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_scatter_ring_for_910_93_executor.h"
12 :
13 : namespace hccl {
14 :
15 0 : CollScatterRingFor91093Executor::CollScatterRingFor91093Executor(
16 0 : const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher>& topoMatcher)
17 0 : : CollScatterExecutor(dispatcher, topoMatcher)
18 : {
19 0 : DMAReduceFlag_ = workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE;
20 0 : }
21 :
22 0 : HcclResult CollScatterRingFor91093Executor::CalcStreamNum(u32& streamNum)
23 : {
24 0 : u32 totalStreamNum
25 0 : = (topoType_ == TopoType::TOPO_TYPE_NP_DOUBLE_RING ? LEVEL0_PLANE_NUM_IN_NPRING_DOUBLE :
26 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE);
27 : // scatter在910_93场景仅支持单算子模式，已有mainstream需要-1
28 0 : streamNum = totalStreamNum - 1;
29 0 : HCCL_INFO("[CollScatterRingFor91093Executor][CalcStreamNum] tag[%s] streamNum[%u]", tag_.c_str(), streamNum);
30 0 : return HCCL_SUCCESS;
31 : }
32 :
33 0 : HcclResult CollScatterRingFor91093Executor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
34 : {
35 0 : TransportMemType inputType = TransportMemType::RESERVED;
36 0 : TransportMemType outputType = TransportMemType::RESERVED;
37 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
38 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
39 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
40 0 : CHK_RET(CalcLevel2CommInfo(inputType, outputType, opTransport));
41 0 : return HCCL_SUCCESS;
42 : }
43 :
44 0 : HcclResult CollScatterRingFor91093Executor::CalcLevel0CommInfo(
45 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
46 : {
47 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_RING_INNER);
48 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
49 0 : return HCCL_SUCCESS;
50 0 : }
51 :
52 0 : HcclResult CollScatterRingFor91093Executor::CalcLevel1CommInfo(
53 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
54 : {
55 0 : CommParaInfo commParaLevel1(COMM_LEVEL1, CommType::COMM_TAG_MAX);
56 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
57 0 : commParaLevel1.commType = CommType::COMM_TAG_NONUNIFORM_HIERARCHICAL_RING;
58 0 : HCCL_INFO("[%s]Calc NHRCommInfo", __func__);
59 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
60 0 : commParaLevel1.commType = CommType::COMM_TAG_NONUNIFORM_BRUCK;
61 0 : HCCL_INFO("[%s]Calc NBCommInfo", __func__);
62 : } else {
63 0 : commParaLevel1.commType = CommType::COMM_TAG_RING_INNER;
64 0 : HCCL_INFO("[%s]Calc RingCommInfo", __func__);
65 : }
66 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel1, opTransport[COMM_LEVEL1], inputType, outputType));
67 :
68 0 : return HCCL_SUCCESS;
69 0 : }
70 :
71 0 : HcclResult CollScatterRingFor91093Executor::CalcLevel2CommInfo(
72 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
73 : {
74 : // 910_93 level2当前仅支持nhr、nb、ring算法
75 0 : CommParaInfo commParaLevel2(COMM_LEVEL2, CommType::COMM_TAG_MAX, root_);
76 :
77 0 : if (algType_.algoLevel2 == AlgTypeLevel2::ALG_LEVEL2_NHR) {
78 0 : commParaLevel2.commType = CommType::COMM_TAG_NONUNIFORM_HIERARCHICAL_RING;
79 0 : HCCL_INFO("[%s]Calc NHRCommInfo", __func__);
80 0 : } else if (algType_.algoLevel2 == AlgTypeLevel2::ALG_LEVEL2_NB) {
81 0 : commParaLevel2.commType = CommType::COMM_TAG_NONUNIFORM_BRUCK;
82 0 : HCCL_INFO("[%s]Calc NBCommInfo", __func__);
83 : } else {
84 0 : commParaLevel2.commType = CommType::COMM_TAG_RING_INNER;
85 0 : HCCL_INFO("[%s]Calc RingCommInfo", __func__);
86 : }
87 :
88 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel2, opTransport[COMM_LEVEL2], inputType, outputType));
89 0 : return HCCL_SUCCESS;
90 0 : }
91 :
92 0 : HcclResult CollScatterRingFor91093Executor::KernelRun(const OpParam& param, ExecMem& execMem)
93 : {
94 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] starts.", __func__);
95 0 : Stream& stream = const_cast<Stream&>(param.stream);
96 :
97 0 : CHK_RET(SalGetDataTypeSize(param.DataDes.dataType, perDataSize_));
98 :
99 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
100 0 : level0CommInfo_ = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
101 :
102 0 : commIndex_ = level0CommInfo_.localRank;
103 :
104 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex_ + 1));
105 0 : level1CommInfo_ = GetSubCommInfo(COMM_LEVEL1, commIndex_);
106 :
107 0 : CHK_RET(CheckCommSize(COMM_LEVEL2, COMM_INDEX_0 + 1));
108 0 : level2CommInfo_ = GetSubCommInfo(COMM_LEVEL2, COMM_INDEX_0);
109 :
110 0 : CHK_RET(KernelRunLevel2(param, execMem, stream));
111 0 : CHK_RET(KernelRunLevel1(param, execMem, stream));
112 0 : CHK_RET(KernelRunLevel0(param, execMem, stream));
113 :
114 0 : if (!DMAReduceFlag_) {
115 : DeviceMem srcMem = execMem.inputMem.range(
116 0 : serverSliceOffset_ + execMem.outputMem.size() * commIndex_, execMem.count * perDataSize_);
117 0 : CHK_SMART_PTR_NULL(srcMem.ptr());
118 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, execMem.outputMem, srcMem, stream));
119 0 : }
120 0 : HCCL_INFO("scatter ring run success");
121 0 : return HCCL_SUCCESS;
122 : }
123 :
124 : /* ***********超节点间scatter*********** */
125 0 : HcclResult CollScatterRingFor91093Executor::KernelRunLevel2(const OpParam& param, ExecMem& execMem, Stream& stream)
126 : {
127 0 : u32 level2RankSize = level2CommInfo_.localRankSize;
128 0 : u32 level2Rank = level2CommInfo_.localRank;
129 0 : subUserRankRootSupperPod_ = topoMatcher_->GetSubRootWithSuperPod(topoAttr_.userRank, param.root);
130 :
131 0 : if (level2RankSize > 1 && subUserRankRootSupperPod_ == topoAttr_.userRank) {
132 0 : u32 planeRootSupperPod = 0;
133 0 : CHK_RET(GetRankByUserRank(COMM_LEVEL2, COMM_INDEX_0, param.root, planeRootSupperPod));
134 0 : std::unique_ptr<AlgTemplateBase> level2TempAlg;
135 0 : if (algType_.algoLevel2 == AlgTypeLevel2::ALG_LEVEL2_NB) {
136 : level2TempAlg
137 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_NB, dispatcher_);
138 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_SCATTER_NB in COMM_LEVEL2", __func__);
139 0 : } else if (algType_.algoLevel2 == AlgTypeLevel2::ALG_LEVEL2_NHR) {
140 : level2TempAlg
141 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_NHR, dispatcher_);
142 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_SCATTER_NHR in COMM_LEVEL2", __func__);
143 : } else {
144 : level2TempAlg
145 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_RING, dispatcher_);
146 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_SCATTER_RING in COMM_LEVEL2", __func__);
147 : }
148 :
149 0 : CHK_SMART_PTR_NULL(level2TempAlg);
150 :
151 0 : u64 level2Count = execMem.inputMem.size() / perDataSize_;
152 0 : CHK_RET(level2TempAlg->Prepare(
153 : execMem.inputMem, execMem.inputMem, execMem.scratchMem, level2Count, param.DataDes.dataType, stream,
154 : HCCL_REDUCE_RESERVED, planeRootSupperPod, std::vector<Slice>(0)));
155 0 : CHK_RET(level2TempAlg->RegisterProfiler(
156 : (level2RankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level2Rank, PROF_STAGE_0, HCCL_EXEC_STEP_NOT_SET,
157 : stream));
158 0 : CHK_RET(RunTemplate(level2TempAlg, level2CommInfo_));
159 0 : }
160 0 : return HCCL_SUCCESS;
161 : }
162 :
163 : /* ***********节点间scatter*********** */
164 0 : HcclResult CollScatterRingFor91093Executor::KernelRunLevel1(const OpParam& param, ExecMem& execMem, Stream& stream)
165 : {
166 0 : u32 level2RankSize = level2CommInfo_.localRankSize;
167 0 : u32 level2Rank = level2CommInfo_.localRank;
168 0 : u32 level1RankSize = level1CommInfo_.localRankSize;
169 0 : u32 level1Rank = level1CommInfo_.localRank;
170 0 : HCCL_DEBUG("level1RankSize:%u level1Rank:%u", level1RankSize, level1Rank);
171 :
172 0 : u64 level1SliceSize = execMem.inputMem.size() / level2RankSize;
173 0 : u64 level1SliceCount = level1SliceSize / perDataSize_;
174 0 : level1SliceOffset_ = level1SliceSize * level2Rank;
175 :
176 0 : CHK_RET(topoMatcher_->GetSubRootForScatter(subUserRankRootSupperPod_, subRoot_));
177 0 : CHK_PRT_RET(
178 : subRoot_ == INVALID_VALUE_RANKID,
179 : HCCL_ERROR(
180 : "[CollScatterRingFor91093Executor][KernelRun]GetSubRootForScatter failed, "
181 : "userRank[%u], root[%u], subRoot[%u]",
182 : topoAttr_.userRank, param.root, subRoot_),
183 : HCCL_E_INTERNAL);
184 0 : HCCL_DEBUG(
185 : "[CollScatterRingFor91093Executor][KernelRun]GetSubRootForScatter, userRank[%u], root[%u], subRoot[%u]",
186 : topoAttr_.userRank, param.root, subRoot_);
187 :
188 0 : if (level1RankSize > 1 && subRoot_ == topoAttr_.userRank) {
189 0 : u32 rootRankLevel1 = 0;
190 0 : CHK_RET(GetRankByUserRank(COMM_LEVEL1, commIndex_, subUserRankRootSupperPod_, rootRankLevel1));
191 :
192 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
193 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
194 : level1TempAlg
195 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_NB, dispatcher_);
196 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_SCATTER_NB in COMM_LEVEL1", __func__);
197 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
198 : level1TempAlg
199 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_NHR, dispatcher_);
200 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_SCATTER_NHR in COMM_LEVEL1", __func__);
201 : } else {
202 : level1TempAlg
203 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_RING, dispatcher_);
204 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_SCATTER_RING in COMM_LEVEL1", __func__);
205 : }
206 0 : CHK_SMART_PTR_NULL(level1TempAlg);
207 :
208 0 : DeviceMem level1InputMem = execMem.inputMem.range(level1SliceOffset_, level1SliceSize);
209 0 : CHK_SMART_PTR_NULL(level1InputMem.ptr());
210 :
211 0 : CHK_RET(level1TempAlg->Prepare(
212 : level1InputMem, level1InputMem, level1InputMem, level1SliceCount, param.DataDes.dataType, stream,
213 : HCCL_REDUCE_RESERVED, rootRankLevel1, std::vector<Slice>(0), level1SliceOffset_));
214 0 : CHK_RET(level1TempAlg->RegisterProfiler(
215 : (level1RankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1Rank, PROF_STAGE_1, HCCL_EXEC_STEP_NOT_SET,
216 : stream));
217 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo_));
218 0 : }
219 0 : return HCCL_SUCCESS;
220 : }
221 :
222 : /* ***********节点内scatter*********** */
223 0 : HcclResult CollScatterRingFor91093Executor::KernelRunLevel0(const OpParam& param, ExecMem& execMem, Stream& stream)
224 : {
225 : // 每个server分配的slice大小
226 0 : u32 level0RankSize = level0CommInfo_.localRankSize;
227 0 : u32 level2RankSize = level2CommInfo_.localRankSize;
228 0 : u32 level1RankSize = level1CommInfo_.localRankSize;
229 0 : u32 level1Rank = level1CommInfo_.localRank;
230 :
231 0 : u64 serverSliceSize = execMem.inputMem.size() / (level1RankSize * level2RankSize);
232 0 : serverSliceOffset_ = serverSliceSize * level1Rank + level1SliceOffset_;
233 0 : HCCL_DEBUG(
234 : "inputMem.size()=%llu, commLevel0->RankSize()=%u, serverSliceSize=%llu, serverSliceOffset=%llu "
235 : "commIndex=%u commLevel1[commIndex]->rank=%u",
236 : execMem.inputMem.size(), level0RankSize, serverSliceSize, serverSliceOffset_, commIndex_, level1Rank);
237 :
238 0 : DeviceMem scatterRingInput = execMem.inputMem.range(serverSliceOffset_, serverSliceSize);
239 0 : CHK_SMART_PTR_NULL(scatterRingInput);
240 :
241 : // 将根节点数据切分成level0RankSize份
242 0 : std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
243 0 : std::vector<std::vector<Slice>> mulRingSlice; // 每个stream使用的数据基于用户buffer的偏移
244 : // 根据数据量算每个环上数据的偏移和大小
245 0 : CHK_RET(PrepareDataSlice(execMem.count, perDataSize_, level0RankSize, dataSegsSlice));
246 :
247 : u32 ringNum;
248 0 : if (topoType_ == TopoType::TOPO_TYPE_NP_DOUBLE_RING) {
249 0 : ringNum = LEVEL0_PLANE_NUM_IN_NPRING_DOUBLE;
250 0 : mulRingSlice = PrepareMultiRingSlice(dataSegsSlice, param.tag, false, topoAttr_.nicList);
251 : } else {
252 0 : ringNum = LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
253 0 : mulRingSlice.push_back(dataSegsSlice);
254 : }
255 0 : CHK_PRT_RET(
256 : mulRingSlice.size() != ringNum,
257 : HCCL_ERROR(
258 : "[CollScatterRingFor91093Executor][KernelRunLevel0]ringNum[%u] != mulRingSlice size[%zu]", ringNum,
259 : mulRingSlice.size()),
260 : HCCL_E_INTERNAL);
261 0 : HCCL_INFO("scatter ring/scatter ring direct: using multiring algo inner-server.");
262 0 : HcomCollOpInfo* scatterOpInfoPtr = nullptr;
263 0 : HcomCollOpInfo scatterOpInfo
264 0 : = {"", nullptr, execMem.outputPtr, param.DataDes.count, param.DataDes.dataType, subRoot_, param.reduceType, 0};
265 0 : if (DMAReduceFlag_) {
266 0 : scatterOpInfoPtr = &scatterOpInfo;
267 : }
268 0 : CHK_RET(MultiRingScatter(
269 : param.tag, scatterRingInput, scatterRingInput, execMem.count, param.DataDes.dataType, mulRingSlice, subRoot_,
270 : stream, scatterOpInfoPtr, serverSliceOffset_));
271 0 : return HCCL_SUCCESS;
272 0 : }
273 0 : HcclResult CollScatterRingFor91093Executor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
274 : {
275 0 : if (CheckCommSize(COMM_LEVEL2, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
276 0 : return HCCL_E_UNAVAIL;
277 : }
278 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL2, COMM_INDEX_0);
279 :
280 0 : return HCCL_SUCCESS;
281 : }
282 :
283 : HcclResult
284 0 : CollScatterRingFor91093Executor::SelectTempAlg(std::unique_ptr<AlgTemplateBase>& level1TempAlg, u32 level1RankSize)
285 : {
286 0 : if (level1RankSize > 1) {
287 0 : if (algType_.algoLevel2 == AlgTypeLevel2::ALG_LEVEL2_NB) {
288 : level1TempAlg
289 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_NB, dispatcher_);
290 0 : CHK_SMART_PTR_NULL(level1TempAlg);
291 0 : } else if (algType_.algoLevel2 == AlgTypeLevel2::ALG_LEVEL2_NHR) {
292 : level1TempAlg
293 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_NHR, dispatcher_);
294 0 : CHK_SMART_PTR_NULL(level1TempAlg);
295 : } else {
296 : level1TempAlg
297 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_SCATTER_RING, dispatcher_);
298 0 : CHK_SMART_PTR_NULL(level1TempAlg);
299 : }
300 0 : return HCCL_SUCCESS;
301 : }
302 0 : return HCCL_E_UNAVAIL;
303 : }
304 : REGISTER_EXEC("ScatterRingFor91093Executor", ScatterRingFor91093, CollScatterRingFor91093Executor);
305 : } // namespace hccl
|
