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_gather_mesh_graph_executor.h"
12 :
13 : namespace hccl {
14 4 : CollAllGatherMeshGraphExecutor::CollAllGatherMeshGraphExecutor(const HcclDispatcher dispatcher,
15 4 : std::unique_ptr<TopoMatcher> &topoMatcher)
16 4 : : CollAllGatherExecutor(dispatcher, topoMatcher)
17 : {
18 4 : DMAReduceFlag_ = false;
19 4 : }
20 :
21 4 : HcclResult CollAllGatherMeshGraphExecutor::CalcStreamNum(u32& streamNum)
22 : {
23 4 : u32 totalStreamNum = topoAttr_.deviceNumPerAggregation > 1U ? topoAttr_.deviceNumPerAggregation - 1U : 1U;
24 4 : streamNum = totalStreamNum - 1U;
25 4 : HCCL_INFO("[CollAllGatherMeshGraphExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
26 : tag_.c_str(), streamNum);
27 4 : return HCCL_SUCCESS;
28 : }
29 :
30 4 : HcclResult CollAllGatherMeshGraphExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
31 : {
32 4 : TransportMemType inputType = TransportMemType::RESERVED;
33 4 : TransportMemType outputType = TransportMemType::RESERVED;
34 4 : CHK_RET(CalcTransportMemType(inputType, outputType));
35 4 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
36 4 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
37 4 : return HCCL_SUCCESS;
38 : }
39 :
40 4 : HcclResult CollAllGatherMeshGraphExecutor::CalcTransportMemType(TransportMemType &inputType, TransportMemType &outputType)
41 : {
42 4 : inputType = TransportMemType::PARAM_INPUT;
43 4 : outputType = TransportMemType::PARAM_OUTPUT;
44 4 : HCCL_INFO("[CollAllGatherMeshGraphExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]",
45 : tag_.c_str(), inputType, outputType);
46 4 : return HCCL_SUCCESS;
47 : }
48 :
49 4 : HcclResult CollAllGatherMeshGraphExecutor::CalcLevel0CommInfo(TransportMemType inputType, TransportMemType outputType,
50 : std::vector<LevelNSubCommTransport>& opTransport)
51 : {
52 4 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
53 12 : commParaLevel0.meshSinglePlane = (topoAttr_.deviceType == DevType::DEV_TYPE_910B) &&
54 8 : topoMatcher_->GetExternalInputHcclDeterministic() == DETERMINISTIC_DISABLE &&
55 4 : (workflowMode_ != HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
56 4 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
57 4 : return HCCL_SUCCESS;
58 4 : }
59 :
60 0 : HcclResult CollAllGatherMeshGraphExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
61 : {
62 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollAllGatherMeshGraphExecutor][KernelRun]AllGather start");
63 :
64 : // 获取子通信域信息
65 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
66 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
67 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
68 0 : u32 level0RankSize = level0CommInfo.localRankSize;
69 0 : u32 commIndex = level0CommInfo.localRank;
70 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
71 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
72 0 : u32 level1RankSize = level1CommInfo.localRankSize;
73 0 : u32 serverIndex = level1CommInfo.localRank;
74 0 : u64 inputMemSize = execMem.inputMem.size();
75 0 : CHK_RET(ActiveSlaveStreams(param.stream));
76 :
77 : // 第一步，将数据从input内存拷贝到output内存的对应位置
78 0 : u32 transposeRankId = commIndex * level1RankSize + serverIndex;
79 0 : u64 initOffset = transposeRankId * inputMemSize;
80 0 : DeviceMem dstMem = execMem.outputMem.range(initOffset, inputMemSize);
81 0 : CHK_SMART_PTR_NULL(dstMem);
82 :
83 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, dstMem, execMem.inputMem, const_cast<Stream&>(param.stream));
84 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
85 : HCCL_ERROR("[CollAllGatherMeshGraphExecutor][KernelRun]AllGather 4PmeshHD memcpy Failed, Offset[%llu], Size[%llu].",
86 : initOffset, inputMemSize), ret);
87 :
88 : // 第二步，各个AI Server 间 AllGather
89 0 : u64 inputDataCount = inputMemSize / perDataSize;
90 0 : u64 level1Offset = commIndex * level1RankSize * inputMemSize;
91 0 : HCCL_INFO("[CollAllGatherMeshGraphExecutor][KernelRun] userRank[%u] commIndex[%u] level1Offset[%llu] level0RankSize[%u] " \
92 : "level1RankSize[%u] outputMemSize[%llu] count[%u]", topoAttr_.userRank, commIndex, level1Offset, level0RankSize, level1RankSize,
93 : execMem.outputMem.size(), execMem.count);
94 0 : DeviceMem level1OutputMem = execMem.outputMem.range(level1Offset, inputMemSize * level1RankSize);
95 0 : CHK_SMART_PTR_NULL(level1OutputMem);
96 :
97 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
98 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING || (topoAttr_.isDiffDeviceModule && topoAttr_.serverNum == 1)) {
99 : // 1-单server-SDMA
100 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
101 0 : HCCL_INFO("AllGather mesh: using ring algo inter-server.");
102 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
103 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
104 0 : HCCL_INFO("AllGather mesh: using nhr algo inter-server.");
105 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
106 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
107 0 : HCCL_INFO("AllGather mesh: using nhr_v1 algo inter-server.");
108 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
109 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
110 0 : HCCL_INFO("AllGather mesh: using nonuniform-bruck algo inter-server.");
111 : } else {
112 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
113 0 : HCCL_INFO("AllGather mesh: algo is [%s] using halving-doubling algo inter-server.", (HCCL_ALGO_LEVEL1_MAP.at(algType_.algoLevel1)).c_str());
114 : }
115 0 : CHK_SMART_PTR_NULL(level1TempAlg);
116 : // 此处虽然带入inputMem作为scratch mem, 但inputMem 不能被使用
117 0 : CHK_RET(level1TempAlg->Prepare(level1OutputMem, level1OutputMem, execMem.inputMem, inputDataCount,
118 : param.DataDes.dataType, param.stream, HcclReduceOp::HCCL_REDUCE_RESERVED, INVALID_VALUE_RANKID,
119 : std::vector<Slice>(COMM_INDEX_0), level1Offset));
120 :
121 0 : u32 rankSize = level1CommInfo.localRankSize;
122 0 : CHK_RET(level1TempAlg->RegisterProfiler((rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + serverIndex,
123 : PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET, param.stream));
124 :
125 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
126 0 : HCCL_INFO("[CollAllGatherMeshGraphExecutor][KernelRun] AllGather level1 run success");
127 :
128 : // 第3步 机内AllGather
129 0 : std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
130 0 : std::vector<std::vector<Slice>> multiStreamSlice; // 每个stream使用的数据基于用户buffer的偏移
131 0 : u32 sliceNum = level0RankSize;
132 0 : CHK_RET(PrepareAllgatherSlice(sliceNum, inputMemSize * level1RankSize, dataSegsSlice));
133 : // mesh算法stream数量为server内rank数减1
134 0 : CHK_RET(AlgTemplateBase::PrepareSliceMeshStreams(dataSegsSlice, sliceNum - 1, multiStreamSlice));
135 :
136 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg;
137 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910B) {
138 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_MESH_ATOMIC,
139 0 : dispatcher_);
140 : } else {
141 0 : HCCL_ERROR("[CollAllGatherMeshGraphExecutor][KernelRun] current device type [%u] not supported.", topoAttr_.deviceType);
142 0 : return HCCL_E_NOT_SUPPORT;
143 : }
144 0 : CHK_SMART_PTR_NULL(level0TempAlg);
145 0 : CHK_RET(level0TempAlg->Prepare(algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux,
146 : topoAttr_.userRank, nullptr, commIndex, level0RankSize));
147 0 : CHK_RET(level0TempAlg->Prepare(execMem.outputMem, execMem.outputMem, execMem.inputMem,
148 : execMem.count * level0RankSize * level1RankSize, param.DataDes.dataType, param.stream, HCCL_REDUCE_RESERVED,
149 : LEVEL0_BRIDGE_RANK_ID, dataSegsSlice, 0));
150 0 : rankSize = level0RankSize;
151 0 : CHK_RET(level0TempAlg->RegisterProfiler((rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + commIndex,
152 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
153 0 : CHK_RET(RunTemplate(level0TempAlg, level0CommInfo));
154 0 : HCCL_INFO("[CollAllGatherMeshGraphExecutor][KernelRun] AllGather level0 run success");
155 :
156 : // 第4步 内存重排
157 0 : std::vector<bool> reorderedDone(topoAttr_.userRankSize, false);
158 0 : for (u32 scratchSliceIndex = 0; scratchSliceIndex < topoAttr_.userRankSize; scratchSliceIndex++) {
159 0 : if (reorderedDone[scratchSliceIndex]) {
160 0 : continue;
161 : }
162 :
163 : // 当前位置应该放的数据实际所在位置
164 0 : u32 srcServerId = scratchSliceIndex / level0RankSize;
165 0 : u32 srcLocalRank = scratchSliceIndex % level0RankSize;
166 0 : u32 srcSliceRankId = srcLocalRank * level1RankSize + srcServerId;
167 0 : if (srcSliceRankId == scratchSliceIndex) {
168 0 : reorderedDone[scratchSliceIndex] = true;
169 0 : continue;
170 : }
171 :
172 0 : u32 dstSliceRankId = scratchSliceIndex;
173 : // 将当前数据拷贝到input暂存
174 0 : u64 srcOutputOffset = scratchSliceIndex * inputMemSize;
175 0 : u64 dstOutputOffset = dstSliceRankId * inputMemSize;
176 0 : DeviceMem currentDataMem = execMem.outputMem.range(dstOutputOffset, inputMemSize);
177 0 : CHK_SMART_PTR_NULL(currentDataMem);
178 0 : ret = HcclD2DMemcpyAsync(dispatcher_, execMem.inputMem, currentDataMem, const_cast<Stream&>(param.stream));
179 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
180 : HCCL_ERROR("[CollAllGatherMeshGraphExecutor][KernelRun] slice[%u] to input mem temp failed",
181 : scratchSliceIndex), ret);
182 : // 环遍历到当前位置退出循环
183 0 : u32 loopCount = 0;
184 0 : while (srcSliceRankId != scratchSliceIndex) {
185 0 : HCCL_INFO("[CollAllGatherMeshGraphExecutor][KernelRun] slice[%u] to slice[%u]", srcSliceRankId, dstSliceRankId);
186 0 : srcOutputOffset = srcSliceRankId * inputMemSize;
187 0 : dstOutputOffset = dstSliceRankId * inputMemSize;
188 0 : DeviceMem srcOutputMem = execMem.outputMem.range(srcOutputOffset, inputMemSize);
189 0 : CHK_SMART_PTR_NULL(srcOutputMem);
190 0 : DeviceMem dstOutputMem = execMem.outputMem.range(dstOutputOffset, inputMemSize);
191 0 : CHK_SMART_PTR_NULL(dstOutputMem);
192 0 : ret = HcclD2DMemcpyAsync(dispatcher_, dstOutputMem, srcOutputMem, const_cast<Stream&>(param.stream));
193 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
194 : HCCL_ERROR("[CollAllGatherMeshGraphExecutor][KernelRun] slice[%u] to slice[%u] failed",
195 : srcSliceRankId, dstSliceRankId), ret);
196 0 : reorderedDone[dstSliceRankId] = true;
197 0 : dstSliceRankId = srcSliceRankId;
198 0 : srcServerId = srcSliceRankId / level0RankSize;
199 0 : srcLocalRank = srcSliceRankId % level0RankSize;
200 0 : srcSliceRankId = srcLocalRank * level1RankSize + srcServerId;
201 :
202 0 : loopCount++;
203 0 : if (loopCount > topoAttr_.userRankSize) {
204 0 : HCCL_ERROR("[CollAllGatherMeshGraphExecutor][KernelRun] ERROR: loop exceeds user rank size");
205 0 : return HCCL_E_INTERNAL;
206 : }
207 0 : }
208 0 : dstOutputOffset = dstSliceRankId * inputMemSize;
209 0 : DeviceMem ringEndMem = execMem.outputMem.range(dstOutputOffset, inputMemSize);
210 0 : CHK_SMART_PTR_NULL(ringEndMem);
211 0 : ret = HcclD2DMemcpyAsync(dispatcher_, ringEndMem, execMem.inputMem, const_cast<Stream&>(param.stream));
212 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
213 : HCCL_ERROR("[CollAllGatherMeshGraphExecutor][KernelRun] temp input to slice[%u] failed",
214 : dstSliceRankId), ret);
215 0 : reorderedDone[dstSliceRankId] = true;
216 0 : }
217 0 : return HCCL_SUCCESS;
218 0 : }
219 0 : HcclResult CollAllGatherMeshGraphExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
220 : {
221 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
222 0 : return HCCL_E_UNAVAIL;
223 : }
224 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
225 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
226 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
227 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
228 :
229 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
230 0 : return HCCL_E_UNAVAIL;
231 : }
232 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
233 :
234 0 : return HCCL_SUCCESS;
235 0 : }
236 :
237 0 : HcclResult CollAllGatherMeshGraphExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
238 : {
239 0 : if (level1RankSize > 1) {
240 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING || (topoAttr_.isDiffDeviceModule && topoAttr_.serverNum == 1)) {
241 : // 1-单server-SDMA
242 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
243 0 : HCCL_INFO("AllGather mesh: using ring algo inter-server.");
244 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
245 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
246 0 : HCCL_INFO("AllGather mesh: using nhr algo inter-server.");
247 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
248 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
249 0 : HCCL_INFO("AllGather mesh: using nhr_v1 algo inter-server.");
250 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
251 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
252 0 : HCCL_INFO("AllGather mesh: using nonuniform-bruck algo inter-server.");
253 : } else {
254 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
255 0 : HCCL_INFO("AllGather mesh: algo is [%s] using halving-doubling algo inter-server.", (HCCL_ALGO_LEVEL1_MAP.at(algType_.algoLevel1)).c_str());
256 : }
257 0 : CHK_SMART_PTR_NULL(level1TempAlg);
258 0 : return HCCL_SUCCESS;
259 : }
260 0 : return HCCL_E_UNAVAIL;
261 : }
262 : REGISTER_EXEC("AllGatherMeshGraphExecutor", AllGatherMeshGraph, CollAllGatherMeshGraphExecutor);
263 : } // namespace hccl
|
