LCOV - code coverage report
Current view: top level - legacy/ascend910/algorithm/base/alg_template/temp_broadcast - broadcast_nb_binary.cc (source / functions) Coverage Total Hit
Test: coverage.info Lines: 0.0 % 108 0
Test Date: 2026-08-04 10:52:23 Functions: 0.0 % 8 0

            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 "broadcast_nb_binary.h"
      12              : #include "device_capacity.h"
      13              : #include "alg_template_register.h"
      14              : 
      15              : namespace hccl {
      16              : constexpr float LATENCY = 60; // 静态时延 60 us;
      17              : 
      18            0 : BroadcastNBBinary::BroadcastNBBinary(const HcclDispatcher dispatcher)
      19            0 :     : NBBase(dispatcher)
      20              : {
      21            0 : }
      22              : 
      23            0 : BroadcastNBBinary::~BroadcastNBBinary()
      24              : {
      25            0 : }
      26              : 
      27            0 : HcclResult BroadcastNBBinary::RunAsync(const u32 rank, const u32 rankSize,
      28              :     const std::vector<std::shared_ptr<Transport> > &links)
      29              : {
      30            0 :     CHK_SMART_PTR_NULL(dispatcher_);
      31            0 :     CHK_PTR_NULL(stream_.ptr());
      32            0 :     HCCL_INFO("BroadcastNBBinary run: rank[%u] totalrank[%u] count[%llu]", rank, rankSize, count_);
      33              : 
      34            0 :     if (rankSize == 1) {
      35            0 :         return HCCL_SUCCESS;
      36              :     }
      37              : 
      38            0 :     CHK_PRT_RET(links.size() < rankSize,
      39              :         HCCL_ERROR("[BroadcastNBBinary][RunAsync]rank[%u] linksize[%llu] is less than rank size", rank, links.size()),
      40              :         HCCL_E_INTERNAL);
      41              : 
      42            0 :     u32 unitSize = DataUnitSize(dataType_);
      43            0 :     CHK_PRT_RET(unitSize == 0, HCCL_ERROR("[BroadcastNBBinary][RunAsync]rank[%u] unit data size is zero", rank),
      44              :         HCCL_E_INTERNAL);
      45              : 
      46            0 :     HcclResult ret = HCCL_SUCCESS;
      47            0 :     if (inputMem_ != outputMem_) {
      48            0 :         ret = HcclD2DMemcpyAsync(dispatcher_, outputMem_, inputMem_, stream_);
      49            0 :         CHK_PRT_RET(ret != HCCL_SUCCESS,
      50              :             HCCL_ERROR("[Run][BroadcastOnRootRank]root rank[%u] memcpy async from input[%p] "\
      51              :             "failed to output[%p]", rank, inputMem_.ptr(), outputMem_.ptr()), ret);
      52              :     }
      53              : 
      54            0 :     CHK_RET(RunBroadcastNBBinary(rank, rankSize, links));
      55            0 :     HCCL_INFO("BroadcastNBBinary finished: rank[%u] end count[%llu]", rank, count_);
      56            0 :     return HCCL_SUCCESS;
      57              : }
      58              : 
      59            0 : HcclResult BroadcastNBBinary::RunBroadcastNBBinary(const u32 rank, const u32 rankSize,
      60              :     const std::vector<std::shared_ptr<Transport> > &links)
      61              : {
      62            0 :     HcclResult ret = HCCL_SUCCESS;
      63            0 :     if (rank == root_) {
      64            0 :         hasData_ = true;
      65            0 :         CHK_PRT_RET(!inputMem_, HCCL_ERROR("[BroadcastNBBinary][RunAsync]rank[%u] inputmem is null", rank), HCCL_E_PTR);
      66              :     } else {
      67            0 :         CHK_PRT_RET(!outputMem_, HCCL_ERROR("[BroadcastNBBinary][RunAsync]rank[%u] outputmem is null", rank),
      68              :             HCCL_E_PTR);
      69              :     }
      70              : 
      71            0 :     HCCL_DEBUG("root[%u], hasData[%u]", root_, hasData_);
      72              :  
      73            0 :     u64 dataBytes = count_ * DataUnitSize(dataType_); // 总数据量
      74            0 :     u32 nSteps = CalcCeilLog2(rankSize); // 通信步数
      75            0 :     u32 deltaRoot = (rank + rankSize - root_) % rankSize; // 与Root节点的距离
      76              :  
      77            0 :     for (u32 step = 0; step < nSteps; step++) {
      78            0 :         if (deltaRoot < u32(1 << step)) { // 该节点需要发送数据
      79            0 :             if ((step != nSteps - 1 || deltaRoot < (rankSize - (1 << step))) && hasData_) {
      80            0 :                 u32 deltaRank = 1 << step;
      81            0 :                 u32 sendTo = (rank + deltaRank) % rankSize;
      82            0 :                 LINK linkRight = links[sendTo];
      83            0 :                 CHK_SMART_PTR_NULL(linkRight);
      84              :  
      85            0 :                 std::vector<Slice> txSlices;
      86            0 :                 txSlices.resize(1);
      87            0 :                 txSlices[0].offset = baseOffset_;
      88            0 :                 txSlices[0].size = dataBytes;
      89              :  
      90            0 :                 CHK_RET(linkRight->RxAck(stream_));
      91            0 :                 ret = Tx(linkRight, txSlices);
      92            0 :                 CHK_PRT_RET(ret != HCCL_SUCCESS,
      93              :                     HCCL_ERROR("[Run][Broadcast]rank[%u] step[%u] Right Link tx slices Failed", rank, step), ret);
      94              :  
      95            0 :                 ret = linkRight->TxWaitDone(stream_);
      96            0 :                 CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Run][Broadcast]TxWaitDone failed"), ret);
      97            0 :                 ret = linkRight->GetLinkType() == LinkType::LINK_HCCS 
      98            0 :                     ? linkRight->WaitFin(stream_) : linkRight->WaitFinAck(stream_); // P2P和Roce场景都需要同步
      99            0 :                 CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Run][Broadcast]WaitFinAck failed"), ret);
     100            0 :             }
     101            0 :         } else if (deltaRoot < u32(1 << (step + 1)) && rank != root_) { // 该节点需要接收数据
     102            0 :             u32 deltaRank = 1 << step;
     103            0 :             u32 recvFrom = (rank + rankSize - deltaRank) % rankSize;
     104            0 :             LINK linkLeft = links[recvFrom];
     105            0 :             CHK_SMART_PTR_NULL(linkLeft);
     106              :  
     107            0 :             std::vector<Slice> rxSlices;
     108            0 :             rxSlices.resize(1);
     109            0 :             rxSlices[0].offset = baseOffset_;
     110            0 :             rxSlices[0].size = dataBytes;
     111              :  
     112            0 :             CHK_RET(linkLeft->TxAck(stream_));
     113            0 :             ret = Rx(linkLeft, rxSlices);
     114            0 :             CHK_PRT_RET(ret != HCCL_SUCCESS,
     115              :                 HCCL_ERROR("[Run][Broadcast]rank[%u] step[%u] Right Link rx slices Failed", rank, step), ret);
     116              :  
     117            0 :             ret = linkLeft->RxWaitDone(stream_);
     118            0 :             CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Run][Broadcast]RxWaitDone failed"), ret);
     119            0 :             ret = linkLeft->GetLinkType() == LinkType::LINK_HCCS 
     120            0 :                 ? linkLeft->PostFin(stream_) : linkLeft->PostFinAck(stream_); // P2P和Roce场景都需要同步
     121            0 :             CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Run][Broadcast]PostFinAck failed"), ret);
     122            0 :             hasData_ = true;
     123            0 :         }
     124              :     }
     125            0 :     return HCCL_SUCCESS;
     126              : }
     127              :  
     128            0 : HcclResult BroadcastNBBinary::Tx(const LINK &link, const std::vector<Slice> &txSlices)
     129              : {
     130            0 :     std::vector<TxMemoryInfo> txMems;
     131            0 :     for (const Slice& txSlice : txSlices) {
     132            0 :         DeviceMem srcMem = outputMem_.range(txSlice.offset, txSlice.size);
     133            0 :         HCCL_DEBUG("tx srcMem[%p] range[%llu] size[%llu] ", srcMem.ptr(), txSlice.offset, txSlice.size);
     134            0 :         txMems.emplace_back(
     135            0 :             TxMemoryInfo { UserMemType::OUTPUT_MEM, txSlice.offset, srcMem.ptr(), txSlice.size });
     136            0 :     }
     137              :  
     138            0 :     CHK_RET(link->TxAsync(txMems, stream_));
     139            0 :     return HCCL_SUCCESS;
     140            0 : }
     141              :  
     142            0 : HcclResult BroadcastNBBinary::Rx(const LINK &link, const std::vector<Slice> &rxSlices)
     143              : {
     144            0 :     std::vector<RxMemoryInfo> rxMems;
     145            0 :     for (const Slice& rxSlice : rxSlices) {
     146            0 :         DeviceMem dstMem = outputMem_.range(rxSlice.offset, rxSlice.size);
     147            0 :         HCCL_DEBUG("rx dstMem[%p] range[%llu], size[%llu] ",  dstMem.ptr(), rxSlice.offset, rxSlice.size);
     148            0 :         rxMems.emplace_back(
     149            0 :             RxMemoryInfo { UserMemType::OUTPUT_MEM, rxSlice.offset, dstMem.ptr(), rxSlice.size });
     150            0 :     }
     151              :  
     152            0 :     CHK_RET(link->RxAsync(rxMems, stream_));
     153            0 :     return HCCL_SUCCESS;
     154            0 : }
     155              :  
     156            0 : bool ShouldUseBinaryBroadcastOfNB(const u64 dataSize, const u32 rankSize, const u32 userRankSize,
     157              :     const float deviceNumPerAggregation)
     158              : {
     159              :     // 小数据量和rank size为2时使用Binary broadcast
     160            0 :     HCCL_INFO("datasize[%llu], ranksize[%u], userRankSize[%u], deviceNumPerAggregation[%f]", dataSize, rankSize,
     161              :         userRankSize, deviceNumPerAggregation);
     162              : 
     163            0 :     constexpr u32 TWO_RANK_SIZE = 2;
     164            0 :     if (rankSize == TWO_RANK_SIZE) {
     165            0 :         return true;
     166              :     }
     167              :  
     168              :     // 通信步数为log_2(rankSize)向上取整
     169            0 :     u32 nSteps = 0;
     170            0 :     for (u32 tmp = rankSize - 1; tmp != 0; tmp >>= 1, ++nSteps) {}
     171              :  
     172              :     float bandWidth; // 网卡出口带宽,用于计算大小包切片策略的阈值
     173            0 :     CHK_RET(GetBandWidthPerNPU(1, userRankSize, static_cast<u32>(deviceNumPerAggregation), bandWidth)); // 单位:GB/s
     174              :     
     175              :     // (公式解释):
     176              :     // bandwidth_ - 网卡出口带宽,单位GB/s
     177              :     // LATENCY - 链路端到端静态时延
     178              :     // * 1000 - 转换为ns
     179            0 :     const float dataSizeBaseNum = bandWidth * LATENCY * 1000;
     180            0 :     constexpr u32 rankSizeOfSmallScale = 4; // 4以下节点数为小规模
     181            0 :     constexpr u32 dataSizeMultiple = 2;     // 通信数据量倍数为2
     182              :  
     183              :     // 计算阈值,用于判断大小包
     184              :     float thresholdDataSize;
     185            0 :     if (rankSize <= rankSizeOfSmallScale) {
     186            0 :         thresholdDataSize = rankSize * dataSizeBaseNum;
     187              :     } else {
     188              :         // 公式:用于计算大包处理的阈值
     189            0 :         thresholdDataSize = (nSteps + nSteps - dataSizeMultiple) / (nSteps - dataSizeMultiple) * dataSizeBaseNum;
     190              :     }
     191              :  
     192            0 :     return dataSize < thresholdDataSize; // 小数据量使用Binary broadcast
     193              : }
     194              :  
     195              : REGISTER_TEMPLATE(TemplateType::TEMPLATE_BROADCAST_NB_BINARY, BroadcastNBBinary);
     196              : }  // namespace hccl
        

Generated by: LCOV version 2.0-1