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_v_mesh_graph_pipeline_executor.h"
      12              : 
      13              : namespace hccl {
      14            0 : CollAllGatherVMeshGraphPipelineExecutor::CollAllGatherVMeshGraphPipelineExecutor(
      15            0 :     const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher)
      16            0 :     : CollAllGatherVExecutor(dispatcher, topoMatcher)
      17            0 : {}
      18              : 
      19            0 : HcclResult CollAllGatherVMeshGraphPipelineExecutor::CalcStreamNum(u32 &streamNum)
      20              : {
      21            0 :     u32 totalStreamNum = topoAttr_.deviceNumPerAggregation + 1U;
      22            0 :     streamNum = totalStreamNum - 1U;
      23            0 :     HCCL_INFO("[CollAllGatherVMeshGraphPipelineExecutor][CalcStreamNum] tag[%s] streamNum[%u]", tag_.c_str(), streamNum);
      24            0 :     return HCCL_SUCCESS;
      25              : }
      26              : 
      27            0 : HcclResult CollAllGatherVMeshGraphPipelineExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport> &opTransport)
      28              : {
      29            0 :     TransportMemType inputType = TransportMemType::RESERVED;
      30            0 :     TransportMemType outputType = TransportMemType::RESERVED;
      31            0 :     CHK_RET(CalcTransportMemType(inputType, outputType));
      32            0 :     CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
      33            0 :     CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
      34            0 :     return HCCL_SUCCESS;
      35              : }
      36              : 
      37            0 : HcclResult CollAllGatherVMeshGraphPipelineExecutor::CalcTransportMemType(
      38              :     TransportMemType &inputType, TransportMemType &outputType)
      39              : {
      40            0 :     inputType = TransportMemType::PARAM_INPUT;
      41            0 :     outputType = TransportMemType::PARAM_OUTPUT;
      42            0 :     HCCL_INFO("[CollAllGatherVMeshGraphPipelineExecutor][CalcTransportMemType]tag[%s] inputType[%d], outputType[%d]",
      43              :         tag_.c_str(), inputType, outputType);
      44            0 :     return HCCL_SUCCESS;
      45              : }
      46              : 
      47            0 : HcclResult CollAllGatherVMeshGraphPipelineExecutor::CalcLevel0CommInfo(
      48              :     TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport> &opTransport)
      49              : {
      50            0 :     CommParaInfo commParaInfo(COMM_LEVEL0, CommType::COMM_TAG_MESH);
      51            0 :     commParaInfo.meshSinglePlane = true;
      52            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL0], inputType, outputType));
      53            0 :     return HCCL_SUCCESS;
      54            0 : }
      55              : 
      56              : // PipeLine模式下使用Ring算法
      57            0 : HcclResult CollAllGatherVMeshGraphPipelineExecutor::CalcLevel1CommInfo(
      58              :     TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport> &opTransport)
      59              : {
      60            0 :     CommParaInfo commParaInfo(COMM_LEVEL1, CommType::COMM_TAG_RING_INNER);
      61            0 :     CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL1], inputType, outputType));
      62            0 :     return HCCL_SUCCESS;
      63            0 : }
      64              : 
      65            0 : HcclResult CollAllGatherVMeshGraphPipelineExecutor::KernelRun(const OpParam &param, ExecMem &execMem)
      66              : {
      67            0 :     HCCL_CONFIG_INFO(HCCL_ALG,
      68              :         "[CollAllGatherVMeshGraphPipelineExecutor][KernelRun]AllGatherMeshGraphPipelineExecutor begins.");
      69              : 
      70              :     // 先获取 comm level0 和 comm level1 的value
      71            0 :     CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
      72            0 :     SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
      73            0 :     u32 commIndex = level0CommInfo.localRank;
      74            0 :     CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
      75            0 :     SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
      76              : 
      77              :     // allgatherv 计算slice，数据分成ranksize份，每份的起始偏移和大小
      78            0 :     u32 perDataSize = SIZE_TABLE[param.VDataDes.dataType];
      79            0 :     std::vector<Slice> outputSlices;
      80            0 :     const auto counts = static_cast<u64*>(param.VDataDes.counts);
      81            0 :     const auto displs = static_cast<u64*>(param.VDataDes.displs);
      82            0 :     for (u32 rank = 0; rank < topoAttr_.userRankSize; ++rank) {
      83            0 :         Slice userslice;
      84            0 :         userslice.offset = displs[rank] * perDataSize;
      85            0 :         userslice.size = counts[rank] * perDataSize;
      86            0 :         outputSlices.emplace_back(std::move(userslice));
      87              :     }
      88              : 
      89              :     // 激活从流
      90            0 :     CHK_RET(ActiveSlaveStreams(param.stream));
      91              : 
      92            0 :     HcomCollOpInfo opInfo = {
      93            0 :         "", execMem.inputPtr, execMem.outputPtr, 0, param.VDataDes.dataType, 0, HCCL_REDUCE_RESERVED};
      94              : 
      95              :     std::unique_ptr<AlgTemplateBase> tempAlg =
      96            0 :         AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_V_GRAPH_PIPELINE, dispatcher_);
      97            0 :     CHK_SMART_PTR_NULL(tempAlg);
      98            0 :     CHK_RET(tempAlg->Prepare(&opInfo, topoAttr_.userRank, execMem.count, execMem.inputMem, execMem.outputMem,
      99              :                              level0CommInfo, level1CommInfo,  const_cast<Stream &>(param.stream),
     100              :                              algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux, outputSlices));
     101            0 :     CHK_RET(tempAlg->RunAsync());
     102            0 :     return HCCL_SUCCESS;
     103            0 : }
     104              : 
     105              : REGISTER_EXEC("AllGatherVMeshGraphPipelineExecutor", AllGatherVGraphPipeline, CollAllGatherVMeshGraphPipelineExecutor);
     106              : 
     107              : }  // namespace hccl