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_mesh_aiv_smallcount_executor.h"
12 :
13 : namespace hccl {
14 0 : CollReduceScatterMeshAivSmallCountExecutor::CollReduceScatterMeshAivSmallCountExecutor(const HcclDispatcher dispatcher,
15 0 : std::unique_ptr<TopoMatcher> &topoMatcher):
16 0 : CollReduceScatterExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = false;
19 0 : desc_.isAivMode = true;
20 0 : desc_.deterministic = 0;
21 0 : }
22 :
23 0 : HcclResult CollReduceScatterMeshAivSmallCountExecutor::CalcStreamNum(u32& streamNum)
24 : {
25 0 : streamNum = 0; // AIV通信不需要申请从流
26 0 : HCCL_INFO("[CollReduceScatterMeshAivSmallCountExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
27 : tag_.c_str(), streamNum);
28 0 : return HCCL_SUCCESS;
29 : }
30 :
31 0 : HcclResult CollReduceScatterMeshAivSmallCountExecutor::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 CollReduceScatterMeshAivSmallCountExecutor::CalcTransportMemType(TransportMemType &inputType,
41 : TransportMemType &outputType)
42 : {
43 0 : inputType = TransportMemType::AIV_INPUT;
44 0 : outputType = TransportMemType::AIV_OUTPUT;
45 0 : HCCL_INFO("[CollReduceScatterMeshAivSmallCountExecutor][CalcTransportMemType] tag[%s] inputType[%d],"
46 : " outputType[%d].", tag_.c_str(), inputType, outputType);
47 0 : return HCCL_SUCCESS;
48 : }
49 :
50 0 : HcclResult CollReduceScatterMeshAivSmallCountExecutor::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 CollReduceScatterMeshAivSmallCountExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
61 : {
62 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
63 :
64 0 : bool isOpBase = (GetWorkflowMode() == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
65 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93 && !isOpBase) {
66 0 : numBlocks = rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3 > MAX_NUM_BLOCKS ?
67 : rankSize * NUM_BLOCKS_THREE_PER_RANK_A3 : rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3;
68 : }
69 :
70 0 : u32 bestNumBlocks = numBlocks;
71 0 : CHK_PRT_RET(numBlocks_ < rankSize && topoAttr_.deviceType == DevType::DEV_TYPE_910_93,
72 : HCCL_WARNING("[CollReduceScatterMeshAivSmallCountExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
73 : numBlocks_, rankSize), HCCL_E_PARA);
74 :
75 0 : CHK_PRT_RET(numBlocks_ < bestNumBlocks && topoAttr_.deviceType != DevType::DEV_TYPE_910_93,
76 : HCCL_WARNING("[CollReduceScatterMeshAivSmallCountExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
77 : numBlocks_, bestNumBlocks), HCCL_E_PARA);
78 :
79 0 : if (numBlocks_ < bestNumBlocks) {
80 0 : numBlocks = numBlocks_ / rankSize * rankSize;
81 : }
82 :
83 0 : HCCL_INFO("[CollReduceScatterMeshAivSmallCountExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]",
84 : numBlocks, numBlocks_, bestNumBlocks);
85 0 : return HCCL_SUCCESS;
86 : }
87 :
88 0 : HcclResult CollReduceScatterMeshAivSmallCountExecutor::GetAivExecParam(const OpParam& param, AlgResourceResponse& algRes, AivSuperKernelArgs &args)
89 : {
90 0 : HcclUs startut = TIME_NOW();
91 0 : tag_ = param.tag;
92 0 : algResResp_ = &algRes;
93 :
94 0 : HcclResult ret = HCCL_SUCCESS;
95 0 : ExecMem execMem;
96 0 : execMem.count = param.DataDes.count;
97 0 : execMem.inputPtr = param.inputPtr;
98 0 : execMem.outputPtr = param.outputPtr;
99 :
100 : // 单算子大数据量
101 0 : execMem.inputMem = algRes.aivInputMem;
102 0 : execMem.outputMem = algRes.aivOutputMem;
103 :
104 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
105 :
106 0 : u32 localRank = level0CommInfo.localRank;
107 0 : u32 localRankSize = level0CommInfo.localRankSize;
108 0 : HCCL_DEBUG("[CollReduceScatterMeshAivSmallCountExecutor][GetAivExecParam] userRank [%d] localRank [%d]",
109 : topoAttr_.userRank, localRank);
110 :
111 0 : for (u32 i = 0; i < localRankSize; i++) {
112 0 : if (i != localRank) {
113 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(args.buffersIn[i])));
114 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(args.buffersOut[i])));
115 : } else {
116 0 : args.buffersIn[i] = execMem.inputMem.ptr();
117 0 : args.buffersOut[i] = execMem.outputMem.ptr();
118 : }
119 : }
120 :
121 0 : HCCL_INFO("SPK, buffersIn [%p] [%p] [%p] [%p]"
122 : "buffersOut [%p] [%p] [%p] [%p]", args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3],
123 : args.buffersOut[0], args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);
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 [CollReduceScatterMeshAivSmallCountExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], op[%llu]", args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);
132 :
133 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
134 : HCCL_ERROR("[CollReduceScatterMeshAivSmallCountExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
135 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
136 :
137 0 : HCCL_INFO("tag[%s], ReduceScatter executor getalgexecparam success, take time [%lld]us.",
138 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
139 0 : return HCCL_SUCCESS;
140 0 : }
141 :
142 0 : HcclResult CollReduceScatterMeshAivSmallCountExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
143 : {
144 0 : HcclUs startut = TIME_NOW();
145 0 : tag_ = param.tag;
146 0 : algResResp_ = &algRes;
147 :
148 0 : HcclResult ret = HCCL_SUCCESS;
149 0 : ExecMem execMem;
150 0 : execMem.count = param.DataDes.count;
151 0 : execMem.inputPtr = param.inputPtr;
152 0 : execMem.outputPtr = param.outputPtr;
153 :
154 : // 单算子大数据量
155 0 : execMem.inputMem = algRes.aivInputMem; // 仍然使用CCLIN
156 0 : execMem.outputMem = algRes.aivOutputMem;
157 0 : execMem.scratchMem = algRes.scratchMem; // 不需要
158 0 : ret = KernelRun(param, execMem);
159 :
160 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
161 : HCCL_ERROR("[CollReduceScatterMeshAivSmallCountExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
162 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
163 :
164 0 : HCCL_INFO("tag[%s], ReduceScatter executor orchestrate success, take time [%lld]us.",
165 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
166 0 : return HCCL_SUCCESS;
167 0 : }
168 :
169 0 : HcclResult CollReduceScatterMeshAivSmallCountExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
170 : {
171 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] ReduceScatter aiv enter", __func__);
172 :
173 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
174 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
175 :
176 : void *buffersIn[MAX_RANK_SIZE];
177 : void *buffersOut[MAX_RANK_SIZE];
178 :
179 0 : u32 localRank = level0CommInfo.localRank;
180 0 : u32 localRankSize = level0CommInfo.localRankSize;
181 0 : HCCL_DEBUG("[CollReduceScatterMeshAivSmallCountExecutor][KernelRun] userRank [%d] localRank [%d]",
182 : topoAttr_.userRank, localRank);
183 :
184 0 : for (u32 i = 0; i < localRankSize; i++) {
185 0 : if (i != localRank) {
186 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(buffersIn[i])));
187 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(buffersOut[i])));
188 : } else {
189 0 : buffersIn[i] = execMem.inputMem.ptr();
190 0 : buffersOut[i] = execMem.outputMem.ptr();
191 : }
192 : }
193 :
194 0 : bool isOpbase = (GetWorkflowMode() == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
195 :
196 0 : AivOpArgs opArgs {
197 0 : HcclCMDType::HCCL_CMD_REDUCE_SCATTER, execMem.inputPtr, execMem.outputPtr, execMem.count,
198 0 : param.DataDes.dataType, param.reduceType, 0, isOpbase
199 0 : };
200 0 : AivTopoArgs topoArgs { localRank, localRankSize, MAX_RANK_SIZE, 0, 1, topoAttr_.deviceType};
201 0 : topoArgs.identify = algoAttr_.identifier;
202 : u32 numBlocks;
203 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, localRankSize) != HCCL_SUCCESS,
204 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
205 : HCCL_E_PARA);
206 0 : numBlocks_ = numBlocks;
207 0 : HCCL_DEBUG("[CollReduceScatterMeshAivSmallCountExecutor][KernelRun]numBlocks is [%u]", numBlocks_);
208 : AivResourceArgs resourceArgs {
209 0 : param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(), numBlocks_, param.aivTag
210 0 : };
211 0 : AivAlgArgs algArgs {};
212 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
213 0 : algArgs.execTimeOutSet = true;
214 0 : struct AivProfilingInfo aivProfilingInfo;
215 0 : aivProfilingInfo.counter = opCounter_;
216 :
217 0 : HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
218 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
219 : HCCL_ERROR("[CollReduceScatterMeshAivSmallCountExecutor][KernelRun]ReduceScatter aiv failed, return[%d]", ret),
220 : ret);
221 :
222 0 : ExtraArgs extraArgs;
223 0 : CHK_RET(SetOpCache(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo, false));
224 :
225 0 : HCCL_INFO("[CollReduceScatterMeshAivSmallCountExecutor][KernelRun]ReduceScatter aiv run success");
226 0 : return HCCL_SUCCESS;
227 0 : }
228 :
229 : REGISTER_EXEC("ReduceScatterMeshAivSmallCountExecutor", ReduceScatterMeshAiv, CollReduceScatterMeshAivSmallCountExecutor);
230 :
231 : } // namespace hccl
|
