LCOV - code coverage report
Current view: top level - legacy/ascend910/algorithm/base/alg_template/temp_reduce_scatter - reduce_scatter_mesh_atomic_opbase.cc (source / functions) Coverage Total Hit
Test: coverage.info Lines: 16.0 % 81 13
Test Date: 2026-08-04 10:52:23 Functions: 44.4 % 9 4

            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 "reduce_scatter_mesh_atomic_opbase.h"
      12              : #include "alg_template_register.h"
      13              : 
      14              : namespace hccl {
      15              : using namespace std;
      16              : 
      17            1 : ReduceScatterMeshDirect::ReduceScatterMeshDirect(const HcclDispatcher dispatcher)
      18            1 :     : AlgTemplateBase(dispatcher)
      19            1 : {}
      20              : 
      21            2 : ReduceScatterMeshDirect::~ReduceScatterMeshDirect() {}
      22              : 
      23            1 : HcclResult ReduceScatterMeshDirect::Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem,
      24              :                                             const u64 count, const HcclDataType dataType, const Stream &stream,
      25              :                                             const HcclReduceOp reductionOp, const u32 root,
      26              :                                             const std::vector<Slice> &slices, const u64 baseOffset,
      27              :                                             const u64 reduceAttrBitMap, std::vector<Stream> &meshStreams,
      28              :                                             std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
      29              :                                             std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
      30              :                                             u32 userRank, const HcomCollOpInfo *opInfo)
      31              : {
      32            1 :     reduceAttr_ = reduceAttrBitMap;
      33            1 :     userRank_ = userRank;
      34            1 :     meshStreams_ = meshStreams;
      35            1 :     meshSignalPtr_ = &meshSignal;
      36            1 :     meshSignalAuxPtr_ = &meshSignalAux;
      37            1 :     opInfo_ = opInfo;
      38            4 :     return AlgTemplateBase::Prepare(inputMem, outputMem, scratchMem, count, dataType, stream, reductionOp,
      39            2 :         root, slices, baseOffset);
      40              : }
      41              : 
      42            0 : HcclResult ReduceScatterMeshDirect::MainRecordSub()
      43              : {
      44            0 :     for (u32 signalIndex = 0; signalIndex < meshSignalAuxPtr_->size(); signalIndex++) {
      45            0 :         CHK_RET(LocalNotify::Post(stream_, dispatcher_, (*meshSignalAuxPtr_)[signalIndex],
      46              :             profilerInput_.stage));
      47              :     }
      48            0 :     return HCCL_SUCCESS;
      49              : }
      50              : 
      51            0 : HcclResult ReduceScatterMeshDirect::SubWaitMain()
      52              : {
      53            0 :     for (u32 streamIndex = 0; streamIndex < meshSignalAuxPtr_->size(); streamIndex++) {
      54            0 :         CHK_RET(LocalNotify::Wait(meshStreams_[streamIndex], dispatcher_,
      55              :             (*meshSignalAuxPtr_)[streamIndex], profilerInput_.stage));
      56              :     }
      57            0 :     return HCCL_SUCCESS;
      58              : }
      59              : 
      60            0 : HcclResult ReduceScatterMeshDirect::MainWaitSub()
      61              : {
      62            0 :     for (u32 signalIndex = 0; signalIndex < meshSignalPtr_->size(); signalIndex++) {
      63            0 :         CHK_RET(LocalNotify::Wait(stream_, dispatcher_, (*meshSignalPtr_)[signalIndex], profilerInput_.stage));
      64              :     }
      65            0 :     return HCCL_SUCCESS;
      66              : }
      67              : 
      68            0 : HcclResult ReduceScatterMeshDirect::SubRecordMain()
      69              : {
      70            0 :     for (u32 streamIndex = 0; streamIndex < meshSignalPtr_->size(); streamIndex++) {
      71            0 :         CHK_RET(LocalNotify::Post(meshStreams_[streamIndex], dispatcher_, (*meshSignalPtr_)[streamIndex],
      72              :             profilerInput_.stage));
      73              :     }
      74            0 :     return HCCL_SUCCESS;
      75              : }
      76              : 
      77            0 : HcclResult ReduceScatterMeshDirect::RunAsync(const u32 rank, const u32 rankSize, const std::vector<LINK> &links)
      78              : {
      79            0 :     HCCL_INFO("ReduceScatterMeshDirect run: rank[%u] totalrank[%u] inputMem[%p] outputMem[%p] count[%llu]", rank,
      80              :         rankSize, inputMem_.ptr(), outputMem_.ptr(), count_);
      81              : 
      82              :     // 数据准备
      83            0 :     u32 unitSize = SIZE_TABLE[dataType_];
      84              : 
      85            0 :     if (slices_.size() == 0) {
      86              :         // slices_为空,临时构造等长slices
      87            0 :         slices_.resize(rankSize);
      88            0 :         u64 curSize = count_ * unitSize;
      89            0 :         u64 sliceSize = (opInfo_->count) * unitSize;
      90              : 
      91            0 :         for (u32 i = 0; i < rankSize; i++) {
      92            0 :             slices_[i].size = curSize;
      93            0 :             slices_[i].offset = (i * sliceSize);
      94              :         }
      95              :     }
      96              : 
      97              :     DeviceMem userMemIn =
      98            0 :         DeviceMem::create(static_cast<char *>(opInfo_->inputAddr) + slices_[rank].offset, slices_[rank].size);
      99            0 :     DeviceMem userMemOut = DeviceMem::create(static_cast<char *>(opInfo_->outputAddr), slices_[rank].size);
     100            0 :     DeviceMem commMemOut = DeviceMem::create(outputMem_.ptr(), outputMem_.size());
     101              :     
     102            0 :     DeviceMem src;
     103            0 :     DeviceMem dst;
     104              :     
     105            0 :     dst = commMemOut.range(0, slices_[rank].size);
     106            0 :     CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, userMemIn, stream_));
     107              :     
     108            0 :     CHK_RET(MainRecordSub());
     109            0 :     CHK_RET(SubWaitMain());
     110              : 
     111              :     // 每个stream只负责一个对端的交互
     112            0 :     HCCL_DEBUG("[ReduceScatterMeshDirect][RunAsync]rankSize is %u", rankSize);
     113            0 :     for (u32 round = 1; round < rankSize; round++) {
     114            0 :         u32 dstRank = (round + rank) % rankSize;
     115            0 :         const LINK &dstLink = links[dstRank];
     116            0 :         Stream &subStream = meshStreams_[round - 1];
     117            0 :         CHK_RET(dstLink->TxAck(subStream));
     118            0 :         CHK_RET(dstLink->RxAck(subStream));
     119              :     }
     120            0 :     CHK_RET(SubRecordMain());
     121            0 :     CHK_RET(MainWaitSub());
     122              :     // 为子图增加一个从stream到主stream的附着点
     123            0 :     DeviceMem srcZero = DeviceMem::create(inputMem_.ptr(), 0);
     124            0 :     DeviceMem dstZero = DeviceMem::create(outputMem_.ptr(), 0);
     125            0 :     CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstZero, srcZero, stream_));
     126              : 
     127            0 :     CHK_RET(SubWaitMain());
     128            0 :     CHK_RET(MainRecordSub());
     129              : 
     130              :     // inline执行notice reduce
     131            0 :     for (u32 round = 1; round < rankSize; round++) {
     132            0 :         u32 dstRank = (round + rank) % rankSize;
     133            0 :         const LINK &dstLink = links[dstRank];
     134            0 :         Stream &subStream = meshStreams_[round - 1];
     135              :         // 本rank要发数据
     136            0 :         void *remMemPtr = nullptr;
     137              :         // 获取远端的commoutMem
     138            0 :         CHK_RET(dstLink->GetRemoteMem(UserMemType::INPUT_MEM, &remMemPtr));
     139            0 :         dst = DeviceMem::create(static_cast<char *>(remMemPtr), slices_[dstRank].size);
     140            0 :         src = DeviceMem::create(static_cast<char *>(opInfo_->inputAddr) + slices_[dstRank].offset, slices_[dstRank].size);
     141            0 :         u64 curCount = slices_[dstRank].size / unitSize;
     142            0 :         CHK_RET(HcclReduceAsync(dispatcher_, static_cast<void *>(src.ptr()), curCount, dataType_, reductionOp_,
     143              :             subStream, static_cast<void *>(dst.ptr()), dstLink->GetRemoteRank(), dstLink->GetLinkType(),
     144              :             INLINE_REDUCE_BIT));
     145              : 
     146            0 :         CHK_RET(dstLink->TxDataSignal(subStream));
     147            0 :         CHK_RET(dstLink->RxDataSignal(subStream));
     148              :     }
     149              : 
     150            0 :     CHK_RET(SubRecordMain());
     151            0 :     CHK_RET(MainWaitSub());
     152              : 
     153              :     // commout--> useroutput
     154            0 :     DeviceMem srcMem = commMemOut.range(0, slices_[rank].size);
     155            0 :     CHK_RET(HcclD2DMemcpyAsync(dispatcher_, userMemOut, srcMem, stream_));
     156              : 
     157            0 :     HCCL_INFO("ReduceScatterMeshDirect finished: rank[%u]", rank);
     158            0 :     return HCCL_SUCCESS;
     159            0 : }
     160              : REGISTER_TEMPLATE(TemplateType::TEMPLATE_REDUCESCATTER_MESH_DIRECT, ReduceScatterMeshDirect);
     161              : }
        

Generated by: LCOV version 2.0-1