LCOV - code coverage report
Current view: top level - legacy/ascend950/service/collective/alg/coll_alg_factory/alg_template/ins_alg_template - ins_temp_reduce_scatter_mesh_1D.cc (source / functions) Coverage Total Hit
Test: coverage.info Lines: 0.0 % 117 0
Test Date: 2026-08-04 10:52:23 Functions: 0.0 % 9 0

            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_temp_reduce_scatter_mesh_1D.h"
      12              : #include "log.h"
      13              : #include "alg_data_trans_wrapper.h"
      14              : 
      15              : namespace Hccl {
      16            0 : InsTempReduceScatterMesh1D::InsTempReduceScatterMesh1D(const RankId virtualRank, const u32 tempRankSize,
      17            0 :     const std::vector<std::vector<RankId>> &tempVTopo, const std::map<RankId, u32> &tempVirtRankMap)
      18            0 :     : InsAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
      19            0 : {}
      20              : 
      21            0 : InsTempReduceScatterMesh1D::~InsTempReduceScatterMesh1D()
      22            0 : {}
      23              : 
      24            0 : HcclResult InsTempReduceScatterMesh1D::CalcRes(AlgTempResReq &tempResReq)
      25              : {
      26            0 :     CHK_PRT_RET(
      27              :         CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq) != HcclResult::HCCL_SUCCESS,
      28              :         HCCL_ERROR("[CollAlgFactory] [InsTempReduceScatterMesh1D] Rank [%d], resLinks calculation error!", myRank_),
      29              :         HcclResult::HCCL_E_INTERNAL);
      30            0 :     auto &linkReq = tempResReq.links;
      31            0 :     u32 pathNum = 0;
      32            0 :     for (auto resReqIter = linkReq.begin(); resReqIter != linkReq.end(); resReqIter++) {
      33            0 :         auto remoteRank = resReqIter->first;
      34            0 :         if (rank2PathNumMap_.find(remoteRank) == rank2PathNumMap_.end() || rank2PathNumMap_[remoteRank] == 0) {
      35            0 :             HCCL_ERROR("[InsTempReduceScatterMesh1D] No path to remoteRank[%d]", remoteRank);
      36            0 :             return HcclResult::HCCL_E_INTERNAL;
      37              :         }
      38            0 :         if (pathNum == 0) {
      39            0 :             pathNum = rank2PathNumMap_[remoteRank];
      40            0 :         } else if (rank2PathNumMap_[remoteRank] != pathNum) {
      41            0 :             HCCL_ERROR("[InsTempReduceScatterMesh1D] Inconsistency pathNum to remoteRanks, Previous consistent pathNum=[%u], mismatched "
      42              :                        "remoteRank=[%d], pathNum=[%u]",
      43              :                 pathNum,
      44              :                 remoteRank,
      45              :                 rank2PathNumMap_[remoteRank]);
      46            0 :             return HcclResult::HCCL_E_INTERNAL;
      47              :         }
      48            0 :         resReqIter->second = pathNum;
      49              :     }
      50              : 
      51              :     // Mesh 需要的 que Num 为 tempVTopo_[0].size()-1
      52            0 :     tempResReq.queNum = (tempVTopo_[0].size() > 1) ? (tempVTopo_[0].size()) * pathNum: pathNum;
      53            0 :     tempResReq.streamNum = tempResReq.queNum;
      54            0 :     tempResReq.queNotifys = CreateMasterSlaveQueNotifiesRequest(tempResReq.queNum);
      55            0 :     QId centerQ = 0;
      56            0 :     tempResReq.localWaitGroupCntNotify.emplace_back(centerQ, 0);
      57            0 :     tempResReq.localBcastPostCntNotify.emplace_back(centerQ, 0);
      58              : 
      59            0 :     return HcclResult::HCCL_SUCCESS;
      60              : }
      61              : 
      62            0 : u64 InsTempReduceScatterMesh1D::CalcScratchMultiple(const BufferType &inBuffType, const BufferType &outBuffType) const
      63              : {
      64              :     (void)inBuffType;
      65              :     (void)outBuffType;
      66            0 :     u64 scratchMultiple = tempRankSize_;
      67            0 :     return scratchMultiple;
      68              : }
      69              : 
      70            0 : HcclResult InsTempReduceScatterMesh1D::GenExtIns(const TempFuncs &tempFuncs, const TemplateDataParams &tempAlgParams,
      71              :     const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      72              : {
      73            0 :     opMode_ = tempFuncs.opMode;
      74            0 :     enableCounterNotify_ = tempFuncs.enableCounterNotify;
      75            0 :     queNum_ = tempInsQues.size();
      76            0 :     HCCL_INFO("[InsTempReduceScatterMesh1D] Run Start");
      77            0 :     uint32_t linkNum = tempLinks.begin()->second.size();
      78            0 :     CHK_PRT_RET(linkNum > tempInsQues.size(), HCCL_ERROR("[CollAlgFactory] [InsTempReduceScatterMesh1D] Rank [%d], requiredQue Error.", myRank_),
      79              :             HcclResult::HCCL_E_INTERNAL);
      80              : 
      81            0 :     if (queNum_ > 1) {
      82            0 :         CHK_RET(PreSyncInterQueues(tempInsQues));
      83              :     }
      84            0 :     CHK_RET(RunReduceScatter(tempLinks, tempInsQues, tempAlgParams));
      85            0 :     HCCL_INFO("[InsTempReduceScatterMesh1D][PostCopy] Rank [%d].", myRank_);
      86              :     // 流间后同步,从流通知主流
      87            0 :     if (queNum_ > 1) {
      88            0 :         CHK_RET(PostSyncInterQueues(tempInsQues));
      89              :     }
      90            0 :     PostCopy(tempAlgParams, tempInsQues);
      91            0 :     return HcclResult::HCCL_SUCCESS;
      92              : }
      93              : 
      94            0 : HcclResult InsTempReduceScatterMesh1D::PostCopy(
      95              :     const TemplateDataParams &tempAlgParams, std::vector<InsQuePtr> &tempInsQues)
      96              : {
      97              :     // 通信结束之后,数据都在 inbuff 上,需要搬运到对应的输出位置。
      98            0 :     u32 rankIdx = tempVirtRankMap_[myRank_];
      99              :     // 如果是单算子模式, 并且是最后一步算子,需要将数据从 inBuff 拷贝到 userOut
     100              :     // 是否需要将数据搬运到 OutBuff 上再搬运到 UserOut 上??
     101            0 :     HCCL_INFO("[InsTempReduceScatterMesh1D][PostCopy], copy from outBuff to userOut");
     102              :     // 先把本卡的数据从input搬运到output
     103            0 :     HCCL_INFO("[InsTempReduceScatterMesh1D][PostCopy]tempAlgParams.repeatNum=%llu", tempAlgParams.repeatNum);
     104            0 :     u64 sliceSize = ((rankIdx == tempRankSize_ - 1) && (tempAlgParams.tailSize != 0)) ? tempAlgParams.tailSize : tempAlgParams.sliceSize;
     105            0 :     for (u32 repeatIdx = 0; repeatIdx < tempAlgParams.repeatNum; repeatIdx++) {
     106            0 :         DataSlice myRankSlice = DataSlice(tempAlgParams.buffInfo.inBuffType, tempAlgParams.buffInfo.inBuffBaseOff +
     107            0 :             repeatIdx * tempAlgParams.inputRepeatStride + rankIdx * tempAlgParams.inputSliceStride, sliceSize);
     108            0 :         DataSlice outputSlice = DataSlice(tempAlgParams.buffInfo.outBuffType, tempAlgParams.buffInfo.outBuffBaseOff + 
     109            0 :             repeatIdx * tempAlgParams.outputRepeatStride, sliceSize);
     110              :         // myRankSlice与outputSlice一致就不进行LocalCopy
     111            0 :         if (!(tempAlgParams.buffInfo.inBuffType == tempAlgParams.buffInfo.outBuffType 
     112            0 :             && myRankSlice.GetOffset() == outputSlice.GetOffset())) {
     113            0 :             CHK_RET(LocalCopy(tempInsQues[0], myRankSlice, outputSlice));
     114              :         }
     115              :         
     116              :         // 把其他卡的数据input累加到output
     117            0 :         for (u32 tmpRank = 0; tmpRank < tempRankSize_; tmpRank++) {
     118            0 :             if (tmpRank != rankIdx) {
     119            0 :                 DataSlice srcDataSlice = DataSlice(tempAlgParams.buffInfo.scratBuffType, tempAlgParams.buffInfo.scratchBuffBaseOff + 
     120            0 :                     repeatIdx * tempAlgParams.outputRepeatStride + tmpRank * sliceSize, sliceSize);
     121            0 :                 DataSlice dstDataSlice = DataSlice(tempAlgParams.buffInfo.outBuffType, tempAlgParams.buffInfo.outBuffBaseOff + 
     122            0 :                     repeatIdx * tempAlgParams.outputRepeatStride, sliceSize);
     123            0 :                 CHK_RET(LocalReduce(tempInsQues[0], srcDataSlice, dstDataSlice, dataType_, redOp_));                                    
     124              :             }
     125              :         }
     126              :     }
     127            0 :     return HcclResult::HCCL_SUCCESS;
     128              : }
     129              : 
     130            0 : HcclResult InsTempReduceScatterMesh1D::RunReduceScatter(const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues,
     131              :     const TemplateDataParams &tempAlgParams)
     132              : {
     133              :     u32 myAlgRank;
     134            0 :     CHK_RET(GetAlgRank(myRank_, tempVTopo_[0], myAlgRank));
     135              :     // 控制mesh通信的rankSize - 1个对端
     136            0 :     u32 queIdx = 0;
     137            0 :     for (u32 rankIdx = 0; rankIdx < tempRankSize_ - 1; rankIdx++) {
     138            0 :         u32 nextRank = (myAlgRank + 1 + rankIdx) % tempRankSize_;
     139            0 :         RankId remoteRank = tempVTopo_[0][nextRank];
     140              :         u32 rmAlgRank;
     141            0 :         CHK_RET(GetAlgRank(remoteRank, tempVTopo_[0], rmAlgRank));
     142            0 :         HCCL_DEBUG("[InsTempReduceScatterMesh1D][RunReduceScatter] myRank[%d], toRank[%d], fromRank[%d], rmAlgRank[%u]", myRank_, remoteRank, remoteRank, rmAlgRank);
     143            0 :         CHK_PRT_RET(tempLinks.at(remoteRank).empty(), HCCL_ERROR("[InsTempReduceScatterMesh1D][RunReduceScatter] Rank [%d], remoteRank[%d] required links Error.", myRank_, remoteRank),
     144              :             HcclResult::HCCL_E_INTERNAL);
     145            0 :         const std::vector<LinkData> &neighborLinkDatas = tempLinks.at(remoteRank);
     146            0 :         u32 linkNum = rank2PathNumMap_.at(remoteRank);
     147            0 :         CHK_PRT_RET(linkNum != neighborLinkDatas.size(), HCCL_ERROR("[InsTempReduceScatterMesh1D][RunReduceScatter] Rank [%d], remoteRank[%d] linkNum != neighborLinkDatas.size().", myRank_, remoteRank),
     148              :             HcclResult::HCCL_E_INTERNAL);
     149            0 :         std::vector<float> dataSplitRate(linkNum);
     150            0 :         CHK_RET(CalcDataSplitRateForLinks(neighborLinkDatas, dataSplitRate));
     151            0 :         for (u32 linkIdx = 0; linkIdx < linkNum; linkIdx++) {
     152            0 :             CHK_PRT_RET(queIdx >= tempInsQues.size(), HCCL_ERROR("[InsTempReduceScatterMesh1D][RunReduceScatter] queIdx [%u] >= tempInsQues.size() [%zu].", queIdx, tempInsQues.size()),
     153              :                 HcclResult::HCCL_E_INTERNAL);
     154            0 :             InsQuePtr currQue = tempInsQues[queIdx + 1];
     155            0 :             queIdx++;
     156            0 :             const LinkData &neighborLinkData = neighborLinkDatas[linkIdx];
     157            0 :             std::vector<DataSlice> txSrcSlices;
     158            0 :             std::vector<DataSlice> txDstSlices;
     159            0 :             std::vector<DataSlice> rxSrcSlices;
     160            0 :             std::vector<DataSlice> rxDstSlices;
     161            0 :             u64 sendSlice = ((rmAlgRank == tempRankSize_ - 1) && (tempAlgParams.tailSize != 0)) ? tempAlgParams.tailSize : tempAlgParams.sliceSize;
     162            0 :             for (u32 repeatIdx = 0; repeatIdx < tempAlgParams.repeatNum; repeatIdx++) {
     163            0 :                 DataSlice rxSrcSlice = DataSlice(tempAlgParams.buffInfo.inBuffType, tempAlgParams.buffInfo.inBuffBaseOff + 
     164            0 :                     repeatIdx * tempAlgParams.inputRepeatStride + myAlgRank * tempAlgParams.inputSliceStride, tempAlgParams.sliceSize); // 接收源
     165            0 :                 DataSlice rxDstSlice = DataSlice(tempAlgParams.buffInfo.scratBuffType, tempAlgParams.buffInfo.scratchBuffBaseOff + 
     166            0 :                     repeatIdx * tempAlgParams.outputRepeatStride + nextRank * tempAlgParams.sliceSize, tempAlgParams.sliceSize); // 接收目标
     167            0 :                 DataSlice txSrcSlice = DataSlice(tempAlgParams.buffInfo.inBuffType, tempAlgParams.buffInfo.inBuffBaseOff + 
     168            0 :                     repeatIdx * tempAlgParams.inputRepeatStride + nextRank * tempAlgParams.inputSliceStride, sendSlice); // 发送源
     169            0 :                 DataSlice txDstSlice = DataSlice(tempAlgParams.buffInfo.scratBuffType, tempAlgParams.buffInfo.scratchBuffBaseOff + 
     170            0 :                     repeatIdx * tempAlgParams.outputRepeatStride + myAlgRank * sendSlice, sendSlice);  // 发送目标
     171              : 
     172            0 :                 txSrcSlices.push_back(CalcDataSliceForLinks(txSrcSlice, dataSplitRate, linkIdx, dataType_));
     173            0 :                 txDstSlices.push_back(CalcDataSliceForLinks(txDstSlice, dataSplitRate, linkIdx, dataType_));
     174            0 :                 rxSrcSlices.push_back(CalcDataSliceForLinks(rxSrcSlice, dataSplitRate, linkIdx, dataType_));
     175            0 :                 rxDstSlices.push_back(CalcDataSliceForLinks(rxDstSlice, dataSplitRate, linkIdx, dataType_));
     176              :             }
     177            0 :             SendRecvInfo sendRecvInfo{{neighborLinkData, neighborLinkData}, {{txSrcSlices, txDstSlices},{rxSrcSlices, rxDstSlices}}};
     178            0 :             CHK_PRT_RET(SendRecv(sendRecvInfo, currQue, 0, true, DmaMode::PUT),
     179              :                         HCCL_ERROR("[InsTempReduceScatterMesh1D] RunReduceScatter SendReduce failed"), HcclResult::HCCL_E_INTERNAL);
     180            0 :         }
     181            0 :     }
     182            0 :     return HcclResult::HCCL_SUCCESS;
     183              : }
     184              : 
     185            0 : RankId InsTempReduceScatterMesh1D::GetRankFromMap(const u32 rankIdx)
     186              : {
     187            0 :     RankId rank = -1;
     188            0 :     HCCL_INFO("[InsTempReduceScatterMesh1D] GetRankFromMap");
     189            0 :     for (auto &pair : tempVirtRankMap_) {
     190            0 :         if (pair.second == rankIdx) {
     191            0 :             rank = pair.first;
     192            0 :             break;
     193              :         }
     194              :     }
     195            0 :     return rank;
     196              : }
     197              : 
     198              : }  // namespace Hccl
        

Generated by: LCOV version 2.0-1