Line data    Source code

       1              : /**
       2              :  * Copyright (c) 2026 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_rank_group.h"
      17              : #include "ccu_ctx_creator_registry.h"
      18              : #include "ccu_context_all_gather_mesh1d_2die.h"
      19              : #include "ccu_temp_all_gather_mesh_1D_2die.h"
      20              : #include "ccu_ins_group.h"
      21              : 
      22              : namespace Hccl {
      23              : 
      24              : static CcuInstRegister<CcuContextAllGatherMesh1D2Die> g_registrarAllGather(
      25              :     CcuInstType::CCU_ALLGATHER_MESH_1D_2DIE);
      26              : 
      27            0 : CcuTempAllGatherMesh1D2Die::CcuTempAllGatherMesh1D2Die(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 : CcuTempAllGatherMesh1D2Die::~CcuTempAllGatherMesh1D2Die()
      35              : {
      36            0 : }
      37              : 
      38            0 : HcclResult CcuTempAllGatherMesh1D2Die::CalcRes(AlgTempResReq &tempResReq)
      39              : {
      40            0 :     tempResReq.queNum = 1;
      41            0 :     tempResReq.streamNum = tempResReq.queNum + 1;
      42            0 :     HCCL_DEBUG("[CalcRes] tempResReq.queNum[%u]", tempResReq.queNum);
      43              : 
      44            0 :     CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
      45            0 :     return HcclResult::HCCL_SUCCESS;
      46              : }
      47              : 
      48            0 : uint64_t CcuTempAllGatherMesh1D2Die::GetMaxSliceSize() const
      49              : {
      50            0 :     return UB_MAX_DATA_SIZE;
      51              : }
      52              : 
      53            0 : HcclResult CcuTempAllGatherMesh1D2Die::GenExtIns(const TempFuncs &tempFuncs, TemplateDataParams &tempAlgParams,
      54              :                                                 const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      55              : {
      56            0 :     CHK_PRT_RET(tempInsQues.empty(),
      57              :         HCCL_ERROR("[CcuTempAllGatherMesh1D2Die] empty queue"), HcclResult::HCCL_E_INTERNAL);
      58            0 :     CHK_PTR_NULL(tempInsQues[0]);
      59            0 :     opMode_ = tempFuncs.opMode;
      60            0 :     CcuInstructionAllGatherMesh1D2Die ccuInsAllGatherMesh1D2Die;
      61            0 :     std::vector<uint64_t> dimSize;
      62            0 :     dimSize.push_back(tempRankSize_);
      63              : 
      64            0 :     uint64_t sliceSize = tempAlgParams.sliceSize;
      65            0 :     uint64_t inputAddr = BufferTypeToAddr(tempAlgParams.buffInfo.inBuffType) + tempAlgParams.buffInfo.inBuffBaseOff;
      66            0 :     uint64_t outputAddr = BufferTypeToAddr(tempAlgParams.buffInfo.outBuffType) + tempAlgParams.buffInfo.outBuffBaseOff;
      67            0 :     uint64_t outputSliceStride = tempAlgParams.outputSliceStride;
      68              :     uint64_t token;
      69            0 :     CHK_RET(GetToken(op_, token));
      70            0 :     HCCL_INFO("[CcuTempAllGatherMesh1D2Die] dimSize[%llu], sliceSize[%llu], inputAddr[%llu],"\
      71              :         "outputAddr[%llu], outputSliceStride[%llu]",
      72              :         dimSize[0], sliceSize, inputAddr, outputAddr, outputSliceStride);
      73              : 
      74              :     // key表示为dieId
      75            0 :     std::map<uint32_t, std::vector<LinkData>> linksDie;
      76            0 :     std::map<uint32_t, RankGroup> rankGroup;
      77              : 
      78            0 :     for (auto link: tempLinks) {
      79            0 :         std::vector<LinkData> linkData = link.second;
      80            0 :         RankId peerRankId = link.first;
      81            0 :         if (linkData.size() == 0){
      82            0 :             continue;
      83              :         }
      84              : 
      85            0 :         linksDie[linkData[0].GetLocalDieId()].push_back(linkData[0]);
      86            0 :         rankGroup[linkData[0].GetLocalDieId()].AddRank(peerRankId);
      87            0 :     }
      88              : 
      89            0 :     HCCL_INFO("[CcuTempAllGatherMesh1D2Die] linksDie0Size[%llu], linksDie1Size[%llu]", linksDie[0].size(), linksDie[1].size());
      90              : 
      91            0 :     rankGroup[0].AddRank(myRank_);
      92            0 :     rankGroup[1].AddRank(myRank_);
      93              : 
      94            0 :     std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
      95            0 :     for (uint32_t dieId = 0; dieId < 2; dieId++) {  // 2Die算法，需要下发 2 条通信指令
      96            0 :         CcuInstructionAllGatherMesh1D2Die ccuInstruction;
      97            0 :         bool withMyRank = linksDie[dieId].size() > linksDie[1 - dieId].size() ? false : true;
      98            0 :         ccuInstruction.Init(myRank_, inputAddr, outputAddr, sliceSize, outputSliceStride, token, withMyRank, op_, tempVTopo_);
      99            0 :         ccuInstruction.SetLinks(linksDie[dieId]);
     100            0 :         ccuInstruction.SetRankGroup(rankGroup[dieId]);
     101            0 :         ccuInstruction.SetCntCkeNum(5);  // 每个transport用5个CKE
     102            0 :         insGroupPtr->Append(std::move(std::make_unique<CcuInstructionAllGatherMesh1D2Die>(ccuInstruction)));
     103            0 :     }
     104            0 :     tempInsQues[0]->Append(std::move(insGroupPtr));  // 只有一条que
     105            0 :     HCCL_INFO("[CcuTempAllGatherMesh1D2Die] Template Run for all steps Ends.");
     106            0 :     return HcclResult::HCCL_SUCCESS;
     107            0 : }
     108              : 
     109              : } // namespace Hccl