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_reduce_mesh2d.h"
17 : #include "ccu_rank_group.h"
18 : #include "ccu_ctx_creator_registry.h"
19 : #include "ccu_context_reduce_mesh2d.h"
20 : #include "ccu_temp_reduce_mesh_2D.h"
21 : #include "ccu_ins_group.h"
22 :
23 : namespace Hccl {
24 :
25 : static CcuInstRegister<CcuContextReduceMesh2D> registrarReduce2D(
26 : CcuInstType::CCU_REDUCE_MESH_2D_DIRECT);
27 :
28 0 : CcuTempReduceMesh2D::CcuTempReduceMesh2D(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() != 2 || tempVTopo_[0].size() <= 1 || 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
34 0 : THROW<InvalidParamsException>(StringFormat("[CcuTempReduceMesh2D] Rank[%d], Invalid tempVTopo "
35 : "Size[%u] or Invalid tempVTopo[0] size [%u] or tempVTopo[1] size [%u].",
36 0 : myRank_, tempVTopo_.size(), tempVTopo_[0].size(),
37 0 : tempVTopo_[1].size()));
38 : }
39 0 : dimSize_.emplace_back(tempVTopo[0].size());
40 0 : dimSize_.emplace_back(tempVTopo[1].size());
41 0 : }
42 :
43 0 : CcuTempReduceMesh2D::~CcuTempReduceMesh2D()
44 : {
45 0 : }
46 :
47 0 : void CcuTempReduceMesh2D::InitReduceInfo(const ReduceOp &reduceOp, const DataType &dataType)
48 : {
49 0 : reduceOp_ = reduceOp;
50 0 : dataType_ = dataType;
51 0 : }
52 :
53 0 : HcclResult CcuTempReduceMesh2D::CalcRes(AlgTempResReq &tempResReq)
54 : {
55 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
56 0 : tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个 stream 给 ccuInsGroup
57 0 : HCCL_INFO("[CalcRes] tempResReq.queNum[%u]", tempResReq.queNum);
58 0 : uint32_t dieNum = tempVTopo_.size();
59 0 : if (dieNum != 2) { // concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
60 0 : HCCL_ERROR("[CcuTempReduceMesh2D] Rank[%d], Invalid IODieNum[%zu].", myRank_, tempVTopo_.size());
61 0 : return HcclResult::HCCL_E_PARA;
62 : }
63 0 : HCCL_INFO("[CcuTempReduceMesh2D] Rank[%d] requiredQueNum[%u] VtopoSize[%u], VtopoSize0[%u] VtopoSize1[%u].",
64 : myRank_, tempResReq.queNum, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size());
65 :
66 : uint32_t myAlgRank;
67 0 : for (u32 dim = 0; dim < tempVTopo_.size(); dim++) {
68 0 : CHK_RET(GetAlgRank(myRank_, tempVTopo_[dim], myAlgRank));
69 0 : for (u32 queIdx = 0; queIdx < tempVTopo_[dim].size() - 1; queIdx++) {
70 : // find neighbors -> virtualRank
71 0 : u32 neighborAlgRank = (myAlgRank + 1 + queIdx) % (tempVTopo_[dim].size());
72 0 : RankId neighborRank = tempVTopo_[dim][neighborAlgRank];
73 0 : HCCL_INFO("[CollAlgFactory] [CcuTempReduceMesh2D] Rank[%d], Dim[%u], NeighborRank[%d].", myRank_,
74 : dim, neighborRank);
75 : // LinkNum
76 0 : tempResReq.links[neighborRank] = 1;
77 : }
78 : }
79 0 : return HcclResult::HCCL_SUCCESS;
80 : }
81 :
82 0 : HcclResult CcuTempReduceMesh2D::CalcSliceInfo(const AllignInfo &allignInfo, const u64 dataSize,
83 : RankSliceInfo &sliceInfoVec)
84 : {
85 : (void)allignInfo;
86 : SliceInfo basicSlice;
87 0 : basicSlice.offset = 0;
88 0 : basicSlice.size = dataSize;
89 0 : HCCL_INFO("[CcuTempReduceMesh2D] [CalcSliceInfo]basicSlice.size[%u]", basicSlice.size);
90 0 : std::vector<SliceInfo> singleRankSliceInfoVector{basicSlice};
91 0 : sliceInfoVec.resize(tempRankSize_, singleRankSliceInfoVector);
92 :
93 0 : return HcclResult::HCCL_SUCCESS;
94 0 : }
95 :
96 0 : HcclResult CcuTempReduceMesh2D::Run(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
97 : const BuffInfo &buffInfo, const ResLinks &tempLinks,
98 : std::vector<InsQuePtr> &tempInsQues)
99 : {
100 0 : opMode_ = tempFuncs.opMode;
101 0 : buffInfo_ = buffInfo;
102 0 : rootId_ = op_.root;
103 : // 分别记录两个Die上的link，构造rankGroup
104 0 : for (auto pair : tempLinks) {
105 0 : if (pair.second.size() == 0 || pair.second[0].GetHop() != 1) { // ESL环境上暂只有直连链路
106 0 : THROW<InvalidParamsException>(StringFormat("[CcuTempReduceMesh2D] Rank[%d]--Peer[%d], InvalidHop[%u].",
107 0 : myRank_, pair.first, pair.second[0].GetHop()));
108 : }
109 0 : if ((pair.first / dimSize_[0] == myRank_ / dimSize_[0]) && pair.second[0].GetHop() == 1) {
110 0 : HCCL_INFO("[CcuTempReduceMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksX", myRank_, pair.first);
111 0 : linksX_.emplace_back(pair.second[0]);
112 0 : } else if ((pair.first % dimSize_[0] == myRank_ % dimSize_[0]) && pair.second[0].GetHop() == 1) {
113 0 : HCCL_INFO("[CcuTempReduceMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksY", myRank_, pair.first);
114 0 : linksY_.emplace_back(pair.second[0]);
115 : } else {
116 0 : THROW<InvalidParamsException>(StringFormat(
117 0 : "[CcuTempReduceMesh2D] Rank[%d], Unexpected peerRank[%d] in tempLinks.", myRank_, pair.first));
118 : }
119 0 : }
120 :
121 0 : RankGroup rankGroupX;
122 0 : RankGroup rankGroupY;
123 0 : AddRanksToGroup(tempVTopo_,rankGroupX,rankGroupY);
124 :
125 0 : std::vector<uint64_t> dimSize;
126 0 : dimSize.push_back(tempRankSize_); // tempRankSize_ 就是rank数量
127 :
128 : // 只传userIn的起始位置，不带偏移，偏移已在offset中包含
129 : uint64_t inputAddr;
130 : // userOut 的位置，需要带上偏移
131 : uint64_t outputAddr;
132 0 : if (opMode_ == OpMode::OPBASE) {
133 0 : if (tempFuncs.isForepart) {
134 0 : inputAddr = BufferTypeToAddr(tempFuncs.usrData.usrInSlices[0].GetType())
135 0 : + tempFuncs.usrData.usrInSlices[0].GetOffset();
136 : } else {
137 0 : inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
138 : }
139 0 : if (tempFuncs.isBottom) {
140 0 : outputAddr = BufferTypeToAddr(tempFuncs.usrData.usrOutSlices[0].GetType())
141 0 : + tempFuncs.usrData.usrOutSlices[0].GetOffset();
142 : } else {
143 0 : outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
144 : }
145 : } else {
146 0 : inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
147 0 : outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
148 : }
149 0 : uint64_t sliceSize = sliceInfoVec[myRank_][0].size;
150 : // 自己需要 reduce 的数据基于userIn的起始位置的偏移
151 0 : uint64_t offset = 0;
152 : uint64_t token;
153 0 : CHK_RET(GetToken(op_, token));
154 0 : std::vector<uint32_t> dimId;
155 0 : dimId.emplace_back(myRank_ % dimSize_[0]);
156 0 : dimId.emplace_back(myRank_ / dimSize_[0]); // dimId里面放本rank在拓扑图上的位置
157 0 : std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
158 :
159 0 : for (uint32_t axisId = 0; axisId < 2; axisId++) { // 2D算法，需要执行两次
160 : // 计算每次编译的偏移量和数据量
161 0 : uint64_t xAxisSize = sliceSize;
162 0 : uint64_t yAxisSize = sliceSize;
163 0 : CcuInstructionReduceMesh2D ccuInsReduceMesh2D;
164 0 : ccuInsReduceMesh2D.Init(dimSize_, static_cast<uint32_t>(myRank_), rootId_, axisId, inputAddr, outputAddr, sliceSize,
165 0 : xAxisSize, yAxisSize, offset, token, op_, tempVTopo_);
166 :
167 0 : HCCL_INFO("[CcuTempReduceMesh2D] Run Init: myRank_[%d], dimSize[%llu], inputAddr[%llu],"\
168 : "outputAddr[%llu], sliceSize[%llu], xAxisSize[%llu], yAxisSize[%llu], offset[%llu], axisId_[%u]",
169 : myRank_, dimSize[0], inputAddr, outputAddr, sliceSize, xAxisSize, yAxisSize, offset, axisId);
170 :
171 0 : ccuInsReduceMesh2D.SetLinks(axisId == 0 ? linksX_ : linksY_);
172 0 : ccuInsReduceMesh2D.SetRankGroup(axisId == 0 ? rankGroupX : rankGroupY);
173 0 : u32 ckeNum = 5;
174 0 : ccuInsReduceMesh2D.SetCntCkeNum(ckeNum); // 每个transport用5个CKE
175 0 : ccuInsReduceMesh2D.Describe();
176 0 : insGroupPtr->Append(std::move(std::make_unique<CcuInstructionReduceMesh2D>(ccuInsReduceMesh2D)));
177 0 : }
178 :
179 0 : tempInsQues[0]->Append(std::move(insGroupPtr)); // 只有一条流
180 :
181 0 : return HcclResult::HCCL_SUCCESS;
182 0 : }
183 : } // namespace Hccl
|
