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_reduce_parallel_executor.h"
12 : #include <cmath>
13 : #include "log.h"
14 : #include "ins_coll_alg_registry.h"
15 : #include "topo_match_mesh_nhr.h"
16 : #include "topo_match_mesh_nhr_pcie.h"
17 : #include "alg_data_trans_wrapper.h"
18 : #include "ins_temp_reduce_nhr.h"
19 : #include "ins_temp_reduce_mesh_1D.h"
20 : #include "ccu_temp_reduce_nhr_1D_mem2mem.h"
21 : #include "ccu_temp_reduce_mesh_1D_mem2mem.h"
22 :
23 : namespace Hccl {
24 : constexpr u64 MAX_OFFLOAD_SCRATCH_SIZE = 200 * 1024 * 1024; // 200M
25 :
26 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
27 0 : InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::InsReduceParallelExecutor()
28 0 : : InsCollAlgBase()
29 : {
30 0 : }
31 :
32 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
33 0 : InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::~InsReduceParallelExecutor()
34 : {
35 0 : }
36 :
37 :
38 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
39 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::CalcResOffload(
40 : const RankGraph *rankGraph, const u64 &dataSize, CollOffloadOpResReq &resReq)
41 : {
42 0 : HCCL_INFO("[InsReduceParallelExecutor] CalcResOffload begins.");
43 : (void)dataSize;
44 0 : u64 scratchMemSize = MAX_OFFLOAD_SCRATCH_SIZE;
45 0 : resReq.requiredScratchMemSize = scratchMemSize; // 200MB
46 : // Topo Match
47 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
48 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
49 0 : CHK_RET(CalcLocalRankSize(myRank_, virtRanks_, intraLocalRankSize_, interLocalRankSize_));
50 0 : InsAlgTemplate0 intraTempAlg(myRank_, intraLocalRankSize_, vTopo_[0], virtRankMap_[0]);
51 0 : InsAlgTemplate1 interTempAlg(myRank_, interLocalRankSize_, vTopo_[1], virtRankMap_[1]);
52 :
53 : // calculate required insQues and prepare queue
54 0 : AlgTempResReq resReqIntra;
55 0 : AlgTempResReq resReqInter;
56 0 : if (enableDetour_) {
57 0 : HCCL_DEBUG("[InsReduceParallelExecutor] Rank[%d], CalcRes with detouring enabled.", myRank_);
58 0 : CHK_RET(intraTempAlg.CalcResDetour(rankGraph, resReqIntra));
59 : } else {
60 0 : HCCL_DEBUG("[InsReduceParallelExecutor] Rank[%d], CalcRes with detouring disabled.", myRank_);
61 0 : CHK_RET(intraTempAlg.CalcRes(resReqIntra));
62 : }
63 :
64 0 : CHK_RET(interTempAlg.CalcRes(resReqInter));
65 0 : resReq.requiredSubQueNum = resReqIntra.streamNum + resReqInter.streamNum - 1;
66 0 : return HcclResult::HCCL_SUCCESS;
67 0 : }
68 :
69 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
70 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::CalcRes(
71 : const RankGraph *rankGraph, CollAlgResReq &algResReq)
72 : {
73 0 : HCCL_INFO("[InsReduceParallelExecutor] CalcRes begins.");
74 : // Topo Match
75 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
76 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
77 0 : algResReq.topoInfo.UpdateMultiLevelTopo(virtRanks_, virtRankMap_, vTopo_);
78 0 : CHK_RET(CalcLocalRankSize(myRank_, virtRanks_, intraLocalRankSize_, interLocalRankSize_));
79 :
80 : // instantiate a template
81 0 : InsAlgTemplate0 intraTempAlg(myRank_, intraLocalRankSize_, vTopo_[0], virtRankMap_[0]);
82 0 : InsAlgTemplate1 interTempAlg(myRank_, interLocalRankSize_, vTopo_[1], virtRankMap_[1]);
83 :
84 : // calculate required insQues and prepare queue
85 0 : AlgTempResReq resReqIntra;
86 0 : AlgTempResReq resReqInter;
87 0 : if (enableDetour_) {
88 0 : HCCL_DEBUG("[InsReduceParallelExecutor] Rank[%d], CalcRes with detouring enabled.", myRank_);
89 0 : CHK_RET(intraTempAlg.CalcResDetour(rankGraph, resReqIntra));
90 : } else {
91 0 : HCCL_DEBUG("[InsReduceParallelExecutor] Rank[%d], CalcRes with detouring disabled.", myRank_);
92 0 : CHK_RET(intraTempAlg.CalcRes(resReqIntra));
93 : }
94 0 : CHK_RET(interTempAlg.CalcRes(resReqInter));
95 :
96 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, resReqIntra.links, algResReq.levelRankPairs));
97 0 : CHK_RET(CalcLinkInfo(myRank_, rankGraph, resReqInter.links, algResReq.levelRankPairs));
98 0 : algResReq.primQueueNum = resReqIntra.streamNum + resReqInter.streamNum;
99 0 : std::vector<std::tuple<QId, QId, u32>> notifyRequests;
100 :
101 0 : u32 slaveNum = algResReq.primQueueNum - 1;
102 0 : notifyRequests.reserve(slaveNum); //每个从流需要1个
103 0 : for (QId q = 1; q < algResReq.primQueueNum; q++) {
104 0 : notifyRequests.emplace_back(std::make_tuple(0, q, 0));
105 0 : notifyRequests.emplace_back(std::make_tuple(q, 0, 0));
106 : }
107 :
108 : // nhr算法只有一个stream
109 0 : for (QId q = resReqIntra.streamNum; q < algResReq.primQueueNum; q++) {
110 0 : if(resReqIntra.streamNum == q){
111 0 : continue;
112 : }
113 0 : notifyRequests.emplace_back(std::make_tuple(resReqIntra.streamNum, q, 0));
114 0 : notifyRequests.emplace_back(std::make_tuple(q, resReqIntra.streamNum, 0));
115 0 : HCCL_DEBUG("[InsReduceParallelExecutor] CalcRes notifyRequests:%u->%u. %u->%u",
116 : resReqIntra.streamNum, q, q, resReqIntra.streamNum);
117 : }
118 :
119 0 : algResReq.queueNotifys = notifyRequests;
120 0 : HCCL_DEBUG("[InsReduceParallelExecutor] algResReq.primQueueNum %u", algResReq.primQueueNum);
121 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, resReqIntra.links, algResReq.links));
122 0 : CHK_RET(CalcResLinks(myRank_, rankGraph, linkPriority_, resReqInter.links, algResReq.links));
123 :
124 0 : return HcclResult::HCCL_SUCCESS;
125 0 : }
126 :
127 : // HOST 侧算法入口，将对应的 instruction 添加到指令队列中
128 : // 传入的insQue为一条主流
129 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
130 0 : void InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::GenTemplateAlgParams0(
131 : const u64 dataOffset, const u64 dataCount, const u64 scratchOffset, TemplateDataParams &tempAlgParams) const
132 : {
133 0 : tempAlgParams.buffInfo.inBuffType = BufferType::INPUT;
134 0 : tempAlgParams.buffInfo.outBuffType = BufferType::OUTPUT;
135 0 : tempAlgParams.buffInfo.scratBuffType = BufferType::SCRATCH;
136 0 : tempAlgParams.buffInfo.inBuffBaseOff = dataOffset;
137 0 : tempAlgParams.buffInfo.outBuffBaseOff = dataOffset;
138 0 : tempAlgParams.buffInfo.scratchBuffBaseOff = scratchOffset;
139 0 : tempAlgParams.sliceSize = dataCount * dataTypeSize_;
140 0 : tempAlgParams.tailSize = tempAlgParams.sliceSize;
141 0 : tempAlgParams.inputSliceStride = 0; // 输入数据仅有 1 个 slice, 不需要 stride
142 0 : tempAlgParams.outputSliceStride = 0;
143 0 : tempAlgParams.repeatNum = 1;
144 0 : return;
145 : }
146 :
147 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
148 0 : void InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::GenTemplateAlgParams1(
149 : const u64 dataOffset, const u64 dataCount, const u64 scratchOffset, TemplateDataParams &tempAlgParams) const
150 : {
151 0 : tempAlgParams.buffInfo.inBuffType = BufferType::OUTPUT;
152 0 : tempAlgParams.buffInfo.outBuffType = BufferType::OUTPUT;
153 0 : tempAlgParams.buffInfo.scratBuffType = BufferType::SCRATCH;
154 0 : tempAlgParams.buffInfo.inBuffBaseOff = dataOffset;
155 0 : tempAlgParams.buffInfo.outBuffBaseOff = dataOffset;
156 0 : tempAlgParams.buffInfo.scratchBuffBaseOff = scratchOffset;
157 0 : tempAlgParams.sliceSize = dataCount * dataTypeSize_;
158 0 : tempAlgParams.tailSize = tempAlgParams.sliceSize;
159 0 : tempAlgParams.inputSliceStride = 0; // 输入数据仅有 1 个 slice, 不需要 stride
160 0 : tempAlgParams.outputSliceStride = 0;
161 0 : tempAlgParams.repeatNum = 1;
162 0 : return;
163 : }
164 :
165 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
166 0 : void InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::GetParallelDataSplitRate(
167 : std::vector<float> &splitDataSize) const
168 : {
169 : // 先做等分，后续根据性能做调整
170 0 : double splitData = 0.5;
171 0 : splitDataSize.push_back(static_cast<float>(splitData));
172 0 : splitDataSize.push_back(static_cast<float>(splitData));
173 0 : return;
174 : }
175 :
176 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
177 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::CalcLocalRoot()
178 : {
179 0 : CHK_PRT_RET(root_ >= rankSize_,
180 : HCCL_ERROR("[CalcLocalRoot] root[%u] is out of rankSize[%u]", root_, rankSize_),
181 : HcclResult::HCCL_E_INTERNAL);
182 :
183 0 : u32 intraLocalRootIdx = root_ % intraLocalRankSize_;
184 0 : intraLocalRoot_ = static_cast<u32>(vTopo_.at(0).at(0).at(intraLocalRootIdx));
185 0 : u32 interLocalRootIdx = root_ / intraLocalRankSize_;
186 0 : interLocalRoot_ = static_cast<u32>(vTopo_.at(1).at(0).at(interLocalRootIdx));
187 :
188 0 : HCCL_INFO("[CalcLocalRoot] localRoot: myRank[%d] intraLocalRoot[%u] interLocalRoot[%u]",
189 : myRank_, intraLocalRoot_, interLocalRoot_);
190 0 : return HcclResult::HCCL_SUCCESS;
191 : }
192 :
193 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
194 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::PrepareResForTemplate(
195 : const RankGraph *rankGraph, InsAlgTemplate0 &tempAlgIntra, InsAlgTemplate1 &tempAlgInter)
196 : {
197 0 : AlgTempResReq resReqIntra;
198 0 : AlgTempResReq resReqInter;
199 0 : if (enableDetour_) {
200 0 : HCCL_DEBUG("[%s] Rank[%d], detouring enabled.", __func__, myRank_);
201 0 : CHK_RET(tempAlgIntra.CalcResDetour(rankGraph, resReqIntra));
202 : } else {
203 0 : HCCL_DEBUG("[%s] Rank[%d], detouring disabled.", __func__, myRank_);
204 0 : CHK_RET(tempAlgIntra.CalcRes(resReqIntra));
205 : }
206 0 : CHK_RET(tempAlgInter.CalcRes(resReqInter));
207 :
208 : // 申请算法模板所需资源
209 0 : if(!(resReqIntra.queNum > 0 && resReqInter.queNum > 0)) {
210 0 : HCCL_ERROR("[InsReduceParallelExecutor]resReqIntra.queNum and resReqInter.queNum must larger than 0.");
211 0 : return HcclResult::HCCL_E_INTERNAL;
212 : }
213 0 : u32 totalQueueNum = resReqIntra.queNum + resReqInter.queNum;
214 0 : CHK_RET(InitQueue(totalQueueNum, reqQue_));
215 0 : for(u32 i = 0 ; i < reqQue_.size(); i++) {
216 0 : if (i < resReqIntra.queNum) {
217 0 : intraQue_.push_back(reqQue_[i]);
218 : } else {
219 0 : interQue_.push_back(reqQue_[i]);
220 : }
221 : }
222 0 : syncQueues_.emplace_back(intraQue_[0]);
223 0 : syncQueues_.emplace_back(interQue_[0]);
224 :
225 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, resReqIntra.links, intraLinks_));
226 0 : CHK_RET(PrepResLinks(myRank_, rankGraph, linkPriority_, resReqInter.links, interLinks_));
227 0 : HCCL_INFO("[InsReduceParallelExecutor] intraLinks_ size[%zu], interLinks_ size[%zu]",
228 : intraLinks_.size(), interLinks_.size());
229 0 : return HCCL_SUCCESS;
230 0 : }
231 :
232 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
233 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::PrepareResForTemplate(
234 : ConnectedLinkMgr *linkMgr, InsAlgTemplate0 &tempAlgIntra, InsAlgTemplate1 &tempAlgInter)
235 : {
236 0 : AlgTempResReq resReqIntra;
237 0 : AlgTempResReq resReqInter;
238 0 : if (enableDetour_) {
239 0 : HCCL_DEBUG("[%s] Rank[%d], CalcRes with detour enabled", __func__, myRank_);
240 0 : CHK_RET(tempAlgIntra.CalcResDetour(linkMgr, resReqIntra));
241 : } else {
242 0 : CHK_RET(tempAlgIntra.CalcRes(resReqIntra));
243 : }
244 0 : CHK_RET(tempAlgInter.CalcRes(resReqInter));
245 :
246 : // 申请算法模板所需资源
247 0 : if(!(resReqIntra.queNum > 0 && resReqInter.queNum > 0)) {
248 0 : HCCL_ERROR("[InsReduceParallelExecutor]resReqIntra.queNum and resReqInter.queNum must > 0.");
249 0 : return HcclResult::HCCL_E_INTERNAL;
250 : }
251 0 : u32 totalQueueNum = resReqIntra.queNum + resReqInter.queNum;
252 0 : CHK_RET(InitQueue(totalQueueNum, reqQue_));
253 0 : for(u32 i = 0 ; i < reqQue_.size(); i++) {
254 0 : if (i < resReqIntra.queNum) {
255 0 : intraQue_.push_back(reqQue_[i]);
256 : } else {
257 0 : interQue_.push_back(reqQue_[i]);
258 : }
259 : }
260 0 : syncQueues_.emplace_back(intraQue_[0]);
261 0 : syncQueues_.emplace_back(interQue_[0]);
262 :
263 0 : CHK_RET(PrepResLinks(myRank_, resReqIntra.links, linkMgr, intraLinks_));
264 0 : CHK_RET(PrepResLinks(myRank_, resReqInter.links, linkMgr, interLinks_));
265 0 : HCCL_INFO("[InsReduceParallelExecutor] intraLinks_ size[%zu], interLinks_ size[%zu]", intraLinks_.size(), interLinks_.size());
266 0 : return HCCL_SUCCESS;
267 0 : }
268 :
269 : // Aicpu展开
270 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
271 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::Orchestrate(
272 : const AlgTopoInfo &topoInfo, const CollAlgOperator &op, const CollAlgParams ¶ms, ConnectedLinkMgr *linkMgr,
273 : InsQuePtr insQue)
274 : {
275 0 : HCCL_INFO("[InsReduceParallelExecutor] AICPU Orchestrate begins.");
276 : // init and check params
277 0 : CHK_RET(Init(op, params, insQue));
278 : // 所以获取取级通信域的信息
279 0 : vTopo_ = topoInfo.vTopo; // 本通信域内的通信平面
280 0 : virtRankMap_ = topoInfo.virtRankMap; // 本通信域内的 rank 映射表
281 0 : virtRanks_ = topoInfo.virtRanks; // 本通信域内的 rank 集合
282 :
283 : // 计算localRankSize和localRoot
284 0 : CHK_RET(CalcLocalRankSize(myRank_, virtRanks_, intraLocalRankSize_, interLocalRankSize_));
285 0 : CHK_RET(CalcLocalRoot());
286 :
287 : // 实例化算法模板类
288 0 : InsAlgTemplate0 tempAlgIntra(myRank_, intraLocalRankSize_, vTopo_[0], virtRankMap_[0]); //server内算法，比如mesh
289 0 : InsAlgTemplate1 tempAlgInter(myRank_, interLocalRankSize_, vTopo_[1], virtRankMap_[1]); //server间算法，比如nhr
290 :
291 0 : tempAlgIntra.SetDmaMode(dmaMode_);
292 0 : tempAlgIntra.InitReduceInfo(redOp_, dataType_);
293 0 : tempAlgIntra.SetRoot(intraLocalRoot_);
294 0 : tempAlgIntra.SetCollOp(op);
295 :
296 0 : tempAlgInter.SetDmaMode(dmaMode_);
297 0 : tempAlgInter.InitReduceInfo(redOp_, dataType_);
298 0 : tempAlgInter.SetRoot(interLocalRoot_);
299 0 : tempAlgInter.SetCollOp(op);
300 :
301 : // 计算算法模板所需资源
302 0 : CHK_RET(PrepareResForTemplate(linkMgr, tempAlgIntra, tempAlgInter));
303 0 : CHK_RET(GenInsQues(tempAlgIntra, tempAlgInter));
304 0 : HCCL_INFO("[InsReduceParallelExecutor] AICPU Orchestrate success.");
305 0 : return HcclResult::HCCL_SUCCESS;
306 0 : }
307 :
308 : // Host展开
309 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
310 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::Orchestrate(
311 : const RankGraph *rankGraph, const CollAlgOperator &op, const CollAlgParams ¶ms, InsQuePtr insQue)
312 : {
313 0 : HCCL_INFO("[InsReduceParallelExecutor] Host Orchestrate begins.");
314 : // init and check params
315 0 : CHK_RET(Init(op, params, insQue));
316 :
317 : // Topo Match
318 0 : AlgTopoMatch topoMatch(myRank_, rankSize_, rankGraph, devType_);
319 0 : CHK_RET(topoMatch.MatchTopo(vTopo_, virtRanks_, virtRankMap_));
320 :
321 : // 计算localRankSize和localRoot
322 0 : CHK_RET(CalcLocalRankSize(myRank_, virtRanks_, intraLocalRankSize_, interLocalRankSize_));
323 0 : CHK_RET(CalcLocalRoot());
324 :
325 : // 实例化算法模板类
326 0 : InsAlgTemplate0 tempAlgIntra(myRank_, intraLocalRankSize_, vTopo_[0], virtRankMap_[0]); //server内算法，比如mesh
327 0 : InsAlgTemplate1 tempAlgInter(myRank_, interLocalRankSize_, vTopo_[1], virtRankMap_[1]); //server间算法，比如nhr
328 :
329 0 : tempAlgIntra.SetDmaMode(dmaMode_);
330 0 : tempAlgIntra.SetCollOp(op);
331 0 : tempAlgIntra.InitReduceInfo(redOp_, dataType_);
332 0 : tempAlgIntra.SetRoot(intraLocalRoot_);
333 :
334 0 : tempAlgInter.SetDmaMode(dmaMode_);
335 0 : tempAlgInter.InitReduceInfo(redOp_, dataType_);
336 0 : tempAlgInter.SetCollOp(op);
337 0 : tempAlgInter.SetRoot(interLocalRoot_);
338 :
339 : // 计算算法模板所需资源
340 0 : CHK_RET(PrepareResForTemplate(rankGraph, tempAlgIntra, tempAlgInter));
341 0 : CHK_RET(GenInsQues(tempAlgIntra, tempAlgInter));
342 0 : HCCL_INFO("[InsReduceParallelExecutor] Host Orchestrate success.");
343 0 : return HcclResult::HCCL_SUCCESS;
344 0 : }
345 :
346 : template <typename AlgTopoMatch, typename InsAlgTemplate0, typename InsAlgTemplate1>
347 0 : HcclResult InsReduceParallelExecutor<AlgTopoMatch, InsAlgTemplate0, InsAlgTemplate1>::GenInsQues(
348 : InsAlgTemplate0 &tempAlgIntra, InsAlgTemplate1 &tempAlgInter)
349 : {
350 0 : std::vector<float> dataSplitSize;
351 0 : GetParallelDataSplitRate(dataSplitSize);
352 0 : u64 alignedSize = 16 * 1024; //假设需要16K对齐
353 0 : BufferType inBuffType = BufferType::INPUT;
354 0 : BufferType outBuffType = BufferType::OUTPUT;
355 0 : u32 intraScatchteMultipleStage0 = tempAlgIntra.CalcScratchMultiple(inBuffType, outBuffType);
356 0 : u32 interScatchteMultipleStage0 = tempAlgInter.CalcScratchMultiple(inBuffType, outBuffType);
357 0 : u32 intraScatchteMultipleStage1 = tempAlgIntra.CalcScratchMultiple(outBuffType, outBuffType);
358 0 : u32 interScatchteMultipleStage1 = tempAlgInter.CalcScratchMultiple(outBuffType, outBuffType);
359 0 : u32 scratchMultipleIntra = static_cast<u32>(std::max(std::ceil(dataSplitSize[0] * intraScatchteMultipleStage0),
360 0 : std::ceil(dataSplitSize[1] * intraScatchteMultipleStage1)));
361 0 : u32 scratchMultipleInter = static_cast<u32>(std::max(std::ceil(dataSplitSize[1] * interScatchteMultipleStage0),
362 0 : std::ceil(dataSplitSize[0] * interScatchteMultipleStage1)));
363 0 : u32 totalScratchMultiple = scratchMultipleIntra + scratchMultipleInter;
364 0 : u64 scratchMemBlockSize = maxTmpMemSize_;
365 0 : if (totalScratchMultiple > 0) {
366 0 : scratchMemBlockSize = (maxTmpMemSize_ / alignedSize / totalScratchMultiple) * alignedSize;
367 : }
368 0 : u64 intraScratchOffset = 0;
369 0 : u64 interScratchOffset = scratchMultipleIntra * scratchMemBlockSize;
370 :
371 : // dataSplitSize为分数，这里maxCountPerLoop对10取整
372 0 : u64 maxCountPerLoop = (std::min(static_cast<u64>(scratchMemBlockSize),
373 0 : static_cast<u64>(UB_MAX_DATA_SIZE)) / dataTypeSize_ / 10) * 10;
374 :
375 0 : u32 loopTimes = dataCount_ / maxCountPerLoop + ((dataCount_ % maxCountPerLoop == 0) ? 0 : 1);
376 :
377 0 : TemplateDataParams tempAlgParamsIntra0, tempAlgParamsInter0;
378 0 : TemplateDataParams tempAlgParamsInter1, tempAlgParamsIntra1;
379 0 : TempFuncs tempFuncs;
380 0 : tempFuncs.opMode = opMode_;
381 0 : tempFuncs.enableCounterNotify = false;
382 0 : tempFuncs.isBottom = true;
383 0 : tempFuncs.isForepart = true;
384 0 : for (u32 loopIndex = 0; loopIndex < loopTimes; loopIndex++) {
385 0 : u64 currCount = (loopIndex == loopTimes - 1) ? (dataCount_ - loopIndex * maxCountPerLoop) : maxCountPerLoop;
386 0 : u64 dataCountPerLoopAixs0 = static_cast<u64>(dataSplitSize[0] * currCount);
387 0 : u64 dataCountPerLoopAixs1 = currCount - dataCountPerLoopAixs0;
388 : //第一步开始前同步
389 :
390 0 : CHK_RET(PreSyncQues(syncQueues_, 0));
391 0 : u64 dataOffset0 = loopIndex * maxCountPerLoop * dataTypeSize_;
392 0 : u64 dataOffset1 = dataOffset0 + dataCountPerLoopAixs0 * dataTypeSize_;
393 : //数据0的server内的mesh算法
394 0 : GenTemplateAlgParams0(dataOffset0, dataCountPerLoopAixs0, intraScratchOffset, tempAlgParamsIntra0);
395 : //把每个template需要的queue传进去，比如stars的mesh要传多条queue
396 0 : CHK_RET(tempAlgIntra.GenExtIns(tempFuncs, tempAlgParamsIntra0, intraLinks_, intraQue_));
397 : //数据1的server间的nhr算法
398 0 : GenTemplateAlgParams0(dataOffset1, dataCountPerLoopAixs1, interScratchOffset, tempAlgParamsInter1);
399 0 : CHK_RET(tempAlgInter.GenExtIns(tempFuncs, tempAlgParamsInter1, interLinks_, interQue_));
400 : //第一步做完后回到主流做尾同步
401 0 : CHK_RET(PostSyncQues(syncQueues_, 0));
402 : // 只有真正root节点的横纵坐标所在的卡，需要做第二步骤，担任过其中一个root节点的，只需要负责发就行了
403 0 : if ((static_cast<u32>(myRank_) != intraLocalRoot_) && (static_cast<u32>(myRank_) != interLocalRoot_)) {
404 0 : continue;
405 : }
406 :
407 : //第二步开始前同步
408 0 : CHK_RET(PreSyncQues(syncQueues_, 0));
409 0 : if (static_cast<u32>(myRank_) == intraLocalRoot_) {
410 : //数据0的server间的nhr算法
411 0 : GenTemplateAlgParams1(dataOffset0, dataCountPerLoopAixs0, interScratchOffset, tempAlgParamsInter0);
412 0 : CHK_RET(tempAlgInter.GenExtIns(tempFuncs, tempAlgParamsInter0, interLinks_, interQue_));
413 : }
414 0 : if (static_cast<u32>(myRank_) == interLocalRoot_) {
415 : //数据1的server内的mesh算法
416 0 : GenTemplateAlgParams1(dataOffset1, dataCountPerLoopAixs1, intraScratchOffset, tempAlgParamsIntra1);
417 0 : CHK_RET(tempAlgIntra.GenExtIns(tempFuncs, tempAlgParamsIntra1, intraLinks_, intraQue_));
418 : }
419 : //尾同步
420 0 : CHK_RET(PostSyncQues(syncQueues_, 0));
421 : }
422 0 : return HcclResult::HCCL_SUCCESS;
423 0 : }
424 :
425 : // 算法注册
426 : INS_REGISTER_IMPL_BY_TWO_TEMPS(OpType::REDUCE, InsReduceParallelMesh1DNHR, InsReduceParallelExecutor, TopoMatchMeshNHR,
427 : InsTempReduceMesh1D, InsTempReduceNHR);
428 : INS_REGISTER_IMPL_BY_TWO_TEMPS(OpType::REDUCE, InsReduceParallelMesh1DNHRPcie, InsReduceParallelExecutor,
429 : TopoMatchMeshNHRPcie, InsTempReduceMesh1D, InsTempReduceNHR);
430 : #ifndef CCL_KERNEL_AICPU
431 : INS_REGISTER_IMPL_BY_TWO_TEMPS(OpType::REDUCE, CcuReduceParallelMesh1DNHR, InsReduceParallelExecutor, TopoMatchMeshNHR,
432 : CcuTempReduceMeshMem2Mem1D, CcuTempReduceNHRMem2Mem1D);
433 : #endif
434 : }
|
