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 "alg_template_register.h"
12 : #include "all_reduce_mesh_opbase.h"
13 :
14 : namespace hccl {
15 3 : AllReduceMeshDirect::AllReduceMeshDirect(const HcclDispatcher dispatcher) : AlgTemplateBase(dispatcher)
16 3 : {}
17 :
18 6 : AllReduceMeshDirect::~AllReduceMeshDirect()
19 6 : {}
20 :
21 3 : HcclResult AllReduceMeshDirect::Prepare(u64 reduceAttrBitMap, std::vector<Stream> &meshStreams,
22 : std::vector<std::shared_ptr<LocalNotify>> &meshSignal, std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
23 : u32 interRank, u32 interRankSize, u32 userRank, HcomCollOpInfo *opInfo)
24 : {
25 3 : reduceAttr_ = reduceAttrBitMap;
26 3 : localRank_ = interRank;
27 3 : localRankSize_ = interRankSize;
28 3 : userRank_ = userRank;
29 3 : meshStreams_ = meshStreams;
30 3 : meshSignal_ = &meshSignal;
31 3 : meshSignalAux_ = &meshSignalAux;
32 3 : opInfo_ = opInfo;
33 3 : return HCCL_SUCCESS;
34 : }
35 :
36 0 : HcclResult AllReduceMeshDirect::MainRecordSub()
37 : {
38 0 : for (u32 signalIndex = 0; signalIndex < meshSignalAux_->size(); signalIndex++) {
39 0 : CHK_RET(LocalNotify::Post(stream_, dispatcher_, (*meshSignalAux_)[signalIndex],
40 : profilerInput_.stage));
41 : }
42 0 : return HCCL_SUCCESS;
43 : }
44 :
45 0 : HcclResult AllReduceMeshDirect::SubWaitMain()
46 : {
47 0 : for (u32 streamIndex = 0; streamIndex < meshSignalAux_->size(); streamIndex++) {
48 0 : CHK_RET(LocalNotify::Wait(meshStreams_[streamIndex], dispatcher_, (*meshSignalAux_)[streamIndex],
49 : profilerInput_.stage));
50 : }
51 0 : return HCCL_SUCCESS;
52 : }
53 :
54 0 : HcclResult AllReduceMeshDirect::MainWaitSub()
55 : {
56 0 : for (u32 signalIndex = 0; signalIndex < meshSignal_->size(); signalIndex++) {
57 0 : CHK_RET(LocalNotify::Wait(stream_, dispatcher_, (*meshSignal_)[signalIndex], profilerInput_.stage));
58 : }
59 0 : return HCCL_SUCCESS;
60 : }
61 :
62 0 : HcclResult AllReduceMeshDirect::SubRecordMain()
63 : {
64 0 : for (u32 streamIndex = 0; streamIndex < meshSignal_->size(); streamIndex++) {
65 0 : CHK_RET(LocalNotify::Post(meshStreams_[streamIndex], dispatcher_, (*meshSignal_)[streamIndex],
66 : profilerInput_.stage));
67 : }
68 0 : return HCCL_SUCCESS;
69 : }
70 :
71 : // 将数据均分，最小单位是128
72 3 : HcclResult AllReduceMeshDirect::PrepareSlice(u64 dataCount, u32 unitSize, u32 sliceNum, std::vector<Slice>& dataSlice)
73 : {
74 3 : Slice temp;
75 3 : u64 totalSize = dataCount * unitSize;
76 3 : dataSlice.clear();
77 3 : dataSlice.reserve(sliceNum);
78 3 : if (sliceNum == 0) {
79 0 : HCCL_ERROR("[Prepare][SliceData]data slice prepare, sliceNum is 0");
80 0 : return HCCL_E_PARA;
81 : }
82 3 : u64 sizePerSlice = (totalSize + sliceNum - 1) / sliceNum; /* 1是为了向上取整 */
83 3 : sizePerSlice = RoundUpWithDivisor(sizePerSlice, HCCL_MIN_SLICE_ALIGN_910B);
84 3 : u64 residueSize = totalSize;
85 3 : u32 i = 0;
86 6 : while (residueSize > 0) {
87 3 : u64 sliceSize = sizePerSlice < residueSize ? sizePerSlice : residueSize;
88 3 : temp.size = sliceSize;
89 3 : temp.offset = totalSize - residueSize;
90 3 : i++;
91 3 : if (sliceSize <= 0) {
92 0 : HCCL_ERROR("[Prepare][SliceData]data_slice_prepare sliceSize[%llu]", sliceSize);
93 0 : return HCCL_E_PARA;
94 : }
95 3 : residueSize -= sliceSize;
96 3 : dataSlice.push_back(temp);
97 : }
98 24 : while (i < sliceNum) {
99 21 : temp.size = 0;
100 21 : temp.offset = totalSize;
101 21 : i++;
102 21 : dataSlice.push_back(temp);
103 : }
104 3 : return HCCL_SUCCESS;
105 : }
106 :
107 0 : HcclResult AllReduceMeshDirect::PrepareAllreduceSliceData()
108 : {
109 0 : return PrepareSlice(count_, DataUnitSize(dataType_), localRankSize_, slices_);
110 : }
111 :
112 : // ringallreduce算法的函数入口
113 0 : HcclResult AllReduceMeshDirect::RunAsync(const u32 rank, const u32 rankSize, const std::vector<LINK> &links)
114 : {
115 0 : HcclResult ret = HCCL_SUCCESS;
116 0 : CHK_SMART_PTR_NULL(dispatcher_);
117 0 : CHK_PTR_NULL(stream_.ptr());
118 0 : HCCL_INFO("AllReduceMeshDirect run: rank[%u] ranksize[%u] inputMem[%p] outputMem[%p] count[%llu]", rank, rankSize,
119 : inputMem_.ptr(), outputMem_.ptr(), count_);
120 :
121 0 : if (links.size() < rankSize) {
122 0 : HCCL_ERROR("[AllReduceMeshDirect][RunAsync]rank[%u] linksize[%llu] is less than rankSize[%u]",
123 : rank, links.size(), rankSize);
124 0 : return HCCL_E_INTERNAL;
125 : }
126 :
127 : // 如果ranksize为1, inline reduce和普通跨片reduce操作一致，从input->output
128 0 : if (rankSize == 1) {
129 0 : if (opInfo_->inputAddr != opInfo_->outputAddr) {
130 0 : DeviceMem userMemIn = DeviceMem::create(opInfo_->inputAddr, count_ * DataUnitSize(dataType_));
131 0 : DeviceMem userMemOut = DeviceMem::create(opInfo_->outputAddr, count_ * DataUnitSize(dataType_));
132 0 : ret = HcclD2DMemcpyAsync(dispatcher_, userMemOut, userMemIn, stream_);
133 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[AllReduceRing][RunAsync]rank[%u] memcpy async failed", rank),
134 : ret);
135 0 : }
136 0 : return ret;
137 : }
138 :
139 0 : ret = PrepareAllreduceSliceData();
140 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
141 : HCCL_ERROR("[AllReduceRing][RunAsync]rank[%u] count[%llu] failed in PrepareSliceData "
142 : "step",
143 : rank, count_),
144 : ret);
145 :
146 0 : ret = RunReduceScatter(rank, rankSize, links);
147 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
148 : HCCL_ERROR("[AllReduceRing][RunAsync]rank[%u] count[%llu] failed in reducescater "
149 : "step",
150 : rank, count_),
151 : ret);
152 :
153 0 : ret = RunAllGather(rank, rankSize, links);
154 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
155 : HCCL_ERROR("[AllReduceRing][RunAsync]rank[%u] count[%llu] failed in AllGather "
156 : "step",
157 : rank, count_),
158 : ret);
159 :
160 0 : HCCL_INFO("AllReduceMeshDirect finished: rank[%u] ranksize[%u].", rank, rankSize);
161 0 : return HCCL_SUCCESS;
162 : }
163 :
164 :
165 0 : HcclResult AllReduceMeshDirect::RunReduceScatter(u32 rank, u32 rankSize, const std::vector<LINK> &links)
166 : {
167 0 : HCCL_INFO("ReduceScatterMeshAtomicOpbase run: rank[%u] totalrank[%u] inputMem[%p] outputMem[%p] count[%llu].", rank,
168 : rankSize, inputMem_.ptr(), outputMem_.ptr(), count_);
169 :
170 : // 数据准备
171 0 : u32 unitSize = DataUnitSize(dataType_);
172 :
173 0 : DeviceMem userMemIn = DeviceMem::create(opInfo_->inputAddr, count_ * unitSize);
174 0 : DeviceMem commMemOut = DeviceMem::create(outputMem_.ptr(), outputMem_.size());
175 :
176 0 : DeviceMem src;
177 0 : DeviceMem dst;
178 :
179 0 : src = DeviceMem::create(static_cast<char *>(opInfo_->inputAddr) + slices_[rank].offset, slices_[rank].size);
180 0 : dst = commMemOut.range(0, slices_[rank].size);
181 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, stream_));
182 :
183 0 : CHK_RET(MainRecordSub());
184 0 : CHK_RET(SubWaitMain());
185 :
186 0 : for (u32 round = 1; round < rankSize; round++) {
187 0 : u32 dstRank = (round + rank) % rankSize;
188 0 : Stream &subStream = meshStreams_[round - 1];
189 0 : CHK_RET(links[dstRank]->TxAck(subStream));
190 0 : CHK_RET(links[dstRank]->RxAck(subStream));
191 : }
192 :
193 0 : CHK_RET(SubRecordMain());
194 0 : CHK_RET(MainWaitSub());
195 0 : DeviceMem srcTmp = DeviceMem::create(inputMem_.ptr(), 0);
196 0 : DeviceMem dstTmp = DeviceMem::create(outputMem_.ptr(), 0);
197 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstTmp, srcTmp, stream_));
198 0 : CHK_RET(SubWaitMain());
199 0 : CHK_RET(MainRecordSub());
200 :
201 0 : for (u32 round = 1; round < rankSize; round++) {
202 0 : u32 dstRank = (round + rank) % rankSize;
203 0 : Stream &subStream = meshStreams_[round - 1];
204 :
205 0 : void *remMemPtr = nullptr;
206 0 : CHK_RET(links[dstRank]->GetRemoteMem(UserMemType::OUTPUT_MEM, &remMemPtr));
207 :
208 0 : dst = DeviceMem::create(static_cast<char *>(remMemPtr), slices_[dstRank].size);
209 0 : src = userMemIn.range(slices_[dstRank].offset, slices_[dstRank].size);
210 0 : CHK_RET(HcclReduceAsync(dispatcher_, static_cast<void *>(src.ptr()), slices_[dstRank].size / unitSize,
211 : dataType_, reductionOp_, subStream, static_cast<void *>(dst.ptr()),
212 : links[dstRank]->GetRemoteRank(), links[dstRank]->GetLinkType(), INLINE_REDUCE_BIT));
213 :
214 0 : CHK_RET(links[dstRank]->TxDataSignal(subStream));
215 0 : CHK_RET(links[dstRank]->RxDataSignal(subStream));
216 : }
217 :
218 0 : CHK_RET(SubRecordMain());
219 0 : CHK_RET(MainWaitSub());
220 0 : return HCCL_SUCCESS;
221 0 : }
222 :
223 0 : HcclResult AllReduceMeshDirect::RunAllGather(u32 rank, u32 rankSize, const std::vector<LINK> &links)
224 : {
225 0 : HCCL_INFO("AllGatherMesh run: rank[%u] totalrank[%u] inputMem[%p] outputMem[%p] count[%llu].", rank, rankSize,
226 : inputMem_.ptr(), outputMem_.ptr(), count_);
227 0 : u32 unitSize = DataUnitSize(dataType_);
228 :
229 0 : DeviceMem userMemOut = DeviceMem::create(opInfo_->outputAddr, count_ * unitSize);
230 0 : DeviceMem commMemOut = DeviceMem::create(outputMem_.ptr(), outputMem_.size());
231 :
232 0 : DeviceMem src;
233 0 : DeviceMem dst;
234 :
235 0 : src = commMemOut.range(0, slices_[rank].size);
236 0 : dst = userMemOut.range(slices_[rank].offset, slices_[rank].size);
237 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, stream_));
238 :
239 0 : CHK_RET(MainRecordSub());
240 0 : CHK_RET(SubWaitMain());
241 :
242 0 : for (u32 round = 1; round < rankSize; round++) {
243 0 : u32 dstRank = BackwardRank(rank, rankSize, round);
244 0 : Stream &subStream = meshStreams_[round - 1];
245 0 : CHK_RET(links[dstRank]->TxAck(subStream));
246 0 : CHK_RET(links[dstRank]->RxAck(subStream));
247 : }
248 :
249 0 : CHK_RET(SubRecordMain());
250 0 : CHK_RET(MainWaitSub());
251 0 : DeviceMem srcTmp = DeviceMem::create(inputMem_.ptr(), 0);
252 0 : DeviceMem dstTmp = DeviceMem::create(outputMem_.ptr(), 0);
253 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstTmp, srcTmp, stream_));
254 0 : CHK_RET(SubWaitMain());
255 0 : CHK_RET(MainRecordSub());
256 :
257 0 : for (u32 round = 1; round < rankSize; round++) {
258 0 : u32 dstRank = BackwardRank(rank, rankSize, round);
259 0 : Stream &subStream = meshStreams_[round - 1];
260 0 : void *remMemPtr = nullptr;
261 0 : CHK_RET(links[dstRank]->GetRemoteMem(UserMemType::OUTPUT_MEM, &remMemPtr));
262 0 : src = DeviceMem::create(static_cast<char *>(remMemPtr), slices_[dstRank].size);
263 0 : dst = userMemOut.range(slices_[dstRank].offset, slices_[dstRank].size);
264 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, subStream,
265 : links[dstRank]->GetRemoteRank(), links[dstRank]->GetLinkType()));
266 0 : CHK_RET(links[dstRank]->TxDataSignal(subStream));
267 0 : CHK_RET(links[dstRank]->RxDataSignal(subStream));
268 : }
269 :
270 0 : CHK_RET(SubRecordMain());
271 0 : CHK_RET(MainWaitSub());
272 :
273 0 : HCCL_INFO("AllGatherMesh finished: rank[%u]", rank);
274 0 : return HCCL_SUCCESS;
275 0 : }
276 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_ALL_REDUCE_MESH_DIRECT, AllReduceMeshDirect);
277 : } // namespace hccl
|
