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_mesh_executor.h"
12 :
13 : namespace hccl {
14 :
15 1 : CollBroadcastMeshExecutor::CollBroadcastMeshExecutor(const HcclDispatcher dispatcher,
16 1 : std::unique_ptr<TopoMatcher> &topoMatcher)
17 1 : : CollBroadcastExecutor(dispatcher, topoMatcher)
18 : {
19 1 : }
20 :
21 0 : HcclResult CollBroadcastMeshExecutor::CalcStreamNum(u32& streamNum)
22 : {
23 0 : u32 totalStreamNum = 0U;
24 0 : if (algType_.algoLevel0 == AlgTypeLevel0::ALG_LEVEL0_4P_MESH) {
25 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_4PMESH;
26 : } else {
27 0 : if ((workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) &&
28 0 : (topoAttr_.deviceType == DevType::DEV_TYPE_910B) && topoAttr_.isSingleMeshAggregation ) {
29 0 : totalStreamNum = topoAttr_.deviceNumPerAggregation;
30 : } else {
31 0 : totalStreamNum = topoAttr_.deviceNumPerAggregation - 1;
32 : }
33 : }
34 0 : streamNum = totalStreamNum > 0 ? totalStreamNum - 1 : 0;
35 :
36 0 : HCCL_INFO("[CollBroadcastMeshExecutor][CalcStreamNum] tag[%s] streamNum_[%u]", tag_.c_str(), streamNum);
37 0 : return HCCL_SUCCESS;
38 : }
39 :
40 0 : HcclResult CollBroadcastMeshExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
41 : {
42 0 : TransportMemType inputType = TransportMemType::RESERVED;
43 0 : TransportMemType outputType = TransportMemType::RESERVED;
44 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
45 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
46 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
47 0 : return HCCL_SUCCESS;
48 : }
49 :
50 0 : HcclResult CollBroadcastMeshExecutor::CalcLevel0CommInfo(TransportMemType inputType,
51 : TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
52 : {
53 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
54 0 : commParaLevel0.meshSinglePlane = true;
55 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
56 0 : return HCCL_SUCCESS;
57 0 : }
58 :
59 0 : HcclResult CollBroadcastMeshExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
60 : {
61 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollBroadcastMeshExecutor][KernelRun] userRank[%u] starts.", topoAttr_.userRank);
62 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
63 :
64 0 : bool isUsedRegister = false;
65 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg1;
66 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
67 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg2;
68 :
69 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
70 0 : u32 commIndex = level0CommInfo.localRank;
71 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
72 :
73 0 : level0TempAlg1 = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_MESH, dispatcher_);
74 0 : CHK_SMART_PTR_NULL(level0TempAlg1);
75 0 : CHK_RET(level0TempAlg1->Prepare(level0CommInfo.localRank, level0CommInfo.localRankSize));
76 0 : level0TempAlg1->CloseBarrier();
77 :
78 : /* 内层topo:all_reduce */
79 : /* 外层所有rank均参与内层的broadcast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 */
80 0 : std::vector<Slice> slice;
81 0 : CHK_RET(GetRankSliceSize(param.DataDes.dataType, execMem.count, level0CommInfo.localRankSize, slice));
82 :
83 0 : CHK_PRT_RET(slice.empty(), HCCL_ERROR("[BroadCastOperator][BroadCastMeshExecutor]got slice is empty"),
84 : HCCL_E_INTERNAL);
85 :
86 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
87 :
88 0 : u64 curSize = execMem.count * SIZE_TABLE[param.DataDes.dataType];
89 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
90 0 : HCCL_DEBUG("broadcast mesh: curSize[%llu] deviceNumPerAggregation[%u] commLevel0Size[%u]",
91 : curSize, topoAttr_.deviceNumPerAggregation, level0CommInfo.localRankSize);
92 0 : if (curSize / topoAttr_.deviceNumPerAggregation <= NHR_BCAST_SMALL_SIZE) {
93 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
94 0 : TemplateType::TEMPLATE_BROADCAST_NHR_ONESHOT, dispatcher_);
95 : } else {
96 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
97 0 : TemplateType::TEMPLATE_BROADCAST_NHR, dispatcher_);
98 : }
99 0 : HCCL_INFO("broadcast mesh: using nhr algo inter-server.");
100 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
101 0 : isUsedRegister = true;
102 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR_V1,
103 0 : dispatcher_);
104 0 : HCCL_INFO("broadcast mesh: using nhr_v1 algo inter-server.");
105 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
106 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
107 0 : HCCL_DEBUG("[CollBroadcastMeshExecutor][KernelRun]level1RankSize is %u", level1RankSize);
108 0 : if (ShouldUseBinaryBroadcastOfNB(curSize / topoAttr_.deviceNumPerAggregation, level1RankSize,
109 0 : topoAttr_.userRankSize, topoAttr_.deviceNumPerAggregation)) {
110 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
111 0 : TemplateType::TEMPLATE_BROADCAST_NB_BINARY, dispatcher_);
112 : } else {
113 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
114 0 : TemplateType::TEMPLATE_BROADCAST_NB, dispatcher_);
115 : }
116 0 : HCCL_INFO("broadcast mesh: using nonuniform-bruck algo inter-server.");
117 : } else {
118 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
119 0 : TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, dispatcher_);
120 0 : HCCL_INFO("broadcast mesh: using Recursive halving-doubling algo inter-server.");
121 : }
122 0 : CHK_SMART_PTR_NULL(level1TempAlg);
123 :
124 : /* 外层topo:all_gather */
125 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910B) {
126 0 : level0TempAlg2 = AlgTemplateRegistry::Instance().GetAlgTemplate(
127 0 : TemplateType::TEMPLATE_ALL_GATHER_MESH_ATOMIC, dispatcher_);
128 : } else {
129 0 : level0TempAlg2 = AlgTemplateRegistry::Instance().GetAlgTemplate(
130 0 : TemplateType::TEMPLATE_ALL_GATHER_MESH, dispatcher_);
131 : }
132 0 : CHK_SMART_PTR_NULL(level0TempAlg2);
133 0 : CHK_RET(level0TempAlg2->Prepare(algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux,
134 : topoAttr_.userRank, nullptr, level0CommInfo.localRank, level0CommInfo.localRankSize));
135 :
136 : /* 节点内执行器 stage0 */
137 0 : u32 rootRank = 0;
138 0 : HcclResult ret = GetRankByUserRank(COMM_LEVEL0, COMM_INDEX_0, param.root, rootRank);
139 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
140 : HCCL_ERROR("[BroadCastOperator][BroadCastMeshExecutor]invalid root[%u] to get userrank", param.root), ret);
141 :
142 0 : if (ret == HCCL_SUCCESS) {
143 0 : CHK_RET(level0TempAlg1->Prepare(execMem.inputMem, execMem.outputMem, execMem.outputMem, execMem.count,
144 : param.DataDes.dataType, param.stream, HCCL_REDUCE_RESERVED, rootRank, slice));
145 :
146 0 : u32 rankSize = level0CommInfo.localRankSize;
147 0 : CHK_RET(level0TempAlg1->RegisterProfiler(
148 : (0 << PROF_RINGINDEX_OFFSET_OF_PLANEID)+(rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) +
149 : level0CommInfo.localRank, PROF_STAGE_0, HCCL_EXEC_STEP_NOT_SET, param.stream));
150 :
151 0 : CHK_RET(RunTemplate(level0TempAlg1, level0CommInfo));
152 : } else {
153 0 : HCCL_ERROR("[BroadCastOperator][BroadCastMeshExecutor]invalid root[%u] to get userrank", param.root);
154 : }
155 0 : HCCL_INFO("[BroadCastOperator][BroadCastMeshExecutor] stage0 run success");
156 0 : u64 hdCount = slice[level0CommInfo.localRank].size / perDataSize;
157 : /* 节点间执行器 stage1 */
158 :
159 0 : u32 subUserrankRoot = topoMatcher_->GetSubRootUserRank(topoAttr_.userRank, param.root);
160 0 : CHK_PRT_RET(subUserrankRoot == INVALID_VALUE_RANKID,
161 : HCCL_ERROR("[BroadCastOperator][BroadCastMeshExecutor]subUserrankRoot[%u] is invalid,userRank[%u],root[%u]",
162 : subUserrankRoot, topoAttr_.userRank, param.root),
163 : HCCL_E_INTERNAL);
164 :
165 0 : u32 subRoot = 0;
166 0 : CHK_RET(GetRankByUserRank(COMM_LEVEL1, commIndex, subUserrankRoot, subRoot));
167 :
168 : // 增加偏移参数
169 0 : if (isUsedRegister) {
170 0 : PrepareData prepareData;
171 0 : prepareData.inputMem = execMem.inputMem;
172 0 : prepareData.outputMem = execMem.outputMem;
173 0 : prepareData.scratchMem = execMem.outputMem;
174 0 : prepareData.count = hdCount;
175 0 : prepareData.dataType = param.DataDes.dataType;
176 0 : prepareData.stream = param.stream;
177 0 : prepareData.reductionOp = HCCL_REDUCE_RESERVED;
178 0 : prepareData.root = subRoot;
179 0 : prepareData.baseOffset = slice[level0CommInfo.localRank].offset;
180 0 : CHK_RET(level1TempAlg->Prepare(prepareData));
181 0 : } else {
182 0 : CHK_RET(level1TempAlg->Prepare(execMem.inputMem, execMem.outputMem, execMem.outputMem, hdCount,
183 : param.DataDes.dataType, param.stream, HCCL_REDUCE_RESERVED, subRoot,
184 : std::vector<Slice>(0), slice[level0CommInfo.localRank].offset));
185 : }
186 :
187 0 : u32 rankSize = level1CommInfo.localRankSize;
188 0 : CHK_RET(level1TempAlg->RegisterProfiler((0 << PROF_RINGINDEX_OFFSET_OF_PLANEID) +
189 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1CommInfo.localRank,
190 : PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET, param.stream));
191 :
192 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
193 0 : HCCL_INFO("[BroadCastOperator][BroadCastMeshExecutor] stage1 run success");
194 :
195 : /* 节点内执行器 stage2 */
196 : {
197 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OPS_KERNEL_INFO_LIB) { // offline
198 0 : for (u32 streamIndex = 0; streamIndex < algResResp_->slaveStreams.size(); streamIndex++) {
199 0 : CHK_RET(StreamActiveManager::GetInstance(topoAttr_.deviceLogicId).StreamActive(
200 : algResResp_->slaveStreams[streamIndex].ptr(), param.stream.ptr()));
201 : }
202 : }
203 0 : CHK_RET(level0TempAlg2->Prepare(execMem.outputMem, execMem.outputMem, execMem.outputMem, execMem.count,
204 : param.DataDes.dataType, param.stream, HCCL_REDUCE_RESERVED, LEVEL0_BRIDGE_RANK_ID, slice));
205 :
206 0 : u32 rankSize = level0CommInfo.localRankSize;
207 0 : CHK_RET(level0TempAlg2->RegisterProfiler((0 << PROF_RINGINDEX_OFFSET_OF_PLANEID) +
208 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
209 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
210 :
211 0 : CHK_RET(RunTemplate(level0TempAlg2, level0CommInfo));
212 : }
213 :
214 0 : HCCL_INFO("[BroadCastOperator][BroadCastMeshExecutor] stage2 run success");
215 0 : return HCCL_SUCCESS;
216 0 : }
217 0 : HcclResult CollBroadcastMeshExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
218 : {
219 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
220 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
221 :
222 0 : u32 commIndex = level0CommInfo.localRank; // 找到rank所在的节点间平面
223 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
224 :
225 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
226 0 : return HCCL_SUCCESS;
227 0 : }
228 :
229 0 : HcclResult CollBroadcastMeshExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
230 : {
231 0 : if (level1RankSize > 1) {
232 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
233 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
234 0 : TemplateType::TEMPLATE_BROADCAST_NHR, dispatcher_);
235 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
236 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_BROADCAST_NHR_V1,
237 0 : dispatcher_);
238 0 : HCCL_INFO("broadcast mesh: using nhr_v1 algo inter-server.");
239 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
240 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
241 0 : TemplateType::TEMPLATE_BROADCAST_NB, dispatcher_);
242 0 : HCCL_INFO("broadcast mesh: using nonuniform-bruck algo inter-server.");
243 : } else {
244 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
245 0 : TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, dispatcher_);
246 0 : HCCL_INFO("broadcast mesh: using Recursive halving-doubling algo inter-server.");
247 : }
248 0 : CHK_SMART_PTR_NULL(level1TempAlg);
249 : }
250 0 : return HCCL_SUCCESS;
251 : }
252 : REGISTER_EXEC("BroadCastMeshExecutor", BroadcastMesh, CollBroadcastMeshExecutor);
253 :
254 : } // namespace hccl
|
