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 <ios>
12 : #include <iostream>
13 :
14 : #include "log.h"
15 :
16 : #include "ccu_instruction_broadcast_mesh2d_mem2mem.h"
17 : #include "ccu_rank_group.h"
18 : #include "ccu_ctx_creator_registry.h"
19 : #include "ccu_ins_group.h"
20 : #include "ccu_context_broadcast_mesh2d_mem2mem.h"
21 : #include "ccu_temp_broadcast_mesh_2D_mem2mem.h"
22 : #include "template_utils.h"
23 : namespace Hccl {
24 :
25 : static CcuInstRegister<CcuContextBroadcastMeshMem2Mem2D>
26 : g_registrarBroadcastMesh2D_Mem2mem(CcuInstType::CCU_BROADCAST_MESH_2D_MEM2MEM);
27 :
28 0 : CcuTempBroadcastMeshMem2Mem2D::CcuTempBroadcastMeshMem2Mem2D(const RankId virtualRank, const u32 tempRankSize,
29 : const std::vector<std::vector<RankId>> &tempVTopo,
30 0 : const std::map<RankId, u32> &tempVirtRankMap)
31 0 : : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
32 : {
33 0 : if (tempVTopo_.size() != NUM_TWO || tempVTopo_[0].size() <= 1
34 0 : || tempVTopo_[1].size() <= 1) { // concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
35 0 : THROW<InvalidParamsException>(
36 0 : StringFormat("[CcuTempBroadcastMeshMem2Mem2D] Rank[%d], Invalid tempVTopo "
37 : "Size[%u] or Invalid tempVTopo[0] size [%u] or tempVTopo[1] size [%u].",
38 0 : myRank_, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size()));
39 : }
40 0 : dimSize_.emplace_back(tempVTopo[0].size());
41 0 : dimSize_.emplace_back(tempVTopo[1].size());
42 0 : }
43 :
44 0 : CcuTempBroadcastMeshMem2Mem2D::~CcuTempBroadcastMeshMem2Mem2D()
45 : {
46 0 : }
47 :
48 0 : HcclResult CcuTempBroadcastMeshMem2Mem2D::CalcRes(AlgTempResReq &tempResReq)
49 : {
50 : // 按照IODienum来确定stream数量，支持2D和2D的template
51 0 : tempResReq.queNum = 1; // 只申请一个insQue，填充一个insGroup，由框架将其中的ins放在多个stream上
52 0 : tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个 stream 给 ccuInsGroup
53 0 : uint32_t dieNum = tempVTopo_.size();
54 0 : if (dieNum != NUM_TWO) { // concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
55 0 : HCCL_ERROR("[CcuTempBroadcastMeshMem2Mem2D] Rank[%d], Invalid IODieNum[%zu].", myRank_, tempVTopo_.size());
56 0 : return HcclResult::HCCL_E_PARA;
57 : }
58 0 : HCCL_INFO(
59 : "[CcuTempBroadcastMeshMem2Mem2D] Rank[%d] requiredQueNum[%u] VtopoSize[%u], VtopoSize0[%u] VtopoSize1[%u].",
60 : myRank_, tempResReq.queNum, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size());
61 :
62 : uint32_t myAlgRank;
63 0 : for (u32 dim = 0; dim < tempVTopo_.size(); dim++) {
64 0 : CHK_RET(GetAlgRank(myRank_, tempVTopo_[dim], myAlgRank));
65 0 : for (u32 queIdx = 0; queIdx < tempVTopo_[dim].size() - 1; queIdx++) {
66 : // find neighbors -> virtualRank
67 0 : u32 neighborAlgRank = (myAlgRank + 1 + queIdx) % (tempVTopo_[dim].size());
68 0 : RankId neighborRank = tempVTopo_[dim][neighborAlgRank];
69 0 : HCCL_INFO("[CollAlgFactory] [CcuTempBroadcastMeshMem2Mem2D] Rank[%d], Dim[%u], NeighborRank[%d].", myRank_,
70 : dim, neighborRank);
71 : // LinkNum
72 0 : tempResReq.links[neighborRank] = 1;
73 : }
74 : }
75 0 : return HcclResult::HCCL_SUCCESS;
76 : }
77 :
78 0 : HcclResult CcuTempBroadcastMeshMem2Mem2D::PrepareLinks(const ResLinks &tempLinks)
79 : {
80 : // 分别记录两个Die上的link，构造rankGroup
81 0 : for (auto pair : tempLinks) {
82 0 : if (pair.second.size() == 0 || pair.second[0].GetHop() != 1) { // ESL环境上暂只有直连链路
83 0 : THROW<InvalidParamsException>(
84 0 : StringFormat("[CcuTempBroadcastMeshMem2Mem2D] Rank[%d]--Peer[%d], InvalidHop[%u].", myRank_, pair.first,
85 0 : pair.second[0].GetHop()));
86 : }
87 0 : if ((pair.first / dimSize_[0] == myRank_ / dimSize_[0]) && pair.second[0].GetHop() == 1) {
88 0 : HCCL_INFO("[CcuTempBroadcastMeshMem2Mem2D][GenExtIns] Rank[%d] insert link to Rank[%d] in linksX", myRank_,
89 : pair.first);
90 0 : linksX_.emplace_back(pair.second[0]);
91 0 : } else if ((pair.first % dimSize_[0] == myRank_ % dimSize_[0]) && pair.second[0].GetHop() == 1) {
92 0 : HCCL_INFO("[CcuTempBroadcastMeshMem2Mem2D][GenExtIns] Rank[%d] insert link to Rank[%d] in linksY", myRank_,
93 : pair.first);
94 0 : linksY_.emplace_back(pair.second[0]);
95 : } else {
96 0 : THROW<InvalidParamsException>(
97 0 : StringFormat("[CcuTempBroadcastMeshMem2Mem2D] Rank[%d], Unexpected peerRank[%d] in tempLinks.", myRank_,
98 0 : pair.first));
99 : }
100 0 : }
101 0 : return HcclResult::HCCL_SUCCESS;
102 : }
103 :
104 0 : HcclResult CcuTempBroadcastMeshMem2Mem2D::GenExtIns(const TempFuncs &tempFuncs,
105 : const TemplateDataParams &templateDataParams,
106 : const ResLinks &tempLinks, std::vector<InsQuePtr> &tempInsQues)
107 : {
108 0 : opMode_ = tempFuncs.opMode;
109 0 : buffInfo_ = templateDataParams.buffInfo;
110 :
111 0 : CHK_RET(PrepareLinks(tempLinks));
112 :
113 0 : RankGroup rankGroupX;
114 0 : RankGroup rankGroupY;
115 0 : AddRanksToGroup(tempVTopo_,rankGroupX,rankGroupY);
116 :
117 0 : uint64_t inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
118 :
119 0 : uint64_t sliceSize = templateDataParams.sliceSize; // 获取本rank需要处理的数据量
120 : uint64_t token;
121 0 : CHK_RET(GetToken(op_, token));
122 :
123 0 : std::vector<uint32_t> dimId;
124 0 : dimId.emplace_back(myRank_ % dimSize_[0]);
125 0 : dimId.emplace_back(myRank_ / dimSize_[0]);
126 0 : std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
127 0 : for (uint32_t axisId = 0; axisId < NUM_TWO; axisId++) { // 2D算法，需要执行两次
128 : // 计算每次编译的偏移量和数据量
129 0 : uint64_t xAxisSize = sliceSize * dimSize_[0] / (dimSize_[axisId] + dimSize_[1 - axisId]);
130 0 : uint64_t yAxisSize = sliceSize - xAxisSize;
131 0 : HCCL_INFO("[CcuTempBroadcastMeshMem2Mem2D] [GenExtIns]: dimSize0[%llu], dimSize1[%llu], myRank_[%d],"
132 : "inputAddr[%llu], sliceSize[%llu], xAxisSize[%llu], yAxisSize[%llu],axisId[%u]",
133 : dimSize_[0], dimSize_[1], myRank_, inputAddr, sliceSize, xAxisSize, yAxisSize, axisId);
134 0 : CcuInstructionBroadcastMeshMem2Mem2D ccuInsBroadcastMeshMem2Mem2D;
135 0 : ccuInsBroadcastMeshMem2Mem2D.Init(dimSize_, static_cast<uint32_t>(myRank_), inputAddr, axisId, sliceSize,
136 0 : xAxisSize, yAxisSize, token, op_, tempVTopo_);
137 0 : ccuInsBroadcastMeshMem2Mem2D.SetLinks(axisId == 0 ? linksX_ : linksY_);
138 0 : ccuInsBroadcastMeshMem2Mem2D.SetRankGroup(axisId == 0 ? rankGroupX : rankGroupY);
139 0 : u32 ckeNum = 5;
140 0 : ccuInsBroadcastMeshMem2Mem2D.SetCntCkeNum(ckeNum); // 每个transport用5个CKE
141 0 : ccuInsBroadcastMeshMem2Mem2D.Describe();
142 0 : insGroupPtr->Append(
143 0 : std::move(std::make_unique<CcuInstructionBroadcastMeshMem2Mem2D>(ccuInsBroadcastMeshMem2Mem2D)));
144 0 : }
145 0 : tempInsQues[0]->Append(std::move(insGroupPtr)); // 只有一条流
146 0 : return HcclResult::HCCL_SUCCESS;
147 0 : }
148 : } // namespace Hccl
|
