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 "log.h"
      12              : 
      13              : #include "alg_data_trans_wrapper.h"
      14              : #include "ins_alg_template/ins_temp_all_reduce_mesh_1D_one_shot.h"
      15              : 
      16              : namespace Hccl {
      17            0 : InsTempAllReduceMesh1DOneShot::InsTempAllReduceMesh1DOneShot(const RankId virtualRank, const u32 tempRankSize,
      18            0 :     const std::vector<std::vector<RankId>> &tempVTopo, const std::map<RankId, u32> &tempVirtRankMap)
      19            0 :     : InsAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
      20              : {
      21            0 : }
      22              : 
      23            0 : InsTempAllReduceMesh1DOneShot::~InsTempAllReduceMesh1DOneShot()
      24              : {
      25            0 : }
      26              : 
      27            0 : u32 InsTempAllReduceMesh1DOneShot::CalcScratchMultiple(BufferType input, BufferType output) const
      28              :  {
      29              :     (void)input;
      30              :     (void)output;
      31              :     // one shot 场景，scratch Buffer 需要是 usrIn的rankSize倍
      32            0 :     return tempRankSize_;
      33              :  }
      34              : 
      35            0 : HcclResult InsTempAllReduceMesh1DOneShot::CalcRes(AlgTempResReq &tempResReq)
      36              : {
      37            0 :     tempResReq.queNum = tempVTopo_[0].size();
      38            0 :     tempResReq.streamNum = tempResReq.queNum;
      39            0 :     tempResReq.queNotifys = CreateMasterSlaveQueNotifiesRequest(tempResReq.queNum);
      40              : 
      41            0 :     QId centerQ = 0;
      42            0 :     tempResReq.localWaitGroupCntNotify.emplace_back(centerQ, 0);
      43            0 :     tempResReq.localBcastPostCntNotify.emplace_back(centerQ, 0);
      44              : 
      45            0 :     CHK_PRT_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq) != HcclResult::HCCL_SUCCESS,
      46              :                 HCCL_ERROR("[CollAlgFactory] [InsTempAllReduceMesh1DOneShot] Rank [%d], resLinks calculation error!", myRank_),
      47              :                 HcclResult::HCCL_E_INTERNAL);
      48              : 
      49            0 :     return HcclResult::HCCL_SUCCESS;
      50              : }
      51              : 
      52            0 : HcclResult InsTempAllReduceMesh1DOneShot::CalcSlice(const u64 dataSize, RankSliceInfo &sliceInfoVec)
      53              : {
      54            0 :     std::vector<SliceInfo> tmp(tempVTopo_.size());
      55            0 :     sliceInfoVec.resize(tempRankSize_, tmp);
      56            0 :     AllignInfo allignInfo = {false, 0, dataType_}; // 参数填充，CalcRsAgSliceInfoMesh实际没用到
      57              : 
      58            0 :     CHK_RET(CalcRsAgSliceInfoMesh(myRank_, tempRankSize_, allignInfo, dataSize, sliceInfoVec));
      59              : 
      60            0 :     return HcclResult::HCCL_SUCCESS;
      61            0 : }
      62              : 
      63            0 : HcclResult InsTempAllReduceMesh1DOneShot::GenExtIns(const TempFuncs &tempFuncs, const TemplateDataParams &tempAlgParams,
      64              :     const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      65              : {
      66            0 :     HCCL_INFO("[InsTempAllReduceMesh1DOneShot][Run] AllReduceMesh1DOneShot begin: rank[%d] start", myRank_);
      67              : 
      68            0 :     opMode_              = tempFuncs.opMode;
      69            0 :     queNum_ = tempVTopo_[0].size();
      70            0 :     CHK_PRT_RET(queNum_ != tempInsQues.size(),
      71              :                 HCCL_ERROR("[CollAlgFactory] [InsTempAllReduceMesh1DOneShot] Rank [%d], requiredQue Error.", myRank_),
      72              :                 HcclResult::HCCL_E_INTERNAL);
      73              : 
      74            0 :     RankSliceInfo sliceInfoVec;
      75            0 :     CHK_RET(CalcSlice(tempAlgParams.sliceSize, sliceInfoVec));
      76              : 
      77            0 :     HCCL_INFO("[InsTempAllReduceMesh1DOneShot][PreCopy] write userIn data directly to the ScratchBuffer, skip precopy");
      78            0 :     CHK_RET(RunAllReduce(tempAlgParams, sliceInfoVec, tempLinks, tempInsQues));
      79            0 :     HCCL_INFO("[InsTempAllReduceMesh1DOneShot][PostCopy] data is already in the userOut, skip postcopy");
      80              : 
      81            0 :     HCCL_INFO("[InsTempAllReduceMesh1DOneShot][Run] AllReduceMesh1DOneShot finished: rank[%d] end", myRank_);
      82            0 :     return HcclResult::HCCL_SUCCESS;
      83            0 : }
      84              : 
      85            0 : HcclResult InsTempAllReduceMesh1DOneShot::RunAllReduce(const TemplateDataParams &tempAlgParams, const RankSliceInfo &sliceInfoVec,
      86              :     const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      87              : {
      88            0 :     HCCL_INFO("[InsTempAllReduceMesh1DOneShot][RunAllReduce] send/recv: rank[%d]", myRank_);
      89              : 
      90              :     // semaphore sync
      91            0 :     if (tempVTopo_[0].size() > 1) {
      92            0 :         CHK_RET(PreSyncInterQueues(tempInsQues));
      93              :     }
      94              : 
      95            0 :     DataSlice usrInSlices = DataSlice(BufferType::INPUT, tempAlgParams.buffInfo.inBuffBaseOff, tempAlgParams.sliceSize);
      96            0 :     DataSlice usrOutSlices = DataSlice(BufferType::OUTPUT, tempAlgParams.buffInfo.outBuffBaseOff, tempAlgParams.sliceSize);
      97              : 
      98              :     // 主流动作
      99            0 :     CHK_RET(LocalCopy(tempInsQues[0], usrInSlices, usrOutSlices));
     100              : 
     101              :     // 从流动作
     102            0 :     for (u32 queIdx = 1; queIdx < queNum_; queIdx++) {
     103            0 :         u32 nextRank = (myRank_ + queIdx) % tempRankSize_; // 让rank和que对应上
     104            0 :         RankId fromRank = nextRank;
     105            0 :         RankId toRank = nextRank;
     106              : 
     107            0 :         const std::vector<LinkData> &linkRecv = tempLinks.at(fromRank);
     108            0 :         const std::vector<LinkData> &linkSend = tempLinks.at(toRank);
     109              : 
     110            0 :         std::vector<DataSlice> txSrcSlices;
     111            0 :         std::vector<DataSlice> txDstSlices;
     112              : 
     113            0 :         u64 txDstOffset   = sliceInfoVec[myRank_][0].offset + tempAlgParams.buffInfo.scratchBuffBaseOff;
     114            0 :         u64 txDstSize     = sliceInfoVec[myRank_][0].size;
     115            0 :         DataSlice txSrcSlice = usrInSlices;
     116            0 :         DataSlice txDstSlice = DataSlice(BufferType::SCRATCH, txDstOffset, txDstSize);
     117            0 :         txSrcSlices.push_back(txSrcSlice);
     118            0 :         txDstSlices.push_back(txDstSlice);
     119              : 
     120            0 :         std::vector<DataSlice> rxSrcSlices;
     121            0 :         std::vector<DataSlice> rxDstSlices;
     122            0 :         u64 rxDstOffset   = sliceInfoVec[fromRank][0].offset + tempAlgParams.buffInfo.scratchBuffBaseOff;
     123            0 :         u64 rxDstSize     = sliceInfoVec[fromRank][0].size;
     124            0 :         DataSlice rxSrcSlice = usrInSlices;
     125            0 :         DataSlice rxDstSlice = DataSlice(BufferType::SCRATCH, rxDstOffset, rxDstSize);
     126            0 :         rxSrcSlices.push_back(rxSrcSlice);
     127            0 :         rxDstSlices.push_back(rxDstSlice);
     128              : 
     129            0 :         TxRxLinks sendRecvLinks(linkSend[0], linkRecv[0]);
     130            0 :         TxRxSlicesList sendRecvSlicesList({txSrcSlices, txDstSlices}, {rxSrcSlices, rxDstSlices});
     131              : 
     132            0 :         SendRecvInfo sendRecvInfo(sendRecvLinks, sendRecvSlicesList);
     133            0 :         CHK_PRT_RET(SendRecv(sendRecvInfo, tempInsQues[queIdx], 0, true, DmaMode::PUT),
     134              :             HCCL_ERROR("[InsTempAllReduceMesh1DOneShot] RunAllReduce SendRecv failed"),
     135              :             HcclResult::HCCL_E_INTERNAL);
     136            0 :     }
     137              : 
     138              :     // semaphore sync
     139            0 :     if (tempVTopo_[0].size() > 1) {
     140            0 :         CHK_RET(PostSyncInterQueues(tempInsQues));
     141              :     }
     142              : 
     143            0 :     HCCL_INFO("[InsTempAllReduceMesh1DOneShot][RunAllReduce] reduce: rank[%d]", myRank_);
     144              : 
     145            0 :     for (u32 rankIdx = 0; rankIdx < tempVTopo_[0].size(); rankIdx++) {
     146            0 :         RankId myRank = myRank_;
     147            0 :         RankId curRank = rankIdx;
     148            0 :         if ( curRank == myRank) {
     149            0 :             continue;
     150              :         }
     151              : 
     152            0 :         u64 curSrcOffset = sliceInfoVec[curRank][0].offset + tempAlgParams.buffInfo.scratchBuffBaseOff;
     153            0 :         u64 curSrcSize = sliceInfoVec[curRank][0].size;
     154            0 :         DataSlice curSrcSlice = DataSlice(BufferType::SCRATCH, curSrcOffset, curSrcSize);
     155            0 :         DataSlice curDstSlice = usrOutSlices;
     156              : 
     157            0 :         CHK_RET(LocalReduce(tempInsQues[0], curSrcSlice, curDstSlice, dataType_, redOp_));
     158              :     }
     159              : 
     160            0 :     return HcclResult::HCCL_SUCCESS;
     161              : }
     162              : 
     163              : } // namespace Hccl