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_all_to_all_mesh_aiv_executor.h"
12 :
13 : namespace hccl {
14 :
15 5 : CollAlltoAllMeshAivExecutor::CollAlltoAllMeshAivExecutor(
16 5 : const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher>& topoMatcher)
17 5 : : CollAlltoAllExecutor(dispatcher, topoMatcher)
18 : {
19 5 : desc_.isAivMode = true;
20 5 : }
21 :
22 5 : HcclResult CollAlltoAllMeshAivExecutor::CalcStreamNum(u32& streamNum)
23 : {
24 5 : streamNum = 0; // AIV通信不需要申请从流
25 5 : HCCL_INFO("[CollAlltoAllMeshAivExecutor][CalcStreamNum] tag[%s] streamNum[%u]", tag_.c_str(), streamNum);
26 5 : return HCCL_SUCCESS;
27 : }
28 :
29 5 : HcclResult CollAlltoAllMeshAivExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
30 : {
31 5 : TransportMemType inputType = TransportMemType::RESERVED;
32 5 : TransportMemType outputType = TransportMemType::RESERVED;
33 5 : CHK_RET(CalcTransportMemType(inputType, outputType));
34 5 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
35 5 : return HCCL_SUCCESS;
36 : }
37 :
38 5 : HcclResult CollAlltoAllMeshAivExecutor::CalcTransportMemType(TransportMemType& inputType, TransportMemType& outputType)
39 : {
40 5 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
41 5 : inputType = TransportMemType::CCL_INPUT;
42 5 : outputType = TransportMemType::AIV_OUTPUT;
43 : } else {
44 0 : inputType = TransportMemType::PARAM_INPUT;
45 0 : outputType = TransportMemType::AIV_OUTPUT;
46 : }
47 5 : HCCL_INFO(
48 : "[CollAlltoAllMeshAivExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]", tag_.c_str(),
49 : inputType, outputType);
50 5 : return HCCL_SUCCESS;
51 : }
52 :
53 5 : HcclResult CollAlltoAllMeshAivExecutor::CalcLevel0CommInfo(
54 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
55 : {
56 5 : CommParaInfo commParaLevel0(COMM_MESH_L0, CommType::COMM_TAG_MESH);
57 5 : commParaLevel0.meshSinglePlane = true;
58 5 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_MESH_L0], inputType, outputType));
59 5 : return HCCL_SUCCESS;
60 5 : }
61 :
62 5 : HcclResult CollAlltoAllMeshAivExecutor::CalNumBlocks(u32& numBlocks, u32 rankSize, u64 dataSize, HcclCMDType cmdType)
63 : {
64 5 : numBlocks = rankSize; // 默认情况使用rankSize个AIV
65 :
66 5 : bool isOpBase = (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
67 5 : if (cmdType == HcclCMDType::HCCL_CMD_ALLTOALL) {
68 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93 && !isOpBase) {
69 0 : numBlocks = rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3 > MAX_NUM_BLOCKS ?
70 : rankSize * NUM_BLOCKS_THREE_PER_RANK_A3 :
71 : rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3;
72 0 : } else if (isOpBase && dataSize >= AIV_ALL_TO_ALL_BIG_SIZE) {
73 0 : numBlocks = NUM_BLOCKS_FACTOR_TWO * rankSize; // 单机场景，单算子AlltoAll使用2倍 rankSize个aiv
74 : }
75 5 : } else if (cmdType == HcclCMDType::HCCL_CMD_ALLTOALLVC || cmdType == HcclCMDType::HCCL_CMD_ALLTOALLV) {
76 5 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93
77 0 : && ((isOpBase && cmdType == HcclCMDType::HCCL_CMD_ALLTOALLV)
78 0 : || (!isOpBase && cmdType == HcclCMDType::HCCL_CMD_ALLTOALLVC))) {
79 : // A3单机单算子场景，numBlocks_为3倍或者4倍的ranksize
80 0 : numBlocks = rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3 > MAX_NUM_BLOCKS ?
81 : rankSize * NUM_BLOCKS_THREE_PER_RANK_A3 :
82 : rankSize * NUM_BLOCKS_FOUR_PER_RANK_A3;
83 5 : } else if (isOpBase) {
84 5 : numBlocks = NUM_BLOCKS_FACTOR_TWO * rankSize; // 单机场景，单算子AlltoAll使用2倍 rankSize个aiv
85 : }
86 : }
87 :
88 5 : u32 bestNumBlocks = numBlocks;
89 5 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910B
90 0 : || (isOpBase && topoAttr_.deviceType == DevType::DEV_TYPE_910_93 && cmdType != HcclCMDType::HCCL_CMD_ALLTOALLV)
91 0 : || (!isOpBase && topoAttr_.deviceType == DevType::DEV_TYPE_910_93
92 0 : && cmdType == HcclCMDType::HCCL_CMD_ALLTOALLV)) {
93 5 : CHK_PRT_RET(
94 : numBlocks_ < numBlocks,
95 : HCCL_WARNING(
96 : "[CollAlltoAllMeshAivExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].", numBlocks_,
97 : numBlocks),
98 : HCCL_E_PARA);
99 0 : } else if (
100 0 : (isOpBase && topoAttr_.deviceType == DevType::DEV_TYPE_910_93 && cmdType == HcclCMDType::HCCL_CMD_ALLTOALLV)
101 0 : || (!isOpBase && topoAttr_.deviceType == DevType::DEV_TYPE_910_93
102 0 : && cmdType != HcclCMDType::HCCL_CMD_ALLTOALLV)) {
103 0 : CHK_PRT_RET(
104 : numBlocks_ < rankSize,
105 : HCCL_WARNING(
106 : "[CollAlltoAllMeshAivExecutor][CalNumBlocks]aivCore[%u] is invalid, at least need [%u].", numBlocks_,
107 : rankSize),
108 : HCCL_E_PARA);
109 0 : if (numBlocks_ < numBlocks) {
110 0 : numBlocks = numBlocks_ / rankSize * rankSize;
111 : }
112 : }
113 :
114 5 : HCCL_INFO(
115 : "[CollAlltoAllMeshAivExecutor][CalNumBlocks] numBlocks is set to [%u], limit[%u], recommanded[%u]", numBlocks,
116 : numBlocks_, bestNumBlocks);
117 5 : return HCCL_SUCCESS;
118 : }
119 :
120 5 : HcclResult CollAlltoAllMeshAivExecutor::Orchestrate(OpParam& param, AlgResourceResponse& algRes)
121 : {
122 5 : HcclUs startut = TIME_NOW();
123 5 : tag_ = param.tag;
124 5 : algResResp_ = &algRes;
125 :
126 5 : HcclResult ret = HCCL_SUCCESS;
127 5 : ExecMem execMem;
128 5 : execMem.count = param.DataDes.count;
129 5 : execMem.inputPtr = param.inputPtr;
130 5 : execMem.outputPtr = param.outputPtr;
131 :
132 5 : if (workflowMode_ != HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
133 0 : execMem.inputMem = algRes.paramInputMem;
134 0 : execMem.outputMem = algRes.aivOutputMem; // 存放flag
135 0 : ret = KernelRun(param, execMem);
136 : } else {
137 5 : execMem.inputMem = algRes.cclInputMem;
138 5 : execMem.outputMem = algRes.aivOutputMem;
139 5 : ret = KernelRun(param, execMem);
140 : }
141 :
142 5 : CHK_PRT_RET(
143 : ret != HCCL_SUCCESS,
144 : HCCL_ERROR(
145 : "[CollAlltoAllMeshAivExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel run failed",
146 : HCCL_ERROR_CODE(ret), param.tag.c_str()),
147 : ret);
148 :
149 5 : HCCL_INFO(
150 : "tag[%s], AlltoAll executor orchestrate success, take time [%lld]us.", param.tag.c_str(),
151 : DURATION_US(TIME_NOW() - startut));
152 5 : return HCCL_SUCCESS;
153 5 : }
154 :
155 0 : HcclResult CollAlltoAllMeshAivExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& adjInfo)
156 : {
157 0 : return HCCL_SUCCESS;
158 : }
159 :
160 0 : HcclResult CollAlltoAllMeshAivExecutor::GetAivExecParam(
161 : const OpParam& param, AlgResourceResponse& algRes, AivSuperKernelArgs& args)
162 : {
163 0 : HcclUs startut = TIME_NOW();
164 0 : tag_ = param.tag;
165 0 : algResResp_ = &algRes;
166 :
167 0 : HcclResult ret = HCCL_SUCCESS;
168 0 : ExecMem execMem;
169 0 : execMem.count = param.All2AllDataDes.sendCount;
170 0 : execMem.inputPtr = param.inputPtr;
171 0 : execMem.outputPtr = param.outputPtr;
172 :
173 : // 单算子大数据量
174 0 : execMem.inputMem = algRes.paramInputMem;
175 0 : execMem.outputMem = algRes.aivOutputMem;
176 :
177 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_MESH_L0, COMM_INDEX_0);
178 :
179 0 : u32 localRank = level0CommInfo.localRank;
180 0 : u32 localRankSize = level0CommInfo.localRankSize;
181 0 : HCCL_DEBUG(
182 : "[CollAlltoAllMeshAivExecutor][GetAivExecParam] userRank [%d] localRank [%d]", topoAttr_.userRank, localRank);
183 :
184 0 : for (u32 i = 0; i < localRankSize; i++) {
185 0 : if (i != localRank) {
186 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(args.buffersIn[i])));
187 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(args.buffersOut[i])));
188 : } else {
189 0 : args.buffersIn[i] = execMem.inputMem.ptr();
190 0 : args.buffersOut[i] = execMem.outputMem.ptr();
191 : }
192 : }
193 0 : args.rank = localRank;
194 0 : args.rankSize = localRankSize;
195 0 : args.len = execMem.count;
196 0 : args.dataType = param.All2AllDataDes.sendType;
197 0 : args.unitSize = SIZE_TABLE[param.All2AllDataDes.sendType];
198 0 : args.devType = static_cast<u32>(topoAttr_.deviceType);
199 0 : CHK_PRT_RET(
200 : ret != HCCL_SUCCESS,
201 : HCCL_ERROR(
202 : "[CollAlltoAllMeshAivExecutor][Orchestrate]errNo[0x%016llx] tag[%s] executor kernel "
203 : "run failed",
204 : HCCL_ERROR_CODE(ret), param.tag.c_str()),
205 : ret);
206 :
207 0 : HCCL_INFO(
208 : "tag[%s], AlltoAll executor getalgexecparam success, take time [%lld]us.", param.tag.c_str(),
209 : DURATION_US(TIME_NOW() - startut));
210 0 : return HCCL_SUCCESS;
211 0 : }
212 :
213 5 : HcclResult CollAlltoAllMeshAivExecutor::KernelRun(const OpParam& param, ExecMem& execMem)
214 : {
215 5 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollAlltoAllMeshAivExecutor][KernelRun]AllToAll aiv enter");
216 :
217 5 : CHK_RET(CheckCommSize(COMM_MESH_L0, COMM_INDEX_0 + 1));
218 5 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_MESH_L0, COMM_INDEX_0);
219 :
220 : void* buffersIn[MAX_RANK_SIZE];
221 : void* buffersOut[MAX_RANK_SIZE];
222 :
223 5 : u32 localRank = level0CommInfo.localRank;
224 5 : u32 localRankSize = level0CommInfo.localRankSize;
225 5 : HCCL_DEBUG("[CollAlltoAllMeshAivExecutor][KernelRun] userRank [%u] localRank [%u]", topoAttr_.userRank, localRank);
226 :
227 10 : for (u32 i = 0; i < localRankSize; i++) {
228 5 : if (i != localRank) {
229 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::INPUT_MEM, &(buffersIn[i])));
230 0 : CHK_RET(level0CommInfo.links[i]->GetRemoteMem(UserMemType::OUTPUT_MEM, &(buffersOut[i])));
231 : } else {
232 5 : buffersIn[i] = execMem.inputMem.ptr();
233 5 : buffersOut[i] = execMem.outputMem.ptr();
234 : }
235 : }
236 :
237 5 : ExtraArgs extraArgs;
238 5 : bool isOpbase = (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
239 : HcclResult ret;
240 5 : u64 dataSize
241 5 : = (param.opType == HcclCMDType::HCCL_CMD_ALLTOALL ?
242 0 : param.All2AllDataDes.sendCount * SIZE_TABLE[param.All2AllDataDes.sendType] :
243 : 0);
244 :
245 5 : AivOpArgs opArgs{
246 : HcclCMDType::HCCL_CMD_ALLTOALL,
247 5 : execMem.inputPtr,
248 5 : execMem.outputPtr,
249 : 0,
250 5 : param.All2AllDataDes.sendType,
251 : HCCL_REDUCE_RESERVED,
252 : 0,
253 5 : isOpbase};
254 10 : AivTopoArgs topoArgs{localRank, localRankSize, MAX_RANK_SIZE, 0, topoAttr_.serverNum, topoAttr_.deviceType};
255 5 : topoArgs.identify = algoAttr_.identifier;
256 15 : AivResourceArgs resourceArgs{param.tag, param.stream.ptr(), buffersIn, buffersOut, execMem.inputMem.size(),
257 5 : 0, param.aivTag};
258 5 : AivAlgArgs algArgs{};
259 5 : algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
260 5 : algArgs.execTimeOutSet = true;
261 5 : struct AivProfilingInfo aivProfilingInfo;
262 5 : aivProfilingInfo.counter = opCounter_;
263 :
264 5 : if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALL
265 0 : && ((isOpbase && dataSize < AIV_ALL_TO_ALL_BIG_SIZE)
266 0 : || (!isOpbase && topoAttr_.deviceType == DevType::DEV_TYPE_910_93))) {
267 0 : opArgs.count = param.All2AllDataDes.sendCount;
268 : u32 numBlocks;
269 0 : CHK_PRT_RET(
270 : CalNumBlocks(numBlocks, localRankSize, dataSize, opArgs.cmdType) != HCCL_SUCCESS,
271 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__), HCCL_E_PARA);
272 0 : numBlocks_ = numBlocks;
273 0 : resourceArgs.numBlocks = numBlocks_;
274 0 : ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
275 5 : } else if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALLVC || param.opType == HcclCMDType::HCCL_CMD_ALLTOALL) {
276 2 : for (u32 i = 0; i < localRankSize; i++) {
277 1 : u64 rankCount = 0;
278 2 : for (u32 j = 0; j < localRankSize; j++) {
279 1 : u64 curSendCount
280 1 : = *(static_cast<const u64*>(param.All2AllDataDes.sendCountMatrix) + i * localRankSize + j);
281 1 : extraArgs.sendCountMatrix[i * localRankSize + j] = curSendCount;
282 1 : rankCount += curSendCount;
283 : }
284 1 : if (rankCount > extraArgs.maxCount) {
285 1 : extraArgs.maxCount = rankCount;
286 : }
287 : }
288 1 : opArgs.count = extraArgs.maxCount;
289 1 : opArgs.cmdType = HcclCMDType::HCCL_CMD_ALLTOALLVC;
290 : u32 numBlocks;
291 1 : CHK_PRT_RET(
292 : CalNumBlocks(numBlocks, localRankSize, dataSize, opArgs.cmdType) != HCCL_SUCCESS,
293 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__), HCCL_E_PARA);
294 1 : numBlocks_ = numBlocks;
295 1 : resourceArgs.numBlocks = numBlocks_;
296 1 : ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
297 1 : } else {
298 8 : for (u32 i = 0; i < localRankSize; i++) {
299 4 : extraArgs.sendCounts[i] = *(static_cast<const u64*>(param.All2AllDataDes.sendCounts) + i);
300 4 : extraArgs.sendDispls[i] = *(static_cast<const u64*>(param.All2AllDataDes.sdispls) + i);
301 4 : extraArgs.recvCounts[i] = *(static_cast<const u64*>(param.All2AllDataDes.recvCounts) + i);
302 4 : extraArgs.recvDispls[i] = *(static_cast<const u64*>(param.All2AllDataDes.rdispls) + i);
303 : }
304 :
305 4 : opArgs.cmdType = HcclCMDType::HCCL_CMD_ALLTOALLV;
306 : u32 numBlocks;
307 4 : CHK_PRT_RET(
308 : CalNumBlocks(numBlocks, localRankSize, dataSize, opArgs.cmdType) != HCCL_SUCCESS,
309 : HCCL_ERROR("[%s] CalNumBlocks failed", __func__), HCCL_E_PARA);
310 4 : numBlocks_ = numBlocks;
311 4 : resourceArgs.numBlocks = numBlocks_;
312 4 : ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
313 : }
314 5 : CHK_PRT_RET(
315 : ret != HCCL_SUCCESS, HCCL_ERROR("[CollAlltoAllMeshAivExecutor][KernelRun]AllToAll aiv failed, return[%d]", ret),
316 : ret);
317 :
318 5 : if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALL && isOpbase) {
319 0 : CHK_RET(SetOpCache(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo, false));
320 : }
321 :
322 5 : HCCL_INFO("[CollAlltoAllMeshAivExecutor][KernelRun]AllToAll aiv run success");
323 5 : return HCCL_SUCCESS;
324 5 : }
325 :
326 : REGISTER_EXEC("AlltoAllMeshAivExecutor", AlltoAllMeshAiv, CollAlltoAllMeshAivExecutor);
327 :
328 : } // namespace hccl
|
