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_deter_executor.h"
12 :
13 : namespace hccl {
14 :
15 0 : CollReduceScatterAivDeterExecutor::CollReduceScatterAivDeterExecutor(const HcclDispatcher dispatcher,
16 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
17 0 : : CollReduceScatterExecutor(dispatcher, topoMatcher)
18 : {
19 0 : DMAReduceFlag_ = false;
20 0 : desc_.isAivMode = true;
21 0 : }
22 :
23 0 : HcclResult CollReduceScatterAivDeterExecutor::CalcStreamNum(u32& streamNum)
24 : {
25 0 : streamNum = 0; // AIV通信不需要申请从流
26 0 : HCCL_INFO("[CollReduceScatterAivDeterExecutor][CalcStreamNum] tag[%s] streamNum[%u].", tag_.c_str(), streamNum);
27 0 : return HCCL_SUCCESS;
28 : }
29 :
30 0 : HcclResult CollReduceScatterAivDeterExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
31 : {
32 0 : TransportMemType inputType = TransportMemType::RESERVED;
33 0 : TransportMemType outputType = TransportMemType::RESERVED;
34 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
35 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
36 0 : return HCCL_SUCCESS;
37 : }
38 :
39 0 : HcclResult CollReduceScatterAivDeterExecutor::CalcTransportMemType(TransportMemType &inputType, TransportMemType &outputType)
40 : {
41 0 : inputType = TransportMemType::CCL_INPUT;
42 0 : outputType = TransportMemType::AIV_OUTPUT;
43 0 : HCCL_INFO("[CollReduceScatterAivDeterExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d].",
44 : tag_.c_str(), inputType, outputType);
45 0 : return HCCL_SUCCESS;
46 : }
47 :
48 0 : HcclResult CollReduceScatterAivDeterExecutor::CalcLevel0CommInfo(TransportMemType inputType,
49 : TransportMemType outputType,
50 : std::vector<LevelNSubCommTransport>& opTransport)
51 : {
52 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
53 0 : commParaLevel0.meshSinglePlane = true;
54 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
55 0 : return HCCL_SUCCESS;
56 0 : }
57 :
58 0 : HcclResult CollReduceScatterAivDeterExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
59 : {
60 0 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
61 :
62 0 : numBlocks = NUM_BLOCKS_FACTOR_TWO * rankSize; // 小数据量使用2倍rankSize的AIV core
63 0 : if (dataSize*rankSize >= AIV_REDUCE_SCATTER_DETER_SMALL_SIZE) {
64 0 : numBlocks = NUM_BLOCKS_FACTOR_THREE * rankSize;
65 : }
66 :
67 0 : HCCL_INFO("[CollReduceScatterAivDeterExecutor][CalNumBlocks] datasize is [%u], numBlocks is set to [%u]", dataSize, numBlocks);
68 0 : return HCCL_SUCCESS;
69 : }
70 :
71 0 : HcclResult CollReduceScatterAivDeterExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
72 : {
73 0 : HcclUs startut = TIME_NOW();
74 0 : tag_ = param.tag;
75 0 : algResResp_ = &algRes;
76 :
77 0 : HcclResult ret = HCCL_SUCCESS;
78 0 : ExecMem execMem;
79 0 : execMem.count = param.DataDes.count;
80 0 : execMem.inputPtr = param.inputPtr;
81 0 : execMem.outputPtr = param.outputPtr;
82 : // 确定性场景不支持图模式
83 0 : execMem.inputMem = algRes.cclInputMem;
84 0 : execMem.outputMem = algRes.aivOutputMem;
85 :
86 0 : ret = KernelRun(param, execMem);
87 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
88 : HCCL_ERROR("[CollReduceScatterAivDeterExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel run failed",
89 : HCCL_ERROR_CODE(ret), param.tag.c_str()), ret);
90 :
91 0 : HCCL_INFO("tag[%s], ReduceScatter executor orchestrate success, take time [%lld]us",
92 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
93 0 : return HCCL_SUCCESS;
94 0 : }
95 :
96 0 : HcclResult CollReduceScatterAivDeterExecutor::GetAivExecParam(const OpParam& param, AlgResourceResponse& algRes, AivSuperKernelArgs &args)
97 : {
98 0 : HcclUs startut = TIME_NOW();
99 0 : tag_ = param.tag;
100 0 : algResResp_ = &algRes;
101 :
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.cclInputMem;
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("[CollReduceScatterAivDeterExecutor][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 :
128 0 : HCCL_INFO("SPK, buffersIn [%p] [%p] [%p] [%p]"
129 : "buffersOut [%p] [%p] [%p] [%p]", args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3],
130 : args.buffersOut[0], args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);
131 0 : args.rank = localRank;
132 0 : args.rankSize = localRankSize;
133 0 : args.len = execMem.count;
134 0 : args.dataType = param.DataDes.dataType;
135 0 : args.unitSize = SIZE_TABLE[param.DataDes.dataType];
136 0 : args.reduceOp = param.reduceType;
137 0 : args.devType = static_cast<u32>(topoAttr_.deviceType);
138 0 : HCCL_INFO("SPK [CollReduceScatterAivDeterExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], op[%llu]", args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);
139 :
140 0 : HCCL_INFO("tag[%s], ReduceScatter executor getalgexecparam success, take time [%lld]us.",
141 : param.tag.c_str(), DURATION_US(TIME_NOW() - startut));
142 0 : return HCCL_SUCCESS;
143 0 : }
144 :
145 0 : HcclResult CollReduceScatterAivDeterExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
146 : {
147 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollReduceScatterAivDeterExecutor][KernelRun]ReduceScatter aiv enter.");
148 :
149 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
150 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
151 :
152 : void *buffersIn[MAX_RANK_SIZE];
153 : void *buffersOut[MAX_RANK_SIZE];
154 :
155 0 : u32 localRank = level0CommInfo.localRank;
156 0 : u32 localRankSize = level0CommInfo.localRankSize;
157 0 : HCCL_DEBUG("[CollReduceScatterAivDeterExecutor][KernelRun] userRank [%u] localRank [%u]", topoAttr_.userRank, localRank);
158 :
159 0 : for (u32 i = 0; i < localRankSize; i++) {
160 0 : if (i != localRank) {
161 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(buffersIn[i])));
162 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(buffersOut[i])));
163 : } else {
164 0 : buffersIn[i] = execMem.inputMem.ptr();
165 0 : buffersOut[i] = execMem.outputMem.ptr();
166 : }
167 : }
168 :
169 0 : bool isOpbase = (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
170 0 : AivOpArgs opArgs {
171 0 : HcclCMDType::HCCL_CMD_REDUCE_SCATTER, execMem.inputPtr, execMem.outputPtr, execMem.count,
172 0 : param.DataDes.dataType, param.reduceType, 0, isOpbase
173 0 : };
174 0 : AivTopoArgs topoArgs { localRank, localRankSize, MAX_RANK_SIZE, 0, 1, topoAttr_.deviceType };
175 0 : topoArgs.identify = algoAttr_.identifier;
176 :
177 0 : u64 dataSize = SIZE_TABLE[param.DataDes.dataType] * execMem.count;
178 : u32 numBlocks;
179 0 : CHK_PRT_RET(CalNumBlocks(numBlocks, localRankSize, dataSize) != HCCL_SUCCESS,
180 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__),
181 : HCCL_E_PARA);
182 0 : numBlocks_ = numBlocks;
183 : AivResourceArgs resourceArgs {
184 0 : param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(), numBlocks_, param.aivTag
185 0 : };
186 0 : AivAlgArgs algArgs {};
187 0 : algArgs.deterministic = 1;
188 0 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
189 0 : algArgs.execTimeOutSet = true;
190 0 : struct AivProfilingInfo aivProfilingInfo;
191 0 : aivProfilingInfo.counter = opCounter_;
192 0 : HCCL_INFO("[CollReduceScatterAivDeterExecutor][KernelRun]ReduceScatter bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());
193 :
194 0 : HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
195 :
196 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
197 : HCCL_ERROR("[CollReduceScatterAivDeterExecutor][KernelRun]ReduceScatter aiv failed, return[%d]", ret), ret);
198 :
199 0 : ExtraArgs extraArgs;
200 0 : CHK_RET(SetOpCache(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo, false));
201 :
202 0 : HCCL_INFO("[CollReduceScatterAivDeterExecutor][KernelRun]ReduceScatter aiv run success.");
203 0 : return HCCL_SUCCESS;
204 0 : }
205 :
206 : REGISTER_EXEC("ReduceScatterAivDeterExecutor", ReduceScatterAivDeter, CollReduceScatterAivDeterExecutor);
207 :
208 : } // namespace hccl
|
