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 <ios>
      12              : #include <iostream>
      13              : 
      14              : #include "log.h"
      15              : 
      16              : #include "ccu_temp_reduce_mesh_1D_mem2mem.h"
      17              : #include "alg_data_trans_wrapper.h"
      18              : #include "ccu_instruction_reduce_mesh1d_mem2mem.h"
      19              : #include "ccu_rank_group.h"
      20              : #include "ccu_ctx_creator_registry.h"
      21              : #include "ccu_context_reduce_mesh1d_mem2mem.h"
      22              : 
      23              : namespace Hccl {
      24              : 
      25              : static CcuInstRegister<CcuContextReduceMeshMem2Mem1D> registrarReduce(CcuInstType::CCU_REDUCE_MESH_1D_MEM2MEM);
      26              : 
      27            0 : CcuTempReduceMeshMem2Mem1D::CcuTempReduceMeshMem2Mem1D(const RankId virtualRank, const u32 tempRankSize,
      28              :                                                        const std::vector<std::vector<RankId>> &tempVTopo,
      29            0 :                                                        const std::map<RankId, u32>            &tempVirtRankMap)
      30            0 :     : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
      31              : {
      32            0 : }
      33              : 
      34            0 : CcuTempReduceMeshMem2Mem1D::~CcuTempReduceMeshMem2Mem1D()
      35              : {
      36            0 : }
      37              : 
      38            0 : void CcuTempReduceMeshMem2Mem1D::InitReduceInfo(const ReduceOp &reduceOp, const DataType &dataType)
      39              : {
      40            0 :     reduceOp_ = reduceOp;
      41            0 :     dataType_ = dataType;
      42            0 : }
      43              : 
      44            0 : HcclResult CcuTempReduceMeshMem2Mem1D::CalcRes(AlgTempResReq &tempResReq)
      45              : {
      46            0 :     tempResReq.queNum    = 1;
      47            0 :     tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个 stream 给 ccuInsGroup
      48            0 :     HCCL_INFO("[CalcRes] tempResReq.queNum[%u]", tempResReq.queNum);
      49            0 :     CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
      50            0 :     return HcclResult::HCCL_SUCCESS;
      51              : }
      52              : 
      53            0 : HcclResult CcuTempReduceMeshMem2Mem1D::GenExtIns(const TempFuncs          &tempFuncs,
      54              :                                                  const TemplateDataParams &templateDataParams,
      55              :                                                  const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      56              : {
      57            0 :     CHK_PRT_RET(tempInsQues.empty(),
      58              :         HCCL_ERROR("[CcuTempReduceMeshMem2Mem1D] empty queue"), HcclResult::HCCL_E_INTERNAL);
      59            0 :     CHK_PTR_NULL(tempInsQues[0]);
      60              : 
      61            0 :     buffInfo_ = templateDataParams.buffInfo;
      62            0 :     opMode_   = tempFuncs.opMode;
      63            0 :     CcuInstructionReduceMeshMem2Mem1D ccuIns;
      64            0 :     std::vector<uint64_t> dimSize;
      65            0 :     dimSize.push_back(tempRankSize_);
      66              : 
      67            0 :     uint32_t                                rankId    = myRank_;
      68            0 :     uint32_t                                rootId    = tempVirtRankMap_[rootId_];
      69              : 
      70            0 :     const CollAlgOperator                  &op        = op_;
      71            0 :     const std::vector<std::vector<RankId>> &tempVTopo = tempVTopo_;
      72              :     uint64_t token;
      73            0 :     CHK_RET(GetToken(op_, token));
      74            0 :     uint64_t inputAddr          = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
      75            0 :     uint64_t outputAddr         = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
      76            0 :     uint64_t repeatNum          = templateDataParams.repeatNum;
      77            0 :     uint64_t inputRepeatStride  = templateDataParams.inputRepeatStride;
      78            0 :     uint64_t outputRepeatStride = templateDataParams.outputRepeatStride;
      79            0 :     uint64_t normalSliceSize    = templateDataParams.sliceSize;
      80            0 :     uint64_t lastSliceSize      = templateDataParams.tailSize;
      81            0 :     uint64_t repeatNumVar       = UINT64_MAX - repeatNum;
      82              : 
      83              :     // 数据切分为sliceNum块，当数据量不能均匀切分时，后面smallDataSliceNum个数据块比前面bigDataSliceNum个数据块每块少1个数据
      84            0 :     uint64_t sliceNum   = tempRankSize_ - 1;
      85            0 :     uint64_t sliceSize  = templateDataParams.sliceSize; // 获取本rank需要处理的数据量
      86            0 :     uint64_t sliceCount = sliceSize / DataTypeSizeGet(op_.dataType);
      87              : 
      88            0 :     uint64_t bigDataSliceNum    = sliceCount % sliceNum;
      89            0 :     uint64_t bigDataSliceSize   = (sliceCount / sliceNum + 1) * DataTypeSizeGet(op_.dataType);
      90            0 :     uint64_t smallDataSliceNum  = sliceNum - sliceCount % sliceNum;
      91            0 :     uint64_t smallDataSliceSize = sliceCount / sliceNum * DataTypeSizeGet(op_.dataType);
      92              :  
      93            0 :     ccuIns.Init(tempVirtRankMap_[myRank_], rootId, op, tempVTopo, inputAddr, outputAddr, token, bigDataSliceNum, bigDataSliceSize, 
      94              :                 smallDataSliceNum, smallDataSliceSize, inputRepeatStride, outputRepeatStride,
      95              :                 normalSliceSize, lastSliceSize, repeatNumVar);
      96              :  
      97            0 :     HCCL_INFO("[CcuTempReduceMeshMem2Mem1D] Run Init: rankId[%u], rootId[%u], inputAddr[%llu], outputAddr[%llu],"
      98              :                "bigDataSliceNum[%llu], bigDataSliceSize[%llu], smallDataSliceNum[%llu], smallDataSliceSize[%llu],"
      99              :                "inputRepeatStride[%llu], outputRepeatStride[%llu], normalSliceSize[%llu], lastSliceSize[%llu], repeatNumVar[%llu]",
     100              :                rankId, rootId, inputAddr, outputAddr, bigDataSliceNum, bigDataSliceSize, smallDataSliceNum, smallDataSliceSize,
     101              :                inputRepeatStride, outputRepeatStride, normalSliceSize, lastSliceSize, repeatNumVar);
     102              :  
     103            0 :     std::vector<LinkData> links;
     104            0 :     for (auto &pair : tempLinks) {
     105            0 :         if (pair.second.empty()) {
     106            0 :             continue;
     107              :         }
     108            0 :         links.push_back(pair.second[0]);
     109              :     }
     110            0 :     HCCL_DEBUG("[CcuTempReduceMeshMem2Mem1D] links.size[%zu]", links.size());
     111            0 :     ccuIns.SetLinks(links);
     112              :  
     113            0 :     RankGroup rankGroup;
     114            0 :     for (auto &peer : tempVTopo_[0]) {
     115            0 :         rankGroup.AddRank(peer);
     116              :     }
     117            0 :     u32 cntCkeNum = 4;
     118            0 :     ccuIns.SetCntCkeNum(cntCkeNum);
     119            0 :     ccuIns.SetRankGroup(rankGroup);
     120            0 :     HCCL_DEBUG("CcuTempReduceMeshMem2Mem1D is [%s]", ccuIns.Describe().c_str());
     121            0 :     tempInsQues[0]->Append(std::move(std::make_unique<CcuInstructionReduceMeshMem2Mem1D>(ccuIns)));
     122              :  
     123            0 :     return HcclResult::HCCL_SUCCESS;
     124            0 : }
     125              :  
     126            0 : HcclResult CcuTempReduceMeshMem2Mem1D::GenExtIns(const RankGraph *rankGraph,
     127              :                                             const TemplateInfo &tmpInfo,
     128              :                                             const std::vector<InsQuePtr> &tempInsQues) const
     129              : {
     130              :     (void)rankGraph;
     131              :     (void)tmpInfo;
     132              :     (void)tempInsQues;
     133              :     // 框架解析aicpuIns，算法的algCompnnetLite在device侧直接调用Run（）
     134            0 :     return HcclResult::HCCL_SUCCESS;
     135              : }
     136              : 
     137              : } // namespace Hccl