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_graph_pipeline_executor.h"
      12              : 
      13              : namespace hccl {
      14              : // 准入条件: pipeLine && 910B && 图模式 && sdmaReduce && rdmaReduce && 多Mesh && MeshTopo && 非确定性
      15            0 : CollAllReduceMeshGraphPipelineExecutor::CollAllReduceMeshGraphPipelineExecutor(const HcclDispatcher dispatcher,
      16            0 :     std::unique_ptr<TopoMatcher> &topoMatcher): CollAllReduceExecutor(dispatcher, topoMatcher)
      17              : {
      18            0 : }
      19              : 
      20            0 : HcclResult CollAllReduceMeshGraphPipelineExecutor::CalcStreamNum(u32& streamNum)
      21              : {
      22            0 :     u32 totalStreamNum = topoAttr_.deviceNumPerAggregation;
      23            0 :     streamNum = totalStreamNum - 1U;
      24            0 :     HCCL_INFO("[CollAllReduceMeshGraphPipelineExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
      25              :         tag_.c_str(), streamNum);
      26            0 :     return HCCL_SUCCESS;
      27              : }
      28              : 
      29            0 : HcclResult CollAllReduceMeshGraphPipelineExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
      30              : {
      31            0 :     TransportMemType inputType = TransportMemType::RESERVED;
      32            0 :     TransportMemType outputType = TransportMemType::RESERVED;
      33            0 :     CHK_RET(CalcTransportMemType(inputType, outputType));
      34            0 :     CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
      35            0 :     CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
      36            0 :     return HCCL_SUCCESS;
      37              : }
      38              : 
      39            0 : HcclResult CollAllReduceMeshGraphPipelineExecutor::CalcTransportMemType(TransportMemType &inputType,
      40              :     TransportMemType &outputType)
      41              : {
      42            0 :     inputType = TransportMemType::PARAM_INPUT;
      43            0 :     outputType = TransportMemType::PARAM_OUTPUT;
      44            0 :     HCCL_INFO("[CollAllReduceMeshGraphPipelineExecutor][CalcTransportMemType]" \
      45              :         "tag[%s] inputType[%d], outputType[%d]",
      46              :         tag_.c_str(), inputType, outputType);
      47            0 :     return HCCL_SUCCESS;
      48              : }
      49              : 
      50            0 : HcclResult CollAllReduceMeshGraphPipelineExecutor::CalcLevel0CommInfo(TransportMemType inputType,
      51              :     TransportMemType outputType,
      52              :     std::vector<LevelNSubCommTransport>& opTransport)
      53              : {
      54            0 :     CommParaInfo commParaInfo(COMM_LEVEL0, CommType::COMM_TAG_MESH);
      55            0 :     commParaInfo.meshSinglePlane = true;
      56            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL0], inputType, outputType));
      57            0 :     return HCCL_SUCCESS;
      58            0 : }
      59              : 
      60              : // PipeLine模式下使用Ring算法
      61            0 : HcclResult CollAllReduceMeshGraphPipelineExecutor::CalcLevel1CommInfo(TransportMemType inputType,
      62              :     TransportMemType outputType,
      63              :     std::vector<LevelNSubCommTransport>& opTransport)
      64              : {
      65            0 :     CommParaInfo commParaInfo(COMM_LEVEL1, CommType::COMM_TAG_RING_INNER);
      66            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL1], inputType, outputType));
      67            0 :     return HCCL_SUCCESS;
      68            0 : }
      69              : 
      70            0 : HcclResult CollAllReduceMeshGraphPipelineExecutor::KernelRun(const OpParam &param, ExecMem &execMem)
      71              : {
      72            0 :     HCCL_CONFIG_INFO(HCCL_ALG,
      73              :         "[CollAllReduceMeshGraphPipelineExecutor][Run]CollAllReduceMeshGraphPipelineExecutor begins.");
      74              : 
      75            0 :     CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
      76            0 :     SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
      77            0 :     u32 commIndex = level0CommInfo.localRank;
      78            0 :     CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
      79            0 :     SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
      80              : 
      81            0 :     CHK_RET(ActiveSlaveStreams(param.stream));
      82              : 
      83            0 :     u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.inputMem, param.DataDes.dataType, param.reduceType);
      84              : 
      85            0 :     HcomCollOpInfo opInfo = {
      86            0 :         "", execMem.inputPtr, execMem.outputPtr, execMem.count, param.DataDes.dataType, param.root, param.reduceType
      87            0 :     };
      88              : 
      89            0 :     std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
      90            0 :         TemplateType::TEMPLATE_ALLREDUCE_GRAPH_PIPELINE, dispatcher_);
      91            0 :     CHK_SMART_PTR_NULL(tempAlg);
      92            0 :     CHK_RET(tempAlg->Prepare(reduceAttr));
      93            0 :     CHK_RET(tempAlg->Prepare(&opInfo, execMem.inputMem, execMem.outputMem, execMem.count,
      94              :         level1CommInfo, level0CommInfo, const_cast<Stream&>(param.stream),
      95              :         algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux));
      96            0 :     CHK_RET(tempAlg->RunAsync());
      97            0 :     return HCCL_SUCCESS;
      98            0 : }
      99              : 
     100              : REGISTER_EXEC("AllReduceMeshGraphPipelineExecutor",
     101              :     AllReduceMeshGraphPipeline, CollAllReduceMeshGraphPipelineExecutor);
     102              : 
     103              : } // namespace hccl