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 "reduce_recursive_hd.h"
13 :
14 : namespace hccl {
15 1 : ReduceRecursiveHalvingDoubling::ReduceRecursiveHalvingDoubling(const HcclDispatcher dispatcher)
16 1 : : RecursiveHalvingDoublingBase(dispatcher)
17 : {
18 1 : }
19 :
20 2 : ReduceRecursiveHalvingDoubling::~ReduceRecursiveHalvingDoubling()
21 : {
22 2 : }
23 :
24 1 : HcclResult ReduceRecursiveHalvingDoubling::Prepare(u64 reduceAttrBitMap, HcomCollOpInfo *opInfo)
25 : {
26 1 : reduceAttr = reduceAttrBitMap;
27 1 : return HCCL_SUCCESS;
28 : }
29 :
30 : // 算法的主入口
31 0 : HcclResult ReduceRecursiveHalvingDoubling::RunAsync(const u32 rank, const u32 rankSize,
32 : const std::vector<std::shared_ptr<Transport> > &links)
33 : {
34 0 : CHK_SMART_PTR_NULL(dispatcher_);
35 0 : CHK_PTR_NULL(stream_.ptr());
36 0 : if (!outputMem_ || !inputMem_) {
37 0 : HCCL_ERROR("[ReduceRecursiveHalvingDoubling][RunAsync]rank[%u] run_async inputmem or outputmem is null",
38 : rank);
39 0 : return HCCL_E_PTR;
40 : }
41 0 : HCCL_INFO("ReduceRecursiveHalvingDoubling run: rank[%u] root[%u] totalrank[%u] inputMem[%p] outputMem[%p]" \
42 : "count[%llu]", rank, root_, rankSize, inputMem_.ptr(), outputMem_.ptr(), count_);
43 :
44 0 : HcclResult ret = HCCL_SUCCESS;
45 :
46 0 : if (rankSize == 1) {
47 0 : if (inputMem_ != outputMem_) {
48 0 : ret = HcclD2DMemcpyAsync(dispatcher_, outputMem_, inputMem_, stream_);
49 : }
50 0 : return ret;
51 : }
52 :
53 0 : senderInfo_.reset(new (std::nothrow) Sender(dataType_, reductionOp_, reduceAttr));
54 0 : CHK_SMART_PTR_NULL(senderInfo_);
55 :
56 0 : reducerInfo_.reset(new (std::nothrow) Reducer(dataType_, reductionOp_, reduceAttr));
57 0 : CHK_SMART_PTR_NULL(reducerInfo_);
58 :
59 0 : bool bRetSize = (links.size() < rankSize);
60 0 : CHK_PRT_RET(bRetSize,
61 : HCCL_ERROR("[ReduceRecursiveHalvingDoubling][RunAsync]rank[%u] linksize[%llu] is error",
62 : rank, links.size()), HCCL_E_INTERNAL);
63 :
64 0 : CHK_RET(CalcPartOneSizeAndBlockSize(rankSize));
65 :
66 0 : u32 bytesPerData = DataUnitSize(dataType_);
67 0 : u64 dataBytes = count_ * bytesPerData;
68 0 : CHK_RET(CalculateSlices(dataBytes));
69 :
70 : // 结果完成需要放在input
71 0 : CHK_RET(ReduceInPartOne(rank, links));
72 :
73 : // 此步骤完成后，结果放在ouput中
74 0 : CHK_RET(ReduceScatterInBlock(rank, rankSize, links));
75 :
76 : // 使用output进行gather
77 0 : CHK_RET(GatherInBlock(rank, rankSize, links));
78 :
79 0 : HCCL_INFO("ReduceRecursiveHalvingDoubling rank[%u] finished", rank);
80 0 : return HCCL_SUCCESS;
81 : }
82 :
83 0 : HcclResult ReduceRecursiveHalvingDoubling::ReduceInPartOne(u32 rank, const std::vector<LINK> &links)
84 : {
85 0 : HCCL_INFO("rank[%u] part1Size_[%u] root[%u]", rank, part1Size_, root_);
86 :
87 0 : if (rank >= part1Size_) { // rank在第二部分，不参与ReduceInPartOne
88 0 : HCCL_INFO("rank[%u] not in part1, don't need reduce", rank);
89 0 : return HCCL_SUCCESS;
90 : }
91 : // root在第二部分，需要选取第一部分偶数rank接收，以0作为判断标准，否则在第一部分，与root奇偶性相同rank接收
92 0 : u32 rootFlag = (root_ >= part1Size_) ? 0 : root_;
93 :
94 0 : if (rank % 2 == rootFlag % 2) { // 1.从下一个rank接收数据到output，2. reduce到本rank的input
95 0 : u32 peerRank = (rank % 2) == 0 ? (rank + 1) : (rank - 1);
96 0 : HCCL_INFO("rank[%u] outputMem receives from PeerRank[%u] inputMem, Offset[%llu], Size[%llu]", \
97 : rank, peerRank, baseOffset_, outputMem_.size());
98 :
99 0 : if (peerRank < links.size()) {
100 0 : CHK_SMART_PTR_NULL(links[peerRank]);
101 :
102 0 : HcclResult ret = links[peerRank]->TxAck(stream_);
103 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
104 : HCCL_ERROR("[Reduce][InPartOne]tx ack to peerrank[%u] failed", peerRank), ret);
105 0 : ret = links[peerRank]->RxAck(stream_);
106 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
107 : HCCL_ERROR("[Reduce][InPartOne]rx ack from peerank[%u] failed", peerRank), ret);
108 :
109 : // 接收数据到本端的 output
110 0 : HCCL_DEBUG("send mem[%p] size[%llu] to peerank[%u]", outputMem_.ptr(), outputMem_.size(), peerRank);
111 0 : ret = links[peerRank]->TxAsync(UserMemType::INPUT_MEM, baseOffset_, outputMem_.ptr(), 0, stream_);
112 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOneToEven]TxAsync: tx async size[%llu] "\
113 : "failed", 0), ret);
114 0 : CHK_RET(reducerInfo_->run(dispatcher_, links[peerRank], baseOffset_,
115 : outputMem_, inputMem_, outputMem_, stream_));
116 0 : ret = links[peerRank]->RxWaitDone(stream_);
117 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]RxWaitDone failed"), ret);
118 : }
119 0 : } else if ((rank % 2) != (rootFlag % 2)) { // 向上一个rank的output发数据 2
120 0 : u32 peerRank = (rank % 2 == 0) ? (rank + 1) : (rank -1);
121 :
122 0 : if (peerRank < links.size()) {
123 0 : CHK_SMART_PTR_NULL(links[peerRank]);
124 0 : HcclResult ret = links[peerRank]->TxAck(stream_);
125 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
126 : HCCL_ERROR("[Reduce][InPartOne]tx ack to peerrank[%u] failed", peerRank), ret);
127 0 : ret = links[peerRank]->RxAck(stream_);
128 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
129 : HCCL_ERROR("[Reduce][InPartOne]rx ack from peerank[%u] failed", peerRank), ret);
130 : // 发送到对端的output
131 0 : HCCL_DEBUG("rank[%u] sends inputMem[%p] to PeerRank[%u] Offset[%llu], Size[%llu]", \
132 : rank, inputMem_.ptr(), peerRank, baseOffset_, inputMem_.size());
133 0 : ret = senderInfo_->run(links[peerRank], baseOffset_, inputMem_, stream_);
134 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
135 : HCCL_ERROR("[Reduce][InPartOne]tx sync to peerank[%u] failed", peerRank), ret);
136 0 : ret = links[peerRank]->RxAsync(UserMemType::OUTPUT_MEM, baseOffset_, inputMem_.ptr(), 0, stream_);
137 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
138 : HCCL_ERROR("[AlgTemplateBase][ExecuteTxSync]ExecuteTxSync: rx async size[%llu] failed", 0), ret);
139 0 : ret = links[peerRank]->DataReceivedAck(stream_);
140 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
141 : HCCL_ERROR("[AlgTemplateBase][ExecuteTxSync]ExecuteTxSync: data received ack failed"), ret);
142 0 : ret = links[peerRank]->TxWaitDone(stream_);
143 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]TxWaitDone failed"), ret);
144 : }
145 : }
146 0 : return HCCL_SUCCESS;
147 : }
148 :
149 :
150 0 : HcclResult ReduceRecursiveHalvingDoubling::ReduceScatterInBlock(u32 rank, u32 rankSize,
151 : const std::vector<LINK> &links)
152 : {
153 0 : u32 rankInBlock = 0;
154 :
155 0 : u32 rootFlag = (root_ >= part1Size_) ? 0 : root_;
156 0 : HCCL_DEBUG("[ReduceRecursiveHalvingDoubling][ReduceScatterInBlock]rootFlag is %u, rankInBlock is %u", rootFlag, rankInBlock);
157 : // 需要根据root判断，让root节点必然参加reducescatter,在第一部分的rank若与root奇偶性不同，直接返回
158 0 : if (rank < part1Size_ && (rank % 2) != (rootFlag % 2)) { // 模2判断奇偶性，本rank处于第一部分，奇偶性与root不同
159 0 : return HCCL_SUCCESS;
160 0 : } else if (rank < part1Size_) { // 模2判断奇偶性，本rank 处于第一部分，奇偶性与root相同
161 0 : rankInBlock = rank / 2; // 除2计算block内的rank值
162 : } else { // 本rank不属于第一部分
163 0 : rankInBlock = rank - part1Size_ / 2; // 除2计算block内的part1的范围
164 : }
165 : // 直接调用block的reducscatterhd算法
166 0 : std::unique_ptr<AlgTemplateBase> executor = AlgTemplateRegistry::Instance().GetAlgTemplate(
167 0 : TemplateType::TEMPLATE_REDUCESCATTER_HD, dispatcher_);
168 0 : CHK_SMART_PTR_NULL(executor);
169 0 : CHK_RET(executor->Prepare(inputMem_, outputMem_, outputMem_, count_, dataType_, stream_,
170 : reductionOp_, -1, slices_, baseOffset_, blockSize_, reduceAttr,
171 : UserMemType::INPUT_MEM, UserMemType::OUTPUT_MEM));
172 :
173 0 : CHK_RET(executor->RegisterProfiler(profilerInput_.planeID, profilerInput_.stage, profilerInput_.step,
174 : stream_));
175 :
176 : // 重新建立reducscatterscatter需要的链接
177 0 : std::vector<LINK> subLinks;
178 0 : CHK_RET(BuildRootSubLinks(links, subLinks, rankSize));
179 :
180 0 : CHK_PRT_RET(subLinks.size() == 0, HCCL_ERROR("[ReduceRecursiveHalvingDoubling][ReduceScatterInBlock]rank[%u] "\
181 : "BuildSubLinks failed", rank), HCCL_E_PARA);
182 :
183 0 : CHK_RET(executor->RunAsync(rankInBlock, blockSize_, subLinks));
184 :
185 0 : return HCCL_SUCCESS;
186 0 : }
187 :
188 0 : HcclResult ReduceRecursiveHalvingDoubling::CalculateStepSlices(const std::vector<Slice> &inputSlices, u32 stepNum,
189 : u32 rank, SliceType type, std::vector<Slice> &sliceOut)
190 : {
191 0 : std::vector<Slice> slice(stepNum);
192 :
193 0 : for (u32 step = 0; step < stepNum; step++) {
194 : // all-gather操作, halving_bitmask从低往高循环, size倍增
195 0 : u32 halvingBitmask = (1 << step);
196 0 : u32 peerRank = rank ^ halvingBitmask;
197 :
198 : // 计算tx_slice/rx_slice
199 0 : u32 sliceId = (type == SliceType::SLICE_TYPE_RX) ? \
200 0 : (peerRank & (~(halvingBitmask - 1))) : (rank & (~(halvingBitmask - 1)));
201 :
202 0 : slice[step].offset = inputSlices[sliceId].offset;
203 0 : CHK_RET(Sum(inputSlices, sliceId, halvingBitmask, slice[step].size));
204 :
205 0 : HCCL_DEBUG("Slice Info: rank[%u], slices[%u].offset=%llu, slices[%u].size=%llu", \
206 : rank, step, slice[step].offset, step, slice[step].size);
207 : }
208 :
209 0 : sliceOut = std::move(slice);
210 0 : return HCCL_SUCCESS;
211 0 : }
212 0 : HcclResult ReduceRecursiveHalvingDoubling::BuildRootSubLinks(const std::vector<LINK> &links,
213 : std::vector<LINK> &subLinks, u32 rankSize) const
214 : {
215 0 : std::vector<LINK>::const_iterator iter = links.begin();
216 0 : subLinks.resize(blockSize_);
217 0 : u32 rootFlag = (root_ >= part1Size_) ? 0 : root_;
218 0 : for (u32 i = 0; i < rankSize; i++) {
219 0 : if (i < part1Size_ && (i % 2) != rootFlag % 2) { // 模2与root模2比较代表当前rank在part1的内且与root奇偶性不同，不参与block内的建链
220 0 : continue;
221 0 : } else if (i < part1Size_) {
222 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
223 0 : if (niter != links.end()) {
224 0 : subLinks[i / 2] = *niter; // 除2计算出在block内的rank号
225 : }
226 : } else {
227 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
228 0 : if (niter != links.end()) {
229 0 : subLinks[i - part1Size_ / 2] = *niter; // rank在part2中，用原始rank减part1除2，计算出在block内的rank号
230 : }
231 : }
232 : }
233 :
234 0 : return HCCL_SUCCESS;
235 : }
236 : // 结果在output中，直接使用oupt进行数据收发
237 0 : HcclResult ReduceRecursiveHalvingDoubling::GatherInBlock(u32 rank, u32 rankSize,
238 : const std::vector<LINK> &links)
239 : {
240 0 : u32 rankInBlock = 0;
241 :
242 0 : u32 rootFlag = (root_ >= part1Size_) ? 0 : root_;
243 0 : if (rank < part1Size_ && (rank % 2) != (rootFlag % 2)) { // 模2判断奇偶性，本rank 处于第一部分，并且和root rank奇偶不同
244 0 : return HCCL_SUCCESS;
245 0 : } else if (rank < part1Size_) { // 模2判断奇偶性，本rank 处于第一部分，并且奇偶性和root相同
246 0 : rankInBlock = rank / 2; // 在block内的rank为实际rank除以2
247 : } else {
248 0 : rankInBlock = rank - part1Size_ / 2; // 除2计算block内的part1的范围
249 : }
250 0 : u32 rootInBlock = (root_ > part1Size_) ? (root_ - part1Size_ / 2) : (root_ / 2);
251 : // 重新建立gather需要的链接
252 0 : std::vector<LINK> subLinks;
253 :
254 0 : CHK_RET(BuildRootSubLinks(links, subLinks, rankSize));
255 :
256 0 : CHK_PRT_RET(subLinks.size() == 0,
257 : HCCL_ERROR("[Gather][InBlock]rank[%u] build sub links failed", rank), HCCL_E_PARA);
258 :
259 0 : CHK_RET(CalculateStepSlices(slices_, round_, rankInBlock, SliceType::SLICE_TYPE_TX, txSlices_));
260 :
261 0 : CHK_RET(CalculateStepSlices(slices_, round_, rankInBlock, SliceType::SLICE_TYPE_RX, rxSlices_));
262 :
263 0 : for (u32 step = 0; step < round_; step++) {
264 0 : u32 peerRankBitmask = (1 << step);
265 0 : u32 opBitmask = peerRankBitmask - 1 ; // 判断本轮是否进行收发
266 : // 断rank是否和root在同一轮次接收发送的block内，第一轮为total，第二轮为1/2，第三轮为1/4....
267 0 : if ((step != 0) && ((rankInBlock & opBitmask) != (rootInBlock & opBitmask))) {
268 0 : return HCCL_SUCCESS; // rank在本轮同root不在一个操作块内，不操作，直接返回
269 : }
270 0 : u32 peerRank = rankInBlock ^ peerRankBitmask;
271 0 : CHK_SMART_PTR_NULL(subLinks[peerRank]);
272 : // 再次判断是否和root在同一1/2，1/4,用来判断数据是收还是发
273 0 : if ((rankInBlock & peerRankBitmask) == (rootInBlock & peerRankBitmask)) {
274 0 : DeviceMem rxMem = outputMem_.range(rxSlices_[step].offset, rxSlices_[step].size);
275 0 : HcclResult ret = subLinks[peerRank]->TxAck(stream_);
276 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Gather][InBlock]rank[%u] tx ack from peerank[%u] failed",
277 : rank, peerRank), ret);
278 0 : ret = subLinks[peerRank]->RxAck(stream_);
279 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Gather][InBlock]rank[%u] rx ack from peerank[%u] failed",
280 : rank, peerRank), ret);
281 :
282 : // 等待对端可以接收数据
283 0 : HCCL_DEBUG("rank[%u] outputMem[%p] receive from PeerRank[%u] outputMem, Offset[%llu], "\
284 : "Size[%llu]", rank, outputMem_.ptr(), peerRank,
285 : baseOffset_ + rxSlices_[step].offset, rxSlices_[step].size);
286 :
287 0 : ret = ExecuteRxSync(subLinks[peerRank], UserMemType::OUTPUT_MEM, baseOffset_ + rxSlices_[step].offset,
288 0 : rxMem.ptr(), rxSlices_[step].size, stream_);
289 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
290 : HCCL_ERROR("[Gather][InBlock]rank[%u] rx sync from PeerRank[%u] failed", rank, peerRank), ret);
291 0 : ret = subLinks[peerRank]->RxWaitDone(stream_);
292 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]RxWaitDone failed"), ret);
293 0 : } else {
294 0 : DeviceMem txMem = outputMem_.range(txSlices_[step].offset, txSlices_[step].size);
295 0 : HcclResult ret = subLinks[peerRank]->TxAck(stream_);
296 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Gather][InBlock]rank[%u] tx ack from peerank[%u] failed",
297 : rank, peerRank), ret);
298 0 : ret = subLinks[peerRank]->RxAck(stream_);
299 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Gather][InBlock]rank[%u] rx ack from peerank[%u] failed",
300 : rank, peerRank), ret);
301 0 : HCCL_DEBUG("rank[%u] outputMem[%p] sends to peerrank[%u] outputmem, offset[%llu], " \
302 : "size[%llu]", rank, outputMem_.ptr(), peerRank,
303 : baseOffset_ + txSlices_[step].offset, txSlices_[step].size);
304 0 : ret = ExecuteTxSync(subLinks[peerRank], UserMemType::OUTPUT_MEM, baseOffset_ + txSlices_[step].offset,
305 0 : txMem.ptr(), txSlices_[step].size, stream_);
306 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Gather][InBlock]rank[%u] tx sync to PeerRank[%u] failed",
307 : rank, peerRank), ret);
308 0 : ret = subLinks[peerRank]->TxWaitDone(stream_);
309 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]TxWaitDone failed"), ret);
310 0 : }
311 : }
312 :
313 0 : return HCCL_SUCCESS;
314 0 : }
315 0 : HcclResult ReduceRecursiveHalvingDoubling::GetNslbAdjInfo(const u32 rank, const u32 rankSize,
316 : const std::vector<LINK> &links,
317 : AdjInfo& nslbAdjInfo)
318 : {
319 0 : u32 nslbRound = 0;
320 0 : u32 base = 1;
321 0 : const u32 minExponent = 1;
322 0 : while ((base << nslbRound) <= rankSize) {
323 0 : nslbRound++;
324 : }
325 0 : if (nslbRound >= minExponent) {
326 0 : nslbRound = nslbRound - minExponent;
327 : }
328 0 : u32 nslbBlockSize = base << nslbRound;
329 : // 获取第一部分：rank数减block数乘2
330 0 : u32 nslbPart1Size = (rankSize - nslbBlockSize) * NSLBDP_REDUCE_MOLD2;
331 : // 2的次幂场景下处理流程
332 0 : if (nslbPart1Size == 0) {
333 0 : u32 stepNum = 0;
334 0 : while ((rankSize >> (stepNum + 1)) != 0) {
335 0 : stepNum++;
336 : }
337 0 : HCCL_DEBUG("[ReduceRecursiveHalvingDoubling]GetNslbAdjInfo start");
338 0 : for (u32 step = 0; step < stepNum; step++) {
339 0 : u32 peerRankBitmask = 1 << (stepNum - step - 1);
340 0 : u32 peerRank = rank ^ peerRankBitmask;
341 0 : NslbDpAdjInfo adjInfoStep = {0};
342 0 : u32 remoteuserRank = links[peerRank]->GetRemoteRank();
343 0 : HCCL_DEBUG("[ReduceRecursiveHalvingDoubling]now step %u, remoteuserRank is %u", step, remoteuserRank);
344 0 : adjInfoStep.dstLocalRankId = remoteuserRank;
345 0 : adjInfoStep.phaseId = step + 1;
346 0 : adjInfoStep.rev = 0;
347 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
348 : }
349 0 : nslbAdjInfo.dstRankNum = stepNum;
350 0 : return HCCL_SUCCESS;
351 : }
352 : // 非2的次幂场景下，被合并部分的奇数rank处理流程
353 0 : if (rank < nslbPart1Size && rank % NSLBDP_REDUCE_MOLD2 == 1) {
354 0 : u32 peerRank = rank - 1;
355 0 : if (peerRank < links.size()) {
356 0 : NslbDpAdjInfo adjInfoStep = {0};
357 0 : adjInfoStep.dstLocalRankId = links[peerRank]->GetRemoteRank();
358 0 : adjInfoStep.phaseId = 1;
359 0 : adjInfoStep.rev = 0;
360 0 : HCCL_INFO("AllGatherHDR-nslb: peerRank[%u]", peerRank);
361 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
362 0 : nslbAdjInfo.dstRankNum = 1;
363 : }
364 0 : return HCCL_SUCCESS;
365 : }
366 : // 针对合并后映射成2的次幂场景处理
367 0 : u32 rankInBlock = 0;
368 0 : if (rank < nslbPart1Size && (rank % NSLBDP_REDUCE_MOLD2) == 0) {
369 0 : rankInBlock = rank / NSLBDP_REDUCE_MOLD2; // 直接除以2即为本rank的在block内的排序
370 : } else {
371 0 : rankInBlock = rank - nslbPart1Size / NSLBDP_REDUCE_MOLD2; // 通过rank减去part1除2的大小即不处于第一部分的block内rank号
372 : }
373 0 : std::vector<LINK> subLinks;
374 0 : std::vector<LINK>::const_iterator iter = links.begin();
375 0 : subLinks.resize(nslbBlockSize);
376 0 : for (u32 i = 0; i < rankSize; i++) {
377 0 : if (i < nslbPart1Size && (i % NSLBDP_REDUCE_MOLD2) == 1) { // 模2余1代表当前rank在part1的奇数位置上，不参与block内的建链
378 0 : continue;
379 0 : } else if (i < nslbPart1Size && (i % NSLBDP_REDUCE_MOLD2) == 0) { // 模2余0代表当前rank在part1的偶数位置上
380 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
381 0 : if (niter != links.end()) {
382 0 : subLinks[i / NSLBDP_REDUCE_MOLD2] = *niter;
383 : }
384 0 : } else {
385 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
386 0 : if (niter != links.end()) {
387 0 : subLinks[i - nslbPart1Size / NSLBDP_REDUCE_MOLD2] = *niter;
388 : }
389 : }
390 : }
391 0 : u32 stepNum = 0;
392 0 : while ((rankSize >> (stepNum + 1)) != 0) {
393 0 : stepNum++;
394 : }
395 : // 映射完成后针对以新的通信域进行邻接表获取
396 0 : for (u32 step = 0; step < stepNum; step++) {
397 0 : u32 peerRankBitmask = 1 << (stepNum - step - 1);
398 0 : u32 peerRank = rankInBlock ^ peerRankBitmask;
399 0 : if (subLinks[peerRank] == nullptr) {
400 0 : continue;
401 : }
402 0 : NslbDpAdjInfo adjInfoStep = {0};
403 0 : u32 remoteuserRank = subLinks[peerRank]->GetRemoteRank();
404 0 : adjInfoStep.dstLocalRankId = remoteuserRank;
405 0 : adjInfoStep.phaseId = step + 1;
406 0 : adjInfoStep.rev = 0;
407 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
408 : }
409 0 : nslbAdjInfo.dstRankNum = stepNum;
410 :
411 0 : if(nslbAdjInfo.nsAdjInfo.size() == 0) {
412 0 : return HCCL_SUCCESS;
413 : }
414 : // 上面处理完成后，紧接着处理合并部分的偶数rank同步到奇数rank增加phaseId
415 0 : if (rank < nslbPart1Size && rank % NSLBDP_REDUCE_MOLD2 == 0) {
416 0 : u32 peerRank = rank + 1;
417 0 : uint16_t phaseSize = nslbAdjInfo.nsAdjInfo.size();
418 0 : if (peerRank < links.size()) {
419 0 : NslbDpAdjInfo adjInfoStep = {0};
420 0 : adjInfoStep.dstLocalRankId = links[peerRank]->GetRemoteRank();
421 0 : adjInfoStep.phaseId = nslbAdjInfo.nsAdjInfo[phaseSize - 1].phaseId + 1;
422 0 : adjInfoStep.rev = 0;
423 0 : HCCL_INFO("Scatter-nslb: peerRank[%u]", peerRank);
424 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
425 0 : nslbAdjInfo.dstRankNum = nslbAdjInfo.nsAdjInfo.size();
426 : }
427 0 : return HCCL_SUCCESS;
428 : }
429 0 : return HCCL_SUCCESS;
430 0 : }
431 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_REDUCE_RECURSIVE_HALVING_DOUBLING, ReduceRecursiveHalvingDoubling);
432 : } // namespace hccl
|
