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_reduce_scatter_mesh_graph_executor.h"
12 :
13 : namespace hccl {
14 :
15 8 : CollReduceScatterMeshGraphExecutor::CollReduceScatterMeshGraphExecutor(const HcclDispatcher dispatcher,
16 8 : std::unique_ptr<TopoMatcher> &topoMatcher)
17 8 : : CollReduceScatterExecutor(dispatcher, topoMatcher)
18 : {
19 8 : DMAReduceFlag_ = false;
20 8 : }
21 :
22 8 : void CollReduceScatterMeshGraphExecutor::ParseParam(const OpParam& param)
23 : {
24 8 : tag_ = param.tag;
25 :
26 : // 910B 图模式非确定计算，inlineReduce使能，MESH拓扑场景下，创建一个mesh平面
27 16 : bool isInlineReduce = IsSupportSDMAReduce(param.inputPtr, param.outputPtr, param.DataDes.dataType,
28 8 : param.reduceType);
29 24 : meshSinglePlane_ = (topoAttr_.deviceType == DevType::DEV_TYPE_910B) &&
30 8 : topoMatcher_->GetExternalInputHcclDeterministic() == DETERMINISTIC_DISABLE &&
31 16 : isInlineReduce && (workflowMode_ != HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
32 :
33 : // 是否需要scratch memory
34 8 : scratchMemFlag_ = true;
35 :
36 : // 记录图模式总数据量
37 8 : totalSize_ = topoAttr_.userRankSize * param.DataDes.count * SIZE_TABLE[param.DataDes.dataType];
38 8 : aicpuUnfoldMode_ = param.aicpuUnfoldMode;
39 8 : }
40 :
41 8 : HcclResult CollReduceScatterMeshGraphExecutor::CalcScratchMemSize(u64& scratchMemSize)
42 : {
43 8 : if (scratchMemFlag_) {
44 8 : scratchMemSize = totalSize_;
45 : } else {
46 0 : scratchMemSize = 0U;
47 : }
48 8 : HCCL_INFO("[CollReduceScatterMeshGraphExecutor][CalcScratchMemSize] tag[%s] scratchMemSize[%llu]",
49 : tag_.c_str(), scratchMemSize);
50 8 : return HCCL_SUCCESS;
51 : }
52 :
53 8 : HcclResult CollReduceScatterMeshGraphExecutor::CalcStreamNum(u32& streamNum)
54 : {
55 8 : u32 totalStreamNum = topoAttr_.deviceNumPerAggregation > 1U ? topoAttr_.deviceNumPerAggregation - 1U : 1U;
56 8 : streamNum = totalStreamNum - 1U;
57 8 : HCCL_INFO("[CollReduceScatterMeshGraphExecutor][CalcStreamNum] tag[%s] streamNum[%u]", tag_.c_str(), streamNum);
58 8 : return HCCL_SUCCESS;
59 : }
60 :
61 8 : HcclResult CollReduceScatterMeshGraphExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
62 : {
63 8 : TransportMemType inputType = TransportMemType::RESERVED;
64 8 : TransportMemType outputType = TransportMemType::RESERVED;
65 8 : CHK_RET(CalcTransportMemType(inputType, outputType));
66 8 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
67 7 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
68 8 : return HCCL_SUCCESS;
69 : }
70 :
71 8 : HcclResult CollReduceScatterMeshGraphExecutor::CalcTransportMemType(TransportMemType &inputType,
72 : TransportMemType &outputType)
73 : {
74 8 : inputType = TransportMemType::SCRATCH;
75 8 : outputType = TransportMemType::PARAM_INPUT;
76 :
77 8 : HCCL_INFO("[CollReduceScatterMeshGraphExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]",
78 : tag_.c_str(), inputType, outputType);
79 8 : return HCCL_SUCCESS;
80 : }
81 :
82 8 : HcclResult CollReduceScatterMeshGraphExecutor::CalcLevel0CommInfo(TransportMemType inputType,
83 : TransportMemType outputType,
84 : std::vector<LevelNSubCommTransport>& opTransport)
85 : {
86 8 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
87 7 : commParaLevel0.meshSinglePlane = meshSinglePlane_;
88 7 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
89 8 : return HCCL_SUCCESS;
90 8 : }
91 :
92 0 : bool CollReduceScatterMeshGraphExecutor::IsHugeData(const u64 curSize, OpParam *param)
93 : {
94 0 : bool hugeData = (curSize * topoAttr_.userRankSize / HCCL_INTERNODE_MAX_DATA_RATE > RDMA_SEND_MAX_SIZE) ||
95 : (curSize > SDMA_SEND_MAX_SIZE);
96 0 : return hugeData;
97 : }
98 :
99 0 : HcclResult CollReduceScatterMeshGraphExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
100 : {
101 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollReduceScatterMeshGraphExecutor][KernelRun] userRank[%u] starts.", topoAttr_.userRank);
102 :
103 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
104 0 : u64 singleRankDataSize = execMem.count * perDataSize;
105 :
106 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
107 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
108 0 : u32 commIndex = level0CommInfo.localRank; // 找到rank所在的节点间平面
109 0 : u32 level0RankSize = level0CommInfo.localRankSize;
110 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
111 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
112 0 : u32 serverIndex = level1CommInfo.localRank;
113 0 : u32 level1RankSize = level1CommInfo.localRankSize;
114 0 : CHK_RET(ActiveSlaveStreams(param.stream));
115 :
116 : /* ******************第1步: input to scratch *******************************/
117 0 : HCCL_INFO("[CollReduceScatterMeshGraphExecutor][KernelRun] userRank[%u], level0RankSize[%u], level1RankSize[%u]", topoAttr_.userRank, level0RankSize, level1RankSize);
118 0 : for (u32 inputSliceId = 0; inputSliceId < topoAttr_.userRankSize; inputSliceId++) {
119 0 : u32 dstServerId = inputSliceId / topoAttr_.deviceNumPerAggregation;
120 0 : u32 dstLocalRank = inputSliceId % topoAttr_.deviceNumPerAggregation;
121 0 : u32 dstSliceId = dstLocalRank * topoAttr_.moduleNum + dstServerId;
122 :
123 0 : u64 srcInputOffset = inputSliceId * singleRankDataSize;
124 0 : u64 dstScratchOffset = dstSliceId * singleRankDataSize;
125 :
126 0 : DeviceMem srcInputMem = execMem.inputMem.range(srcInputOffset, singleRankDataSize);
127 0 : CHK_SMART_PTR_NULL(srcInputMem);
128 0 : DeviceMem dstScratchMem = execMem.scratchMem.range(dstScratchOffset, singleRankDataSize);
129 0 : CHK_SMART_PTR_NULL(dstScratchMem);
130 :
131 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, dstScratchMem, srcInputMem, const_cast<Stream&>(param.stream));
132 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
133 : HCCL_ERROR("[CollReduceScatterMeshGraphExecutor][KernelRun] rank[%u] slice[%u] to slice[%u] failed",
134 : topoAttr_.userRank, inputSliceId, dstSliceId), ret);
135 0 : }
136 :
137 : /* ******************第2步: intranode *******************************/
138 0 : u32 sliceNum = level0CommInfo.localRankSize;
139 : // 根据数据量算每个环上数据的偏移和大小，把做完hd的slice均分成RankSize份
140 0 : std::vector<Slice> dataSegsSlice;
141 0 : u32 level0ReduceCount = execMem.count * level1RankSize;
142 0 : CHK_RET(PrepareReduceScatterSliceData(level0ReduceCount, perDataSize, sliceNum, dataSegsSlice));
143 :
144 : // 每个server分配的slice大小
145 0 : u64 serverSliceSize = execMem.inputMem.size();
146 : // 每个服务器对应的偏移
147 0 : u64 serverSliceOffset = 0;
148 :
149 0 : HCCL_DEBUG("inputMem.size=%llu, level0CommInfo.localRankSize=%u, serverSliceSize=%llu, serverSliceOffset=%llu "\
150 : "commIndex=%u level1CommInfo.localRank=%u", execMem.inputMem.size(), level0CommInfo.localRankSize,
151 : serverSliceSize, serverSliceOffset, commIndex, level1CommInfo.localRank);
152 :
153 0 : DeviceMem reduceScatterMeshInput = execMem.scratchMem.range(serverSliceOffset, serverSliceSize);
154 0 : CHK_SMART_PTR_NULL(reduceScatterMeshInput);
155 0 : DeviceMem reduceScatterMeshOutput = execMem.inputMem.range(serverSliceOffset, serverSliceSize);
156 0 : CHK_SMART_PTR_NULL(reduceScatterMeshOutput);
157 :
158 0 : HcomCollOpInfo *opInfoPtr = nullptr;
159 :
160 0 : if (topoMatcher_->GetExternalInputHcclDeterministic() == DETERMINISTIC_DISABLE &&
161 0 : (param.DataDes.dataType != HCCL_DATA_TYPE_INT64) &&
162 0 : (topoAttr_.deviceType == DevType::DEV_TYPE_910B && param.reduceType != HCCL_REDUCE_PROD)) {
163 0 : CHK_RET(MultiStreamReduceScatterMeshAtomic(param.tag, reduceScatterMeshInput, reduceScatterMeshOutput, // 非确定性
164 : level0ReduceCount, param.DataDes.dataType, param.reduceType, dataSegsSlice, const_cast<Stream&>(param.stream),
165 : COMM_LEVEL0, serverSliceOffset, opInfoPtr));
166 : } else {
167 0 : std::vector<std::vector<Slice> > multiStreamSlice; // 每个stream使用的数据基于用户buffer的偏移
168 : // mesh算法stream数量为rank数减1
169 0 : CHK_RET(AlgTemplateBase::PrepareSliceMeshStreams(dataSegsSlice, sliceNum - 1, multiStreamSlice));
170 0 : CHK_RET(MultiStreamReduceScatterMesh(param.tag, reduceScatterMeshInput, reduceScatterMeshOutput, // 确定性
171 : level0ReduceCount, param.DataDes.dataType, param.reduceType, multiStreamSlice,
172 : const_cast<Stream&>(param.stream), COMM_LEVEL0, serverSliceOffset));
173 0 : }
174 :
175 : /* ******************第3步: internode *******************************/
176 :
177 0 : if (level1RankSize > 1) {
178 0 : u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
179 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
180 0 : u64 ringCount = execMem.count;
181 0 : u64 level1SliceSize = execMem.inputMem.size() / level0RankSize;
182 : // 每个服务器对应的偏移
183 0 : u64 level1SliceOffset = commIndex * level1SliceSize;
184 0 : DeviceMem level1ReduceScatterInput = execMem.scratchMem.range(level1SliceOffset, level1SliceSize);
185 0 : CHK_SMART_PTR_NULL(level1ReduceScatterInput);
186 0 : DeviceMem level1ReduceScatterScratch = execMem.inputMem.range(level1SliceOffset, level1SliceSize);
187 0 : CHK_SMART_PTR_NULL(level1ReduceScatterScratch);
188 0 : HCCL_INFO("[CollReduceScatterMeshGraphExecutor][KernelRun] rank[%u] level 1 sliceSize[%llu] sliceOffset[%llu]", topoAttr_.userRank, level1SliceSize, level1SliceOffset);
189 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
190 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_RING, dispatcher_);
191 0 : CHK_SMART_PTR_NULL(level1TempAlg);
192 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
193 0 : HCCL_INFO("ReduceScatter mesh: using ring algo inter-server.");
194 0 : CHK_RET(level1TempAlg->Prepare(level1ReduceScatterInput, level1ReduceScatterInput, level1ReduceScatterScratch, ringCount,
195 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), level1SliceOffset));
196 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
197 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_NHR, dispatcher_);
198 0 : HCCL_INFO("ReduceScatter mesh: using nhr algo inter-server.");
199 0 : CHK_SMART_PTR_NULL(level1TempAlg);
200 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr, false));
201 0 : CHK_RET(level1TempAlg->Prepare(level1ReduceScatterInput, level1ReduceScatterInput, level1ReduceScatterScratch, ringCount,
202 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), level1SliceOffset));
203 0 : level1TempAlg->CloseBarrier();
204 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
205 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_NHR_V1, dispatcher_);
206 0 : HCCL_INFO("ReduceScatter mesh: using nhr_v1 algo inter-server.");
207 0 : CHK_SMART_PTR_NULL(level1TempAlg);
208 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
209 0 : CHK_RET(level1TempAlg->Prepare(level1ReduceScatterInput, level1ReduceScatterInput, level1ReduceScatterScratch, ringCount,
210 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), level1SliceOffset));
211 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
212 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_NB, dispatcher_);
213 0 : HCCL_INFO("ReduceScatter mesh: using nonuniform-bruck algo inter-server.");
214 0 : CHK_SMART_PTR_NULL(level1TempAlg);
215 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
216 0 : CHK_RET(level1TempAlg->Prepare(level1ReduceScatterInput, level1ReduceScatterInput, level1ReduceScatterScratch, ringCount,
217 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), level1SliceOffset));
218 : } else {
219 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_RECURSIVE_HD, dispatcher_);
220 0 : CHK_SMART_PTR_NULL(level1TempAlg);
221 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
222 0 : HCCL_INFO("ReduceScatter mesh: algo is [%s] using halving-doubling algo inter-server.", (HCCL_ALGO_LEVEL1_MAP.at(algType_.algoLevel1)).c_str());
223 0 : u64 inputDataCount = level1SliceSize / perDataSize; // count是output的数据个数
224 0 : CHK_RET(level1TempAlg->Prepare(level1ReduceScatterInput, level1ReduceScatterInput, level1ReduceScatterScratch, inputDataCount,
225 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), level1SliceOffset));
226 : }
227 0 : CHK_RET(level1TempAlg->RegisterProfiler(
228 : (level1RankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1CommInfo.localRank,
229 : PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET, param.stream));
230 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
231 0 : }
232 :
233 : /* *******************第4步: 节点内reducescatter ******************************************/
234 0 : u32 transposeRankId = commIndex * level1RankSize + serverIndex;
235 0 : u64 finalOffset = transposeRankId * singleRankDataSize;
236 0 : DeviceMem srcScratchMem = execMem.scratchMem.range(finalOffset, singleRankDataSize);
237 0 : CHK_SMART_PTR_NULL(srcScratchMem);
238 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, execMem.outputMem, srcScratchMem, const_cast<Stream&>(param.stream));
239 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
240 : HCCL_ERROR("[CollReduceScatterMeshGraphExecutor][KernelRun] rank[%u] memcpy failed, offset[%llu], size[%llu]",
241 : topoAttr_.userRank, finalOffset, singleRankDataSize), ret);
242 :
243 0 : return HCCL_SUCCESS;
244 0 : }
245 :
246 : REGISTER_EXEC("ReduceScatterMeshGraphExecutor", ReduceScatterMeshGraph, CollReduceScatterMeshGraphExecutor);
247 : }
|
