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