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_reduce_scatter_mesh_graph_pipeline_executor.h"
      12              : 
      13              : namespace hccl {
      14            0 : CollReduceScatterMeshGraphPipelineExecutor::CollReduceScatterMeshGraphPipelineExecutor(
      15            0 :     const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher)
      16            0 :     : CollReduceScatterExecutor(dispatcher, topoMatcher)
      17            0 : {}
      18              : 
      19            0 : void CollReduceScatterMeshGraphPipelineExecutor::ParseParam(const OpParam &param)
      20              : {
      21            0 :     tag_ = param.tag;
      22            0 :     aicpuUnfoldMode_ = param.aicpuUnfoldMode;
      23            0 : }
      24              : 
      25            0 : HcclResult CollReduceScatterMeshGraphPipelineExecutor::CalcStreamNum(u32 &streamNum)
      26              : {
      27            0 :     streamNum = topoAttr_.deviceNumPerAggregation;
      28            0 :     HCCL_INFO(
      29              :         "[CollReduceScatterMeshGraphPipelineExecutor][CalcStreamNum] tag[%s] streamNum[%u]", tag_.c_str(), streamNum);
      30            0 :     return HCCL_SUCCESS;
      31              : }
      32              : 
      33            0 : HcclResult CollReduceScatterMeshGraphPipelineExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport> &opTransport)
      34              : {
      35            0 :     TransportMemType inputType = TransportMemType::RESERVED;
      36            0 :     TransportMemType outputType = TransportMemType::RESERVED;
      37            0 :     CHK_RET(CalcTransportMemType(inputType, outputType));
      38            0 :     CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
      39            0 :     CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
      40            0 :     return HCCL_SUCCESS;
      41              : }
      42              : 
      43            0 : HcclResult CollReduceScatterMeshGraphPipelineExecutor::CalcTransportMemType(
      44              :     TransportMemType &inputType, TransportMemType &outputType)
      45              : {
      46            0 :     inputType = TransportMemType::PARAM_INPUT;
      47            0 :     outputType = TransportMemType::PARAM_OUTPUT;
      48            0 :     HCCL_INFO("[CollReduceScatterMeshGraphPipelineExecutor][CalcTransportMemType]tag[%s] inputType[%d], outputType[%d]",
      49              :         tag_.c_str(),
      50              :         inputType,
      51              :         outputType);
      52            0 :     return HCCL_SUCCESS;
      53              : }
      54              : 
      55            0 : HcclResult CollReduceScatterMeshGraphPipelineExecutor::CalcLevel0CommInfo(
      56              :     TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport> &opTransport)
      57              : {
      58            0 :     CommParaInfo commParaInfo(COMM_LEVEL0, CommType::COMM_TAG_MESH);
      59            0 :     commParaInfo.meshSinglePlane = true;
      60            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL0], inputType, outputType));
      61            0 :     return HCCL_SUCCESS;
      62            0 : }
      63              : 
      64              : // PipeLine模式下使用Ring算法
      65            0 : HcclResult CollReduceScatterMeshGraphPipelineExecutor::CalcLevel1CommInfo(
      66              :     TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport> &opTransport)
      67              : {
      68            0 :     CommParaInfo commParaInfo(COMM_LEVEL1, CommType::COMM_TAG_RING_INNER);
      69            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL1], inputType, outputType));
      70            0 :     return HCCL_SUCCESS;
      71            0 : }
      72              : 
      73            0 : HcclResult CollReduceScatterMeshGraphPipelineExecutor::KernelRun(const OpParam &param, ExecMem &execMem)
      74              : {
      75            0 :     HCCL_CONFIG_INFO(HCCL_ALG,
      76              :         "[CollReduceScatterMeshGraphPipelineExecutor][KernelRun] userRank[%u] starts.", topoAttr_.userRank);
      77            0 :     CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
      78            0 :     SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
      79            0 :     u32 commIndex = level0CommInfo.localRank;
      80            0 :     CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
      81            0 :     SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
      82              : 
      83            0 :     u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
      84            0 :     DeviceMem userInMem = DeviceMem::create(param.inputPtr, param.DataDes.count * unitSize);
      85            0 :     u64 reduceAttr = GetReduceAttr(userInMem, userInMem, param.DataDes.dataType, param.reduceType);
      86              : 
      87            0 :     CHK_RET(ActiveSlaveStreams(param.stream));
      88              : 
      89            0 :     std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
      90            0 :         TemplateType::TEMPLATE_REDUCESCATTER_GRAPH_PIPELINE, dispatcher_);
      91            0 :     CHK_SMART_PTR_NULL(tempAlg);
      92              : 
      93            0 :     HcomCollOpInfo opInfo = {"",
      94            0 :         execMem.inputPtr,
      95            0 :         execMem.outputPtr,
      96            0 :         param.DataDes.count,
      97            0 :         param.DataDes.dataType,
      98            0 :         param.root,
      99            0 :         param.reduceType};
     100              : 
     101            0 :     CHK_RET(tempAlg->Prepare(&opInfo,
     102              :         execMem.inputMem,
     103              :         param.DataDes.count,
     104              :         0,
     105              :         0,
     106              :         level0CommInfo,
     107              :         level1CommInfo,
     108              :         const_cast<Stream &>(param.stream),
     109              :         algResResp_->slaveStreams,
     110              :         algResResp_->notifiesMain,
     111              :         algResResp_->notifiesAux,
     112              :         reduceAttr));
     113            0 :     CHK_RET(tempAlg->RunAsync());
     114              : 
     115            0 :     return HCCL_SUCCESS;
     116            0 : }
     117              : 
     118              : REGISTER_EXEC("ReduceScatterMeshGraphPipelineExecutor", ReduceScatterMeshGraphPipeline,
     119              :     CollReduceScatterMeshGraphPipelineExecutor);
     120              : }  // namespace hccl