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_scatter_recursive_hd.h"
13 :
14 : namespace hccl {
15 5 : ReduceScatterRecursiveHalvingDoubling::ReduceScatterRecursiveHalvingDoubling(
16 5 : const HcclDispatcher dispatcher) : RecursiveHalvingDoublingBase(dispatcher)
17 : {
18 5 : }
19 :
20 10 : ReduceScatterRecursiveHalvingDoubling::~ReduceScatterRecursiveHalvingDoubling()
21 : {
22 10 : }
23 :
24 5 : HcclResult ReduceScatterRecursiveHalvingDoubling::Prepare(u64 reduceAttrBitMap, HcomCollOpInfo *opInfo)
25 : {
26 : (void)opInfo;
27 5 : reduceAttr = reduceAttrBitMap;
28 5 : return HCCL_SUCCESS;
29 : }
30 :
31 : // reducescatter recursiveHD 入口函数
32 0 : HcclResult ReduceScatterRecursiveHalvingDoubling::RunAsync(const u32 rank, const u32 rankSize,
33 : const std::vector<std::shared_ptr<Transport> > &links)
34 : {
35 0 : CHK_SMART_PTR_NULL(dispatcher_);
36 0 : CHK_PTR_NULL(stream_.ptr());
37 0 : HCCL_INFO("run: rank[%u] totalrank[%u] inputMem[%p] outputMem[%p] count[%llu]", rank, rankSize,
38 : inputMem_.ptr(), outputMem_.ptr(), count_);
39 0 : if (!outputMem_ || !inputMem_) {
40 0 : HCCL_ERROR("[ReduceScatterRecursiveHalvingDoubling][RunAsync]rank[%u] run_async inputmem or outputmem is null",
41 : rank);
42 0 : return HCCL_E_PTR;
43 : }
44 :
45 0 : HcclResult ret = HCCL_SUCCESS;
46 :
47 0 : if (rankSize == 1) {
48 0 : if (inputMem_ != outputMem_) {
49 0 : ret = HcclD2DMemcpyAsync(dispatcher_, outputMem_, inputMem_, stream_);
50 : }
51 0 : return ret;
52 : }
53 :
54 : // 创建reducer & sender
55 0 : senderInfo_.reset(new (std::nothrow) Sender(dataType_, reductionOp_, reduceAttr));
56 0 : CHK_SMART_PTR_NULL(senderInfo_);
57 :
58 0 : reducerInfo_.reset(new (std::nothrow) Reducer(dataType_, reductionOp_, reduceAttr));
59 0 : CHK_SMART_PTR_NULL(reducerInfo_);
60 :
61 0 : bool bRetSize = (links.size() < rankSize);
62 0 : CHK_PRT_RET(bRetSize, HCCL_ERROR("[ReduceScatterRecursiveHalvingDoubling][RunAsync]rank[%u] linksize[%llu] is "\
63 : "error", rank, links.size()), HCCL_E_INTERNAL);
64 :
65 0 : ret = CalcPartOneSizeAndBlockSize(rankSize);
66 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
67 : HCCL_ERROR("[ReduceScatterRecursiveHalvingDoubling][RunAsync]calculate part1size[%u] "\
68 : "and blocksize[%u] Failed! rankSize[%u]", part1Size_, blockSize_, rankSize), ret);
69 :
70 0 : HCCL_DEBUG("rank[%u] calculate par1size[%u] blocksize[%u] ranksize[%u]", rank, part1Size_, blockSize_, rankSize);
71 :
72 0 : CHK_RET(ReduceInPartOne(rank, links));
73 :
74 0 : CHK_RET(CalculateSlices(dataBytes_, rankSize));
75 :
76 0 : CHK_RET(ReduceScatterInBlock(rank, rankSize, links));
77 :
78 0 : CHK_RET(ScatterInPartOne(rank, rankSize, links));
79 :
80 0 : HCCL_INFO("ReduceScatterRecursiveHalvingDoubling finished: rank[%u]", rank);
81 0 : return HCCL_SUCCESS;
82 : }
83 :
84 :
85 0 : HcclResult ReduceScatterRecursiveHalvingDoubling::CalculateSlices(u64 dataBytes, const u32 rankSize) const
86 : {
87 0 : CHK_PRT_RET((blockSize_ == 0), HCCL_ERROR("[Calculate][Slices]calculate_slices para error"), HCCL_E_INTERNAL);
88 :
89 0 : slices_.resize(blockSize_);
90 0 : u64 bytesPerSlice = dataBytes / rankSize; // input大小 / server数 = 服务器内rank数*count (4p mesh以4*count为粒度)
91 0 : u32 i = 0;
92 0 : u32 halfPart1Size = (part1Size_ / 2); // 除2计算一半part1的大小
93 :
94 : /* 先给属于part1的block rank分配slice。每个rank有两份数据 */
95 0 : while (i < halfPart1Size) {
96 0 : slices_[i].size = 2 * bytesPerSlice; // 乘2计算2倍数据大小
97 0 : slices_[i].offset = i * 2 * bytesPerSlice; // 乘2计算2倍数据大小
98 0 : i++;
99 : }
100 :
101 : /* 再给剩余的block rank分配slice。每个rank有一份数据 */
102 0 : while (i < blockSize_) {
103 0 : slices_[i].size = bytesPerSlice;
104 0 : slices_[i].offset = (i * bytesPerSlice) + (halfPart1Size * bytesPerSlice);
105 0 : i++;
106 : }
107 :
108 0 : return HCCL_SUCCESS;
109 : }
110 :
111 0 : HcclResult ReduceScatterRecursiveHalvingDoubling::ReduceInPartOne(u32 rank, const std::vector<LINK> &links)
112 : {
113 0 : if (rank < part1Size_ && rank % 2 == 0) { // 模2判断奇偶性，rank属于第一部分，并且为偶数rank
114 0 : u32 peerRank = rank + 1;
115 0 : HCCL_DEBUG("rank[%u] outputmem receives from peerrank[%u] inputmem, offset[%llu], size[%llu]", \
116 : rank, peerRank, baseOffset_, scratchMem_.size());
117 0 : if (peerRank < links.size()) {
118 0 : CHK_SMART_PTR_NULL(links[peerRank]);
119 0 : HcclResult ret = links[peerRank]->TxAck(stream_);
120 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
121 : HCCL_ERROR("[Reduce][InPartOneToEven]rank[%u] tx ack from peerank[%u] failed", rank, peerRank), ret);
122 0 : ret = links[peerRank]->RxAck(stream_);
123 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
124 : HCCL_ERROR("[Reduce][InPartOneToEven]rank[%u] rx ack from peerank[%u] failed", rank, peerRank), ret);
125 : // 接收数据到本端的input
126 0 : HCCL_DEBUG("send mem[%p] size[%llu] to peerank[%u]", \
127 : scratchMem_.ptr(), scratchMem_.size(), peerRank);
128 0 : ret = links[peerRank]->TxAsync(UserMemType::INPUT_MEM, baseOffset_, scratchMem_.ptr(), 0, stream_);
129 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOneToEven]TxAsync: tx async size[%llu] "\
130 : "failed", 0), ret);
131 0 : CHK_RET(reducerInfo_->run(dispatcher_, links[peerRank], baseOffset_,
132 : inputMem_, inputMem_, scratchMem_, stream_));
133 0 : ret = links[peerRank]->RxWaitDone(stream_);
134 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]RxWaitDone failed"), ret);
135 : }
136 0 : } else if (rank < part1Size_ && rank % 2 == 1) { // 向上一个rank的input发数据 2
137 0 : u32 peerRank = rank - 1;
138 0 : if (peerRank < links.size()) {
139 0 : CHK_SMART_PTR_NULL(links[peerRank]);
140 0 : HcclResult ret = links[peerRank]->TxAck(stream_);
141 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
142 : HCCL_ERROR("[Reduce][InPartOneToEven]rank[%u] tx ack from peerank[%u] failed", rank, peerRank), ret);
143 0 : ret = links[peerRank]->RxAck(stream_);
144 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
145 : HCCL_ERROR("[Reduce][InPartOneToEven]rank[%u] rx ack from peerank[%u] failed", rank, peerRank), ret);
146 : // 发送到对端的input
147 0 : HCCL_DEBUG("rank[%u] sends inputMem[%p] to peerrank[%u] offset[%llu], size[%llu]", \
148 : rank, inputMem_.ptr(), peerRank, baseOffset_, inputMem_.size());
149 0 : ret = senderInfo_->run(links[peerRank], baseOffset_, inputMem_, stream_);
150 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]tx sync to peerank[%u] failed",
151 : peerRank), ret);
152 0 : ret = links[peerRank]->RxAsync(UserMemType::OUTPUT_MEM, baseOffset_, inputMem_.ptr(), 0, stream_);
153 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
154 : HCCL_ERROR("[AlgTemplateBase][ExecuteTxSync]ExecuteTxSync: rx async size[%llu] failed", 0), ret);
155 0 : ret = links[peerRank]->DataReceivedAck(stream_);
156 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
157 : HCCL_ERROR("[AlgTemplateBase][ExecuteTxSync]ExecuteTxSync: data received ack failed"), ret);
158 0 : ret = links[peerRank]->TxWaitDone(stream_);
159 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]TxWaitDone failed"), ret);
160 : }
161 : }
162 :
163 0 : return HCCL_SUCCESS;
164 : }
165 :
166 :
167 0 : HcclResult ReduceScatterRecursiveHalvingDoubling::ReduceScatterInBlock(u32 rank, u32 rankSize,
168 : const std::vector<LINK> &links)
169 : {
170 0 : u32 rankInBlock = 0;
171 0 : if (rank < part1Size_ && (rank % 2) == 1) { // rank号对2求余，rank为奇数
172 0 : return HCCL_SUCCESS;
173 0 : } else if (rank < part1Size_ && (rank % 2) == 0) { // rank对2求余，rank为偶数
174 0 : rankInBlock = rank / 2; // 直接除以2即为本rank的在block内的排序
175 : } else {
176 0 : rankInBlock = rank - part1Size_ / 2; // 通过rank减去part1除2的大小即不处于第一部分的block内rank号
177 : }
178 :
179 0 : std::unique_ptr<AlgTemplateBase> executor = AlgTemplateRegistry::Instance().GetAlgTemplate(
180 0 : TemplateType::TEMPLATE_REDUCESCATTER_HD, dispatcher_);
181 0 : CHK_SMART_PTR_NULL(executor);
182 0 : CHK_RET(executor->Prepare(inputMem_, inputMem_, scratchMem_, count_, dataType_, stream_,
183 : reductionOp_, root_, slices_, baseOffset_, blockSize_, reduceAttr,
184 : UserMemType::INPUT_MEM, UserMemType::OUTPUT_MEM));
185 :
186 0 : CHK_RET(executor->RegisterProfiler(profilerInput_.planeID, profilerInput_.stage, profilerInput_.step, stream_));
187 :
188 0 : std::vector<LINK> subLinks;
189 0 : CHK_RET(BuildSubLinks(links, subLinks, rankSize));
190 :
191 0 : CHK_PRT_RET(subLinks.size() == 0,
192 : HCCL_ERROR("[ReduceScatterRecursiveHalvingDoubling][ReduceScatterInBlock]rank[%u] "\
193 : "build sub links failed", rank), HCCL_E_PARA);
194 0 : CHK_RET(executor->RunAsync(rankInBlock, blockSize_, subLinks));
195 0 : return HCCL_SUCCESS;
196 0 : }
197 :
198 0 : HcclResult ReduceScatterRecursiveHalvingDoubling::ScatterInPartOne(u32 rank, u32 rankSize,
199 : const std::vector<LINK> &links)
200 : {
201 0 : u32 bytesPerData = DataUnitSize(dataType_);
202 0 : u64 dataBytes = count_ * bytesPerData;
203 0 : u64 bytesPerSlice = dataBytes / rankSize;
204 :
205 0 : if (rank < part1Size_ && rank % 2 == 0) { // 模2计算奇偶性，偶数rank把自己第二份数据给下一个奇数rank
206 0 : u32 peerRank = rank + 1;
207 0 : if (peerRank < links.size()) {
208 0 : CHK_SMART_PTR_NULL(links[peerRank]);
209 0 : HcclResult ret = links[peerRank]->TxAck(stream_);
210 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
211 : HCCL_ERROR("[Scatter][InPartOneToEven]rank[%u] tx ack from peerank[%u] failed", rank, peerRank), ret);
212 0 : ret = links[peerRank]->RxAck(stream_);
213 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
214 : HCCL_ERROR("[Scatter][InPartOneToEven]rank[%u] rx ack from peerank[%u] failed", rank, peerRank), ret);
215 : // 发送到对端的input
216 0 : HCCL_DEBUG("rank[%u] sends inputmem[%p] to peerrank[%u] Offset[%llu], Size[%llu]", \
217 : rank, inputMem_.ptr(), peerRank, baseOffset_, inputMem_.size());
218 :
219 0 : u64 offset = peerRank * bytesPerSlice; // 计算对端rank的slice偏移
220 0 : void *srcAddr = reinterpret_cast<s8 *>(inputMem_.ptr()) + offset;
221 0 : ret = ExecuteTxSync(links[peerRank], UserMemType::INPUT_MEM, offset + baseOffset_, srcAddr, bytesPerSlice,
222 0 : stream_);
223 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Scatter][InPartOne]tx sync to peerank[%u] failed",
224 : peerRank), ret);
225 0 : ret = links[peerRank]->TxWaitDone(stream_);
226 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]TxWaitDone failed"), ret);
227 : }
228 0 : } else if (rank < part1Size_ && rank % 2 == 1) { // 模2计算奇偶性，奇数rank接收偶数rank发过来下半份的数据
229 0 : u32 peerRank = rank - 1;
230 0 : if (peerRank < links.size()) {
231 0 : CHK_SMART_PTR_NULL(links[peerRank]);
232 0 : HcclResult ret = links[peerRank]->TxAck(stream_);
233 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
234 : HCCL_ERROR("[Scatter][InPartOneToEven]rank[%u] tx ack from peerank[%u] failed", rank, peerRank), ret);
235 0 : ret = links[peerRank]->RxAck(stream_);
236 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
237 : HCCL_ERROR("[Scatter][InPartOneToEven]rank[%u] rx ack from peerank[%u] failed", rank, peerRank), ret);
238 : // 接收数据到本端的 inputMem_
239 0 : HCCL_DEBUG("rx mem[%p] size[%llu] from peerank[%u]", inputMem_.ptr(), inputMem_.size(), peerRank);
240 :
241 0 : u64 offset = rank * bytesPerSlice; // 本rank slice偏移
242 0 : void *dstAddr = reinterpret_cast<s8 *>(inputMem_.ptr()) + offset;
243 0 : ret = ExecuteRxSync(links[peerRank], UserMemType::INPUT_MEM, offset + baseOffset_, dstAddr, bytesPerSlice,
244 0 : stream_);
245 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Scatter][InPartOne]rx sync from peerank[%u] failed",
246 : peerRank), ret);
247 0 : ret = links[peerRank]->RxWaitDone(stream_);
248 0 : CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[Reduce][InPartOne]RxWaitDone failed"), ret);
249 : }
250 : }
251 :
252 0 : return HCCL_SUCCESS;
253 : }
254 :
255 0 : HcclResult ReduceScatterRecursiveHalvingDoubling::GetNslbAdjInfo(const u32 rank, const u32 rankSize,
256 : const std::vector<LINK> &links,
257 : AdjInfo& nslbAdjInfo)
258 : {
259 0 : u32 nslbRound = 0;
260 0 : u32 base = 1;
261 0 : const u32 minExponent = 1;
262 0 : while ((base << nslbRound) <= rankSize) {
263 0 : nslbRound++;
264 : }
265 0 : if (nslbRound >= minExponent) {
266 0 : nslbRound = nslbRound - minExponent;
267 : }
268 0 : u32 nslbBlockSize = base << nslbRound;
269 : // 获取第一部分：rank数减block数乘2
270 0 : u32 nslbPart1Size = (rankSize - nslbBlockSize) * 2;
271 : // 2的次幂场景下处理流程
272 0 : if (nslbPart1Size == 0) {
273 0 : u32 stepNum = 0;
274 0 : while ((rankSize >> (stepNum + 1)) != 0) {
275 0 : stepNum++;
276 : }
277 0 : for (u32 step = 0; step < stepNum; step++) {
278 0 : u32 peerRankBitmask = 1 << (stepNum - step - 1);
279 0 : u32 peerRank = rank ^ peerRankBitmask;
280 0 : NslbDpAdjInfo adjInfoStep = {0};
281 0 : u32 remoteuserRank = links[peerRank]->GetRemoteRank();
282 0 : adjInfoStep.dstLocalRankId = remoteuserRank;
283 0 : adjInfoStep.phaseId = step + 1;
284 0 : adjInfoStep.rev = 0;
285 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
286 : }
287 0 : nslbAdjInfo.dstRankNum = stepNum;
288 0 : return HCCL_SUCCESS;
289 : }
290 : // 非2的次幂场景下，被合并部分的奇数rank处理流程
291 0 : if (rank < nslbPart1Size && rank % NSLBDP_REDUCE_SCATTER_MOLD2 == 1) {
292 0 : u32 peerRank = rank - 1;
293 0 : if (peerRank < links.size()) {
294 0 : NslbDpAdjInfo adjInfoStep = {0};
295 0 : adjInfoStep.dstLocalRankId = links[peerRank]->GetRemoteRank();
296 0 : adjInfoStep.phaseId = 1;
297 0 : adjInfoStep.rev = 0;
298 0 : HCCL_INFO("AllGatherHDR-nslb: peerRank[%u]", peerRank);
299 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
300 0 : nslbAdjInfo.dstRankNum = 1;
301 : }
302 0 : return HCCL_SUCCESS;
303 : }
304 : // 针对合并后映射成2的次幂场景处理
305 0 : u32 rankInBlock = 0;
306 0 : if (rank < nslbPart1Size && (rank % NSLBDP_REDUCE_SCATTER_MOLD2) == 0) {
307 0 : rankInBlock = rank / NSLBDP_REDUCE_SCATTER_MOLD2; // 直接除以2即为本rank的在block内的排序
308 : } else {
309 0 : rankInBlock = rank - nslbPart1Size / NSLBDP_REDUCE_SCATTER_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
310 : }
311 0 : std::vector<LINK> subLinks;
312 0 : std::vector<LINK>::const_iterator iter = links.begin();
313 0 : subLinks.resize(nslbBlockSize);
314 0 : for (u32 i = 0; i < rankSize; i++) {
315 0 : if (i < nslbPart1Size && (i % NSLBDP_REDUCE_SCATTER_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
316 0 : continue;
317 0 : } else if (i < nslbPart1Size && (i % NSLBDP_REDUCE_SCATTER_MOLD2) == 0) { // 模2余0代表当前rank在part1的偶数位置上
318 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
319 0 : if (niter != links.end()) {
320 0 : subLinks[i / NSLBDP_REDUCE_SCATTER_MOLD2] = *niter;
321 : }
322 0 : } else {
323 0 : std::vector<LINK>::const_iterator niter = std::next(iter, i);
324 0 : if (niter != links.end()) {
325 0 : subLinks[i - nslbPart1Size / NSLBDP_REDUCE_SCATTER_MOLD2] = *niter;
326 : }
327 : }
328 : }
329 0 : u32 stepNum = 0;
330 0 : while ((rankSize >> (stepNum + 1)) != 0) {
331 0 : stepNum++;
332 : }
333 : // 映射完成后针对以新的通信域进行邻接表获取
334 0 : u32 begin = 1;
335 0 : for (u32 step = 0; step < stepNum; step++) {
336 0 : u32 peerRankBitmask = 1 << (stepNum - step - 1);
337 0 : u32 peerRank = rankInBlock ^ peerRankBitmask;
338 0 : if (subLinks[peerRank] == nullptr) {
339 0 : continue;
340 : }
341 0 : NslbDpAdjInfo adjInfoStep = {0};
342 0 : u32 remoteuserRank = subLinks[peerRank]->GetRemoteRank();
343 0 : adjInfoStep.dstLocalRankId = remoteuserRank;
344 0 : adjInfoStep.phaseId = step + begin + 1;
345 0 : adjInfoStep.rev = 0;
346 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
347 : }
348 0 : nslbAdjInfo.dstRankNum = nslbAdjInfo.nsAdjInfo.size();
349 :
350 0 : if(nslbAdjInfo.nsAdjInfo.size() == 0) {
351 0 : return HCCL_SUCCESS;
352 : }
353 : // 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
354 0 : if (rank < nslbPart1Size && rank % NSLBDP_REDUCE_SCATTER_MOLD2 == 0) {
355 0 : u32 peerRank = rank + 1;
356 0 : uint16_t phaseSize = nslbAdjInfo.nsAdjInfo.size();
357 0 : if (peerRank < links.size()) {
358 0 : NslbDpAdjInfo adjInfoStep = {0};
359 0 : adjInfoStep.dstLocalRankId = links[peerRank]->GetRemoteRank();
360 0 : adjInfoStep.phaseId = nslbAdjInfo.nsAdjInfo[phaseSize - 1].phaseId + 1;
361 0 : adjInfoStep.rev = 0;
362 0 : HCCL_INFO("Scatter-nslb: peerRank[%u]", peerRank);
363 0 : nslbAdjInfo.nsAdjInfo.push_back(adjInfoStep);
364 0 : nslbAdjInfo.dstRankNum = nslbAdjInfo.nsAdjInfo.size();
365 : }
366 0 : return HCCL_SUCCESS;
367 : }
368 0 : return HCCL_SUCCESS;
369 0 : }
370 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_REDUCESCATTER_RECURSIVE_HD, ReduceScatterRecursiveHalvingDoubling);
371 : } // namespace hccl
|
