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 "coll_reduce_scatter_mesh_opbase_small_count_deterministic_executor.h"
12 :
13 : const u32 RANK_SIZE_FOUR = 4;
14 : namespace hccl {
15 : // 准入条件: 确定性&小数据量
16 0 : CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor(const HcclDispatcher dispatcher,
17 0 : std::unique_ptr<TopoMatcher> &topoMatcher): CollReduceScatterExecutor(dispatcher, topoMatcher)
18 : {
19 0 : DMAReduceFlag_ = true;
20 0 : CCLMemSlice_ = false;
21 0 : }
22 :
23 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CalcStreamNum(u32& streamNum)
24 : {
25 : u32 totalStreamNum;
26 0 : if (IsPowerOfTwo(topoAttr_.deviceNumPerAggregation)) {
27 : // level0 为HD staged
28 0 : totalStreamNum = 2U;
29 : } else {
30 0 : totalStreamNum = 1U;
31 : }
32 :
33 0 : streamNum = totalStreamNum - 1U;
34 0 : HCCL_INFO("[CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
35 : tag_.c_str(), streamNum);
36 0 : return HCCL_SUCCESS;
37 : }
38 :
39 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
40 : {
41 0 : TransportMemType inputType = TransportMemType::RESERVED;
42 0 : TransportMemType outputType = TransportMemType::RESERVED;
43 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
44 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
45 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
46 0 : return HCCL_SUCCESS;
47 : }
48 :
49 0 : bool CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::IsPowerOfTwo(u32 num)
50 : {
51 0 : return (num & (num - 1)) == 0;
52 : }
53 :
54 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CalcTransportMemType(TransportMemType &inputType,
55 : TransportMemType &outputType)
56 : {
57 0 : inputType = TransportMemType::CCL_INPUT;
58 0 : outputType = TransportMemType::CCL_OUTPUT;
59 0 : HCCL_INFO("[CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor][CalcTransportMemType]" \
60 : "tag[%s] inputType[%d], outputType[%d]",
61 : tag_.c_str(), inputType, outputType);
62 0 : return HCCL_SUCCESS;
63 : }
64 :
65 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CalcLevel0CommInfo(TransportMemType inputType,
66 : TransportMemType outputType,
67 : std::vector<LevelNSubCommTransport>& opTransport)
68 : {
69 : CommType commType;
70 0 : if (topoAttr_.deviceNumPerAggregation >= RANK_SIZE_FOUR && IsPowerOfTwo(topoAttr_.deviceNumPerAggregation)) {
71 : // HD stage
72 0 : commType = CommType::COMM_TAG_HALVING_DOUBLING;
73 : } else {
74 : // NHR
75 0 : commType = CommType::COMM_TAG_WHOLE_NHR;
76 : }
77 0 : CommParaInfo commParaInfo(COMM_LEVEL0, commType);
78 0 : commParaInfo.meshSinglePlane = false;
79 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL0], inputType, outputType));
80 0 : return HCCL_SUCCESS;
81 0 : }
82 :
83 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CalcLevel1CommInfo(TransportMemType inputType,
84 : TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
85 : {
86 0 : HCCL_INFO("[%s][CalcLevel1CommInfo]tag[%s] start", __func__, tag_.c_str());
87 0 : CommParaInfo commParaLevel1(COMM_LEVEL1, CommType::COMM_TAG_MAX);
88 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
89 0 : commParaLevel1.commType = CommType::COMM_TAG_RING_INNER;
90 0 : HCCL_INFO("[%s][CalcLevel1CommInfo]tag[%s] Calc RingCommInfo", __func__, tag_.c_str());
91 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
92 0 : commParaLevel1.commType = CommType::COMM_TAG_NONUNIFORM_HIERARCHICAL_RING;
93 0 : HCCL_INFO("[%s][CalcLevel1CommInfo]tag[%s] Calc NHRCommInfo", __func__, tag_.c_str());
94 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
95 0 : commParaLevel1.commType = CommType::COMM_TAG_NONUNIFORM_BRUCK;
96 0 : HCCL_INFO("[%s][CalcLevel1CommInfo]tag[%s] Calc NBCommInfo", __func__, tag_.c_str());
97 : } else {
98 0 : commParaLevel1.commType = CommType::COMM_TAG_HALVING_DOUBLING;
99 0 : HCCL_INFO("[%s][CalcLevel1CommInfo]tag[%s] Calc HDCommInfo", __func__, tag_.c_str());
100 : }
101 0 : commParaLevel1.forceRdma = false;
102 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel1, opTransport[commParaLevel1.commPlane], inputType, outputType));
103 0 : HCCL_INFO("[%s][CalcLevel1CommInfo]tag[%s] Calc CommInfo Finish", __func__, tag_.c_str());
104 :
105 0 : return HCCL_SUCCESS;
106 0 : }
107 :
108 0 : u64 CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CalcLoopMaxCount(const u32 unitSize)
109 : {
110 : // 中转内存单次最多能够接受的output count
111 0 : u64 maxCountPerLoop = inCCLbufferSize_ / (topoAttr_.userRankSize * unitSize);
112 0 : return maxCountPerLoop;
113 : }
114 :
115 0 : bool CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::IsHugeData(const u64 curSize, OpParam *param)
116 : {
117 0 : bool hugeData = (curSize * topoAttr_.userRankSize / HCCL_INTERNODE_MAX_DATA_RATE > RDMA_SEND_MAX_SIZE) ||
118 : (curSize > SDMA_SEND_MAX_SIZE);
119 0 : return hugeData;
120 : }
121 :
122 0 : bool CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::IsSmallData(const u64 totalSize, const u64 curSize)
123 : {
124 : // 小数据量才选到该执行器，默认为true
125 0 : return true;
126 : }
127 :
128 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
129 : {
130 0 : HCCL_CONFIG_INFO(HCCL_ALG,
131 : "[CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor][Run]CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor begins.");
132 0 : u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
133 0 : u64 curSize = execMem.count * unitSize; // 单位：字节
134 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
135 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
136 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, level0CommInfo.localRank + 1));
137 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, level0CommInfo.localRank);
138 :
139 0 : execMem.inputMem = execMem.inputMem.range(0, curSize * topoAttr_.userRankSize);
140 0 : execMem.outputMem = execMem.outputMem.range(0, curSize * topoAttr_.userRankSize);
141 0 : u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
142 0 : CHK_RET(RunAlgLevel1(param, reduceAttr, execMem, level1CommInfo));
143 0 : CHK_RET(RunAlgLevel0(param, reduceAttr, execMem, level0CommInfo, level1CommInfo));
144 0 : return HCCL_SUCCESS;
145 0 : }
146 :
147 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::CopyFromUserInToCclIn(const OpParam ¶m,
148 : ExecMem &execMem)
149 : {
150 0 : u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
151 0 : const bool preloadCopyOpt = IsPreloadCopyOptimizeCondition(param, execMem);
152 0 : DeviceMem dstMem;
153 0 : DeviceMem srcMem;
154 0 : if (preloadCopyOpt) {
155 : // 中转内存大小足够时，一次性搬完
156 0 : const u64 copySize = execMem.count * unitSize * topoAttr_.userRankSize;
157 0 : dstMem = execMem.inputMem.range(0, copySize);
158 0 : srcMem = DeviceMem::create(static_cast<u8 *>(execMem.inputPtr), copySize);
159 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstMem, srcMem, const_cast<Stream&>(param.stream)));
160 : } else {
161 0 : u64 copySizeOnce = execMem.count * unitSize;
162 0 : for (u32 i = 0; i < topoAttr_.userRankSize; i++) {
163 : // 拷贝input上每个slice的数据到中转内存，源端每个slice的size固定为output的size
164 0 : dstMem = execMem.inputMem.range(copySizeOnce * i, copySizeOnce);
165 0 : srcMem = DeviceMem::create(static_cast<u8 *>(execMem.inputPtr) + param.DataDes.count * unitSize * i,
166 0 : copySizeOnce);
167 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstMem, srcMem, const_cast<Stream&>(param.stream)));
168 : }
169 : }
170 0 : return HCCL_SUCCESS;
171 0 : }
172 :
173 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::RunAlgLevel1(const OpParam ¶m, u64 reduceAttr,
174 : ExecMem &execMem, SubCommInfo &level1CommInfo)
175 : {
176 0 : u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
177 :
178 0 : CHK_RET(CopyFromUserInToCclIn(param, execMem));
179 :
180 : // 第一步：节点间
181 0 : std::unique_ptr<AlgTemplateBase> level1TempAlg;
182 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_RING) {
183 : level1TempAlg
184 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_RING, dispatcher_);
185 0 : CHK_SMART_PTR_NULL(level1TempAlg);
186 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
187 0 : HCCL_INFO("ReduceScatter smallcount deterministic: using ring algo inter-server.");
188 0 : u64 ringSize = execMem.inputMem.size() / level1CommInfo.localRankSize;
189 0 : u64 ringCount = ringSize / unitSize;
190 0 : CHK_RET(level1TempAlg->Prepare(execMem.inputMem, execMem.inputMem, execMem.scratchMem, ringCount,
191 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0)));
192 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
193 : level1TempAlg
194 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_NHR, dispatcher_);
195 0 : HCCL_INFO("ReduceScatter smallcount deterministic: using nhr algo inter-server.");
196 0 : CHK_SMART_PTR_NULL(level1TempAlg);
197 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr, false));
198 0 : u64 ringSize = execMem.inputMem.size() / level1CommInfo.localRankSize;
199 0 : u64 ringCount = ringSize / unitSize;
200 0 : CHK_RET(level1TempAlg->Prepare(execMem.inputMem, execMem.inputMem, execMem.scratchMem, ringCount,
201 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0)));
202 0 : level1TempAlg->CloseBarrier();
203 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
204 : level1TempAlg
205 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_NB, dispatcher_);
206 0 : HCCL_INFO("ReduceScatter smallcount deterministic: using nonuniform-bruck algo inter-server.");
207 0 : CHK_SMART_PTR_NULL(level1TempAlg);
208 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
209 0 : u64 ringSize = execMem.inputMem.size() / level1CommInfo.localRankSize;
210 0 : u64 ringCount = ringSize / unitSize;
211 0 : CHK_RET(level1TempAlg->Prepare(execMem.inputMem, execMem.inputMem, execMem.scratchMem, ringCount,
212 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0)));
213 : } else {
214 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
215 0 : TemplateType::TEMPLATE_REDUCESCATTER_RECURSIVE_HD, dispatcher_);
216 0 : CHK_SMART_PTR_NULL(level1TempAlg);
217 0 : CHK_RET(level1TempAlg->Prepare(reduceAttr));
218 0 : HCCL_INFO("ReduceScatter smallcount deterministic: using halving-doubling algo inter-server.");
219 0 : u64 inputDataCount = execMem.inputMem.size() / unitSize; // count是output的数据个数
220 0 : CHK_RET(level1TempAlg->Prepare(execMem.inputMem, execMem.inputMem, execMem.scratchMem, inputDataCount,
221 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0)));
222 : }
223 0 : CHK_RET(level1TempAlg->RegisterProfiler(
224 : (level1CommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level1CommInfo.localRank, PROF_STAGE_0,
225 : HCCL_EXEC_STEP_NOT_SET, param.stream));
226 0 : CHK_RET(RunTemplate(level1TempAlg, level1CommInfo));
227 0 : return HCCL_SUCCESS;
228 0 : }
229 :
230 0 : HcclResult CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::RunAlgLevel0(const OpParam ¶m, u64 reduceAttr,
231 : ExecMem &execMem, SubCommInfo &level0CommInfo, SubCommInfo &level1CommInfo)
232 : {
233 0 : u32 unitSize = SIZE_TABLE[param.DataDes.dataType];
234 : // 第二步：节点内
235 : // 根据数据量算每个环上数据的偏移和大小，把做完hd的slice均分成RankSize份
236 0 : std::vector<Slice> dataSegsSlice;
237 0 : CHK_RET(PrepareReduceScatterSliceData(execMem.count, unitSize, level0CommInfo.localRankSize, dataSegsSlice));
238 :
239 : // 每个server分配的slice大小
240 0 : u64 serverSliceSize = execMem.inputMem.size() / level1CommInfo.localRankSize;
241 : // 每个服务器对应的偏移
242 0 : u64 serverSliceOffset = serverSliceSize * level1CommInfo.localRank;
243 :
244 0 : HCCL_DEBUG("inputMem.size=%llu, level0CommInfo.localRankSize=%u, serverSliceSize=%llu, serverSliceOffset=%llu "\
245 : "level0CommInfo.localRank=%u level1CommInfo.localRank=%u", execMem.inputMem.size(), level0CommInfo.localRankSize,
246 : serverSliceSize, serverSliceOffset, level0CommInfo.localRank, level1CommInfo.localRank);
247 :
248 0 : DeviceMem reduceScatterMeshInput = execMem.inputMem.range(serverSliceOffset, serverSliceSize);
249 0 : CHK_SMART_PTR_NULL(reduceScatterMeshInput);
250 0 : DeviceMem reduceScatterMeshOutput = execMem.outputMem.range(0, serverSliceSize);
251 0 : CHK_SMART_PTR_NULL(reduceScatterMeshOutput);
252 :
253 0 : HcomCollOpInfo opInfo = {"", reduceScatterMeshInput.ptr(), execMem.outputPtr, param.DataDes.count, param.DataDes.dataType,
254 0 : param.root, param.reduceType, 0};
255 0 : std::unique_ptr<AlgTemplateBase> level0TempAlg;
256 : // HD stage 模板中ranksize必须大于等于4
257 0 : if (level0CommInfo.localRankSize >= RANK_SIZE_FOUR && IsPowerOfTwo(level0CommInfo.localRankSize)) {
258 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_REDUCESCATTER_HDSTAGE,
259 0 : dispatcher_);
260 0 : CHK_SMART_PTR_NULL(level0TempAlg);
261 0 : CHK_RET(level0TempAlg->Prepare(reduceScatterMeshInput, reduceScatterMeshOutput, reduceScatterMeshOutput, execMem.count,
262 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), 0,
263 : reduceAttr, algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux,
264 : topoAttr_.userRank, &opInfo));
265 0 : HCCL_INFO("ReduceScatter smallcount deterministic: using hd stage algo inter-server.");
266 : } else {
267 0 : level0TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
268 0 : TemplateType::TEMPLATE_REDUCESCATTER_NHR, dispatcher_);
269 0 : CHK_SMART_PTR_NULL(level0TempAlg);
270 0 : CHK_RET(level0TempAlg->Prepare(reduceAttr, false));
271 0 : u64 ringSize = reduceScatterMeshInput.size() / level0CommInfo.localRankSize;
272 0 : u64 ringCount = ringSize / unitSize;
273 0 : CHK_RET(level0TempAlg->Prepare(reduceScatterMeshInput, reduceScatterMeshInput, reduceScatterMeshOutput, ringCount,
274 : param.DataDes.dataType, param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, std::vector<Slice>(0), serverSliceOffset));
275 0 : level0TempAlg->CloseBarrier();
276 0 : HCCL_INFO("ReduceScatter smallcount deterministic: using nhr algo inter-server.");
277 : }
278 0 : CHK_RET(level0TempAlg->RegisterProfiler(
279 : (level0CommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0CommInfo.localRank,
280 : PROF_STAGE_2, HCCL_EXEC_STEP_NOT_SET, param.stream));
281 0 : CHK_RET(RunTemplate(level0TempAlg, level0CommInfo));
282 0 : if (level0CommInfo.localRankSize < RANK_SIZE_FOUR || !IsPowerOfTwo(level0CommInfo.localRankSize)) {
283 0 : DeviceMem srcMem = execMem.inputMem.range(serverSliceOffset + dataSegsSlice[level0CommInfo.localRank].offset,
284 0 : execMem.count * unitSize);
285 0 : DeviceMem dstMem = DeviceMem::create(execMem.outputPtr, execMem.count * unitSize);
286 0 : CHK_SMART_PTR_NULL(srcMem);
287 0 : CHK_SMART_PTR_NULL(dstMem);
288 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstMem, srcMem, const_cast<Stream&>(param.stream)));
289 0 : }
290 0 : return HCCL_SUCCESS;
291 0 : }
292 :
293 0 : bool CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor::IsPreloadCopyOptimizeCondition(const OpParam ¶m,
294 : ExecMem &execMem)
295 : {
296 : // 通信buffer足够大时，将user in到ccl的拷贝任务合并成一个
297 0 : return param.DataDes.count == execMem.count;
298 : }
299 :
300 : REGISTER_EXEC("ReduceScatterMeshOpbaseSmallCountDeterministicExecutor",
301 : ReduceScatterMeshOpbaseSmallCountDeterministic, CollReduceScatterMeshOpbaseSmallCountDeterministicExecutor);
302 :
303 : } // namespace hccl
|
