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(const HcclDispatcher dispatcher,
16 0 : 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 = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
30 : 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(TransportMemType inputType,
56 : 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 ¶m, 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 = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
75 : 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("[CollBroadcastRingExecutor][KernelRun] execMem.count[%llu], perDataSize[%u], sliceNum[%u], ringNum[%u] ",
85 : execMem.count, perDataSize, sliceNum, ringNum);
86 :
87 : /* 外层:scatter */
88 : // 将每slice再切分成4份，按各ring的dev顺序排列
89 0 : if (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) {
90 : // 构造ring algorithm对应的scatter实例
91 0 : multRingsSliceZero = PrepareMultiRingSlice(dataSegsSlice, param.tag, false, topoAttr_.nicList);
92 0 : CHK_PRT_RET(multRingsSliceZero.size() != ringNum, HCCL_ERROR("[CollBroadcastRingExecutor]"\
93 : "ringNum[%u] !=multRingsSliceZero size[%zu]", ringNum, multRingsSliceZero.size()), HCCL_E_INTERNAL);
94 : } else {
95 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
96 : }
97 :
98 0 : HcomCollOpInfo *scatterOpInfoPtr = nullptr;
99 0 : HcomCollOpInfo scatterOpInfo = {
100 0 : "", execMem.inputPtr, nullptr, execMem.count, param.DataDes.dataType, param.root
101 0 : };
102 0 : if (DMAReduceFlag_) {
103 0 : scatterOpInfoPtr = &scatterOpInfo;
104 : }
105 0 : CHK_RET(MultiRingScatter(param.tag, execMem.inputMem, execMem.outputMem, execMem.count, param.DataDes.dataType,
106 : multRingsSliceZero, param.root, param.stream, scatterOpInfoPtr));
107 :
108 0 : HCCL_INFO("broadcast 8PringHD stage0 run success");
109 :
110 : // step2: 节点间的broadcast
111 : u64 hdSize;
112 : u32 segmentIdx;
113 : u32 commIndex;
114 0 : CHK_RET(PrepareLevel1CommInfo(segmentIdx, commIndex, hdSize, level0CommInfo, multRingsSliceZero, param.tag));
115 :
116 0 : u64 hdCount = hdSize / perDataSize;
117 0 : auto nicList = topoAttr_.nicList;
118 0 : bool isMultiNic = topoType_ == TopoType::TOPO_TYPE_8P_RING && nicList.size() != DEVICE_EIGHT;
119 0 : std::vector<u32>::iterator iterNic = std::find(nicList.begin(), nicList.end(), topoAttr_.devicePhyId);
120 0 : bool innRunRet = isMultiNic && (iterNic == nicList.end());
121 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
122 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
123 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
124 0 : u64 curSize = execMem.count * SIZE_TABLE[param.DataDes.dataType];
125 0 : bool isUsedRegister = false;
126 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
127 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
128 0 : HCCL_DEBUG("broadcast ring: curSize[%llu] deviceNumPerAggregation[%u] commLevel0Size[%u].",
129 : curSize, topoAttr_.deviceNumPerAggregation, level0CommInfo.localRankSize);
130 0 : if (curSize / topoAttr_.deviceNumPerAggregation <= NHR_BCAST_SMALL_SIZE) {
131 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
132 0 : TemplateType::TEMPLATE_BROADCAST_NHR_ONESHOT, dispatcher_);
133 : } else {
134 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
135 0 : TemplateType::TEMPLATE_BROADCAST_NHR, dispatcher_);
136 : }
137 0 : HCCL_INFO("broadcast ring: using nhr algo inter-server");
138 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
139 0 : isUsedRegister = true;
140 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR_V1,
141 0 : dispatcher_);
142 0 : HCCL_INFO("broadcast ring: using nhr_v1 algo inter-server");
143 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
144 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
145 0 : if (ShouldUseBinaryBroadcastOfNB(curSize / topoAttr_.deviceNumPerAggregation, level1RankSize,
146 0 : topoAttr_.userRankSize, topoAttr_.deviceNumPerAggregation)) {
147 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
148 0 : TemplateType::TEMPLATE_BROADCAST_NB_BINARY, dispatcher_);
149 : } else {
150 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
151 0 : TemplateType::TEMPLATE_BROADCAST_NB, dispatcher_);
152 : }
153 0 : HCCL_INFO("broadcast ring: using nonuniform-bruck algo inter-server");
154 : } else {
155 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
156 0 : TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, dispatcher_);
157 0 : HCCL_INFO("broadcast ring: using Recursive halving-doubling algo inter-server.");
158 : }
159 0 : CHK_SMART_PTR_NULL(level1TempAlg);
160 :
161 0 : u32 subUserrankRoot = topoMatcher_->GetSubRootUserRank(topoAttr_.userRank, param.root);
162 0 : CHK_PRT_RET(subUserrankRoot == INVALID_VALUE_RANKID,
163 : HCCL_ERROR("[BroadCastOperator][BroadCastRingExecutor]subUserrankRoot[%u] is invalid,userRank[%u],root[%u]",
164 : subUserrankRoot, topoAttr_.userRank, param.root), HCCL_E_INTERNAL);
165 0 : u32 planeRoot = 0;
166 0 : u32 level1RankSize = level1CommInfo.localRankSize;
167 0 : u32 level1LocalRank = level1CommInfo.localRank;
168 0 : CHK_RET(GetRankByUserRank(COMM_LEVEL1, commIndex, subUserrankRoot, planeRoot));
169 :
170 : // 节点间的hd 使用环0来记录
171 0 : if (isUsedRegister) {
172 0 : PrepareData prepareData;
173 0 : prepareData.inputMem = execMem.inputMem;
174 0 : prepareData.outputMem = execMem.inputMem;
175 0 : prepareData.scratchMem = execMem.outputMem;
176 0 : prepareData.count = hdCount;
177 0 : prepareData.dataType = param.DataDes.dataType;
178 0 : prepareData.stream = param.stream;
179 0 : prepareData.reductionOp = HCCL_REDUCE_RESERVED;
180 0 : prepareData.root = planeRoot;
181 0 : prepareData.baseOffset = dataSegsSlice[segmentIdx].offset;
182 0 : CHK_RET(level1TempAlg->Prepare(prepareData));
183 0 : } else {
184 0 : CHK_RET(level1TempAlg->Prepare(execMem.inputMem, execMem.inputMem, execMem.outputMem, hdCount, param.DataDes.dataType,
185 : param.stream, HCCL_REDUCE_RESERVED, planeRoot, std::vector<Slice>(0), dataSegsSlice[segmentIdx].offset));
186 : }
187 :
188 0 : CHK_RET(level1TempAlg->RegisterProfiler((level1RankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1LocalRank, \
189 : PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET, param.stream));
190 :
191 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
192 0 : }
193 0 : HCCL_INFO("broadcast 8PringHD stage1 run success");
194 :
195 : // step3: 节点内的allgatherring
196 0 : HcomCollOpInfo *allgatherOpInfoPtr = nullptr;
197 0 : HcomCollOpInfo allgatherOpInfo = {
198 0 : "", nullptr, execMem.outputPtr, execMem.count, param.DataDes.dataType, param.root, HCCL_REDUCE_RESERVED
199 0 : };
200 0 : if (DMAReduceFlag_) {
201 0 : allgatherOpInfoPtr = &allgatherOpInfo;
202 : }
203 0 : CHK_RET(MultiRingAllGather(param.tag, execMem.inputMem, execMem.outputMem, hdCount, param.DataDes.dataType,
204 : multRingsSliceZero, param.stream, PROF_STAGE_2, 0, allgatherOpInfoPtr));
205 :
206 0 : HCCL_INFO("broadcast 8PringHD stage2 run success");
207 0 : return HCCL_SUCCESS;
208 0 : }
209 0 : HcclResult CollBroadcastRingExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
210 : {
211 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
212 0 : return HCCL_E_UNAVAIL;
213 : }
214 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
215 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
216 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
217 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
218 :
219 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
220 0 : return HCCL_E_UNAVAIL;
221 : }
222 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
223 :
224 0 : return HCCL_SUCCESS;
225 0 : }
226 :
227 0 : HcclResult CollBroadcastRingExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
228 : {
229 0 : if (level1RankSize > 1) {
230 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
231 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
232 0 : TemplateType::TEMPLATE_BROADCAST_NHR, dispatcher_);
233 0 : HCCL_INFO("broadcast ring: using nhr algo inter-server.");
234 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
235 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR_V1,
236 0 : dispatcher_);
237 0 : HCCL_INFO("broadcast ring: using nhr_v1 algo inter-server.");
238 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
239 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
240 0 : TemplateType::TEMPLATE_BROADCAST_NB, dispatcher_);
241 0 : HCCL_INFO("broadcast ring: using nonuniform-bruck algo inter-server.");
242 : } else {
243 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
244 0 : TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, dispatcher_);
245 0 : HCCL_INFO("broadcast ring: using Recursive halving-doubling algo inter-server.");
246 : }
247 0 : CHK_SMART_PTR_NULL(level1TempAlg);
248 0 : return HCCL_SUCCESS;
249 : }
250 0 : return HCCL_E_UNAVAIL;
251 : }
252 :
253 : REGISTER_EXEC("BroadCastRingExecutor", BroadcastRing, CollBroadcastRingExecutor);
254 :
255 : } // namespace hccl
|
