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_all_reduce_mesh_executor.h"
12 :
13 : namespace hccl {
14 :
15 4 : CollAllReduceMeshExecutor::CollAllReduceMeshExecutor(const HcclDispatcher dispatcher,
16 4 : std::unique_ptr<TopoMatcher> &topoMatcher)
17 4 : : CollAllReduceExecutor(dispatcher, topoMatcher)
18 : {
19 4 : DMAReduceFlag_ = false;
20 4 : }
21 :
22 4 : void CollAllReduceMeshExecutor::ParseParam(const OpParam& param)
23 : {
24 4 : tag_ = param.tag;
25 7 : bool isInlineReduce = IsSupportSDMAReduce(param.inputPtr, param.outputPtr,
26 4 : param.DataDes.dataType, param.reduceType);
27 9 : meshSinglePlane_ = (topoAttr_.deviceType == DevType::DEV_TYPE_910B) &&
28 3 : topoMatcher_->GetExternalInputHcclDeterministic() == DETERMINISTIC_DISABLE &&
29 6 : isInlineReduce && (workflowMode_ != HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
30 3 : aicpuUnfoldMode_ = param.aicpuUnfoldMode;
31 3 : }
32 :
33 4 : HcclResult CollAllReduceMeshExecutor::CalcStreamNum(u32& streamNum)
34 : {
35 4 : u32 totalStreamNum = topoAttr_.deviceNumPerAggregation > 1U ? topoAttr_.deviceNumPerAggregation - 1U : 1U;
36 4 : streamNum = totalStreamNum - 1U;
37 4 : HCCL_INFO("[CollAllReduceMeshExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
38 : tag_.c_str(), streamNum);
39 4 : return HCCL_SUCCESS;
40 : }
41 :
42 4 : HcclResult CollAllReduceMeshExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
43 : {
44 4 : TransportMemType inputType = TransportMemType::RESERVED;
45 4 : TransportMemType outputType = TransportMemType::RESERVED;
46 4 : CHK_RET(CalcTransportMemType(inputType, outputType));
47 4 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
48 4 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
49 4 : return HCCL_SUCCESS;
50 : }
51 :
52 4 : HcclResult CollAllReduceMeshExecutor::CalcTransportMemType(TransportMemType &inputType, TransportMemType &outputType)
53 : {
54 4 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
55 0 : inputType = TransportMemType::CCL_INPUT;
56 0 : outputType = TransportMemType::CCL_OUTPUT;
57 : } else {
58 4 : inputType = TransportMemType::PARAM_INPUT;
59 4 : outputType = TransportMemType::PARAM_OUTPUT;
60 : }
61 4 : HCCL_INFO("[CollAllReduceMeshExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]",
62 : tag_.c_str(), inputType, outputType);
63 4 : return HCCL_SUCCESS;
64 : }
65 :
66 4 : HcclResult CollAllReduceMeshExecutor::CalcLevel0CommInfo(TransportMemType inputType,
67 : TransportMemType outputType,
68 : std::vector<LevelNSubCommTransport>& opTransport)
69 : {
70 4 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
71 4 : commParaLevel0.meshSinglePlane = meshSinglePlane_;
72 4 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
73 4 : return HCCL_SUCCESS;
74 4 : }
75 :
76 0 : bool CollAllReduceMeshExecutor::IsHugeData(const u64 curSize)
77 : {
78 0 : bool hugeData = curSize / topoAttr_.deviceNumPerAggregation / HCCL_INTERNODE_MAX_DATA_RATE > RDMA_SEND_MAX_SIZE ||
79 : curSize > SDMA_SEND_MAX_SIZE;
80 0 : return hugeData;
81 : }
82 :
83 0 : bool CollAllReduceMeshExecutor::IsSmallData(const u64 totalSize, const u64 curSize)
84 : {
85 0 : bool smallData = IsAllReduceSmallData(curSize);
86 0 : return smallData;
87 : }
88 :
89 0 : HcclResult CollAllReduceMeshExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
90 : {
91 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollAllReduceMeshExecutor][KernelRun] userRank[%u] starts.", topoAttr_.userRank);
92 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
93 :
94 0 : std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
95 0 : std::vector<std::vector<Slice> > multiStreamSlice; // 每个stream使用的数据基于用户buffer的偏移
96 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
97 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg;
98 :
99 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
100 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
101 :
102 0 : u32 sliceNum = level0CommInfo.localRankSize;
103 : // 根据数据量算每个环上数据的偏移和大小
104 0 : CHK_RET(AlgTemplateBase::PrepareSliceData(execMem.count, perDataSize, sliceNum, 0, dataSegsSlice));
105 : // mesh算法stream数量为server内rank数减1
106 :
107 0 : CHK_RET(ActiveSlaveStreams(param.stream));
108 :
109 0 : if (topoMatcher_->GetExternalInputHcclDeterministic() == DETERMINISTIC_DISABLE &&
110 0 : (param.DataDes.dataType != HCCL_DATA_TYPE_INT64) &&
111 0 : (topoAttr_.deviceType == DevType::DEV_TYPE_910B && param.reduceType != HCCL_REDUCE_PROD)) {
112 0 : CHK_RET(MultiStreamReduceScatterMeshAtomic(param.tag, execMem.inputMem, execMem.outputMem, execMem.count,
113 : param.DataDes.dataType, param.reduceType, dataSegsSlice, const_cast<Stream&>(param.stream), COMM_LEVEL0));
114 : } else {
115 0 : CHK_RET(AlgTemplateBase::PrepareSliceMeshStreams(dataSegsSlice, sliceNum - 1, multiStreamSlice));
116 0 : CHK_RET(MultiStreamReduceScatterMesh(param.tag, execMem.inputMem, execMem.outputMem, execMem.count,
117 : param.DataDes.dataType, param.reduceType, multiStreamSlice,
118 : const_cast<Stream&>(param.stream), COMM_LEVEL0));
119 : }
120 :
121 0 : HCCL_INFO("AllReduce meshhd stage0 run success.");
122 :
123 : /* 内层topo:all_reduce */
124 : /* 外层所有rank均参与内层的allReduce计算，所以此处对rank不作限制，但是每个rank需找到自己所在的内层通信域 */
125 0 : u32 commIndex = level0CommInfo.localRank;
126 0 : CHK_PRT_RET(commIndex >= dataSegsSlice.size(),
127 : HCCL_ERROR("[CollAllReduceMeshExecutor][Run]commIndex[%u] >= dataSegsSlice size[%zu]", commIndex,
128 : dataSegsSlice.size()), HCCL_E_INTERNAL);
129 :
130 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
131 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
132 :
133 0 : DeviceMem allreduceInput = execMem.inputMem.range(dataSegsSlice[commIndex].offset, dataSegsSlice[commIndex].size);
134 0 : CHK_SMART_PTR_NULL(allreduceInput);
135 0 : DeviceMem allreduceOutput = execMem.outputMem.range(dataSegsSlice[commIndex].offset, dataSegsSlice[commIndex].size);
136 0 : CHK_SMART_PTR_NULL(allreduceOutput);
137 :
138 0 : u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
139 :
140 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
141 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RING, dispatcher_);
142 0 : HCCL_INFO("AllReduce mesh: using ring algo inter-server.");
143 0 : CHK_SMART_PTR_NULL(level1TempAlg);
144 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
145 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
146 0 : u64 curSize = execMem.count * perDataSize; // 单位 byte
147 0 : HCCL_DEBUG("AllReduce mesh: curSize[%llu] deviceNumPerAggregation[%u] commLevel0Size[%u]",
148 : curSize, topoAttr_.deviceNumPerAggregation, level0CommInfo.localRankSize);
149 0 : if (curSize / topoAttr_.deviceNumPerAggregation <= NHR_ALLREDUCE_SMALL_SIZE) {
150 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR_ONESHOT, dispatcher_);
151 : } else {
152 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR, dispatcher_);
153 : }
154 0 : HCCL_INFO("AllReduce mesh: using nhr algo inter-server.");
155 0 : CHK_SMART_PTR_NULL(level1TempAlg);
156 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
157 0 : level1TempAlg->CloseBarrier();
158 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
159 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR_V1, dispatcher_);
160 0 : HCCL_INFO("AllReduce mesh: using nhr_v1 algo inter-server.");
161 0 : CHK_SMART_PTR_NULL(level1TempAlg);
162 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
163 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
164 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
165 0 : TemplateType::TEMPLATE_ALL_REDUCE_NB, dispatcher_);
166 0 : HCCL_INFO("AllReduce mesh: using nb algo inter-server.");
167 0 : CHK_SMART_PTR_NULL(level1TempAlg);
168 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
169 : } else {
170 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
171 0 : TemplateType::TEMPLATE_ALL_REDUCE_RECURSIVE_HALVING_DOUBLING, dispatcher_);
172 0 : HCCL_INFO("AllReduce mesh: using Recursive halving-doubling algo inter-server.");
173 0 : CHK_SMART_PTR_NULL(level1TempAlg);
174 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
175 : }
176 0 : CHK_SMART_PTR_NULL(level1TempAlg);
177 :
178 0 : u32 rankSize = level1CommInfo.localRankSize;
179 : // 节点间的hd 使用环0来记录
180 :
181 0 : u64 hdCount = dataSegsSlice[commIndex].size / perDataSize;
182 0 : CHK_RET(level1TempAlg->Prepare(allreduceInput, allreduceOutput, allreduceOutput, hdCount,
183 : param.DataDes.dataType, param.stream, param.reduceType,
184 : LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), dataSegsSlice[commIndex].offset));
185 :
186 0 : CHK_RET(level1TempAlg->RegisterProfiler((rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) +
187 : level1CommInfo.localRank, PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET, param.stream));
188 :
189 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
190 :
191 0 : HCCL_INFO("AllReduce meshhd stage1 run success.");
192 :
193 : /* 外层topo:all_gather */
194 :
195 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910B) {
196 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_MESH_ATOMIC,
197 0 : dispatcher_);
198 : } else {
199 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_MESH,
200 0 : dispatcher_);
201 : }
202 :
203 0 : CHK_SMART_PTR_NULL(level0TempAlg);
204 0 : CHK_RET(level0TempAlg->Prepare(algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux,
205 : topoAttr_.userRank, nullptr, level0CommInfo.localRank, level0CommInfo.localRankSize));
206 :
207 : /* 节点内执行器 stage2 */
208 : {
209 0 : u32 rankSize = level0CommInfo.localRankSize;
210 0 : CHK_RET(level0TempAlg->Prepare(execMem.outputMem, execMem.outputMem, execMem.inputMem, execMem.count,
211 : param.DataDes.dataType, param.stream, param.reduceType,
212 : LEVEL0_BRIDGE_RANK_ID, dataSegsSlice, 0));
213 :
214 0 : CHK_RET(level0TempAlg->RegisterProfiler(
215 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
216 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
217 :
218 0 : CHK_RET(RunTemplate(level0TempAlg, level0CommInfo));
219 : }
220 :
221 0 : HCCL_INFO("AllReduce meshhd stage2 run success");
222 0 : return HCCL_SUCCESS;
223 0 : }
224 :
225 :
226 0 : HcclResult CollAllReduceMeshExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
227 : {
228 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
229 0 : return HCCL_E_UNAVAIL;
230 : }
231 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
232 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
233 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
234 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
235 :
236 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
237 0 : return HCCL_E_UNAVAIL;
238 : }
239 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
240 :
241 0 : return HCCL_SUCCESS;
242 0 : }
243 :
244 0 : HcclResult CollAllReduceMeshExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
245 : {
246 0 : if (level1RankSize > 1) {
247 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
248 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RING, dispatcher_);
249 0 : HCCL_INFO("AllReduce mesh: using ring algo inter-server.");
250 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
251 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR, dispatcher_);
252 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
253 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR_V1, dispatcher_);
254 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_AHC) {
255 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_AHC, dispatcher_);
256 0 : HCCL_INFO("AllReduce mesh: using ahc algo inter-server.");
257 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_AHC_BROKE) {
258 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_AHC_BROKE, dispatcher_);
259 0 : HCCL_INFO("AllReduce mesh: using ahc-broke algo inter-server.");
260 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
261 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
262 0 : TemplateType::TEMPLATE_ALL_REDUCE_NB, dispatcher_);
263 0 : HCCL_INFO("AllReduce mesh: using nb algo inter-server.");
264 : } else {
265 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
266 0 : TemplateType::TEMPLATE_ALL_REDUCE_RECURSIVE_HALVING_DOUBLING, dispatcher_);
267 0 : HCCL_INFO("AllReduce mesh: using Recursive halving-doubling algo inter-server.");
268 : }
269 0 : CHK_SMART_PTR_NULL(level1TempAlg);
270 0 : return HCCL_SUCCESS;
271 : }
272 0 : return HCCL_E_UNAVAIL;
273 : }
274 : REGISTER_EXEC("AllReduceMeshExecutor", AllReduceMesh, CollAllReduceMeshExecutor);
275 :
276 : } // namespace hccl
|
