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_all_reduce_mesh2d_one_shot.h"
12 : #include "ccu_instruction_all_reduce_mesh2d_one_shot.h"
13 :
14 : namespace Hccl {
15 :
16 : constexpr int INPUT_XN_ID = 0;
17 : constexpr int SCRATCH_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 : constexpr int CKE_IDX_4 = 4;
24 : constexpr uint32_t AXIS_NUM = 2;
25 :
26 0 : CcuContextAllReduceMesh2DOneShot::CcuContextAllReduceMesh2DOneShot(const CcuCtxArg &arg,
27 0 : const std::vector<CcuTransport*> &transports, const CcuTransportGroup &group)
28 0 : : CcuContextAlgBase(arg, transports, group)
29 : {
30 0 : const CcuCtxArgAllReduceMesh2DOneShot *ctxArg = dynamic_cast<const CcuCtxArgAllReduceMesh2DOneShot *>(&arg);
31 0 : if (ctxArg == nullptr) {
32 0 : THROW<NullPtrException>(StringFormat("CcuContextAllReduceMesh2DOneShot::ctxArg ptr is null"));
33 : }
34 0 : dimSize_ = ctxArg->dimSize_;
35 0 : axisId_ = ctxArg->axisId_;
36 0 : rankId_ = ctxArg->rankId_;
37 0 : dataType_ = ctxArg->op_.dataType;
38 0 : outputDataType_ = ctxArg->op_.outputDataType;
39 0 : reduceOp_ = ctxArg->op_.reduceOp;
40 0 : if (outputDataType_ == DataType::INVALID) {
41 0 : outputDataType_ = dataType_;
42 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] outputDataType is [INVALID], set outputDataType to[%s]",
43 : outputDataType_.Describe().c_str());
44 : }
45 :
46 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] Init, CtxArgs are dimSize0[%llu], dimSize1[%llu], axisId[%u], "
47 : "rankId[%llu], dataType[%s], outputDataType[%s], reduceOp[%s]",
48 : dimSize_[0], dimSize_[1], axisId_, rankId_, dataType_.Describe().c_str(),
49 : outputDataType_.Describe().c_str(), reduceOp_.Describe().c_str());
50 0 : uint32_t max_dimSize = 2;
51 0 : if (dimSize_.size() != max_dimSize or axisId_ > 1) {
52 0 : THROW<NullPtrException>(StringFormat("[CcuContextAllReduceMesh2DOneShot] dimSize[%u] or axisId[%u] is invalid",
53 : dimSize_.size(), axisId_));
54 : }
55 0 : CHK_PRT_THROW(dimSize_[0] == 0 || dimSize_[1] == 0,
56 : HCCL_ERROR("[CcuContextAllReduceMesh2DOneShot] dimSize0[%llu] or dimSize1[%llu] is zero",
57 : dimSize_[0], dimSize_[1]),
58 : InvalidParamsException, "dimSize[0] or dimSize[1] is invalid");
59 :
60 0 : myRankIdxInAxis_.push_back(rankId_ % dimSize_[0]); // 本 rank 在第 0 维上的 index
61 0 : myRankIdxInAxis_.push_back(rankId_ / dimSize_[0]); // 本 rank 在第 1 维上的 index
62 :
63 0 : myRankIdxInCurrentAxis_ = myRankIdxInAxis_[axisId_];
64 0 : currentAxisRankSize_ = dimSize_[axisId_];
65 :
66 : // 同步信号初始化
67 0 : currAxisSignalName_ = "CcuContextAllReduceMesh2DOneShotAxisSync_" + std::to_string(axisId_);
68 0 : otherAxisSignalName_ = "CcuContextAllReduceMesh2DOneShotAxisSync_" + std::to_string(1 - axisId_);
69 0 : currAxisSignal_ = CreateMaskSignal();
70 0 : ExportMaskSignal(currAxisSignal_, currAxisSignalName_);
71 0 : otherAxisSignal_ = ImportMaskSignal(otherAxisSignalName_);
72 :
73 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] Init, myRankIdx0[%llu], myRankIdx1[%llu], "
74 : "myRankIdxInCurrentAxis[%llu], currentAxisRankSize[%llu]",
75 : myRankIdxInAxis_[0], myRankIdxInAxis_[1], myRankIdxInCurrentAxis_, currentAxisRankSize_);
76 0 : }
77 :
78 0 : void CcuContextAllReduceMesh2DOneShot::Algorithm()
79 : {
80 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] AllReduceMesh2DOneShot run");
81 0 : uint16_t selfBit = 1 << myRankIdxInCurrentAxis_;
82 0 : uint16_t allBit = ((1 << currentAxisRankSize_) - 1) & (~(1 << myRankIdxInCurrentAxis_));
83 :
84 0 : InitVariables();
85 :
86 0 : LoadArgs();
87 :
88 : // 前同步
89 0 : for (auto t : transports) {
90 0 : WriteVariableWithSignal(*t, inputAddr_[myRankIdxInCurrentAxis_], INPUT_XN_ID, CKE_IDX_1, selfBit);
91 0 : WriteVariableWithSignal(*t, scratchAddr_[myRankIdxInCurrentAxis_], SCRATCH_XN_ID, CKE_IDX_2, selfBit);
92 0 : WriteVariableWithSignal(*t, token_[myRankIdxInCurrentAxis_], TOKEN_XN_ID, CKE_IDX_3, selfBit);
93 : }
94 :
95 0 : GroupWait(*transportGroup, CKE_IDX_1, allBit);
96 0 : GroupWait(*transportGroup, CKE_IDX_2, allBit);
97 0 : GroupWait(*transportGroup, CKE_IDX_3, allBit);
98 :
99 : // OneShot Step1
100 0 : CcuRep::Variable& Step1Offset = (axisId_ == 0) ? xSliceOffset_ : ySliceOffset_;
101 0 : GroupOpSize& Step1GoSize = (axisId_ == 0) ? xGoSize_ : yGoSize_;
102 :
103 0 : DoGroupReduce(inputAddr_, scratchAddr_[myRankIdxInCurrentAxis_], Step1Offset, Step1GoSize);
104 :
105 0 : DoAxisSync(0);
106 0 : DoGroupSync(CKE_IDX_4, selfBit, allBit);
107 0 : DoAxisSync(1);
108 :
109 : // OneShot Step2
110 0 : CcuRep::Variable& Step2Offset = (axisId_ == 0) ? ySliceOffset_ : xSliceOffset_ ;
111 0 : GroupOpSize& Step2GoSize = (axisId_ == 0) ? yGoSize_ : xGoSize_;
112 :
113 0 : DoGroupReduce(scratchAddr_, outputAddr_[myRankIdxInCurrentAxis_], Step2Offset, Step2GoSize);
114 :
115 0 : DoAxisSync(0);
116 0 : DoGroupSync(CKE_IDX_0, selfBit, allBit);
117 0 : DoAxisSync(1);
118 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] AllReduceMesh2DOneShot end");
119 0 : return;
120 : }
121 :
122 0 : void CcuContextAllReduceMesh2DOneShot::DoGroupSync(int ckeIdx, uint16_t selfBit, uint16_t allBit)
123 : {
124 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] DoGroupSync Starts, ckeIdx[%d], selfBit[%u], allBit[%u]",
125 : ckeIdx, selfBit, allBit);
126 0 : for (auto t : transports) {
127 0 : RemotePost(*t, ckeIdx, selfBit);
128 : }
129 0 : GroupWait(*transportGroup, ckeIdx, allBit);
130 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] DoGroupSync Ends");
131 0 : return;
132 : }
133 :
134 0 : void CcuContextAllReduceMesh2DOneShot::DoGroupReduce(std::vector<CcuRep::Variable> &srcBase,
135 : CcuRep::Variable &dstBase, CcuRep::Variable &offset, GroupOpSize &goSize)
136 : {
137 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] DoGroupReduce Starts");
138 : // 从轴上所有的对端的对应位置读取数据
139 0 : std::vector<CcuRep::Memory> srcAddrs;
140 0 : for (uint64_t rankIdx = 0; rankIdx < currentAxisRankSize_; rankIdx++) {
141 0 : srcAddrs.push_back(CreateMemory());
142 : }
143 0 : uint32_t rmtId = 0;
144 0 : uint32_t curId = 0;
145 0 : for (uint64_t rankIdx = 0; rankIdx < currentAxisRankSize_; rankIdx++) {
146 0 : if (rankIdx != myRankIdxInCurrentAxis_) {
147 0 : curId = rmtId;
148 0 : rmtId++;
149 : } else {
150 0 : curId = currentAxisRankSize_ - 1;
151 : }
152 0 : srcAddrs[curId].addr = srcBase[rankIdx];
153 0 : srcAddrs[curId].addr += offset;
154 0 : srcAddrs[curId].token = token_[rankIdx];
155 : }
156 : // Reduce 到本端
157 0 : CcuRep::Memory dstAddr = CreateMemory();
158 0 : dstAddr.addr = dstBase;
159 0 : dstAddr.addr += offset;
160 0 : dstAddr.token = token_[myRankIdxInCurrentAxis_];
161 0 : GroupReduce(transports, dstAddr, srcAddrs, goSize, dataType_, outputDataType_, reduceOp_);
162 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] DoGroupReduce Ends");
163 0 : return;
164 0 : }
165 :
166 0 : void CcuContextAllReduceMesh2DOneShot::DoAxisSync(uint32_t signalIdx)
167 : {
168 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] DoAxisSync Starts, signalIdx[%u]", signalIdx);
169 0 : uint32_t sendBit = 1 << axisId_;
170 0 : uint32_t waitBit = 1 << (1 - axisId_);
171 0 : sendBit = sendBit << (AXIS_NUM * signalIdx);
172 0 : waitBit = waitBit << (AXIS_NUM * signalIdx);
173 0 : LocalCtxPost(otherAxisSignal_, sendBit);
174 0 : LocalWait(currAxisSignal_, waitBit);
175 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] DoAxisSync Ends");
176 0 : return;
177 : }
178 :
179 0 : void CcuContextAllReduceMesh2DOneShot::InitVariables()
180 : {
181 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] InitVariables Starts");
182 : // 初始化资源
183 0 : uint16_t transportIdx = 0;
184 0 : if (transports.size() == 0) {
185 0 : THROW<NullPtrException>(StringFormat("CcuContextAllReduceMesh2DOneShot transports is empty"));
186 : }
187 : // 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
188 0 : for (uint64_t peerId = 0; peerId < currentAxisRankSize_; peerId++) {
189 0 : if (peerId == myRankIdxInCurrentAxis_) {
190 0 : inputAddr_.push_back(CreateVariable());
191 0 : scratchAddr_.push_back(CreateVariable());
192 0 : token_.push_back(CreateVariable());
193 : } else {
194 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] MyRank[%u], PeerId[%llu], TransportId[%u]",
195 : myRankIdxInCurrentAxis_, peerId, transportIdx);
196 0 : CHK_PRT_RET(transports[transportIdx] == nullptr || transportIdx >= transports.size(),
197 : HCCL_ERROR("[CcuContextAllReduceMesh2DOneShot] Algorithm transport ptr is null or transportIdx is out of bounds"),);
198 0 : inputAddr_.push_back(CreateVariable((*transports[transportIdx]), INPUT_XN_ID));
199 0 : scratchAddr_.push_back(CreateVariable((*transports[transportIdx]), SCRATCH_XN_ID));
200 0 : token_.push_back(CreateVariable((*transports[transportIdx]), TOKEN_XN_ID));
201 0 : transportIdx++;
202 : }
203 0 : outputAddr_.push_back(CreateVariable());
204 : }
205 :
206 0 : xSliceOffset_ = CreateVariable();
207 0 : ySliceOffset_ = CreateVariable();
208 0 : xGoSize_ = CreateGroupOpSize();
209 0 : yGoSize_ = CreateGroupOpSize();
210 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] InitVariables Ends");
211 0 : return;
212 : }
213 :
214 0 : void CcuContextAllReduceMesh2DOneShot::LoadArgs()
215 : {
216 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] LoadArgs Starts");
217 0 : Load(inputAddr_[myRankIdxInCurrentAxis_]);
218 0 : Load(outputAddr_[myRankIdxInCurrentAxis_]);
219 0 : Load(token_[myRankIdxInCurrentAxis_]);
220 0 : Load(scratchAddr_[myRankIdxInCurrentAxis_]);
221 0 : Load(xSliceOffset_);
222 0 : Load(ySliceOffset_);
223 0 : Load(xGoSize_);
224 0 : Load(yGoSize_);
225 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] LoadArgs Eends");
226 0 : return;
227 : }
228 :
229 0 : std::vector<uint64_t> CcuContextAllReduceMesh2DOneShot::GeneArgs(const CcuTaskArg &arg)
230 : {
231 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] GeneArgs Starts");
232 0 : const CcuTaskArgAllReduceMesh2DOneShot *taskArg = dynamic_cast<const CcuTaskArgAllReduceMesh2DOneShot *>(&arg);
233 0 : if (taskArg == nullptr) {
234 0 : THROW<NullPtrException>(StringFormat("CcuContextAllReduceMesh2DOneShot::taskArg ptr is null"));
235 : }
236 0 : uint64_t tokenInfo = taskArg->token_;
237 :
238 0 : uint64_t inputAddr = taskArg->inputAddr_;
239 0 : uint64_t outputAddr = taskArg->outputAddr_;
240 0 : uint64_t scratchAddr = taskArg->scratchAddr_;
241 :
242 0 : uint64_t xSliceOffset = taskArg->xSliceOffset_;
243 0 : uint64_t ySliceOffset = taskArg->ySliceOffset_;
244 :
245 0 : auto xGoSize = CalGoSize(taskArg->xSliceSize_);
246 0 : auto yGoSize = CalGoSize(taskArg->ySliceSize_);
247 :
248 0 : HCCL_INFO("[CcuContextAllReduceMesh2DOneShot] GeneArgs, TaskArgs are inputAddr[%llu], "
249 : "outputAddr[%llu], scratchAddr[%llu], xSliceSize[%llu], ySliceSize[%llu], xSliceOffset[%llu], "
250 : "ySliceOffset[%llu]",
251 : inputAddr, outputAddr, scratchAddr, taskArg->xSliceSize_, taskArg->ySliceSize_, xSliceOffset,
252 : ySliceOffset);
253 :
254 0 : std::vector<uint64_t> taskArgList = {inputAddr, outputAddr, tokenInfo, scratchAddr, xSliceOffset, ySliceOffset};
255 : // push goSize
256 0 : for (auto goSize : {xGoSize, yGoSize}) {
257 0 : for (auto val : goSize) {
258 0 : taskArgList.push_back(val);
259 : }
260 0 : }
261 0 : return taskArgList;
262 0 : }
263 : }
|
