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_ring_executor.h"
12 :
13 : namespace hccl {
14 :
15 0 : CollAllReduceRingExecutor::CollAllReduceRingExecutor(const HcclDispatcher dispatcher,
16 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
17 0 : : CollAllReduceExecutor(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 CollAllReduceRingExecutor::CalcStreamNum(u32& streamNum)
28 : {
29 0 : u32 totalStreamNum = 1U;
30 0 : if (algType_.algoLevel0 == AlgTypeLevel0::ALG_LEVEL0_8P_RING) {
31 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_8PRING;
32 0 : } else if (algType_.algoLevel0 == AlgTypeLevel0::ALG_LEVEL0_NP_SINGLE_RING) {
33 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93) {
34 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
35 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_NPRING_SINGLE * STREAM_NUM_FOR_DMAREDUCE_ONE_RING;
36 : } else {
37 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
38 : }
39 : }
40 : }
41 0 : streamNum = totalStreamNum - 1;
42 0 : HCCL_INFO("[CollAllReduceRingExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
43 : tag_.c_str(), streamNum);
44 0 : return HCCL_SUCCESS;
45 : }
46 :
47 0 : HcclResult CollAllReduceRingExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
48 : {
49 0 : TransportMemType inputType = TransportMemType::RESERVED;
50 0 : TransportMemType outputType = TransportMemType::RESERVED;
51 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
52 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
53 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
54 0 : return HCCL_SUCCESS;
55 : }
56 :
57 0 : HcclResult CollAllReduceRingExecutor::CalcTransportMemType(TransportMemType &inputType, TransportMemType &outputType)
58 : {
59 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
60 0 : inputType = TransportMemType::CCL_INPUT;
61 0 : outputType = TransportMemType::CCL_OUTPUT;
62 : } else {
63 0 : inputType = TransportMemType::PARAM_INPUT;
64 0 : outputType = TransportMemType::PARAM_OUTPUT;
65 : }
66 0 : HCCL_INFO("[CollAllReduceRingExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d].",
67 : tag_.c_str(), inputType, outputType);
68 0 : return HCCL_SUCCESS;
69 : }
70 :
71 0 : HcclResult CollAllReduceRingExecutor::CalcLevel0CommInfo(TransportMemType inputType,
72 : TransportMemType outputType,
73 : std::vector<LevelNSubCommTransport>& opTransport)
74 : {
75 0 : HCCL_INFO("[CollAllReduceRingExecutor][CalcLevel0CommInfo]tag[%s] start.", tag_.c_str());
76 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_RING_INNER);
77 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
78 0 : HCCL_INFO("[CollAllReduceRingExecutor][CalcLevel0CommInfo]tag[%s] Calc RingComm finish.", tag_.c_str());
79 0 : return HCCL_SUCCESS;
80 0 : }
81 :
82 0 : bool CollAllReduceRingExecutor::IsHugeData(const u64 curSize)
83 : {
84 0 : bool hugeData = curSize / topoAttr_.deviceNumPerAggregation / HCCL_INTERNODE_MAX_DATA_RATE > RDMA_SEND_MAX_SIZE ||
85 : curSize > SDMA_SEND_MAX_SIZE;
86 0 : return hugeData;
87 : }
88 :
89 0 : bool CollAllReduceRingExecutor::IsSmallData(const u64 totalSize, const u64 curSize)
90 : {
91 0 : bool smallData = DMAReduceFlag_ ? false : IsAllReduceSmallData(curSize);
92 0 : return smallData;
93 : }
94 :
95 0 : HcclResult CollAllReduceRingExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
96 : {
97 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollAllReduceRingExecutor][Run]The CollAllReduceRingExecutor starts.");
98 0 : u32 perDataSize = 0;
99 0 : CHK_RET(SalGetDataTypeSize(param.DataDes.dataType, perDataSize));
100 :
101 0 : std::vector<std::vector<Slice> > multRingsSliceZero; // 数据基于该rank上环0的偏移
102 0 : std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
103 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
104 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
105 :
106 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, ringNum));
107 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
108 0 : u32 sliceNum = level0CommInfo.localRankSize;
109 : // 根据数据量计算每个环上数据的偏移和大小
110 0 : CHK_RET(AlgTemplateBase::PrepareSliceData(execMem.count, perDataSize, sliceNum, 0, dataSegsSlice));
111 :
112 : /* 三步算法step1：外层 - 节点内 reduce-scatter */
113 : /* 外层:reduce_scatter */
114 0 : if (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) {
115 : // 构造ring algorithm对应的reduce-scatter实例
116 0 : multRingsSliceZero = PrepareMultiRingSlice(dataSegsSlice, param.tag, false, topoAttr_.nicList);
117 0 : CHK_PRT_RET(multRingsSliceZero.size() != ringNum, HCCL_ERROR("[CollAllReduceRingExecutor]"\
118 : "ringNum[%u] !=multRingsSliceZero size[%zu]", ringNum, multRingsSliceZero.size()), HCCL_E_INTERNAL);
119 : } else {
120 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
121 : }
122 :
123 0 : HcomCollOpInfo *reduceScatterOpInfoPtr = nullptr;
124 : // 第一步的reducescatter输出放在CCL buffer上，通过设置nullptr指示不做最后一步的DMA削减动作
125 0 : HcomCollOpInfo reduceScatterOpInfo = {
126 0 : "", execMem.inputPtr, nullptr, execMem.count, param.DataDes.dataType, param.root, param.reduceType
127 0 : };
128 0 : if (DMAReduceFlag_) {
129 0 : reduceScatterOpInfoPtr = &reduceScatterOpInfo;
130 : }
131 0 : CHK_RET(MultiRingReduceScatter(param.tag, execMem.inputMem, execMem.outputMem, execMem.count,
132 : param.DataDes.dataType, param.reduceType, multRingsSliceZero, param.stream,
133 : PROF_STAGE_0, 0, reduceScatterOpInfoPtr));
134 :
135 0 : HCCL_INFO("AllReduce ringhd stage0 run success");
136 :
137 : /* 三步算法step2: 内层 - 节点间 allreduce */
138 : u64 hdSize;
139 : u32 segmentIdx;
140 : u32 commIndex;
141 0 : CHK_RET(PrepareLevel1CommInfo(segmentIdx, commIndex, hdSize, level0CommInfo, multRingsSliceZero, param.tag));
142 :
143 0 : u64 hdCount = hdSize / perDataSize;
144 0 : auto nicList = topoAttr_.nicList;
145 0 : bool isMultiNic = topoType_ == TopoType::TOPO_TYPE_8P_RING && nicList.size() != DEVICE_EIGHT;
146 0 : std::vector<u32>::iterator iterNic = std::find(nicList.begin(), nicList.end(), topoAttr_.devicePhyId);
147 0 : bool innRunRet = isMultiNic && (iterNic == nicList.end());
148 0 : HCCL_DEBUG("[CollAllReduceRingExecutor][KernelRun]isMultiNic is %d, innRunRet is %d", isMultiNic, innRunRet);
149 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
150 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
151 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
152 :
153 0 : DeviceMem allreduceInput = execMem.inputMem.range(dataSegsSlice[segmentIdx].offset, hdSize);
154 0 : CHK_SMART_PTR_NULL(allreduceInput);
155 0 : DeviceMem allreduceOutput = execMem.outputMem.range(dataSegsSlice[segmentIdx].offset, hdSize);
156 0 : CHK_SMART_PTR_NULL(allreduceOutput);
157 :
158 0 : u64 reduceAttr = GetReduceAttr(allreduceInput, allreduceOutput, param.DataDes.dataType, param.reduceType);
159 :
160 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
161 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
162 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RING,
163 0 : dispatcher_);
164 0 : HCCL_INFO("AllReduce ring: using ring algo inter-server.");
165 0 : CHK_SMART_PTR_NULL(level1TempAlg);
166 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
167 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
168 0 : u64 curSize = execMem.count * SIZE_TABLE[param.DataDes.dataType]; // 单位 byte
169 0 : HCCL_DEBUG("AllReduce ring: curSize[%llu] deviceNumPerAggregation[%u] commLevel0Size[%u]",
170 : curSize, topoAttr_.deviceNumPerAggregation, level0CommInfo.localRankSize);
171 0 : if (curSize / topoAttr_.deviceNumPerAggregation <= NHR_ALLREDUCE_SMALL_SIZE) {
172 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR_ONESHOT, dispatcher_);
173 : } else {
174 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR, dispatcher_);
175 : }
176 0 : HCCL_INFO("AllReduce ring: using nhr algo inter-server.");
177 0 : CHK_SMART_PTR_NULL(level1TempAlg);
178 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
179 0 : level1TempAlg->CloseBarrier();
180 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
181 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR_V1, dispatcher_);
182 0 : HCCL_INFO("AllReduce ring: using nhr_v1 algo inter-server.");
183 0 : CHK_SMART_PTR_NULL(level1TempAlg);
184 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
185 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
186 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NB,
187 0 : dispatcher_);
188 0 : HCCL_INFO("AllReduce ring: using nonuniform-bruck algo inter-server.");
189 0 : CHK_SMART_PTR_NULL(level1TempAlg);
190 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
191 : } else {
192 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RECURSIVE_HALVING_DOUBLING, dispatcher_);
193 0 : HCCL_INFO("AllReduce ring: using Recursive halving-doubling algo inter-server.");
194 0 : CHK_SMART_PTR_NULL(level1TempAlg);
195 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
196 : }
197 :
198 0 : CHK_SMART_PTR_NULL(level1TempAlg);
199 :
200 : // 节点间的hd 使用环0来记录
201 0 : CHK_RET(level1TempAlg->Prepare(
202 : allreduceInput, allreduceOutput, allreduceOutput, hdCount,
203 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID,
204 : std::vector<Slice>(0), dataSegsSlice[segmentIdx].offset));
205 :
206 0 : CHK_RET(level1TempAlg->RegisterProfiler(
207 : (level1CommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1CommInfo.localRank,
208 : PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET, param.stream));
209 :
210 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
211 0 : }
212 0 : HCCL_INFO("AllReduce ringhd stage1 run success");
213 :
214 : /* 三步算法step3：外层 - 节点内 allgather */
215 0 : HcomCollOpInfo *allgatherOpInfoPtr = nullptr;
216 : // 第三步的allgather输入放在CCL buffer上，通过设置nullptr指示要从CCL buffer获取输入
217 0 : HcomCollOpInfo allgatherOpInfo = {
218 0 : "", nullptr, execMem.outputPtr, execMem.count, param.DataDes.dataType, param.root, param.reduceType
219 0 : };
220 0 : if (DMAReduceFlag_) {
221 0 : allgatherOpInfoPtr = &allgatherOpInfo;
222 : }
223 0 : CHK_RET(MultiRingAllGather(param.tag, execMem.inputMem, execMem.outputMem, hdCount, param.DataDes.dataType,
224 : multRingsSliceZero, param.stream, PROF_STAGE_2, 0, allgatherOpInfoPtr));
225 0 : HCCL_INFO("AllReduce ringhd stage2 run success");
226 0 : return HCCL_SUCCESS;
227 0 : }
228 0 : HcclResult CollAllReduceRingExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
229 : {
230 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
231 0 : return HCCL_E_UNAVAIL;
232 : }
233 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
234 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
235 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
236 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
237 :
238 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
239 0 : return HCCL_E_UNAVAIL;
240 : }
241 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
242 :
243 0 : return HCCL_SUCCESS;
244 0 : }
245 :
246 0 : HcclResult CollAllReduceRingExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
247 : {
248 0 : if (level1RankSize > 1) {
249 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
250 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RING,
251 0 : dispatcher_);
252 0 : HCCL_INFO("AllReduce ring: using ring algo inter-server.");
253 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
254 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR, dispatcher_);
255 0 : HCCL_INFO("AllReduce ring: using nhr algo inter-server.");
256 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
257 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NHR_V1, dispatcher_);
258 0 : HCCL_INFO("AllReduce ring: using nhr_v1 algo inter-server.");
259 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_AHC) {
260 : // 获取通信域分组信息
261 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_AHC, dispatcher_);
262 0 : HCCL_INFO("AllReduce ring: using ahc algo inter-server.");
263 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_AHC_BROKE) {
264 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_AHC_BROKE, dispatcher_);
265 0 : HCCL_INFO("AllReduce ring: using ahc-broke algo inter-server.");
266 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
267 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_NB,
268 0 : dispatcher_);
269 0 : HCCL_INFO("AllReduce ring: using nonuniform-bruck algo inter-server.");
270 : } else {
271 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RECURSIVE_HALVING_DOUBLING, dispatcher_);
272 0 : HCCL_INFO("AllReduce ring: using Recursive halving-doubling algo inter-server.");
273 : }
274 0 : CHK_SMART_PTR_NULL(level1TempAlg);
275 0 : return HCCL_SUCCESS;
276 : }
277 0 : return HCCL_E_UNAVAIL;
278 : }
279 : REGISTER_EXEC("AllReduceRingExecutor", AllReduceRing, CollAllReduceRingExecutor);
280 :
281 : } // namespace hccl
|
