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_mesh_aiv_for_910_93_executor.h"
12 :
13 : namespace hccl {
14 0 : CollAllGatherMeshAivFor91093Executor::CollAllGatherMeshAivFor91093Executor(const HcclDispatcher dispatcher,
15 0 : std::unique_ptr<TopoMatcher> &topoMatcher):
16 0 : CollAllGatherExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = false;
19 0 : desc_.isAivMode = true;
20 0 : desc_.isAivCrossNode = true;
21 0 : }
22 :
23 0 : HcclResult CollAllGatherMeshAivFor91093Executor::CalcStreamNum(u32& streamNum)
24 : {
25 0 : streamNum = 0; // AIV通信不需要申请从流
26 0 : HCCL_INFO("[CollAllGatherMeshAivFor91093Executor][CalcStreamNum] tag[%s] streamNum[%u].",
27 : tag_.c_str(), streamNum);
28 0 : return HCCL_SUCCESS;
29 : }
30 :
31 0 : HcclResult CollAllGatherMeshAivFor91093Executor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
32 : {
33 0 : TransportMemType inputType = TransportMemType::RESERVED;
34 0 : TransportMemType outputType = TransportMemType::RESERVED;
35 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
36 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
37 0 : return HCCL_SUCCESS;
38 : }
39 :
40 0 : HcclResult CollAllGatherMeshAivFor91093Executor::CalcTransportMemType(TransportMemType &inputType,
41 : TransportMemType &outputType)
42 : {
43 0 : inputType = (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE ?
44 : TransportMemType::CCL_INPUT : TransportMemType::PARAM_INPUT);
45 0 : outputType = TransportMemType::AIV_OUTPUT;
46 0 : HCCL_INFO("[CollAllGatherMeshAivFor91093Executor][CalcTransportMemType] tag[%s] inputType[%d],"
47 : " outputType[%d].", tag_.c_str(), inputType, outputType);
48 0 : return HCCL_SUCCESS;
49 : }
50 :
51 0 : HcclResult CollAllGatherMeshAivFor91093Executor::CalcLevel0CommInfo(TransportMemType inputType,
52 : TransportMemType outputType,
53 : std::vector<LevelNSubCommTransport>& opTransport)
54 : {
55 0 : CommParaInfo commCombinePara(COMM_COMBINE_ORDER, CommType::COMM_TAG_MESH);
56 0 : commCombinePara.meshSinglePlane = true;
57 0 : CHK_RET(CalcCommPlaneInfo(tag_, commCombinePara, opTransport[COMM_COMBINE_ORDER], inputType, outputType));
58 :
59 0 : LevelNSubCommTransport &commTransportLevel0 = opTransport[COMM_COMBINE_ORDER];
60 0 : for (u32 subCommIndex = 0; subCommIndex < commTransportLevel0.size(); subCommIndex++) {
61 0 : for (auto &transportRequest : commTransportLevel0[subCommIndex].transportRequests) {
62 0 : transportRequest.isUsedRdma = false;
63 : }
64 : }
65 0 : return HCCL_SUCCESS;
66 0 : }
67 :
68 0 : HcclResult CollAllGatherMeshAivFor91093Executor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
69 : {
70 0 : u32 minNumBlocks = (rankSize + MAX_TARGET_NUM - 1) / MAX_TARGET_NUM;
71 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
72 :
73 : // A3超节点内多机场景
74 : // 多核并行优化
75 0 : if (rankSize > HALF_MAX_NUM_BLOCKS || dataSize < AIV_A3_CROSSNODE_TINY_SIZE) {
76 0 : numBlocks = rankSize;
77 0 : } else if (rankSize > ONE_THIRD_MAX_NUM_BLOCKS || dataSize < AIV_A3_CROSSNODE_SMALL_SIZE) {
78 0 : numBlocks = rankSize * NUM_BLOCKS_FACTOR_TWO;
79 0 : } else if (rankSize > ONE_FOURTH_MAX_NUM_BLOCKS) {
80 0 : numBlocks = rankSize * NUM_BLOCKS_FACTOR_THREE;
81 : } else {
82 0 : numBlocks = rankSize * NUM_BLOCKS_FACTOR_FOUR;
83 : }
84 0 : HCCL_DEBUG("[CollAllGatherMeshAivFor91093Executor][CalNumBlocks]aivCore at least is [%u]", minNumBlocks);
85 0 : u32 bestNumBlocks = numBlocks;
86 0 : numBlocks = numBlocks_ < rankSize ? numBlocks_ : (numBlocks_ < numBlocks ? numBlocks_ / rankSize * rankSize : numBlocks);
87 :
88 0 : CHK_PRT_RET(numBlocks < minNumBlocks,
89 : HCCL_ERROR("[CollAllGatherMeshAivFor91093Executor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
90 : numBlocks_, minNumBlocks),
91 : HCCL_E_PARA);
92 :
93 0 : HCCL_INFO("[CollAllGatherMeshAivFor91093Executor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u].",
94 : numBlocks, numBlocks_, bestNumBlocks);
95 0 : return HCCL_SUCCESS;
96 : }
97 :
98 0 : HcclResult CollAllGatherMeshAivFor91093Executor::GetAivExecParam(const OpParam& param, AlgResourceResponse& algRes, AivSuperKernelArgs &args)
99 : {
100 0 : HcclUs startut = TIME_NOW();
101 0 : tag_ = param.tag;
102 0 : algResResp_ = &algRes;
103 :
104 0 : HcclResult ret = HCCL_SUCCESS;
105 0 : ExecMem execMem;
106 0 : execMem.count = param.DataDes.count;
107 0 : execMem.inputPtr = param.inputPtr;
108 0 : execMem.outputPtr = param.outputPtr;
109 :
110 0 : execMem.inputMem = algRes.paramInputMem;
111 0 : execMem.outputMem = algRes.aivOutputMem;
112 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_COMBINE_ORDER, COMM_INDEX_0);
113 :
114 0 : u32 localRank = level0CommInfo.localRank;
115 0 : u32 localRankSize = level0CommInfo.localRankSize;
116 0 : HCCL_DEBUG("[CollAllGatherMeshAivFor91093Executor][GetAivExecParam] userRank [%d] localRank [%d]",
117 : topoAttr_.userRank, localRank);
118 :
119 0 : args.buffersIn[0] = execMem.inputMem.ptr();
120 0 : args.buffersOut[0] = execMem.outputMem.ptr();
121 0 : constexpr u32 BUFFER_IDX_ONE = 1;
122 0 : args.buffersOut[BUFFER_IDX_ONE] = algRes.aivCommInfoMem.ptr(); // 通信域信息
123 :
124 0 : args.rank = localRank;
125 0 : args.rankSize = localRankSize;
126 0 : args.len = execMem.count;
127 0 : args.dataType = param.DataDes.dataType;
128 0 : args.unitSize = SIZE_TABLE[param.DataDes.dataType];
129 0 : args.reduceOp = param.reduceType;
130 0 : args.devType = static_cast<u32>(topoAttr_.deviceType);
131 0 : HCCL_INFO("SPK [CollAllGatherMeshAivFor91093Executor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], op[%llu]",
132 : args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);
133 :
134 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
135 : HCCL_ERROR("[CollAllGatherMeshAivFor91093Executor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
136 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
137 :
138 0 : HCCL_INFO("tag[%s], AllGather executor getalgexecparam success, take time [%lld]us.",
139 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
140 0 : return HCCL_SUCCESS;
141 0 : }
142 :
143 0 : HcclResult CollAllGatherMeshAivFor91093Executor::PrepareCommInfoToDevice(AlgResourceResponse& algResource)
144 : {
145 0 : HCCL_INFO("[CollAllGatherMeshAivFor91093Executor][PrepareCommInfoToDevice]AllGather aiv copy comm info to device.");
146 0 : CHK_RET(CopyAivCommInfoToDevice(COMM_COMBINE_ORDER, COMM_INDEX_0, algResource));
147 0 : return HCCL_SUCCESS;
148 : }
149 :
150 0 : HcclResult CollAllGatherMeshAivFor91093Executor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
151 : {
152 0 : HcclUs startut = TIME_NOW();
153 0 : tag_ = param.tag;
154 0 : algResResp_ = &algRes;
155 :
156 0 : ExecMem execMem;
157 0 : execMem.count = param.DataDes.count;
158 0 : execMem.inputPtr = param.inputPtr;
159 0 : execMem.outputPtr = param.outputPtr;
160 :
161 0 : execMem.inputMem = (workflowMode_ != HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE ?
162 0 : algRes.paramInputMem : algRes.cclInputMem);
163 0 : execMem.outputMem = algRes.aivOutputMem;
164 0 : HcclResult ret = KernelRun(param, execMem);
165 :
166 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
167 : HCCL_ERROR("[CollAllGatherMeshAivFor91093Executor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
168 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
169 :
170 0 : HCCL_INFO("tag[%s], AllGather executor orchestrate success, take time [%lld]us.",
171 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
172 0 : return HCCL_SUCCESS;
173 0 : }
174 :
175 0 : HcclResult CollAllGatherMeshAivFor91093Executor::KernelRun(const OpParam ¶m, ExecMem &execMem)
176 : {
177 0 : HCCL_CONFIG_INFO(HCCL_ALG,"[CollAllGatherMeshAivFor91093Executor][KernelRun]AllGather aiv enter.");
178 :
179 0 : CHK_RET(CheckCommSize(COMM_COMBINE_ORDER, COMM_INDEX_0 + 1));
180 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_COMBINE_ORDER, COMM_INDEX_0);
181 :
182 : void *buffersIn[MAX_RANK_SIZE];
183 : void *buffersOut[MAX_RANK_SIZE];
184 :
185 0 : u32 localRank = level0CommInfo.localRank;
186 0 : u32 localRankSize = level0CommInfo.localRankSize;
187 0 : HCCL_DEBUG("[CollAllGatherMeshAivFor91093Executor][KernelRun] userRank [%d] localRank [%d].",
188 : topoAttr_.userRank, localRank);
189 :
190 0 : buffersIn[0] = execMem.inputMem.ptr();
191 0 : buffersOut[0] = execMem.outputMem.ptr();
192 0 : constexpr u32 BUFFER_IDX_ONE = 1;
193 0 : buffersOut[BUFFER_IDX_ONE] = algResResp_->aivCommInfoMem.ptr(); // 通信域信息
194 :
195 0 : bool isOpbase = (GetWorkflowMode() == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
196 0 : AivOpArgs opArgs {
197 0 : HcclCMDType::HCCL_CMD_ALLGATHER, execMem.inputPtr, execMem.outputPtr, execMem.count,
198 0 : param.DataDes.dataType, param.reduceType, 0, isOpbase
199 0 : };
200 :
201 0 : AivTopoArgs topoArgs { localRank, localRankSize, MAX_RANK_SIZE, 0, topoAttr_.serverNum, topoAttr_.deviceType, algoAttr_.identifier };
202 : u32 numBlocks;
203 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, localRankSize, opArgs.count * SIZE_TABLE[opArgs.dataType]) != HCCL_SUCCESS,
204 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
205 : HCCL_E_PARA);
206 0 : numBlocks_ = numBlocks;
207 : AivResourceArgs resourceArgs {
208 0 : param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(), numBlocks_, param.aivTag
209 0 : };
210 0 : AivAlgArgs algArgs {};
211 0 : algArgs.argsType = KernelArgsType::ARGS_TYPE_SIMPLE;
212 0 : if(numBlocks_ >= localRankSize) {
213 0 : algArgs.step = localRankSize;
214 : } else {
215 0 : algArgs.step = numBlocks_;
216 : }
217 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
218 0 : algArgs.execTimeOutSet = true;
219 0 : struct AivProfilingInfo aivProfilingInfo;
220 0 : aivProfilingInfo.counter = opCounter_;
221 :
222 0 : HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
223 :
224 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
225 : HCCL_ERROR("[CollAllGatherMeshAivFor91093Executor][KernelRun]AllGather aiv failed, return[%d]", ret),
226 : ret);
227 :
228 0 : ExtraArgs extraArgs;
229 0 : CHK_RET(SetOpCache(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo, true));
230 :
231 0 : HCCL_INFO("[CollAllGatherMeshAivFor91093Executor][KernelRun]AllGather aiv run success");
232 0 : return HCCL_SUCCESS;
233 0 : }
234 :
235 : REGISTER_EXEC("AllGatherMeshAivFor91093Executor", AllGatherMeshAivFor91093, CollAllGatherMeshAivFor91093Executor);
236 :
237 : } // namespace hccl
|