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_gather_mesh_opbase_pipeline_executor.h"
12 :
13 : namespace hccl {
14 0 : CollAllGatherMeshOpbasePipelineExecutor::CollAllGatherMeshOpbasePipelineExecutor(const HcclDispatcher dispatcher,
15 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
16 0 : : CollAllGatherExecutor(dispatcher, topoMatcher)
17 : {
18 0 : DMAReduceFlag_ = true;
19 0 : }
20 :
21 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::CalcStreamNum(u32& streamNum)
22 : {
23 0 : u32 totalStreamNum = topoAttr_.deviceNumPerAggregation + 1U;
24 0 : streamNum = totalStreamNum - 1U;
25 0 : HCCL_INFO("[CollAllGatherMeshOpbasePipelineExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
26 : tag_.c_str(), streamNum);
27 0 : return HCCL_SUCCESS;
28 : }
29 :
30 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::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 CollAllGatherMeshOpbasePipelineExecutor::CalcTransportMemType(TransportMemType &inputType,
41 : TransportMemType &outputType)
42 : {
43 0 : inputType = TransportMemType::CCL_INPUT;
44 0 : outputType = TransportMemType::CCL_OUTPUT;
45 0 : HCCL_INFO("[CollAllGatherMeshOpbasePipelineExecutor][CalcTransportMemType]" \
46 : "tag[%s] inputType[%d], outputType[%d]",
47 : tag_.c_str(), inputType, outputType);
48 0 : return HCCL_SUCCESS;
49 : }
50 :
51 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::CalcLevel0CommInfo(TransportMemType inputType,
52 : TransportMemType outputType,
53 : std::vector<LevelNSubCommTransport>& opTransport)
54 : {
55 0 : CommParaInfo commParaInfo(COMM_LEVEL0, CommType::COMM_TAG_MESH);
56 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL0], inputType, outputType));
57 0 : return HCCL_SUCCESS;
58 0 : }
59 :
60 : // PipeLine模式下使用Ring算法
61 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::CalcLevel1CommInfo(TransportMemType inputType,
62 : TransportMemType outputType,
63 : std::vector<LevelNSubCommTransport>& opTransport)
64 : {
65 0 : CommParaInfo commParaInfo(COMM_LEVEL1, CommType::COMM_TAG_RING_INNER);
66 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[COMM_LEVEL1], inputType, outputType));
67 0 : return HCCL_SUCCESS;
68 0 : }
69 :
70 0 : u64 CollAllGatherMeshOpbasePipelineExecutor::CalcLoopMaxCount(const u64 cclBuffSize, const u32 unitSize)
71 : {
72 0 : u64 maxCountPerLoop = (cclBuffSize - HCCL_MIN_SLICE_ALIGN_910B) / unitSize;
73 0 : return maxCountPerLoop;
74 : }
75 :
76 0 : bool CollAllGatherMeshOpbasePipelineExecutor::IsHugeData(const u64 curSize)
77 : {
78 0 : bool hugeData = curSize > RDMA_SEND_MAX_SIZE || curSize > SDMA_SEND_MAX_SIZE;
79 0 : return hugeData;
80 : }
81 :
82 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
83 : {
84 0 : HCCL_CONFIG_INFO(HCCL_ALG,
85 : "[CollAllGatherMeshOpbasePipelineExecutor][KernelRun]AllGatherMeshOpbasePipelineExecutor begins.");
86 :
87 : // step 1 先获取 comm level0 \ comm level1 的value
88 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
89 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
90 0 : u32 commIndex = level0CommInfo.localRank;
91 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
92 0 : SubCommInfo level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
93 :
94 : // DMA消减场景，打包opInfo
95 0 : HcomCollOpInfo opInfo = {
96 0 : "", execMem.inputPtr, execMem.outputPtr, param.DataDes.count, param.DataDes.dataType, 0, HCCL_REDUCE_RESERVED
97 0 : };
98 :
99 0 : std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
100 0 : TemplateType::TEMPLATE_ALL_GATHER_PIPELINE, dispatcher_);
101 0 : CHK_SMART_PTR_NULL(tempAlg);
102 0 : CHK_RET(tempAlg->Prepare(&opInfo, topoAttr_.userRank, execMem.count, execMem.inputMem, execMem.outputMem,
103 : level0CommInfo, level1CommInfo, const_cast<Stream&>(param.stream),
104 : algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux));
105 0 : CHK_RET(tempAlg->RunAsync());
106 0 : return HCCL_SUCCESS;
107 0 : }
108 :
109 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
110 : {
111 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
112 0 : SubCommInfo level0CommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
113 0 : u32 ringNum = (topoType_ == TopoType::TOPO_TYPE_8P_RING) ? LEVEL0_PLANE_NUM_IN_8PRING :
114 : LEVEL0_PLANE_NUM_IN_NPRING_SINGLE;
115 0 : u32 commIndex = (ringNum == LEVEL0_PLANE_NUM_IN_8PRING) ? topoAttr_.devicePhyId : level0CommInfo.localRank;
116 :
117 0 : CHK_RET(CheckCommSize(COMM_LEVEL1, commIndex + 1));
118 :
119 0 : level1CommInfo = GetSubCommInfo(COMM_LEVEL1, commIndex);
120 0 : HCCL_DEBUG("[CollAllGatherMeshOpbasePipelineExecutor]Getlevel1CommRank for level1CommInfo");
121 0 : return HCCL_SUCCESS;
122 0 : }
123 :
124 0 : HcclResult CollAllGatherMeshOpbasePipelineExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize)
125 : {
126 0 : if (level1RankSize > 1) {
127 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
128 0 : TemplateType::TEMPLATE_ALL_GATHER_PIPELINE, dispatcher_);
129 0 : CHK_SMART_PTR_NULL(level1TempAlg);
130 0 : return HCCL_SUCCESS;
131 : }
132 0 : return HCCL_E_UNAVAIL;
133 : }
134 : REGISTER_EXEC("AllGatherMeshOpbasePipelineExecutor", AllGatherOpbasePipeline, CollAllGatherMeshOpbasePipelineExecutor);
135 :
136 : } // namespace hccl
|
