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 "bcast_recursive_halvingdoubling.h"
12 : #include <cmath>
13 : #include "alg_template_register.h"
14 :
15 : namespace hccl {
16 0 : BcastRecursiveHalvingDoubling::BcastRecursiveHalvingDoubling(const HcclDispatcher dispatcher)
17 0 : : RecursiveHalvingDoublingBase(dispatcher), hasData_(false)
18 : {
19 0 : }
20 :
21 0 : BcastRecursiveHalvingDoubling::~BcastRecursiveHalvingDoubling()
22 : {
23 0 : }
24 :
25 : // recursiveHD broadcast算法主入口
26 0 : HcclResult BcastRecursiveHalvingDoubling::RunAsync(const u32 rank, const u32 rankSize,
27 : const std::vector<std::shared_ptr<Transport> > &links)
28 : {
29 0 : CHK_SMART_PTR_NULL(dispatcher_);
30 0 : CHK_PTR_NULL(stream_.ptr());
31 0 : CHK_PRT_RET(!inputMem_, HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] run_async inputmem is null",
32 : rank), HCCL_E_PTR);
33 :
34 0 : HCCL_INFO("BcastRecursiveHalvingDoubling run: rank[%u] rootRank[%u] totalrank[%u]"\
35 : " inputMem[%p] outputMem[%p] count[%llu]", \
36 : rank, root_, rankSize, inputMem_.ptr(), outputMem_.ptr(), count_);
37 :
38 0 : if (rankSize == 1) {
39 0 : return HCCL_SUCCESS;
40 : }
41 :
42 0 : if (rank == root_) {
43 0 : hasData_ = true;
44 : }
45 :
46 0 : CHK_PRT_RET(links.size() < rankSize,
47 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] linksize[%llu] is less than rankSize[%u]",
48 : rank, links.size(), rankSize), HCCL_E_INTERNAL);
49 :
50 : // 计算recursive算法第一部分相关参数
51 0 : HcclResult ret = CalcPartOneSizeAndBlockSize(rankSize);
52 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] Calculate "\
53 : "Par1Size[%u] And BlockSize[%u] Failed! rankSize[%u]", rank, part1Size_, blockSize_,
54 : rankSize), ret);
55 :
56 0 : HCCL_DEBUG("rank[%u] BroadcastInBlock... blockSize_[%u], part1Size_[%u]", \
57 : rank, blockSize_, part1Size_);
58 :
59 : // 先进行block内部的bcast
60 0 : CHK_RET(BroadcastInBlock(rank, links));
61 :
62 0 : HCCL_DEBUG("rank[%u] BroadcastOutOfBlock", rank);
63 :
64 0 : if (rank < part1Size_ && (rank % 2 == 0)) { // 模2是否为0判断rank奇偶性
65 0 : CHK_RET(EvenNumberRankProcess(rank, links));
66 0 : } else if (rank < part1Size_ && (rank % 2 == 1)) { // 模2是否为1判断rank奇偶性
67 0 : CHK_RET(OddNumberRankProcess(rank, links));
68 : }
69 :
70 0 : HCCL_INFO("BcastRecursiveHalvingDoubling finished: rank[%u] finished", rank);
71 0 : return HCCL_SUCCESS;
72 : }
73 :
74 0 : HcclResult BcastRecursiveHalvingDoubling::ReceiveData(const u32 destRank,
75 : const std::vector<std::shared_ptr<Transport> > &links)
76 : {
77 0 : if (destRank < links.size()) {
78 0 : if (links[destRank] == nullptr) {
79 0 : HCCL_ERROR("[Receive][Data]errNo[0x%016llx] links[destRank[%u]] ptr is NULL, return HCCL_E_PTR",
80 : HCCL_ERROR_CODE(HCCL_E_PTR), destRank);
81 0 : return HCCL_E_PTR;
82 : }
83 0 : HcclResult ret = links[destRank]->TxAck(stream_);
84 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Receive][Data]tx ack to dstrank[%u] failed", destRank), ret);
85 0 : ret = links[destRank]->RxAck(stream_);
86 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Receive][Data]rx ack to dstrank[%u] failed", destRank), ret);
87 :
88 0 : u64 dataBytes = count_ * DataUnitSize(dataType_);
89 0 : DeviceMem rcvMem = inputMem_.range(baseOffset_, dataBytes);
90 0 : HCCL_DEBUG("rx async from dstrank[%u] with rcvMem[%p] inputmem's offset[%llu] size[%llu]", \
91 : destRank, rcvMem.ptr(), baseOffset_, dataBytes);
92 :
93 0 : ret = ExecuteRxSync(links[destRank], UserMemType::INPUT_MEM, baseOffset_, rcvMem.ptr(), dataBytes, stream_);
94 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Receive][Data]rx sync from rank[%u] failed",
95 : destRank), ret);
96 0 : }
97 0 : return HCCL_SUCCESS;
98 : }
99 :
100 0 : HcclResult BcastRecursiveHalvingDoubling::SendData(const u32 destRank,
101 : const std::vector<std::shared_ptr<Transport> > &links)
102 : {
103 0 : if (destRank < links.size()) {
104 0 : if (links[destRank] == nullptr) {
105 0 : HCCL_ERROR("[Send][Data]errNo[0x%016llx] links[destRank[%u]] ptr is NULL, return HCCL_E_PTR",
106 : HCCL_ERROR_CODE(HCCL_E_PTR), destRank);
107 0 : return HCCL_E_PTR;
108 : }
109 0 : HcclResult ret = links[destRank]->TxAck(stream_);
110 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Send][Data]tx ack from rank[%u] failed", destRank), ret);
111 0 : ret = links[destRank]->RxAck(stream_);
112 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Send][Data]rx ack from rank[%u] failed", destRank), ret);
113 :
114 0 : u64 dataBytes = count_ * DataUnitSize(dataType_);
115 0 : DeviceMem sendMem = inputMem_.range(baseOffset_, dataBytes);
116 0 : HCCL_DEBUG("tx async to dstrank[%u] from sendMem[%p] inputmem's offset[%llu] size[%llu]", \
117 : destRank, sendMem.ptr(), baseOffset_, dataBytes);
118 :
119 0 : ret = ExecuteTxSync(links[destRank], UserMemType::INPUT_MEM, baseOffset_, sendMem.ptr(), dataBytes, stream_);
120 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Send][Data]tx sync to rank[%u] failed", destRank),
121 : ret);
122 0 : }
123 0 : return HCCL_SUCCESS;
124 : }
125 :
126 0 : u32 BcastRecursiveHalvingDoubling::GetRankIndexReal(const u32 rankInBlock) const
127 : {
128 0 : u32 res = 0;
129 : /* 如果根节点在第一部分的偶数位置或其他部分 */
130 0 : if ((root_ < part1Size_ && (root_ % 2) == 0) || root_ >= part1Size_) { // 模2判断奇偶性
131 0 : if (rankInBlock < part1Size_ / 2) { // 除2计算block内part1的rank范围
132 0 : res = rankInBlock * 2; // 乘2计算block内part1的rank范围
133 0 : return res;
134 : } else {
135 0 : res = part1Size_ / 2 + rankInBlock; // 除2加rankInBlock计算真实的rank值
136 0 : return res;
137 : }
138 : } else {
139 0 : if (rankInBlock < part1Size_ / 2) { // 除2计算block内part1的rank范围
140 0 : res = rankInBlock * 2 + 1; // 乘2加1计算计算真实的rank值
141 0 : return res;
142 : } else {
143 0 : res = part1Size_ / 2 + rankInBlock; // 除2加rankInBlock计算真实的rank值
144 0 : return res;
145 : }
146 : }
147 : }
148 :
149 0 : u32 BcastRecursiveHalvingDoubling::GetRankIndexInBlock(const u32 rank) const
150 : {
151 : // root在第一部分，并且root是偶数rank，或者root在第二部分
152 0 : if ((root_ < part1Size_ && (root_ % 2) == 0) || root_ >= part1Size_) { // 模2判断奇偶性
153 : // rank在第一部分，并且本rank是偶数rank，除以2就为在block内的index
154 0 : if (rank < part1Size_ && rank % 2 == 0) { // 模2判断奇偶性
155 0 : return rank / 2; // 除2计算block内rank值
156 0 : } else if (rank < part1Size_ && rank % 2 == 1) { // 模2判断奇偶性，奇数的话不在block内
157 0 : return INVALID_VALUE_RANKID;
158 : } else {
159 0 : return rank - part1Size_ / 2; // 除2计算block内part1的rank范围
160 : }
161 : } else { // root在第一部分属于奇数rank
162 0 : if (rank < part1Size_ && rank % 2 == 0) { // 模2判断奇偶性，偶数rank不在block内
163 0 : return INVALID_VALUE_RANKID;
164 0 : } else if (rank < part1Size_ && rank % 2 == 1) { // 模2判断奇偶性，为奇计算block内rank号
165 0 : return (rank - 1) / 2; // 通过减1再除2得到block内rank号
166 : } else {
167 0 : return rank - part1Size_ / 2; // 除2计算block内part1的rank范围
168 : }
169 : }
170 : }
171 :
172 0 : HcclResult BcastRecursiveHalvingDoubling::BroadcastInBlock(const u32 rank,
173 : const std::vector<std::shared_ptr<Transport> > &links)
174 : {
175 0 : u32 rankInBlock = GetRankIndexInBlock(rank);
176 0 : if (rankInBlock == INVALID_VALUE_RANKID) { // 非block内的节点，不做操作
177 0 : return HCCL_SUCCESS;
178 : }
179 :
180 0 : u32 rootInBlock = GetRankIndexInBlock(root_);
181 0 : for (u32 i = 0; i < round_; i++) {
182 0 : u32 peerRankBitmask = 1 << (round_ - i - 1); // 进入此条件，round必然不小于1
183 0 : u32 peerRankInBlock = rankInBlock ^ peerRankBitmask;
184 0 : u32 andOprand = (1 << (round_ - i - 1)) - 1;
185 0 : u32 peerRankReal = GetRankIndexReal(peerRankInBlock);
186 :
187 0 : HcclResult ret = HCCL_SUCCESS;
188 : // 本rank在第round轮需要接收数据
189 0 : if (((rankInBlock & andOprand) == (rootInBlock & andOprand)) && (rank != root_) &&
190 0 : !hasData_) {
191 0 : HCCL_DEBUG("rank[%u] receive memsize[%llu] from rank[%u] in round[%u]", \
192 : rank, DataUnitSize(dataType_)*count_, peerRankReal, i);
193 0 : ret = ReceiveData(peerRankReal, links);
194 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[BcastRecursiveHalvingDoubling][BroadcastInBlock]rank[%u] "\
195 : "Receive Data from rank[%u] failed.", rank, peerRankReal), ret);
196 0 : hasData_ = true;
197 0 : if (peerRankReal < links.size()) {
198 0 : ret = links[peerRankReal]->RxWaitDone(stream_);
199 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
200 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][BroadcastInBlock]RxWaitDone failed"), ret);
201 : }
202 0 : continue;
203 : }
204 :
205 : // 需要向目的rank发送数据，前提是收到数据后(root 节点每轮都发)
206 0 : if (hasData_) {
207 0 : HCCL_DEBUG("rank[%u] send mem[%llu] to rank[%u] in round:%u", \
208 : rank, DataUnitSize(dataType_)*count_, peerRankReal, i);
209 0 : ret = SendData(peerRankReal, links);
210 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
211 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][BroadcastInBlock]rank[%u] Send "\
212 : "Data to rank[%u] failed.", rank, peerRankReal), ret);
213 : }
214 0 : if (peerRankReal < links.size()) {
215 0 : ret = links[peerRankReal]->TxWaitDone(stream_);
216 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
217 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][BroadcastInBlock]TxWaitDone failed"), ret);
218 : }
219 : }
220 0 : return HCCL_SUCCESS;
221 : }
222 :
223 0 : HcclResult BcastRecursiveHalvingDoubling::EvenNumberRankProcess(const u32 rank,
224 : const std::vector<std::shared_ptr<Transport> > &links)
225 : {
226 : HcclResult ret;
227 0 : if (root_ % 2 == 0 || root_ >= part1Size_) { // 模2是否为0判断rank_奇偶性
228 0 : HCCL_DEBUG("rank[%u] stream[%p] send memsize[%llu] to rank[%u]", \
229 : rank, stream_.ptr(), DataUnitSize(dataType_)*count_, rank + 1);
230 : // 该rank需要向第一部分的后续奇数rank发送数据
231 0 : ret = SendData(rank + 1, links);
232 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
233 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] stream[%p] Send data to "\
234 : "Rank[%u] failed", rank, stream_.ptr(), rank + 1), ret);
235 0 : if (rank + 1 < links.size()) {
236 0 : ret = links[rank + 1]->TxWaitDone(stream_);
237 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
238 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]TxWaitDone failed"), ret);
239 : }
240 : } else {
241 0 : HCCL_DEBUG("rank[%u] stream[%p] receive memsize[%llu] from rank[%u]", \
242 : rank, stream_.ptr(), DataUnitSize(dataType_)*count_, rank + 1);
243 :
244 : // root为奇数，本rank为偶数，需要从邻接的奇数rank接收数据
245 0 : ret = ReceiveData(rank + 1, links);
246 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
247 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] stream[%p] Receive data "\
248 : "from Rank[%u] failed", rank, stream_.ptr(), rank + 1), ret);
249 0 : if (rank + 1 < links.size()) {
250 0 : ret = links[rank + 1]->RxWaitDone(stream_);
251 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
252 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]RxWaitDone failed"), ret);
253 : }
254 : }
255 0 : return HCCL_SUCCESS;
256 : }
257 :
258 0 : HcclResult BcastRecursiveHalvingDoubling::OddNumberRankProcess(const u32 rank,
259 : const std::vector<std::shared_ptr<Transport> > &links)
260 : {
261 : HcclResult ret;
262 0 : if (root_ % 2 == 0 || root_ >= part1Size_) { // 模2是否为0判断rank_奇偶性
263 0 : HCCL_DEBUG("rank[%u] stream[%p] receive memsize[%llu] from rank[%u]", \
264 : rank, stream_.ptr(), DataUnitSize(dataType_)*count_, rank - 1);
265 :
266 : // root是偶数节点，rank从前面邻接的偶数节点接收数据
267 0 : ret = ReceiveData(rank - 1, links);
268 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
269 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] stream[%p] Receive data "\
270 : "from Rank[%u] failed", rank, stream_.ptr(), rank - 1), ret);
271 0 : if (rank - 1 < links.size()) {
272 0 : ret = links[rank - 1]->RxWaitDone(stream_);
273 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
274 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]RxWaitDone failed"), ret);
275 : }
276 : } else {
277 0 : HCCL_DEBUG("rank[%u] stream[%p] send memsize[%llu] to rank[%u]", \
278 : rank, stream_.ptr(), DataUnitSize(dataType_)*count_, rank - 1);
279 0 : ret = SendData(rank - 1, links);
280 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
281 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]rank[%u] stream[%p] Send data to "\
282 : "Rank[%u] failed", rank, stream_.ptr(), rank - 1), ret);
283 0 : if (rank - 1 < links.size()) {
284 0 : ret = links[rank - 1]->TxWaitDone(stream_);
285 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
286 : HCCL_ERROR("[BcastRecursiveHalvingDoubling][RunAsync]TxWaitDone failed"), ret);
287 : }
288 : }
289 0 : return HCCL_SUCCESS;
290 : }
291 0 : HcclResult BcastRecursiveHalvingDoubling::GetNslbAdjInfo(const u32 rank, const u32 rankSize,
292 : const std::vector<LINK> &links, AdjInfo& nslbAdjInfo)
293 : {
294 0 : u32 nslbRound = 0;
295 0 : u32 base = 1;
296 0 : const u32 minExponent = 1;
297 0 : while ((base << nslbRound) <= rankSize) {
298 0 : nslbRound++;
299 : }
300 0 : if (nslbRound >= minExponent) {
301 0 : nslbRound = nslbRound - minExponent;
302 : }
303 0 : u32 nslbBlockSize = base << nslbRound;
304 : // 获取第一部分：rank数减block数乘2
305 0 : u32 nslbPart1Size = (rankSize - nslbBlockSize) * NSLBDP_BCAST_MOLD2;
306 :
307 0 : u32 rankInBlock = 0;
308 0 : if (rank < nslbPart1Size && (rank % NSLBDP_BCAST_MOLD2) == 0) {
309 0 : rankInBlock = rank / NSLBDP_BCAST_MOLD2; // 直接除以2即为本rank的在block内的排序
310 : } else {
311 0 : rankInBlock = rank - nslbPart1Size / NSLBDP_BCAST_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
312 : }
313 : // 2的次幂场景下处理流程
314 0 : if (nslbPart1Size == 0) {
315 0 : u32 stepNum = 0;
316 0 : while ((rankSize >> (stepNum + 1)) != 0) {
317 0 : stepNum++;
318 : }
319 0 : bool begin = false;
320 0 : if (rank == 0) {
321 0 : begin = true;
322 : }
323 0 : for (u32 step = 0; step < stepNum; step++) {
324 0 : u32 peerRankBitmask = 1 << (stepNum - step - 1); // 进入此条件，round必然不小于1
325 0 : u32 peerRankInBlock = rankInBlock ^ peerRankBitmask;
326 0 : u32 andOprand = (1 << (stepNum - step - 1)) - 1;
327 :
328 : // 本rank在第round轮需要接收数据
329 0 : if (((rankInBlock & andOprand) == 0) && (rank != 0) && !begin) {
330 0 : begin = true;
331 0 : continue;
332 : }
333 0 : if (begin) {
334 0 : NslbDpAdjInfo adjInfoStep = {0};
335 0 : u32 remoteuserRank = links[peerRankInBlock]->GetRemoteRank();
336 0 : adjInfoStep.dstLocalRankId = remoteuserRank;
337 0 : adjInfoStep.phaseId = step + 1;
338 0 : adjInfoStep.rev = 0;
339 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
340 : }
341 : }
342 0 : nslbAdjInfo.dstRankNum = nslbAdjInfo.nsAdjInfo.size();
343 0 : return HCCL_SUCCESS;
344 : }
345 : // 非2的次幂场景下，被合并部分的奇数rank处理流程
346 0 : if (rank < nslbPart1Size && rank % NSLBDP_BCAST_MOLD2 == 1) {
347 0 : u32 peerRank = rank - 1;
348 0 : if (peerRank < links.size()) {
349 0 : NslbDpAdjInfo adjInfoStep = {0};
350 0 : adjInfoStep.dstLocalRankId = links[peerRank]->GetRemoteRank();
351 0 : adjInfoStep.phaseId = 1;
352 0 : adjInfoStep.rev = 0;
353 0 : HCCL_INFO("AllGatherHDR-nslb: peerRank[%u]", peerRank);
354 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
355 0 : nslbAdjInfo.dstRankNum = 1;
356 : }
357 0 : return HCCL_SUCCESS;
358 : }
359 :
360 0 : std::vector<LINK> subLinks;
361 0 : std::vector<LINK>::const_iterator iter = links.begin();
362 0 : subLinks.resize(nslbBlockSize);
363 0 : for (u32 i = 0; i < rankSize; i++) {
364 0 : if (i < nslbPart1Size && (i % NSLBDP_BCAST_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
365 0 : continue;
366 0 : } else if (i < nslbPart1Size && (i % NSLBDP_BCAST_MOLD2) == 0) { // 模2余0代表当前rank在part1的偶数位置上
367 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
368 0 : if (niter != links.end()) {
369 0 : subLinks[i / NSLBDP_BCAST_MOLD2] = *niter;
370 : }
371 0 : } else {
372 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
373 0 : if (niter != links.end()) {
374 0 : subLinks[i - nslbPart1Size / NSLBDP_BCAST_MOLD2] = *niter;
375 : }
376 : }
377 : }
378 0 : u32 stepNum = 0;
379 0 : while ((nslbBlockSize >> (stepNum + 1)) != 0) {
380 0 : stepNum++;
381 : }
382 : // 映射完成后针对以新的通信域进行邻接表获取
383 0 : u32 begin = 1;
384 0 : for (u32 step = 0; step < stepNum; step++) {
385 0 : u32 peerRankBitmask = 1 << (stepNum - step - 1);
386 0 : u32 peerRank = rankInBlock ^ peerRankBitmask;
387 0 : NslbDpAdjInfo adjInfoStep = {0};
388 0 : if (subLinks[peerRank] == nullptr) {
389 0 : continue;
390 : }
391 0 : u32 remoteuserRank = subLinks[peerRank]->GetRemoteRank();
392 0 : adjInfoStep.dstLocalRankId = remoteuserRank;
393 0 : adjInfoStep.phaseId = step + begin + 1;
394 0 : adjInfoStep.rev = 0;
395 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
396 : }
397 0 : nslbAdjInfo.dstRankNum = nslbAdjInfo.nsAdjInfo.size();
398 0 : HCCL_DEBUG("[BcastRecursiveHalvingDoubling]dstRankNum is %u", nslbAdjInfo.dstRankNum);
399 :
400 0 : if(nslbAdjInfo.nsAdjInfo.size() == 0) {
401 0 : return HCCL_SUCCESS;
402 : }
403 : // 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
404 0 : if (rank < nslbPart1Size && rank % NSLBDP_BCAST_MOLD2 == 0) {
405 0 : u32 peerRank = rank + 1;
406 0 : uint16_t phaseSize = nslbAdjInfo.nsAdjInfo.size();
407 0 : if (peerRank < links.size()) {
408 0 : NslbDpAdjInfo adjInfoStep = {0};
409 0 : adjInfoStep.dstLocalRankId = links[peerRank]->GetRemoteRank();
410 0 : adjInfoStep.phaseId = nslbAdjInfo.nsAdjInfo[phaseSize - 1].phaseId + 1;
411 0 : adjInfoStep.rev = 0;
412 0 : HCCL_INFO("Scatter-nslb: peerRank[%u]", peerRank);
413 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
414 0 : nslbAdjInfo.dstRankNum = nslbAdjInfo.nsAdjInfo.size();
415 : }
416 0 : return HCCL_SUCCESS;
417 : }
418 0 : return HCCL_SUCCESS;
419 0 : }
420 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_BROADCAST_RECURSIVE_HD, BcastRecursiveHalvingDoubling);
421 : } // namespace hccl
|
