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_ring_executor.h"
12 :
13 : namespace hccl {
14 0 : CollAllGatherRingExecutor::CollAllGatherRingExecutor(const HcclDispatcher dispatcher,
15 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
16 0 : : CollAllGatherExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = false;
19 0 : }
20 :
21 0 : HcclResult CollAllGatherRingExecutor::CalcStreamNum(u32& streamNum)
22 : {
23 0 : u32 totalStreamNum = 1U;
24 0 : if (algType_.algoLevel0 == AlgTypeLevel0::ALG_LEVEL0_8P_RING) {
25 0 : totalStreamNum = LEVEL0_PLANE_NUM_IN_8PRING;
26 : }
27 0 : streamNum = totalStreamNum - 1;
28 0 : HCCL_INFO("[CollAllGatherRingExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
29 : tag_.c_str(), streamNum);
30 0 : return HCCL_SUCCESS;
31 : }
32 :
33 0 : HcclResult CollAllGatherRingExecutor::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 : return HCCL_SUCCESS;
41 : }
42 :
43 0 : HcclResult CollAllGatherRingExecutor::CalcTransportMemType(TransportMemType &inputType, TransportMemType &outputType)
44 : {
45 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
46 0 : inputType = TransportMemType::CCL_INPUT;
47 0 : outputType = TransportMemType::CCL_OUTPUT;
48 : } else {
49 0 : inputType = TransportMemType::PARAM_INPUT;
50 0 : outputType = TransportMemType::PARAM_OUTPUT;
51 : }
52 0 : HCCL_INFO("[CollAllGatherRingExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]",
53 : tag_.c_str(), inputType, outputType);
54 0 : return HCCL_SUCCESS;
55 : }
56 :
57 0 : HcclResult CollAllGatherRingExecutor::CalcLevel0CommInfo(TransportMemType inputType,
58 : TransportMemType outputType,
59 : std::vector<LevelNSubCommTransport>& opTransport)
60 : {
61 0 : HCCL_INFO("[CollAllGatherRingExecutor][CalcLevel0CommInfo]tag[%s] start", tag_.c_str());
62 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_RING_INNER);
63 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
64 0 : HCCL_INFO("[CollAllGatherRingExecutor][CalcLevel0CommInfo]tag[%s] Calc RingComm finish", tag_.c_str());
65 0 : return HCCL_SUCCESS;
66 0 : }
67 :
68 0 : u64 CollAllGatherRingExecutor::CalcLoopMaxCount(const u64 cclBuffSize, const u32 unitSize)
69 : {
70 0 : u64 maxCountPerLoop = cclBuffSize / (topoAttr_.userRankSize * unitSize);
71 0 : return maxCountPerLoop;
72 : }
73 :
74 0 : HcclResult CollAllGatherRingExecutor::SelectAlgorithmTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg)
75 : {
76 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
77 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
78 0 : TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
79 0 : HCCL_INFO("AllGather ring: using ring algo inter-server.");
80 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
81 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
82 0 : TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
83 0 : HCCL_INFO("AllGather ring: using nhr algo inter-server.");
84 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
85 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
86 0 : TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
87 0 : HCCL_INFO("AllGather ring: using nhr_v1 algo inter-server.");
88 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
89 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
90 0 : TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
91 0 : HCCL_INFO("AllGather ring: using nonuniform-bruck algo inter-server.");
92 : } else {
93 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
94 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
95 0 : HCCL_INFO("AllGather ring: using halving-doubling algo inter-server.");
96 : }
97 0 : CHK_SMART_PTR_NULL(level1TempAlg);
98 0 : return HCCL_SUCCESS;
99 : }
100 :
101 0 : HcclResult CollAllGatherRingExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
102 : {
103 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollAllGatherRingExecutor][KernelRun]The AllGatherRingExecutor starts.");
104 0 : u32 perDataSize = 0;
105 0 : CHK_RET(SalGetDataTypeSize(param.DataDes.dataType, perDataSize));
106 0 : CHK_PRT_RET(perDataSize == 0,
107 : HCCL_ERROR("[CollAllGatherRingExecutor][KernelRun]errNo[0x%016llx] datatype[%d] is invalid",
108 : HCCL_ERROR_CODE(HCCL_E_PARA), param.DataDes.dataType), HCCL_E_PARA);
109 :
110 : // 获取子通信域的信息
111 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
112 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
113 :
114 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, ringNum));
115 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
116 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
117 0 : u32 level0RankSize = level0CommInfo.localRankSize;
118 :
119 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
120 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
121 0 : u32 serverIndex = level1CommInfo.localRank;
122 :
123 : // 第一步，如果非DMA消减，将数据从input内存拷贝到output内存的对应位置
124 0 : u64 inputMemSize = execMem.inputMem.size();
125 0 : u64 baseOffset = serverIndex * inputMemSize * level0RankSize;
126 0 : u64 level0Offset = commIndex * inputMemSize;
127 0 : DeviceMem dstMem = execMem.outputMem.range(baseOffset + level0Offset, inputMemSize);
128 0 : CHK_SMART_PTR_NULL(dstMem);
129 :
130 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, dstMem, execMem.inputMem, const_cast<Stream&>(param.stream));
131 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
132 : HCCL_ERROR("[CollAllGatherRingExecutor][KernelRun]AllGather 8PringHD memcpy Failed, "
133 : "Offset[%llu], Size[%llu]", baseOffset + level0Offset, inputMemSize), ret);
134 :
135 : // 第二步，各个AI Server 内 multi ring AllGather
136 0 : std::vector<Slice> dataSegsSlice; // 数据分成ranksize份，每份的起始偏移和大小
137 0 : std::vector<std::vector<Slice>> multRingsSliceZero; // 数据基于该rank上环0的偏移
138 0 : u32 sliceNum = level0RankSize;
139 0 : CHK_RET(PrepareAllgatherSlice(sliceNum, inputMemSize, dataSegsSlice));
140 :
141 : // 多环数据切分
142 0 : if (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) {
143 0 : multRingsSliceZero = PrepareMultiRingSlice(dataSegsSlice, param.tag);
144 : } else {
145 0 : multRingsSliceZero.push_back(dataSegsSlice);
146 : }
147 0 : CHK_PRT_RET(multRingsSliceZero.size() != ringNum,
148 : HCCL_ERROR("[CollAllGatherRingExecutor][KernelRun]ringNum[%u] != multRingsSliceZero size[%zu]",
149 : ringNum, multRingsSliceZero.size()), HCCL_E_INTERNAL);
150 :
151 : // 抽取当前用于多环AllGather 的output内存数据
152 0 : DeviceMem currentOutputMem = execMem.outputMem.range(baseOffset, inputMemSize * level0RankSize);
153 0 : CHK_SMART_PTR_NULL(currentOutputMem);
154 :
155 0 : CHK_RET(ActiveSlaveStreams(param.stream));
156 :
157 0 : CHK_RET(MultiRingAllGather(param.tag, execMem.inputMem, currentOutputMem, execMem.count, param.DataDes.dataType,
158 : multRingsSliceZero, param.stream, PROF_STAGE_1, baseOffset, nullptr));
159 :
160 0 : HCCL_INFO("AllGather 8PringHD level0 run success");
161 :
162 : // 第三步， AI server 间 recursive halving doubling AllGather
163 0 : u64 hdSize = 0;
164 0 : std::vector<u32> nicList = const_cast<std::vector<u32>&>(topoAttr_.nicList);
165 0 : std::vector<u32>::iterator iterNic = std::find(nicList.begin(), nicList.end(), topoAttr_.devicePhyId);
166 0 : if (iterNic != nicList.end()) {
167 0 : hdSize = inputMemSize * level0RankSize;
168 : }
169 0 : u64 hdCount = hdSize / perDataSize;
170 :
171 0 : bool isMultiNic = topoType_ == TopoType::TOPO_TYPE_8P_RING && nicList.size() != DEVICE_EIGHT;
172 0 : bool innRunRet = isMultiNic && (iterNic == nicList.end());
173 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
174 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
175 0 : CHK_RET(SelectAlgorithmTempAlg(level1TempAlg));
176 :
177 : // 此处虽然带入inputMem作为scratch mem, 但inputMem 不能被使用
178 0 : CHK_RET(level1TempAlg->Prepare(execMem.outputMem, execMem.outputMem, execMem.inputMem, hdCount,
179 : param.DataDes.dataType, param.stream, HCCL_REDUCE_RESERVED, INVALID_VALUE_RANKID,
180 : std::vector<Slice>(COMM_INDEX_0), 0));
181 :
182 0 : u32 rankSize = level1CommInfo.localRankSize;
183 0 : CHK_RET(level1TempAlg->RegisterProfiler((rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + serverIndex,
184 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
185 :
186 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
187 0 : }
188 0 : HCCL_INFO("AllGather 8PringHD level1 run success");
189 :
190 : // 网口裁剪：AI server 内多网口的allgather
191 0 : if (topoType_ == TopoType::TOPO_TYPE_8P_RING && nicList.size() != DEVICE_EIGHT) {
192 0 : CHK_RET(ActiveSlaveStreams(param.stream)); // 为什么要active两遍
193 :
194 0 : u32 perDataSize = 0;
195 0 : CHK_RET(SalGetDataTypeSize(param.DataDes.dataType, perDataSize));
196 0 : u64 tempCount = execMem.outputMem.size() / perDataSize;
197 0 : CHK_RET(AlgTemplateBase::PrepareSliceData(tempCount, perDataSize, sliceNum, 0, dataSegsSlice));
198 0 : multRingsSliceZero = PrepareMultiRingSlice(dataSegsSlice, param.tag, false, nicList);
199 0 : CHK_PRT_RET(multRingsSliceZero.size() != ringNum, HCCL_ERROR("[CollAllGatherRingExecutor][KernelRun]"\
200 : "ringNum[%u] != multRingsSliceZero size[%zu]", ringNum, multRingsSliceZero.size()), HCCL_E_INTERNAL);
201 :
202 0 : CHK_RET(MultiRingAllGather(param.tag, execMem.outputMem, execMem.outputMem, tempCount / DEVICE_EIGHT,
203 : param.DataDes.dataType, multRingsSliceZero, param.stream, PROF_STAGE_1));
204 :
205 0 : HCCL_INFO("AllGather 8PringHD level1 chunk run success");
206 : }
207 0 : return HCCL_SUCCESS;
208 0 : }
209 0 : HcclResult CollAllGatherRingExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
210 : {
211 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
212 0 : return HCCL_E_UNAVAIL;
213 : }
214 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
215 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
216 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
217 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
218 :
219 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
220 0 : return HCCL_E_UNAVAIL;
221 : }
222 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
223 :
224 0 : return HCCL_SUCCESS;
225 0 : }
226 :
227 0 : HcclResult CollAllGatherRingExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
228 : {
229 0 : if (level1RankSize > 1) {
230 0 : CHK_RET(SelectAlgorithmTempAlg(level1TempAlg));
231 : }
232 0 : return HCCL_E_UNAVAIL;
233 : }
234 :
235 : REGISTER_EXEC("AllGatherRingExecutor", AllGatherRing, CollAllGatherRingExecutor);
236 :
237 : } // namespace hccl
|
