank_ + round) % rankSize_;
281 0 : u32 recvRank = (rankSize_ + localRank_ - round) % rankSize_;
282 0 : Stream &subStream = (streamIndex == 0) ? stream_ : subStreams_[streamIndex - 1];
283 0 : CHK_SMART_PTR_NULL(links[sendRank]);
284 0 : CHK_SMART_PTR_NULL(links[recvRank]);
285 0 : CHK_RET(links[recvRank]->TxAck(subStream));
286 0 : CHK_RET(links[sendRank]->RxAck(subStream));
287 0 : streamIndex++;
288 : }
289 :
290 0 : CHK_RET(SubRecordMain(all2allfirstSubStreamId, all2allSubStreamNum_));
291 0 : CHK_RET(MainWaitSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
292 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, subStreams_[lRMainStreamId_], dispatcher_));
293 :
294 0 : CHK_RET(MainRecordSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
295 0 : CHK_RET(SubWaitMain(all2allfirstSubStreamId, all2allSubStreamNum_));
296 0 : streamIndex = 0;
297 0 : for (u32 cnt = 0, round = idGroup * numInGroup; round < rankSize_ && cnt < numInGroup; ++round, ++cnt) {
298 0 : if (round == 0) {
299 0 : continue;
300 : }
301 0 : u32 sendRank = (localRank_ + round) % rankSize_;
302 0 : u32 recvRank = (rankSize_ + localRank_ - round) % rankSize_;
303 0 : Stream &subStream = (streamIndex == 0) ? stream_ : subStreams_[streamIndex - 1];
304 0 : CHK_SMART_PTR_NULL(links[sendRank]);
305 0 : CHK_SMART_PTR_NULL(links[recvRank]);
306 :
307 0 : u64 sliceSize = memBlockInfo.size[sendRank];
308 0 : if (sliceSize != 0) {
309 0 : u64 userMemInOffset = memBlockInfo.userInputOffsets[sendRank];
310 0 : DeviceMem src = DeviceMem::create(static_cast<u8 *>(inputMemPtr_) + userMemInOffset, sliceSize);
311 0 : u32 outputIndex = CalcOutputIndex(sendRank);
312 0 : u64 dstOffset = 0;
313 0 : void *remMemPtr = nullptr;
314 0 : if (isNeedSpaceBorrow_ && isLastBlockData(outputIndex) && !(isLastRank(sendRank) && isLastGroup(groupId))) {
315 0 : CHK_RET(links[sendRank]->GetRemoteMem(scratchMemType_, &remMemPtr));
316 0 : dstOffset = memBlockInfo.outputOffsets[sendRank];
317 : } else {
318 0 : CHK_RET(links[sendRank]->GetRemoteMem(outputMemType_, &remMemPtr));
319 0 : dstOffset = memBlockInfo.outputOffsets[outputIndex];
320 : }
321 0 : DeviceMem dst = DeviceMem::create(static_cast<u8 *>(remMemPtr) + dstOffset, sliceSize);
322 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, subStream, links[sendRank]->GetRemoteRank(),
323 : links[sendRank]->GetLinkType()));
324 0 : }
325 0 : CHK_RET(links[sendRank]->TxDataSignal(subStream));
326 0 : CHK_RET(links[recvRank]->RxDataSignal(subStream));
327 0 : streamIndex++;
328 : }
329 :
330 : // 从流通知主流完成拷贝
331 0 : CHK_RET(SubRecordMain(all2allfirstSubStreamId, all2allSubStreamNum_));
332 0 : CHK_RET(MainWaitSub(stream_, all2allfirstSubStreamId, all2allSubStreamNum_));
333 : }
334 :
335 0 : return HCCL_SUCCESS;
336 : }
337 :
338 0 : HcclResult ReduceScatterPlantLocalReduce::RunLocalReduce(u32 groupId, const MemBlockInfo& memBlockInfo)
339 : {
340 0 : u32 reduceStep = static_cast<u32>(std::ceil(log2(rankSize_)));
341 0 : u64 srcOffset = memBlockInfo.inputOffsets[localRank_];
342 0 : u64 sliceSize = memBlockInfo.size[localRank_];
343 0 : u32 dataUnitSize = DataUnitSize(dataType_);
344 0 : if (dataUnitSize == 0) {
345 0 : HCCL_ERROR("[ReduceScatterPlantLocalReduce][RunLocalReduce]data type[%s] out of range[%d, %d]",
346 : GetDataTypeEnumStr(dataType_).c_str(), HCCL_DATA_TYPE_INT8, static_cast<int>(HCCL_DATA_TYPE_RESERVED) - 1);
347 0 : return HCCL_E_INTERNAL;
348 : }
349 0 : u64 count = sliceSize / dataUnitSize;
350 :
351 0 : for (u32 round = 0; round < reduceStep; round++) {
352 0 : u32 tailIndex = std::min(rankSize_, static_cast<u32>(1 << static_cast<int>(reduceStep - round))) - 1;
353 0 : u32 headIndex = static_cast<u32>(1 << static_cast<int>((reduceStep - round - 1)));
354 0 : u32 reduceSubStreamNum = std::min(tailIndex - headIndex, DEVICE_EIGHT / FACTOR_NUM_TWO - 1);
355 : // LR主流通知从流可以开始接受数据
356 0 : for (u32 offset = 0; offset < reduceSubStreamNum; offset++) {
357 0 : u32 streamId = lRMainStreamId_ + offset + 1;
358 : // 只有reduce任务 > 1时才需要主从流同步: LR主流通知从流, 从流Wait LR主流
359 0 : CHK_RET(LocalNotify::Post(subStreams_[lRMainStreamId_], dispatcher_, (*meshSignalAuxPtr_)[streamId],
360 : profilerInput_.stage));
361 0 : CHK_RET(LocalNotify::Wait(subStreams_[streamId], dispatcher_, (*meshSignalAuxPtr_)[streamId],
362 : profilerInput_.stage));
363 : }
364 :
365 : // LocalReduce操作
366 0 : for (u32 offset = 0; offset <= tailIndex - headIndex; offset++) {
367 0 : u32 inputIndex = CalcOutputIndex(headIndex + offset); // reduce的源数据offset
368 0 : u32 outputIndex = CalcOutputIndex(offset); // reduce的目标offset
369 0 : u32 streamOffset = offset % (reduceSubStreamNum + 1);
370 0 : Stream &subStream = subStreams_[lRMainStreamId_ + streamOffset];
371 0 : if (sliceSize == 0) {
372 0 : continue;
373 : }
374 : void *srcPtr;
375 : void *dstPtr;
376 0 : if (isNeedSpaceBorrow_ && !(isLastRank(localRank_) && isLastGroup(groupId)) && isLastBlockData(inputIndex)) {
377 0 : srcPtr = static_cast<u8 *>(inputMem_.ptr()) + srcOffset;
378 : } else {
379 0 : srcPtr = static_cast<u8 *>(outputMem_.ptr()) + memBlockInfo.outputOffsets[inputIndex];
380 : }
381 :
382 0 : if (isNeedSpaceBorrow_ && !(isLastRank(localRank_) && isLastGroup(groupId)) && isLastBlockData(outputIndex)) {
383 0 : dstPtr = static_cast<u8 *>(inputMem_.ptr()) + srcOffset;
384 : } else {
385 0 : dstPtr = static_cast<u8 *>(outputMem_.ptr()) + memBlockInfo.outputOffsets[outputIndex];
386 : }
387 :
388 0 : CHK_RET(HcclReduceAsync(dispatcher_, srcPtr, count, dataType_, reductionOp_, subStream, dstPtr,
389 : INVALID_VALUE_RANKID, LinkType::LINK_ONCHIP, INLINE_REDUCE_BIT));
390 : }
391 :
392 : // 从流通知LR主流可以开始下一轮
393 0 : for (u32 offset = 0; offset < reduceSubStreamNum; offset++) {
394 0 : u32 streamId = lRMainStreamId_ + offset + 1;
395 : // 只有reduce任务 > 1时才需要主从流同步: LR主流通知从流, 从流Wait LR主流
396 0 : CHK_RET(LocalNotify::Post(subStreams_[streamId], dispatcher_, (*meshSignalPtr_)[streamId],
397 : profilerInput_.stage));
398 0 : CHK_RET(LocalNotify::Wait(subStreams_[lRMainStreamId_], dispatcher_, (*meshSignalPtr_)[streamId],
399 : profilerInput_.stage));
400 : }
401 :
402 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, subStreams_[lRMainStreamId_], dispatcher_));
403 : }
404 :
405 0 : return HCCL_SUCCESS;
406 : }
407 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_REDUCESCATTER_PLANT_LOCAL_REDUCE, ReduceScatterPlantLocalReduce);
408 : } // namespace hccl
|
