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_reduce_mesh_mid_count_executor.h"
      12              : 
      13              : namespace hccl {
      14              : 
      15            1 : CollAllReduceMeshMidCountExecutor::CollAllReduceMeshMidCountExecutor(const HcclDispatcher dispatcher,
      16            1 :                                                                      std::unique_ptr<TopoMatcher> &topoMatcher)
      17            1 :     : CollAllReduceExecutor(dispatcher, topoMatcher)
      18              : {
      19            1 :     CCLMemSlice_ = false;
      20            1 :     DMAReduceFlag_ = true;
      21            1 : }
      22              : 
      23            1 : HcclResult CollAllReduceMeshMidCountExecutor::CalcStreamNum(u32& streamNum)
      24              : {
      25            1 :     u32 totalStreamNum = INVALID_UINT;
      26            1 :     if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
      27            1 :         totalStreamNum = topoAttr_.deviceNumPerAggregation;
      28              :     } else {
      29            0 :         totalStreamNum = topoAttr_.deviceNumPerAggregation - 1U;
      30              :     }
      31            1 :     streamNum = totalStreamNum - 1U;
      32            1 :     HCCL_INFO("[CollAllReduceMeshMidCountExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
      33              :         tag_.c_str(), streamNum);
      34            1 :     return HCCL_SUCCESS;
      35              : }
      36              : 
      37            1 : HcclResult CollAllReduceMeshMidCountExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
      38              : {
      39            1 :     TransportMemType inputType = TransportMemType::RESERVED;
      40            1 :     TransportMemType outputType = TransportMemType::RESERVED;
      41            1 :     CHK_RET(CalcTransportMemType(inputType, outputType));
      42            1 :     CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
      43            1 :     return HCCL_SUCCESS;
      44              : }
      45              : 
      46            1 : HcclResult CollAllReduceMeshMidCountExecutor::CalcTransportMemType(TransportMemType &inputType,
      47              :     TransportMemType &outputType)
      48              : {
      49            1 :     if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
      50            1 :         inputType = TransportMemType::CCL_INPUT;
      51            1 :         outputType = TransportMemType::CCL_OUTPUT;
      52              :     } else {
      53            0 :         inputType = TransportMemType::PARAM_INPUT;  // userInput
      54            0 :         outputType = TransportMemType::PARAM_OUTPUT;
      55              :     }
      56            1 :     HCCL_INFO("[CollAllReduceMeshMidCountExecutor][CalcTransportMemType]tag[%s] inputType[%d], outputType[%d].",
      57              :         tag_.c_str(), inputType, outputType);
      58            1 :     return HCCL_SUCCESS;
      59              : }
      60              : 
      61            1 : HcclResult CollAllReduceMeshMidCountExecutor::CalcLevel0CommInfo(TransportMemType inputType,
      62              :     TransportMemType outputType,
      63              :     std::vector<LevelNSubCommTransport>& opTransport)
      64              : {
      65            1 :     CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
      66            1 :     commParaLevel0.meshSinglePlane = true;
      67            1 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
      68            1 :     return HCCL_SUCCESS;
      69            1 : }
      70              : 
      71            1 : bool CollAllReduceMeshMidCountExecutor::IsHugeData(const u64 curSize)
      72              : {
      73              :     // 只有server内通信，多QP哈希散列下不刷新子图
      74            1 :     bool hugeData = curSize > SDMA_SEND_MAX_SIZE;
      75            1 :     return hugeData;
      76              : }
      77              : 
      78            1 : bool CollAllReduceMeshMidCountExecutor::IsSmallData(const u64 totalSize, const u64 curSize)
      79              : {
      80            1 :     bool smallData = totalSize <= HCCL_SMALL_COUNT_128_KB;
      81            1 :     return smallData;
      82              : }
      83              : 
      84            1 : HcclResult CollAllReduceMeshMidCountExecutor::KernelRun(const OpParam &param, ExecMem &execMem)
      85              : {
      86            1 :     HCCL_CONFIG_INFO(HCCL_ALG,
      87              :         "[CollAllReduceMeshMidCountExecutor][KernelRun] userRank[%u] starts.", topoAttr_.userRank);
      88            1 :     std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
      89              : 
      90            1 :     CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
      91            1 :     SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
      92              : 
      93            1 :     CHK_RET(ActiveSlaveStreams(param.stream));
      94              : 
      95            1 :     u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
      96            1 :     HcomCollOpInfo opInfo = {
      97            1 :         "", execMem.inputPtr, execMem.outputPtr, execMem.count, param.DataDes.dataType, param.root, param.reduceType
      98            1 :     };
      99              : 
     100            1 :     std::unique_ptr<AlgTemplateBase> level0TempAlg;
     101            1 :     level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_LOCAL_REDUCE, dispatcher_);
     102            1 :     CHK_SMART_PTR_NULL(level0TempAlg);
     103            1 :     CHK_RET(level0TempAlg->Prepare(reduceAttr, algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux, 
     104              :         level0CommInfo.localRank, level0CommInfo.localRankSize, topoAttr_.userRank, &opInfo));
     105              : 
     106            3 :     CHK_RET(level0TempAlg->Prepare(execMem.outputMem, execMem.outputMem, execMem.outputMem, execMem.count,
     107              :         param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, dataSegsSlice, 0));
     108            1 :     CHK_RET(level0TempAlg->RegisterProfiler(
     109              :         (level0CommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
     110              :         PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
     111            1 :     CHK_RET(RunTemplate(level0TempAlg, level0CommInfo));
     112            1 :     HCCL_INFO("AllReduce mid count run success.");
     113              : 
     114            1 :     return HCCL_SUCCESS;
     115            1 : }
     116              : 
     117              : REGISTER_EXEC("AllReduceMeshMidCountLoopExecutor", AllReduceMeshMidCount, CollAllReduceMeshMidCountExecutor);
     118              : 
     119              : } // namespace hccl