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_broadcast_nhr1d_mem2mem.h"
12 :
13 : namespace Hccl {
14 :
15 : constexpr uint16_t TOKEN_XN_ID = 2;
16 : constexpr uint16_t OUTPUT_XN_ID = 1;
17 : 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
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 : CcuContextBroadcastNHRMem2Mem1D::CcuContextBroadcastNHRMem2Mem1D(const CcuCtxArg &arg, const std::vector<CcuTransport *> &transports,
27 0 : const CcuTransportGroup &group)
28 0 : : CcuContextAlgBase(arg, transports, group)
29 : {
30 0 : const CcuCtxArgBroadcastNHRMem2Mem1D *ctxArg = dynamic_cast<const CcuCtxArgBroadcastNHRMem2Mem1D *>(&arg);
31 0 : rankId_ = ctxArg->rankId_;
32 0 : axisId_ = ctxArg->axisId_;
33 0 : axisSize_ = ctxArg->axisSize_;
34 0 : dimSize_ = ctxArg->dimSize_[0];
35 0 : localAxisSignalName_ = "CcuContextBroadcastNHR1DDieSync_" + std::to_string(axisId_);
36 0 : anotherAxisSignalName_ = "CcuContextBroadcastNHR1DDieSync_" + std::to_string(1 - axisId_);
37 0 : stepInfoVector_ = ctxArg->stepInfoVector_;
38 0 : indexMap_ = ctxArg->indexMap_;
39 0 : localSize_ = indexMap_.size();
40 0 : myRankIdx_ = indexMap_.size();
41 0 : dataType_ = ctxArg->op_.dataType;
42 0 : signalNum_ = (dimSize_ + RANK_NUM_PER_CKE - 1) / RANK_NUM_PER_CKE; // 每个CKE有16个bit
43 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] CtxArg: rankId_[%u], axisId_[%u], axisSize_[%u], dimSize_[%u], localSize_[%u], "
44 : "signalNum_[%u], dataType[%s]",
45 : rankId_, axisId_, axisSize_, dimSize_, localSize_, signalNum_, dataType_.Describe().c_str());
46 0 : }
47 :
48 0 : void CcuContextBroadcastNHRMem2Mem1D::InitResources()
49 : {
50 0 : die0Size_ = CreateVariable();
51 0 : die1Size_ = CreateVariable();
52 0 : die0SliceSize_ = CreateVariable();
53 0 : die1SliceSize_ = CreateVariable();
54 0 : die0LastSliceSize_ = CreateVariable();
55 0 : die1LastSliceSize_ = CreateVariable();
56 0 : localAxisSignal_ = CreateMaskSignal();
57 0 : localSignal_ = CreateMaskSignal();
58 :
59 0 : if (axisSize_ > 1) {
60 0 : anotherAxisSignal_ = CreateMaskSignal();
61 :
62 0 : ExportMaskSignal(localAxisSignal_, localAxisSignalName_);
63 0 : anotherAxisSignal_ = ImportMaskSignal(anotherAxisSignalName_);
64 : }
65 :
66 0 : input_ = CreateVariable();
67 0 : for (uint32_t transportIdx = 0; transportIdx < localSize_; transportIdx++) {
68 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] MyRank[%u], TransportId[%u]", rankId_, transportIdx);
69 0 : CHK_PRT_RET(transports[transportIdx] == nullptr,
70 : HCCL_ERROR("[CcuContextBroadcastNHRMem2Mem1D] Algorithm transport ptr is null"), );
71 0 : output_.push_back(
72 0 : CreateVariable((*transports[transportIdx]), OUTPUT_XN_ID)); // 获取transport中id=1的Var来传递output
73 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
74 : }
75 0 : output_.push_back(CreateVariable());
76 0 : token_.push_back(CreateVariable());
77 :
78 0 : srcMem_ = CreateMemory();
79 0 : dstMem_ = CreateMemory();
80 :
81 0 : CcuRep::Variable tmpSliceOffset = CreateVariable();
82 0 : tmpSliceOffset = 0;
83 :
84 0 : for (u64 i = 0; i < dimSize_; i++) {
85 0 : sliceOffset_.push_back(CreateVariable());
86 0 : sliceOffset_[i] = tmpSliceOffset;
87 0 : tmpSliceOffset += axisId_ == 0? die0SliceSize_: die1SliceSize_;
88 : }
89 :
90 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] InitResources finished");
91 0 : }
92 :
93 0 : void CcuContextBroadcastNHRMem2Mem1D::LoadArgs()
94 : {
95 0 : Load(input_);
96 0 : Load(output_[myRankIdx_]);
97 0 : Load(token_[myRankIdx_]);
98 0 : Load(die0Size_);
99 0 : Load(die1Size_);
100 0 : Load(die0SliceSize_);
101 0 : Load(die1SliceSize_);
102 0 : Load(die0LastSliceSize_);
103 0 : Load(die1LastSliceSize_);
104 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] LoadArgs run finished");
105 0 : }
106 :
107 0 : void CcuContextBroadcastNHRMem2Mem1D::AxisSync(uint32_t signalIndex)
108 : {
109 0 : const uint32_t DIE_NUM = 2;
110 0 : if (signalIndex > 1) {
111 0 : THROW<InvalidParamsException>(
112 0 : StringFormat("[CcuContextBroadcastNHRMem2Mem1D] Unexpected SignalInex[%u]", signalIndex));
113 : }
114 0 : LocalCtxPost(anotherAxisSignal_, 1 << (axisId_ + signalIndex * DIE_NUM));
115 0 : LocalWait(localAxisSignal_, 1 << (1 - axisId_ + signalIndex * DIE_NUM));
116 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] AxisSync run finished");
117 0 : return;
118 : }
119 :
120 0 : void CcuContextBroadcastNHRMem2Mem1D::PreSync()
121 : {
122 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] PreSync start");
123 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
124 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
125 0 : for (auto t : transports) {
126 0 : WriteVariableWithSignal(*t, output_[localSize_], OUTPUT_XN_ID, selfSignalId + signalNum_ * CKE_IDX_1, selfBit);
127 0 : WriteVariableWithSignal(*t, token_[localSize_], TOKEN_XN_ID, selfSignalId + signalNum_ * CKE_IDX_2, selfBit);
128 : }
129 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
130 0 : for (auto &pair : indexMap_) {
131 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
132 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
133 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
134 : }
135 0 : for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) {
136 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_1, waitBitVector[sId]);
137 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_2, waitBitVector[sId]);
138 : }
139 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] PreSync end");
140 0 : }
141 :
142 0 : void CcuContextBroadcastNHRMem2Mem1D::PostSync()
143 : {
144 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
145 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
146 0 : for (auto &t : transports) {
147 0 : RemotePost(*t, selfSignalId + signalNum_ * CKE_IDX_0, selfBit);
148 : }
149 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
150 0 : for (auto &pair : indexMap_) {
151 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
152 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
153 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
154 : }
155 0 : for (uint32_t sId = 0; sId < signalNum_; sId++) {
156 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_0, waitBitVector[sId]);
157 : }
158 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] PostSync run finished");
159 0 : }
160 :
161 0 : void CcuContextBroadcastNHRMem2Mem1D::DoScatterNHR()
162 : {
163 0 : const uint32_t NHR_NUM = 2;
164 0 : for (u64 i = 0; i < stepInfoVector_.size() / NHR_NUM; i++) {
165 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[i];
166 0 : DoScatterNHRSingleStep(nhrStepInfo);
167 : }
168 0 : }
169 :
170 0 : void CcuContextBroadcastNHRMem2Mem1D::DoScatterNHRSingleStep(const NHRStepInfo &nhrStepInfo)
171 : {
172 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
173 0 : const std::vector<u32> &recvSliceIdxList = nhrStepInfo.rxSliceIdxs;
174 : // 只需要发
175 0 : if(sendSliceIdxList.size() != 0){
176 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
177 0 : u32 sendSliceIdx = 0;
178 0 : CcuTransport *sendTransport = transports[toRankIdx];
179 0 : srcMem_.token = token_[myRankIdx_];
180 0 : dstMem_.token = token_[toRankIdx];
181 :
182 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
183 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
184 :
185 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
186 0 : sendSliceIdx = sendSliceIdxList[i];
187 :
188 0 : if (i != 0) {
189 0 : if (i % RANK_NUM_PER_CKE == 0) {
190 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
191 : }
192 : }
193 :
194 0 : if (nhrStepInfo.step == 0) {
195 : // 只有第0步的源数据从input中取
196 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] nhrStepInfo 0.");
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 : DoSendRecvSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
208 : }
209 :
210 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
211 :
212 : // 通知toRank数据写入完毕
213 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_4, selfBit, true);
214 : }
215 :
216 : //只需要收
217 0 : if(recvSliceIdxList.size() != 0){
218 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
219 0 : CcuTransport *recvTransport = transports[fromRankIdx];
220 :
221 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
222 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
223 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_4, recvBit);
224 : }
225 0 : }
226 :
227 0 : void CcuContextBroadcastNHRMem2Mem1D::DoSendRecvSlice(const u32 &toRank, CcuRep::Memory &src, CcuRep::Memory &dst,
228 : const u32 &sendSliceIdx, u32 signalIndex)
229 : {
230 0 : CcuTransport *sendTransport = transports[indexMap_[toRank]];
231 : bool islastSlice;
232 :
233 : // 添加 die1 偏移
234 0 : if (axisId_ == 1) {
235 0 : src.addr += die0Size_;
236 0 : dst.addr += die0Size_;
237 : }
238 :
239 0 : islastSlice = (sendSliceIdx + 1 == dimSize_);
240 0 : const CcuRep::Variable &sliceSize = axisId_ == 0? (islastSlice? die0LastSliceSize_ : die0SliceSize_)
241 : : (islastSlice? die1LastSliceSize_ : die1SliceSize_);
242 :
243 0 : CCU_IF(sliceSize == 0)
244 : {
245 0 : LocalPost(localSignal_, 1 << signalIndex);
246 0 : }
247 :
248 0 : CCU_IF(sliceSize != 0)
249 : {
250 0 : Write(*sendTransport, dst, src, sliceSize, localSignal_, 1 << signalIndex);
251 0 : }
252 0 : }
253 :
254 0 : void CcuContextBroadcastNHRMem2Mem1D::DoAllGatherNHR()
255 : {
256 0 : const uint32_t NHR_NUM = 2;
257 0 : for (u64 i = stepInfoVector_.size() / NHR_NUM; i < stepInfoVector_.size(); i++) {
258 0 : const NHRStepInfo &nhrStepInfo = stepInfoVector_[i];
259 0 : DoAllGatherNHRSingleStep(nhrStepInfo);
260 : }
261 0 : }
262 :
263 0 : void CcuContextBroadcastNHRMem2Mem1D::DoAllGatherNHRSingleStep(const NHRStepInfo &nhrStepInfo)
264 : {
265 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
266 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
267 0 : u32 sendSliceIdx = 0;
268 0 : CcuTransport *sendTransport = transports[toRankIdx];
269 0 : CcuTransport *recvTransport = transports[fromRankIdx];
270 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
271 0 : dstMem_.token = token_[toRankIdx];
272 0 : srcMem_.token = token_[myRankIdx_];
273 :
274 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
275 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
276 :
277 0 : for (u32 i = 0; i < sendSliceIdxList.size(); i++) {
278 0 : sendSliceIdx = sendSliceIdxList[i];
279 :
280 0 : if (i != 0) {
281 0 : if (i % RANK_NUM_PER_CKE == 0) {
282 0 : LocalWait(localSignal_, (1 << RANK_NUM_PER_CKE) - 1);
283 : }
284 : }
285 :
286 0 : srcMem_.addr = output_[myRankIdx_];
287 0 : srcMem_.addr += sliceOffset_[sendSliceIdx];
288 :
289 0 : dstMem_.addr = output_[toRankIdx];
290 0 : dstMem_.addr += sliceOffset_[sendSliceIdx];
291 0 : DoSendRecvSlice(nhrStepInfo.toRank, srcMem_, dstMem_, sendSliceIdx, i % RANK_NUM_PER_CKE);
292 : }
293 :
294 0 : LocalWait(localSignal_, (1 << (sendSliceIdxList.size() % RANK_NUM_PER_CKE)) - 1);
295 :
296 0 : if (nhrStepInfo.step + 1 != stepInfoVector_.size()) { // 最后一步不需要同步
297 : // 通知toRank，写入完毕
298 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_3, selfBit, true);
299 : // 等待fromRank通知写入完毕
300 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
301 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
302 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_3, recvBit);
303 : }
304 :
305 0 : HCCL_DEBUG("[DoAllGatherNHRSingleStep] rank %u step %u, toRank=%u, fromRank=%u, nSlice=%lu",
306 : rankId_, nhrStepInfo.step, nhrStepInfo.toRank, nhrStepInfo.fromRank, sendSliceIdxList.size());
307 0 : }
308 :
309 0 : void CcuContextBroadcastNHRMem2Mem1D::Algorithm()
310 : {
311 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] BroadcastNHR1D run");
312 :
313 0 : InitResources();
314 0 : LoadArgs();
315 0 : if (axisSize_ > 1) {
316 0 : AxisSync(FST_AXIS_ID);
317 : }
318 0 : PreSync();
319 0 : DoScatterNHR();
320 0 : DoAllGatherNHR();
321 0 : PostSync();
322 0 : if (axisSize_ > 1) {
323 0 : AxisSync(SEC_AXIS_ID);
324 : }
325 :
326 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] BroadcastNHR1D end");
327 0 : return;
328 : }
329 :
330 0 : std::vector<uint64_t> CcuContextBroadcastNHRMem2Mem1D::GeneArgs(const CcuTaskArg &arg)
331 : {
332 0 : const CcuTaskArgBroadcastNHRMem2Mem1D *taskArg = dynamic_cast<const CcuTaskArgBroadcastNHRMem2Mem1D *>(&arg);
333 0 : if (taskArg == nullptr) {
334 0 : THROW<NullPtrException>(StringFormat("CcuContextBroadcastNHRMem2Mem1D::taskArg ptr is null"));
335 : }
336 : // input&output&buffer地址
337 0 : uint64_t inputAddr = taskArg->inputAddr_;
338 0 : uint64_t outputAddr = taskArg->outputAddr_;
339 0 : uint64_t token = taskArg->token_;
340 0 : uint64_t die0Size = taskArg->die0Size_;
341 0 : uint64_t die1Size = taskArg->die1Size_;
342 0 : uint64_t die0SliceSize = taskArg->die0SliceSize_;
343 0 : uint64_t die1SliceSize = taskArg->die1SliceSize_;
344 0 : uint64_t die0LastSliceSize = taskArg->die0LastSliceSize_;
345 0 : uint64_t die1LastSliceSize = taskArg->die1LastSliceSize_;
346 :
347 0 : HCCL_INFO("[CcuContextBroadcastNHRMem2Mem1D] TaskArgs: inputAddr[%llu], outputAddr[%llu], "
348 : "die0Size[%llu], die1Size[%llu], die0SliceSize[%llu], die1SliceSize[%llu],"
349 : "die0LastSliceSize[%llu], die1LastSliceSize[%llu]",
350 : inputAddr, outputAddr, die0Size, die1Size, die0SliceSize, die1SliceSize,
351 : die0LastSliceSize, die1LastSliceSize);
352 :
353 : return {inputAddr, outputAddr,
354 : token, die0Size, die1Size,
355 : die0SliceSize, die1SliceSize, die0LastSliceSize,
356 0 : die1LastSliceSize};
357 : }
358 :
359 : } // namespace Hccl
|
