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 "coll_broadcast_ring_executor.h"
12 :
13 : namespace hccl {
14 :
15 0 : CollBroadcastRingExecutor::CollBroadcastRingExecutor(
16 0 : const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher>& topoMatcher)
17 0 : : CollBroadcastExecutor(dispatcher, topoMatcher)
18 : {
19 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE
20 0 : && topoAttr_.deviceType == DevType::DEV_TYPE_910_93) {
21 0 : DMAReduceFlag_ = true;
22 : } else {
23 0 : DMAReduceFlag_ = false;
24 : }
25 0 : }
26 :
27 0 : HcclResult CollBroadcastRingExecutor::CalcStreamNum(u32& streamNum)
28 : {
29 0 : u32 totalStreamNum
30 0 : = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
31 :
32 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93) {
33 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
34 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_NPRING_SINGLE * STREAM_NUM_FOR_DMAREDUCE_ONE_RING;
35 : } else {
36 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
37 : }
38 : }
39 :
40 0 : streamNum = totalStreamNum - 1;
41 0 : HCCL_INFO("[CollBroadcastRingExecutor][CalcStreamNum] tag[%s] streamNum_[%u]", tag_.c_str(), streamNum);
42 0 : return HCCL_SUCCESS;
43 : }
44 :
45 0 : HcclResult CollBroadcastRingExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
46 : {
47 0 : TransportMemType inputType = TransportMemType::RESERVED;
48 0 : TransportMemType outputType = TransportMemType::RESERVED;
49 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
50 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
51 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
52 0 : return HCCL_SUCCESS;
53 : }
54 :
55 0 : HcclResult CollBroadcastRingExecutor::CalcLevel0CommInfo(
56 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
57 : {
58 0 : HCCL_INFO("[CollBroadcastRingExecutor][CalcLevel0CommInfo]tag[%s] start", tag_.c_str());
59 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_RING_INNER);
60 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
61 0 : HCCL_INFO("[CollBroadcastRingExecutor][CalcLevel0CommInfo]tag[%s] Calc RingComm finish", tag_.c_str());
62 0 : return HCCL_SUCCESS;
63 0 : }
64 :
65 0 : HcclResult CollBroadcastRingExecutor::KernelRun(const OpParam& param, ExecMem& execMem)
66 : {
67 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollBroadcastRingExecutor][KernelRun]The CollBroadcastRingExecutor starts.");
68 0 : u32 perDataSize = 0;
69 0 : CHK_RET(SalGetDataTypeSize(param.DataDes.dataType, perDataSize));
70 :
71 0 : std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
72 0 : std::vector<std::vector<Slice>> multRingsSliceZero; // 数据基于该rank上环0的偏移
73 : // step1: 节点内的scatter
74 0 : u32 ringNum
75 0 : = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
76 :
77 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, ringNum));
78 :
79 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
80 : // 按ranksize得到内存切分slice数
81 0 : u32 sliceNum = level0CommInfo.localRankSize;
82 : // 将根节点数据切分成sliceNum份
83 0 : CHK_RET(AlgTemplateBase::PrepareSliceData(execMem.count, perDataSize, sliceNum, 0, dataSegsSlice));
84 0 : HCCL_DEBUG(
85 : "[CollBroadcastRingExecutor][KernelRun] execMem.count[%llu], perDataSize[%u], sliceNum[%u], ringNum[%u] ",
86 : execMem.count, perDataSize, sliceNum, ringNum);
87 :
88 : /* 外层:scatter */
89 : // 将每slice再切分成4份，按各ring的dev顺序排列
90 0 : if (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) {
91 : // 构造ring algorithm对应的scatter实例
92 0 : multRingsSliceZero = PrepareMultiRingSlice(dataSegsSlice, param.tag, false, topoAttr_.nicList);
93 0 : CHK_PRT_RET(
94 : multRingsSliceZero.size() != ringNum,
95 : HCCL_ERROR(
96 : "[CollBroadcastRingExecutor]"
97 : "ringNum[%u] !=multRingsSliceZero size[%zu]",
98 : ringNum, multRingsSliceZero.size()),
99 : HCCL_E_INTERNAL);
100 : } else {
101 0 : multRingsSliceZero.push_back(dataSegsSlice); // 应该offset全为0，而大小和dataSegsSlice中一样,里面的offset不使用
102 : }
103 :
104 0 : HcomCollOpInfo* scatterOpInfoPtr = nullptr;
105 0 : HcomCollOpInfo scatterOpInfo = {"", execMem.inputPtr, nullptr, execMem.count, param.DataDes.dataType, param.root};
106 0 : if (DMAReduceFlag_) {
107 0 : scatterOpInfoPtr = &scatterOpInfo;
108 : }
109 0 : CHK_RET(MultiRingScatter(
110 : param.tag, execMem.inputMem, execMem.outputMem, execMem.count, param.DataDes.dataType, multRingsSliceZero,
111 : param.root, param.stream, scatterOpInfoPtr));
112 :
113 0 : HCCL_INFO("broadcast 8PringHD stage0 run success");
114 :
115 : // step2: 节点间的broadcast
116 : u64 hdSize;
117 : u32 segmentIdx;
118 : u32 commIndex;
119 0 : CHK_RET(PrepareLevel1CommInfo(segmentIdx, commIndex, hdSize, level0CommInfo, multRingsSliceZero, param.tag));
120 :
121 0 : u64 hdCount = hdSize / perDataSize;
122 0 : auto nicList = topoAttr_.nicList;
123 0 : bool isMultiNic = topoType_ == TopoType::TOPO_TYPE_8P_RING && nicList.size() != DEVICE_EIGHT;
124 0 : std::vector<u32>::iterator iterNic = std::find(nicList.begin(), nicList.end(), topoAttr_.devicePhyId);
125 0 : bool innRunRet = isMultiNic && (iterNic == nicList.end());
126 0 : if (!innRunRet) { // 满足以下条件, 不做server间通信: 1. 8P ring的拓扑 2. 网口不满配 3. 当前device不出网口
127 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
128 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
129 0 : u64 curSize = execMem.count * SIZE_TABLE[param.DataDes.dataType];
130 0 : bool isUsedRegister = false;
131 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
132 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
133 0 : HCCL_DEBUG(
134 : "broadcast ring: curSize[%llu] deviceNumPerAggregation[%u] commLevel0Size[%u].", curSize,
135 : topoAttr_.deviceNumPerAggregation, level0CommInfo.localRankSize);
136 0 : if (curSize / topoAttr_.deviceNumPerAggregation <= NHR_BCAST_SMALL_SIZE) {
137 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
138 0 : TemplateType::TEMPLATE_BROADCAST_NHR_ONESHOT, dispatcher_);
139 : } else {
140 : level1TempAlg
141 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR, dispatcher_);
142 : }
143 0 : HCCL_INFO("broadcast ring: using nhr algo inter-server");
144 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
145 0 : isUsedRegister = true;
146 : level1TempAlg
147 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR_V1, dispatcher_);
148 0 : HCCL_INFO("broadcast ring: using nhr_v1 algo inter-server");
149 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
150 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
151 0 : if (ShouldUseBinaryBroadcastOfNB(
152 0 : curSize / topoAttr_.deviceNumPerAggregation, level1RankSize, topoAttr_.userRankSize,
153 0 : topoAttr_.deviceNumPerAggregation)) {
154 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
155 0 : TemplateType::TEMPLATE_BROADCAST_NB_BINARY, dispatcher_);
156 : } else {
157 : level1TempAlg
158 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NB, dispatcher_);
159 : }
160 0 : HCCL_INFO("broadcast ring: using nonuniform-bruck algo inter-server");
161 : } else {
162 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
163 0 : TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, dispatcher_);
164 0 : HCCL_INFO("broadcast ring: using Recursive halving-doubling algo inter-server.");
165 : }
166 0 : CHK_SMART_PTR_NULL(level1TempAlg);
167 :
168 0 : u32 subUserrankRoot = topoMatcher_->GetSubRootUserRank(topoAttr_.userRank, param.root);
169 0 : CHK_PRT_RET(
170 : subUserrankRoot == INVALID_VALUE_RANKID,
171 : HCCL_ERROR(
172 : "[BroadCastOperator][BroadCastRingExecutor]subUserrankRoot[%u] is invalid,userRank[%u],root[%u]",
173 : subUserrankRoot, topoAttr_.userRank, param.root),
174 : HCCL_E_INTERNAL);
175 0 : u32 planeRoot = 0;
176 0 : u32 level1RankSize = level1CommInfo.localRankSize;
177 0 : u32 level1LocalRank = level1CommInfo.localRank;
178 0 : CHK_RET(GetRankByUserRank(COMM_LEVEL1, commIndex, subUserrankRoot, planeRoot));
179 :
180 : // 节点间的hd 使用环0来记录
181 0 : if (isUsedRegister) {
182 0 : PrepareData prepareData;
183 0 : prepareData.inputMem = execMem.inputMem;
184 0 : prepareData.outputMem = execMem.inputMem;
185 0 : prepareData.scratchMem = execMem.outputMem;
186 0 : prepareData.count = hdCount;
187 0 : prepareData.dataType = param.DataDes.dataType;
188 0 : prepareData.stream = param.stream;
189 0 : prepareData.reductionOp = HCCL_REDUCE_RESERVED;
190 0 : prepareData.root = planeRoot;
191 0 : prepareData.baseOffset = dataSegsSlice[segmentIdx].offset;
192 0 : CHK_RET(level1TempAlg->Prepare(prepareData));
193 0 : } else {
194 0 : CHK_RET(level1TempAlg->Prepare(
195 : execMem.inputMem, execMem.inputMem, execMem.outputMem, hdCount, param.DataDes.dataType, param.stream,
196 : HCCL_REDUCE_RESERVED, planeRoot, std::vector<Slice>(0), dataSegsSlice[segmentIdx].offset));
197 : }
198 :
199 0 : CHK_RET(level1TempAlg->RegisterProfiler(
200 : (level1RankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1LocalRank, PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET,
201 : param.stream));
202 :
203 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
204 0 : }
205 0 : HCCL_INFO("broadcast 8PringHD stage1 run success");
206 :
207 : // step3: 节点内的allgatherring
208 0 : HcomCollOpInfo* allgatherOpInfoPtr = nullptr;
209 0 : HcomCollOpInfo allgatherOpInfo
210 0 : = {"", nullptr, execMem.outputPtr, execMem.count, param.DataDes.dataType, param.root, HCCL_REDUCE_RESERVED};
211 0 : if (DMAReduceFlag_) {
212 0 : allgatherOpInfoPtr = &allgatherOpInfo;
213 : }
214 0 : CHK_RET(MultiRingAllGather(
215 : param.tag, execMem.inputMem, execMem.outputMem, hdCount, param.DataDes.dataType, multRingsSliceZero,
216 : param.stream, PROF_STAGE_2, 0, allgatherOpInfoPtr));
217 :
218 0 : HCCL_INFO("broadcast 8PringHD stage2 run success");
219 0 : return HCCL_SUCCESS;
220 0 : }
221 0 : HcclResult CollBroadcastRingExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
222 : {
223 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
224 0 : return HCCL_E_UNAVAIL;
225 : }
226 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
227 0 : u32 ringNum
228 0 : = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
229 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
230 :
231 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
232 0 : return HCCL_E_UNAVAIL;
233 : }
234 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
235 :
236 0 : return HCCL_SUCCESS;
237 0 : }
238 :
239 0 : HcclResult CollBroadcastRingExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase>& level1TempAlg, u32 level1RankSize)
240 : {
241 0 : if (level1RankSize > 1) {
242 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
243 : level1TempAlg
244 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR, dispatcher_);
245 0 : HCCL_INFO("broadcast ring: using nhr algo inter-server.");
246 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
247 : level1TempAlg
248 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR_V1, dispatcher_);
249 0 : HCCL_INFO("broadcast ring: using nhr_v1 algo inter-server.");
250 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
251 : level1TempAlg
252 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NB, dispatcher_);
253 0 : HCCL_INFO("broadcast ring: using nonuniform-bruck algo inter-server.");
254 : } else {
255 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
256 0 : TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, dispatcher_);
257 0 : HCCL_INFO("broadcast ring: using Recursive halving-doubling algo inter-server.");
258 : }
259 0 : CHK_SMART_PTR_NULL(level1TempAlg);
260 0 : return HCCL_SUCCESS;
261 : }
262 0 : return HCCL_E_UNAVAIL;
263 : }
264 :
265 : REGISTER_EXEC("BroadCastRingExecutor", BroadcastRing, CollBroadcastRingExecutor);
266 :
267 : } // namespace hccl
|
