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_mesh2d.h"
16 : #include "ccu_rank_group.h"
17 : #include "ccu_ctx_creator_registry.h"
18 : #include "ccu_context_all_to_all_mesh2d.h"
19 : #include "ccu_ins_group.h"
20 :
21 : namespace Hccl {
22 :
23 : static CcuInstRegister<CcuContextAlltoAllMesh2D> registerAlltoAll(CcuInstType::CCU_ALLTOALL_MESH_2D_DIRECT);
24 :
25 5 : CcuTempAlltoAllMesh2D::CcuTempAlltoAllMesh2D(const RankId virtualRank, const u32 tempRankSize,
26 : const std::vector<std::vector<RankId>> &tempVTopo,
27 5 : const std::map<RankId, u32> &tempVirtRankMap)
28 5 : : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
29 : {
30 : // 填充框内的维度大小
31 5 : 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("[CcuTempAlltoAllMesh2D] 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 5 : dimSize_.emplace_back(tempVTopo[0].size());
38 5 : dimSize_.emplace_back(tempVTopo[1].size());
39 5 : }
40 :
41 5 : CcuTempAlltoAllMesh2D::~CcuTempAlltoAllMesh2D()
42 : {
43 5 : }
44 :
45 2 : void CcuTempAlltoAllMesh2D::SetA2ASendRecvInfo(const A2ASendRecvInfo &sendRecvInfo)
46 : {
47 2 : localSendRecvInfo_ = sendRecvInfo;
48 2 : return;
49 : }
50 :
51 3 : HcclResult CcuTempAlltoAllMesh2D::CalcRes(AlgTempResReq &tempResReq)
52 : {
53 3 : 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 3 : tempResReq.streamNum = tempResReq.queNum + 1; // 多申请一个 stream 给 ccuInsGroup
55 3 : uint32_t dieNum = tempVTopo_.size();
56 3 : 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("[CcuTempAlltoAllMesh2D] Rank[%d], Invalid IODieNum[%u].",
58 : myRank_, dieNum));
59 : }
60 9 : HCCL_INFO("[CcuTempAlltoAllMesh2D] 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 9 : for (u32 dim = 0; dim < tempVTopo_.size(); dim++) {
65 6 : CHK_RET(GetAlgRank(myRank_, tempVTopo_[dim], myAlgRank));
66 14 : for (u32 queIdx = 0; queIdx < tempVTopo_[dim].size() - 1; queIdx++) {
67 : // find neighbors -> virtualRank
68 8 : u32 neighborAlgRank = (myAlgRank + 1 + queIdx) % (tempVTopo_[dim].size());
69 8 : RankId neighborRank = tempVTopo_[dim][neighborAlgRank];
70 24 : HCCL_INFO("[CollAlgFactory] [CcuTempAlltoAllMesh2D] Rank[%d], Dim[%u], NeighborRank[%d].", myRank_,
71 : dim, neighborRank);
72 :
73 : // LinkNum
74 8 : tempResReq.links[neighborRank] = 1;
75 : }
76 : }
77 :
78 3 : return HcclResult::HCCL_SUCCESS;
79 : }
80 :
81 2 : HcclResult CcuTempAlltoAllMesh2D::FillLinks(const ResLinks &tempLinks)
82 : {
83 8 : for (auto pair : tempLinks) {
84 6 : if (pair.second.size() == 0) { // ESL环境上暂只有直连链路
85 0 : THROW<InvalidParamsException>(
86 0 : StringFormat("[CcuTempAlltoAllMesh2D] Rank[%d]--Peer[%d].", myRank_, pair.first));
87 : }
88 6 : if (pair.first / dimSize_[0] == myRank_ / dimSize_[0]) {
89 12 : HCCL_INFO("[CcuTempAlltoAllMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksX", myRank_, pair.first);
90 4 : linksX_.emplace_back(pair.second[0]);
91 2 : } else if (pair.first % dimSize_[0] == myRank_ % dimSize_[0]) {
92 6 : HCCL_INFO("[CcuTempAlltoAllMesh2D][Run] Rank[%d] insert link to Rank[%d] in linksY", myRank_, pair.first);
93 2 : linksY_.emplace_back(pair.second[0]);
94 : } else {
95 0 : THROW<InvalidParamsException>(StringFormat(
96 0 : "[CcuTempAlltoAllMesh2D] Rank[%d], Unexpected peerRank[%d] in tempLinks.", myRank_, pair.first));
97 : }
98 6 : }
99 8 : for (auto &peer : tempVTopo_[0]) {
100 6 : rankGroupX_.AddRank(peer);
101 : }
102 6 : for (auto &peer : tempVTopo_[1]) {
103 4 : rankGroupY_.AddRank(peer);
104 : }
105 :
106 2 : return HcclResult::HCCL_SUCCESS;
107 : }
108 :
109 2 : HcclResult CcuTempAlltoAllMesh2D::RunOneStep(uint64_t sendRecvSize, uint64_t maxTransportSize, uint32_t sendRecvTimes,
110 : uint32_t step, std::vector<InsQuePtr> &tempInsQues)
111 : {
112 6 : HCCL_INFO("[CcuTempAlltoAllMesh2D][Run] Rank[%d], Step[%u], sendRecvTimes[%u].", myRank_, step, sendRecvTimes);
113 :
114 2 : uint64_t inputAddr = op_.inputMem->GetAddr();
115 2 : uint64_t outputAddr = op_.outputMem->GetAddr();
116 2 : uint64_t scratchAddr = op_.scratchMem->GetAddr();
117 2 : uint32_t typeSize = DataTypeSizeGet(op_.all2AllDataDes.sendType);
118 2 : uint64_t sendStrideSize = 0 * typeSize;
119 2 : uint64_t recvStrideSize = 0 * typeSize;
120 : uint64_t token;
121 2 : CHK_RET(GetToken(op_, token));
122 :
123 2 : uint64_t stepSize = (step == sendRecvTimes - 1) ? (sendRecvSize - step * maxTransportSize) : maxTransportSize;
124 2 : uint64_t aSize = static_cast<uint64_t>((stepSize / 2) / typeSize) * typeSize; // 暂定X和Y方向每轮传输大小一致，按照count对齐
125 2 : uint64_t bSize = stepSize - aSize;
126 2 : uint64_t baseOffset = step * maxTransportSize; // 已传输完成的数据量
127 2 : std::unique_ptr<CcuInsGroup> insGroupPtr = std::make_unique<CcuInsGroup>();
128 6 : for (uint32_t axisId = 0; axisId < 2; axisId++) { // 2D算法，需要执行两次
129 12 : HCCL_INFO("[CcuTempAlltoAllMesh2D][Run] Rank[%d], Step[%u], axisId[%u], aSize[%llu], bSize[%llu], baseOffset[%llu].",
130 : myRank_, step, axisId, aSize, bSize, baseOffset);
131 :
132 4 : CcuInstructionAlltoAllMesh2D ins = CcuInstructionAlltoAllMesh2D(op_, dimSize_, tempVTopo_);
133 4 : ins.Init(myRank_, inputAddr, outputAddr, scratchAddr, axisId, sendStrideSize, recvStrideSize,
134 4 : localSendRecvInfo_.sendLength[0], aSize, bSize, baseOffset, token);
135 4 : ins.SetLinks(axisId == 0 ? linksX_ : linksY_);
136 4 : ins.SetRankGroup(axisId == 0 ? rankGroupX_ : rankGroupY_);
137 4 : ins.SetCntCkeNum(4); // 每个transport用4个CKE
138 4 : insGroupPtr->Append(std::move(std::make_unique<CcuInstructionAlltoAllMesh2D>(ins)));
139 4 : }
140 2 : tempInsQues[0]->Append(std::move(insGroupPtr)); // 只有一条流
141 :
142 2 : return HcclResult::HCCL_SUCCESS;
143 2 : }
144 :
145 2 : HcclResult CcuTempAlltoAllMesh2D::Run(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
146 : const BuffInfo &buffInfo, const ResLinks &tempLinks,
147 : std::vector<InsQuePtr> &tempInsQues)
148 : {
149 : // 分别记录两个Die上的link，构造rankGroup
150 : (void)tempFuncs;
151 : (void)sliceInfoVec;
152 : (void)buffInfo;
153 2 : CHK_RET(FillLinks(tempLinks));
154 :
155 : // 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
156 6 : HCCL_INFO("[CcuTempAlltoAllMesh2D] dataType[%s] sendType[%s]", op_.dataType.Describe().c_str(),
157 : op_.all2AllDataDes.sendType.Describe().c_str());
158 2 : uint32_t typeSize = DataTypeSizeGet(op_.all2AllDataDes.sendType);
159 2 : uint64_t sendRecvSize = localSendRecvInfo_.sendLength[0];
160 2 : uint64_t blockBufferSize = static_cast<uint64_t>((scratchBufferSize_ / tempRankSize_ / 2) / typeSize) * typeSize; // 分2组buffer
161 2 : uint64_t maxTransportSize = min(min(CalcLGMaxTransSize(), UB_MAX_TRANS_SIZE), blockBufferSize);
162 2 : uint32_t sendRecvTimes = (sendRecvSize + maxTransportSize - 1) / maxTransportSize;
163 6 : HCCL_INFO("[CollAlgFactory][Run] Rank[%d], blockBufferSize[%llu], sendRecvTimes[%u].",
164 : myRank_, blockBufferSize, sendRecvTimes);
165 :
166 : uint64_t token;
167 2 : CHK_RET(GetToken(op_, token));
168 6 : HCCL_INFO("[CcuTempAlltoAllMesh2D] Rank[%d], input[%llu], output[%llu], scratch[%llu], sendStride[%llu], \
169 : recvStride[%llu].",
170 : myRank_, op_.inputMem->GetAddr(), op_.outputMem->GetAddr(), op_.scratchMem->GetAddr(), 0, 0);
171 :
172 2 : if (tempInsQues.size() == 0) {
173 0 : HCCL_ERROR("[CcuTempAlltoAllMesh2D][Run] invalid tempInsQues size is [%zu].", tempInsQues.size());
174 0 : return HcclResult::HCCL_E_PARA;
175 : }
176 4 : for (uint32_t step = 0; step < sendRecvTimes; step++) { // 零数据量时会跳过
177 2 : if (tempRankSize_ == 1) {
178 : // alltoall算子的单P场景单独处理
179 0 : DataSlice usrInSlice = DataSlice(BufferType::INPUT, 0, sendRecvSize);
180 0 : DataSlice usrOutSlice = DataSlice(BufferType::OUTPUT, 0, sendRecvSize);
181 0 : std::unique_ptr<Instruction> insLocalCopy = std::make_unique<InsLocalCopy>(usrInSlice, usrOutSlice);
182 0 : tempInsQues[0]->Append(std::move(insLocalCopy));
183 0 : HCCL_INFO("[CcuTempAlltoAllMesh2D] rankSize = 1, use InsLocalCopy for sliceSize[%llu].", sendRecvSize);
184 0 : break;
185 0 : }
186 2 : CHK_RET(RunOneStep(sendRecvSize, maxTransportSize, sendRecvTimes, step, tempInsQues));
187 : }
188 :
189 2 : return HcclResult::HCCL_SUCCESS;
190 : }
191 :
192 2 : HcclResult CcuTempAlltoAllMesh2D::GetScratchBufferInfo(const uint64_t scratchBufferSize, DataType dataType)
193 : {
194 2 : scratchBufferSize_ = scratchBufferSize;
195 2 : dataType_ = dataType;
196 2 : return HcclResult::HCCL_SUCCESS;
197 : }
198 :
199 : } // namespace Hccl
|
