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_mem2mem.h"
12 : #include "ccu_instruction_reduce_mesh1d_mem2mem.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 = CcuCtxArgReduceMeshMem2Mem1D;
25 : using CurrentTaskArg = CcuTaskArgReduceMeshMem2Mem1D;
26 :
27 0 : CcuContextReduceMeshMem2Mem1D::CcuContextReduceMeshMem2Mem1D(const CcuCtxArg &arg,
28 : const std::vector<CcuTransport *> &transports,
29 0 : const CcuTransportGroup &group)
30 0 : : CcuContextAlgBase(arg, transports, group)
31 : {
32 0 : const CurrentCtxArg *ctxArg = dynamic_cast<const CurrentCtxArg *>(&arg);
33 0 : if (ctxArg == nullptr) {
34 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceMeshMem2Mem1D::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 0 : if (outputDataType_ == DataType::INVALID) {
41 0 : outputDataType_ = dataType_;
42 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] outputDataType is [INVALID], set outputDataType to[%s]",
43 : outputDataType_.Describe().c_str());
44 : }
45 0 : if (ctxArg->dimSize_.size() > 0) {
46 0 : rankSize_ = ctxArg->dimSize_[0];
47 : }
48 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] CtxArg: rankId[%u] rankSize[%u]",
49 : rankId_, rankSize_);
50 0 : reduceOp_ = ctxArg->op_.reduceOp;
51 0 : rootId_ = ctxArg->rootId_;
52 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] init end, ctxArg->dimSize size[%zu] rankSize[%llu]",
53 : ctxArg->dimSize_.size(), rankSize_);
54 0 : }
55 :
56 0 : void CcuContextReduceMeshMem2Mem1D::InitResource()
57 : {
58 0 : if (transports.size() == 0) {
59 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceMeshMem2Mem1D transports is empty"));
60 : }
61 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D]transports.size: [%u]", transports.size());
62 : // 初始化资源
63 0 : uint16_t transportIdx = 0;
64 : // 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
65 0 : for (uint64_t peerId = 0; peerId < rankSize_; peerId++) {
66 0 : if (peerId == rankId_) {
67 0 : input_.push_back(CreateVariable());
68 0 : output_.push_back(CreateVariable());
69 0 : token_.push_back(CreateVariable());
70 : } else {
71 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] MyRank[%u], PeerId[%llu], TransportId[%hu]", rankId_, peerId,
72 : transportIdx);
73 : // 判断transport是否为空，为空直接报错
74 0 : CHK_PRT_RET(transports[transportIdx] == nullptr || transportIdx >= transports.size(),
75 : HCCL_ERROR("[CcuContextReduceMeshMem2Mem1D] Algorithm transport ptr is null or transportIdx is out of bounds"),);
76 0 : input_.push_back(
77 0 : CreateVariable((*transports[transportIdx]), CKE_IDX_0)); // 获取transport中id=1的Var来传递input
78 0 : output_.push_back(
79 0 : CreateVariable((*transports[transportIdx]), CKE_IDX_1)); // 获取transport中id=2的Var来传递output
80 0 : token_.push_back(CreateVariable((*transports[transportIdx]), CKE_IDX_2));
81 0 : transportIdx++;
82 : }
83 : }
84 0 : for (uint16_t roundId = 0; roundId < (rankSize_ - 1); roundId++) {
85 0 : chunkSize_.push_back(CreateVariable());
86 : }
87 0 : inputRepeatStride_ = CreateVariable();
88 0 : outputRepeatStride_ = CreateVariable();
89 0 : normalSliceSize_ = CreateVariable();
90 0 : lastSliceSize_ = CreateVariable();
91 0 : repeatNumVar_ = CreateVariable();
92 0 : flag_ = CreateVariable();
93 0 : isInputOutputEqual_= CreateVariable();
94 0 : selfBit_ = 1 << rankId_;
95 0 : allBit_ = ((1 << rankSize_) - 1) & (~(1 << rankId_));
96 :
97 0 : srcMem_ = CreateMemory();
98 0 : dstMem_ = CreateMemory();
99 0 : locMask_ = CreateMaskSignal();
100 0 : localGoSize_ = CreateGroupOpSize();
101 0 : chunkOffset_ = CreateVariable();
102 0 : AllocGoResource(CCU_MS_LOCAL_COPY_LOOP_COUNT, LOCAL_COPY_MS_PER_LOOP);
103 : }
104 :
105 0 : void CcuContextReduceMeshMem2Mem1D::LoadArgs()
106 : {
107 0 : Load(input_[rankId_]);
108 0 : Load(output_[rankId_]);
109 0 : Load(token_[rankId_]);
110 0 : Load(isInputOutputEqual_);
111 0 : Load(inputRepeatStride_);
112 0 : Load(outputRepeatStride_);
113 0 : Load(normalSliceSize_);
114 0 : Load(lastSliceSize_);
115 0 : Load(repeatNumVar_);
116 0 : for (uint16_t i = 0; i < (rankSize_ - 1); i++) {
117 0 : Load(chunkSize_[i]);
118 : }
119 0 : Load(localGoSize_);
120 0 : }
121 :
122 0 : void CcuContextReduceMeshMem2Mem1D::PreSync()
123 : {
124 : // 互换内存信息
125 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMeshMem2Mem1D LocalPost begin");
126 0 : for (auto t : transports) {
127 0 : WriteVariableWithSignal(*t, input_[rankId_], INPUT_XN_ID, CKE_IDX_1, selfBit_); // index = 1，传递input信息
128 0 : WriteVariableWithSignal(*t, output_[rankId_], OUTPUT_XN_ID, CKE_IDX_2, selfBit_); // index = 0，传递output信息
129 0 : WriteVariableWithSignal(*t, token_[rankId_], TOKEN_XN_ID, CKE_IDX_3, selfBit_); // index = 2，传递token信息
130 : }
131 0 : GroupWait(*transportGroup, CKE_IDX_1, allBit_);
132 0 : GroupWait(*transportGroup, CKE_IDX_2, allBit_);
133 0 : GroupWait(*transportGroup, CKE_IDX_3, allBit_);
134 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMeshMem2Mem1D wait all end");
135 0 : }
136 :
137 0 : void CcuContextReduceMeshMem2Mem1D::PostSync()
138 : {
139 0 : for (auto t : transports) {
140 0 : RemotePost(*t, CKE_IDX_0, selfBit_);
141 : }
142 0 : GroupWait(*transportGroup, CKE_IDX_0, allBit_);
143 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMesh1D Reduce groupwait end");
144 0 : }
145 :
146 0 : void CcuContextReduceMeshMem2Mem1D::DoRepeatReduce(const std::vector<CcuRep::Variable> &srcAddr,
147 : const CcuRep::Variable &dstAddr)
148 : {
149 : // 从远程设备读取数据并逐步归约到本地设备
150 0 : CHK_PRT_THROW(
151 : srcAddr.size() != transports.size() + 1,
152 : HCCL_ERROR("[ReadReduceRmtToLoc] srcAddr.size[%zu] != transports size[%zu] +1", srcAddr.size(), transports.size()),
153 : InvalidParamsException, "Invalid srcAddr size");
154 :
155 0 : dstMem_.addr = dstAddr;
156 0 : dstMem_.token = token_[rankId_];
157 :
158 0 : srcMem_.addr = srcAddr[rankId_];
159 0 : srcMem_.token = token_[rankId_];
160 0 : CCU_IF (flag_ != 0) {
161 : // 非第一轮执行时，src 和 dst 已经初始化，需要添加偏移量
162 0 : dstMem_.addr += outputRepeatStride_;
163 0 : srcMem_.addr += inputRepeatStride_;
164 0 : }
165 0 : CCU_IF (isInputOutputEqual_ == 0) {
166 0 : GroupCopy(dstMem_, srcMem_, localGoSize_);
167 0 : }
168 0 : for (uint16_t i = 0; i < (rankSize_ - 1); i++) { // 外层循环控制step
169 : // 读不同rank的不同chunk
170 0 : for (uint16_t rmtId = 0; rmtId < rankSize_; ++rmtId) {
171 0 : if (rmtId == rootId_) {
172 0 : continue;
173 : }
174 0 : chunkOffset_ = 0;
175 0 : dstMem_.addr = dstAddr;
176 0 : srcMem_.addr = srcAddr[rmtId];
177 0 : srcMem_.token = token_[rmtId];
178 :
179 0 : CCU_IF (flag_ != 0) {
180 : // 非第一轮执行时，src 和 dst 已经初始化，需要添加偏移量
181 0 : dstMem_.addr += outputRepeatStride_;
182 0 : srcMem_.addr += inputRepeatStride_;
183 0 : }
184 0 : uint16_t chkId = 0;
185 0 : if (rmtId < rankId_) {
186 0 : chkId = (i + rmtId) % (rankSize_ - 1);
187 : } else {
188 0 : chkId = (i + rmtId - 1) % (rankSize_ - 1);
189 : }
190 0 : uint16_t transId = rmtId < rootId_ ? rmtId : rmtId - 1;
191 0 : HCCL_DEBUG(
192 : "[ReadReduceRmtToLoc] debug rankId[%llu], root[%llu] chkId[%llu], rmtId[%llu] transId[%llu]",
193 : rankId_, rootId_, chkId, rmtId, transId);
194 :
195 : // 计算一下offset 0~(chikd-1)
196 0 : for (uint16_t j = 0; j < chkId; ++j) {
197 0 : chunkOffset_ += chunkSize_[j];
198 : }
199 : // 更新对应的addr
200 0 : srcMem_.addr += chunkOffset_;
201 0 : dstMem_.addr += chunkOffset_;
202 :
203 0 : CCU_IF(chunkSize_[chkId] == 0)
204 : {
205 0 : LocalPost(locMask_, 1 << rmtId);
206 0 : }
207 :
208 0 : CCU_IF(chunkSize_[chkId] != 0)
209 : {
210 0 : ReadReduce(*transports[transId], dstMem_, srcMem_, chunkSize_[chkId], dataType_, reduceOp_, locMask_,
211 0 : 1 << rmtId);
212 0 : }
213 : }
214 0 : LocalWait(locMask_, allBit_);
215 : }
216 :
217 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMeshMem2Mem1D ReadReduce end");
218 0 : }
219 :
220 0 : void CcuContextReduceMeshMem2Mem1D::Algorithm()
221 : {
222 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMeshMem2Mem1D run");
223 0 : InitResource();
224 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMeshMem2Mem1D load input variables, id: [%u]", rankId_);
225 0 : LoadArgs();
226 0 : PreSync();
227 0 : CCU_IF(normalSliceSize_ != 0) // 所有rank
228 : {
229 0 : if (rankId_ == rootId_) {
230 0 : CcuRep::Variable repeatNumAdd = CreateVariable();
231 0 : repeatNumAdd = 1;
232 0 : flag_ = 0;
233 0 : CCU_WHILE(repeatNumVar_ != UINT64_MAX) { // 循环repeatNum_次
234 : // root要去读每个rank每个chunk的数据
235 0 : DoRepeatReduce(input_, output_[rankId_]);
236 0 : repeatNumVar_ += repeatNumAdd;
237 0 : flag_ = 1;
238 0 : }
239 0 : }
240 0 : }
241 0 : PostSync();
242 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] ReduceMeshMem2Mem1D end");
243 0 : return;
244 : }
245 :
246 :
247 0 : std::vector<uint64_t> CcuContextReduceMeshMem2Mem1D::GeneArgs(const CcuTaskArg &arg)
248 : {
249 0 : const CcuTaskArgReduceMeshMem2Mem1D *taskArg = dynamic_cast<const CcuTaskArgReduceMeshMem2Mem1D *>(&arg);
250 0 : if (taskArg == nullptr) {
251 0 : THROW<NullPtrException>(StringFormat("CcuContextReduceMeshMem2Mem1D::taskArg ptr is null"));
252 : }
253 0 : uint64_t inputAddr = taskArg->inputAddr_;
254 0 : uint64_t outputAddr = taskArg->outputAddr_;
255 0 : uint64_t tokenInfo = taskArg->token_;
256 :
257 0 : uint64_t bigDataSliceNum = taskArg->bigDataSliceNum_;
258 0 : uint64_t bigDataSliceSize = taskArg->bigDataSliceSize_;
259 0 : uint64_t smallDataSliceNum = taskArg->smallDataSliceNum_;
260 0 : uint64_t smallDataSliceSize = taskArg->smallDataSliceSize_;
261 0 : uint64_t inputRepeatStride = taskArg->inputRepeatStride_;
262 0 : uint64_t outputRepeatStride = taskArg->outputRepeatStride_;
263 0 : uint64_t normalSliceSize = taskArg->normalSliceSize_;
264 0 : uint64_t lastSliceSize = taskArg->lastSliceSize_;
265 0 : uint64_t repeatNumVar = taskArg->repeatNumVar_;
266 0 : uint64_t isInputOutputEqual = (inputAddr == outputAddr) ? 1: 0;
267 : std::vector<uint64_t> taskArgs = {
268 : inputAddr,
269 : outputAddr,
270 : tokenInfo,
271 : isInputOutputEqual,
272 : inputRepeatStride,
273 : outputRepeatStride,
274 : normalSliceSize,
275 : lastSliceSize,
276 : repeatNumVar,
277 0 : };
278 0 : for (uint64_t i = 0; i < bigDataSliceNum; i++) {
279 0 : taskArgs.push_back(bigDataSliceSize);
280 : }
281 0 : for (uint64_t i = 0; i < smallDataSliceNum; i++) {
282 0 : taskArgs.push_back(smallDataSliceSize);
283 : }
284 :
285 0 : auto localGoSize = CalGoSize(normalSliceSize);
286 0 : taskArgs.push_back(localGoSize[0]);
287 0 : taskArgs.push_back(localGoSize[1]);
288 0 : taskArgs.push_back(localGoSize[2]);
289 0 : taskArgs.push_back(localGoSize[3]);
290 0 : HCCL_INFO("[CcuContextReduceMeshMem2Mem1D] TaskArgs: inputAddr[%llu], outputAddr[%llu], inputRepeatStride[%llu], "
291 : "outputRepeatStride[%llu], normalSliceSize[%llu], lastSliceSize[%llu], repeatNumVar[%llu], "
292 : "bigDataSliceNum[%llu], bigDataSliceSize[%llu], smallDataSliceNum[%llu], smallDataSliceSize[%llu], ",
293 : inputAddr, outputAddr, inputRepeatStride, outputRepeatStride, normalSliceSize, lastSliceSize, repeatNumVar,
294 : bigDataSliceNum, bigDataSliceSize, smallDataSliceNum, smallDataSliceSize);
295 0 : return taskArgs;
296 0 : }
297 :
298 : } // namespace Hccl
|
