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 <ios>
12 : #include <iostream>
13 :
14 : #include "log.h"
15 :
16 : #include "ccu_temp_all_to_all_v_mesh_1D.h"
17 : #include "alg_data_trans_wrapper.h"
18 : #include "ccu_instruction_all_to_all_v_mesh1d.h"
19 : #include "ccu_rank_group.h"
20 : #include "ccu_ctx_creator_registry.h"
21 : #include "ccu_context_all_to_all_v_mesh1d.h"
22 :
23 : namespace Hccl {
24 :
25 : static CcuInstRegister<CcuContextAllToAllVMesh1D> registrarAllToAllV(CcuInstType::CCU_ALLTOALLV_MESH_1D_DIRECT);
26 :
27 0 : CcuTempAlltoAllVMesh1D::CcuTempAlltoAllVMesh1D(
28 : const RankId virtualRank, const u32 tempRankSize, const std::vector<std::vector<RankId>>& tempVTopo,
29 0 : const std::map<RankId, u32>& tempVirtRankMap)
30 0 : : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
31 0 : {}
32 :
33 0 : CcuTempAlltoAllVMesh1D::~CcuTempAlltoAllVMesh1D() {}
34 :
35 0 : HcclResult CcuTempAlltoAllVMesh1D::CalcRes(AlgTempResReq& tempResReq)
36 : {
37 0 : tempResReq.queNum = 1;
38 0 : tempResReq.streamNum = tempResReq.queNum;
39 0 : HCCL_INFO("[CalcRes] tempResReq.queNum[%u]", tempResReq.queNum);
40 0 : CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
41 0 : return HcclResult::HCCL_SUCCESS;
42 : }
43 :
44 : HcclResult
45 0 : CcuTempAlltoAllVMesh1D::CalcSliceInfo(const AllignInfo& allignInfo, const u64 dataSize, RankSliceInfo& sliceInfoVec)
46 : {
47 0 : std::vector<SliceInfo> tmp(tempVTopo_.size());
48 0 : sliceInfoVec.resize(tempRankSize_, tmp);
49 :
50 0 : CHK_RET(CalcRsAgSliceInfoMesh(myRank_, tempRankSize_, allignInfo, dataSize, sliceInfoVec));
51 :
52 0 : return HcclResult::HCCL_SUCCESS;
53 0 : }
54 :
55 0 : HcclResult CcuTempAlltoAllVMesh1D::GetScratchBufferInfo(const uint64_t scratchBufferSize, DataType dataType)
56 : {
57 : (void)scratchBufferSize;
58 : (void)dataType;
59 0 : return HcclResult::HCCL_SUCCESS;
60 : }
61 :
62 0 : void CcuTempAlltoAllVMesh1D::SetA2ASendRecvInfo(const A2ASendRecvInfo& sendRecvInfo)
63 : {
64 0 : localSendRecvInfo_ = sendRecvInfo;
65 0 : }
66 :
67 0 : HcclResult CcuTempAlltoAllVMesh1D::SetBuffBlockSize(const u64 buffBlockSize)
68 : {
69 0 : CHK_PRT_RET(
70 : buffBlockSize == 0, HCCL_ERROR("[CcuTempAlltoAllVMesh1D][SetBuffBlockSize] buffBlockSize should not be zero"),
71 : HcclResult::HCCL_E_PARA);
72 0 : buffBlockSize_ = buffBlockSize;
73 0 : return HcclResult::HCCL_SUCCESS;
74 : }
75 :
76 0 : HcclResult CcuTempAlltoAllVMesh1D::SetConcurrentSendRecvNum(const u32 concurrentSendRecvNum)
77 : {
78 0 : CHK_PRT_RET(
79 : concurrentSendRecvNum == 0,
80 : HCCL_ERROR("[CcuTempAlltoAllVMesh1D][SetConcurrentSendRecvNum] concurrentSendRecvNum"
81 : "should not be zero"),
82 : HcclResult::HCCL_E_PARA);
83 0 : concurrentSendRecvNum_ = concurrentSendRecvNum;
84 0 : return HcclResult::HCCL_SUCCESS;
85 : }
86 :
87 0 : uint64_t CcuTempAlltoAllVMesh1D::CalcSendRecvNumSubStep()
88 : {
89 0 : sendNumSubStep_.clear();
90 0 : recvNumSubStep_.clear();
91 0 : uint64_t numSubStep = 0;
92 0 : u32 rankSize = localSendRecvInfo_.sendLength.size();
93 :
94 0 : for (u32 destRank = 0; destRank < rankSize; destRank++) {
95 : uint64_t currRankSendSubStep
96 0 : = ((localSendRecvInfo_.sendLength[destRank] + UB_MAX_DATA_SIZE - 1) / UB_MAX_DATA_SIZE);
97 0 : sendNumSubStep_[destRank] = currRankSendSubStep;
98 :
99 : uint64_t currRankRecvSubStep
100 0 : = ((localSendRecvInfo_.recvLength[destRank] + UB_MAX_DATA_SIZE - 1) / UB_MAX_DATA_SIZE);
101 0 : recvNumSubStep_[destRank] = currRankRecvSubStep;
102 0 : HCCL_INFO(
103 : "[CcuTempAlltoAllVMesh1D][CalcNumSubStep] myRank [%d] currRankSendSubStep[%llu]"
104 : "currRankRecvSubStep[%llu]",
105 : myRank_, currRankSendSubStep, currRankRecvSubStep);
106 0 : numSubStep = std::max(numSubStep, std::max(currRankSendSubStep, currRankRecvSubStep));
107 : }
108 0 : HCCL_INFO("[CcuTempAlltoAllVMesh1D][CalcNumSubStep] myRank [%d] max communication step[%u]", myRank_, numSubStep);
109 0 : return numSubStep;
110 : }
111 :
112 0 : HcclResult CcuTempAlltoAllVMesh1D::Run(
113 : const TempFuncs& tempFuncs, const RankSliceInfo& sliceInfoVec, const BuffInfo& buffInfo, const ResLinks& tempLinks,
114 : std::vector<InsQuePtr>& tempInsQues)
115 : {
116 0 : HCCL_INFO("[CcuTempAlltoAllVMesh1D] Run");
117 : (void)sliceInfoVec;
118 : (void)tempFuncs;
119 : (void)buffInfo;
120 :
121 0 : if (tempRankSize_ == 1) {
122 : // ccu-alltoall算子的单P场景单独处理
123 0 : CHK_PRT_RET(
124 : localSendRecvInfo_.sendLength[myRank_] != localSendRecvInfo_.recvLength[myRank_],
125 : HCCL_ERROR(
126 : "[CcuTempAlltoAllVMesh1D] rankSize = 1, sendLength[%llu] and recvLength[%llu]"
127 : "should be equal.",
128 : localSendRecvInfo_.sendLength[myRank_], localSendRecvInfo_.recvLength[myRank_]),
129 : HcclResult::HCCL_E_PARA);
130 0 : CHK_PRT_RET(
131 : localSendRecvInfo_.sendLength[myRank_] == 0,
132 : HCCL_INFO("[CcuTempAlltoAllVMesh1D] Single Rank and DataSlice size is 0, no need to process."),
133 : HcclResult::HCCL_SUCCESS);
134 :
135 0 : DataSlice usrInSlice = DataSlice(BufferType::INPUT, 0, localSendRecvInfo_.sendLength[myRank_]);
136 0 : DataSlice usrOutSlice = DataSlice(BufferType::OUTPUT, 0, localSendRecvInfo_.recvLength[myRank_]);
137 :
138 0 : std::unique_ptr<Instruction> insLocalCopy = std::make_unique<InsLocalCopy>(usrInSlice, usrOutSlice);
139 0 : tempInsQues[0]->Append(std::move(insLocalCopy));
140 0 : HCCL_INFO(
141 : "[CcuTempAlltoAllVMesh1D] rankSize = 1, use InsLocalCopy for sliceSize[%llu].",
142 : localSendRecvInfo_.sendLength[myRank_]);
143 0 : return HcclResult::HCCL_SUCCESS;
144 0 : }
145 :
146 0 : CcuInstructionAllToAllVMesh1D ccuInsAllToAllVMesh1D;
147 :
148 0 : std::vector<uint64_t> dimSize;
149 0 : dimSize.push_back(tempRankSize_);
150 :
151 0 : std::vector<uint64_t> sliceSize;
152 0 : sliceSize.reserve(localSendRecvInfo_.sendLength.size());
153 0 : for (auto& slice : localSendRecvInfo_.sendLength) {
154 0 : sliceSize.push_back(slice);
155 : }
156 :
157 0 : uint64_t inputAddr = op_.inputMem == nullptr ? 0 : static_cast<uint64_t>(op_.inputMem->GetAddr()); // usrIN起始地址
158 0 : uint64_t outputAddr = op_.outputMem == nullptr ? 0 : static_cast<uint64_t>(op_.outputMem->GetAddr());
159 0 : uint64_t srcOffset = 0; // alltoallv假设src起始地址为发送rank的对应块起始地址
160 0 : uint64_t dstOffset = 0; // alltoallv假设dst起始地址为接收rank的对应块起始地址
161 0 : uint64_t token = 0;
162 0 : GetToken(op_, token);
163 :
164 0 : ccuInsAllToAllVMesh1D.Init(
165 0 : static_cast<uint32_t>(myRank_), inputAddr, outputAddr, sliceSize, token, srcOffset, dstOffset, op_, tempVTopo_,
166 0 : localSendRecvInfo_, loadFromMem_);
167 0 : HCCL_INFO(
168 : "[CcuTempAlltoAllVMesh1D] Run Init: LoadFromMem[%d], myRank_[%d], dimSize[%llu], inputAddr[%llu],"
169 : "outputAddr[%llu], sliceSize size[%llu], srcOffset[%llu], dstOffset[%llu]",
170 : loadFromMem_, myRank_, dimSize[0], inputAddr, outputAddr, sliceSize.size(), srcOffset, dstOffset);
171 : // init links
172 0 : std::vector<LinkData> links;
173 0 : for (auto& pair : tempLinks) {
174 0 : if (pair.second.empty()) {
175 0 : continue;
176 : }
177 0 : links.push_back(pair.second[0]);
178 : }
179 0 : HCCL_INFO("[CcuTempAlltoAllVMesh1D] links.size[%zu]", links.size());
180 0 : ccuInsAllToAllVMesh1D.SetLinks(links);
181 0 : RankGroup rankGroup;
182 0 : if (tempVTopo_.size() == 0 || tempInsQues.size() == 0 || tempVTopo_[0].size() == 0) {
183 0 : HCCL_ERROR(
184 : "[CcuTempAlltoAllVMesh1D] invalid tempVTopo size is [%u] or invalid tempInsQues size is [%u] or "
185 : "invalid tempVTopo_[0].size is [%u].",
186 : tempVTopo_.size(), tempInsQues.size(), tempVTopo_[0].size());
187 0 : return HcclResult::HCCL_E_PARA;
188 : }
189 0 : for (auto& peer : tempVTopo_[0]) {
190 0 : rankGroup.AddRank(peer);
191 : }
192 0 : u32 cntCkeNum = 3; // 默认为3，实际alltoallv只需要2个
193 0 : ccuInsAllToAllVMesh1D.SetCntCkeNum(cntCkeNum);
194 0 : ccuInsAllToAllVMesh1D.SetRankGroup(rankGroup);
195 0 : HCCL_INFO("ccuInsAllToAllVMesh1D is [%s]", ccuInsAllToAllVMesh1D.Describe().c_str());
196 0 : ccuInsAllToAllVMesh1D.Describe();
197 0 : tempInsQues[0]->Append(std::move(std::make_unique<CcuInstructionAllToAllVMesh1D>(ccuInsAllToAllVMesh1D)));
198 :
199 0 : return HcclResult::HCCL_SUCCESS;
200 0 : }
201 : } // namespace Hccl
|
