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_two_shot_mem2mem.h"
      17              : #include "alg_data_trans_wrapper.h"
      18              : #include "ccu_instruction_reduce_mesh1d_two_shot_mem2mem.h"
      19              : #include "ccu_rank_group.h"
      20              : #include "ccu_ctx_creator_registry.h"
      21              : #include "ccu_context_reduce_mesh1d_two_shot_mem2mem.h"
      22              : 
      23              : namespace Hccl {
      24              : 
      25              : static CcuInstRegister<CcuContextReduceMeshTwoShotMem2Mem1D> registrarReduce(CcuInstType::CCU_REDUCE_MESH_1D_TWO_SHOT_MEM2MEM);
      26              : 
      27            0 : CcuTempReduceMeshTwoShotMem2Mem1D::CcuTempReduceMeshTwoShotMem2Mem1D(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 : CcuTempReduceMeshTwoShotMem2Mem1D::~CcuTempReduceMeshTwoShotMem2Mem1D()
      35              : {
      36            0 : }
      37              : 
      38            0 : void CcuTempReduceMeshTwoShotMem2Mem1D::InitReduceInfo(const ReduceOp &reduceOp, const DataType &dataType)
      39              : {
      40            0 :     reduceOp_ = reduceOp;
      41            0 :     dataType_ = dataType;
      42            0 : }
      43              : 
      44            0 : HcclResult CcuTempReduceMeshTwoShotMem2Mem1D::CalcSlice(const u64 dataSize, RankSliceInfo &sliceInfoVec)
      45              : {
      46            0 :     std::vector<SliceInfo> tmp(tempVTopo_.size());
      47            0 :     sliceInfoVec.resize(tempRankSize_, tmp);
      48              : 
      49            0 :     u64 unitAllignSize = DataTypeSizeGet(dataType_);
      50            0 :     u64 chunkSize      = RoundUp(dataSize, (tempRankSize_ * unitAllignSize)) * unitAllignSize;
      51            0 :     HCCL_INFO("chunkSize[%llu], dataSize[%llu], tempRankSize_[%u], unitAllignSize[%llu]", chunkSize, dataSize,
      52              :               tempRankSize_, unitAllignSize);
      53            0 :     u64 accumOff = 0;
      54            0 :     for (u32 rankIdx = 0; rankIdx < tempRankSize_; rankIdx++) {
      55            0 :         u64       currChunkSize  = ((dataSize - accumOff) > chunkSize) ? chunkSize : (dataSize - accumOff);
      56            0 :         SliceInfo slice          = {accumOff, currChunkSize};
      57            0 :         sliceInfoVec[rankIdx][0] = slice;
      58            0 :         accumOff += currChunkSize;
      59              :     }
      60              : 
      61            0 :     CHK_PRT_RET(
      62              :         (sliceInfoVec[tempRankSize_ - 1][0].offset + sliceInfoVec[tempRankSize_ - 1][0].size != dataSize),
      63              :         HCCL_ERROR(
      64              :             "[CcuTempReduceMeshTwoShotMem2Mem1D] chunkSize:[%llu], Rank:[%d], SliceInfo calculation error!",
      65              :             chunkSize, myRank_),
      66              :         HcclResult::HCCL_E_INTERNAL);
      67            0 :     return HcclResult::HCCL_SUCCESS;
      68            0 : }
      69              : 
      70            0 : HcclResult CcuTempReduceMeshTwoShotMem2Mem1D::CalcRes(AlgTempResReq &tempResReq)
      71              : {
      72            0 :     tempResReq.queNum    = 1;
      73            0 :     tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个 stream 给 ccuInsGroup
      74            0 :     HCCL_INFO("[CalcRes] tempResReq.queNum[%u]", tempResReq.queNum);
      75            0 :     CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
      76            0 :     return HcclResult::HCCL_SUCCESS;
      77              : }
      78              : 
      79            0 : HcclResult CcuTempReduceMeshTwoShotMem2Mem1D::GenExtIns(const TempFuncs          &tempFuncs,
      80              :                                                  const TemplateDataParams &templateDataParams,
      81              :                                                  const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      82              : {
      83            0 :     buffInfo_ = templateDataParams.buffInfo;
      84            0 :     opMode_   = tempFuncs.opMode;
      85            0 :     CcuInstructionReduceMeshTwoShotMem2Mem1D ccuIns;
      86            0 :     std::vector<uint64_t> dimSize;
      87            0 :     dimSize.push_back(tempRankSize_);
      88            0 :     RankSliceInfo sliceInfoVec;
      89            0 :     CHK_RET(CalcSlice(templateDataParams.sliceSize, sliceInfoVec));
      90              : 
      91            0 :     uint32_t                                rankId    = tempVirtRankMap_[myRank_];
      92            0 :     uint32_t                                rootId    = tempVirtRankMap_[rootId_];
      93              : 
      94            0 :     const CollAlgOperator                  &op        = op_;
      95            0 :     const std::vector<std::vector<RankId>> &tempVTopo = tempVTopo_;
      96            0 :     uint64_t inputAddr          = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
      97            0 :     uint64_t outputAddr         = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
      98            0 :     uint64_t scratchAddr        = BufferTypeToAddr(buffInfo_.scratBuffType) + buffInfo_.scratchBuffBaseOff;
      99              :     uint64_t token;
     100            0 :     CHK_RET(GetToken(op_, token));
     101            0 :     uint64_t normalSliceSize    = sliceInfoVec[0][0].size;
     102            0 :     uint64_t lastSliceSize      = sliceInfoVec[tempRankSize_ - 1][0].size;
     103            0 :     uint64_t mySliceSize        = sliceInfoVec[rankId][0].size;
     104              : 
     105            0 :     ccuIns.Init(rankId, rootId, op, tempVTopo, inputAddr, outputAddr, scratchAddr, token,
     106              :                  normalSliceSize, lastSliceSize, mySliceSize);
     107              :  
     108            0 :     HCCL_INFO("[CcuTempReduceMeshTwoShotMem2Mem1D] Run Init: rankId[%u], rootId[%u], inputAddr[%llu], outputAddr[%llu],"
     109              :                "scratchAddr[%llu], normalSliceSize[%llu], lastSliceSize[%llu], mySliceSize[%llu]",
     110              :                rankId, rootId, inputAddr, outputAddr, scratchAddr, normalSliceSize, lastSliceSize, mySliceSize);
     111              :  
     112            0 :     std::vector<LinkData> links;
     113            0 :     for (auto &pair : tempLinks) {
     114            0 :         if (pair.second.empty()) {
     115            0 :             continue;
     116              :         }
     117            0 :         links.push_back(pair.second[0]);
     118              :     }
     119            0 :     HCCL_DEBUG("[CcuTempReduceMeshTwoShotMem2Mem1D] links.size[%zu]", links.size());
     120            0 :     ccuIns.SetLinks(links);
     121              :  
     122            0 :     RankGroup rankGroup;
     123            0 :     for (auto &peer : tempVTopo_[0]) {
     124            0 :         rankGroup.AddRank(peer);
     125              :     }
     126            0 :     u32 cntCkeNum = 4;
     127            0 :     ccuIns.SetCntCkeNum(cntCkeNum);
     128            0 :     ccuIns.SetRankGroup(rankGroup);
     129            0 :     HCCL_DEBUG("CcuTempReduceMeshTwoShotMem2Mem1D is [%s]", ccuIns.Describe().c_str());
     130            0 :     tempInsQues[0]->Append(std::move(std::make_unique<CcuInstructionReduceMeshTwoShotMem2Mem1D>(ccuIns)));
     131              :  
     132            0 :     return HcclResult::HCCL_SUCCESS;
     133            0 : }
     134              :  
     135            0 : HcclResult CcuTempReduceMeshTwoShotMem2Mem1D::GenExtIns(const RankGraph *rankGraph,
     136              :                                             const TemplateInfo &tmpInfo,
     137              :                                             const std::vector<InsQuePtr> &tempInsQues) const
     138              : {
     139              :     (void)rankGraph;
     140              :     (void)tmpInfo;
     141              :     (void)tempInsQues;
     142              :     // 框架解析aicpuIns，算法的algCompnnetLite在device侧直接调用Run（）
     143            0 :     return HcclResult::HCCL_SUCCESS;
     144              : }
     145              : 
     146              : } // namespace Hccl