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 : #include "ccu_temp_all_to_all_v_mesh_2D.h"
16 : #include "ccu_rank_group.h"
17 : #include "ccu_ctx_creator_registry.h"
18 : #include "ccu_context_all_to_all_v_mesh2d.h"
19 : #include "ccu_ins_group.h"
20 :
21 : namespace Hccl {
22 :
23 : static CcuInstRegister<CcuContextAllToAllVMesh2D> g_registerAlltoAllV(CcuInstType::CCU_ALLTOALLV_MESH_2D_DIRECT);
24 :
25 0 : CcuTempAlltoAllVMesh2D::CcuTempAlltoAllVMesh2D(const RankId virtualRank, const u32 tempRankSize,
26 : const std::vector<std::vector<RankId>> &tempVTopo,
27 0 : const std::map<RankId, u32> &tempVirtRankMap)
28 0 : : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
29 : {
30 : // 填充框内的维度大小
31 0 : if (tempVTopo_.size() != 2 || tempVTopo_[0].size() <= 1 || tempVTopo_[1].size() <= 1) { // concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
32 0 : THROW<InvalidParamsException>(StringFormat("[CcuTempAlltoAllVMesh2D] Rank[%d], Invalid tempVTopo "
33 : "Size[%u] or Invalid tempVTopo[0] size [%u] or tempVTopo[1] size [%u].",
34 0 : myRank_, tempVTopo_.size(), tempVTopo_[0].size(),
35 0 : tempVTopo_[1].size()));
36 : }
37 0 : dimSize_.emplace_back(tempVTopo[0].size());
38 0 : dimSize_.emplace_back(tempVTopo[1].size());
39 0 : }
40 :
41 0 : CcuTempAlltoAllVMesh2D::~CcuTempAlltoAllVMesh2D()
42 : {
43 0 : }
44 :
45 0 : void CcuTempAlltoAllVMesh2D::SetA2ASendRecvInfo(const A2ASendRecvInfo &sendRecvInfo)
46 : {
47 0 : localSendRecvInfo_ = sendRecvInfo;
48 0 : return;
49 : }
50 :
51 0 : HcclResult CcuTempAlltoAllVMesh2D::CalcRes(AlgTempResReq &tempResReq)
52 : {
53 0 : tempResReq.queNum = 1; // 只申请一个insQue，填充一个insGroup，由框架将其中的ins放在多个stream上
54 0 : tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个 stream 给 ccuInsGroup
55 0 : uint32_t dieNum = tempVTopo_.size();
56 0 : if (dieNum != 2) { // concurrmesh的topoMatch返回的vTopo大小应当为2，对应X轴和Y轴的大小
57 0 : THROW<InvalidParamsException>(StringFormat("[CcuTempAlltoAllVMesh2D] Rank[%d], Invalid IODieNum[%u].",
58 : myRank_, dieNum));
59 : }
60 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D] Rank[%d] requiredQueNum[%u] VtopoSize[%u], VtopoSize0[%u] VtopoSize1[%u].",
61 : myRank_, tempResReq.queNum, tempVTopo_.size(), tempVTopo_[0].size(), tempVTopo_[1].size());
62 :
63 : uint32_t myAlgRank;
64 0 : for (u32 dim = 0; dim < tempVTopo_.size(); dim++) {
65 0 : CHK_RET(GetAlgRank(myRank_, tempVTopo_[dim], myAlgRank));
66 0 : for (u32 queIdx = 0; queIdx < tempVTopo_[dim].size() - 1; queIdx++) {
67 : // find neighbors -> virtualRank
68 0 : u32 neighborAlgRank = (myAlgRank + 1 + queIdx) % (tempVTopo_[dim].size());
69 0 : RankId neighborRank = tempVTopo_[dim][neighborAlgRank];
70 0 : HCCL_INFO("[CollAlgFactory] [CcuTempAlltoAllVMesh2D] Rank[%d], Dim[%u], NeighborRank[%d].", myRank_,
71 : dim, neighborRank);
72 :
73 : // LinkNum
74 0 : tempResReq.links[neighborRank] = 1;
75 : }
76 : }
77 :
78 0 : return HcclResult::HCCL_SUCCESS;
79 : }
80 :
81 0 : uint64_t CcuTempAlltoAllVMesh2D::CalcSendRecvNumSubStep(uint64_t sliceSize)
82 : {
83 0 : sendNumSubStep_.clear();
84 0 : recvNumSubStep_.clear();
85 0 : uint64_t numSubStep = 0;
86 0 : if (localSendRecvInfo_.sendLength.size() != localSendRecvInfo_.recvLength.size()) {
87 0 : THROW<InvalidParamsException>(
88 0 : StringFormat("[CcuTempAlltoAllVMesh2D][CalcSendRecvNumSubStep] Rank[%d] sendLength size[%u] is not equal to"
89 : "recvLength size[%u]",
90 : myRank_, localSendRecvInfo_.sendLength.size(), localSendRecvInfo_.recvLength.size()));
91 : }
92 0 : u32 rankSize = localSendRecvInfo_.sendLength.size();
93 0 : if (rankSize == 0 || sliceSize == 0) {
94 0 : THROW<InvalidParamsException>(StringFormat(
95 : "[CcuTempAlltoAllVMesh2D][CalcSendRecvNumSubStep] Invalid rankSize [%u] or invalid slicesize[%u].",
96 : rankSize, sliceSize));
97 : }
98 0 : for (u32 destRank = 0; destRank < rankSize; destRank++) {
99 0 : uint64_t currRankSendSubStep = ((localSendRecvInfo_.sendLength[destRank] + sliceSize - 1) / sliceSize);
100 0 : sendNumSubStep_[destRank] = currRankSendSubStep;
101 :
102 : uint64_t currRankRecvSubStep =
103 0 : ((localSendRecvInfo_.recvLength[destRank] + sliceSize- 1) / sliceSize);
104 0 : recvNumSubStep_[destRank] = currRankRecvSubStep;
105 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D][CalcNumSubStep] myRank [%d] currRankSendSubStep[%llu]" \
106 : "currRankRecvSubStep[%llu]", myRank_, currRankSendSubStep, currRankRecvSubStep);
107 0 : numSubStep = std::max(numSubStep, std::max(currRankSendSubStep, currRankRecvSubStep));
108 : }
109 0 : HCCL_INFO("[CcuTempAlltoAllVMesh1D][CalcNumSubStep] myRank [%d] max communication step[%u]",
110 : myRank_, numSubStep);
111 0 : return numSubStep;
112 : }
113 :
114 0 : HcclResult CcuTempAlltoAllVMesh2D::FillLinks(const ResLinks &tempLinks)
115 : {
116 0 : for (auto pair : tempLinks) {
117 0 : if (pair.second.size() == 0) { // ESL环境上暂只有直连链路
118 0 : THROW<InvalidParamsException>(
119 0 : StringFormat("[CcuTempAlltoAllVMesh2D] Rank[%d]--Peer[%d].", myRank_, pair.first));
120 : }
121 0 : if (pair.first / dimSize_[0] == myRank_ / dimSize_[0]) {
122 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksX", myRank_, pair.first);
123 0 : linksX_.emplace_back(pair.second[0]);
124 0 : } else if (pair.first % dimSize_[0] == myRank_ % dimSize_[0]) {
125 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksY", myRank_, pair.first);
126 0 : linksY_.emplace_back(pair.second[0]);
127 : } else {
128 0 : THROW<InvalidParamsException>(StringFormat(
129 0 : "[CcuTempAlltoAllVMesh2D] Rank[%d], Unexpected peerRank[%d] in tempLinks.", myRank_, pair.first));
130 : }
131 0 : }
132 0 : return HcclResult::HCCL_SUCCESS;
133 : }
134 :
135 0 : HcclResult CcuTempAlltoAllVMesh2D::FillRankGroup()
136 : {
137 0 : for (auto &peer : tempVTopo_[0]) {
138 0 : rankGroupX_.AddRank(peer);
139 : }
140 0 : for (auto &peer : tempVTopo_[1]) {
141 0 : rankGroupY_.AddRank(peer);
142 : }
143 0 : return HcclResult::HCCL_SUCCESS;
144 : }
145 :
146 0 : HcclResult CcuTempAlltoAllVMesh2D::CalcSliceSize(uint32_t sendRecvTime, uint64_t maxTransportSize)
147 : {
148 0 : sendSliceSize_.resize(tempRankSize_);
149 0 : recvSliceSize_.resize(tempRankSize_);
150 0 : for (u32 j = 0; j < tempRankSize_; j++) {
151 0 : if ((sendRecvTime + 1) < sendNumSubStep_[j]) {
152 0 : sendSliceSize_[j] = maxTransportSize;
153 0 : } else if ((sendRecvTime + 1) == sendNumSubStep_[j]) {
154 0 : sendSliceSize_[j] = localSendRecvInfo_.sendLength[j] - sliceBias_;
155 : } else {
156 0 : sendSliceSize_[j] = 0;
157 : }
158 : }
159 0 : for (u32 j = 0; j < tempRankSize_; j++) {
160 0 : if ((sendRecvTime + 1) < recvNumSubStep_[j]) {
161 0 : recvSliceSize_[j] = maxTransportSize;
162 0 : } else if ((sendRecvTime + 1) == recvNumSubStep_[j]) {
163 0 : recvSliceSize_[j] = localSendRecvInfo_.recvLength[j] - sliceBias_;
164 : } else {
165 0 : recvSliceSize_[j] = 0;
166 : }
167 : }
168 0 : return HcclResult::HCCL_SUCCESS;
169 : }
170 :
171 0 : HcclResult CcuTempAlltoAllVMesh2D::Run(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
172 : const BuffInfo &buffInfo, const ResLinks &tempLinks,
173 : std::vector<InsQuePtr> &tempInsQues)
174 : {
175 0 : if (tempVTopo_.size() == 0 || tempInsQues.size() == 0) {
176 0 : THROW<NullPtrException>(StringFormat(
177 : "[CcuTempAlltoAllVMesh2D][Run] invalid tempVTopo size is [%u] or invalid tempInsQues size is [%u].",
178 : tempVTopo_.size(), tempInsQues.size()));
179 : }
180 : // 分别记录两个Die上的link，构造rankGroup
181 : (void)tempFuncs;
182 : (void)sliceInfoVec;
183 : (void)buffInfo;
184 0 : CHK_RET(FillLinks(tempLinks));
185 0 : CHK_RET(FillRankGroup());
186 :
187 : // scratch分两组，每组rankSize份，放一个分片，按照传输大小限制与buffer大小限制分多轮执行算子
188 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D] dataType[%d] sendType[%d]", op_.dataType, op_.all2AllVDataDes.sendType);
189 0 : if (!op_.scratchMem) {
190 0 : HCCL_ERROR("[CcuTempAlltoAllVMesh2D][Run] Rank[%d] inputmem or outputmem or scratchMem is null", myRank_);
191 0 : return HcclResult::HCCL_E_PTR;
192 : }
193 :
194 0 : uint64_t inputAddr = op_.inputMem == nullptr ? 0 : op_.inputMem->GetAddr();
195 0 : uint64_t outputAddr = op_.outputMem == nullptr ? 0 : op_.outputMem->GetAddr();
196 0 : uint64_t scratchAddr = op_.scratchMem->GetAddr();
197 : uint64_t token;
198 0 : CHK_RET(GetToken(op_, token));
199 0 : uint32_t typeSize = DataTypeSizeGet(op_.all2AllVDataDes.sendType);
200 : // scratchmem需要切成blockSize大小的格子,发送的数据块blockBufferSize是scratchmem每一格大小的两倍
201 0 : uint32_t blockSize = (tempVTopo_[0].size() - 1) * (tempVTopo_[1].size() - 1) * 2;
202 0 : uint64_t blockBufferSize = static_cast<uint64_t>((scratchBufferSize_ / blockSize) / typeSize) * typeSize * 2;
203 :
204 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D] Rank[%d], input[%llu], output[%llu], scratch[%llu], blockSize[%llu]," \
205 : "blockBufferSize[%llu].", myRank_, inputAddr, outputAddr, scratchAddr, blockSize, blockBufferSize);
206 :
207 0 : if (tempRankSize_ == 1) {
208 : // alltoallv算子的单P场景单独处理
209 0 : DataSlice usrInSlice = DataSlice(BufferType::INPUT, 0, localSendRecvInfo_.sendLength[0]);
210 0 : DataSlice usrOutSlice = DataSlice(BufferType::OUTPUT, 0, localSendRecvInfo_.sendLength[0]);
211 0 : std::unique_ptr<Instruction> insLocalCopy = std::make_unique<InsLocalCopy>(usrInSlice, usrOutSlice);
212 0 : tempInsQues[0]->Append(std::move(insLocalCopy));
213 0 : HCCL_INFO("[CcuTempAlltoAllVMesh2D] rankSize = 1, use InsLocalCopy for sliceSize[%llu].", localSendRecvInfo_.sendLength[0]);
214 0 : }
215 :
216 0 : uint64_t scratchSliceSize = blockBufferSize / 2;
217 0 : uint64_t scratchSliceBias = scratchSliceSize * (tempVTopo_[0].size() - 1) * (tempVTopo_[1].size() - 1);
218 0 : std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
219 0 : for (uint32_t axisId = 0; axisId < 2; axisId++) { // 2D算法，需要执行两次
220 0 : CcuInstructionAllToAllVMesh2D ins = CcuInstructionAllToAllVMesh2D(op_, dimSize_, tempVTopo_);
221 0 : ins.Init(myRank_, axisId, inputAddr, outputAddr, scratchAddr, token, scratchSliceSize,
222 0 : scratchSliceBias, localSendRecvInfo_);
223 0 : ins.SetLinks(axisId == 0 ? linksX_ : linksY_);
224 0 : ins.SetRankGroup(axisId == 0 ? rankGroupX_ : rankGroupY_);
225 0 : u32 ckeNum = 5 + 2 * std::max(dimSize_[0], dimSize_[1]);
226 0 : ins.SetCntCkeNum(ckeNum);
227 0 : insGroupPtr->Append(std::move(std::make_unique<CcuInstructionAllToAllVMesh2D>(ins)));
228 0 : }
229 0 : tempInsQues[0]->Append(std::move(insGroupPtr)); // 只有一条流
230 :
231 0 : return HcclResult::HCCL_SUCCESS;
232 0 : }
233 :
234 0 : HcclResult CcuTempAlltoAllVMesh2D::GetScratchBufferInfo(const uint64_t scratchBufferSize, DataType dataType)
235 : {
236 0 : scratchBufferSize_ = scratchBufferSize;
237 0 : dataType_ = dataType;
238 0 : return HcclResult::HCCL_SUCCESS;
239 : }
240 :
241 : } // namespace Hccl
|
