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 "ccu_context_reduce_nhr1d_mem2mem.h"
12 :
13 : namespace Hccl {
14 :
15 : constexpr uint16_t CKE_IDX_0 = 0; // 后同步
16 : constexpr uint16_t CKE_IDX_1 = 1; // 前同步addr
17 : constexpr uint16_t CKE_IDX_2 = 2; // 前同步token
18 : constexpr uint16_t CKE_IDX_3 = 3; // NHR step同步信号0，用于RS前同步，AG后同步
19 : constexpr uint16_t CKE_IDX_4 = 4; // NHR step同步信号1，用于RS后同步
20 : constexpr uint16_t FST_AXIS_ID = 0;
21 : constexpr uint16_t SEC_AXIS_ID = 1;
22 : constexpr uint16_t RANK_NUM_PER_CKE = 16; // 本rank给远端置位时应当写的CKE，16个对端一个CKE
23 : constexpr uint16_t OUTPUT_XN_ID = 1;
24 : constexpr uint16_t TOKEN_XN_ID = 2;
25 :
26 0 : CcuContextReduceNHR1DMem2mem::CcuContextReduceNHR1DMem2mem(const CcuCtxArg &arg, const std::vector<CcuTransport *> &transports,
27 0 : const CcuTransportGroup &group)
28 0 : : CcuContextAlgBase(arg, transports, group)
29 : {
30 0 : const CcuCtxArgReduceNHR1D *ctxArg = dynamic_cast<const CcuCtxArgReduceNHR1D *>(&arg);
31 0 : if (ctxArg == nullptr) {
32 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceNHR1DMem2mem::ctxArg ptr is null"));
33 : }
34 0 : rankId_ = ctxArg->rankId_;
35 0 : rootId_ = ctxArg->rootId_;
36 0 : axisId_ = ctxArg->axisId_;
37 0 : axisSize_ = ctxArg->axisSize_;
38 0 : dimSize_ = ctxArg->dimSize_[0];
39 0 : localAxisSignalName_ = "CcuContextAllReduceNHR1DDieSync_" + std::to_string(axisId_);
40 0 : anotherAxisSignalName_ = "CcuContextAllReduceNHR1DDieSync_" + std::to_string(1 - axisId_);
41 0 : stepInfoVector_ = ctxArg->stepInfoVector_;
42 0 : indexMap_ = ctxArg->indexMap_;
43 0 : localSize_ = indexMap_.size();
44 0 : myRankIdx_ = indexMap_.size();
45 0 : dataType_ = ctxArg->op_.dataType;
46 0 : reduceOp_ = ctxArg->op_.reduceOp;
47 0 : signalNum_ = (dimSize_ + RANK_NUM_PER_CKE - 1) / RANK_NUM_PER_CKE; // 每个CKE有16个bit
48 0 : HCCL_INFO("[CcuContextReduceNHR1DMem2mem] CtxArg: rankId_[%u], axisId_[%u], axisSize_[%u], dimSize_[%u], localSize_[%u], "
49 : "signalNum_[%u], dataType[%s], reduceOp[%s]",
50 : rankId_, axisId_, axisSize_, dimSize_, localSize_, signalNum_, dataType_.Describe().c_str(),
51 : reduceOp_.Describe().c_str());
52 0 : }
53 :
54 0 : void CcuContextReduceNHR1DMem2mem::LoadArgs()
55 : {
56 0 : Load(input_);
57 0 : Load(output_[myRankIdx_]);
58 0 : Load(token_[myRankIdx_]);
59 0 : Load(isInputOutputEqual_);
60 0 : Load(die0Size_);
61 0 : Load(die1Size_);
62 0 : Load(die0SliceSize_);
63 0 : Load(die1SliceSize_);
64 0 : Load(die0LastSliceSize_);
65 0 : Load(die1LastSliceSize_);
66 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] LoadArgs run finished");
67 0 : }
68 :
69 0 : void CcuContextReduceNHR1DMem2mem::InitResources()
70 : {
71 0 : die0Size_ = CreateVariable();
72 0 : die1Size_ = CreateVariable();
73 0 : die0LastSliceSize_ = CreateVariable();
74 0 : die1LastSliceSize_ = CreateVariable();
75 0 : die0SliceSize_ = CreateVariable();
76 0 : die1SliceSize_ = CreateVariable();
77 0 : localAxisSignal_ = CreateMaskSignal();
78 0 : localSignal_ = CreateMaskSignal();
79 0 : if (axisSize_ > 1) {
80 0 : ExportMaskSignal(localAxisSignal_, localAxisSignalName_);
81 0 : anotherAxisSignal_ = ImportMaskSignal(anotherAxisSignalName_);
82 : }
83 :
84 0 : input_ = CreateVariable();
85 0 : for (uint32_t transportIdx = 0; transportIdx < localSize_; transportIdx++) {
86 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] MyRank[%u], TransportId[%u]", rankId_, transportIdx);
87 0 : CHK_PRT_RET(transports[transportIdx] == nullptr,
88 : HCCL_ERROR("[CcuContextReduceNHR1DMem2mem] Algorithm transport ptr is null"), );
89 0 : output_.push_back(
90 0 : CreateVariable((*transports[transportIdx]), OUTPUT_XN_ID)); // 获取transport中id=1的Var来传递output
91 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
92 : }
93 0 : output_.push_back(CreateVariable());
94 0 : token_.push_back(CreateVariable());
95 :
96 0 : srcMem_ = CreateMemory();
97 0 : dstMem_ = CreateMemory();
98 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] InitResources finished");
99 : }
100 :
101 0 : void CcuContextReduceNHR1DMem2mem::PreSync()
102 : {
103 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] PreSync start");
104 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
105 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
106 0 : for (auto t : transports) {
107 0 : WriteVariableWithSignal(*t, output_[localSize_], OUTPUT_XN_ID, selfSignalId + signalNum_ * CKE_IDX_1, selfBit);
108 0 : WriteVariableWithSignal(*t, token_[localSize_], TOKEN_XN_ID, selfSignalId + signalNum_ * CKE_IDX_2, selfBit);
109 : }
110 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
111 0 : for (auto &pair : indexMap_) {
112 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
113 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
114 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
115 : }
116 0 : for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) {
117 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_1, waitBitVector[sId]);
118 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_2, waitBitVector[sId]);
119 : }
120 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] PreSync end");
121 0 : }
122 :
123 0 : void CcuContextReduceNHR1DMem2mem::PostSync()
124 : {
125 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
126 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
127 0 : for (auto &t : transports) {
128 0 : RemotePost(*t, selfSignalId + signalNum_ * CKE_IDX_0, selfBit);
129 : }
130 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
131 0 : for (auto &pair : indexMap_) {
132 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
133 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
134 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
135 : }
136 0 : for (uint32_t sId = 0; sId < signalNum_; sId++) {
137 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_0, waitBitVector[sId]);
138 : }
139 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] PostSync run finished");
140 0 : }
141 :
142 0 : void CcuContextReduceNHR1DMem2mem::AxisSync(uint32_t signalIndex)
143 : {
144 0 : const uint32_t DIE_NUM = 2;
145 0 : if (signalIndex > 1) {
146 0 : THROW<InvalidParamsException>(
147 0 : StringFormat("[CcuContextReduceNHR1DMem2mem] Unexpected SignalInex[%u]", signalIndex));
148 : }
149 0 : LocalCtxPost(anotherAxisSignal_, 1 << (axisId_ + signalIndex * DIE_NUM));
150 0 : LocalWait(localAxisSignal_, 1 << (1 - axisId_ + signalIndex * DIE_NUM));
151 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] AxisSync run finished");
152 0 : return;
153 : }
154 :
155 0 : void CcuContextReduceNHR1DMem2mem::DoReduceScatterNHR()
156 : {
157 0 : const uint32_t NHR_NUM = 2;
158 0 : for (u64 i = 0; i < stepInfoVector_.size() / NHR_NUM; i++) {
159 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[i];
160 0 : DoReduceScatterNHRSingleStep(nhrStepInfo);
161 : }
162 0 : }
163 :
164 0 : void CcuContextReduceNHR1DMem2mem::DoReduceScatterNHRSingleStep(const NHRStepInfo &nhrStepInfo)
165 : {
166 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
167 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
168 0 : u32 sendSliceIdx = 0;
169 0 : CcuTransport *sendTransport = transports[toRankIdx];
170 0 : CcuTransport *recvTransport = transports[fromRankIdx];
171 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
172 0 : srcMem_.token = token_[myRankIdx_];
173 0 : dstMem_.token = token_[toRankIdx];
174 0 : uint16_t sendSignalId = nhrStepInfo.toRank / RANK_NUM_PER_CKE;
175 0 : uint16_t sendBit = 1 << (nhrStepInfo.toRank % RANK_NUM_PER_CKE);
176 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
177 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
178 0 : if (nhrStepInfo.step != 0) {
179 : // 通知fromRank，可以写入
180 0 : RemotePost(*recvTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
181 :
182 : // 等待toRank通知其可以写入
183 0 : RemoteWait(*sendTransport, sendSignalId + signalNum_ * CKE_IDX_3, sendBit);
184 : }
185 :
186 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
187 0 : sendSliceIdx = sendSliceIdxList[i];
188 :
189 0 : if (i != 0) {
190 0 : if (i % RANK_NUM_PER_CKE == 0) {
191 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
192 : }
193 : }
194 :
195 0 : if (nhrStepInfo.step == 0) {
196 : // 只有第0步的源数据从input中取
197 0 : srcMem_.addr = input_;
198 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
199 : } else {
200 0 : srcMem_.addr = output_[myRankIdx_];
201 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
202 : }
203 :
204 0 : dstMem_.addr = output_[toRankIdx];
205 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
206 :
207 0 : DoWriteReduceSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
208 : }
209 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
210 : // 通知toRank数据写入完毕
211 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_4, selfBit, true);
212 : // 等待fromRank通知数据写入完毕
213 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
214 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
215 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_4, recvBit);
216 :
217 0 : HCCL_DEBUG("[DoReduceScatterNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu",
218 : rankId_, nhrStepInfo.step, nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
219 0 : }
220 :
221 0 : void CcuContextReduceNHR1DMem2mem::DoWriteReduceSlice(const u32 &toRank, CcuRep::Memory &src, CcuRep::Memory &dst,
222 : const u32 &sendSliceIdx, u32 signalIndex)
223 : {
224 0 : CcuTransport *sendTransport = transports[indexMap_[toRank]];
225 : bool islastSlice;
226 :
227 : // 添加 die1 偏移
228 0 : if (axisId_ == 1) {
229 0 : dst.addr += die0Size_;
230 0 : src.addr += die0Size_;
231 : }
232 :
233 : // allreduce切片的最后一块slice，大小可能不一致
234 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
235 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
236 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
237 :
238 0 : CCU_IF(sliceSize != 0) {
239 0 : WriteReduce(*sendTransport, dst, src, sliceSize, dataType_, reduceOp_, localSignal_, 1 << signalIndex);
240 0 : }
241 0 : CCU_IF(sliceSize == 0) {
242 0 : LocalPost(localSignal_, 1 << signalIndex);
243 0 : }
244 0 : }
245 :
246 0 : void CcuContextReduceNHR1DMem2mem::DoGatherNHR()
247 : {
248 0 : const uint32_t NHR_NUM = 2;
249 0 : for (u64 i = stepInfoVector_.size() / NHR_NUM; i < stepInfoVector_.size(); i++) {
250 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[i];
251 0 : DoGatherNHRSingleStep(nhrStepInfo);
252 : }
253 0 : }
254 :
255 0 : void CcuContextReduceNHR1DMem2mem::DoGatherNHRSingleStep(const NHRStepInfo &nhrStepInfo)
256 : {
257 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
258 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
259 0 : CcuTransport *recvTransport = transports[fromRankIdx];
260 0 : CcuTransport *sendTransport = transports[toRankIdx];
261 0 : u32 sendSliceIdx = 0;
262 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
263 0 : srcMem_.token = token_[myRankIdx_];
264 0 : dstMem_.token = token_[toRankIdx];
265 :
266 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
267 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
268 :
269 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
270 0 : sendSliceIdx = sendSliceIdxList[i];
271 :
272 0 : if (i != 0) {
273 0 : if (i % RANK_NUM_PER_CKE == 0) {
274 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
275 : }
276 : }
277 :
278 0 : srcMem_.addr = output_[myRankIdx_];
279 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
280 :
281 0 : dstMem_.addr = output_[toRankIdx];
282 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
283 0 : DoSendRecvSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
284 : }
285 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
286 :
287 0 : if (nhrStepInfo.step + 1 != stepInfoVector_.size()) { // 最后一步不需要同步
288 : // 通知toRank，写入完毕
289 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
290 : // 等待fromRank通知写入完毕
291 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
292 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
293 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_3, recvBit);
294 : }
295 :
296 0 : HCCL_DEBUG("[DoAllGatherNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu",
297 : rankId_, nhrStepInfo.step, nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
298 0 : }
299 :
300 0 : void CcuContextReduceNHR1DMem2mem::DoSendRecvSlice(const u32 &toRank, CcuRep::Memory &src, CcuRep::Memory &dst,
301 : const u32 &sendSliceIdx, u32 signalIndex)
302 : {
303 0 : CcuTransport *sendTransport = transports[indexMap_[toRank]];
304 : bool islastSlice;
305 :
306 : // 添加 die1 偏移
307 0 : if (axisId_ == 1) {
308 0 : src.addr += die0Size_;
309 0 : dst.addr += die0Size_;
310 : }
311 :
312 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
313 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
314 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
315 :
316 0 : CCU_IF(sliceSize != 0) {
317 0 : Write(*sendTransport, dst, src, sliceSize, localSignal_, 1 << signalIndex);
318 0 : }
319 0 : CCU_IF(sliceSize == 0) {
320 0 : LocalPost(localSignal_, 1 << signalIndex);
321 0 : }
322 0 : }
323 :
324 0 : void CcuContextReduceNHR1DMem2mem::LocalCopySlices()
325 : {
326 0 : u32 nonTxSliceIdx = 0;
327 0 : CcuRep::Variable tmpSliceOffset = CreateVariable();
328 0 : tmpSliceOffset = 0;
329 :
330 0 : for (u64 i = 0; i < dimSize_; i++) {
331 0 : sliceOffset_.push_back(CreateVariable());
332 0 : sliceOffset_[i] = tmpSliceOffset;
333 0 : tmpSliceOffset += axisId_ == 0? die0SliceSize_: die1SliceSize_;
334 : }
335 :
336 : // 当input == output时，不需要拷贝
337 0 : CCU_IF(isInputOutputEqual_ == 0)
338 : {
339 : // 将step0中不需要写的slice，拷贝到本rank的output中
340 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[0];
341 0 : const std::vector<u32> &nonTxSliceIdxList = GetNonTxSliceIdxs(nhrStepInfo.txSliceIdxs);
342 0 : for (u32 i = 0; i < nonTxSliceIdxList.size(); i++) {
343 0 : nonTxSliceIdx = nonTxSliceIdxList[i];
344 :
345 0 : if (i != 0) {
346 0 : if (i % RANK_NUM_PER_CKE == 0) {
347 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
348 : }
349 : }
350 :
351 0 : srcMem_.addr = input_;
352 0 : srcMem_.addr += sliceOffset_[nonTxSliceIdx];
353 0 : srcMem_.token = token_[myRankIdx_];
354 :
355 0 : dstMem_.addr = output_[myRankIdx_];
356 0 : dstMem_.addr += sliceOffset_[nonTxSliceIdx];
357 0 : dstMem_.token = token_[myRankIdx_];
358 0 : DoLocalCopySlice(srcMem_, dstMem_, nonTxSliceIdx, i);
359 : }
360 0 : LocalWait(localSignal_, (1 << (nonTxSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
361 0 : }
362 0 : }
363 :
364 0 : std::vector<u32> CcuContextReduceNHR1DMem2mem::GetNonTxSliceIdxs(const std::vector<u32> &txSliceIdxs) const
365 : {
366 0 : std::vector<bool> isTx(dimSize_, false);
367 0 : for (u32 idx : txSliceIdxs) {
368 0 : if (idx < dimSize_) {
369 0 : isTx[idx] = true;
370 : }
371 : }
372 :
373 0 : std::vector<u32> nonTxSliceIdxs;
374 0 : for (u32 idx = 0; idx < dimSize_; ++idx) {
375 0 : if (!isTx[idx]) {
376 0 : nonTxSliceIdxs.push_back(idx);
377 : }
378 : }
379 :
380 0 : return nonTxSliceIdxs;
381 0 : }
382 :
383 0 : void CcuContextReduceNHR1DMem2mem::DoLocalCopySlice(CcuRep::Memory &src, CcuRep::Memory &dst,
384 : const u32 ©SliceIdx, u32 signalIndex)
385 : {
386 : bool islastSlice;
387 : // 添加 die1 偏移
388 0 : if (axisId_ == 1) {
389 0 : dst.addr += die0Size_;
390 0 : src.addr += die0Size_;
391 : }
392 :
393 0 : islastSlice = (copySliceIdx + 1 == dimSize_);
394 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
395 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
396 :
397 0 : CCU_IF(sliceSize != 0) {
398 0 : LocalCopy(dst, src, sliceSize, localSignal_, 1 << signalIndex);
399 0 : }
400 0 : CCU_IF(sliceSize == 0) {
401 0 : LocalPost(localSignal_, 1 << signalIndex);
402 0 : }
403 0 : }
404 :
405 0 : void CcuContextReduceNHR1DMem2mem::Algorithm()
406 : {
407 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] AllReduceNHR1D run");
408 :
409 0 : InitResources();
410 0 : LoadArgs();
411 0 : if (axisSize_ > 1) {
412 0 : AxisSync(FST_AXIS_ID);
413 : }
414 0 : LocalCopySlices();
415 0 : PreSync();
416 0 : DoReduceScatterNHR();
417 0 : DoGatherNHR();
418 0 : PostSync();
419 0 : if (axisSize_ > 1) {
420 0 : AxisSync(SEC_AXIS_ID);
421 : }
422 :
423 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] AllReduceNHR1D end");
424 0 : return;
425 : }
426 :
427 0 : std::vector<uint64_t> CcuContextReduceNHR1DMem2mem::GeneArgs(const CcuTaskArg &arg)
428 : {
429 0 : const CcuTaskArgReduceNHR1D *taskArg = dynamic_cast<const CcuTaskArgReduceNHR1D *>(&arg);
430 0 : if (taskArg == nullptr) {
431 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceNHR1DMem2mem::taskArg ptr is null"));
432 : }
433 : // input&output&buffer地址
434 0 : uint64_t inputAddr = taskArg->inputAddr_;
435 0 : uint64_t outputAddr = taskArg->outputAddr_;
436 0 : uint64_t token = taskArg->token_;
437 0 : uint64_t isInputOutputEqual = taskArg->isInputOutputEqual_;
438 0 : uint64_t die0Size = taskArg->die0Size_;
439 0 : uint64_t die1Size = taskArg->die1Size_;
440 0 : uint64_t die0SliceSize = taskArg->die0SliceSize_;
441 0 : uint64_t die1SliceSize = taskArg->die1SliceSize_;
442 0 : uint64_t die0LastSliceSize = taskArg->die0LastSliceSize_;
443 0 : uint64_t die1LastSliceSize = taskArg->die1LastSliceSize_;
444 :
445 0 : HCCL_INFO("[CcuContextReduceNHR1DMem2mem] TaskArgs: inputAddr[%llu], outputAddr[%llu], "
446 : "die0Size[%llu], die1Size[%llu], die0SliceSize[%llu], die1SliceSize[%llu],"
447 : "die0LastSliceSize[%llu], die1LastSliceSize[%llu], isInputOutputEqual is [%lu]",
448 : inputAddr, outputAddr, die0Size, die1Size, die0SliceSize, die1SliceSize,
449 : die0LastSliceSize, die1LastSliceSize, isInputOutputEqual);
450 :
451 : return {inputAddr, outputAddr, token,
452 : isInputOutputEqual, die0Size, die1Size,
453 : die0SliceSize, die1SliceSize, die0LastSliceSize,
454 0 : die1LastSliceSize};
455 : }
456 : } // namespace Hccl
|
