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_all_reduce_parallel_executor_opt.h"
12 :
13 : #include "log.h"
14 :
15 : #include "ins_coll_alg_registry.h"
16 :
17 : #include "topo_match_mesh_nhr.h"
18 : #include "alg_data_trans_wrapper.h"
19 :
20 : #include "ins_temp_reduce_scatter_mesh_1D.h"
21 :
22 : #include "ins_temp_all_gather_mesh.h"
23 :
24 : namespace Hccl {
25 : constexpr u64 MAX_OFFLOAD_SCRATCH_SIZE = 200 * 1024 * 1024; // 200M
26 :
27 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
28 0 : InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::InsAllReduceParallelExecutorV2()
29 0 : : InsCollAlgBase()
30 : {
31 0 : }
32 :
33 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
34 0 : InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::~InsAllReduceParallelExecutorV2()
35 : {
36 0 : }
37 :
38 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
39 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::CalcResOffload(const RankGraph *rankGraph, const u64 &dataSize,
40 : CollOffloadOpResReq &resReq)
41 : {
42 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] CalcResOffload begins.");
43 0 : resReq.requiredScratchMemSize = MAX_OFFLOAD_SCRATCH_SIZE; // 200MB
44 : // Topo Match
45 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
46 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
47 0 : CHK_RET(CalcLocalRankSize());
48 :
49 0 : InsAlgTemplate0 intraTempAlgRS(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
50 0 : InsAlgTemplate1 interTempAlgRS(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
51 0 : InsAlgTemplate2 intraTempAlgAG(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
52 0 : InsAlgTemplate3 interTempAlgAG(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
53 :
54 : // 设置链路信息
55 0 : std::vector<map<u32, u32>> rank2PathNumMap;
56 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] CalcResOffload SetPathNumMap");
57 0 : CHK_RET(SetPathNumMapByRankGraphMultiLevel(rankGraph, virtRanks_, myRank_, rank2PathNumMap));
58 0 : intraTempAlgRS.setPathNumMap(rank2PathNumMap[0]);
59 0 : interTempAlgRS.setPathNumMap(rank2PathNumMap[1]);
60 0 : intraTempAlgAG.setPathNumMap(rank2PathNumMap[0]);
61 0 : interTempAlgAG.setPathNumMap(rank2PathNumMap[1]);
62 :
63 0 : AlgTempResReq resReqIntraRS;
64 0 : AlgTempResReq resReqInterRS;
65 :
66 0 : CHK_RET(intraTempAlgRS.CalcRes(resReqIntraRS));
67 0 : CHK_RET(interTempAlgRS.CalcRes(resReqInterRS));
68 :
69 0 : AlgTempResReq resReqIntraAG;
70 0 : AlgTempResReq resReqInterAG;
71 :
72 0 : CHK_RET(intraTempAlgAG.CalcRes(resReqIntraAG));
73 0 : CHK_RET(interTempAlgAG.CalcRes(resReqInterAG));
74 :
75 : // 算法从流数量 = Σ(temp的que数量 + temp的从流数量 * temp调用次数) - 算法主流数量
76 0 : resReq.requiredSubQueNum = std::max((resReqIntraAG.queNum + resReqInterAG.queNum), (resReqIntraRS.queNum + resReqInterRS.queNum)) - 1;
77 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2::CalcResOffload]requiredSubQueNum = %llu", resReq.requiredSubQueNum);
78 :
79 0 : return HcclResult::HCCL_SUCCESS;
80 0 : }
81 :
82 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
83 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::CalcRes(const RankGraph *rankGraph, CollAlgResReq &algResReq)
84 : {
85 0 : HCCL_INFO("[InsFourTemplateAllReduceExecutor] CalcRes begins.");
86 :
87 : // 拓扑匹配
88 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
89 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
90 0 : algResReq.topoInfo.UpdateMultiLevelTopo(virtRanks_, virtRankMap_, vTopo_);
91 0 : CHK_RET(CalcLocalRankSize());
92 :
93 : // 创建四个模板实例
94 0 : InsAlgTemplate0 tempAlgRSIntra(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
95 0 : InsAlgTemplate1 tempAlgRSInter(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
96 0 : InsAlgTemplate2 tempAlgAGIntra(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
97 0 : InsAlgTemplate3 tempAlgAGInter(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
98 :
99 : // 设置链路信息
100 0 : std::vector<map<u32, u32>> rank2PathNumMap;
101 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] CalcResOffload SetPathNumMap");
102 0 : CHK_RET(SetPathNumMapByRankGraphMultiLevel(rankGraph, virtRanks_, myRank_, rank2PathNumMap));
103 0 : tempAlgRSIntra.setPathNumMap(rank2PathNumMap[0]);
104 0 : tempAlgRSInter.setPathNumMap(rank2PathNumMap[1]);
105 0 : tempAlgAGIntra.setPathNumMap(rank2PathNumMap[0]);
106 0 : tempAlgAGInter.setPathNumMap(rank2PathNumMap[1]);
107 :
108 : // 计算各模板资源需求
109 0 : AlgTempResReq resReqRSIntra, resReqRSInter, resReqAGIntra, resReqAGInter;
110 :
111 0 : CHK_RET(tempAlgRSIntra.CalcRes(resReqRSIntra));
112 0 : CHK_RET(tempAlgRSInter.CalcRes(resReqRSInter));
113 0 : CHK_RET(tempAlgAGIntra.CalcRes(resReqAGIntra));
114 0 : CHK_RET(tempAlgAGInter.CalcRes(resReqAGInter));
115 :
116 : // 计算链接信息
117 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, resReqRSIntra.links, algResReq.levelRankPairs));
118 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, resReqRSInter.links, algResReq.levelRankPairs));
119 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, resReqAGIntra.links, algResReq.levelRankPairs));
120 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, resReqAGInter.links, algResReq.levelRankPairs));
121 :
122 0 : algResReq.primQueueNum = std::max(resReqRSIntra.queNum, resReqAGIntra.queNum) +
123 0 : std::max(resReqRSInter.queNum, resReqAGInter.queNum);
124 0 : HCCL_INFO("[InsFourTemplateAllReduceExecutor::CalcRes] primQueueNum = %u", algResReq.primQueueNum);
125 :
126 0 : std::vector<std::tuple<QId, QId, u32>> notifyRequests;
127 0 : for (QId q = 1; q < algResReq.primQueueNum; q++) {
128 0 : notifyRequests.emplace_back(std::make_tuple(0, q, 0));
129 0 : notifyRequests.emplace_back(std::make_tuple(q, 0, 0));
130 : }
131 :
132 0 : u32 tempMasterQId = std::max(resReqRSIntra.queNum, resReqAGIntra.queNum);
133 0 : for (QId q = tempMasterQId + 1; q < algResReq.primQueueNum; q++) {
134 0 : notifyRequests.emplace_back(std::make_tuple(tempMasterQId, q, 0));
135 0 : notifyRequests.emplace_back(std::make_tuple(q, tempMasterQId, 0));
136 : }
137 :
138 0 : algResReq.queueNotifys = notifyRequests;
139 :
140 : // 计算链接资源
141 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, resReqRSIntra.links, algResReq.links));
142 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, resReqRSInter.links, algResReq.links));
143 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, resReqAGIntra.links, algResReq.links));
144 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, resReqAGInter.links, algResReq.links));
145 :
146 0 : return HcclResult::HCCL_SUCCESS;
147 0 : }
148 :
149 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
150 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GetParallelDataSplitRate(
151 : std::vector<float> &splitDataSize) const
152 : {
153 0 : double splitData = 0.5;
154 0 : splitDataSize.push_back(splitData);
155 0 : splitDataSize.push_back(splitData);
156 0 : return;
157 : }
158 :
159 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
160 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::CalcLocalRankSize()
161 : {
162 0 : uint64_t virtRanks_2 = 2;
163 0 : CHK_PRT_RET(virtRanks_.size() < virtRanks_2,
164 : HCCL_ERROR("[CalcLocalRankSize] virtRanks level num is smaller than 2."),
165 : HcclResult::HCCL_E_INTERNAL);
166 :
167 0 : rankSizeLevel0_ = virtRanks_.at(0).size();
168 0 : rankSizeLevel1_ = virtRanks_.at(1).size();
169 0 : rankSize_ = rankSizeLevel0_ * rankSizeLevel1_;
170 :
171 : // 计算当前 rank 在各层级中的索引
172 0 : if (virtRankMap_[0].find(myRank_) != virtRankMap_[0].end()) {
173 0 : rankIdxLevel0_ = virtRankMap_[0][myRank_];
174 : } else {
175 0 : HCCL_ERROR("[CalcLocalRankSize] rank [%d] is not in level 0 topo", myRank_);
176 0 : return HcclResult::HCCL_E_INTERNAL;
177 : }
178 0 : if (virtRankMap_[1].find(myRank_) != virtRankMap_[1].end()) {
179 0 : rankIdxLevel1_ = virtRankMap_[1][myRank_];
180 : } else {
181 0 : HCCL_ERROR("[CalcLocalRankSize] rank [%d] is not in level 1 topo", myRank_);
182 0 : return HcclResult::HCCL_E_INTERNAL;
183 : }
184 :
185 0 : HCCL_INFO("[CalcLocalRankSize] localRankSize: myRank[%d] rankSizeLevel0_[%u] rankSizeLevel1_[%u] "
186 : "rankIdxLevel0_[%llu] rankIdxLevel1_[%llu]",
187 : myRank_, rankSizeLevel0_, rankSizeLevel1_, rankIdxLevel0_, rankIdxLevel1_);
188 0 : return HcclResult::HCCL_SUCCESS;
189 : };
190 :
191 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
192 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::PrepareResForTemplate(
193 : const RankGraph *rankGraph, InsAlgTemplate0 &tempAlgIntraRS, InsAlgTemplate1 &tempAlgInterRS, InsAlgTemplate2 &tempAlgIntraAG, InsAlgTemplate3 &tempAlgInterAG)
194 : {
195 0 : AlgTempResReq resReqIntraRS, resReqInterRS, resReqIntraAG, resReqInterAG;
196 :
197 0 : CHK_RET(tempAlgIntraRS.CalcRes(resReqIntraRS));
198 0 : CHK_RET(tempAlgInterRS.CalcRes(resReqInterRS));
199 0 : CHK_RET(tempAlgIntraAG.CalcRes(resReqIntraAG));
200 0 : CHK_RET(tempAlgInterAG.CalcRes(resReqInterAG));
201 :
202 0 : CHK_RET(CalcQue(resReqIntraRS, resReqInterRS, resReqIntraAG, resReqInterAG));
203 :
204 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, resReqIntraRS.links, intraRSLinks_));
205 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, resReqInterRS.links, interRSLinks_));
206 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, resReqIntraAG.links, intraAGLinks_));
207 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, resReqInterAG.links, interAGLinks_));
208 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] intraRSLinks_ size[%zu], interRSLinks_ size[%zu], intraAGLinks_ size[%zu], interAGLinks_ size[%zu]",
209 : intraRSLinks_.size(), interRSLinks_.size(), intraAGLinks_.size(), interAGLinks_.size());
210 0 : return HCCL_SUCCESS;
211 0 : }
212 :
213 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
214 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::PrepareResForTemplate(ConnectedLinkMgr *linkMgr,
215 : InsAlgTemplate0 &tempAlgIntraRS,
216 : InsAlgTemplate1 &tempAlgInterRS,
217 : InsAlgTemplate2 &tempAlgIntraAG,
218 : InsAlgTemplate3 &tempAlgInterAG)
219 : {
220 0 : AlgTempResReq resReqIntraRS, resReqInterRS, resReqIntraAG, resReqInterAG;
221 :
222 0 : CHK_RET(tempAlgIntraRS.CalcRes(resReqIntraRS));
223 0 : CHK_RET(tempAlgInterRS.CalcRes(resReqInterRS));
224 0 : CHK_RET(tempAlgIntraAG.CalcRes(resReqIntraAG));
225 0 : CHK_RET(tempAlgInterAG.CalcRes(resReqInterAG));
226 :
227 0 : CHK_RET(CalcQue(resReqIntraRS, resReqInterRS, resReqIntraAG, resReqInterAG));
228 :
229 0 : CHK_RET(PrepResLinks(myRank_, resReqIntraRS.links, linkMgr, intraRSLinks_));
230 0 : CHK_RET(PrepResLinks(myRank_, resReqInterRS.links, linkMgr, interRSLinks_));
231 0 : CHK_RET(PrepResLinks(myRank_, resReqIntraAG.links, linkMgr, intraAGLinks_));
232 0 : CHK_RET(PrepResLinks(myRank_, resReqInterAG.links, linkMgr, interAGLinks_));
233 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] intraRSLinks_ size[%zu], interRSLinks_ size[%zu], intraAGLinks_ size[%zu], interAGLinks_ size[%zu]",
234 : intraRSLinks_.size(), interRSLinks_.size(), intraAGLinks_.size(), interAGLinks_.size());
235 0 : return HCCL_SUCCESS;
236 0 : }
237 :
238 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
239 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenRSIntraParams0(
240 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
241 : {
242 0 : u64 sliceCount = dataCount / rankSize_;
243 0 : u64 sliceBytes = sliceCount * dataTypeSize_;
244 0 : u64 tailSize = dataCount * dataTypeSize_ - sliceBytes * (rankSize_ - 1);
245 0 : SetTemplateDataParams(params,
246 : BufferType::INPUT, BufferType::OUTPUT,
247 0 : dataOffset, dataOffset + rankIdxLevel0_ * sliceBytes, scratchOff,
248 : sliceBytes, sliceBytes, sliceBytes,
249 0 : rankSizeLevel1_, sliceBytes * rankSizeLevel0_, sliceBytes * rankSizeLevel0_,
250 : tailSize);
251 0 : }
252 :
253 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
254 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenRSInterParams0(
255 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
256 : {
257 0 : u64 sliceCount = dataCount / rankSize_;
258 0 : u64 sliceBytes = sliceCount * dataTypeSize_;
259 0 : u64 tailSize = dataCount * dataTypeSize_ - sliceBytes * (rankSize_ - 1);
260 0 : SetTemplateDataParams(params,
261 : BufferType::OUTPUT, BufferType::OUTPUT,
262 0 : dataOffset + rankIdxLevel0_ * sliceBytes, dataOffset + myRank_ * sliceBytes, scratchOff,
263 0 : sliceBytes, sliceBytes * rankSizeLevel0_, sliceBytes * rankSizeLevel0_,
264 : 1, 0, 0,
265 : tailSize);
266 0 : }
267 :
268 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
269 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenAGInterParams0(
270 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
271 : {
272 0 : u64 sliceCount = dataCount / rankSize_;
273 0 : u64 sliceBytes = sliceCount * dataTypeSize_;
274 0 : u64 tailSize = dataCount * dataTypeSize_ - sliceBytes * (rankSize_ - 1);
275 0 : SetTemplateDataParams(params,
276 : BufferType::OUTPUT, BufferType::OUTPUT,
277 0 : dataOffset + rankIdxLevel0_ * sliceBytes, dataOffset + rankIdxLevel0_ * sliceBytes, scratchOff,
278 0 : sliceBytes, sliceBytes * rankSizeLevel0_, sliceBytes * rankSizeLevel0_,
279 : 1, 0, 0,
280 : tailSize);
281 0 : }
282 :
283 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
284 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenAGIntraParams0(
285 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
286 : {
287 0 : u64 sliceCount = dataCount / rankSize_;
288 0 : u64 sliceBytes = sliceCount * dataTypeSize_;
289 0 : u64 tailSize = dataCount * dataTypeSize_ - sliceBytes * (rankSize_ - 1);
290 0 : SetTemplateDataParams(params,
291 : BufferType::OUTPUT, BufferType::OUTPUT,
292 : dataOffset, dataOffset, scratchOff,
293 : sliceBytes, sliceBytes, sliceBytes,
294 0 : rankSizeLevel1_, rankSizeLevel0_ * sliceBytes, rankSizeLevel0_ * sliceBytes,
295 : tailSize);
296 0 : }
297 :
298 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
299 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenRSInterParams1(
300 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
301 : {
302 0 : u64 sliceCount = dataCount / rankSize_ * rankSizeLevel0_;
303 0 : u64 sliceBytes = sliceCount * dataTypeSize_;
304 0 : u64 tailSize = dataCount * dataTypeSize_ - sliceBytes * (rankSizeLevel1_ - 1);
305 0 : SetTemplateDataParams(params,
306 : BufferType::INPUT, BufferType::OUTPUT,
307 0 : dataOffset, dataOffset + rankIdxLevel1_ * sliceBytes, scratchOff,
308 : sliceBytes, sliceBytes, sliceBytes,
309 : 1, 0, 0,
310 : tailSize);
311 0 : }
312 :
313 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
314 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenRSIntraParams1(
315 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
316 : {
317 0 : HCCL_INFO("InsAllReduceParallelExecutorV2 GenRSIntraParams1 start");
318 0 : u64 dataCountTmp = CalcDataCountTmp1(dataCount);
319 0 : u64 sliceBytes = sliceCount_ * dataTypeSize_;
320 0 : SetTemplateDataParams(params,
321 : BufferType::OUTPUT, BufferType::OUTPUT,
322 0 : dataOffset + rankIdxLevel1_ * sliceBytes * rankSizeLevel0_, dataOffset + sliceBytes * myRank_, scratchOff,
323 : sliceBytes, sliceBytes, sliceBytes,
324 : 1, 0, 0,
325 0 : (dataCountTmp - sliceCount_ * (rankSizeLevel0_ - 1)) * dataTypeSize_);
326 0 : HCCL_INFO("InsAllReduceParallelExecutorV2 GenRSIntraParams1 end");
327 0 : }
328 :
329 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
330 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenAGIntraParams1(
331 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
332 : {
333 0 : HCCL_INFO("InsAllReduceParallelExecutorV2 GenAGIntraParams1 start");
334 0 : u64 dataCountTmp = CalcDataCountTmp1(dataCount);
335 0 : u64 sliceBytes = sliceCount_ * dataTypeSize_;
336 0 : SetTemplateDataParams(params,
337 : BufferType::OUTPUT, BufferType::OUTPUT,
338 0 : dataOffset + rankIdxLevel1_ * sliceBytes * rankSizeLevel0_,
339 0 : dataOffset + rankIdxLevel1_ * sliceBytes * rankSizeLevel0_,
340 0 : scratchOff + rankIdxLevel1_ * rankSizeLevel0_ * sliceBytes,
341 : sliceBytes, sliceBytes, sliceBytes,
342 : 1, 0, 0,
343 0 : (dataCountTmp - sliceCount_ * (rankSizeLevel0_ - 1)) * dataTypeSize_);
344 0 : HCCL_INFO("InsAllReduceParallelExecutorV2 GenAGIntraParams1 end");
345 0 : }
346 :
347 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
348 0 : void InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenAGInterParams1(
349 : const u64 dataOffset, const u64 dataCount, const u64 scratchOff, TemplateDataParams ¶ms) const
350 : {
351 0 : u64 sliceCount = dataCount / rankSize_ * rankSizeLevel0_;
352 0 : u64 sliceBytes = sliceCount * dataTypeSize_;
353 0 : u64 tailSize = dataCount * dataTypeSize_ - sliceBytes * (rankSizeLevel1_ - 1);
354 0 : SetTemplateDataParams(params,
355 : BufferType::OUTPUT, BufferType::OUTPUT,
356 : dataOffset, dataOffset, scratchOff,
357 : sliceBytes, sliceBytes, sliceBytes,
358 : 1, 0, 0,
359 : tailSize);
360 0 : HCCL_INFO("InsAllReduceParallelExecutorV2 GenAGInterParams1 end");
361 0 : }
362 :
363 : /*
364 : *@Desc: HOST算法编排
365 : */
366 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
367 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::Orchestrate(
368 : const RankGraph *rankGraph, const CollAlgOperator &op, const CollAlgParams ¶ms, InsQuePtr insQue)
369 : {
370 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] Host Orchestrate begins.");
371 :
372 : // init and check params
373 0 : CHK_RET(Init(op, params, insQue));
374 :
375 : // Topo Match
376 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
377 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
378 :
379 0 : CHK_RET(CalcLocalRankSize());
380 :
381 : // 实例化算法模板类
382 0 : InsAlgTemplate0 tempAlgIntraRS(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
383 0 : InsAlgTemplate1 tempAlgInterRS(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
384 0 : InsAlgTemplate2 tempAlgIntraAG(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
385 0 : InsAlgTemplate3 tempAlgInterAG(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
386 :
387 0 : std::vector<map<u32, u32>> rank2PathNumMap;
388 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] Orchestrate SetPathNumMap");
389 0 : CHK_RET(SetPathNumMapByRankGraphMultiLevel(rankGraph, virtRanks_, myRank_, rank2PathNumMap));
390 0 : tempAlgIntraRS.setPathNumMap(rank2PathNumMap[0]);
391 0 : tempAlgInterRS.setPathNumMap(rank2PathNumMap[1]);
392 0 : tempAlgIntraAG.setPathNumMap(rank2PathNumMap[0]);
393 0 : tempAlgInterAG.setPathNumMap(rank2PathNumMap[1]);
394 :
395 0 : InitAlgCommonParams(tempAlgIntraRS, tempAlgInterRS, tempAlgIntraAG, tempAlgInterAG, op);
396 :
397 : // 计算算法模板所需资源
398 0 : CHK_RET(PrepareResForTemplate(rankGraph, tempAlgIntraRS, tempAlgInterRS, tempAlgIntraAG, tempAlgInterAG));
399 0 : CHK_RET(GenInsQues(tempAlgIntraRS, tempAlgInterRS, tempAlgIntraAG, tempAlgInterAG));
400 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] Orchestrate success.");
401 :
402 0 : return HcclResult::HCCL_SUCCESS;
403 0 : }
404 :
405 : /*
406 : *@Desc: AICPU算法编排
407 : */
408 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
409 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::Orchestrate(
410 : const AlgTopoInfo &topoInfo, const CollAlgOperator &op, const CollAlgParams ¶ms, ConnectedLinkMgr *linkMgr,
411 : InsQuePtr insQue)
412 : {
413 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] AICPU Orchestrate begins.");
414 : // init and check params
415 0 : CHK_RET(Init(op, params, insQue));
416 : // 获取当前通信域的信息
417 0 : vTopo_ = topoInfo.vTopo;
418 0 : virtRankMap_ = topoInfo.virtRankMap;
419 0 : virtRanks_ = topoInfo.virtRanks;
420 0 : CHK_RET(CalcLocalRankSize());
421 :
422 : // 实例化算法模板类
423 0 : InsAlgTemplate0 tempAlgIntraRS(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
424 0 : InsAlgTemplate1 tempAlgInterRS(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
425 0 : InsAlgTemplate2 tempAlgIntraAG(myRank_, rankSizeLevel0_, vTopo_[0], virtRankMap_[0]);
426 0 : InsAlgTemplate3 tempAlgInterAG(myRank_, rankSizeLevel1_, vTopo_[1], virtRankMap_[1]);
427 :
428 0 : std::vector<map<u32, u32>> rank2PathNumMap;
429 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] Orchestrate SetPathNumMap");
430 0 : CHK_RET(SetPathNumMapByLinkMgrMultiLevel(linkMgr, virtRanks_, myRank_, rank2PathNumMap));
431 0 : tempAlgIntraRS.setPathNumMap(rank2PathNumMap[0]);
432 0 : tempAlgInterRS.setPathNumMap(rank2PathNumMap[1]);
433 0 : tempAlgIntraAG.setPathNumMap(rank2PathNumMap[0]);
434 0 : tempAlgInterAG.setPathNumMap(rank2PathNumMap[1]);
435 :
436 0 : InitAlgCommonParams(tempAlgIntraRS, tempAlgInterRS, tempAlgIntraAG, tempAlgInterAG, op);
437 :
438 : // 计算算法模板所需资源
439 0 : CHK_RET(PrepareResForTemplate(linkMgr, tempAlgIntraRS, tempAlgInterRS, tempAlgIntraAG, tempAlgInterAG));
440 0 : CHK_RET(GenInsQues(tempAlgIntraRS, tempAlgInterRS, tempAlgIntraAG, tempAlgInterAG));
441 0 : HCCL_INFO("[InsAllReduceParallelExecutorV2] Orchestrate success.");
442 :
443 0 : return HcclResult::HCCL_SUCCESS;
444 0 : }
445 :
446 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1, typename InsAlgTemplate2, typename InsAlgTemplate3>
447 0 : HcclResult InsAllReduceParallelExecutorV2<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1, InsAlgTemplate2, InsAlgTemplate3>::GenInsQues(
448 : InsAlgTemplate0 &tempAlgIntraRS, InsAlgTemplate1 &tempAlgInterRS, InsAlgTemplate2 &tempAlgIntraAG, InsAlgTemplate3 &tempAlgInterAG)
449 : {
450 0 : std::vector<float> dataSplitSize;
451 0 : GetParallelDataSplitRate(dataSplitSize);
452 0 : u64 alignedSize = 16 * 1024; // 16K 对齐
453 0 : u64 UB_DATA_SIZE_LIMIT = static_cast<u64>(UB_MAX_DATA_SIZE) * rankSize_ / rankSizeLevel0_;
454 :
455 0 : u64 dataCount0 = (static_cast<u64>((dataCount_ * dataSplitSize[0])) / rankSize_) * rankSize_;
456 0 : u64 dataCount1 = dataCount_ - dataCount0;
457 :
458 0 : u64 scratchSize0 = (static_cast<u64>(maxTmpMemSize_ * dataSplitSize[0]) / alignedSize / dataTypeSize_) * alignedSize * dataTypeSize_;
459 0 : scratchSize0 = std::min(scratchSize0, UB_DATA_SIZE_LIMIT);
460 0 : u64 maxCountPerLoop0 = scratchSize0 / dataTypeSize_ / rankSize_ * rankSize_;
461 0 : maxCountPerLoop0 = std::min(dataCount0, maxCountPerLoop0);
462 :
463 0 : u64 scratchSize1 = ((maxTmpMemSize_ - scratchSize0) / alignedSize / dataTypeSize_) * alignedSize * dataTypeSize_;
464 0 : scratchSize1 = std::min(scratchSize1, UB_DATA_SIZE_LIMIT);
465 0 : u64 maxCountPerLoop1 = scratchSize1 / dataTypeSize_ / rankSize_ * rankSize_;
466 0 : maxCountPerLoop1 = std::min(dataCount1, maxCountPerLoop1);
467 :
468 0 : u32 loopTimes0 = 0;
469 0 : if (maxCountPerLoop0 != 0) {
470 0 : loopTimes0 = dataCount0 / maxCountPerLoop0 + ((dataCount0 % maxCountPerLoop0 == 0) ? 0 : 1);
471 : }
472 :
473 0 : u32 loopTimes1 = 0;
474 0 : if (maxCountPerLoop1 != 0) {
475 0 : loopTimes1 = dataCount1 / maxCountPerLoop1 + ((dataCount1 % maxCountPerLoop1 == 0) ? 0 : 1);
476 : }
477 0 : u32 loopTimes = std::max(loopTimes0, loopTimes1);
478 :
479 0 : u64 scratchOffset0 = 0;
480 0 : u64 scratchOffset1 = scratchSize0;
481 :
482 0 : TempFuncs tempFuncs;
483 0 : tempFuncs.opMode = opMode_;
484 0 : tempFuncs.enableCounterNotify = false;
485 :
486 0 : TemplateDataParams rsIntra0Params, rsInter1Params;
487 0 : TemplateDataParams rsInter0Params, rsIntra1Params;
488 0 : TemplateDataParams agIntra0Params, agInter1Params;
489 0 : TemplateDataParams agInter0Params, agIntra1Params;
490 :
491 0 : for (u32 loopIndex = 0; loopIndex < loopTimes; loopIndex++) {
492 0 : u64 currCount0 = 0;
493 0 : u64 dataOffset0 = 0;
494 0 : if (loopIndex < loopTimes) {
495 0 : currCount0 = (loopIndex == loopTimes - 1) ? (dataCount0 - loopIndex * maxCountPerLoop0) : maxCountPerLoop0;
496 0 : dataOffset0 = loopIndex * maxCountPerLoop0 * dataTypeSize_;
497 : }
498 :
499 0 : u64 currCount1 = 0;
500 0 : u64 dataOffset1 = 0;
501 0 : if (loopIndex < loopTimes) {
502 0 : currCount1 = (loopIndex == loopTimes - 1) ? (dataCount1 - loopIndex * maxCountPerLoop1) : maxCountPerLoop1;
503 0 : dataOffset1 = loopIndex * maxCountPerLoop1 * dataTypeSize_ + dataCount0 * dataTypeSize_;
504 : }
505 :
506 : // 计算统一sliceCount数量
507 0 : u64 totalSliceCount = currCount1 / rankSize_ * rankSizeLevel0_;
508 0 : sliceCount_ = (currCount1 >= rankSize_)
509 0 : ? totalSliceCount / rankSizeLevel0_
510 0 : : totalSliceCount / rankSize_ * rankSizeLevel1_;
511 :
512 : // ────────────── Phase 1: RS-1 ──────────────
513 : // 前半: 框内 Mesh RS, 后半: 框间 NHR RS
514 0 : CHK_RET(PreSyncQues(syncQueues_, 0));
515 :
516 0 : GenRSIntraParams0(dataOffset0, currCount0, scratchOffset0, rsIntra0Params);
517 0 : CHK_RET(tempAlgIntraRS.GenExtIns(tempFuncs, rsIntra0Params, intraRSLinks_, intraQue_));
518 :
519 0 : GenRSInterParams1(dataOffset1, currCount1, scratchOffset1, rsInter1Params);
520 0 : CHK_RET(tempAlgInterRS.GenExtIns(tempFuncs, rsInter1Params, interRSLinks_, interQue_));
521 :
522 0 : CHK_RET(PostSyncQues(syncQueues_, 0));
523 :
524 : // ────────────── Phase 2: RS-2 ──────────────
525 : // 前半: 框间 NHR RS, 后半: 框内 Mesh RS
526 0 : CHK_RET(PreSyncQues(syncQueues_, 0));
527 :
528 0 : GenRSInterParams0(dataOffset0, currCount0, scratchOffset0, rsInter0Params);
529 0 : CHK_RET(tempAlgInterRS.GenExtIns(tempFuncs, rsInter0Params, interRSLinks_, interQue_));
530 :
531 0 : GenRSIntraParams1(dataOffset1, currCount1, scratchOffset1, rsIntra1Params);
532 0 : CHK_RET(tempAlgIntraRS.GenExtIns(tempFuncs, rsIntra1Params, intraRSLinks_, intraQue_));
533 :
534 0 : CHK_RET(PostSyncQues(syncQueues_, 0));
535 :
536 : // ────────────── Phase 3: AG-1 ──────────────
537 : // 前半: 框间 NHR AG, 后半: 框内 Mesh AG
538 0 : CHK_RET(PreSyncQues(syncQueues_, 0));
539 :
540 0 : GenAGInterParams0(dataOffset0, currCount0, scratchOffset0, agInter0Params);
541 0 : CHK_RET(tempAlgInterAG.GenExtIns(tempFuncs, agInter0Params, interAGLinks_, interQue_));
542 :
543 0 : GenAGIntraParams1(dataOffset1, currCount1, scratchOffset1, agIntra1Params);
544 0 : CHK_RET(tempAlgIntraAG.GenExtIns(tempFuncs, agIntra1Params, intraAGLinks_, intraQue_));
545 :
546 0 : CHK_RET(PostSyncQues(syncQueues_, 0));
547 :
548 : // ────────────── Phase 4: AG-2 ──────────────
549 : // 前半: 框内 Mesh AG, 后半: 框间 NHR AG
550 0 : CHK_RET(PreSyncQues(syncQueues_, 0));
551 :
552 0 : GenAGIntraParams0(dataOffset0, currCount0, scratchOffset0, agIntra0Params);
553 0 : CHK_RET(tempAlgIntraAG.GenExtIns(tempFuncs, agIntra0Params, intraAGLinks_, intraQue_));
554 :
555 0 : GenAGInterParams1(dataOffset1, currCount1, scratchOffset1, agInter1Params);
556 0 : CHK_RET(tempAlgInterAG.GenExtIns(tempFuncs, agInter1Params, interAGLinks_, interQue_));
557 :
558 0 : CHK_RET(PostSyncQues(syncQueues_, 0));
559 : }
560 :
561 0 : return HcclResult::HCCL_SUCCESS;
562 0 : }
563 :
564 : // 算法注册
565 : INS_REGISTER_IMPL_BY_FOUR_TEMPS(OpType::ALLREDUCE, InsAllReduceFourTemplateMesh1DNHR, InsAllReduceParallelExecutorV2,
566 : TopoMatchMeshNHR, InsTempReduceScatterMesh1D, InsTempReduceScatterMesh1D, InsTempAllGatherMesh1D, InsTempAllGatherMesh1D);
567 : }
|
