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_scatter_mesh2d.h"
17 : #include "ccu_rank_group.h"
18 : #include "ccu_ctx_creator_registry.h"
19 : #include "ccu_ins_group.h"
20 : #include "ccu_context_reduce_scatter_mesh2d.h"
21 : #include "ccu_temp_reduce_scatter_mesh_2D.h"
22 :
23 : namespace Hccl {
24 :
25 : static CcuInstRegister<CcuContextReduceScatterMesh2D> g_registrarReduceScatterMesh2D(
26 : CcuInstType::CCU_REDUCE_SCATTER_MESH_2D_DIRECT);
27 :
28 0 : CcuTempReduceScatterMesh2D::CcuTempReduceScatterMesh2D(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("[CcuTempReducescatterMesh2D] 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 : CcuTempReduceScatterMesh2D::~CcuTempReduceScatterMesh2D()
44 : {
45 0 : }
46 :
47 0 : void CcuTempReduceScatterMesh2D::InitReduceInfo(const ReduceOp &reduceOp, const DataType &dataType) {
48 0 : reduceOp_ = reduceOp;
49 0 : dataType_ = dataType;
50 0 : }
51 :
52 0 : HcclResult CcuTempReduceScatterMesh2D::CalcSliceInfo(const AllignInfo &allignInfo, const u64 dataSize,
53 : RankSliceInfo &sliceInfoVec)
54 : {
55 0 : std::vector<SliceInfo> tmp(tempVTopo_.size());
56 0 : sliceInfoVec.resize(tempRankSize_, tmp);
57 0 : CHK_RET(CalcRsAgSliceInfoMesh(myRank_, tempRankSize_, allignInfo, dataSize, sliceInfoVec));
58 0 : return HcclResult::HCCL_SUCCESS;
59 0 : }
60 :
61 0 : HcclResult CcuTempReduceScatterMesh2D::CalcRes(AlgTempResReq &tempResReq)
62 : {
63 : // 按照IODienum来确定stream数量，支持2D和2D的template
64 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
65 0 : tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个stream给ccuInsGroup
66 0 : uint32_t dieNum = tempVTopo_.size();
67 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
68 0 : HCCL_ERROR("[CcuTempReducescatterMesh2D] Rank[%d], Invalid IODieNum[%zu].", myRank_, tempVTopo_.size());
69 0 : return HcclResult::HCCL_E_PARA;
70 : }
71 0 : HCCL_INFO("[CcuTempReducescatterMesh2D] Rank[%d] requiredQueNum[%u] VtopoSize[%u], VtopoSize0[%u] VtopoSize1[%u].",
72 : myRank_, tempResReq.queNum, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size());
73 :
74 : uint32_t myAlgRank;
75 0 : for (u32 dim = 0; dim < tempVTopo_.size(); dim++) {
76 0 : CHK_RET(GetAlgRank(myRank_, tempVTopo_[dim], myAlgRank));
77 0 : for (u32 queIdx = 0; queIdx < tempVTopo_[dim].size() - 1; queIdx++) {
78 : // find neighbors -> virtualRank
79 0 : u32 neighborAlgRank = (myAlgRank + 1 + queIdx) % (tempVTopo_[dim].size());
80 0 : RankId neighborRank = tempVTopo_[dim][neighborAlgRank];
81 0 : HCCL_INFO("[CollAlgFactory] [CcuTempReducescatterMesh2D] Rank[%d], Dim[%u], NeighborRank[%d].", myRank_,
82 : dim, neighborRank);
83 :
84 : // LinkNum
85 0 : tempResReq.links[neighborRank] = 1;
86 : }
87 : }
88 0 : return HcclResult::HCCL_SUCCESS;
89 : }
90 :
91 0 : HcclResult CcuTempReduceScatterMesh2D::Run(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
92 : const BuffInfo &buffInfo, const ResLinks &tempLinks,
93 : std::vector<InsQuePtr> &tempInsQues)
94 : {
95 0 : opMode_ = tempFuncs.opMode;
96 0 : buffInfo_ = buffInfo;
97 :
98 : // 分别记录两个Die上的link，构造rankGroup
99 0 : for (auto &pair : tempLinks) {
100 0 : if (pair.second.size() == 0 || pair.second[0].GetHop() != 1) { // ESL环境上暂只有直连链路
101 0 : THROW<InvalidParamsException>(StringFormat("[CcuTempAlltoAllMesh2D] Rank[%d]--Peer[%d], InvalidHop[%u].",
102 0 : myRank_, pair.first, pair.second[0].GetHop()));
103 : }
104 0 : if ((pair.first / dimSize_[0] == myRank_ / dimSize_[0]) && pair.second[0].GetHop() == 1) {
105 0 : HCCL_INFO("[CcuTempAlltoAllMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksX", myRank_, pair.first);
106 0 : linksX_.emplace_back(pair.second[0]);
107 0 : } else if ((pair.first % dimSize_[0] == myRank_ % dimSize_[0]) && pair.second[0].GetHop() == 1) {
108 0 : HCCL_INFO("[CcuTempAlltoAllMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksY", myRank_, pair.first);
109 0 : linksY_.emplace_back(pair.second[0]);
110 : } else {
111 0 : THROW<InvalidParamsException>(StringFormat(
112 0 : "[CcuTempAlltoAllMesh2D] Rank[%d], Unexpected peerRank[%d] in tempLinks.", myRank_, pair.first));
113 : }
114 : }
115 :
116 0 : RankGroup rankGroupX;
117 0 : RankGroup rankGroupY;
118 0 : AddRanksToGroup(tempVTopo_,rankGroupX,rankGroupY);
119 :
120 : uint64_t inputAddr;
121 : uint64_t outputAddr;
122 : uint64_t offset;
123 0 : uint64_t outputSize = static_cast<uint64_t>(op_.outputMem->GetSize());
124 0 : if (opMode_ == OpMode::OPBASE) {
125 0 : if (tempFuncs.isForepart) {
126 0 : inputAddr = BufferTypeToAddr(tempFuncs.usrData.usrInSlices[myRank_].GetType());
127 : // 当前loop的size大小
128 0 : offset = tempFuncs.usrData.usrOutSlices[0].GetOffset();
129 : } else {
130 0 : inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
131 : // 从inBuff获取数据，只需要加上rank偏移
132 0 : offset = sliceInfoVec[myRank_][0].offset;
133 : }
134 0 : if (tempFuncs.isBottom) {
135 0 : outputAddr = BufferTypeToAddr(tempFuncs.usrData.usrOutSlices[0].GetType())
136 0 : + tempFuncs.usrData.usrOutSlices[0].GetOffset();
137 : } else {
138 0 : outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
139 : }
140 : } else {
141 0 : offset = tempFuncs.usrData.usrOutSlices[0].GetOffset();
142 0 : inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
143 0 : outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff + tempFuncs.usrData.usrOutSlices[0].GetOffset();
144 : }
145 0 : uint64_t sliceSize = sliceInfoVec[myRank_][0].size; // 获取本rank需要处理的数据量
146 : uint64_t token;
147 0 : CHK_RET(GetToken(op_, token));
148 :
149 0 : std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
150 :
151 0 : for (uint32_t axisId = 0; axisId < 2; axisId++) { // 2D算法，需要执行两次
152 : // 计算每次编译的偏移量和数据量
153 0 : uint64_t sliceCount = sliceSize / DataTypeSizeGet(dataType_);
154 0 : uint64_t xAxisSize = (sliceCount * dimSize_[0] / (dimSize_[axisId] + dimSize_[1 - axisId])) * DataTypeSizeGet(dataType_);
155 0 : uint64_t yAxisSize = sliceSize - xAxisSize;
156 :
157 0 : CcuInstructionReduceScatterMesh2D ccuInsReduceScatterMesh2D;
158 0 : ccuInsReduceScatterMesh2D.Init(dimSize_, static_cast<uint32_t>(myRank_), inputAddr, outputAddr, axisId, outputSize,
159 0 : xAxisSize, yAxisSize, offset, token, op_, tempVTopo_);
160 :
161 0 : HCCL_INFO("[CcuTempReduceScatterMesh2D] Init: dimSize0[%llu], dimSize1[%llu], myRank_[%d], inputAddr[%llu],"\
162 : "outputAddr[%llu], sliceSize[%llu], xAxisSize[%llu], yAxisSize[%llu], offset[%llu]",
163 : dimSize_[0], dimSize_[1], myRank_, inputAddr, outputAddr, outputSize, xAxisSize, yAxisSize, offset);
164 :
165 0 : ccuInsReduceScatterMesh2D.SetLinks(axisId == 0 ? linksX_ : linksY_);
166 0 : ccuInsReduceScatterMesh2D.SetRankGroup(axisId == 0 ? rankGroupX : rankGroupY);
167 0 : u32 ckeNum = 5;
168 0 : ccuInsReduceScatterMesh2D.SetCntCkeNum(ckeNum); // 每个transport用4个CKE
169 0 : ccuInsReduceScatterMesh2D.Describe();
170 0 : insGroupPtr->Append(std::move(std::make_unique<CcuInstructionReduceScatterMesh2D>(ccuInsReduceScatterMesh2D)));
171 0 : }
172 :
173 0 : tempInsQues[0]->Append(std::move(insGroupPtr)); // 只有1条流
174 :
175 0 : return HcclResult::HCCL_SUCCESS;
176 0 : }
177 : } // namespace Hccl
|
