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_scatter_nhr1d_mem2mem.h"
12 :
13 : namespace Hccl {
14 :
15 : // 注意型号量变化
16 : constexpr uint16_t INPUT_XN_ID = 0;
17 : constexpr uint16_t TOKEN_XN_ID = 2;
18 : constexpr uint16_t CKE_IDX_0 = 0;
19 : constexpr uint16_t CKE_IDX_1 = 1;
20 : constexpr uint16_t CKE_IDX_2 = 2;
21 : constexpr uint16_t CKE_IDX_3 = 3;
22 : constexpr uint16_t CKE_IDX_4 = 4;
23 : constexpr uint16_t FST_AXIS_ID = 0;
24 : constexpr uint16_t SEC_AXIS_ID = 1;
25 : 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
26 : constexpr uint16_t LINK_SIZE = 2;
27 :
28 0 : CcuContextReduceScatterNHR1DMem2Mem::CcuContextReduceScatterNHR1DMem2Mem(const CcuCtxArg &arg,
29 : const std::vector<CcuTransport *> &transports,
30 0 : const CcuTransportGroup &group)
31 0 : : CcuContextAlgBase(arg, transports, group)
32 : {
33 0 : const CcuCtxArgReduceScatterNHR1D *ctxArg = dynamic_cast<const CcuCtxArgReduceScatterNHR1D *>(&arg);
34 0 : rankId_ = ctxArg->rankId_;
35 0 : axisId_ = ctxArg->axisId_;
36 0 : dimSize_ = ctxArg->dimSize_[0];
37 0 : localAxisSignalName_ = "CcuContextReduceScatterNHR1DDieSync_" + std::to_string(axisId_);
38 0 : anotherAxisSignalName_ = "CcuContextReduceScatterNHR1DDieSync_" + std::to_string(1 - axisId_);
39 0 : stepInfoVector_ = ctxArg->stepInfoVector_;
40 0 : indexMap_ = ctxArg->indexMap_;
41 0 : localSize_ = indexMap_.size();
42 0 : myRankIdx_ = indexMap_.size();
43 0 : reduceOp_ = ctxArg->op_.reduceOp;
44 0 : dataType_ = ctxArg->op_.dataType;
45 0 : outputDataType_ = ctxArg->op_.outputDataType;
46 0 : linkNum_ = ctxArg->linkNum_;
47 0 : if (outputDataType_ == DataType::INVALID) {
48 0 : outputDataType_ = dataType_;
49 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] outputDataType is [INVALID], set outputDataType to[%s]",
50 : outputDataType_.Describe().c_str());
51 : }
52 0 : signalNum_ = (dimSize_ + RANK_NUM_PER_CKE - 1) / RANK_NUM_PER_CKE; // 每个CKE有16个bit
53 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] CtxArg: rankId_[%u], axisId_[%u], dimSize_[%u], localSize_[%u], "
54 : "dataType[%s], outputDataType[%s], reduceOp[%s], signalNum_[%u]",
55 : rankId_, axisId_, dimSize_, localSize_, dataType_.Describe().c_str(),
56 : outputDataType_.Describe().c_str(), reduceOp_.Describe().c_str(), signalNum_);
57 0 : }
58 :
59 0 : void CcuContextReduceScatterNHR1DMem2Mem::LoadArgs()
60 : {
61 0 : Load(input_[myRankIdx_]);
62 0 : Load(output_);
63 0 : Load(token_[myRankIdx_]);
64 0 : Load(die0Size_);
65 0 : Load(die1Size_);
66 0 : Load(inputSliceStride_);
67 0 : Load(outputSliceStride_);
68 0 : Load(inputRepeatStride_);
69 0 : Load(outputRepeatStride_);
70 0 : Load(repeatNumVar_);
71 0 : Load(isBottom_);
72 0 : repeatNumVarTemp_ = repeatNumVar_;
73 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] LoadArgs run finished");
74 0 : }
75 :
76 0 : void CcuContextReduceScatterNHR1DMem2Mem::InitResources()
77 : {
78 0 : die0Size_ = CreateVariable();
79 0 : die1Size_ = CreateVariable();
80 0 : sliceSize_ = CreateVariable();
81 0 : inputSliceStride_ = CreateVariable();
82 0 : outputSliceStride_ = CreateVariable();
83 0 : inputRepeatStride_ = CreateVariable();
84 0 : outputRepeatStride_ = CreateVariable();
85 0 : localAxisSignal_ = CreateMaskSignal();
86 0 : anotherAxisSignal_ = CreateMaskSignal();
87 0 : localSignal_ = CreateMaskSignal();
88 0 : repeatNumVar_ = CreateVariable();
89 0 : repeatNumVarTemp_ = CreateVariable();
90 0 : isBottom_ = CreateVariable();
91 :
92 0 : if (linkNum_ == LINK_SIZE) {
93 0 : ExportMaskSignal(localAxisSignal_, localAxisSignalName_);
94 0 : anotherAxisSignal_ = ImportMaskSignal(anotherAxisSignalName_);
95 : }
96 :
97 0 : output_ = CreateVariable();
98 0 : for (uint32_t transportIdx = 0; transportIdx < localSize_; transportIdx++) {
99 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] MyRank[%u], TransportId[%u]", rankId_, transportIdx);
100 0 : CHK_PRT_RET(transports[transportIdx] == nullptr,
101 : HCCL_ERROR("[CcuContextReduceScatterNHR1DMem2Mem] Algorithm transport ptr is null"),);
102 0 : input_.push_back(
103 0 : CreateVariable((*transports[transportIdx]), INPUT_XN_ID)); // 获取transport中id=0的Var来传递input
104 :
105 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
106 : }
107 0 : input_.push_back(CreateVariable());
108 0 : token_.push_back(CreateVariable());
109 :
110 0 : repeatInputOffset_ = CreateVariable();
111 0 : repeatOutputOffset_ = CreateVariable();
112 0 : myrankInputSliceOffset_ = CreateVariable();
113 :
114 0 : srcMem_ = CreateMemory();
115 0 : dstMem_ = CreateMemory();
116 0 : flag_ = CreateVariable();
117 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] InitResources finished");
118 : }
119 :
120 0 : void CcuContextReduceScatterNHR1DMem2Mem::PreSync()
121 : {
122 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] PreSync start");
123 : // 本rank用哪一个CKE
124 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
125 : // 本rank用CKE的哪一位
126 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
127 0 : for (auto t : transports) {
128 0 : WriteVariableWithSignal(*t, input_[localSize_], INPUT_XN_ID, selfSignalId + signalNum_ * CKE_IDX_1, selfBit);
129 0 : WriteVariableWithSignal(*t, token_[localSize_], TOKEN_XN_ID, selfSignalId + signalNum_ * CKE_IDX_2, selfBit);
130 : }
131 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
132 0 : for (auto &pair : indexMap_) {
133 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
134 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
135 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
136 : }
137 0 : for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) {
138 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_1, waitBitVector[sId]);
139 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_2, waitBitVector[sId]);
140 : }
141 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] PreSync end");
142 0 : }
143 :
144 0 : void CcuContextReduceScatterNHR1DMem2Mem::PostSync()
145 : {
146 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
147 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
148 0 : for (auto &t : transports) {
149 0 : RemotePost(*t, selfSignalId + signalNum_ * CKE_IDX_0, selfBit);
150 : }
151 0 : std::vector<uint16_t> waitBitVector(signalNum_, 0);
152 0 : for (auto &pair : indexMap_) {
153 0 : uint16_t pairSignalId = pair.first / RANK_NUM_PER_CKE;
154 0 : uint16_t pairBit = 1 << (pair.first % RANK_NUM_PER_CKE);
155 0 : waitBitVector[pairSignalId] = waitBitVector[pairSignalId] | pairBit;
156 : }
157 0 : for (uint32_t sId = 0; sId < waitBitVector.size(); sId++) {
158 0 : GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_0, waitBitVector[sId]);
159 : }
160 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] PostSync run finished");
161 0 : }
162 :
163 0 : void CcuContextReduceScatterNHR1DMem2Mem::AxisSync(uint32_t signalIndex)
164 : {
165 0 : const uint32_t DIE_NUM = 2;
166 0 : if (signalIndex > 1) {
167 0 : THROW<InvalidParamsException>(
168 0 : StringFormat("[CcuContextReduceScatterNHR1DMem2Mem] Unexpected SignalInex[%u]", signalIndex));
169 : }
170 0 : LocalCtxPost(anotherAxisSignal_, 1 << (axisId_ + signalIndex * DIE_NUM));
171 0 : LocalWait(localAxisSignal_, 1 << (1 - axisId_ + signalIndex * DIE_NUM));
172 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] AxisSync run finished");
173 0 : return;
174 : }
175 :
176 0 : void CcuContextReduceScatterNHR1DMem2Mem::DoRepeatReduceScatterNHR()
177 : {
178 0 : CcuRep::Variable tmpSliceOffset = CreateVariable();
179 0 : tmpSliceOffset = 0;
180 : // 用来记录每个rank要读取的rank的sliceIdx的偏移
181 : // 后面会用inputAddr来加上这个偏移获取sliceIdx的地址
182 0 : std::vector<CcuRep::Variable> inputSliceOffset;
183 0 : CCU_IF(isBottom_ == 1) {
184 0 : for (u64 i = 0; i < dimSize_; i++) {
185 0 : inputSliceOffset.push_back(CreateVariable());
186 0 : inputSliceOffset[i] = tmpSliceOffset;
187 0 : tmpSliceOffset += inputSliceStride_;
188 : }
189 0 : }
190 0 : CCU_IF(isBottom_ == 0) {
191 0 : for (u64 i = 0; i < dimSize_; i++) {
192 0 : inputSliceOffset.push_back(CreateVariable());
193 0 : inputSliceOffset[i] = tmpSliceOffset;
194 0 : tmpSliceOffset += inputRepeatStride_;
195 : }
196 0 : }
197 :
198 0 : for (auto &nhrStepInfo : stepInfoVector_) {
199 0 : DoRepeatReduceScatterNHRSingleStep(nhrStepInfo, inputSliceOffset);
200 : }
201 : // 因为所有的修改都是在input上进行的，所以最后需要把input上的数据搬到output上
202 0 : dstMem_.addr = output_;
203 0 : dstMem_.token = token_[myRankIdx_];
204 0 : srcMem_.addr = input_[myRankIdx_];
205 0 : srcMem_.addr += inputSliceOffset[rankId_];
206 0 : srcMem_.token = token_[myRankIdx_];
207 :
208 0 : CcuRep::Variable repeatNumAdd2 = CreateVariable();
209 0 : repeatNumAdd2 = 1;
210 0 : CCU_WHILE(repeatNumVar_ != UINT64_MAX) {
211 0 : repeatNumVar_ += repeatNumAdd2;
212 0 : CCU_IF(flag_ == 1) {
213 0 : CCU_IF(isBottom_ == 0) {
214 0 : srcMem_.addr += inputSliceStride_;
215 0 : dstMem_.addr += outputRepeatStride_;
216 0 : }
217 0 : CCU_IF(isBottom_ == 1) {
218 0 : srcMem_.addr += inputRepeatStride_;
219 0 : dstMem_.addr += outputRepeatStride_;
220 0 : }
221 0 : }
222 0 : CCU_IF(flag_ == 0) {
223 0 : if (axisId_ == 1) {
224 0 : srcMem_.addr += die0Size_;
225 0 : dstMem_.addr += die0Size_;
226 : }
227 0 : }
228 0 : CcuRep::Variable &localSliceSize = (axisId_ == 0) ? die0Size_ : die1Size_;
229 0 : LocalCopy(dstMem_, srcMem_, localSliceSize, localSignal_, 1);
230 0 : LocalWait(localSignal_, 1);
231 0 : flag_ = 1;
232 0 : }
233 0 : }
234 :
235 0 : void CcuContextReduceScatterNHR1DMem2Mem::DoRepeatReduceScatterNHRSingleStep(const NHRStepInfo &nhrStepInfo,
236 : const std::vector<CcuRep::Variable> &inputSliceOffset)
237 : {
238 0 : u32& toRankIdx = indexMap_[nhrStepInfo.toRank];
239 0 : u32& fromRankIdx = indexMap_[nhrStepInfo.fromRank];
240 0 : CcuTransport *sendTransport = transports[toRankIdx];
241 0 : CcuTransport *recvTransport = transports[fromRankIdx];
242 0 : const std::vector<u32> &sendSliceIdxList = nhrStepInfo.txSliceIdxs;
243 0 : dstMem_.token = token_[toRankIdx];
244 0 : srcMem_.token = token_[myRankIdx_];
245 :
246 : // 被写之前告诉写自己的rank自己准备好了-前同步
247 0 : uint16_t recvSignalIdPrev = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
248 0 : uint16_t recvBitPrev = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
249 0 : RemotePost(*recvTransport, recvSignalIdPrev + signalNum_ * CKE_IDX_3, recvBitPrev, true);
250 :
251 0 : uint16_t selfSignalIdPrev = rankId_ / RANK_NUM_PER_CKE;
252 0 : uint16_t selfBitPrev = 1 << (rankId_ % RANK_NUM_PER_CKE);
253 0 : RemoteWait(*sendTransport, selfSignalIdPrev + signalNum_ * CKE_IDX_3, selfBitPrev);
254 :
255 0 : for (const u32 &sendSliceIdx : sendSliceIdxList) {
256 0 : dstMem_.addr = input_[toRankIdx];
257 0 : dstMem_.addr += inputSliceOffset[sendSliceIdx];
258 0 : srcMem_.addr = input_[myRankIdx_];
259 0 : srcMem_.addr += inputSliceOffset[sendSliceIdx];
260 0 : DoRepeatSendRecvSlices(nhrStepInfo.toRank, srcMem_, dstMem_);
261 : }
262 :
263 : // 写之后告诉对面写完了-后同步
264 0 : uint16_t selfSignalId = rankId_ / RANK_NUM_PER_CKE;
265 0 : uint16_t selfBit = 1 << (rankId_ % RANK_NUM_PER_CKE);
266 0 : RemotePost(*sendTransport, selfSignalId + signalNum_ * CKE_IDX_4, selfBit, true);
267 :
268 0 : uint16_t recvSignalId = nhrStepInfo.fromRank / RANK_NUM_PER_CKE;
269 0 : uint16_t recvBit = 1 << (nhrStepInfo.fromRank % RANK_NUM_PER_CKE);
270 0 : RemoteWait(*recvTransport, recvSignalId + signalNum_ * CKE_IDX_4, recvBit);
271 0 : }
272 :
273 0 : void CcuContextReduceScatterNHR1DMem2Mem::DoRepeatSendRecvSlices(const u32 &toRank, CcuRep::Memory &src,
274 : CcuRep::Memory &dst)
275 : {
276 0 : CcuRep::Variable repeatNumAdd = CreateVariable();
277 0 : repeatNumAdd = 1;
278 0 : flag_ = 0;
279 0 : CcuTransport *sendTransport = transports[indexMap_[toRank]];
280 0 : repeatNumVarTemp_ = repeatNumVar_;
281 0 : CCU_WHILE(repeatNumVarTemp_ != UINT64_MAX) {
282 0 : CCU_IF(repeatNumVarTemp_ != UINT64_MAX) {
283 0 : repeatNumVarTemp_ += repeatNumAdd;
284 0 : }
285 :
286 0 : CCU_IF(flag_ == 1) {
287 0 : CCU_IF(isBottom_ == 0) {
288 0 : src.addr += inputSliceStride_;
289 0 : dst.addr += inputSliceStride_;
290 0 : }
291 0 : CCU_IF(isBottom_ == 1) {
292 0 : src.addr += inputRepeatStride_;
293 0 : dst.addr += inputRepeatStride_;
294 0 : }
295 0 : }
296 0 : CCU_IF(flag_ == 0) {
297 0 : if (axisId_ == 1) {
298 0 : src.addr += die0Size_;
299 0 : dst.addr += die0Size_;
300 : }
301 0 : }
302 0 : sliceSize_ = (axisId_ == 0) ? die0Size_ : die1Size_;
303 0 : WriteReduce(*sendTransport, dst, src, sliceSize_, dataType_,
304 0 : reduceOp_, localSignal_, 1);
305 0 : LocalWait(localSignal_, (1 << 1) - 1);
306 0 : flag_ = 1;
307 0 : }
308 0 : flag_ = 0;
309 0 : }
310 :
311 0 : void CcuContextReduceScatterNHR1DMem2Mem::Algorithm()
312 : {
313 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] CcuContextReduceScatterNHR1DMem2Mem run.");
314 0 : InitResources();
315 0 : LoadArgs();
316 0 : if (linkNum_ == LINK_SIZE) {
317 0 : AxisSync(FST_AXIS_ID);
318 : }
319 0 : PreSync();
320 0 : DoRepeatReduceScatterNHR();
321 0 : PostSync();
322 0 : if (linkNum_ == LINK_SIZE) {
323 0 : AxisSync(SEC_AXIS_ID);
324 : }
325 :
326 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] CcuContextReduceScatterNHR1DMem2Mem end.");
327 0 : return;
328 : }
329 :
330 0 : std::vector<uint64_t> CcuContextReduceScatterNHR1DMem2Mem::GeneArgs(const CcuTaskArg &arg)
331 : {
332 0 : const CcuTaskArgReduceScatterNHR1D *taskArg = dynamic_cast<const CcuTaskArgReduceScatterNHR1D *>(&arg);
333 0 : if (taskArg == nullptr) {
334 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceScatterNHR1DMem2Mem::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 inputSliceStride = taskArg->inputSliceStride_;
343 0 : uint64_t outputSliceStride = taskArg->outputSliceStride_;
344 0 : uint64_t inputRepeatStride = taskArg->inputRepeatStride_;
345 0 : uint64_t outputRepeatStride = taskArg->outputRepeatStride_;
346 0 : uint64_t repeatNumVar = taskArg->repeatNum_;
347 0 : uint64_t isBottom = taskArg->isBottom_;
348 :
349 0 : HCCL_INFO("[CcuContextReduceScatterNHR1DMem2Mem] TaskArgs: inputAddr[%llu], outputAddr[%llu],"
350 : "die0Size[%llu], die1Size[%llu],"
351 : "inputSliceStride[%llu], outputSliceStride[%llu], inputRepeatStride[%llu], outputRepeatStride[%llu]",
352 : inputAddr, outputAddr, die0Size, die1Size,
353 : inputSliceStride, outputSliceStride, inputRepeatStride, outputRepeatStride);
354 :
355 : return {inputAddr, outputAddr, token,
356 : die0Size, die1Size, inputSliceStride,
357 : outputSliceStride, inputRepeatStride, outputRepeatStride,
358 0 : repeatNumVar, isBottom};
359 : }
360 : } // namespace Hccl
|
