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_reduce_mesh_aiv_smallcount_executor.h"
12 :
13 : namespace hccl {
14 :
15 0 : CollAllReduceMeshAivSmallCountExecutor::CollAllReduceMeshAivSmallCountExecutor(const HcclDispatcher dispatcher,
16 0 : std::unique_ptr<TopoMatcher> &topoMatcher): CollAllReduceExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = false;
19 0 : desc_.isAivMode = true;
20 0 : desc_.deterministic = 0;
21 0 : }
22 :
23 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::CalcStreamNum(u32& streamNum)
24 : {
25 0 : streamNum = 0; // AIV通信不需要申请从流
26 0 : HCCL_INFO("[CollAllReduceMeshAivSmallCountExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
27 : tag_.c_str(), streamNum);
28 0 : return HCCL_SUCCESS;
29 : }
30 :
31 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::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 CollAllReduceMeshAivSmallCountExecutor::CalcTransportMemType(TransportMemType &inputType,
41 : TransportMemType &outputType)
42 : {
43 0 : inputType = TransportMemType::AIV_INPUT;
44 0 : outputType = TransportMemType::AIV_OUTPUT;
45 0 : HCCL_INFO("[CollAllReduceMeshAivSmallCountExecutor][CalcTransportMemType] tag[%s] inputType[%d],"
46 : " outputType[%d].", tag_.c_str(), inputType, outputType);
47 0 : return HCCL_SUCCESS;
48 : }
49 :
50 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::CalcLevel0CommInfo(TransportMemType inputType,
51 : TransportMemType outputType,
52 : std::vector<LevelNSubCommTransport>& opTransport)
53 : {
54 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
55 0 : commParaLevel0.meshSinglePlane = true;
56 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
57 0 : return HCCL_SUCCESS;
58 0 : }
59 :
60 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
61 : {
62 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
63 0 : u32 bestNumBlocks = numBlocks;
64 :
65 0 : CHK_PRT_RET(numBlocks_ < numBlocks,
66 : HCCL_WARNING("[CollAllReduceMeshAivSmallCountExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
67 : numBlocks_, numBlocks), HCCL_E_PARA);
68 :
69 0 : HCCL_INFO("[CollAllReduceMeshAivSmallCountExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]",
70 : numBlocks, numBlocks_, bestNumBlocks);
71 0 : return HCCL_SUCCESS;
72 : }
73 :
74 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::GetAivExecParam(const OpParam& param, AlgResourceResponse& algRes, AivSuperKernelArgs &args)
75 : {
76 0 : HcclUs startut = TIME_NOW();
77 0 : tag_ = param.tag;
78 0 : algResResp_ = &algRes;
79 :
80 0 : HcclResult ret = HCCL_SUCCESS;
81 0 : ExecMem execMem;
82 0 : execMem.count = param.DataDes.count;
83 0 : execMem.inputPtr = param.inputPtr;
84 0 : execMem.outputPtr = param.outputPtr;
85 :
86 : // 单算子大数据量
87 0 : execMem.inputMem = algRes.aivInputMem;
88 0 : execMem.outputMem = algRes.aivOutputMem;
89 :
90 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
91 :
92 0 : u32 localRank = level0CommInfo.localRank;
93 0 : u32 localRankSize = level0CommInfo.localRankSize;
94 0 : HCCL_DEBUG("[CollAllReduceMeshAivSmallCountExecutor][GetAivExecParam] userRank [%d] localRank [%d]",
95 : topoAttr_.userRank, localRank);
96 :
97 0 : for (u32 i = 0; i < localRankSize; i++) {
98 0 : if (i != localRank) {
99 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(args.buffersIn[i])));
100 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(args.buffersOut[i])));
101 : } else {
102 0 : args.buffersIn[i] = execMem.inputMem.ptr();
103 0 : args.buffersOut[i] = execMem.outputMem.ptr();
104 : }
105 : }
106 :
107 0 : HCCL_INFO("SPK, buffersIn [%p] [%p] [%p] [%p]"
108 : "buffersOut [%p] [%p] [%p] [%p]", args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3],
109 : args.buffersOut[0], args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);
110 0 : args.rank = localRank;
111 0 : args.rankSize = localRankSize;
112 0 : args.len = execMem.count;
113 0 : args.dataType = param.DataDes.dataType;
114 0 : args.unitSize = SIZE_TABLE[param.DataDes.dataType];
115 0 : args.reduceOp = param.reduceType;
116 0 : args.devType = static_cast<u32>(topoAttr_.deviceType);
117 0 : HCCL_INFO("SPK [CollAllReduceMeshAivSmallCountExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], op[%llu]", args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);
118 :
119 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
120 : HCCL_ERROR("[CollAllReduceMeshAivSmallCountExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
121 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
122 :
123 0 : HCCL_INFO("tag[%s], AllReduce executor getalgexecparam 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 CollAllReduceMeshAivSmallCountExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
129 : {
130 0 : HcclUs startut = TIME_NOW();
131 0 : tag_ = param.tag;
132 0 : algResResp_ = &algRes;
133 :
134 0 : ExecMem execMem;
135 0 : HcclResult ret = HCCL_SUCCESS;
136 0 : execMem.count = param.DataDes.count;
137 0 : execMem.inputPtr = param.inputPtr;
138 0 : execMem.outputPtr = param.outputPtr;
139 :
140 : // 小数据量都用aiv
141 0 : execMem.inputMem = algRes.aivInputMem;
142 0 : execMem.outputMem = algRes.aivOutputMem;
143 0 : execMem.scratchMem = algRes.scratchMem; // 不需要
144 0 : ret = KernelRun(param, execMem);
145 :
146 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
147 : HCCL_ERROR("[CollAllReduceMeshAivSmallCountExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
148 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
149 :
150 0 : HCCL_INFO("tag[%s], AllReduce executor orchestrate success, take time [%lld]us.",
151 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
152 0 : return HCCL_SUCCESS;
153 0 : }
154 :
155 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& adjInfo)
156 : {
157 0 : return HCCL_SUCCESS;
158 : }
159 :
160 0 : HcclResult CollAllReduceMeshAivSmallCountExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
161 : {
162 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] AllReduce aiv enter.", __func__);
163 :
164 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
165 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
166 :
167 : void *buffersIn[MAX_RANK_SIZE];
168 : void *buffersOut[MAX_RANK_SIZE];
169 :
170 0 : u32 localRank = level0CommInfo.localRank;
171 0 : u32 localRankSize = level0CommInfo.localRankSize;
172 0 : HCCL_DEBUG("[CollAllReduceMeshAivSmallCountExecutor][KernelRun] userRank [%u] localRank [%u]", topoAttr_.userRank, localRank);
173 :
174 0 : for (u32 i = 0; i < localRankSize; i++) {
175 0 : if (i != localRank) {
176 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(buffersIn[i])));
177 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(buffersOut[i])));
178 : } else {
179 0 : buffersIn[i] = execMem.inputMem.ptr();
180 0 : buffersOut[i] = execMem.outputMem.ptr();
181 : }
182 : }
183 :
184 0 : bool isOpbase = (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
185 0 : AivOpArgs opArgs {
186 0 : HcclCMDType::HCCL_CMD_ALLREDUCE, execMem.inputPtr, execMem.outputPtr, execMem.count,
187 0 : param.DataDes.dataType, param.reduceType, 0, isOpbase
188 0 : };
189 0 : AivTopoArgs topoArgs { localRank, localRankSize, MAX_RANK_SIZE, 0, 1, topoAttr_.deviceType, algoAttr_.identifier};
190 : u32 numBlocks;
191 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, localRankSize) != HCCL_SUCCESS,
192 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
193 : HCCL_E_PARA);
194 0 : numBlocks_ = numBlocks;
195 : AivResourceArgs resourceArgs {
196 0 : param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(), numBlocks_, param.aivTag
197 0 : };
198 0 : AivAlgArgs algArgs {};
199 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
200 0 : algArgs.execTimeOutSet = true;
201 0 : struct AivProfilingInfo aivProfilingInfo;
202 0 : aivProfilingInfo.counter = opCounter_;
203 0 : HCCL_DEBUG("[CollAllReduceMeshAivSmallCountExecutor][KernelRun]numBlocks is %u", numBlocks_);
204 :
205 0 : HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
206 :
207 0 : ExtraArgs extraArgs;
208 0 : CHK_RET(SetOpCache(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo, false));
209 :
210 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
211 : HCCL_ERROR("[CollAllReduceMeshAivSmallCountExecutor][KernelRun]AllReduce aiv failed, return[%d]", ret), ret);
212 :
213 0 : HCCL_INFO("[CollAllReduceMeshAivSmallCountExecutor][KernelRun]AllReduce aiv run success.");
214 0 : return HCCL_SUCCESS;
215 0 : }
216 :
217 : REGISTER_EXEC("AllReduceMeshAivSmallCountExecutor", AllReduceMeshAivSmallCount,
218 : CollAllReduceMeshAivSmallCountExecutor);
219 :
220 : } // namespace hccl
|
