LCOV - code coverage report
Current view: top level - legacy/ascend950/service/collective/alg/coll_alg_factory/alg_template/ccu_alg_template - ccu_temp_all_gather_mesh_1D_detour.cc (source / functions) Coverage Total Hit
Test: coverage.info Lines: 0.0 % 122 0
Test Date: 2026-08-04 10:52:23 Functions: 0.0 % 10 0

            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_instruction_all_gather_mesh1d_detour.h"
      17              : #include "ccu_rank_group.h"
      18              : #include "ccu_ctx_creator_registry.h"
      19              : #include "executor_utils.h"
      20              : #include "ccu_context_all_gather_mesh1d_detour.h"
      21              : #include "ccu_temp_all_gather_mesh_1D_detour.h"
      22              : 
      23              : namespace Hccl {
      24              : 
      25              : static CcuInstRegister<CcuContextAllGatherMeshDetour1D> g_registrarAllGather(CcuInstType::CCU_ALLGATHER_MESH_1D_DETOUR);
      26              : 
      27            0 : CcuTempAllGatherMeshDetour1D::CcuTempAllGatherMeshDetour1D(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 : CcuTempAllGatherMeshDetour1D::~CcuTempAllGatherMeshDetour1D()
      35              : {
      36            0 : }
      37              : 
      38            0 : HcclResult CcuTempAllGatherMeshDetour1D::CalcResDetour(const RankGraph *rankGraph, AlgTempResReq &tempResReq)
      39              : {
      40              :     // 当前仅支持2P或4P
      41            0 :     CHK_PRT_RET(tempRankSize_ != 2 && tempRankSize_ != 4,
      42              :         HCCL_ERROR("[CcuTempAllGatherMeshDetour1D] Invalid RankSize[%u].", tempRankSize_), HcclResult::HCCL_E_INTERNAL);
      43              : 
      44            0 :     tempResReq.queNum = 1;  // 当前只有一个ccu mission,暂定1条流
      45            0 :     tempResReq.streamNum = tempResReq.queNum;
      46            0 :     HCCL_INFO("[CcuTempAllGatherMeshDetour1D][CalcResDetour] tempResReq.queNum[%u]", tempResReq.queNum);
      47              :     u32 myAlgRank;
      48            0 :     CHK_RET(GetAlgRank(myRank_, tempVTopo_[0], myAlgRank));
      49              : 
      50            0 :     for (u32 queIdx = 0; queIdx < tempVTopo_[0].size() - 1; queIdx++) {
      51              :         // find neighbors : virtualRank
      52            0 :         RankId neighborRank = tempVTopo_[0][(myAlgRank + 1 + queIdx) % tempRankSize_];
      53            0 :         uint32_t linkNum = GetPathsFromRankGraph(rankGraph, myRank_, neighborRank).size();
      54            0 :         tempResReq.links[neighborRank] = linkNum;
      55            0 :         HCCL_INFO("[CcuTempAllGatherMeshDetour1D][CalcResDetour] RankSize[%u], MyRank[%d]--Neighbor[%d], linkNum[%u]",
      56              :             tempRankSize_, myRank_, neighborRank, linkNum);
      57              : 
      58              :         // 2P支持2,3,4条link,4P支持2条link,注意绕路link分两条
      59            0 :         CHK_PRT_RET((tempRankSize_ == 2 && (linkNum <= 1 || linkNum > 1 + 3 * 2)) ||
      60              :                     (tempRankSize_ == 4 && linkNum != 1 + 1 * 2),  // 4P场景下,1条直连,绕路拆成2条
      61              :             HCCL_ERROR("[CcuTempAllGatherMeshDetour1D][CalcResDetour] Invalid linkNum[%u] for RankSize[%u].", linkNum, tempRankSize_),
      62              :                 HcclResult::HCCL_E_INTERNAL);
      63            0 :         if (queIdx == 0) {
      64            0 :             detourPathNum_ = (tempRankSize_ == 2) ? (linkNum - 1) / 2 : 1;  // 2P时去掉直连有2N条绕路link,对应N个绕路路径
      65            0 :             pathNumPerPeer_ = (tempRankSize_ == 2) ? (detourPathNum_ + 1) : detourPathNum_ + 2;  // 4P直连有2条,固定3条
      66            0 :             HCCL_INFO("[CcuTempAllGatherMeshDetour1D][CalcResDetour] detourPathNum[%u], pathNum[%u]", detourPathNum_, pathNumPerPeer_);
      67              :         }
      68              :     }
      69              : 
      70            0 :     return HcclResult::HCCL_SUCCESS;
      71              : }
      72              : 
      73            0 : HcclResult CcuTempAllGatherMeshDetour1D::CalcResDetour(ConnectedLinkMgr *linkMgr, AlgTempResReq &tempResReq)
      74              : {
      75              :     (void)linkMgr;
      76              :     (void)tempResReq;
      77            0 :     HCCL_ERROR("[InsCollAlgFactory] Unsupported interface of resource calculation!");
      78            0 :     return HcclResult::HCCL_E_INTERNAL;
      79              : }
      80              : 
      81            0 : HcclResult CcuTempAllGatherMeshDetour1D::CalcSliceInfo(const AllignInfo &allignInfo, const u64 dataSize,
      82              :                                             RankSliceInfo &sliceInfoVec)
      83              : {
      84            0 :     std::vector<SliceInfo> tmp(tempVTopo_.size());
      85            0 :     sliceInfoVec.resize(tempRankSize_, tmp);
      86              : 
      87            0 :     CHK_RET(CalcRsAgSliceInfoMesh(myRank_, tempRankSize_, allignInfo, dataSize, sliceInfoVec));
      88              : 
      89            0 :     return HcclResult::HCCL_SUCCESS;
      90            0 : }
      91              : 
      92            0 : void CcuTempAllGatherMeshDetour1D::CalcDetourOffset(
      93              :     uint64_t sliceSize, uint64_t &tailOffset, uint64_t &tailSize, uint64_t &loopIterNum)
      94              : {
      95            0 :     constexpr uint64_t MS_SIZE = 4096;
      96            0 :     uint64_t loopSize = pathNumPerPeer_ * MS_SIZE * CcuRep::CCU_MS_DEFAULT_LOOP_COUNT;  // 整块迭代
      97            0 :     loopIterNum = sliceSize / loopSize;
      98            0 :     tailSize = sliceSize % loopSize;
      99            0 :     tailOffset = sliceSize - tailSize;
     100              : 
     101            0 :     singleTransportSize_ = 0;
     102            0 :     lengths_.clear();  // 多轮情况下每轮都需要清零
     103            0 :     for (uint32_t i = 0; i < pathNumPerPeer_; i++) {
     104            0 :         lengths_.emplace_back(MS_SIZE);
     105            0 :         singleTransportSize_ += MS_SIZE;
     106              :     }
     107            0 :     return;
     108              : }
     109              : 
     110            0 : void CcuTempAllGatherMeshDetour1D::ProcessLinks(std::vector<LinkData> &links, const ResLinks &tempLinks) const
     111              : {
     112              :     // 整理links,要区分sendOnly与recvOnly,根据读写操作选择不同的绕路link
     113              :     // 固定2P用2-4条链路,每个链路用一个ms;4P用2条链路,其中直连用2个ms,绕路用1个
     114            0 :     std::vector<LinkData> directLinks;
     115            0 :     std::vector<LinkData> sendLinks;  // sendOnly
     116            0 :     std::vector<LinkData> recvLinks;  // recvOnly
     117            0 :     for (auto &pair : tempLinks) {
     118            0 :         if (pair.second.empty()) {
     119            0 :             continue;
     120              :         }
     121            0 :         HCCL_INFO("[CcuTempAllGatherMeshDetour1D][ProcessLinks] rankId[%d], linkSize[%zu]", pair.first, pair.second.size());
     122            0 :         for (uint32_t i = 0; i < pair.second.size(); i++) {
     123            0 :             LinkData curLink = pair.second[i];
     124            0 :             if (curLink.GetHop() == 1) {
     125            0 :                 directLinks.emplace_back(curLink);
     126            0 :             } else if (curLink.GetDirection() == LinkDirection::SEND_ONLY) {
     127            0 :                 sendLinks.emplace_back(curLink);
     128            0 :             } else if (curLink.GetDirection() == LinkDirection::RECV_ONLY) {
     129            0 :                 recvLinks.emplace_back(curLink);
     130              :             } else {
     131            0 :                 THROW<InvalidParamsException>(StringFormat(
     132            0 :                     "[CcuTempAllGatherMeshDetour1D][ProcessLinks] Rank[%d]--Peer[%d]--link[%d], unexpected link type.", myRank_, pair.first, i));
     133              :             }
     134              :         }
     135              :     }
     136              : 
     137              :     // 校验link
     138            0 :     if (sendLinks.size() != recvLinks.size() || directLinks.size() != tempRankSize_ - 1 ||
     139            0 :         sendLinks.size() % directLinks.size() != 0 || recvLinks.size() % directLinks.size() != 0) {
     140            0 :         THROW<InvalidParamsException>(StringFormat(
     141              :             "[CcuTempAllGatherMeshDetour1D] directSize[%zu]-sendSize[%zu]-recvSize[%zu].", directLinks.size(), sendLinks.size(), recvLinks.size()));
     142              :     }
     143            0 :     for (uint32_t i = 0; i < directLinks.size(); i++) {
     144            0 :         HCCL_INFO("Peer[%d][%s]", directLinks[i].GetRemoteRankId(), directLinks[i].GetDirection().Describe().c_str());
     145            0 :         links.emplace_back(directLinks[i]);
     146              :     }
     147            0 :     for (uint32_t i = 0; i < sendLinks.size(); i++) {
     148            0 :         HCCL_INFO("Peer[%d][%s]", sendLinks[i].GetRemoteRankId(), sendLinks[i].GetDirection().Describe().c_str());
     149            0 :         links.emplace_back(sendLinks[i]);
     150              :     }
     151            0 :     for (uint32_t i = 0; i < recvLinks.size(); i++) {
     152            0 :         HCCL_INFO("Peer[%d][%s]", recvLinks[i].GetRemoteRankId(), recvLinks[i].GetDirection().Describe().c_str());
     153            0 :         links.emplace_back(recvLinks[i]);
     154              :     }
     155              : 
     156            0 :     return;
     157            0 : }
     158              : 
     159            0 : void CcuTempAllGatherMeshDetour1D::GetAddrInfo(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
     160              :     uint64_t &inputAddr, uint64_t &outputAddr, uint64_t &offset)
     161              : {
     162            0 :     if (opMode_ == OpMode::OPBASE) {
     163            0 :         if (tempFuncs.isForepart) {
     164            0 :             inputAddr = BufferTypeToAddr(tempFuncs.usrData.usrInSlices[0].GetType())
     165            0 :                 + tempFuncs.usrData.usrInSlices[0].GetOffset();
     166              :         } else {
     167            0 :             inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
     168              :         }
     169            0 :         if (tempFuncs.isBottom) {
     170            0 :             outputAddr = BufferTypeToAddr(tempFuncs.usrData.usrOutSlices[0].GetType());
     171              :             // 需要加上 UserOUt 的偏移,包含了 loop 偏移和 rank 偏移
     172            0 :             offset = tempFuncs.usrData.usrOutSlices[myRank_].GetOffset();
     173              :         } else {
     174            0 :             outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
     175              :             // 从 inBuff 获取数据,只需要加上 rank 偏移
     176            0 :             offset = sliceInfoVec[myRank_][0].offset;
     177              :         }
     178              :     } else {
     179              :         // 图模式没有 tempFuncs.usrData,直接通过 buffInfo_ 获取输入输出地址
     180            0 :         inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff + tempFuncs.usrData.usrInSlices[0].GetOffset();
     181            0 :         outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
     182            0 :         offset = tempFuncs.usrData.usrOutSlices[myRank_].GetOffset();
     183              :     }
     184              : 
     185            0 :     return;
     186              : }
     187              : 
     188            0 : HcclResult CcuTempAllGatherMeshDetour1D::Run(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
     189              :                                           const BuffInfo &buffInfo, const ResLinks &tempLinks,
     190              :                                           std::vector<InsQuePtr> &tempInsQues)
     191              : {
     192            0 :     CHK_PRT_RET(tempInsQues.empty(),
     193              :         HCCL_ERROR("[CcuTempAllGatherMeshDetour1D] empty queue"), HcclResult::HCCL_E_INTERNAL);
     194            0 :     CHK_PTR_NULL(tempInsQues[0]);
     195            0 :     opMode_ = tempFuncs.opMode;
     196            0 :     buffInfo_ = buffInfo;
     197            0 :     CcuInstructionAllGatherMeshDetour1D ccuInsAllGatherMeshDetour1D;
     198              : 
     199            0 :     std::vector<uint64_t> dimSize;
     200            0 :     dimSize.push_back(tempRankSize_);
     201              : 
     202              :     uint64_t inputAddr;
     203              :     uint64_t outputAddr;
     204              :     uint64_t offset;
     205            0 :     GetAddrInfo(tempFuncs, sliceInfoVec, inputAddr, outputAddr, offset);
     206              : 
     207            0 :     std::vector<LinkData> links;
     208            0 :     uint64_t sliceSize = sliceInfoVec[myRank_][0].size;  // 获取本rank需要处理的数据量
     209              :     uint64_t token;
     210            0 :     CHK_RET(GetToken(op_, token));
     211              :     uint64_t tailOffset;
     212              :     uint64_t tailSize;
     213              :     uint64_t loopIterNum;
     214            0 :     CalcDetourOffset(sliceSize, tailOffset, tailSize, loopIterNum);
     215            0 :     ProcessLinks(links, tempLinks);
     216              : 
     217            0 :     ccuInsAllGatherMeshDetour1D.InitDetourInfo(
     218            0 :         static_cast<uint32_t>(myRank_), inputAddr, outputAddr, token, offset, tailOffset, tailSize, loopIterNum,
     219            0 :         lengths_, singleTransportSize_, detourPathNum_, pathNumPerPeer_, op_, tempVTopo_);
     220              : 
     221            0 :     HCCL_INFO("[CcuTempAllGatherMeshDetour1D] Run Init: myRank_[%d], dimSize[%llu], inputAddr[%llu],outputAddr[%llu],"\
     222              : "sliceSize[%llu], baseOffset[%llu], tailOffset[%llu], tailSize[%llu], loopIterNum[%llu],"\
     223              : "singleTransportSize[%llu], detourPathNum[%u], pathNumPerPeer[%u], links.size[%llu]",
     224              :         myRank_, dimSize[0], inputAddr, outputAddr, sliceSize, offset, tailOffset, tailSize, loopIterNum,
     225              :         singleTransportSize_, detourPathNum_, pathNumPerPeer_, links.size());
     226              : 
     227            0 :     ccuInsAllGatherMeshDetour1D.SetLinks(links);
     228              : 
     229            0 :     RankGroup rankGroup;
     230            0 :     for (auto &peer : tempVTopo_[0]) {
     231            0 :         rankGroup.AddRank(peer);
     232              :     }
     233            0 :     u32 cntCkeNum = 3;
     234            0 :     ccuInsAllGatherMeshDetour1D.SetCntCkeNum(cntCkeNum);
     235            0 :     ccuInsAllGatherMeshDetour1D.SetRankGroup(rankGroup);
     236            0 :     HCCL_INFO("ccuInsAllGatherMeshDetour1D is [%s]", ccuInsAllGatherMeshDetour1D.Describe().c_str());
     237            0 :     ccuInsAllGatherMeshDetour1D.Describe();
     238            0 :     tempInsQues[0]->Append(std::move(
     239            0 :         std::make_unique<CcuInstructionAllGatherMeshDetour1D>(ccuInsAllGatherMeshDetour1D)));
     240              : 
     241            0 :     return HcclResult::HCCL_SUCCESS;
     242            0 : }
     243              : } // namespace Hccl
        

Generated by: LCOV version 2.0-1