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_aiv_rdma_executor.h"
12 : #include "alg_template_register.h"
13 :
14 : namespace hccl {
15 : constexpr u32 A_X_SIZE = 16;
16 :
17 0 : CollReduceScatterAivRdmaExecutor::CollReduceScatterAivRdmaExecutor(const HcclDispatcher dispatcher,
18 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
19 0 : : CollReduceScatterExecutor(dispatcher, topoMatcher)
20 : {
21 0 : DMAReduceFlag_ = false;
22 0 : desc_.isAivMode = true;
23 0 : }
24 :
25 0 : void CollReduceScatterAivRdmaExecutor::ParseParam(const OpParam& param)
26 : {
27 0 : tag_ = param.tag;
28 0 : root_ = param.root;
29 0 : opType_ = param.opType;
30 : // 记录图模式总数据量
31 0 : totalSize_ = topoAttr_.userRankSize * param.DataDes.count * SIZE_TABLE[param.DataDes.dataType];
32 0 : }
33 :
34 0 : HcclResult CollReduceScatterAivRdmaExecutor::CalcScratchMemSize(u64& scratchMemSize)
35 : {
36 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
37 0 : scratchMemSize = 0U;
38 : } else {
39 0 : scratchMemSize = totalSize_;
40 : }
41 0 : HCCL_INFO("[CollReduceScatterAivRdmaExecutor][CalcScratchMemSize] tag[%s] scratchMemSize[%llu]",
42 : tag_.c_str(), scratchMemSize);
43 0 : return HCCL_SUCCESS;
44 : }
45 :
46 0 : HcclResult CollReduceScatterAivRdmaExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
47 : {
48 0 : TransportMemType inputType = TransportMemType::RESERVED;
49 0 : TransportMemType outputType = TransportMemType::RESERVED;
50 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
51 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
52 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
53 0 : return HCCL_SUCCESS;
54 : }
55 :
56 0 : HcclResult CollReduceScatterAivRdmaExecutor::CalcTransportMemType(TransportMemType &inputType,
57 : TransportMemType &outputType)
58 : {
59 : // 使用AIVIN，标记区在AIVIN末尾，单算子模式用CCLOUT，图模式用USEROUT
60 0 : inputType = TransportMemType::AIV_INPUT;
61 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
62 0 : outputType = TransportMemType::CCL_OUTPUT;
63 : }else {
64 0 : outputType = TransportMemType::SCRATCH;
65 : }
66 :
67 0 : HCCL_INFO("[CollReduceScatterAivRdmaExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]",
68 : tag_.c_str(), inputType, outputType);
69 0 : return HCCL_SUCCESS;
70 : }
71 :
72 0 : HcclResult CollReduceScatterAivRdmaExecutor::CalcLevel0CommInfo(TransportMemType inputType,
73 : TransportMemType outputType,
74 : std::vector<LevelNSubCommTransport>& opTransport)
75 : {
76 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
77 0 : commParaLevel0.meshSinglePlane = true;
78 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
79 0 : return HCCL_SUCCESS;
80 0 : }
81 :
82 0 : HcclResult CollReduceScatterAivRdmaExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
83 : {
84 0 : numBlocks = rankSize; // 多机场景，单算子ReduceScatter使用rankSize个aiv
85 0 : u32 bestNumBlocks = numBlocks;
86 :
87 0 : CHK_PRT_RET(numBlocks_ < numBlocks,
88 : HCCL_WARNING("[CollReduceScatterAivRdmaExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
89 : numBlocks_, numBlocks), HCCL_E_PARA);
90 :
91 0 : HCCL_INFO("[CollReduceScatterAivRdmaExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]",
92 : numBlocks, numBlocks_, bestNumBlocks);
93 0 : return HCCL_SUCCESS;
94 : }
95 :
96 0 : HcclResult CollReduceScatterAivRdmaExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
97 : {
98 0 : HCCL_INFO("[CollReduceScatterAivRdmaExecutor][Orchestrate]start");
99 :
100 0 : HcclUs startut = TIME_NOW();
101 0 : tag_ = param.tag;
102 0 : algResResp_ = &algRes;
103 :
104 : // OutputMem单算子模式用CCLOUT，图模式用USEROUT
105 0 : ExecMem execMem;
106 0 : execMem.count = param.DataDes.count;
107 0 : execMem.inputPtr = param.inputPtr;
108 0 : execMem.outputPtr = param.outputPtr;
109 0 : execMem.inputMem = algRes.aivInputMem;
110 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
111 0 : execMem.outputMem = algRes.cclOutputMem;
112 0 : execMem.scratchMem = algRes.cclOutputMem;
113 : } else {
114 0 : execMem.outputMem = algRes.paramOutputMem;
115 0 : execMem.scratchMem = algRes.scratchMem;
116 : }
117 0 : HcclResult ret = KernelRun(param, execMem);
118 :
119 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
120 : HCCL_ERROR("[CollReduceScatterAivRdmaExecutor]errNo[0x%016llx] tag[%s] executor kernel run failed",
121 : HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
122 :
123 0 : HCCL_INFO("tag[%s], ReduceScatter executor orchestrate success, take time [%lld]us.",
124 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
125 0 : return HCCL_SUCCESS;
126 0 : }
127 :
128 0 : HcclResult CollReduceScatterAivRdmaExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
129 : {
130 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollReduceScatterAivRdmaExecutor][KernelRun]ReduceScatter aiv enter");
131 :
132 0 : HcclWorkflowMode workflow = workflowMode_;
133 0 : bool isOpbase = (workflow == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
134 :
135 : // 获取通信域信息
136 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
137 0 : SubCommInfo outerCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
138 0 : u32 commIndex = outerCommInfo.localRank;
139 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
140 0 : SubCommInfo innerCommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
141 :
142 : /* 第一步 节点内重排序RS */
143 : // 数据准备，按照server内rankSize切片
144 0 : u32 perDataSize = SIZE_TABLE[param.DataDes.dataType];
145 0 : u64 perRankSize = param.DataDes.count * perDataSize;
146 0 : std::vector<hccl::LINK> intraLinks = outerCommInfo.links; //机间
147 0 : std::vector<hccl::LINK> interLinks = innerCommInfo.links; //机内
148 0 : u32 intraRankSize = outerCommInfo.localRankSize;
149 0 : u32 intraRankId = outerCommInfo.localRank;
150 :
151 : // reduce scatter阶段，inputMem0-31m做数据区，32M开始后的1M做标记区
152 : void* dataBuffers[MAX_RANK_SIZE];
153 : void* flagBuffers[MAX_RANK_SIZE]; // 标记区的具体偏移在kernel中决定
154 0 : CHK_RET(PrepareAivBuffers(intraRankSize, intraRankId, 0, execMem.inputMem, execMem.inputMem, intraLinks,
155 : dataBuffers, flagBuffers, UserMemType::INPUT_MEM, UserMemType::INPUT_MEM, 0, HCCL_MID_COUNT_32_MB));
156 :
157 0 : u32 serverNum = innerCommInfo.localRankSize;
158 : // 先做本地拷贝到AIVIN再跨片拷贝；output统一为reduceScatterInput的位置，即buffer中原位
159 0 : AivOpArgs opArgs {
160 0 : HcclCMDType::HCCL_CMD_REDUCE_SCATTER, execMem.inputPtr, execMem.outputPtr, execMem.count,
161 0 : param.DataDes.dataType, param.reduceType, 0, isOpbase
162 0 : };
163 : AivTopoArgs topoArgs {
164 0 : intraRankId, intraRankSize, topoAttr_.isDiffDeviceModule ? topoAttr_.devicePhyId : A_X_SIZE,
165 0 : 0, serverNum, topoAttr_.deviceType, algoAttr_.identifier
166 0 : };
167 : u32 numBlocks;
168 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, intraRankSize) != HCCL_SUCCESS,
169 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
170 : HCCL_E_PARA);
171 0 : numBlocks_ = numBlocks;
172 : AivResourceArgs resourceArgs {
173 0 : param.tag, param.stream.ptr(), dataBuffers, flagBuffers, execMem.inputMem.size(), numBlocks_, param.aivTag
174 0 : };
175 0 : AivAlgArgs algArgs {0};
176 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
177 0 : algArgs.execTimeOutSet = true;
178 0 : struct AivProfilingInfo aivProfilingInfo;
179 0 : aivProfilingInfo.counter = opCounter_;
180 :
181 0 : CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
182 : /* 第二步 节点间RS */
183 0 : auto autoSelectedAlgTypeLevel1 = static_cast<u32>(algType_.algoLevel1);
184 0 : ReduceType reduceType = ((param.reduceType != HCCL_REDUCE_PROD) &&
185 0 : (param.DataDes.dataType != HCCL_DATA_TYPE_INT64)) ?
186 : ReduceType::INLINE_REDUCE : ReduceType::TBE_REDUCE;
187 0 : auto opMeta = HcclOpMetaInfo::GetOneForReduceScatter(autoSelectedAlgTypeLevel1,
188 0 : param.DataDes.dataType, reduceType, false, false, CopyPattern::BCOPY, false, 0, true);
189 0 : CHK_RET(InitTask(dispatcher_, const_cast<Stream&>(param.stream), opMeta.isEnableCache, opMeta.GetCacheKey()));
190 :
191 0 : u32 innerRankSize = innerCommInfo.localRankSize;
192 0 : DeviceMem inputMem = execMem.inputMem;
193 0 : if (innerRankSize > 1) {
194 : //execMem.inputMem要改成按平面制定的初始位置
195 0 : u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
196 0 : std::unique_ptr<ExecutorBase> innerExecutor;
197 0 : std::vector<Slice> dataSegsSlice;
198 0 : dataSegsSlice.resize(innerRankSize);
199 0 : for (u32 i = 0; i < innerRankSize; i++) {
200 0 : dataSegsSlice[i].size = perRankSize;
201 0 : dataSegsSlice[i].offset = (commIndex * innerRankSize + i) * perRankSize;
202 : }
203 0 : u64 count = param.DataDes.count;
204 0 : u64 baseOffset = 0;
205 0 : DeviceMem inputMem = execMem.inputMem;
206 0 : DeviceMem scratchMem = execMem.scratchMem;
207 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
208 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
209 0 : TemplateType::TEMPLATE_REDUCESCATTER_RING, dispatcher_);
210 0 : CHK_SMART_PTR_NULL(innerExecutor);
211 0 : CHK_RET(innerExecutor->Prepare(reduceAttr));
212 0 : HCCL_INFO("ReduceScatter mesh: using ring algo inter-server.");
213 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
214 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
215 0 : TemplateType::TEMPLATE_REDUCESCATTER_NHR, dispatcher_);
216 0 : CHK_SMART_PTR_NULL(innerExecutor);
217 0 : CHK_RET(innerExecutor->Prepare(reduceAttr, false));
218 0 : innerExecutor->CloseBarrier();
219 0 : HCCL_INFO("ReduceScatter mesh: using nhr algo inter-server.");
220 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
221 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
222 0 : TemplateType::TEMPLATE_REDUCESCATTER_NHR_V1, dispatcher_);
223 0 : CHK_SMART_PTR_NULL(innerExecutor);
224 0 : CHK_RET(innerExecutor->Prepare(reduceAttr));
225 0 : HCCL_INFO("ReduceScatter mesh: using nhr_v1 algo inter-server.");
226 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
227 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
228 0 : TemplateType::TEMPLATE_REDUCESCATTER_NB, dispatcher_);
229 0 : CHK_SMART_PTR_NULL(innerExecutor);
230 0 : CHK_RET(innerExecutor->Prepare(reduceAttr));
231 0 : HCCL_INFO("ReduceScatter mesh: using nonuniform-bruck algo inter-server.");
232 : } else {
233 0 : count = count * innerRankSize;
234 0 : baseOffset = commIndex * innerRankSize * perRankSize;
235 0 : inputMem = execMem.inputMem.range(commIndex * innerRankSize * perRankSize, perRankSize * innerRankSize);
236 0 : scratchMem = execMem.scratchMem.range(commIndex * innerRankSize * perRankSize, perRankSize * innerRankSize);
237 0 : innerExecutor = AlgTemplateRegistry::Instance().GetAlgTemplate(
238 0 : TemplateType::TEMPLATE_REDUCESCATTER_RECURSIVE_HD, dispatcher_);
239 0 : CHK_SMART_PTR_NULL(innerExecutor);
240 0 : CHK_RET(innerExecutor->Prepare(reduceAttr));
241 0 : HCCL_INFO("ReduceScatter mesh: using halving-doubling algo inter-server.");
242 : }
243 0 : CHK_RET(innerExecutor->Prepare(inputMem, inputMem, scratchMem, count,
244 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, dataSegsSlice, baseOffset));
245 0 : CHK_RET(innerExecutor->RegisterProfiler(
246 : (innerRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + innerCommInfo.localRank,
247 : PROF_STAGE_0, HCCL_EXEC_STEP_NOT_SET, param.stream));
248 :
249 0 : CHK_RET(RunTemplate(innerExecutor, innerCommInfo));
250 0 : HCCL_INFO("[CollReduceScatterAivRdmaExecutor] rdma stage run success.");
251 0 : }
252 : /* 第三步 最后D2D拷贝 */
253 :
254 : // 如果使用CCL buffer，需要将CCL buffer in中的结果拷贝到user buffer out
255 0 : DeviceMem srcMem = execMem.inputMem.range(perRankSize * (commIndex * serverNum + innerCommInfo.localRank), perRankSize);
256 0 : DeviceMem dstMem = DeviceMem::create(execMem.outputPtr, perRankSize);
257 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstMem, srcMem, const_cast<Stream&>(param.stream)));
258 :
259 0 : CHK_RET(LaunchTask(dispatcher_, const_cast<Stream&>(param.stream)));
260 :
261 0 : HCCL_INFO("[CollReduceScatterAivRdmaExecutor][KernelRun]ReduceScatter aiv run success");
262 0 : return HCCL_SUCCESS;
263 0 : }
264 :
265 : REGISTER_EXEC("ReduceScatterAivRdmaExecutor", ReduceScatterAivRdma, CollReduceScatterAivRdmaExecutor);
266 :
267 : } // namespace hccl
|
