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