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(
17 0 : const HcclDispatcher dispatcher, 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(
39 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
40 : {
41 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
42 0 : commParaLevel0.meshSinglePlane = true;
43 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
44 0 : return HCCL_SUCCESS;
45 0 : }
46 :
47 0 : HcclResult CollAllGatherAivRdmaExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
48 : {
49 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
50 0 : u32 bestNumBlocks = numBlocks;
51 :
52 0 : CHK_PRT_RET(
53 : numBlocks_ < numBlocks,
54 : HCCL_WARNING(
55 : "[CollAllGatherAivRdmaExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].", numBlocks_,
56 : numBlocks),
57 : HCCL_E_PARA);
58 :
59 0 : HCCL_INFO(
60 : "[CollAllGatherAivRdmaExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]", numBlocks,
61 : numBlocks_, bestNumBlocks);
62 0 : return HCCL_SUCCESS;
63 : }
64 :
65 0 : HcclResult CollAllGatherAivRdmaExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
66 : {
67 0 : HcclUs startut = TIME_NOW();
68 0 : tag_ = param.tag;
69 0 : algResResp_ = &algRes;
70 0 : u64 totalSize = param.DataDes.count * SIZE_TABLE[param.DataDes.dataType];
71 :
72 0 : ExecMem execMem;
73 0 : execMem.count = param.DataDes.count;
74 0 : execMem.scratchMem = algRes.scratchMem;
75 0 : execMem.inputPtr = param.inputPtr;
76 0 : execMem.outputPtr = param.outputPtr;
77 :
78 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
79 0 : execMem.inputMem = algRes.cclOutputMem;
80 0 : execMem.outputMem = algRes.aivInputMem; // 使用aivIn的第33M作为Flag区
81 : } else {
82 0 : execMem.inputMem = algRes.aivInputMem;
83 0 : execMem.outputMem = DeviceMem::create(algRes.paramOutputMem.ptr(), totalSize * topoAttr_.userRankSize);
84 : }
85 :
86 0 : HcclResult ret = KernelRun(param, execMem);
87 0 : CHK_PRT_RET(
88 : ret != HCCL_SUCCESS,
89 : HCCL_ERROR(
90 : "[CollAllGatherAivRdmaExecutor]errNo[0x%016llx] tag[%s] executor kernel run failed", HCCL_ERROR_CODE(ret),
91 : param.tag.c_str()),
92 : ret);
93 :
94 0 : HCCL_INFO(
95 : "tag[%s], AllGather executor orchestrate success, take time [%lld]us.", param.tag.c_str(),
96 : DURATION_US(TIME_NOW() - startut));
97 0 : return HCCL_SUCCESS;
98 0 : }
99 :
100 0 : HcclResult CollAllGatherAivRdmaExecutor::KernelRun(const OpParam& param, ExecMem& execMem)
101 : {
102 0 : HCCL_INFO("[CollAllGatherAivRdmaExecutor][KernelRun]AllGather aiv enter");
103 0 : HcclWorkflowMode workflow = workflowMode_;
104 0 : bool isOpbase = (workflow == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
105 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
106 :
107 : // 获取子通信域信息
108 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
109 0 : SubCommInfo outerCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
110 0 : u32 outerRankSize = outerCommInfo.localRankSize;
111 0 : u32 commIndex = outerCommInfo.localRank;
112 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
113 0 : SubCommInfo innerCommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
114 0 : u32 serverIndex = innerCommInfo.localRank;
115 0 : u64 inputMemSize = perDataSize * param.DataDes.count;
116 :
117 0 : auto autoSelectedAlgTypeLevel1 = static_cast<u32>(algType_.algoLevel1);
118 0 : bool hugeData = IsHugeData(inputMemSize);
119 0 : auto opMeta = HcclOpMetaInfo::GetOneForAllGather(
120 : autoSelectedAlgTypeLevel1, hugeData, false, CopyPattern::BCOPY, false, true);
121 :
122 0 : CHK_RET(InitTask(dispatcher_, const_cast<Stream&>(param.stream), opMeta.isEnableCache, opMeta.GetCacheKey()));
123 0 : DeviceMem dstMem;
124 0 : u64 baseOffset = serverIndex * inputMemSize;
125 0 : if (isOpbase) {
126 0 : dstMem = execMem.inputMem.range(baseOffset, inputMemSize);
127 0 : CHK_SMART_PTR_NULL(dstMem);
128 : } else {
129 0 : baseOffset = serverIndex * inputMemSize * outerRankSize;
130 0 : u64 outerOffset = commIndex * inputMemSize;
131 0 : dstMem = execMem.inputMem.range(baseOffset + outerOffset, inputMemSize);
132 0 : CHK_SMART_PTR_NULL(dstMem);
133 : }
134 :
135 : // STP1，将数据从userinput内存拷贝到ccloutput内存的对应位置
136 0 : DeviceMem srcMem = DeviceMem::create(static_cast<u8*>(param.inputPtr), inputMemSize);
137 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, dstMem, srcMem, const_cast<Stream&>(param.stream));
138 0 : CHK_PRT_RET(
139 : ret != HCCL_SUCCESS,
140 : HCCL_ERROR(
141 : "[CollAllGatherAivRdmaExecutor][KernelRun]AllGather 4PmeshHD memcpy Failed, Offset[%llu], Size[%llu].",
142 : baseOffset, inputMemSize),
143 : ret);
144 : // STP2， AI server 间 recursive halving doubling AllGather
145 0 : u64 hdCount = inputMemSize / perDataSize;
146 0 : std::unique_ptr<ExecutorBase> innerExecutor;
147 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING
148 0 : || (topoAttr_.isDiffDeviceModule && topoAttr_.serverNum == 1)) {
149 : // 1-单server-SDMA
150 : innerExecutor
151 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
152 0 : HCCL_INFO("AllGather mesh: using ring algo inter-server.");
153 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
154 : innerExecutor
155 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
156 0 : HCCL_INFO("AllGather mesh: using nhr algo inter-server.");
157 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
158 : innerExecutor
159 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
160 0 : HCCL_INFO("AllGather mesh: using nhr_v1 algo inter-server.");
161 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
162 : innerExecutor
163 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
164 0 : HCCL_INFO("AllGather mesh: using nonuniform-bruck algo inter-server.");
165 : } else {
166 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
167 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
168 0 : HCCL_INFO("AllGather mesh: using halving-doubling algo inter-server.");
169 : }
170 0 : CHK_SMART_PTR_NULL(innerExecutor);
171 :
172 0 : DeviceMem rdmadstMem;
173 0 : if (isOpbase) {
174 0 : rdmadstMem = execMem.inputMem.range(0, inputMemSize * innerCommInfo.localRankSize);
175 0 : CHK_SMART_PTR_NULL(rdmadstMem);
176 : } else {
177 0 : hdCount = inputMemSize * outerRankSize / perDataSize;
178 0 : rdmadstMem = execMem.inputMem;
179 : }
180 :
181 0 : CHK_RET(innerExecutor->Prepare(
182 : rdmadstMem, rdmadstMem, execMem.inputMem, hdCount, param.DataDes.dataType, param.stream,
183 : HcclReduceOp::HCCL_REDUCE_RESERVED, INVALID_VALUE_RANKID, std::vector<Slice>(COMM_INDEX_0), 0));
184 :
185 0 : u32 rankSize = innerCommInfo.localRankSize;
186 0 : CHK_RET(innerExecutor->RegisterProfiler(
187 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + serverIndex, PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET,
188 : param.stream));
189 :
190 0 : CHK_RET(RunTemplate(innerExecutor, innerCommInfo));
191 0 : CHK_RET(LaunchTask(dispatcher_, const_cast<Stream&>(param.stream)));
192 :
193 : // STP3 机内重排
194 : void* buffersIn[MAX_RANK_SIZE];
195 : void* buffersOut[MAX_RANK_SIZE];
196 :
197 0 : u32 localRank = outerCommInfo.localRank;
198 0 : u32 localRankSize = outerCommInfo.localRankSize;
199 :
200 : void* tmpAivBufferData;
201 : void* tmpAivBufferFlag;
202 :
203 0 : for (u32 i = 0; i < localRankSize; i++) {
204 0 : if (i != localRank) {
205 0 : CHK_RET(outerCommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(tmpAivBufferData)));
206 0 : CHK_RET(outerCommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(tmpAivBufferFlag)));
207 0 : buffersIn[i] = static_cast<u8*>(tmpAivBufferData);
208 0 : buffersOut[i] = static_cast<u8*>(tmpAivBufferFlag) + HCCL_MID_COUNT_32_MB;
209 : } else {
210 0 : buffersIn[i] = static_cast<u8*>(execMem.inputMem.ptr());
211 0 : buffersOut[i] = static_cast<u8*>(execMem.outputMem.ptr()) + HCCL_MID_COUNT_32_MB;
212 : }
213 : }
214 :
215 0 : u32 serverNum = innerCommInfo.localRankSize;
216 :
217 0 : AivOpArgs opArgs{
218 : HcclCMDType::HCCL_CMD_ALLGATHER,
219 0 : execMem.inputPtr,
220 0 : execMem.outputPtr,
221 0 : param.DataDes.count,
222 0 : param.DataDes.dataType,
223 : HCCL_REDUCE_RESERVED,
224 : 0,
225 0 : isOpbase};
226 : AivTopoArgs topoArgs{
227 : localRank,
228 : localRankSize,
229 0 : topoAttr_.isDiffDeviceModule ? topoAttr_.devicePhyId : A_X_SIZE,
230 : 0,
231 : serverNum,
232 0 : topoAttr_.deviceType,
233 0 : algoAttr_.identifier};
234 : u32 numBlocks;
235 0 : CHK_PRT_RET(
236 : CalNumBlocks(numBlocks, localRankSize) != HCCL_SUCCESS, HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
237 : HCCL_E_PARA);
238 0 : numBlocks_ = numBlocks;
239 0 : AivResourceArgs resourceArgs{param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(),
240 0 : numBlocks_, param.aivTag};
241 0 : AivAlgArgs algArgs{0};
242 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
243 0 : algArgs.execTimeOutSet = true;
244 :
245 0 : struct AivProfilingInfo aivProfilingInfo;
246 0 : aivProfilingInfo.counter = opCounter_;
247 :
248 0 : CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
249 0 : HCCL_INFO("[CollAllGatherAivRdmaExecutor][KernelRun]allGather aiv run success.");
250 0 : return HCCL_SUCCESS;
251 0 : }
252 :
253 0 : HcclResult CollAllGatherAivRdmaExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
254 : {
255 0 : if (CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1) != HCCL_SUCCESS) {
256 0 : return HCCL_E_UNAVAIL;
257 : }
258 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
259 0 : u32 ringNum
260 0 : = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
261 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
262 :
263 0 : if (CheckCommSize(COMM_LEVEL1, commIndex + 1) != HCCL_SUCCESS) {
264 0 : return HCCL_E_UNAVAIL;
265 : }
266 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
267 :
268 0 : return HCCL_SUCCESS;
269 0 : }
270 :
271 : HcclResult
272 0 : CollAllGatherAivRdmaExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase>& level1TempAlg, u32 level1RankSize)
273 : {
274 0 : if (level1RankSize > 1) {
275 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING
276 0 : || (topoAttr_.isDiffDeviceModule && topoAttr_.serverNum == 1)) {
277 : level1TempAlg
278 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
279 0 : HCCL_INFO("AllGather mesh: using ring algo inter-server.");
280 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
281 : level1TempAlg
282 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
283 0 : HCCL_INFO("AllGather mesh: using nhr algo inter-server.");
284 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
285 : level1TempAlg
286 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
287 0 : HCCL_INFO("AllGather mesh: using nhr_v1 algo inter-server.");
288 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
289 : level1TempAlg
290 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
291 0 : HCCL_INFO("AllGather mesh: using nonuniform-bruck algo inter-server.");
292 : } else {
293 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
294 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
295 0 : HCCL_INFO("AllGather mesh: using halving-doubling algo inter-server.");
296 : }
297 0 : CHK_SMART_PTR_NULL(level1TempAlg);
298 0 : return HCCL_SUCCESS;
299 : }
300 0 : return HCCL_E_UNAVAIL;
301 : }
302 :
303 : REGISTER_EXEC("AllGatherAivRdmaExecutor", AllGatherAivRdma, CollAllGatherAivRdmaExecutor);
304 :
305 : } // namespace hccl
|
