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_aiv_rdma_executor.h"
12 :
13 : namespace hccl {
14 : constexpr u32 A_X_SIZE = 16;
15 :
16 0 : CollAllGatherAivRdmaExecutor::CollAllGatherAivRdmaExecutor(const HcclDispatcher dispatcher,
17 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
18 0 : : CollAllGatherExecutor(dispatcher, topoMatcher)
19 : {
20 0 : DMAReduceFlag_ = false;
21 0 : desc_.isAivMode = true;
22 0 : }
23 :
24 0 : HcclResult CollAllGatherAivRdmaExecutor::CalcStreamNum(u32& streamNum)
25 : {
26 0 : streamNum=0;
27 0 : HCCL_INFO("[CollAllGatherAivRdmaExecutor][CalcStreamNum] tag[%s] streamNum[%u]", tag_.c_str(), streamNum);
28 0 : return HCCL_SUCCESS;
29 : }
30 :
31 0 : HcclResult CollAllGatherAivRdmaExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
32 : {
33 0 : CHK_RET(CalcLevel0CommInfo(TransportMemType::CCL_OUTPUT, TransportMemType::AIV_INPUT, opTransport));
34 0 : CHK_RET(CalcLevel1CommInfo(TransportMemType::CCL_INPUT, TransportMemType::CCL_OUTPUT, opTransport));
35 0 : return HCCL_SUCCESS;
36 : }
37 :
38 0 : HcclResult CollAllGatherAivRdmaExecutor::CalcLevel0CommInfo(TransportMemType inputType,
39 : TransportMemType outputType,
40 : std::vector<LevelNSubCommTransport>& opTransport)
41 : {
42 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
43 0 : commParaLevel0.meshSinglePlane = true;
44 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
45 0 : return HCCL_SUCCESS;
46 0 : }
47 :
48 0 : HcclResult CollAllGatherAivRdmaExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
49 : {
50 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
51 0 : u32 bestNumBlocks = numBlocks;
52 :
53 0 : CHK_PRT_RET(numBlocks_ < numBlocks,
54 : HCCL_WARNING("[CollAllGatherAivRdmaExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
55 : numBlocks_, numBlocks), HCCL_E_PARA);
56 :
57 0 : HCCL_INFO("[CollAllGatherAivRdmaExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]",
58 : numBlocks, numBlocks_, bestNumBlocks);
59 0 : return HCCL_SUCCESS;
60 : }
61 :
62 0 : HcclResult CollAllGatherAivRdmaExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
63 : {
64 0 : HcclUs startut = TIME_NOW();
65 0 : tag_ = param.tag;
66 0 : algResResp_ = &algRes;
67 0 : u64 totalSize = param.DataDes.count * SIZE_TABLE[param.DataDes.dataType];
68 :
69 0 : ExecMem execMem;
70 0 : execMem.count = param.DataDes.count;
71 0 : execMem.scratchMem = algRes.scratchMem;
72 0 : execMem.inputPtr = param.inputPtr;
73 0 : execMem.outputPtr = param.outputPtr;
74 :
75 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
76 0 : execMem.inputMem = algRes.cclOutputMem;
77 0 : execMem.outputMem = algRes.aivInputMem; // 使用aivIn的第33M作为Flag区
78 : } else {
79 0 : execMem.inputMem = algRes.aivInputMem;
80 0 : execMem.outputMem = DeviceMem::create(algRes.paramOutputMem.ptr(), totalSize * topoAttr_.userRankSize);
81 : }
82 :
83 0 : HcclResult ret = KernelRun(param, execMem);
84 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
85 : HCCL_ERROR("[CollAllGatherAivRdmaExecutor]errNo[0x%016llx] tag[%s] executor kernel run failed",
86 : HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
87 :
88 0 : HCCL_INFO("tag[%s], AllGather executor orchestrate success, take time [%lld]us.",
89 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
90 0 : return HCCL_SUCCESS;
91 0 : }
92 :
93 0 : HcclResult CollAllGatherAivRdmaExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
94 : {
95 0 : HCCL_INFO("[CollAllGatherAivRdmaExecutor][KernelRun]AllGather aiv enter");
96 0 : HcclWorkflowMode workflow = workflowMode_;
97 0 : bool isOpbase = (workflow == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
98 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
99 :
100 : //获取子通信域信息
101 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
102 0 : SubCommInfo outerCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
103 0 : u32 outerRankSize = outerCommInfo.localRankSize;
104 0 : u32 commIndex = outerCommInfo.localRank;
105 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
106 0 : SubCommInfo innerCommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
107 0 : u32 serverIndex = innerCommInfo.localRank;
108 0 : u64 inputMemSize = perDataSize * param.DataDes.count;
109 :
110 0 : auto autoSelectedAlgTypeLevel1 = static_cast<u32>(algType_.algoLevel1);
111 0 : bool hugeData = IsHugeData(inputMemSize);
112 0 : auto opMeta = HcclOpMetaInfo::GetOneForAllGather(autoSelectedAlgTypeLevel1, hugeData, false, CopyPattern::BCOPY,
113 : false, true);
114 :
115 0 : CHK_RET(InitTask(dispatcher_, const_cast<Stream&>(param.stream), opMeta.isEnableCache, opMeta.GetCacheKey()));
116 0 : DeviceMem dstMem;
117 0 : u64 baseOffset = serverIndex * inputMemSize;
118 0 : if (isOpbase) {
119 0 : dstMem = execMem.inputMem.range(baseOffset, inputMemSize);
120 0 : CHK_SMART_PTR_NULL(dstMem);
121 : }
122 : else {
123 0 : baseOffset = serverIndex * inputMemSize * outerRankSize;
124 0 : u64 outerOffset = commIndex * inputMemSize;
125 0 : dstMem = execMem.inputMem.range(baseOffset + outerOffset, inputMemSize);
126 0 : CHK_SMART_PTR_NULL(dstMem);
127 : }
128 :
129 : // STP1，将数据从userinput内存拷贝到ccloutput内存的对应位置
130 0 : DeviceMem srcMem = DeviceMem::create(static_cast<u8 *>(param.inputPtr), inputMemSize);
131 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, dstMem, srcMem, const_cast<Stream&>(param.stream));
132 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
133 : HCCL_ERROR("[CollAllGatherAivRdmaExecutor][KernelRun]AllGather 4PmeshHD memcpy Failed, Offset[%llu], Size[%llu].",
134 : baseOffset, inputMemSize), ret);
135 : // STP2， AI server 间 recursive halving doubling AllGather
136 0 : u64 hdCount = inputMemSize / perDataSize;
137 0 : std::unique_ptr<ExecutorBase> innerExecutor;
138 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING || (topoAttr_.isDiffDeviceModule && topoAttr_.serverNum == 1)) {
139 : // 1-单server-SDMA
140 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
141 0 : TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
142 0 : HCCL_INFO("AllGather mesh: using ring algo inter-server.");
143 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
144 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
145 0 : TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
146 0 : HCCL_INFO("AllGather mesh: using nhr algo inter-server.");
147 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
148 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
149 0 : TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
150 0 : HCCL_INFO("AllGather mesh: using nhr_v1 algo inter-server.");
151 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
152 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
153 0 : TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
154 0 : HCCL_INFO("AllGather mesh: using nonuniform-bruck algo inter-server.");
155 : } else {
156 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
157 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
158 0 : HCCL_INFO("AllGather mesh: using halving-doubling algo inter-server.");
159 : }
160 0 : CHK_SMART_PTR_NULL(innerExecutor);
161 :
162 0 : DeviceMem rdmadstMem;
163 0 : if (isOpbase) {
164 0 : rdmadstMem = execMem.inputMem.range(0, inputMemSize * innerCommInfo.localRankSize);
165 0 : CHK_SMART_PTR_NULL(rdmadstMem);
166 : }
167 : else {
168 0 : hdCount = inputMemSize*outerRankSize / perDataSize;
169 0 : rdmadstMem = execMem.inputMem;
170 : }
171 :
172 0 : CHK_RET(innerExecutor->Prepare(rdmadstMem, rdmadstMem, execMem.inputMem, hdCount,
173 : param.DataDes.dataType, param.stream, HcclReduceOp::HCCL_REDUCE_RESERVED, INVALID_VALUE_RANKID,
174 : std::vector<Slice>(COMM_INDEX_0), 0));
175 :
176 0 : u32 rankSize = innerCommInfo.localRankSize;
177 0 : CHK_RET(innerExecutor->RegisterProfiler((rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + serverIndex,
178 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
179 :
180 0 : CHK_RET(RunTemplate(innerExecutor, innerCommInfo));
181 0 : CHK_RET(LaunchTask(dispatcher_, const_cast<Stream&>(param.stream)));
182 :
183 : //STP3 机内重排
184 : void *buffersIn[MAX_RANK_SIZE];
185 : void *buffersOut[MAX_RANK_SIZE];
186 :
187 0 : u32 localRank = outerCommInfo.localRank;
188 0 : u32 localRankSize = outerCommInfo.localRankSize;
189 :
190 : void* tmpAivBufferData;
191 : void* tmpAivBufferFlag;
192 :
193 0 : for (u32 i = 0; i < localRankSize; i++) {
194 0 : if (i != localRank) {
195 0 : CHK_RET(outerCommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(tmpAivBufferData)));
196 0 : CHK_RET(outerCommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(tmpAivBufferFlag)));
197 0 : buffersIn[i] = static_cast<u8 *>(tmpAivBufferData);
198 0 : buffersOut[i] = static_cast<u8 *>(tmpAivBufferFlag) + HCCL_MID_COUNT_32_MB;
199 : } else {
200 0 : buffersIn[i] = static_cast<u8 *>(execMem.inputMem.ptr());
201 0 : buffersOut[i] = static_cast<u8 *>(execMem.outputMem.ptr()) + HCCL_MID_COUNT_32_MB;
202 : }
203 : }
204 :
205 0 : u32 serverNum = innerCommInfo.localRankSize;
206 :
207 0 : AivOpArgs opArgs {
208 0 : HcclCMDType::HCCL_CMD_ALLGATHER, execMem.inputPtr, execMem.outputPtr, param.DataDes.count,
209 0 : param.DataDes.dataType, HCCL_REDUCE_RESERVED, 0, isOpbase
210 0 : };
211 : AivTopoArgs topoArgs {
212 0 : localRank, localRankSize, topoAttr_.isDiffDeviceModule ? topoAttr_.devicePhyId : A_X_SIZE,
213 0 : 0, serverNum, topoAttr_.deviceType, algoAttr_.identifier
214 0 : };
215 : u32 numBlocks;
216 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, localRankSize) != HCCL_SUCCESS,
217 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
218 : HCCL_E_PARA);
219 0 : numBlocks_ = numBlocks;
220 : AivResourceArgs resourceArgs {
221 0 : param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(), numBlocks_, param.aivTag
222 0 : };
223 0 : AivAlgArgs algArgs {0};
224 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
225 0 : algArgs.execTimeOutSet = true;
226 :
227 0 : struct AivProfilingInfo aivProfilingInfo;
228 0 : aivProfilingInfo.counter = opCounter_;
229 :
230 0 : CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
231 0 : HCCL_INFO("[CollAllGatherAivRdmaExecutor][KernelRun]allGather aiv run success.");
232 0 : return HCCL_SUCCESS;
233 0 : }
234 :
235 0 : HcclResult CollAllGatherAivRdmaExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
236 : {
237 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
238 0 : return HCCL_E_UNAVAIL;
239 : }
240 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
241 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
242 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
243 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
244 :
245 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
246 0 : return HCCL_E_UNAVAIL;
247 : }
248 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
249 :
250 0 : return HCCL_SUCCESS;
251 0 : }
252 :
253 0 : HcclResult CollAllGatherAivRdmaExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
254 : {
255 0 : if (level1RankSize > 1) {
256 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING || (topoAttr_.isDiffDeviceModule && topoAttr_.serverNum == 1)) {
257 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
258 0 : TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
259 0 : HCCL_INFO("AllGather mesh: using ring algo inter-server.");
260 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
261 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
262 0 : TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
263 0 : HCCL_INFO("AllGather mesh: using nhr algo inter-server.");
264 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
265 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
266 0 : TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
267 0 : HCCL_INFO("AllGather mesh: using nhr_v1 algo inter-server.");
268 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
269 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
270 0 : TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
271 0 : HCCL_INFO("AllGather mesh: using nonuniform-bruck algo inter-server.");
272 : } else {
273 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
274 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
275 0 : HCCL_INFO("AllGather mesh: using halving-doubling algo inter-server.");
276 : }
277 0 : CHK_SMART_PTR_NULL(level1TempAlg);
278 0 : return HCCL_SUCCESS;
279 : }
280 0 : return HCCL_E_UNAVAIL;
281 : }
282 :
283 : REGISTER_EXEC("AllGatherAivRdmaExecutor", AllGatherAivRdma, CollAllGatherAivRdmaExecutor);
284 :
285 : } // namespace hccl
|
