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_all_reduce_nhr1d_mem2mem.h"
12 :
13 : namespace Hccl {
14 :
15 : 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
16 : constexpr uint16_t OUTPUT_XN_ID = 1;
17 : constexpr uint16_t TOKEN_XN_ID = 2;
18 : constexpr uint16_t CKE_IDX_0 = 0; // 后同步
19 : constexpr uint16_t CKE_IDX_1 = 1; // 前同步addr
20 : constexpr uint16_t CKE_IDX_2 = 2; // 前同步token
21 : constexpr uint16_t CKE_IDX_3 = 3; // NHR step同步信号0，用于RS前同步，AG后同步
22 : constexpr uint16_t CKE_IDX_4 = 4; // NHR step同步信号1，用于RS后同步
23 : constexpr uint16_t FST_AXIS_ID = 0;
24 : constexpr uint16_t SEC_AXIS_ID = 1;
25 :
26 0 : CcuContextAllReduceNHR1D::CcuContextAllReduceNHR1D(const CcuCtxArg &arg, const std::vector<CcuTransport *> &transports,
27 0 : const CcuTransportGroup &group)
28 0 : : CcuContextAlgBase(arg, transports, group)
29 : {
30 0 : const CcuCtxArgAllReduceNHR1D *ctxArg = dynamic_cast<const CcuCtxArgAllReduceNHR1D *>(&arg);
31 0 : rankId_ = ctxArg->rankId_;
32 0 : axisId_ = ctxArg->axisId_;
33 0 : axisSize_ = ctxArg->axisSize_;
34 0 : dimSize_ = ctxArg->dimSize_[0];
35 0 : stepInfoVector_ = ctxArg->stepInfoVector_;
36 0 : indexMap_ = ctxArg->indexMap_;
37 0 : localSize_ = indexMap_.size();
38 0 : myRankIdx_ = indexMap_.size();
39 0 : dataType_ = ctxArg->op_.dataType;
40 0 : reduceOp_ = ctxArg->op_.reduceOp;
41 0 : signalNum_ = (dimSize_ + RANK_NUM_PER_CKE - 1) / RANK_NUM_PER_CKE; // 每个CKE有16个bit
42 0 : HCCL_INFO("[CcuContextAllReduceNHR1D] CtxArg: rankId_[%u], axisId_[%u], axisSize_[%u], dimSize_[%u], localSize_[%u], "
43 : "signalNum_[%u], dataType[%s], reduceOp[%s]",
44 : rankId_, axisId_, axisSize_, dimSize_, localSize_, signalNum_, dataType_.Describe().c_str(),
45 : reduceOp_.Describe().c_str());
46 0 : }
47 :
48 0 : void CcuContextAllReduceNHR1D::LoadArgs()
49 : {
50 0 : Load(input_);
51 0 : Load(output_[myRankIdx_]);
52 0 : Load(token_[myRankIdx_]);
53 0 : Load(isInputOutputEqual_);
54 0 : Load(die0Size_);
55 0 : Load(die1Size_);
56 0 : Load(die0SliceSize_);
57 0 : Load(die1SliceSize_);
58 0 : Load(die0LastSliceSize_);
59 0 : Load(die1LastSliceSize_);
60 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] LoadArgs run finished");
61 0 : }
62 :
63 0 : void CcuContextAllReduceNHR1D::InitResources()
64 : {
65 0 : isInputOutputEqual_ = CreateVariable();
66 0 : die0SliceSize_ = CreateVariable();
67 0 : die1SliceSize_ = CreateVariable();
68 0 : die0LastSliceSize_ = CreateVariable();
69 0 : die1LastSliceSize_ = CreateVariable();
70 0 : die0Size_ = CreateVariable();
71 0 : die1Size_ = CreateVariable();
72 0 : localSignal_ = CreateMaskSignal();
73 :
74 0 : input_ = CreateVariable();
75 0 : for (uint32_t transportIdx = 0; transportIdx < localSize_; transportIdx++) {
76 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] MyRank[%u], TransportId[%u]", rankId_, transportIdx);
77 0 : CHK_PRT_RET(transports[transportIdx] == nullptr,
78 : HCCL_ERROR("[CcuContextAllReduceNHR1D] Algorithm transport ptr is null"), );
79 0 : output_.push_back(
80 0 : CreateVariable((*transports[transportIdx]), OUTPUT_XN_ID)); // 获取transport中id=1的Var来传递output
81 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
82 : }
83 0 : output_.push_back(CreateVariable());
84 0 : token_.push_back(CreateVariable());
85 :
86 0 : srcMem_ = CreateMemory();
87 0 : dstMem_ = CreateMemory();
88 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] InitResources finished");
89 : }
90 :
91 0 : void CcuContextAllReduceNHR1D::PreSync()
92 : {
93 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] PreSync start");
94 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
95 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
96 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
97 0 : for (auto t : transports) {
98 0 : WriteVariableWithSignal(*t, output_[localSize_], OUTPUT_XN_ID, selfSignalId + signalNum_ * CKE_IDX_1, selfBit);
99 0 : WriteVariableWithSignal(*t, token_[localSize_], TOKEN_XN_ID, selfSignalId + signalNum_ * CKE_IDX_2, selfBit);
100 : }
101 0 : for (auto &pair : indexMap_) {
102 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
103 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
104 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
105 : }
106 0 : for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) {
107 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_1, waitBitVector[sId]);
108 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_2, waitBitVector[sId]);
109 : }
110 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] PreSync end");
111 0 : }
112 :
113 0 : void CcuContextAllReduceNHR1D::PostSync()
114 : {
115 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
116 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
117 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
118 0 : for (auto &t : transports) {
119 0 : RemotePost(*t, selfSignalId + signalNum_ * CKE_IDX_0, selfBit);
120 : }
121 0 : for (auto &pair : indexMap_) {
122 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
123 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
124 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
125 : }
126 0 : for (uint32_t sId = 0; sId < signalNum_; sId++) {
127 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_0, waitBitVector[sId]);
128 : }
129 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] PostSync run finished");
130 0 : }
131 :
132 0 : void CcuContextAllReduceNHR1D::AxisSync(uint32_t signalIndex)
133 : {
134 0 : const uint32_t DIE_NUM = 2;
135 0 : if (signalIndex > 1) {
136 0 : THROW<InvalidParamsException>(
137 0 : StringFormat("[CcuContextAllReduceNHR1D] Unexpected SignalInex[%u]", signalIndex));
138 : }
139 0 : LocalCtxPost(anotherAxisSignal_, 1 << (axisId_ + signalIndex * DIE_NUM));
140 0 : LocalWait(localAxisSignal_, 1 << (1 - axisId_ + signalIndex * DIE_NUM));
141 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] AxisSync run finished");
142 0 : return;
143 : }
144 :
145 0 : void CcuContextAllReduceNHR1D::DoReduceScatterNHR()
146 : {
147 0 : const uint32_t NHR_NUM = 2;
148 0 : for (u64 i = 0; i < stepInfoVector_.size() / NHR_NUM; i++) {
149 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[i];
150 0 : DoReduceScatterNHRSingleStep(nhrStepInfo);
151 : }
152 0 : }
153 :
154 0 : void CcuContextAllReduceNHR1D::DoReduceScatterNHRSingleStep(const NHRStepInfo &nhrStepInfo)
155 : {
156 0 : u32 sendSliceIdx = 0;
157 0 : u32 &fromRankIdx = indexMap_[nhrStepInfo.fromRank];
158 0 : u32 &toRankIdx = indexMap_[nhrStepInfo.toRank];
159 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
160 0 : srcMem_.token = token_[myRankIdx_];
161 0 : dstMem_.token = token_[toRankIdx];
162 0 : CcuTransport *sendTransport = transports[toRankIdx];
163 0 : CcuTransport *recvTransport = transports[fromRankIdx];
164 :
165 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
166 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
167 0 : uint16_t sendSignalId = nhrStepInfo.toRank / RANK_NUM_PER_CKE;
168 0 : uint16_t sendBit = 1 << (nhrStepInfo.toRank % RANK_NUM_PER_CKE);
169 0 : if (nhrStepInfo.step != 0) {
170 : // 通知fromRank，可以写入
171 0 : RemotePost(*recvTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
172 :
173 : // 等待toRank通知其可以写入
174 0 : RemoteWait(*sendTransport, sendSignalId + signalNum_ * CKE_IDX_3, sendBit);
175 : }
176 :
177 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
178 0 : sendSliceIdx = sendSliceIdxList[i];
179 0 : if (i != 0) {
180 0 : if (i % RANK_NUM_PER_CKE == 0) {
181 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
182 : }
183 : }
184 0 : if (nhrStepInfo.step == 0) {
185 : // 只有第0步的源数据从input中取
186 0 : HCCL_INFO("[CcuContextAllReduceNHR1D] nhrStepInfo 0.");
187 0 : srcMem_.addr = input_;
188 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
189 : } else {
190 0 : srcMem_.addr = output_[myRankIdx_];
191 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
192 : }
193 0 : dstMem_.addr = output_[toRankIdx];
194 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
195 0 : DoWriteReduceSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
196 : }
197 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
198 :
199 : // 通知toRank数据写入完毕
200 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_4, selfBit, true);
201 : // 等待fromRank通知数据写入完毕
202 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
203 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
204 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_4, recvBit);
205 0 : HCCL_DEBUG("[DoReduceScatterNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu",
206 : rankId_, nhrStepInfo.step, nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
207 0 : }
208 :
209 0 : void CcuContextAllReduceNHR1D::DoWriteReduceSlice(const u32 &toRank, CcuRep::Memory &src, CcuRep::Memory &dst,
210 : const u32 &sendSliceIdx, u32 signalIndex)
211 : {
212 : bool islastSlice;
213 : // 添加 die1 偏移
214 0 : if (axisId_ == 1) {
215 0 : dst.addr += die0Size_;
216 0 : src.addr += die0Size_;
217 : }
218 0 : CcuTransport *sendTransport = transports[indexMap_[toRank]];
219 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
220 :
221 : // allreduce切片的最后一块slice，大小可能不一致
222 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
223 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
224 0 : CCU_IF(sliceSize != 0)
225 : {
226 0 : WriteReduce(*sendTransport, dst, src, sliceSize, dataType_, reduceOp_, localSignal_, 1 << signalIndex);
227 0 : }
228 0 : CCU_IF(sliceSize == 0)
229 : {
230 0 : LocalPost(localSignal_, 1 << signalIndex);
231 0 : }
232 0 : }
233 :
234 0 : void CcuContextAllReduceNHR1D::DoAllGatherNHR()
235 : {
236 0 : const uint32_t NHR_NUM = 2;
237 0 : for (u64 i = stepInfoVector_.size() / NHR_NUM; i < stepInfoVector_.size(); i++) {
238 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[i];
239 0 : DoAllGatherNHRSingleStep(nhrStepInfo);
240 : }
241 0 : }
242 :
243 0 : void CcuContextAllReduceNHR1D::DoAllGatherNHRSingleStep(const NHRStepInfo &nhrStepInfo)
244 : {
245 0 : u32 sendSliceIdx = 0;
246 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
247 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
248 0 : CcuTransport *sendTransport = transports[toRankIdx];
249 0 : CcuTransport *recvTransport = transports[fromRankIdx];
250 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
251 0 : srcMem_.token = token_[myRankIdx_];
252 0 : dstMem_.token = token_[toRankIdx];
253 :
254 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
255 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
256 :
257 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
258 0 : sendSliceIdx = sendSliceIdxList[i];
259 :
260 0 : if (i != 0) {
261 0 : if (i % RANK_NUM_PER_CKE == 0) {
262 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
263 : }
264 : }
265 :
266 0 : srcMem_.addr = output_[myRankIdx_];
267 0 : dstMem_.addr = output_[toRankIdx];
268 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
269 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
270 0 : DoSendRecvSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
271 : }
272 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
273 :
274 0 : if (nhrStepInfo.step + 1 != stepInfoVector_.size()) { // 最后一步不需要同步
275 : // 通知toRank，写入完毕
276 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
277 : // 等待fromRank通知写入完毕
278 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
279 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
280 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_3, recvBit);
281 : }
282 :
283 0 : HCCL_DEBUG("[DoAllGatherNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu",
284 : rankId_, nhrStepInfo.step, nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
285 0 : }
286 :
287 0 : void CcuContextAllReduceNHR1D::DoSendRecvSlice(const u32 &toRank, CcuRep::Memory &src, CcuRep::Memory &dst,
288 : const u32 &sendSliceIdx, u32 signalIndex)
289 : {
290 : bool islastSlice;
291 0 : CcuTransport *sendTransport = transports[indexMap_[toRank]];
292 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
293 : // 添加 die1 偏移
294 0 : if (axisId_ == 1) {
295 0 : dst.addr += die0Size_;
296 0 : src.addr += die0Size_;
297 : }
298 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
299 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
300 0 : CCU_IF(sliceSize != 0)
301 : {
302 0 : Write(*sendTransport, dst, src, sliceSize, localSignal_, 1 << signalIndex);
303 0 : }
304 0 : CCU_IF(sliceSize == 0)
305 : {
306 0 : LocalPost(localSignal_, 1 << signalIndex);
307 0 : }
308 0 : }
309 :
310 0 : void CcuContextAllReduceNHR1D::LocalCopySlices()
311 : {
312 0 : u32 nonTxSliceIdx = 0;
313 0 : CcuRep::Variable tmpSliceOffset = CreateVariable();
314 0 : tmpSliceOffset = 0;
315 :
316 0 : for (u64 i = 0; i < dimSize_; i++) {
317 0 : sliceOffset_.push_back(CreateVariable());
318 0 : sliceOffset_[i] = tmpSliceOffset;
319 0 : tmpSliceOffset += axisId_ == 0? die0SliceSize_: die1SliceSize_;
320 : }
321 :
322 : // 当input == output时，不需要拷贝
323 0 : CCU_IF(isInputOutputEqual_ == 0)
324 : {
325 : // 将step0中不需要写的slice，拷贝到本rank的output中
326 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[0];
327 0 : const std::vector<u32> &nonTxSliceIdxList = GetNonTxSliceIdxs(nhrStepInfo.txSliceIdxs);
328 0 : for (u32 i = 0; i < nonTxSliceIdxList.size(); i++) {
329 0 : nonTxSliceIdx = nonTxSliceIdxList[i];
330 :
331 0 : if (i != 0) {
332 0 : if (i % RANK_NUM_PER_CKE == 0) {
333 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
334 : }
335 : }
336 :
337 0 : srcMem_.addr = input_;
338 0 : dstMem_.addr = output_[myRankIdx_];
339 0 : srcMem_.addr += sliceOffset_[nonTxSliceIdx];
340 0 : dstMem_.addr += sliceOffset_[nonTxSliceIdx];
341 0 : srcMem_.token = token_[myRankIdx_];
342 0 : dstMem_.token = token_[myRankIdx_];
343 0 : DoLocalCopySlice(srcMem_, dstMem_, nonTxSliceIdx, i);
344 : }
345 0 : LocalWait(localSignal_, (1 << (nonTxSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
346 0 : }
347 0 : }
348 :
349 0 : std::vector<u32> CcuContextAllReduceNHR1D::GetNonTxSliceIdxs(const std::vector<u32> &txSliceIdxs) const
350 : {
351 0 : std::vector<bool> isTx(dimSize_, false);
352 0 : for (u32 idx : txSliceIdxs) {
353 0 : if (idx < dimSize_) {
354 0 : isTx[idx] = true;
355 : }
356 : }
357 :
358 0 : std::vector<u32> nonTxSliceIdxs;
359 0 : for (u32 idx = 0; idx < dimSize_; ++idx) {
360 0 : if (!isTx[idx]) {
361 0 : nonTxSliceIdxs.push_back(idx);
362 : }
363 : }
364 :
365 0 : return nonTxSliceIdxs;
366 0 : }
367 :
368 0 : void CcuContextAllReduceNHR1D::DoLocalCopySlice(CcuRep::Memory &src, CcuRep::Memory &dst,
369 : const u32 ©SliceIdx, u32 signalIndex)
370 : {
371 : bool islastSlice;
372 : // 添加 die1 偏移
373 0 : if (axisId_ == 1) {
374 0 : src.addr += die0Size_;
375 0 : dst.addr += die0Size_;
376 : }
377 :
378 0 : islastSlice = (copySliceIdx + 1 == dimSize_);
379 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
380 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
381 0 : CCU_IF(sliceSize == 0)
382 : {
383 0 : LocalPost(localSignal_, 1 << signalIndex);
384 0 : }
385 :
386 0 : CCU_IF(sliceSize != 0)
387 : {
388 0 : LocalCopy(dst, src, sliceSize, localSignal_, 1 << signalIndex);
389 0 : }
390 0 : }
391 :
392 0 : void CcuContextAllReduceNHR1D::Algorithm()
393 : {
394 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] AllReduceNHR1D run");
395 :
396 0 : InitResources();
397 0 : LoadArgs();
398 0 : LocalCopySlices();
399 0 : PreSync();
400 0 : DoReduceScatterNHR();
401 0 : DoAllGatherNHR();
402 0 : PostSync();
403 :
404 0 : HCCL_DEBUG("[CcuContextAllReduceNHR1D] AllReduceNHR1D end");
405 0 : return;
406 : }
407 :
408 0 : std::vector<uint64_t> CcuContextAllReduceNHR1D::GeneArgs(const CcuTaskArg &arg)
409 : {
410 0 : const CcuTaskArgAllReduceNHR1D *taskArg = dynamic_cast<const CcuTaskArgAllReduceNHR1D *>(&arg);
411 0 : if (taskArg == nullptr) {
412 0 : THROW<NullPtrException>(StringFormat("CcuContextAllReduceNHR1D::taskArg ptr is null"));
413 : }
414 : // input&output&buffer地址
415 0 : uint64_t inputAddr = taskArg->inputAddr_;
416 0 : uint64_t outputAddr = taskArg->outputAddr_;
417 0 : uint64_t token = taskArg->token_;
418 0 : uint64_t isInputOutputEqual = taskArg->isInputOutputEqual_;
419 0 : uint64_t die0Size = taskArg->die0Size_;
420 0 : uint64_t die1Size = taskArg->die1Size_;
421 0 : uint64_t die0SliceSize = taskArg->die0SliceSize_;
422 0 : uint64_t die1SliceSize = taskArg->die1SliceSize_;
423 0 : uint64_t die0LastSliceSize = taskArg->die0LastSliceSize_;
424 0 : uint64_t die1LastSliceSize = taskArg->die1LastSliceSize_;
425 :
426 0 : HCCL_INFO("[CcuContextAllReduceNHR1D] TaskArgs: inputAddr[%llu], outputAddr[%llu], "
427 : "die0Size[%llu], die1Size[%llu], die0SliceSize[%llu], die1SliceSize[%llu],"
428 : "die0LastSliceSize[%llu], die1LastSliceSize[%llu]",
429 : inputAddr, outputAddr, die0Size, die1Size, die0SliceSize, die1SliceSize,
430 : die0LastSliceSize, die1LastSliceSize);
431 :
432 : return {inputAddr, outputAddr, token,
433 : isInputOutputEqual, die0Size, die1Size,
434 : die0SliceSize, die1SliceSize, die0LastSliceSize,
435 0 : die1LastSliceSize};
436 : }
437 : } // namespace Hccl
|
