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_aligned_all_gather_double_ring_for_910_93_executor.h"
      12              : #include "hccl_types.h"
      13              : 
      14              : namespace hccl {
      15            2 : CollAlignedAllGatherDoubleRingFor91093Executor::CollAlignedAllGatherDoubleRingFor91093Executor(
      16              :     const HcclDispatcher dispatcher,
      17            2 :     std::unique_ptr<TopoMatcher> &topoMatcher)
      18            2 :     : CollAllGatherRingFor91093Executor(dispatcher, topoMatcher)
      19              : {
      20            2 :     DMAReduceFlag_ = workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE;
      21            2 : }
      22              : 
      23           18 : HcclResult CollAlignedAllGatherDoubleRingFor91093Executor::RunIntraSeverAllGather(
      24              :     const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
      25              :     const u64 count, const HcclDataType &dataType, const std::vector<std::vector<Slice>> &multRingsSliceZero,
      26              :     const Stream &stream, s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
      27              :     const std::vector<std::vector<Slice>> &multRingsUserMemSlice)
      28              : {
      29           18 :     CHK_RET(DoubleRingAllGather(tag, inputMem, outputMem, count, dataType,
      30              :         multRingsSliceZero, stream, profStage, baseOffset, opInfo, multRingsUserMemSlice));
      31           18 :     return HCCL_SUCCESS;
      32              : }
      33              : 
      34           18 : HcclResult CollAlignedAllGatherDoubleRingFor91093Executor::DoubleRingAllGather(
      35              :     const std::string &tag, DeviceMem inputMem, DeviceMem outputMem,
      36              :     const u64 count, const HcclDataType dataType, const std::vector<std::vector<Slice> > multRingsSliceZero,
      37              :     Stream stream, s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
      38              :     const std::vector<std::vector<Slice>> multRingsUserMemSlice)
      39              : {
      40           18 :     HCCL_CONFIG_INFO(HCCL_ALG, "[CollAlignedAllGatherDoubleRingFor91093Executor]userRank[%u], count[%llu]",
      41              :         topoAttr_.userRank, count);
      42              : 
      43              :     (void)tag;
      44           18 :     HCCL_INFO("[CollAlignedAllGatherDoubleRingFor91093Executor][DoubleRingAllGather] DoubleRingAllGather starts");
      45           18 :     HcclResult ret = HCCL_SUCCESS;
      46           18 :     u32 ringNum = multRingsSliceZero.size();
      47           18 :     CHK_RET(CheckCommSize(COMM_LEVEL0, ringNum));
      48              :     // 拿到ring环映射关系
      49           18 :     SubCommInfo level0ZeroCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
      50           18 :     auto nicList = topoAttr_.nicList;
      51              :     std::vector<std::vector<u32>> multiRingsOrder =
      52           18 :         GetRingsOrderByTopoType(level0ZeroCommInfo.localRankSize, topoType_, nicList);
      53              :     // 生成两个ring上的userMemOut_上对应的slices
      54           18 :     std::vector<std::vector<Slice>> userMemOutputSlicesOfDoubleRing;
      55           18 :     CHK_RET(CollectMultiRingsUserMemSlices(ringNum, dataType, opInfo, multRingsSliceZero,
      56              :         multiRingsOrder, multRingsUserMemSlice, userMemOutputSlicesOfDoubleRing));
      57              :     // 生成两个ring上的rankOrder
      58           18 :     std::vector<std::vector<u32>> rankOrders;
      59           18 :     CHK_RET(CollectMultiRingsRankOrder(ringNum, multiRingsOrder, rankOrders));
      60              :     // 初始化executor
      61           18 :     std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
      62           18 :         TemplateType::TEMPLATE_ALIGNED_ALL_GATHER_DOUBLE_RING, dispatcher_);
      63           18 :     HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALIGNED_ALL_GATHER_DOUBLE_RING in COMM_LEVEL0", __func__);
      64           18 :     CHK_SMART_PTR_NULL(tempAlg);
      65           18 :     CHK_RET(tempAlg->Prepare(const_cast<HcomCollOpInfo *>(opInfo), topoAttr_.userRank, algResResp_->slaveStreams,
      66              :         algResResp_->notifiesMain, algResResp_->notifiesAux, rankOrders, userMemOutputSlicesOfDoubleRing));
      67              : 
      68           18 :     ret = tempAlg->Prepare(outputMem, outputMem, inputMem, count, dataType, stream, multRingsSliceZero,
      69              :         HCCL_REDUCE_RESERVED, LEVEL0_BRIDGE_RANK_ID, baseOffset);
      70           18 :     CHK_PRT_RET(ret != HCCL_SUCCESS,
      71              :         HCCL_ERROR("[CollAlignedAllGatherDoubleRingFor91093Executor][DoubleRingAllGather]Double ring "
      72              :         "AllGather failed, return[%d]", ret), ret);
      73           18 :     u32 ringIndexOp = COMM_INDEX_0;
      74           18 :     u32 rankSize = level0ZeroCommInfo.localRankSize;
      75           18 :     ret = tempAlg->RegisterProfiler(
      76           18 :         ((ringIndexOp + 1) << PROF_RINGINDEX_OFFSET_OF_PLANEID) +
      77           18 :         (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0ZeroCommInfo.localRank,
      78              :         profStage, HCCL_EXEC_STEP_NOT_SET, stream);
      79           18 :     CHK_PRT_RET(ret != HCCL_SUCCESS,
      80              :         HCCL_ERROR("[CollAlignedAllGatherDoubleRingFor91093Executor][DoubleRingAllGather]Double ring "
      81              :         "AllGather failed, return[%d]", ret), ret);
      82              : 
      83              :     // 空拷贝用于后续操作附着
      84           18 :     CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem, outputMem, stream, dispatcher_));
      85           18 :     ret = RunTemplate(tempAlg, level0ZeroCommInfo);
      86           18 :     CHK_PRT_RET(ret != HCCL_SUCCESS,
      87              :         HCCL_ERROR("[CollAlignedAllGatherDoubleRingFor91093Executor][DoubleRingAllGather] Double ring "
      88              :                    "AllGather failed, return[%d]", ret), ret);
      89              :     // 添加空task,保证执行时不乱序
      90           18 :     CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem, outputMem, stream, dispatcher_));
      91           18 :     return HCCL_SUCCESS;
      92           18 : }
      93              : 
      94              : REGISTER_EXEC("AlignedAllGatherDoubleRingFor91093Executor", AlignedAllGatherDoubleRingFor91093,
      95              :     CollAlignedAllGatherDoubleRingFor91093Executor);
      96              : 
      97              : } // namespace hccl