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_reduce_mesh1d.h"
12 : #include "ccu_instruction_reduce_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 : constexpr int CKE_IDX_3 = 3;
23 :
24 : using CurrentCtxArg = CcuCtxArgReduceMesh1D;
25 : using CurrentTaskArg = CcuTaskArgReduceMesh1D;
26 :
27 0 : CcuContextReduceMesh1D::CcuContextReduceMesh1D(const CcuCtxArg &arg, const std::vector<CcuTransport *> &transports,
28 0 : const CcuTransportGroup &group)
29 0 : : CcuContextAlgBase(arg, transports, group)
30 : {
31 0 : HCCL_DEBUG("[CcuContextReduceMesh1D] Enter Constructor");
32 0 : const CurrentCtxArg *ctxArg = dynamic_cast<const CurrentCtxArg *>(&arg);
33 0 : if (ctxArg == nullptr) {
34 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceMesh1D::ctxArg ptr is null"));
35 : }
36 0 : rankId_ = ctxArg->rankId_;
37 0 : rankSize_ = ctxArg->dimSize_[0];
38 0 : dataType_ = ctxArg->op_.dataType;
39 0 : outputDataType_ = ctxArg->op_.outputDataType;
40 :
41 0 : if (outputDataType_ == DataType::INVALID) {
42 0 : outputDataType_ = dataType_;
43 0 : HCCL_INFO("[CcuContextReduceMesh1D] outputDataType is [INVALID], set outputDataType to[%s]",
44 : outputDataType_.Describe().c_str());
45 : }
46 :
47 0 : if (ctxArg->dimSize_.size() > 0) {
48 0 : rankSize_ = ctxArg->dimSize_[0];
49 : }
50 :
51 0 : HCCL_INFO("[CcuContextReduceMesh1D] CtxArg: rankId[%u] rankSize[%u]",
52 : rankId_, rankSize_);
53 :
54 0 : reduceOp_ = ctxArg->op_.reduceOp;
55 0 : rootId_ = ctxArg->rootId_;
56 0 : HCCL_INFO("[CcuContextReduceMesh1D] init end, ctxArg->dimSize size[%u] rankSize[%llu]", ctxArg->dimSize_.size(), rankSize_);
57 0 : }
58 :
59 0 : void CcuContextReduceMesh1D::InitResource()
60 : {
61 0 : if (transports.size() == 0) {
62 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceMesh1D transports is empty"));
63 : }
64 0 : HCCL_INFO("[CcuContextReduceMesh1D]transports.size: [%u]", transports.size());
65 0 : uint16_t transportIdx = 0;
66 : // 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
67 0 : for (uint16_t peerId = 0; peerId < rankSize_; peerId++) {
68 0 : if (peerId == rankId_) {
69 0 : input_.push_back(CreateVariable());
70 0 : output_.push_back(CreateVariable());
71 0 : token_.push_back(CreateVariable());
72 : } else {
73 0 : HCCL_DEBUG("[CcuContextReduceMesh1D] MyRank[%u], PeerId[%hu], TransportId[%hu]",
74 : rankId_, peerId, transportIdx);
75 : // 判断transport是否为空，为空直接报错
76 0 : CHK_PRT_THROW(transports[transportIdx] == nullptr,
77 : HCCL_ERROR("[CcuContextReduceMesh1D] [InitResource] transports[%u] is nullptr", transportIdx),
78 : NullPtrException, "transport is null");
79 0 : input_.push_back(CreateVariable((*transports[transportIdx]), INPUT_XN_ID)); // 获取transport中id=1的Var来传递input
80 0 : output_.push_back(CreateVariable((*transports[transportIdx]), OUTPUT_XN_ID)); // 获取transport中id=2的Var来传递output
81 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
82 0 : transportIdx++;
83 : }
84 : }
85 :
86 0 : groupOpSize_ = CreateGroupOpSize();
87 :
88 0 : currentRankSliceInputOffset_ = CreateVariable();
89 0 : currentRankSliceOutputOffset_ = CreateVariable();
90 0 : repeatNum_ = CreateVariable();
91 0 : inputRepeatStride_ = CreateVariable();
92 0 : outputRepeatStride_ = CreateVariable();
93 :
94 0 : normalSliceSize_ = CreateVariable();
95 0 : lastSliceSize_ = CreateVariable();
96 0 : repeatNumVar_ = CreateVariable();
97 0 : flag_ = CreateVariable();
98 :
99 0 : selfBit_ = 1 << rankId_;
100 0 : allBit_ = ((1 << rankSize_) - 1) & (~(1 << rankId_));
101 :
102 0 : localMem_ = CreateMemory();
103 0 : reomteMem_.reserve(rankSize_);
104 0 : for (uint32_t rankIdx = 0; rankIdx < rankSize_; rankIdx++) {
105 0 : reomteMem_.push_back(CreateMemory());
106 : }
107 :
108 0 : localSignal_ = CreateMaskSignal();
109 0 : return;
110 : }
111 :
112 0 : void CcuContextReduceMesh1D::LoadArgs()
113 : {
114 0 : Load(input_[rankId_]);
115 0 : Load(output_[rankId_]);
116 0 : Load(token_[rankId_]);
117 0 : Load(currentRankSliceInputOffset_);
118 0 : Load(currentRankSliceOutputOffset_);
119 0 : Load(repeatNum_);
120 0 : Load(inputRepeatStride_);
121 0 : Load(outputRepeatStride_);
122 0 : Load(normalSliceSize_);
123 0 : Load(lastSliceSize_);
124 0 : Load(repeatNumVar_);
125 0 : Load(groupOpSize_);
126 0 : return;
127 : }
128 :
129 0 : void CcuContextReduceMesh1D::PreSync()
130 : {
131 0 : for (auto t : transports) {
132 0 : HCCL_INFO("[CcuContextReduceMesh1D] ReduceMesh1D LocalPost begin");
133 0 : WriteVariableWithSignal(*t, input_[rankId_], INPUT_XN_ID, CKE_IDX_1, selfBit_); // index = 1，传递input信息
134 0 : WriteVariableWithSignal(*t, output_[rankId_], OUTPUT_XN_ID, CKE_IDX_2, selfBit_); // index = 0，传递output信息
135 0 : WriteVariableWithSignal(*t, token_[rankId_], TOKEN_XN_ID, CKE_IDX_3, selfBit_); // index = 2，传递token信息
136 : }
137 0 : GroupWait(*transportGroup, CKE_IDX_1, allBit_);
138 0 : GroupWait(*transportGroup, CKE_IDX_2, allBit_);
139 0 : GroupWait(*transportGroup, CKE_IDX_3, allBit_);
140 0 : HCCL_INFO("[CcuContextReduceMesh1D] ReduceMesh1D wait all end");
141 0 : return;
142 : }
143 :
144 0 : void CcuContextReduceMesh1D::PostSync()
145 : {
146 0 : for (auto t : transports) {
147 0 : RemotePost(*t, CKE_IDX_0, selfBit_);
148 : }
149 0 : GroupWait(*transportGroup, CKE_IDX_0, allBit_);
150 0 : HCCL_INFO("[CcuContextReduceMesh1D] ReduceMesh1D Reduce groupwait end");
151 0 : }
152 :
153 0 : void CcuContextReduceMesh1D::DoRepeatReduce()
154 : {
155 0 : std::vector<CcuRep::Memory> &src = reomteMem_;
156 0 : CcuRep::Memory &dst = localMem_;
157 :
158 0 : dst.addr = output_[rankId_];
159 0 : dst.token = token_[rankId_];
160 0 : uint32_t curId = 0;
161 0 : for (uint32_t rankIdx = 0; rankIdx < rankSize_; rankIdx++) {
162 0 : if (rankIdx != rootId_) {
163 0 : src[curId].addr = input_[rankIdx];
164 0 : src[curId].token = token_[rankIdx];
165 0 : curId++;
166 : } else {
167 0 : continue;
168 : }
169 : }
170 0 : src[rankSize_ - 1].addr = input_[rankId_];
171 0 : src[rankSize_ - 1].token = token_[rankId_];
172 :
173 0 : CCU_IF (flag_ != 0) {
174 : // 非第一轮执行时，src 和 dst 已经初始化，需要添加偏移量
175 0 : dst.addr += outputRepeatStride_;
176 0 : for (auto &s : src) {
177 0 : s.addr += inputRepeatStride_;
178 : }
179 0 : }
180 0 : GroupReduce(transports, dst, src, groupOpSize_, dataType_, outputDataType_, reduceOp_);
181 0 : }
182 :
183 0 : void CcuContextReduceMesh1D::Algorithm()
184 : {
185 0 : HCCL_INFO("[CcuContextReduceMesh1D] ReduceMesh1D run");
186 0 : InitResource();
187 0 : LoadArgs();
188 0 : PreSync();
189 0 : if (rankId_ == rootId_) {
190 0 : CcuRep::Variable repeatNumAdd = CreateVariable();
191 0 : repeatNumAdd = 1;
192 0 : flag_ = 0;
193 0 : CCU_WHILE(repeatNumVar_ != UINT64_MAX) { // 循环repeatNum_次
194 0 : DoRepeatReduce();
195 0 : repeatNumVar_ += repeatNumAdd;
196 0 : flag_ = 1;
197 0 : }
198 0 : }
199 0 : PostSync();
200 0 : HCCL_INFO("[CcuContextReduceMesh1D] ReduceMesh1D end");
201 0 : return;
202 : }
203 :
204 0 : std::vector<uint64_t> CcuContextReduceMesh1D::GeneArgs(const CcuTaskArg &arg)
205 : {
206 0 : const CurrentTaskArg *taskArg = dynamic_cast<const CurrentTaskArg *>(&arg);
207 : // 空指针校验
208 0 : if (taskArg == nullptr) {
209 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceMesh1D::taskArg ptr is null"));
210 : }
211 0 : uint64_t currentRankSliceInputOffset = taskArg->inputSliceStride_ * rankId_;
212 0 : uint64_t currentRankSliceOutputOffset = taskArg->outputSliceStride_ * rankId_;
213 0 : uint64_t repeatNum = taskArg->repeatNum_;
214 0 : uint64_t inputRepeatStride = taskArg->inputRepeatStride_;
215 0 : uint64_t outputRepeatStride = taskArg->outputRepeatStride_;
216 0 : uint64_t normalSliceSize = taskArg->normalSliceSize_;
217 0 : uint64_t lastSliceSize = taskArg->lastSliceSize_;
218 0 : uint64_t repeatNumVar = taskArg->repeatNumVar_;
219 0 : uint64_t inputAddr = taskArg->inputAddr_;
220 0 : uint64_t outputAddr = taskArg->outputAddr_;
221 0 : uint64_t tokenInfo = taskArg->token_;
222 :
223 0 : auto goSize = CalGoSize(normalSliceSize);
224 :
225 : std::vector<uint64_t> taskArgs = {
226 : inputAddr,
227 : outputAddr,
228 : tokenInfo,
229 : currentRankSliceInputOffset,
230 : currentRankSliceOutputOffset,
231 : repeatNum,
232 : inputRepeatStride,
233 : outputRepeatStride,
234 : normalSliceSize,
235 : lastSliceSize,
236 : repeatNumVar,
237 0 : goSize[0],
238 0 : goSize[1],
239 0 : goSize[2],
240 0 : goSize[3],
241 0 : };
242 :
243 0 : HCCL_INFO("[CcuContextReduceMesh1D] TaskArgs: inputAddr[%llu], outputAddr[%llu], currentRankSliceInputOffset[%llu], "
244 : "currentRankSliceOutputOffset[%llu], repeatNum[%llu], inputRepeatStride[%llu], outputRepeatStride[%llu], "
245 : "normalSliceSize[%llu], lastSliceSize[%llu], repeatNumVar[%llu], goSize[0][%llu], goSize[1][%llu], goSize[2][%llu], goSize[3][%llu], ",
246 : inputAddr, outputAddr, currentRankSliceInputOffset, currentRankSliceOutputOffset, repeatNum, inputRepeatStride,
247 : outputRepeatStride, normalSliceSize, lastSliceSize, repeatNumVar, goSize[0], goSize[1], goSize[2], goSize[3]);
248 :
249 0 : return taskArgs;
250 0 : }
251 : } // namespace Hccl
|
