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_all_gather_mesh_2D_mem2mem.h"
      17              : #include "alg_data_trans_wrapper.h"
      18              : #include "ccu_instruction_all_gather_mesh2d_mem2mem.h"
      19              : #include "ccu_rank_group.h"
      20              : #include "ccu_ctx_creator_registry.h"
      21              : #include "ccu_context_all_gather_mesh2d_mem2mem.h"
      22              : #include "ccu_ins_group.h"
      23              : 
      24              : namespace Hccl {
      25              : 
      26              : static CcuInstRegister<CcuContextAllGatherMeshMem2Mem2D> registrarAllGather(CcuInstType::CCU_ALLGATHER_MESH_2D_MEM2MEM);
      27              : 
      28            0 : CcuTempAllGatherMeshMem2Mem2D::CcuTempAllGatherMeshMem2Mem2D(const RankId virtualRank, const u32 tempRankSize,
      29            0 :     const std::vector<std::vector<RankId>> &tempVTopo, const std::map<RankId, u32> &tempVirtRankMap)
      30            0 :     : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
      31              : {
      32              :     // 填充框内的维度大小， concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
      33            0 :     if (tempVTopo_.size() != DIM_SIZE || tempVTopo_[0].size() <= 1 || tempVTopo_[1].size() <= 1) {
      34            0 :         THROW<InvalidParamsException>(StringFormat("[CcuTempAllGatherMeshMem2Mem2D] Rank[%d], Invalid tempVTopo "
      35              :             "Size[%u] or Invalid tempVTopo[0] size [%u] or tempVTopo[1] size [%u].",
      36            0 :             myRank_, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size()));
      37              :     }
      38            0 :     dimSize_.emplace_back(tempVTopo[0].size());
      39            0 :     dimSize_.emplace_back(tempVTopo[1].size());
      40            0 : }
      41              : 
      42            0 : CcuTempAllGatherMeshMem2Mem2D::~CcuTempAllGatherMeshMem2Mem2D()
      43              : {
      44            0 : }
      45              : 
      46            0 : HcclResult CcuTempAllGatherMeshMem2Mem2D::CalcRes(AlgTempResReq &tempResReq)
      47              : {
      48            0 :     tempResReq.queNum = 1;  // 只申请一个insQue，填充一个insGroup，由框架将其中的ins放在多个stream上
      49            0 :     tempResReq.streamNum = tempResReq.queNum + 1;  // 多申请一个 stream 给 ccuInsGroup
      50            0 :     uint32_t dieNum = tempVTopo_.size();
      51            0 :     if (dieNum != 2) {  // concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
      52            0 :         THROW<InvalidParamsException>(StringFormat("[CcuTempAllGatherMeshMem2Mem2D] Rank[%d], Invalid IODieNum[%u].",
      53              :             myRank_, dieNum));
      54              :     }
      55            0 :     HCCL_INFO("[CcuTempAllGatherMeshMem2Mem2D] Rank[%d] requiredQueNum[%u] VtopoSize[%u], VtopoSize0[%u] "
      56              :         "VtopoSize1[%u].", myRank_, tempResReq.queNum, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size());
      57              : 
      58              :     uint32_t myAlgRank;
      59            0 :     for (u32 dim = 0; dim < tempVTopo_.size(); dim++) {
      60            0 :         CHK_RET(GetAlgRank(myRank_, tempVTopo_[dim], myAlgRank));
      61            0 :         for (u32 queIdx = 0; queIdx < tempVTopo_[dim].size() - 1; queIdx++) {
      62              :             // find neighbors -> virtualRank
      63            0 :             u32    neighborAlgRank = (myAlgRank + 1 + queIdx) % (tempVTopo_[dim].size());
      64            0 :             RankId neighborRank    = tempVTopo_[dim][neighborAlgRank];
      65            0 :             HCCL_INFO("[CollAlgFactory] [CcuTempAllGatherMeshMem2Mem2D] Rank[%d], Dim[%u], NeighborRank[%d].", myRank_,
      66              :                        dim, neighborRank);
      67              : 
      68              :             // LinkNum
      69            0 :             tempResReq.links[neighborRank] = 1;
      70              :         }
      71              :     }
      72              : 
      73            0 :     return HcclResult::HCCL_SUCCESS;
      74              : }
      75              : 
      76            0 : HcclResult CcuTempAllGatherMeshMem2Mem2D::GenExtIns(const TempFuncs &tempFuncs,
      77              :     const TemplateDataParams &templateDataParams, const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      78              : {
      79            0 :     opMode_ = tempFuncs.opMode;
      80            0 :     buffInfo_ = templateDataParams.buffInfo;
      81              : 
      82              :     // 分别记录两个Die上的link，构造rankGroup
      83            0 :     for (auto pair : tempLinks) {
      84            0 :         if (pair.second.size() == 0 || pair.second[0].GetHop() != 1) {  // ESL环境上暂只有直连链路
      85            0 :             THROW<InvalidParamsException>(StringFormat("[CcuTempAllGatherMeshMem2Mem2D] Rank[%d]--Peer[%d], "
      86            0 :                 "InvalidHop[%u].", myRank_, pair.first, pair.second[0].GetHop()));
      87              :         }
      88            0 :         if ((pair.first / dimSize_[0] == myRank_ / dimSize_[0]) && pair.second[0].GetHop() == 1) {
      89            0 :             linksX_.emplace_back(pair.second[0]);
      90            0 :         } else if ((pair.first % dimSize_[0] == myRank_ % dimSize_[0]) && pair.second[0].GetHop() == 1) {
      91            0 :             linksY_.emplace_back(pair.second[0]);
      92              :         } else {
      93            0 :             THROW<InvalidParamsException>(StringFormat(
      94              :                 "[CcuTempAllGatherMeshMem2Mem2D] Rank[%d], Unexpected peerRank[%d] in tempLinks.", myRank_,
      95            0 :                 pair.first));
      96              :         }
      97            0 :     }
      98              : 
      99            0 :     RankGroup rankGroupX;
     100            0 :     RankGroup rankGroupY;
     101            0 :     AddRanksToGroup(tempVTopo_,rankGroupX,rankGroupY);
     102              : 
     103            0 :     uint64_t inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
     104            0 :     uint64_t outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
     105            0 :     uint64_t offset = static_cast<uint64_t>(op_.inputMem->GetSize());
     106            0 :     uint64_t sliceSize = templateDataParams.sliceSize;
     107              :     uint64_t token;
     108            0 :     CHK_RET(GetToken(op_, token));
     109              : 
     110            0 :     uint64_t aSize = sliceSize * dimSize_[0] / (dimSize_[0] + dimSize_[1]);
     111            0 :     uint64_t bSize = sliceSize - aSize;
     112            0 :     std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
     113            0 :     for (uint32_t axisId = 0; axisId < 2; axisId++) {  // 2D算法，需要执行两次
     114            0 :         HCCL_INFO("[CcuTempAllGatherMeshMem2Mem2D][GenExtIns] Rank[%d], axisId[%u], aSize[%llu], bSize[%llu].",
     115              :             myRank_, axisId, aSize, bSize);
     116              : 
     117            0 :         CcuInstructionAllGatherMeshMem2Mem2D ins = CcuInstructionAllGatherMeshMem2Mem2D();
     118            0 :         ins.Init(myRank_, axisId, inputAddr, outputAddr, aSize, bSize, sliceSize, offset, token, op_, tempVTopo_);
     119            0 :         ins.SetLinks(axisId == 0 ? linksX_ : linksY_);
     120            0 :         ins.SetRankGroup(axisId == 0 ? rankGroupX : rankGroupY);
     121            0 :         u32 cntCkeNum = 5;  // 每个transport用5个CKE
     122            0 :         ins.SetCntCkeNum(cntCkeNum);
     123            0 :         ins.Describe();
     124            0 :         insGroupPtr->Append(std::move(std::make_unique<CcuInstructionAllGatherMeshMem2Mem2D>(ins)));
     125            0 :     }
     126            0 :     tempInsQues[0]->Append(std::move(insGroupPtr));  // 只有一条流
     127              : 
     128            0 :     return HcclResult::HCCL_SUCCESS;
     129            0 : }
     130              : } // namespace Hccl