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 "hccl_aiv_utils.h"
      12              : #include "aiv_ins.h"
      13              : #include "aiv_temp_all_to_all_mesh_1D.h"
      14              : #include "executor_utils.h"
      15              : 
      16              : namespace Hccl {
      17              : 
      18            0 : AivTempAlltoAllMesh1D::AivTempAlltoAllMesh1D(const RankId virtualRank, const u32 tempRankSize,
      19            0 :     const std::vector<std::vector<RankId>> &tempVTopo, const std::map<RankId, u32> &tempVirtRankMap)
      20            0 :     : AivAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
      21              : {
      22            0 : }
      23              : 
      24            0 : AivTempAlltoAllMesh1D::~AivTempAlltoAllMesh1D()
      25              : {
      26            0 : }
      27              : 
      28            0 : u32 AivTempAlltoAllMesh1D::CalcScratchMultiple(BufferType inBuffType, BufferType outBuffType)
      29              : {
      30              :     (void)inBuffType;
      31              :     (void)outBuffType;
      32              :     // 单算子和图模式一致，AlltoAll的usrIn、scratchBuffer，usrOut大小一致
      33            0 :     return 1;
      34              : }
      35              : 
      36            0 : HcclResult AivTempAlltoAllMesh1D::CalNumBlocks(u32& numBlocks, u64 dataSize, u32 numBlocksLimit)
      37              : {
      38              :     (void) dataSize;
      39            0 :     HCCL_INFO("[AivTempAlltoAllMesh1D] Limit core num[%u]", numBlocksLimit);
      40              : 
      41              :     // 小于1的场景
      42            0 :     if (numBlocksLimit < 1) {
      43            0 :         numBlocks = numBlocksLimit;
      44            0 :         return HcclResult::HCCL_SUCCESS;
      45              :     }
      46              : 
      47            0 :     if (numBlocksLimit >= tempRankSize_) {
      48            0 :         numBlocks = numBlocksLimit / tempRankSize_ * tempRankSize_;
      49              :     } else {
      50            0 :         u32 rankPerCore = (tempRankSize_ + numBlocksLimit - 1) / numBlocksLimit;  // 向上取整
      51            0 :         numBlocks = (tempRankSize_ + rankPerCore - 1) / rankPerCore;  // 向上取整
      52              :     }
      53              : 
      54            0 :     HCCL_INFO("[AivTempAlltoAllMesh1D] Actually use core num[%u]", numBlocks);
      55            0 :     return HcclResult::HCCL_SUCCESS;
      56              : }
      57              : 
      58            0 : HcclResult AivTempAlltoAllMesh1D::CalcRes(AlgTempResReq &tempResReq)
      59              : {
      60            0 :     tempResReq.queNum = 1;
      61            0 :     tempResReq.streamNum = tempResReq.queNum;
      62            0 :     HCCL_INFO("[AivTempAlltoAllMesh1D] Calculate resource, stream number is[%u],", tempResReq.streamNum);
      63            0 :     CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
      64            0 :     return HcclResult::HCCL_SUCCESS;
      65              : }
      66              : 
      67            0 : HcclResult AivTempAlltoAllMesh1D::GenExtIns(const TempFuncs &tempFuncs, const TemplateDataParams &templateDataParams,
      68              :     const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
      69              : {
      70            0 :     HCCL_INFO("[AivTempAlltoAllMesh1D] Run algorithm start: rank[%d]", myRank_);
      71            0 :     CHK_PRT_RET(tempInsQues.empty(),
      72              :         HCCL_ERROR("[AivTempAlltoAllMesh1D] empty queue"), HcclResult::HCCL_E_INTERNAL);
      73            0 :     CHK_PTR_NULL(tempInsQues[0]);
      74            0 :     std::vector<LinkData> allLinks;
      75            0 :     for (auto iter = tempLinks.begin(); iter != tempLinks.end(); ++iter) {
      76            0 :         allLinks.emplace_back(iter->second.at(0));
      77              :     }
      78              : 
      79            0 :     IncSliceId();  // 自动增长sliceId，传入aivTag
      80              : 
      81            0 :     AivOpArgs aivAlltoAllArgs;
      82            0 :     aivAlltoAllArgs.cmdType = HcclCMDType::HCCL_CMD_ALLTOALL;
      83            0 :     aivAlltoAllArgs.input = templateDataParams.buffInfo.inBuffBaseOff;
      84            0 :     aivAlltoAllArgs.output = templateDataParams.buffInfo.outBuffBaseOff;
      85            0 :     aivAlltoAllArgs.rank = u32(myRank_);
      86            0 :     aivAlltoAllArgs.rankSize = tempRankSize_;
      87            0 :     aivAlltoAllArgs.count = templateDataParams.sliceSize / DataTypeSizeGet(dataType_);
      88            0 :     aivAlltoAllArgs.dataType = dataType_;
      89            0 :     aivAlltoAllArgs.op = reduceOp_;
      90            0 :     aivAlltoAllArgs.root = root_;
      91            0 :     aivAlltoAllArgs.aivTag = sliceId_;  // 传入aivTag，Lauch时重新组装为aivTag
      92            0 :     aivAlltoAllArgs.isOpBase = (tempFuncs.opMode == OpMode::OPBASE);
      93            0 :     aivAlltoAllArgs.xRankSize = tempVTopo_[0].size();
      94            0 :     aivAlltoAllArgs.yRankSize = 0;
      95            0 :     aivAlltoAllArgs.zRankSize = 0;
      96            0 :     u64 dataSize = op_.dataCount * DataTypeSizeGet(dataType_);
      97            0 :     CHK_RET(CalNumBlocks(aivAlltoAllArgs.numBlocks, dataSize, op_.numBlocksLimit));
      98            0 :     for (u32 i = 0; i < tempVTopo_[0].size(); i ++){
      99            0 :         aivAlltoAllArgs.topo_[i] = tempVTopo_[0][i];
     100              :     }
     101            0 :     if (tempVTopo_.size() > 1){
     102            0 :         aivAlltoAllArgs.yRankSize = tempVTopo_[1].size();
     103            0 :         for (u32 i = 0; i < tempVTopo_[1].size(); i++){
     104            0 :             aivAlltoAllArgs.topo_[TOPO_LEN_Y_OFFSET + i] = tempVTopo_[1][i];
     105              :         }
     106              :     }
     107            0 :     if (tempVTopo_.size() == MAX_DIM_NUM){
     108            0 :         aivAlltoAllArgs.zRankSize = tempVTopo_[MAX_DIM_NUM - 1].size();
     109            0 :         for (u32 i = 0; i < tempVTopo_[MAX_DIM_NUM - 1].size(); i++){
     110            0 :             aivAlltoAllArgs.topo_[TOPO_LEN_Z_OFFSET + i] = tempVTopo_[MAX_DIM_NUM - 1][i];
     111              :         }
     112              :     }
     113              : 
     114            0 :     aivAlltoAllArgs.inputSliceStride = templateDataParams.inputSliceStride;
     115            0 :     aivAlltoAllArgs.outputSliceStride = templateDataParams.outputSliceStride;
     116            0 :     aivAlltoAllArgs.repeatNum = templateDataParams.repeatNum;
     117            0 :     aivAlltoAllArgs.inputRepeatStride = templateDataParams.inputRepeatStride;
     118            0 :     aivAlltoAllArgs.outputRepeatStride = templateDataParams.outputRepeatStride;
     119              : 
     120            0 :     std::unique_ptr<Instruction> aivInsAlltoAllMesh1D = std::make_unique<AivInstruction>(allLinks, aivAlltoAllArgs);
     121              : 
     122            0 :     tempInsQues[0]->Append(std::move(aivInsAlltoAllMesh1D));
     123              : 
     124            0 :     HCCL_INFO("[AivTempAlltoAllMesh1D] Run algorithm end: rank[%d]", myRank_);
     125            0 :     return HcclResult::HCCL_SUCCESS;
     126            0 : }
     127              : 
     128              : }  // namespace Hccl