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_temp_all_reduce_aicpu_reduce.h"
      15              : 
      16              : namespace Hccl {
      17            0 : InsTempAllReduceAicpuReduce::InsTempAllReduceAicpuReduce(const RankId virtualRank, const u32 tempRankSize,
      18              :                                    const std::vector<std::vector<RankId>> &tempVTopo,
      19            0 :                                    const std::map<RankId, u32>            &tempVirtRankMap)
      20            0 :     : InsAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
      21              : {
      22            0 : }
      23              : 
      24            0 : InsTempAllReduceAicpuReduce::~InsTempAllReduceAicpuReduce()
      25              : {
      26            0 : }
      27              : 
      28            0 : HcclResult InsTempAllReduceAicpuReduce::CalcRes(AlgTempResReq &tempResReq)
      29              : {
      30            0 :     tempResReq.queNum = tempVTopo_[0].size() - 1;
      31            0 :     tempResReq.streamNum = tempResReq.queNum;
      32            0 :     tempResReq.queNotifys = CreateMasterSlaveQueNotifiesRequest(tempResReq.queNum);
      33            0 :     QId centerQ = 0;
      34            0 :     tempResReq.localWaitGroupCntNotify.emplace_back(centerQ, 0);
      35            0 :     tempResReq.localBcastPostCntNotify.emplace_back(centerQ, 0);
      36            0 :     CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
      37            0 :     HCCL_DEBUG("[InsTempAllReduceAicpuReduce]CalcRes: queNum[%u], myRank[%d], tempRankSize[%u]", tempResReq.queNum, myRank_, tempRankSize_);
      38            0 :     return HcclResult::HCCL_SUCCESS;
      39              : }
      40              : 
      41            0 : u32 InsTempAllReduceAicpuReduce::CalcScratchMultiple(BufferType inBuffType, BufferType outBuffType) const
      42              : {   
      43              :     (void) inBuffType;
      44              :     (void) outBuffType;
      45            0 :     return tempRankSize_;
      46              : }
      47              : 
      48            0 : HcclResult InsTempAllReduceAicpuReduce::RunAllGatherMesh(const TempFuncs &tempFuncs, const TemplateDataParams &templateDataParams,
      49              :                         const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      50              : {   
      51              :     (void) tempFuncs;
      52              :     // 本端rank数据从本端input -> 本端scratch 
      53            0 :     u64       srcOffset = templateDataParams.buffInfo.inBuffBaseOff;
      54            0 :     u64       srcSize   = templateDataParams.sliceSize;
      55            0 :     u64       dstOffset = templateDataParams.sliceSize * myRank_;
      56            0 :     DataSlice srcSlice  = DataSlice(BufferType::INPUT, srcOffset, srcSize);
      57            0 :     DataSlice dstSlice  = DataSlice(BufferType::SCRATCH, dstOffset, srcSize);
      58            0 :     std::unique_ptr<Instruction> insLocalCopy = std::make_unique<InsLocalCopy>(srcSlice, dstSlice);
      59            0 :     tempInsQues[0]->Append(std::move(insLocalCopy));
      60              : 
      61            0 :     CHK_RET(PreSyncInterQueues(tempInsQues));
      62              :     // 本端rank数据从本端input -> 对端scratch 
      63            0 :     std::vector<DataSlice> txSrcSlices;
      64            0 :     std::vector<DataSlice> txDstSlices;
      65              :     DataSlice currSendSliceSrc
      66            0 :         = DataSlice(BufferType::INPUT, templateDataParams.buffInfo.inBuffBaseOff, templateDataParams.sliceSize);
      67              :     DataSlice currSendSliceDst
      68            0 :         = DataSlice(BufferType::SCRATCH, templateDataParams.sliceSize * myRank_, templateDataParams.sliceSize);
      69            0 :     txSrcSlices.push_back(currSendSliceSrc);
      70            0 :     txDstSlices.push_back(currSendSliceDst);
      71              :     u32 myAlgRank;
      72            0 :     CHK_RET(GetAlgRank(myRank_, tempVTopo_[0], myAlgRank));
      73            0 :     for (u32 queIdx = 0; queIdx < tempVTopo_[0].size() - 1; queIdx++) {
      74            0 :         RankId neighborRank = tempVTopo_[0][(myAlgRank + 1 + queIdx) % tempRankSize_];
      75            0 :         LinkData neighborLinkData = tempLinks.at(neighborRank)[0];
      76            0 :         TxRxLinks sendRecvLinks(neighborLinkData, neighborLinkData);
      77              :         DataSlice currRecvSliceSrc
      78            0 :             = DataSlice(BufferType::SCRATCH, templateDataParams.sliceSize * neighborRank, templateDataParams.sliceSize);
      79              :         DataSlice currRecvSliceDst
      80            0 :             = DataSlice(BufferType::SCRATCH, templateDataParams.sliceSize * neighborRank, templateDataParams.sliceSize);
      81            0 :         std::vector<DataSlice> rxSrcSlices;
      82            0 :         std::vector<DataSlice> rxDstSlices;
      83            0 :         rxSrcSlices.push_back(currRecvSliceSrc);
      84            0 :         rxDstSlices.push_back(currRecvSliceDst);
      85            0 :         TxRxSlicesList sendRecvSlicesList({txSrcSlices, txDstSlices}, {rxSrcSlices, rxDstSlices});
      86            0 :         SendRecvInfo sendRecvInfo(sendRecvLinks, sendRecvSlicesList);
      87            0 :         CHK_PRT_RET(SendRecv(sendRecvInfo, tempInsQues[queIdx], 0, true, dmaMode_), HCCL_ERROR("[InsTempAllReduceAicpuReduce] RunAllGather sendrecv failed"),
      88              :                     HcclResult::HCCL_E_INTERNAL);
      89            0 :     }
      90            0 :     CHK_RET(PostSyncInterQueues(tempInsQues));
      91            0 :     return HCCL_SUCCESS;
      92            0 : }
      93              : 
      94            0 : HcclResult InsTempAllReduceAicpuReduce::RunAicpuLocalReduce(const TemplateDataParams &templateDataParams, std::vector<InsQuePtr> &tempInsQues)
      95              : {   
      96            0 :     DataSlice dataSlice = DataSlice(BufferType::SCRATCH, 0, templateDataParams.sliceSize);
      97            0 :     for (u32 rankId = 1; rankId < tempRankSize_; rankId++) {
      98            0 :         DataSlice addSlice = DataSlice(BufferType::SCRATCH, templateDataParams.sliceSize * rankId, templateDataParams.sliceSize);
      99            0 :         AicpuReduce(tempInsQues[0], addSlice, dataSlice, dataType_, redOp_);
     100              :     }
     101            0 :     DataSlice outputSlice = DataSlice(BufferType::OUTPUT, templateDataParams.buffInfo.inBuffBaseOff, templateDataParams.sliceSize);
     102            0 :     LocalCopy(tempInsQues[0], dataSlice, outputSlice);
     103            0 :     return HCCL_SUCCESS;
     104              : }
     105              : 
     106            0 : HcclResult InsTempAllReduceAicpuReduce::GenExtIns(const TempFuncs &tempFuncs, const TemplateDataParams &templateDataParams,
     107              :                         const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
     108              : {
     109            0 :     HCCL_INFO("[InsTempAllReduceAicpuReduce] Run start");
     110            0 :     if (IsPcieLink(tempLinks)) {
     111            0 :         dmaMode_ = DmaMode::GET;
     112              :     } else {
     113            0 :         dmaMode_ = DmaMode::PUT;
     114              :     }
     115            0 :     if (tempVTopo_[0].size() == 1) {
     116            0 :         return HcclResult::HCCL_SUCCESS;
     117              :     }
     118            0 :     opMode_              = tempFuncs.opMode;
     119            0 :     queNum_ = tempVTopo_[0].size() - 1;
     120            0 :     CHK_PRT_RET(queNum_ != tempInsQues.size(),
     121              :         HCCL_ERROR("[CollAlgFactory] [InsTempAllReduceAicpuReduce] Rank [%d], requiredQue Error.", myRank_),
     122              :         HcclResult::HCCL_E_INTERNAL);
     123              : 
     124            0 :     RunAllGatherMesh(tempFuncs, templateDataParams, tempLinks, tempInsQues);
     125            0 :     StreamSync(tempInsQues);
     126            0 :     RunAicpuLocalReduce(templateDataParams, tempInsQues);
     127            0 :     HCCL_INFO("[InsTempAllReduceAicpuReduce] Run finished");
     128            0 :     return HCCL_SUCCESS;
     129              : }
     130              : 
     131              : } // namespace Hccl