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_executor.h"
12 :
13 : namespace hccl {
14 0 : CollReduceScatterMeshAivExecutor::CollReduceScatterMeshAivExecutor(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 CollReduceScatterMeshAivExecutor::CalcStreamNum(u32& streamNum)
24 : {
25 0 : streamNum = 0; // AIV通信不需要申请从流
26 0 : HCCL_INFO("[CollReduceScatterMeshAivExecutor][CalcStreamNum] tag[%s] streamNum[%u].",
27 : tag_.c_str(), streamNum);
28 0 : return HCCL_SUCCESS;
29 : }
30 :
31 0 : HcclResult CollReduceScatterMeshAivExecutor::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 CollReduceScatterMeshAivExecutor::CalcTransportMemType(TransportMemType &inputType,
41 : TransportMemType &outputType)
42 : {
43 0 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
44 0 : inputType = TransportMemType::CCL_INPUT;
45 0 : outputType = TransportMemType::AIV_OUTPUT;
46 : } else {
47 0 : inputType = TransportMemType::PARAM_INPUT;
48 0 : outputType = TransportMemType::AIV_OUTPUT;
49 : }
50 0 : HCCL_INFO("[CollReduceScatterMeshAivExecutor][CalcTransportMemType] tag[%s] inputType[%d],"
51 : " outputType[%d].", tag_.c_str(), inputType, outputType);
52 0 : return HCCL_SUCCESS;
53 : }
54 :
55 0 : HcclResult CollReduceScatterMeshAivExecutor::CalcLevel0CommInfo(TransportMemType inputType,
56 : TransportMemType outputType,
57 : std::vector<LevelNSubCommTransport>& opTransport)
58 : {
59 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
60 0 : commParaLevel0.meshSinglePlane = true;
61 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
62 0 : return HCCL_SUCCESS;
63 0 : }
64 :
65 0 : HcclResult CollReduceScatterMeshAivExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
66 : {
67 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
68 :
69 0 : bool isOpBase = (GetWorkflowMode() == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
70 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93 && !isOpBase) {
71 0 : numBlocks = rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3 > MAX_NUM_BLOCKS ?
72 : rankSize * NUM_BLOCKS_THREE_PER_RANK_A3 : rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3;
73 0 : } else if (isOpBase) {
74 0 : numBlocks = NUM_BLOCKS_FACTOR_TWO * rankSize; // 单机场景，单算子ReduceScatter大数据使用2倍 rankSize个aiv
75 : }
76 :
77 0 : u32 bestNumBlocks = numBlocks;
78 0 : CHK_PRT_RET(numBlocks_ < rankSize && topoAttr_.deviceType == DevType::DEV_TYPE_910_93,
79 : HCCL_WARNING("[CollReduceScatterMeshAivExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
80 : numBlocks_, rankSize), HCCL_E_PARA);
81 :
82 0 : CHK_PRT_RET(numBlocks_ < bestNumBlocks && (topoAttr_.deviceType != DevType::DEV_TYPE_910_93 || isOpBase),
83 : HCCL_WARNING("[CollReduceScatterMeshAivExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].",
84 : numBlocks_, bestNumBlocks), HCCL_E_PARA);
85 :
86 0 : if (numBlocks_ < bestNumBlocks) {
87 0 : numBlocks = numBlocks_ / rankSize * rankSize;
88 : }
89 :
90 0 : HCCL_INFO("[CollReduceScatterMeshAivExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]",
91 : numBlocks, numBlocks_, bestNumBlocks);
92 0 : return HCCL_SUCCESS;
93 : }
94 :
95 0 : HcclResult CollReduceScatterMeshAivExecutor::GetAivExecParam(const OpParam& param, AlgResourceResponse& algRes, AivSuperKernelArgs &args)
96 : {
97 0 : HcclUs startut = TIME_NOW();
98 0 : tag_ = param.tag;
99 0 : algResResp_ = &algRes;
100 :
101 0 : HcclResult ret = HCCL_SUCCESS;
102 0 : ExecMem execMem;
103 0 : execMem.count = param.DataDes.count;
104 0 : execMem.inputPtr = param.inputPtr;
105 0 : execMem.outputPtr = param.outputPtr;
106 :
107 : // 单算子大数据量
108 0 : execMem.inputMem = algRes.paramInputMem;
109 0 : execMem.outputMem = algRes.aivOutputMem;
110 :
111 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
112 :
113 0 : u32 localRank = level0CommInfo.localRank;
114 0 : u32 localRankSize = level0CommInfo.localRankSize;
115 0 : HCCL_DEBUG("[CollReduceScatterMeshAivExecutor][GetAivExecParam] userRank [%d] localRank [%d]",
116 : topoAttr_.userRank, localRank);
117 :
118 0 : for (u32 i = 0; i < localRankSize; i++) {
119 0 : if (i != localRank) {
120 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(args.buffersIn[i])));
121 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(args.buffersOut[i])));
122 : } else {
123 0 : args.buffersIn[i] = execMem.inputMem.ptr();
124 0 : args.buffersOut[i] = execMem.outputMem.ptr();
125 : }
126 : }
127 0 : HCCL_INFO("SPK, buffersIn [%p] [%p] [%p] [%p]"
128 : "buffersOut [%p] [%p] [%p] [%p]", args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3],
129 : args.buffersOut[0], args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);
130 0 : args.rank = localRank;
131 0 : args.rankSize = localRankSize;
132 0 : args.len = execMem.count;
133 0 : args.dataType = param.DataDes.dataType;
134 0 : args.unitSize = SIZE_TABLE[param.DataDes.dataType];
135 0 : args.reduceOp = param.reduceType;
136 0 : args.devType = static_cast<u32>(topoAttr_.deviceType);
137 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
138 : HCCL_ERROR("[CollReduceScatterMeshAivExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
139 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
140 0 : HCCL_INFO("SPK [CollReduceScatterMeshAivExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], op[%llu], devType[%u]",
141 : args.rank, args.rankSize, args.len, args.dataType, args.reduceOp, args.devType);
142 :
143 0 : HCCL_INFO("tag[%s], ReduceScatter executor getalgexecparam success, take time [%lld]us.",
144 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
145 0 : return HCCL_SUCCESS;
146 0 : }
147 :
148 0 : HcclResult CollReduceScatterMeshAivExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
149 : {
150 0 : HcclUs startut = TIME_NOW();
151 0 : tag_ = param.tag;
152 0 : algResResp_ = &algRes;
153 :
154 0 : HcclResult ret = HCCL_SUCCESS;
155 0 : ExecMem execMem;
156 0 : execMem.count = param.DataDes.count;
157 0 : execMem.inputPtr = param.inputPtr;
158 0 : execMem.outputPtr = param.outputPtr;
159 :
160 0 : if (workflowMode_ != HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
161 0 : execMem.inputMem = algRes.paramInputMem;
162 0 : execMem.outputMem = algRes.aivOutputMem;
163 0 : ret = KernelRun(param, execMem);
164 : } else {
165 0 : execMem.inputMem = algRes.cclInputMem; // 仍然使用CCLIN
166 0 : execMem.outputMem = algRes.aivOutputMem;
167 0 : ret = KernelRun(param, execMem);
168 : }
169 :
170 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
171 : HCCL_ERROR("[CollReduceScatterMeshAivExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
172 : "run failed", HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
173 :
174 0 : HCCL_INFO("tag[%s], ReduceScatter executor orchestrate success, take time [%lld]us.",
175 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
176 0 : return HCCL_SUCCESS;
177 0 : }
178 :
179 0 : HcclResult CollReduceScatterMeshAivExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
180 : {
181 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] ReduceScatter aiv enter.", __func__);
182 :
183 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
184 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
185 :
186 : void *buffersIn[MAX_RANK_SIZE];
187 : void *buffersOut[MAX_RANK_SIZE];
188 :
189 0 : u32 localRank = level0CommInfo.localRank;
190 0 : u32 localRankSize = level0CommInfo.localRankSize;
191 0 : HCCL_DEBUG("[CollReduceScatterMeshAivExecutor][KernelRun] userRank [%d] localRank [%d]", topoAttr_.userRank, localRank);
192 :
193 0 : for (u32 i = 0; i < localRankSize; i++) {
194 0 : if (i != localRank) {
195 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(buffersIn[i])));
196 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(buffersOut[i])));
197 : } else {
198 0 : buffersIn[i] = execMem.inputMem.ptr();
199 0 : buffersOut[i] = execMem.outputMem.ptr();
200 : }
201 : }
202 : u32 numBlocks;
203 0 : bool isOpbase = (GetWorkflowMode() == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
204 0 : HCCL_DEBUG("[CollReduceScatterMeshAivExecutor][KernelRun]isOpbase is %d", isOpbase);
205 :
206 0 : AivOpArgs opArgs {
207 0 : HcclCMDType::HCCL_CMD_REDUCE_SCATTER, execMem.inputPtr, execMem.outputPtr, execMem.count,
208 0 : param.DataDes.dataType, param.reduceType, 0, isOpbase
209 0 : };
210 0 : AivTopoArgs topoArgs { localRank, localRankSize, MAX_RANK_SIZE, 0, 1, topoAttr_.deviceType};
211 0 : topoArgs.identify = algoAttr_.identifier;
212 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, localRankSize) != HCCL_SUCCESS,
213 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
214 : HCCL_E_PARA);
215 0 : numBlocks_ = numBlocks;
216 : AivResourceArgs resourceArgs {
217 0 : param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(), numBlocks_, param.aivTag
218 0 : };
219 0 : struct AivProfilingInfo aivProfilingInfo;
220 0 : AivAlgArgs algArgs {};
221 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
222 0 : algArgs.execTimeOutSet = true;
223 0 : aivProfilingInfo.counter = opCounter_;
224 :
225 0 : HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
226 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
227 : HCCL_ERROR("[CollReduceScatterMeshAivExecutor][KernelRun]ReduceScatter aiv failed, return[%d]", ret),
228 : ret);
229 :
230 0 : ExtraArgs extraArgs;
231 0 : CHK_RET(SetOpCache(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo, false));
232 :
233 0 : HCCL_INFO("[CollReduceScatterMeshAivExecutor][KernelRun]ReduceScatter aiv run success.");
234 0 : return HCCL_SUCCESS;
235 0 : }
236 :
237 : REGISTER_EXEC("ReduceScatterMeshAivExecutor", ReduceScatterMeshAiv, CollReduceScatterMeshAivExecutor);
238 :
239 : } // namespace hccl
|
