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 "ccu_context_scatter_mesh1d.h"
12 : #include "ccu_instruction_scatter_mesh1d.h"
13 :
14 : namespace Hccl {
15 :
16 : constexpr int INPUT_XN_ID = 0;
17 : constexpr int OUTPUT_XN_ID = 1;
18 : constexpr int TOKEN_XN_ID = 2;
19 : constexpr int CKE_IDX_0 = 0;
20 : constexpr int CKE_IDX_1 = 1;
21 : constexpr int CKE_IDX_2 = 2;
22 :
23 : using CurrentCtxArg = CcuCtxArgScatterMesh1D;
24 : using CurrentTaskArg = CcuTaskArgScatterMesh1D;
25 :
26 0 : CcuContextScatterMesh1D::CcuContextScatterMesh1D(
27 0 : const CcuCtxArg &arg, const std::vector<CcuTransport *> &transports, const CcuTransportGroup &group)
28 0 : : CcuContextAlgBase(arg, transports, group)
29 : {
30 0 : HCCL_INFO("[CcuContextScatterMesh1D] Enter Constructor");
31 0 : const CurrentCtxArg *ctxArg = dynamic_cast<const CurrentCtxArg *>(&arg);
32 0 : if (ctxArg == nullptr) {
33 0 : THROW<NullPtrException>(StringFormat("CcuContextScatterMesh1D::ctxArg ptr is null"));
34 : }
35 0 : rankId_ = ctxArg->rankId_;
36 0 : rankSize_ = ctxArg->dimSize_[0];
37 0 : rootId_ = ctxArg->rootId_;
38 :
39 0 : HCCL_INFO("[CcuContextScatterMesh1D] CtxArg: rankId[%u] rankSize[%u]", rankId_, rankSize_);
40 0 : }
41 :
42 0 : void CcuContextScatterMesh1D::InitResource()
43 : {
44 0 : selfBit_ = 1 << rankId_;
45 0 : allBit_ = ((1 << rankSize_) - 1) & (~(1 << rankId_));
46 :
47 0 : if (transports.size() == 0) {
48 0 : THROW<NullPtrException>(StringFormat("CcuContextScatterMesh1D transports is empty"));
49 : }
50 0 : HCCL_INFO("[CcuContextScatterMesh1D]transports.size: [%u]", transports.size());
51 0 : uint16_t transportIdx = 0;
52 : // 按照rank号从小到大遍历transports，遇到本rank就填充本地资源，否则依次取远端资源，要求给框架返回的Link同样是按顺序排列的
53 0 : for (uint64_t peerId = 0; peerId < rankSize_; peerId++) {
54 0 : if (peerId == rankId_) {
55 0 : output_.push_back(CreateVariable());
56 0 : token_.push_back(CreateVariable());
57 : } else {
58 0 : HCCL_INFO("[CcuContextScatterMesh1D] MyRank[%u], PeerId[%llu], TransportId[%u]",
59 : rankId_, peerId, transportIdx);
60 0 : CHK_PRT_RET(transports[transportIdx] == nullptr || transportIdx >= transports.size(),
61 : HCCL_ERROR("[CcuContextScatterMesh1D] Algorithm transport ptr is null or transportIdx is out of bounds"),);
62 0 : output_.push_back(CreateVariable((*transports[transportIdx]), OUTPUT_XN_ID)); // 获取transport中id=1的Var来传递output
63 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
64 0 : transportIdx++;
65 : }
66 : }
67 0 : input_ = CreateVariable();
68 0 : groupOpSize_ = CreateGroupOpSize();
69 0 : currentRankSliceInputOffset_ = CreateVariable();
70 0 : currentRankSliceOutputOffset_ = CreateVariable();
71 0 : inputRepeatStride_ = CreateVariable();
72 0 : outputRepeatStride_ = CreateVariable();
73 0 : normalSliceSize_ = CreateVariable();
74 0 : lastSliceSize_ = CreateVariable();
75 0 : repeatNumVar_ = CreateVariable();
76 0 : flag_ = CreateVariable();
77 0 : flag_ = 0;
78 0 : localMem_.reserve(rankSize_);
79 0 : remoteMem_.reserve(rankSize_);
80 0 : for (uint64_t rankIdx = 0; rankIdx < rankSize_; rankIdx++) {
81 0 : remoteMem_.push_back(CreateMemory());
82 0 : localMem_.push_back(CreateMemory());
83 : }
84 :
85 0 : localSignal_ = CreateMaskSignal();
86 0 : return;
87 : }
88 :
89 0 : void CcuContextScatterMesh1D::LoadArgs()
90 : {
91 0 : Load(input_);
92 0 : Load(output_[rankId_]);
93 0 : Load(token_[rankId_]);
94 0 : Load(currentRankSliceInputOffset_);
95 0 : Load(currentRankSliceOutputOffset_);
96 0 : Load(inputRepeatStride_);
97 0 : Load(outputRepeatStride_);
98 0 : Load(normalSliceSize_);
99 0 : Load(lastSliceSize_);
100 0 : Load(repeatNumVar_);
101 0 : Load(groupOpSize_);
102 0 : return;
103 : }
104 :
105 0 : void CcuContextScatterMesh1D::Algorithm()
106 : {
107 0 : HCCL_INFO("[CcuContextScatterMesh1D] ScatterMesh1D run");
108 0 : InitResource();
109 0 : LoadArgs();
110 0 : PreSync();
111 0 : if (rankId_ == rootId_) {
112 0 : RunSendScatter();
113 : } else {
114 0 : HCCL_INFO("[CcuContextScatterMesh1D] RunRecvScatter local rank[%u], root rank[%u], do nothing", rankId_, rootId_);
115 : }
116 0 : PostSync();
117 0 : HCCL_INFO("[CcuContextScatterMesh1D] ScatterMesh1D end");
118 0 : return;
119 : }
120 :
121 0 : void CcuContextScatterMesh1D::PostSync()
122 : {
123 0 : for (auto t : transports) {
124 0 : RemotePost(*t, CKE_IDX_0, selfBit_);
125 : }
126 0 : GroupWait(*transportGroup, CKE_IDX_0, allBit_);
127 0 : HCCL_INFO("[CcuContextScatterMesh1D] ScatterMesh1D groupwait end");
128 0 : }
129 :
130 0 : void CcuContextScatterMesh1D::PreSync()
131 : {
132 0 : for (auto t : transports) {
133 0 : HCCL_INFO("[CcuContextScatterMesh1D] ScatterMesh1D LocalPost begin");
134 0 : WriteVariableWithSignal(*t, output_[rankId_], CKE_IDX_1, CKE_IDX_1, selfBit_); // index = 1，传递output信息
135 0 : WriteVariableWithSignal(*t, token_[rankId_], CKE_IDX_2, CKE_IDX_2, selfBit_); // index = 2，传递token信息
136 : }
137 0 : GroupWait(*transportGroup, CKE_IDX_1, allBit_); // index = 1，传递output信息
138 0 : GroupWait(*transportGroup, CKE_IDX_2, allBit_); // index = 2，传递token信息
139 0 : HCCL_INFO("[CcuContextScatterMesh1D] ScatterMesh1D wait all end");
140 0 : return;
141 : }
142 :
143 0 : void CcuContextScatterMesh1D::RunSendScatter()
144 : {
145 0 : uint16_t fullBit = ((1 << rankSize_) - 1);
146 0 : HCCL_INFO("[CcuContextScatterMesh1D] RunSendScatter local rank[%u], root rank[%u], start send data", rankId_, rootId_);
147 0 : std::vector<CcuRep::Memory> &dst = remoteMem_; // 在initresource里面pushback
148 0 : std::vector<CcuRep::Memory> &src = localMem_;
149 :
150 0 : for (uint64_t curId = 0; curId < rankSize_; curId++) {
151 0 : src[curId].token = token_[curId];
152 0 : dst[curId].token = token_[curId];
153 :
154 0 : src[curId].addr = input_;
155 0 : for (uint64_t i = 0; i < curId; i++) {
156 0 : src[curId].addr += currentRankSliceInputOffset_;
157 : }
158 0 : dst[curId].addr = output_[curId];
159 : }
160 0 : CcuRep::Variable repeatNumAdd = CreateVariable();
161 0 : repeatNumAdd = 1;
162 0 : CCU_WHILE(repeatNumVar_ != UINT64_MAX) { // 循环repeatNum_次
163 0 : repeatNumVar_ += repeatNumAdd;
164 0 : CCU_IF(flag_ != 0) {
165 : // 非第一轮执行时，src 和 dst 已经初始化，需要添加偏移量
166 0 : for (auto &s : src) {
167 0 : s.addr += inputRepeatStride_;
168 : }
169 0 : for (auto &d : dst) {
170 0 : d.addr += outputRepeatStride_;
171 : }
172 0 : }
173 0 : flag_ = 1;
174 0 : CcuRep::MaskSignal locSig = CreateMaskSignal();
175 0 : uint16_t transportIdx = 0;
176 0 : for (uint64_t rankIdx = 0; rankIdx < rankSize_; rankIdx++) {
177 0 : if ( rankIdx == rankId_ ) {
178 0 : LocalCopy(dst[rankIdx], src[rankIdx], normalSliceSize_, locSig, 1 << rankIdx);
179 : } else {
180 0 : Write(*transports[transportIdx], dst[rankIdx], src[rankIdx], normalSliceSize_, locSig, 1 << rankIdx);
181 0 : transportIdx++;
182 : }
183 : }
184 0 : LocalWait(locSig, fullBit);
185 0 : }
186 0 : return;
187 0 : }
188 :
189 0 : std::vector<uint64_t> CcuContextScatterMesh1D::GeneArgs(const CcuTaskArg &arg)
190 : {
191 0 : const CurrentTaskArg *taskArg = dynamic_cast<const CurrentTaskArg *>(&arg);
192 0 : if (taskArg == nullptr) {
193 0 : THROW<NullPtrException>(StringFormat("CcuContextScatterMesh1D::taskArg ptr is null"));
194 : }
195 0 : uint64_t inputAddr = taskArg->inputAddr_;
196 0 : uint64_t outputAddr = taskArg->outputAddr_;
197 0 : uint64_t tokenInfo = taskArg->token_;
198 0 : uint64_t currentRankSliceInputOffset = taskArg->inputSliceStride_;
199 0 : uint64_t currentRankSliceOutputOffset = taskArg->outputSliceStride_;
200 0 : uint64_t inputRepeatStride = taskArg->inputRepeatStride_;
201 0 : uint64_t outputRepeatStride = taskArg->outputRepeatStride_;
202 0 : uint64_t normalSliceSize = taskArg->normalSliceSize_;
203 0 : uint64_t lastSliceSize = taskArg->lastSliceSize_;
204 0 : uint64_t repeatNumVar = taskArg->repeatNumVar_;
205 0 : auto goSize = CalGoSize(normalSliceSize);
206 :
207 : std::vector<uint64_t> taskArgs = {
208 : inputAddr,
209 : outputAddr,
210 : tokenInfo,
211 : currentRankSliceInputOffset,
212 : currentRankSliceOutputOffset,
213 : inputRepeatStride,
214 : outputRepeatStride,
215 : normalSliceSize,
216 : lastSliceSize,
217 : repeatNumVar,
218 0 : goSize[0],
219 0 : goSize[1],
220 0 : goSize[2],
221 0 : goSize[3]
222 0 : };
223 :
224 0 : HCCL_INFO("[CcuContextScatterMesh1D] TaskArgs: inputAddr[%llu], outputAddr[%llu], "
225 : "currentRankSliceInputOffset[%llu], currentRankSliceOutputOffset[%llu], "
226 : "inputRepeatStride[%llu], outputRepeatStride[%llu], normalSliceSize[%llu], lastSliceSize[%llu],"
227 : "repeatNumVar[%llu], goSize[0][%llu], goSize[1][%llu], goSize[2][%llu], goSize[3][%llu]",
228 : inputAddr, outputAddr, currentRankSliceInputOffset, currentRankSliceOutputOffset, inputRepeatStride,
229 : outputRepeatStride, normalSliceSize, lastSliceSize, repeatNumVar, goSize[0], goSize[1], goSize[2], goSize[3]);
230 0 : return taskArgs;
231 0 : }
232 : }
|
