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 "coll_all_gather_semi_ring_executor.h"
      12              : 
      13              : namespace hccl {
      14              : 
      15            0 : CollAllGatherSemiRingExecutor::CollAllGatherSemiRingExecutor(const HcclDispatcher dispatcher,
      16            0 :     std::unique_ptr<TopoMatcher> &topoMatcher)
      17            0 :     : CollAllGatherRingFor91093Executor(dispatcher, topoMatcher)
      18              : {
      19            0 : }
      20              : 
      21            0 : HcclResult CollAllGatherSemiRingExecutor::CalcNotifyNum(u32 streamNum, u32 &notifyNum)
      22              : {
      23              :     // notify数量是从流的两倍 + 新增带notifyId的notify资源
      24            0 :     notifyNum = 2U * streamNum + (topoAttr_.deviceNumPerAggregation + 4U);
      25            0 :     HCCL_INFO("[CollAllGatherSemiRingExecutor][CalcNotifyNum]tag[%s] notifyNum_ is [%u]", tag_.c_str(), notifyNum);
      26            0 :     return HCCL_SUCCESS;
      27              : }
      28              : 
      29            0 : HcclResult CollAllGatherSemiRingExecutor::CalcStreamNum(u32& streamNum)
      30              : {
      31            0 :     u32 totalStreamNum = LEVEL0_PLANE_NUM_IN_NPRING_DOUBLE + 1U;
      32              : 
      33            0 :     streamNum = totalStreamNum - 1U;
      34            0 :     HCCL_INFO("[CollAllGatherSemiRingExecutor][CalcStreamNum] tag[%s] streamNum_[%u]",
      35              :         tag_.c_str(), streamNum);
      36            0 :     return HCCL_SUCCESS;
      37              : }
      38              : 
      39            0 : HcclResult CollAllGatherSemiRingExecutor::CalcLevel0CommInfo(TransportMemType inputType,
      40              :     TransportMemType outputType,
      41              :     std::vector<LevelNSubCommTransport>& opTransport)
      42              : {
      43            0 :     CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
      44            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
      45              : 
      46            0 :     LevelNSubCommTransport &commTransportLevel0 = opTransport[COMM_LEVEL0];
      47            0 :     for (u32 subCommIndex = 0; subCommIndex < commTransportLevel0.size(); subCommIndex++) {
      48            0 :         for (auto &transportRequest : commTransportLevel0[subCommIndex].transportRequests) {
      49            0 :             transportRequest.notifyNum = topoAttr_.deviceNumPerAggregation + 4U; //只传递额外的notify个数
      50            0 :             HCCL_INFO("[CollAllGatherSemiRingExecutor][CalcLevel0CommInfo] set extral notifyNum[%u]",
      51              :                 transportRequest.notifyNum);
      52              :         }
      53              :     }
      54            0 :     return HCCL_SUCCESS;
      55            0 : }
      56              : 
      57            0 : HcclResult CollAllGatherSemiRingExecutor::DoubleRingMidCountAllGather(
      58              :     const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
      59              :     const u64 count, const HcclDataType &dataType, const std::vector<std::vector<Slice>> &multRingsSliceZero,
      60              :     const Stream &stream, s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
      61              :     const std::vector<std::vector<Slice>> &multRingsUserMemSlice)
      62              : {
      63              :     (void) multRingsSliceZero;
      64              :     (void) baseOffset;
      65              :     (void) opInfo;
      66              :     (void) tag;
      67            0 :     HCCL_CONFIG_INFO(HCCL_ALG,
      68              :         "[CollAllGatherSemiRingExecutor][KernelRun]CollAllGatherSemiRingExecutor starts.");
      69              :     
      70            0 :     CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
      71            0 :     SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
      72              : 
      73              :     // 执行
      74            0 :     std::unique_ptr<AlgTemplateBase> executor = AlgTemplateRegistry::Instance().GetAlgTemplate(
      75            0 :         TemplateType::TEMPLATE_ALL_GATHER_UNIFIED_MARCH, dispatcher_);
      76            0 :     HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALL_GATHER_UNIFIED_MARCH in COMM_LEVEL0", __func__);
      77            0 :     CHK_SMART_PTR_NULL(executor);
      78              : 
      79            0 :     CHK_RET(executor->Prepare(stream, level0CommInfo, algResResp_->paramInputMem, algResResp_->paramOutputMem,
      80              :         inputMem, outputMem, count * SIZE_TABLE[dataType], algResResp_->slaveStreams, algResResp_->notifiesMain,
      81              :         algResResp_->notifiesAux, multRingsUserMemSlice));
      82              : 
      83            0 :     HcclResult ret = executor->RegisterProfiler(
      84              :         ((COMM_INDEX_0 + 1) << PROF_RINGINDEX_OFFSET_OF_PLANEID) +
      85            0 :         (level0CommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
      86              :         profStage, HCCL_EXEC_STEP_NOT_SET, stream);
      87            0 :     CHK_PRT_RET(ret != HCCL_SUCCESS,
      88              :         HCCL_ERROR("[CollAllGatherSemiRingExecutor][DoubleRingMidCountAllGather]Double ring "
      89              :         "AllGather failed, return[%d]", ret), ret);
      90              : 
      91            0 :     CHK_RET(executor->RunAsync());
      92              : 
      93            0 :     HCCL_INFO("[CollAllGatherSemiRingExecutor] AllGather double ring level1 run success");
      94            0 :     return ret;
      95            0 : }
      96              : 
      97            0 : HcclResult CollAllGatherSemiRingExecutor::RunIntraSeverAllGather(
      98              :     const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
      99              :     const u64 count, const HcclDataType &dataType, const std::vector<std::vector<Slice>> &multRingsSliceZero,
     100              :     const Stream &stream, s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
     101              :     const std::vector<std::vector<Slice>> &multRingsUserMemSlice)
     102              : {
     103            0 :     CHK_RET(DoubleRingMidCountAllGather(tag, inputMem, outputMem, count, dataType,
     104              :         multRingsSliceZero, stream, profStage, baseOffset, opInfo, multRingsUserMemSlice));
     105            0 :     return HCCL_SUCCESS;
     106              : }
     107              : 
     108              : REGISTER_EXEC("AllGatherSemiRingExecutor", AllGatherDoubleRingMidCount,
     109              :     CollAllGatherSemiRingExecutor);
     110              : 
     111              : } // namespace hccl