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 "ins_v2_all_gather_sole_executor.h"
12 :
13 : #include "log.h"
14 :
15 : #include "ins_coll_alg_registry.h"
16 :
17 : #include "topo_match_mesh.h"
18 : #include "topo_match_nhr.h"
19 : #include "topo_match_concurr_mesh.h"
20 :
21 : #include "ins_temp_all_gather_mesh.h"
22 : #include "ins_temp_all_gather_mesh_2D.h"
23 : #include "ins_temp_all_gather_nhr.h"
24 : #ifndef CCL_KERNEL_AICPU
25 : #include "ccu_temp_all_gather_mesh_1D_mem2mem_with_stride.h"
26 : #include "ccu_temp_all_gather_nhr_1D_mem2mem.h"
27 : #include "ccu_temp_all_gather_mesh_2D_mem2mem.h"
28 : #include "aiv_temp_all_gather_mesh_1D.h"
29 : #include "ccu_temp_all_gather_mesh_1D_2die.h"
30 : #endif
31 :
32 : namespace Hccl {
33 : constexpr u64 MAX_OFFLOAD_SCRATCH_SIZE = 200 * 1024 * 1024; // 200M
34 :
35 : template <typename AlgTopoMatch, typename InsAlgTemplate>
36 0 : InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::InsV2AllGatherSoleExecutor() : InsCollAlgBase()
37 0 : {}
38 :
39 : template <typename AlgTopoMatch, typename InsAlgTemplate>
40 0 : InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::~InsV2AllGatherSoleExecutor()
41 0 : {}
42 :
43 : template <typename AlgTopoMatch, typename InsAlgTemplate>
44 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::InitCommInfo(const RankGraph *rankGraph)
45 : {
46 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
47 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
48 0 : return HcclResult::HCCL_SUCCESS;
49 0 : }
50 :
51 : template <typename AlgTopoMatch, typename InsAlgTemplate>
52 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::InitCommInfo(const AlgTopoInfo &topoInfo)
53 : {
54 0 : if (topoInfo.vTopo.size() < 1) {
55 0 : return HcclResult::HCCL_E_INTERNAL;
56 : }
57 0 : vTopo_ = topoInfo.vTopo[0]; // 本通信域内的通信平面
58 0 : virtRankMap_ = topoInfo.virtRankMap[0]; // 本通信域内的 rank 映射表
59 0 : virtRanks_ = topoInfo.virtRanks[0]; // 本通信域内的 rank 集合
60 0 : return HcclResult::HCCL_SUCCESS;
61 : }
62 :
63 : template <typename AlgTopoMatch, typename InsAlgTemplate>
64 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::CreateTemplates(
65 : std::shared_ptr<InsAlgTemplate> &algTemplatePtr)
66 : {
67 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor][CreateTemplates]");
68 0 : algTemplatePtr = std::make_shared<InsAlgTemplate>(myRank_, rankSize_, vTopo_, virtRankMap_);
69 0 : CHK_PTR_NULL(algTemplatePtr); // 检查是否成功分配内存
70 0 : algTemplatePtr->SetDmaMode(dmaMode_);
71 0 : algTemplatePtr->SetDataType(dataType_);
72 0 : algTemplatePtr->SetCollOp(op_);
73 0 : return HcclResult::HCCL_SUCCESS;
74 : }
75 :
76 : template <typename AlgTopoMatch, typename InsAlgTemplate>
77 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::GetTemplateResRequest(
78 : const RankGraph *rankGraph, std::shared_ptr<InsAlgTemplate> &algTemplate, AlgTempResReq &tempResReq) const
79 : {
80 0 : if (enableDetour_) {
81 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor] [%s] Rank[%d], CalcRes with detouring enabled.", __func__, myRank_);
82 0 : CHK_RET(algTemplate->CalcResDetour(rankGraph, tempResReq));
83 : } else {
84 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor] [%s] Rank[%d], CalcRes with detouring disabled.", __func__, myRank_);
85 0 : CHK_RET(algTemplate->CalcRes(tempResReq));
86 : }
87 0 : return HcclResult::HCCL_SUCCESS;
88 : }
89 :
90 : template <typename AlgTopoMatch, typename InsAlgTemplate>
91 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::GetTemplateResRequest(
92 : ConnectedLinkMgr *linkMgr, std::shared_ptr<InsAlgTemplate> &algTemplate, AlgTempResReq &tempResReq) const
93 : {
94 0 : if (enableDetour_) {
95 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor] [%s] Rank[%d], CalcRes with detouring enabled.", __func__, myRank_);
96 0 : CHK_RET(algTemplate->CalcResDetour(linkMgr, tempResReq));
97 : } else {
98 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor] [%s] Rank[%d], CalcRes with detouring disabled.", __func__, myRank_);
99 0 : CHK_RET(algTemplate->CalcRes(tempResReq));
100 : }
101 0 : return HcclResult::HCCL_SUCCESS;
102 : }
103 :
104 : template <typename AlgTopoMatch, typename InsAlgTemplate>
105 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::CalNumBlocks(u32& numBlocks, u64 dataSize, u32 numBlocksLimit)
106 : {
107 0 : std::shared_ptr<InsAlgTemplate> algTemplate = nullptr;
108 0 : CHK_RET(CreateTemplates(algTemplate));
109 0 : CHK_RET(algTemplate->CalNumBlocks(numBlocks, dataSize, numBlocksLimit));
110 0 : return HcclResult::HCCL_SUCCESS;
111 0 : }
112 :
113 : // HOST 侧算法入口
114 : template <typename AlgTopoMatch, typename InsAlgTemplate>
115 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::Orchestrate(
116 : const RankGraph *rankGraph, const CollAlgOperator &op, const CollAlgParams ¶ms, InsQuePtr insQue)
117 : {
118 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor][Orchestrate] Orchestrate HOST Start");
119 0 : CHK_RET(Init(op, params, insQue));
120 0 : CHK_RET(InitCommInfo(rankGraph));
121 :
122 0 : std::shared_ptr<InsAlgTemplate> algTemplate = nullptr;
123 0 : CHK_RET(CreateTemplates(algTemplate));
124 :
125 0 : AlgTempResReq tempResReq;
126 0 : CHK_RET(GetTemplateResRequest(rankGraph, algTemplate, tempResReq));
127 :
128 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor][Orchestrate] Rank[%d], template [%s], requiredQue Num [%u].",
129 : myRank_,
130 : algTemplate->Describe().c_str(),
131 : tempResReq.queNum);
132 0 : CHK_RET(InitQueue(tempResReq.queNum, tempInsQue_));
133 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, tempResReq.links, tempResLinks_));
134 0 : CHK_RET(OrchestrateLoop(algTemplate));
135 0 : return HcclResult::HCCL_SUCCESS;
136 0 : }
137 :
138 : // AICPU 侧算法入口
139 : template <typename AlgTopoMatch, typename InsAlgTemplate>
140 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::Orchestrate(const AlgTopoInfo &topoInfo,
141 : const CollAlgOperator &op, const CollAlgParams ¶ms, ConnectedLinkMgr *linkMgr, InsQuePtr insQue)
142 : {
143 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor][Orchestrate] Orchestrate AICPU Start");
144 0 : CHK_RET(Init(op, params, insQue));
145 0 : dataType_ = op.dataType;
146 :
147 0 : CHK_RET(InitCommInfo(topoInfo));
148 :
149 0 : std::shared_ptr<InsAlgTemplate> algTemplate = nullptr;
150 0 : CHK_RET(CreateTemplates(algTemplate));
151 0 : algTemplate->SetDataType(dataType_);
152 0 : std::map<u32, u32>rank2PathNumMap;
153 0 : CHK_RET(SetPathNumMapByLinkMgrMultiLevel(linkMgr, virtRanks_, myRank_, rank2PathNumMap));
154 0 : algTemplate->setPathNumMap(rank2PathNumMap);
155 :
156 0 : AlgTempResReq tempResReq;
157 0 : CHK_RET(GetTemplateResRequest(linkMgr, algTemplate, tempResReq));
158 :
159 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor][Orchestrate] Rank[%d], template [%s], requiredQue Num [%u].",
160 : myRank_,
161 : algTemplate->Describe().c_str(),
162 : tempResReq.queNum);
163 0 : CHK_RET(InitQueue(tempResReq.queNum, tempInsQue_));
164 0 : CHK_RET(PrepResLinks(myRank_, tempResReq.links, linkMgr, tempResLinks_));
165 0 : CHK_RET(OrchestrateLoop(algTemplate));
166 0 : return HcclResult::HCCL_SUCCESS;
167 0 : }
168 :
169 : // 切分数据并调用 template
170 : template <typename AlgTopoMatch, typename InsAlgTemplate>
171 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::OrchestrateLoop(
172 : std::shared_ptr<InsAlgTemplate> algTemplate)
173 : {
174 0 : HCCL_INFO("[InsV2AllGatherSoleExecutor][OrchestrateOpbase] Start, template[%s]", algTemplate->Describe().c_str());
175 0 : u32 dataSizePerVolume = DataTypeSizeGet(dataType_);
176 0 : dataSize_ = dataCount_ * dataSizePerVolume;
177 :
178 0 : TemplateDataParams tempAlgParams;
179 0 : tempAlgParams.inputRepeatStride = 0;
180 0 : tempAlgParams.outputRepeatStride = 0;
181 0 : tempAlgParams.buffInfo.inBuffType = BufferType::INPUT;
182 0 : tempAlgParams.buffInfo.outBuffType = BufferType::OUTPUT;
183 0 : tempAlgParams.buffInfo.scratBuffType = BufferType::SCRATCH;
184 0 : tempAlgParams.repeatNum = 1; // 不需要重复
185 :
186 0 : TempFuncs tempFuncs;
187 0 : tempFuncs.isForepart = true;
188 0 : tempFuncs.isBottom = true;
189 0 : tempFuncs.opMode = opMode_;
190 0 : tempFuncs.enableCounterNotify = IsEnableCounterNotify();
191 :
192 0 : u64 maxDataSizePerLoop = 0;
193 0 : u64 transportBoundDataSize = UB_MAX_DATA_SIZE; // algTemplate->CalcLoopMaxCount();
194 :
195 0 : u32 templateScratchMultiplier = algTemplate->CalcScratchMultiple(BufferType::INPUT, BufferType::OUTPUT);
196 0 : if (templateScratchMultiplier != 0) {
197 0 : u64 scratchBoundDataSize = maxTmpMemSize_ / templateScratchMultiplier;
198 0 : maxDataSizePerLoop = std::min(transportBoundDataSize, scratchBoundDataSize);
199 : } else {
200 0 : maxDataSizePerLoop = transportBoundDataSize;
201 : }
202 0 : u64 maxDataCountPerLoop = maxDataSizePerLoop / dataTypeSize_;
203 0 : HCCL_INFO("[InsV2AllGatherSoleExecutor][OrchestrateOpbase] maxDataCountPerLoop[%llu], maxDataSizePerLoop[%llu], "
204 : "transportBoundDataSize[%llu], templateScratchMultiplier[%llu]",
205 : maxDataCountPerLoop,
206 : maxDataSizePerLoop,
207 : transportBoundDataSize,
208 : templateScratchMultiplier);
209 0 : CHK_PRT_RET(maxDataCountPerLoop == 0,
210 : HCCL_ERROR("[InsV2AllGatherSoleExecutor][OrchestrateOpbase] maxDataCountPerLoop is 0"),
211 : HCCL_E_INTERNAL);
212 :
213 0 : u64 processedDataCount = 0; // 已经处理的数据count
214 0 : u64 loopTimes = dataCount_ / maxDataCountPerLoop + static_cast<u64>(dataCount_ % maxDataCountPerLoop != 0);
215 0 : for (u64 loop = 0; loop < loopTimes; loop++) {
216 0 : u64 currDataCount = (loop == loopTimes - 1) ? dataCount_ - processedDataCount : maxDataCountPerLoop;
217 :
218 0 : tempAlgParams.buffInfo.inBuffBaseOff = processedDataCount * dataTypeSize_;
219 0 : tempAlgParams.buffInfo.outBuffBaseOff = processedDataCount * dataTypeSize_;
220 0 : tempAlgParams.buffInfo.scratchBuffBaseOff = 0;
221 0 : tempAlgParams.sliceSize = currDataCount * dataTypeSize_;
222 0 : tempAlgParams.tailSize = tempAlgParams.sliceSize;
223 0 : tempAlgParams.inputSliceStride = 0; // 输入数据仅有 1 个 slice, 不需要 stride
224 0 : tempAlgParams.outputSliceStride = dataSize_; // 每张卡的数据间隔为算子输入大小
225 :
226 0 : CHK_RET(algTemplate->GenExtIns(tempFuncs, tempAlgParams, tempResLinks_, tempInsQue_));
227 0 : processedDataCount += currDataCount;
228 : }
229 :
230 0 : return HcclResult::HCCL_SUCCESS;
231 0 : }
232 :
233 : template <typename AlgTopoMatch, typename InsAlgTemplate>
234 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::CalcRes(
235 : const RankGraph *rankGraph, CollAlgResReq &algResReq)
236 : {
237 : // Topo Match
238 0 : CHK_RET(InitCommInfo(rankGraph));
239 :
240 : // instantiate a template
241 0 : InsAlgTemplate tempAlg(myRank_, rankSize_, vTopo_, virtRankMap_);
242 :
243 0 : std::shared_ptr<InsAlgTemplate> algTemplate = nullptr;
244 0 : CHK_RET(CreateTemplates(algTemplate));
245 :
246 : // 通过判断哪层通信域能有到所有remoteRank的path，判断当前算法跑在哪一层
247 0 : std::map<u32, u32>rank2PathNumMap;
248 0 : HCCL_INFO("[InsV2AllGatherSoleExecutor] CalcRes SetPathNumMap");
249 0 : CHK_RET(SetPathNumMapByRankGraphMultiLevel(rankGraph, virtRanks_, myRank_, rank2PathNumMap));
250 0 : algTemplate->setPathNumMap(rank2PathNumMap);
251 0 : AlgTempResReq tempResReq;
252 0 : CHK_RET(GetTemplateResRequest(rankGraph, algTemplate, tempResReq));
253 :
254 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, tempResReq.links, algResReq.levelRankPairs));
255 0 : algResReq.topoInfo.UpdateSingleLevelTopo(virtRanks_, virtRankMap_, vTopo_);
256 0 : algResReq.queueNotifys = tempResReq.queNotifys;
257 0 : algResReq.primQueueNum = tempResReq.streamNum;
258 0 : algResReq.localWaitGroupCntNotify = tempResReq.localWaitGroupCntNotify;
259 0 : algResReq.localBcastPostCntNotify = tempResReq.localBcastPostCntNotify;
260 0 : HCCL_DEBUG("[InsV2AllGatherSoleExecutor] [%s] Rank[%d], requiredQueNum [%u].", __func__, myRank_, algResReq.primQueueNum);
261 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, tempResReq.links, algResReq.links));
262 :
263 0 : return HcclResult::HCCL_SUCCESS;
264 0 : }
265 :
266 : template <typename AlgTopoMatch, typename InsAlgTemplate>
267 0 : HcclResult InsV2AllGatherSoleExecutor<AlgTopoMatch, InsAlgTemplate>::CalcResOffload(
268 : const RankGraph *rankGraph, const u64 &dataSize, CollOffloadOpResReq &resReq)
269 : {
270 : // Topo Match
271 : (void)dataSize;
272 0 : CHK_RET(InitCommInfo(rankGraph));
273 :
274 0 : std::shared_ptr<InsAlgTemplate> algTemplate = nullptr;
275 0 : CHK_RET(CreateTemplates(algTemplate));
276 :
277 : // 通过判断哪层通信域能有到所有remoteRank的path，判断当前算法跑在哪一层
278 0 : std::map<u32, u32>rank2PathNumMap;
279 0 : HCCL_INFO("[InsV2AllGatherSoleExecutor] CalcResOffload SetPathNumMap");
280 0 : CHK_RET(SetPathNumMapByRankGraphMultiLevel(rankGraph, virtRanks_, myRank_, rank2PathNumMap));
281 0 : algTemplate->setPathNumMap(rank2PathNumMap);
282 0 : AlgTempResReq tempResReq;
283 0 : CHK_RET(GetTemplateResRequest(rankGraph, algTemplate, tempResReq));
284 0 : resReq.requiredScratchMemSize = UB_MAX_DATA_SIZE;
285 0 : resReq.requiredSubQueNum = tempResReq.streamNum - 1;
286 :
287 0 : return HcclResult::HCCL_SUCCESS;
288 0 : }
289 :
290 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, InsAllGatherMesh, InsV2AllGatherSoleExecutor, TopoMatchMesh,
291 : InsTempAllGatherMesh1D);
292 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, InsAllGatherMesh2D, InsV2AllGatherSoleExecutor, TopoMatchConcurrMesh,
293 : InsTempAllGatherMesh2D);
294 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, InsAllGatherNHR, InsV2AllGatherSoleExecutor, TopoMatchNHR,
295 : InsTempAllGatherNHR);
296 : #ifndef CCL_KERNEL_AICPU
297 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, AivAllGatherMesh1D, InsV2AllGatherSoleExecutor,
298 : TopoMatchMesh, AivTempAllGatherMesh1D);
299 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, CcuAllGatherMesh1DMem2MemWithStride, InsV2AllGatherSoleExecutor,
300 : TopoMatchMesh, CcuTempAllGatherMesh1DMem2MemWithStride);
301 : INS_REGISTER_IMPL_BY_TEMP(
302 : OpType::ALLGATHER, CcuAllGatherNHR1D, InsV2AllGatherSoleExecutor, TopoMatchMesh, CcuTempAllGatherNHRMem2Mem1D);
303 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, CcuAllGatherMeshMem2Mem2D, InsV2AllGatherSoleExecutor,
304 : TopoMatchConcurrMesh, CcuTempAllGatherMeshMem2Mem2D);
305 : INS_REGISTER_IMPL_BY_TEMP(OpType::ALLGATHER, CcuAllGatherMesh1D2Die, InsV2AllGatherSoleExecutor,
306 : TopoMatchMesh, CcuTempAllGatherMesh1D2Die);
307 : #endif
308 : } // namespace Hccl
|
