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_reduce_mesh_opbase_pipeline_executor.h"
12 :
13 : namespace hccl {
14 : // 准入条件: pipeLine && 910B && 单算子 && sdmaReduce && rdmaReduce && 多Mesh && MeshTopo && 非确定性
15 0 : CollAllReduceMeshOpbasePipelineExecutor::CollAllReduceMeshOpbasePipelineExecutor(const HcclDispatcher dispatcher,
16 0 : std::unique_ptr<TopoMatcher> &topoMatcher): CollAllReduceExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = true;
19 0 : }
20 :
21 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::CalcStreamNum(u32& streamNum)
22 : {
23 0 : u32 totalStreamNum = topoAttr_.deviceNumPerAggregation + 1U;
24 0 : streamNum = totalStreamNum - 1U;
25 0 : HCCL_INFO("[CollAllReduceMeshOpbasePipelineExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
26 : tag_.c_str(), streamNum);
27 0 : return HCCL_SUCCESS;
28 : }
29 :
30 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
31 : {
32 0 : TransportMemType inputType = TransportMemType::RESERVED;
33 0 : TransportMemType outputType = TransportMemType::RESERVED;
34 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
35 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
36 0 : CHK_RET(CalcLevel1CommInfo(inputType, outputType, opTransport));
37 0 : return HCCL_SUCCESS;
38 : }
39 :
40 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::CalcTransportMemType(TransportMemType &inputType,
41 : TransportMemType &outputType)
42 : {
43 0 : inputType = TransportMemType::CCL_INPUT;
44 0 : outputType = TransportMemType::CCL_OUTPUT;
45 0 : HCCL_INFO("[CollAllReduceMeshOpbasePipelineExecutor][CalcTransportMemType]" \
46 : "tag[%s] inputType[%d], outputType[%d]",
47 : tag_.c_str(), inputType, outputType);
48 0 : return HCCL_SUCCESS;
49 : }
50 :
51 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::CalcLevel0CommInfo(TransportMemType inputType,
52 : TransportMemType outputType,
53 : std::vector<LevelNSubCommTransport>& opTransport)
54 : {
55 0 : CommParaInfo commParaInfo(COMM_LEVEL0, CommType::COMM_TAG_MESH);
56 0 : commParaInfo.meshSinglePlane = true;
57 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL0], inputType, outputType));
58 0 : return HCCL_SUCCESS;
59 0 : }
60 :
61 : // PipeLine模式下使用Ring算法
62 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::CalcLevel1CommInfo(TransportMemType inputType,
63 : TransportMemType outputType,
64 : std::vector<LevelNSubCommTransport>& opTransport)
65 : {
66 0 : CommParaInfo commParaInfo(COMM_LEVEL1, CommType::COMM_TAG_RING_INNER);
67 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL1], inputType, outputType));
68 0 : return HCCL_SUCCESS;
69 0 : }
70 :
71 0 : u64 CollAllReduceMeshOpbasePipelineExecutor::CalcLoopMaxCount(const u64 cclBuffSize, const u32 unitSize)
72 : {
73 0 : if (cclBuffSize <= HCCL_MIN_SLICE_ALIGN_910B) {
74 0 : return 0;
75 : }
76 0 : u64 maxCountPerLoop = (cclBuffSize - HCCL_MIN_SLICE_ALIGN_910B) /
77 0 : unitSize * topoAttr_.userRankSize;
78 0 : return maxCountPerLoop;
79 : }
80 :
81 0 : bool CollAllReduceMeshOpbasePipelineExecutor::IsHugeData(const u64 curSize)
82 : {
83 0 : bool hugeData = curSize / topoAttr_.deviceNumPerAggregation / HCCL_INTERNODE_MAX_DATA_RATE > RDMA_SEND_MAX_SIZE ||
84 : curSize > SDMA_SEND_MAX_SIZE;
85 0 : return hugeData;
86 : }
87 :
88 0 : bool CollAllReduceMeshOpbasePipelineExecutor::IsSmallData(const u64 totalSize, const u64 curSize)
89 : {
90 0 : return false;
91 : }
92 :
93 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
94 : {
95 0 : HCCL_CONFIG_INFO(HCCL_ALG,
96 : "[CollAllReduceMeshOpbasePipelineExecutor][Run]CollAllReduceMeshOpbasePipelineExecutor begins.");
97 :
98 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
99 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
100 0 : u32 commIndex = level0CommInfo.localRank;
101 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
102 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
103 :
104 0 : u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, param.DataDes.dataType, param.reduceType);
105 :
106 0 : HcomCollOpInfo opInfo = {
107 0 : "", execMem.inputPtr, execMem.outputPtr, execMem.count, param.DataDes.dataType, param.root, param.reduceType
108 0 : };
109 :
110 0 : std::unique_ptr<AlgTemplateBase> tempAlg;
111 0 : tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_OPBASE_PIPELINE, dispatcher_);
112 0 : CHK_SMART_PTR_NULL(tempAlg);
113 0 : CHK_RET(tempAlg->Prepare(reduceAttr));
114 0 : CHK_RET(tempAlg->Prepare(&opInfo, execMem.inputMem, execMem.outputMem, execMem.count,
115 : level1CommInfo, level0CommInfo, const_cast<Stream&>(param.stream),
116 : algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux));
117 0 : CHK_RET(tempAlg->RunAsync());
118 0 : return HCCL_SUCCESS;
119 0 : }
120 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
121 : {
122 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
123 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
124 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
125 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
126 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
127 :
128 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
129 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
130 :
131 0 : return HCCL_SUCCESS;
132 0 : }
133 :
134 0 : HcclResult CollAllReduceMeshOpbasePipelineExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
135 : {
136 0 : if (level1RankSize > 1) {
137 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_REDUCE_OPBASE_PIPELINE, dispatcher_);
138 0 : CHK_SMART_PTR_NULL(level1TempAlg);
139 0 : return HCCL_SUCCESS;
140 : }
141 0 : return HCCL_E_UNAVAIL;
142 : }
143 : REGISTER_EXEC("AllReduceMeshOpbasePipelineExecutor",
144 : AllReduceMeshOpbasePipeline, CollAllReduceMeshOpbasePipelineExecutor);
145 :
146 : } // namespace hccl
|
