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_opbase_mid_count_deterministic_executor.h"
12 :
13 : namespace hccl {
14 0 : CollAllReduceMeshOpbaseMidCountDeterministicExecutor::CollAllReduceMeshOpbaseMidCountDeterministicExecutor(
15 0 : const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher)
16 0 : : CollAllReduceExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = true;
19 0 : }
20 :
21 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::CalcStreamNum(u32& streamNum)
22 : {
23 0 : const u32 level0AlltoallStreamNum = topoAttr_.deviceNumPerAggregation - 1;
24 0 : const u32 level0LocalReduceStreamNum = 1 << static_cast<int>(std::floor(log2(topoAttr_.deviceNumPerAggregation)));
25 0 : streamNum = level0AlltoallStreamNum + level0LocalReduceStreamNum;
26 :
27 0 : HCCL_INFO("[%s]tag[%s] level0AlltoallStreamNum[%u], level0LocalReduceStreamNum[%u], streamNum[%u]", __func__,
28 : tag_.c_str(), level0AlltoallStreamNum, level0LocalReduceStreamNum, streamNum);
29 0 : return HCCL_SUCCESS;
30 : }
31 :
32 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
33 : {
34 0 : TransportMemType inputType = TransportMemType::RESERVED;
35 0 : TransportMemType outputType = TransportMemType::RESERVED;
36 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
37 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
38 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
39 0 : return HCCL_SUCCESS;
40 : }
41 :
42 :
43 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::CalcTransportMemType(TransportMemType &inputType,
44 : TransportMemType &outputType)
45 : {
46 0 : inputType = TransportMemType::CCL_INPUT;
47 0 : outputType = TransportMemType::CCL_OUTPUT;
48 0 : HCCL_INFO("[CollAllReduceMeshOpbaseMidCountDeterministicExecutor][CalcTransportMemType]" \
49 : "tag[%s] inputType[%d], outputType[%d]",
50 : tag_.c_str(), inputType, outputType);
51 0 : return HCCL_SUCCESS;
52 : }
53 :
54 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::CalcLevel0CommInfo(TransportMemType inputType,
55 : TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
56 : {
57 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
58 0 : commParaLevel0.meshSinglePlane = true;
59 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
60 0 : return HCCL_SUCCESS;
61 0 : }
62 :
63 0 : bool CollAllReduceMeshOpbaseMidCountDeterministicExecutor::IsHugeData(const u64 curSize)
64 : {
65 0 : bool hugeData = curSize / topoAttr_.deviceNumPerAggregation / HCCL_INTERNODE_MAX_DATA_RATE > RDMA_SEND_MAX_SIZE ||
66 : curSize > SDMA_SEND_MAX_SIZE;
67 0 : HCCL_DEBUG("[%s]isHugeData[%d], curSize[%llu], topoAttr_.deviceNumPerAggregation[%u]",
68 : __func__, hugeData, curSize, topoAttr_.deviceNumPerAggregation);
69 0 : return hugeData;
70 : }
71 :
72 0 : bool CollAllReduceMeshOpbaseMidCountDeterministicExecutor::IsSmallData(const u64 totalSize, const u64 curSize)
73 : {
74 0 : bool smallData = IsAllReduceSmallData(curSize);
75 0 : return smallData;
76 : }
77 :
78 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::PrepareSlicesInfo(const OpParam ¶m,
79 : const ExecMem &execMem, std::vector<Slice>& dataSegsSlice, GroupSlicesInfo& groupSlicesInfo, const u32 sliceSize)
80 : {
81 0 : const u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
82 0 : MemBlockInfo memInfo;
83 0 : u64 sizePerBlock = (execMem.count + sliceSize - 1) / sliceSize * perDataSize;
84 0 : sizePerBlock = AlgTemplateBase::RoundUpWithDivisor(sizePerBlock, HCCL_MIN_SLICE_ALIGN);
85 0 : const u64 totalSize = execMem.count * perDataSize;
86 0 : u64 sizeRemain = totalSize;
87 0 : for (u32 dataId = 0; dataId < sliceSize; dataId ++) {
88 0 : u64 size = (sizeRemain > sizePerBlock) ? sizePerBlock : sizeRemain;
89 0 : u64 offset = totalSize - sizeRemain;
90 0 : memInfo.size.push_back(size);
91 0 : memInfo.userInputOffsets.push_back(offset);
92 0 : memInfo.inputOffsets.push_back(offset);
93 0 : memInfo.outputOffsets.push_back(offset);
94 0 : Slice slice{offset, size};
95 0 : dataSegsSlice.emplace_back(std::move(slice));
96 0 : sizeRemain -= size;
97 : }
98 0 : groupSlicesInfo.push_back(memInfo);
99 0 : return HCCL_SUCCESS;
100 0 : }
101 :
102 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::RunReduceScatterLevel0(const OpParam ¶m,
103 : ExecMem &execMem, GroupSlicesInfo& groupSlicesInfo)
104 : {
105 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
106 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg;
107 0 : const u32 all2allOffset = 0;
108 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
109 0 : TemplateType::TEMPLATE_REDUCESCATTER_PLANT_LOCAL_REDUCE, dispatcher_);
110 0 : CHK_SMART_PTR_NULL(level0TempAlg);
111 :
112 0 : CHK_RET(level0TempAlg->Prepare(execMem.inputPtr, execMem.inputMem, execMem.outputMem, param.stream,
113 : algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux, groupSlicesInfo,
114 : param.reduceType, all2allOffset, param.DataDes.dataType, true, true));
115 :
116 0 : CHK_RET(level0TempAlg->RegisterProfiler(
117 : (level0CommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
118 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
119 0 : CHK_RET(RunTemplate(level0TempAlg, level0CommInfo));
120 0 : return HCCL_SUCCESS;
121 0 : }
122 :
123 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::RunAllReduceLevel1(const OpParam ¶m,
124 : ExecMem &execMem, const std::vector<Slice>& dataSegsSlice)
125 : {
126 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
127 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
128 0 : const u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
129 0 : const u32 commIndex = level0CommInfo.localRank;
130 :
131 0 : CHK_PRT_RET(commIndex >= dataSegsSlice.size(),
132 : HCCL_ERROR("[%s]commIndex[%u] >= dataSegsSlice size[%zu]", __func__, commIndex,
133 : dataSegsSlice.size()), HCCL_E_INTERNAL);
134 :
135 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
136 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
137 :
138 0 : DeviceMem allreduceInput = execMem.inputMem.range(dataSegsSlice[commIndex].offset, dataSegsSlice[commIndex].size);
139 0 : CHK_SMART_PTR_NULL(allreduceInput);
140 0 : DeviceMem allreduceOutput = execMem.outputMem.range(dataSegsSlice[commIndex].offset, dataSegsSlice[commIndex].size);
141 0 : CHK_SMART_PTR_NULL(allreduceOutput);
142 :
143 0 : const u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
144 :
145 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
146 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_RING,
147 0 : dispatcher_);
148 0 : HCCL_INFO("[%s]: using ring algo inter-server.", __func__);
149 0 : CHK_SMART_PTR_NULL(level1TempAlg);
150 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
151 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
152 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
153 0 : TemplateType::TEMPLATE_ALL_REDUCE_NHR, dispatcher_);
154 0 : HCCL_INFO("[%s]: using nhr algo inter-server.", __func__);
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(
160 0 : TemplateType::TEMPLATE_ALL_REDUCE_NHR_V1, dispatcher_);
161 0 : HCCL_INFO("[%s]: using nhr_v1 algo inter-server.", __func__);
162 0 : CHK_SMART_PTR_NULL(level1TempAlg);
163 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
164 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
165 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
166 0 : TemplateType::TEMPLATE_ALL_REDUCE_NB, dispatcher_);
167 0 : HCCL_INFO("[%s]: using nb algo inter-server.", __func__);
168 0 : CHK_SMART_PTR_NULL(level1TempAlg);
169 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
170 : } else {
171 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
172 0 : TemplateType::TEMPLATE_ALL_REDUCE_RECURSIVE_HALVING_DOUBLING, dispatcher_);
173 0 : HCCL_INFO("[%s]: using Recursive halving-doubling algo inter-server.", __func__);
174 0 : CHK_SMART_PTR_NULL(level1TempAlg);
175 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
176 : }
177 0 : CHK_SMART_PTR_NULL(level1TempAlg);
178 :
179 0 : const u32 rankSize = level1CommInfo.localRankSize;
180 :
181 0 : const 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 0 : return HCCL_SUCCESS;
191 0 : }
192 :
193 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::RunAllGatherLevel0(const OpParam ¶m,
194 : ExecMem &execMem, const std::vector<Slice>& dataSegsSlice)
195 : {
196 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
197 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg;
198 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_MESH_ATOMIC,
199 0 : dispatcher_);
200 :
201 0 : CHK_SMART_PTR_NULL(level0TempAlg);
202 0 : CHK_RET(level0TempAlg->Prepare(algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux,
203 : topoAttr_.userRank, nullptr, level0CommInfo.localRank, level0CommInfo.localRankSize));
204 :
205 0 : u32 rankSize = level0CommInfo.localRankSize;
206 0 : CHK_RET(level0TempAlg->Prepare(execMem.outputMem, execMem.outputMem, execMem.inputMem, execMem.count,
207 : param.DataDes.dataType, param.stream, param.reduceType,
208 : LEVEL0_BRIDGE_RANK_ID, dataSegsSlice, 0));
209 :
210 0 : CHK_RET(level0TempAlg->RegisterProfiler(
211 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
212 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
213 :
214 0 : CHK_RET(RunTemplate(level0TempAlg, level0CommInfo));
215 0 : return HCCL_SUCCESS;
216 0 : }
217 :
218 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
219 : {
220 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s]userRank[%u] starts.", __func__, topoAttr_.userRank);
221 :
222 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, 1));
223 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
224 0 : const u32 sliceNum = level0CommInfo.localRankSize;
225 :
226 0 : std::vector<Slice> dataSegsSlice;
227 0 : GroupSlicesInfo groupSlicesInfoLevel0;
228 0 : CHK_RET(PrepareSlicesInfo(param, execMem, dataSegsSlice, groupSlicesInfoLevel0, sliceNum));
229 0 : CHK_RET(ActiveSlaveStreams(param.stream));
230 :
231 : /* STAGE 0: level 0 reduce scatter - plant local reduce */
232 0 : CHK_RET(RunReduceScatterLevel0(param, execMem, groupSlicesInfoLevel0));
233 0 : HCCL_INFO("[%s]AllReduce stage0 run success.", __func__);
234 :
235 : /* STAGE 1: level1 all_reduce - auto selected */
236 0 : CHK_RET(RunAllReduceLevel1(param, execMem, dataSegsSlice));
237 0 : HCCL_INFO("[%s]AllReduce stage1 run success.", __func__);
238 :
239 : /* STAGE 2: level0 all_gather - mesh atomic */
240 0 : CHK_RET(RunAllGatherLevel0(param, execMem, dataSegsSlice));
241 0 : HCCL_INFO("[%s]AllReduce stage2 run success", __func__);
242 :
243 0 : const u64 curSize = execMem.count * SIZE_TABLE[param.DataDes.dataType];
244 0 : DeviceMem outCommMem = execMem.outputMem.range(0, curSize);
245 0 : DeviceMem outMem(execMem.outputPtr, curSize);
246 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, outMem, outCommMem, const_cast<Stream &>(param.stream)));
247 :
248 0 : return HCCL_SUCCESS;
249 0 : }
250 :
251 0 : HcclResult CollAllReduceMeshOpbaseMidCountDeterministicExecutor::RunLoopInner(OpParam ¶m,
252 : const ReduceType &reduceType, ExecMem &execMem)
253 : {
254 0 : const u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
255 0 : const u64 curSize = execMem.count * unitSize;
256 0 : HCCL_DEBUG("[%s]inputMem[%p][%llu], outputMem[%p][%llu], " \
257 : "intputPtr[%p], outputPtr[%p], curCount[%llu], curSize[%llu]",
258 : __func__, execMem.inputMem.ptr(), execMem.inputMem.size(), execMem.outputMem.ptr(), execMem.outputMem.size(),
259 : execMem.inputPtr, execMem.outputPtr, execMem.count, curSize);
260 0 : CHK_PRT_RET((execMem.count == 0),
261 : HCCL_ERROR("[%s]In OP_BASE curCount is zero.", __func__), HCCL_E_PARA);
262 :
263 : /* init task */
264 0 : const auto autoSelectedAlgTypeLevel1 = static_cast<u32>(algType_.algoLevel1);
265 0 : const bool hugeData = IsHugeData(curSize);
266 0 : const bool smallData = IsSmallData(param.DataDes.count * unitSize, curSize);
267 0 : u64 sliceNum = 0;
268 0 : CHK_RET(GetSliceNum(execMem.count * unitSize, smallData, sliceNum, unitSize));
269 0 : const bool dataSplit = true;
270 0 : const u8 deterministic = topoMatcher_->GetExternalInputHcclDeterministic();
271 0 : const CopyPattern copy = CopyPattern::ZCOPY;
272 0 : const auto opMeta = HcclOpMetaInfo::GetOneForAllReduce(autoSelectedAlgTypeLevel1,
273 : param.DataDes.dataType, reduceType, smallData, 1, hugeData, copy, sliceNum,
274 : false, true, dataSplit, deterministic);
275 0 : CHK_RET(InitTask(dispatcher_, param.stream, opMeta.isEnableCache, opMeta.GetCacheKey()));
276 :
277 : /* kernel run */
278 0 : CHK_RET(KernelRun(param, execMem));
279 0 : CHK_RET(LaunchTaskExtend(dispatcher_,
280 : const_cast<Stream &>(param.stream),
281 : const_cast<std::vector<Stream> &>(algResResp_->slaveStreams)));
282 :
283 0 : return HCCL_SUCCESS;
284 : }
285 :
286 : REGISTER_EXEC("AllReduceMeshOpbaseMidCountDeterministicExecutor",
287 : AllReduceMeshOpbaseMidCountDeterministic, CollAllReduceMeshOpbaseMidCountDeterministicExecutor);
288 :
289 : } // namespace hccl
|
