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              : #ifndef COLL_RUN_ALLTOALLV_TWO_LEVEL_PIPELINE_H
      12              : #define COLL_RUN_ALLTOALLV_TWO_LEVEL_PIPELINE_H
      13              : #include "coll_all_to_all_executor.h"
      14              : namespace hccl {
      15              : class CollRunAlltoAllVStaged : public CollAlltoAllExecutor {
      16              : 
      17              : public:
      18              :     CollRunAlltoAllVStaged(const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher);
      19            0 :     ~CollRunAlltoAllVStaged() override = default;
      20              : 
      21              :     HcclResult CheckNeedRecreateComm(u64 lastScratchMemSize, bool& needRecreateAlltoallComm) override;
      22              :     HcclResult CheckNeedCreateVirtualLinks(AlgResourceRequest &resourceRequest) override;
      23              :     HcclResult ParallelTaskLoaderProcess(const std::string &tag, Stream &stream, SubCommInfo &level0CommInfo,
      24              :         std::vector<Stream> &ringStreams);
      25              : 
      26              : private:
      27              :     HcclResult CalcStreamNum(u32& streamNum) override;
      28              :     void CalcWorkSpaceMemSize(const AlltoAllUserRankInfo &userRankInfo,
      29              :         const std::vector<SendRecvInfo> &allMeshAggregationSendRecvInfo, u64 &workspaceMemSize,
      30              :         u32 meshAggregationRankSize);
      31              :     HcclResult CalcScratchMemSize(u64& scratchMemSize) override;
      32              : 
      33              :     HcclResult CalcLevel0CommInfo(TransportMemType inputType, TransportMemType outputType,
      34              :         std::vector<LevelNSubCommTransport>& opTransport) override;
      35              :     HcclResult CalcLevel1CommInfo(TransportMemType inputType, TransportMemType outputType,
      36              :         std::vector<LevelNSubCommTransport>& opTransport) override;
      37              :     HcclResult CalcLevel2CommInfo(TransportMemType inputType, TransportMemType outputType,
      38              :         std::vector<LevelNSubCommTransport>& opTransport) override;
      39              :     HcclResult CalStagedAlltoallVCommInfo(TransportMemType inputType, TransportMemType outputType,
      40              :         std::vector<LevelNSubCommTransport>& opTransport);
      41              :     HcclResult CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport) override;
      42              :     HcclResult KernelRun(const OpParam &param, ExecMem &execMem) override;
      43              : 
      44              :     HcclResult PrepareAlltoAllVStaged1(DeviceMem &sendBuf, DeviceMem &recvBuf, DeviceMem &scratchMem,
      45              :         std::map<u32, std::list<OneSendRecvAddrInfo>> &sendAddrInfosIntra,
      46              :         std::map<u32, std::list<OneSendRecvAddrInfo>> &recvAddrInfosIntra,
      47              :         Stream &stream, const std::string &tag, std::unique_ptr<AlgTemplateBase> &alltoallLevel0,
      48              :         ExecMem &execMem);
      49              :     void CalcInterMeshAggregationRecvRemoteOffset(const AlltoAllUserRankInfo &userRankInfo,
      50              :         const std::vector<SendRecvInfo> &allSendRecvInfo, u32 index, u64 &remoteOffset, u32 meshAggregationRankSize);
      51              :     void CalcInterMeshAggregationAlltoAllMemInfo(
      52              :         const AlltoAllUserRankInfo &userRankInfo, const std::vector<SendRecvInfo> &allSendRecvInfo,
      53              :         std::map<u32, std::list<OneSendRecvAddrInfo>> &sendAddrInfosInter,
      54              :         std::map<u32, std::list<OneSendRecvAddrInfo>> &recvAddrInfosInter,
      55              :         u32 meshAggregationRankSize);
      56              :     HcclResult PrepareAlltoAllVStaged2(DeviceMem &recvBuf, DeviceMem &scratchMem,
      57              :         std::map<u32, std::list<OneSendRecvAddrInfo>> &sendAddrInfosInter,
      58              :         std::map<u32, std::list<OneSendRecvAddrInfo>> &recvAddrInfosInter,
      59              :         Stream &stream, const std::string &tag, std::unique_ptr<AlgTemplateBase> &alltoallLevel1,
      60              :         ExecMem &execMem);
      61              : };
      62              : 
      63              : } // namespace hccl
      64              : 
      65              : #endif