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(
27 0 : const CcuCtxArg& arg, const std::vector<CcuTransport*>& transports, 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(
49 : "[CcuContextReduceNHR1DMem2mem] CtxArg: rankId_[%u], axisId_[%u], axisSize_[%u], dimSize_[%u], localSize_[%u], "
50 : "signalNum_[%u], dataType[%s], reduceOp[%s]",
51 : rankId_, axisId_, axisSize_, dimSize_, localSize_, signalNum_, dataType_.Describe().c_str(),
52 : reduceOp_.Describe().c_str());
53 0 : }
54 :
55 0 : void CcuContextReduceNHR1DMem2mem::LoadArgs()
56 : {
57 0 : Load(input_);
58 0 : Load(output_[myRankIdx_]);
59 0 : Load(token_[myRankIdx_]);
60 0 : Load(isInputOutputEqual_);
61 0 : Load(die0Size_);
62 0 : Load(die1Size_);
63 0 : Load(die0SliceSize_);
64 0 : Load(die1SliceSize_);
65 0 : Load(die0LastSliceSize_);
66 0 : Load(die1LastSliceSize_);
67 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] LoadArgs run finished");
68 0 : }
69 :
70 0 : void CcuContextReduceNHR1DMem2mem::InitResources()
71 : {
72 0 : die0Size_ = CreateVariable();
73 0 : die1Size_ = CreateVariable();
74 0 : die0LastSliceSize_ = CreateVariable();
75 0 : die1LastSliceSize_ = CreateVariable();
76 0 : die0SliceSize_ = CreateVariable();
77 0 : die1SliceSize_ = CreateVariable();
78 0 : localAxisSignal_ = CreateMaskSignal();
79 0 : localSignal_ = CreateMaskSignal();
80 0 : if (axisSize_ > 1) {
81 0 : ExportMaskSignal(localAxisSignal_, localAxisSignalName_);
82 0 : anotherAxisSignal_ = ImportMaskSignal(anotherAxisSignalName_);
83 : }
84 :
85 0 : input_ = CreateVariable();
86 0 : isInputOutputEqual_ = CreateVariable();
87 0 : for (uint32_t transportIdx = 0; transportIdx < localSize_; transportIdx++) {
88 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] MyRank[%u], TransportId[%u]", rankId_, transportIdx);
89 0 : CHK_PRT_RET(
90 : transports[transportIdx] == nullptr,
91 : HCCL_ERROR("[CcuContextReduceNHR1DMem2mem] Algorithm transport ptr is null"), );
92 0 : output_.push_back(
93 0 : CreateVariable((*transports[transportIdx]), OUTPUT_XN_ID)); // 获取transport中id=1的Var来传递output
94 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
95 : }
96 0 : output_.push_back(CreateVariable());
97 0 : token_.push_back(CreateVariable());
98 :
99 0 : srcMem_ = CreateMemory();
100 0 : dstMem_ = CreateMemory();
101 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] InitResources finished");
102 : }
103 :
104 0 : void CcuContextReduceNHR1DMem2mem::PreSync()
105 : {
106 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] PreSync start");
107 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
108 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
109 0 : for (auto t : transports) {
110 0 : WriteVariableWithSignal(*t, output_[localSize_], OUTPUT_XN_ID, selfSignalId + signalNum_ * CKE_IDX_1, selfBit);
111 0 : WriteVariableWithSignal(*t, token_[localSize_], TOKEN_XN_ID, selfSignalId + signalNum_ * CKE_IDX_2, selfBit);
112 : }
113 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
114 0 : for (auto& pair : indexMap_) {
115 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
116 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
117 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
118 : }
119 0 : for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) {
120 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_1, waitBitVector[sId]);
121 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_2, waitBitVector[sId]);
122 : }
123 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] PreSync end");
124 0 : }
125 :
126 0 : void CcuContextReduceNHR1DMem2mem::PostSync()
127 : {
128 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
129 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
130 0 : for (auto& t : transports) {
131 0 : RemotePost(*t, selfSignalId + signalNum_ * CKE_IDX_0, selfBit);
132 : }
133 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
134 0 : for (auto& pair : indexMap_) {
135 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
136 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
137 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
138 : }
139 0 : for (uint32_t sId = 0; sId < signalNum_; sId++) {
140 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_0, waitBitVector[sId]);
141 : }
142 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] PostSync run finished");
143 0 : }
144 :
145 0 : void CcuContextReduceNHR1DMem2mem::AxisSync(uint32_t signalIndex)
146 : {
147 0 : const uint32_t DIE_NUM = 2;
148 0 : if (signalIndex > 1) {
149 0 : THROW<InvalidParamsException>(
150 0 : StringFormat("[CcuContextReduceNHR1DMem2mem] Unexpected SignalInex[%u]", signalIndex));
151 : }
152 0 : LocalCtxPost(anotherAxisSignal_, 1 << (axisId_ + signalIndex * DIE_NUM));
153 0 : LocalWait(localAxisSignal_, 1 << (1 - axisId_ + signalIndex * DIE_NUM));
154 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] AxisSync run finished");
155 0 : return;
156 : }
157 :
158 0 : void CcuContextReduceNHR1DMem2mem::DoReduceScatterNHR()
159 : {
160 0 : const uint32_t NHR_NUM = 2;
161 0 : for (u64 i = 0; i < stepInfoVector_.size() / NHR_NUM; i++) {
162 0 : const NHRStepInfo& nhrStepInfo = stepInfoVector_[i];
163 0 : DoReduceScatterNHRSingleStep(nhrStepInfo);
164 : }
165 0 : }
166 :
167 0 : void CcuContextReduceNHR1DMem2mem::DoReduceScatterNHRSingleStep(const NHRStepInfo& nhrStepInfo)
168 : {
169 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
170 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
171 0 : u32 sendSliceIdx = 0;
172 0 : CcuTransport* sendTransport = transports[toRankIdx];
173 0 : CcuTransport* recvTransport = transports[fromRankIdx];
174 0 : const std::vector<u32>& sendSliceIdxList = nhrStepInfo.txSliceIdxs;
175 0 : srcMem_.token = token_[myRankIdx_];
176 0 : dstMem_.token = token_[toRankIdx];
177 0 : uint16_t sendSignalId = nhrStepInfo.toRank / RANK_NUM_PER_CKE;
178 0 : uint16_t sendBit = 1 << (nhrStepInfo.toRank % RANK_NUM_PER_CKE);
179 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
180 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
181 0 : if (nhrStepInfo.step != 0) {
182 : // 通知fromRank，可以写入
183 0 : RemotePost(*recvTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
184 :
185 : // 等待toRank通知其可以写入
186 0 : RemoteWait(*sendTransport, sendSignalId + signalNum_ * CKE_IDX_3, sendBit);
187 : }
188 :
189 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
190 0 : sendSliceIdx = sendSliceIdxList[i];
191 :
192 0 : if (i != 0) {
193 0 : if (i % RANK_NUM_PER_CKE == 0) {
194 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
195 : }
196 : }
197 :
198 0 : if (nhrStepInfo.step == 0) {
199 : // 只有第0步的源数据从input中取
200 0 : srcMem_.addr = input_;
201 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
202 : } else {
203 0 : srcMem_.addr = output_[myRankIdx_];
204 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
205 : }
206 :
207 0 : dstMem_.addr = output_[toRankIdx];
208 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
209 :
210 0 : DoWriteReduceSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
211 : }
212 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
213 : // 通知toRank数据写入完毕
214 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_4, selfBit, true);
215 : // 等待fromRank通知数据写入完毕
216 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
217 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
218 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_4, recvBit);
219 :
220 0 : HCCL_DEBUG(
221 : "[DoReduceScatterNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu", rankId_, nhrStepInfo.step,
222 : nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
223 0 : }
224 :
225 0 : void CcuContextReduceNHR1DMem2mem::DoWriteReduceSlice(
226 : const u32& toRank, CcuRep::Memory& src, CcuRep::Memory& dst, const u32& sendSliceIdx, u32 signalIndex)
227 : {
228 0 : CcuTransport* sendTransport = transports[indexMap_[toRank]];
229 : bool islastSlice;
230 :
231 : // 添加 die1 偏移
232 0 : if (axisId_ == 1) {
233 0 : dst.addr += die0Size_;
234 0 : src.addr += die0Size_;
235 : }
236 :
237 : // allreduce切片的最后一块slice，大小可能不一致
238 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
239 0 : const CcuRep::Variable& sliceSize = axisId_ == 0 ? (islastSlice ? die0LastSliceSize_ : die0SliceSize_) :
240 : (islastSlice ? die1LastSliceSize_ : die1SliceSize_);
241 :
242 0 : CCU_IF(sliceSize != 0)
243 : {
244 0 : WriteReduce(*sendTransport, dst, src, sliceSize, dataType_, reduceOp_, localSignal_, 1 << signalIndex);
245 0 : }
246 0 : CCU_IF(sliceSize == 0) { LocalPost(localSignal_, 1 << signalIndex); }
247 0 : }
248 :
249 0 : void CcuContextReduceNHR1DMem2mem::DoGatherNHR()
250 : {
251 0 : const uint32_t NHR_NUM = 2;
252 0 : for (u64 i = stepInfoVector_.size() / NHR_NUM; i < stepInfoVector_.size(); i++) {
253 0 : const NHRStepInfo& nhrStepInfo = stepInfoVector_[i];
254 0 : DoGatherNHRSingleStep(nhrStepInfo);
255 : }
256 0 : }
257 :
258 0 : void CcuContextReduceNHR1DMem2mem::DoGatherNHRSingleStep(const NHRStepInfo& nhrStepInfo)
259 : {
260 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
261 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
262 0 : CcuTransport* recvTransport = transports[fromRankIdx];
263 0 : CcuTransport* sendTransport = transports[toRankIdx];
264 0 : u32 sendSliceIdx = 0;
265 0 : const std::vector<u32>& sendSliceIdxList = nhrStepInfo.txSliceIdxs;
266 0 : srcMem_.token = token_[myRankIdx_];
267 0 : dstMem_.token = token_[toRankIdx];
268 :
269 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
270 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
271 :
272 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
273 0 : sendSliceIdx = sendSliceIdxList[i];
274 :
275 0 : if (i != 0) {
276 0 : if (i % RANK_NUM_PER_CKE == 0) {
277 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
278 : }
279 : }
280 :
281 0 : srcMem_.addr = output_[myRankIdx_];
282 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
283 :
284 0 : dstMem_.addr = output_[toRankIdx];
285 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
286 0 : DoSendRecvSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
287 : }
288 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
289 :
290 0 : if (nhrStepInfo.step + 1 != stepInfoVector_.size()) { // 最后一步不需要同步
291 : // 通知toRank，写入完毕
292 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
293 : // 等待fromRank通知写入完毕
294 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
295 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
296 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_3, recvBit);
297 : }
298 :
299 0 : HCCL_DEBUG(
300 : "[DoAllGatherNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu", rankId_, nhrStepInfo.step,
301 : nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
302 0 : }
303 :
304 0 : void CcuContextReduceNHR1DMem2mem::DoSendRecvSlice(
305 : const u32& toRank, CcuRep::Memory& src, CcuRep::Memory& dst, const u32& sendSliceIdx, u32 signalIndex)
306 : {
307 0 : CcuTransport* sendTransport = transports[indexMap_[toRank]];
308 : bool islastSlice;
309 :
310 : // 添加 die1 偏移
311 0 : if (axisId_ == 1) {
312 0 : src.addr += die0Size_;
313 0 : dst.addr += die0Size_;
314 : }
315 :
316 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
317 0 : const CcuRep::Variable& sliceSize = axisId_ == 0 ? (islastSlice ? die0LastSliceSize_ : die0SliceSize_) :
318 : (islastSlice ? die1LastSliceSize_ : die1SliceSize_);
319 :
320 0 : CCU_IF(sliceSize != 0) { Write(*sendTransport, dst, src, sliceSize, localSignal_, 1 << signalIndex); }
321 0 : CCU_IF(sliceSize == 0) { LocalPost(localSignal_, 1 << signalIndex); }
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(
384 : CcuRep::Memory& src, CcuRep::Memory& dst, const u32& copySliceIdx, 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) { LocalCopy(dst, src, sliceSize, localSignal_, 1 << signalIndex); }
398 0 : CCU_IF(sliceSize == 0) { LocalPost(localSignal_, 1 << signalIndex); }
399 0 : }
400 :
401 0 : void CcuContextReduceNHR1DMem2mem::Algorithm()
402 : {
403 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] AllReduceNHR1D run");
404 :
405 0 : InitResources();
406 0 : LoadArgs();
407 0 : if (axisSize_ > 1) {
408 0 : AxisSync(FST_AXIS_ID);
409 : }
410 0 : LocalCopySlices();
411 0 : PreSync();
412 0 : DoReduceScatterNHR();
413 0 : DoGatherNHR();
414 0 : PostSync();
415 0 : if (axisSize_ > 1) {
416 0 : AxisSync(SEC_AXIS_ID);
417 : }
418 :
419 0 : HCCL_DEBUG("[CcuContextReduceNHR1DMem2mem] AllReduceNHR1D end");
420 0 : return;
421 : }
422 :
423 0 : std::vector<uint64_t> CcuContextReduceNHR1DMem2mem::GeneArgs(const CcuTaskArg& arg)
424 : {
425 0 : const CcuTaskArgReduceNHR1D* taskArg = dynamic_cast<const CcuTaskArgReduceNHR1D*>(&arg);
426 0 : if (taskArg == nullptr) {
427 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceNHR1DMem2mem::taskArg ptr is null"));
428 : }
429 : // input&output&buffer地址
430 0 : uint64_t inputAddr = taskArg->inputAddr_;
431 0 : uint64_t outputAddr = taskArg->outputAddr_;
432 0 : uint64_t token = taskArg->token_;
433 0 : uint64_t isInputOutputEqual = taskArg->isInputOutputEqual_;
434 0 : uint64_t die0Size = taskArg->die0Size_;
435 0 : uint64_t die1Size = taskArg->die1Size_;
436 0 : uint64_t die0SliceSize = taskArg->die0SliceSize_;
437 0 : uint64_t die1SliceSize = taskArg->die1SliceSize_;
438 0 : uint64_t die0LastSliceSize = taskArg->die0LastSliceSize_;
439 0 : uint64_t die1LastSliceSize = taskArg->die1LastSliceSize_;
440 :
441 0 : HCCL_INFO(
442 : "[CcuContextReduceNHR1DMem2mem] TaskArgs: inputAddr[%llu], outputAddr[%llu], "
443 : "die0Size[%llu], die1Size[%llu], die0SliceSize[%llu], die1SliceSize[%llu],"
444 : "die0LastSliceSize[%llu], die1LastSliceSize[%llu], isInputOutputEqual is [%lu]",
445 : inputAddr, outputAddr, die0Size, die1Size, die0SliceSize, die1SliceSize, die0LastSliceSize, die1LastSliceSize,
446 : isInputOutputEqual);
447 :
448 : return {inputAddr, outputAddr, token, isInputOutputEqual, die0Size,
449 0 : die1Size, die0SliceSize, die1SliceSize, die0LastSliceSize, die1LastSliceSize};
450 : }
451 : } // namespace Hccl
|
