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_comm_executor.h"
12 :
13 : namespace hccl {
14 5 : CollAllGatherCommExecutor::CollAllGatherCommExecutor(
15 5 : const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher>& topoMatcher)
16 5 : : CollAllGatherExecutor(dispatcher, topoMatcher)
17 : {
18 5 : DMAReduceFlag_ = false;
19 5 : }
20 :
21 5 : HcclResult CollAllGatherCommExecutor::CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport)
22 : {
23 5 : TransportMemType inputType = TransportMemType::RESERVED;
24 5 : TransportMemType outputType = TransportMemType::RESERVED;
25 5 : CHK_RET(CalcTransportMemType(inputType, outputType));
26 5 : CHK_RET(CalcCombinedCommInfo(inputType, outputType, opTransport));
27 5 : return HCCL_SUCCESS;
28 : }
29 :
30 5 : HcclResult CollAllGatherCommExecutor::CalcTransportMemType(TransportMemType& inputType, TransportMemType& outputType)
31 : {
32 5 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
33 0 : inputType = TransportMemType::CCL_INPUT;
34 0 : outputType = TransportMemType::CCL_OUTPUT;
35 : } else {
36 5 : inputType = TransportMemType::PARAM_INPUT;
37 5 : outputType = TransportMemType::PARAM_OUTPUT;
38 : }
39 5 : HCCL_INFO(
40 : "[CollAllGatherCommExecutor][CalcTransportMemType] tag[%s] inputType[%d], outputType[%d]", tag_.c_str(),
41 : inputType, outputType);
42 5 : return HCCL_SUCCESS;
43 : }
44 :
45 5 : HcclResult CollAllGatherCommExecutor::CalcCombinedCommInfo(
46 : TransportMemType inputType, TransportMemType outputType, std::vector<LevelNSubCommTransport>& opTransport)
47 : {
48 5 : CommPlane commPlane = COMM_COMBINE;
49 5 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93) {
50 0 : commPlane = COMM_COMBINE_ORDER;
51 : }
52 :
53 5 : CommParaInfo commParaInfo(commPlane, CommType::COMM_TAG_MAX);
54 5 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
55 0 : commParaInfo.commType = CommType::COMM_TAG_NONUNIFORM_HIERARCHICAL_RING;
56 5 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
57 0 : commParaInfo.commType = CommType::COMM_TAG_NONUNIFORM_HIERARCHICAL_RING_V1;
58 5 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
59 0 : commParaInfo.commType = CommType::COMM_TAG_NONUNIFORM_BRUCK;
60 5 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_HD) {
61 0 : commParaInfo.commType = CommType::COMM_TAG_HALVING_DOUBLING;
62 : } else {
63 5 : commParaInfo.commType = CommType::COMM_TAG_RING_INNER;
64 : }
65 5 : CHK_RET(CalcCommPlaneInfo(tag_, commParaInfo, opTransport[commPlane], inputType, outputType));
66 :
67 5 : return HCCL_SUCCESS;
68 5 : }
69 :
70 5 : HcclResult CollAllGatherCommExecutor::KernelRun(const OpParam& param, ExecMem& execMem)
71 : {
72 5 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollAllGatherCommExecutor][KernelRun]AllGather enter");
73 5 : CommPlane commPlane = COMM_COMBINE;
74 5 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93) {
75 0 : commPlane = COMM_COMBINE_ORDER;
76 : }
77 :
78 5 : CHK_RET(CheckCommSize(commPlane, COMM_INDEX_0 + 1));
79 5 : SubCommInfo combinedCommInfo = GetSubCommInfo(commPlane, COMM_INDEX_0);
80 :
81 : // 构造ring algorithm对应的all_gather实例
82 5 : std::unique_ptr<AlgTemplateBase> tempAlg;
83 5 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
84 0 : tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
85 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALL_GATHER_NHR in COMM_COMBINE_ORDER", __func__);
86 5 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
87 0 : tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
88 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALL_GATHER_NHRV1 in COMM_COMBINE_ORDER", __func__);
89 5 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
90 0 : tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
91 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALL_GATHER_NB in COMM_COMBINE_ORDER", __func__);
92 5 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_HD) {
93 0 : u8* dstPtr = static_cast<u8*>(execMem.outputMem.ptr()) + execMem.inputMem.size() * combinedCommInfo.localRank;
94 0 : DeviceMem dstMem = DeviceMem::create(dstPtr, execMem.inputMem.size());
95 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dstMem, execMem.inputMem, const_cast<Stream&>(param.stream)));
96 0 : tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
97 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
98 0 : HCCL_CONFIG_INFO(
99 : HCCL_ALG, "[%s] Run TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING in COMM_COMBINE_ORDER", __func__);
100 0 : } else {
101 5 : tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
102 5 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALL_GATHER_RING in COMM_COMBINE_ORDER", __func__);
103 : }
104 5 : CHK_SMART_PTR_NULL(tempAlg);
105 :
106 25 : CHK_RET(tempAlg->Prepare(
107 : execMem.inputMem, execMem.outputMem, execMem.outputMem, execMem.count, param.DataDes.dataType, param.stream,
108 : HCCL_REDUCE_RESERVED, INVALID_VALUE_RANKID));
109 :
110 5 : CHK_RET(RunTemplate(tempAlg, combinedCommInfo));
111 :
112 5 : return HCCL_SUCCESS;
113 5 : }
114 0 : HcclResult CollAllGatherCommExecutor::Getlevel1CommRank(SubCommInfo& level1CommInfo)
115 : {
116 0 : HCCL_INFO("[CollAllGatherCommExecutor] Entry Getlevel1CommRank.");
117 0 : CommPlane commPlane = COMM_COMBINE;
118 0 : if (topoAttr_.deviceType == DevType::DEV_TYPE_910_93) {
119 0 : commPlane = COMM_COMBINE_ORDER;
120 0 : HCCL_INFO("nslbdp AllGather comm: Getlevel1CommRank.");
121 : }
122 0 : CHK_RET(CheckCommSize(commPlane, COMM_INDEX_0 + 1));
123 0 : level1CommInfo = GetSubCommInfo(commPlane, COMM_INDEX_0);
124 :
125 0 : return HCCL_SUCCESS;
126 : }
127 :
128 0 : HcclResult CollAllGatherCommExecutor::SelectTempAlg(std::unique_ptr<AlgTemplateBase>& level1TempAlg, u32 level1RankSize)
129 : {
130 0 : HCCL_INFO("[CollAllGatherCommExecutor] Entry SelectTempAlg, algoLevel1 = [%u].", algType_.algoLevel1);
131 0 : if (level1RankSize > 1) {
132 0 : if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR) {
133 : level1TempAlg
134 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHR, dispatcher_);
135 0 : HCCL_INFO("allgather comm: using nhr algo inter-server.");
136 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NHR_V1) {
137 : level1TempAlg
138 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NHRV1, dispatcher_);
139 0 : HCCL_INFO("allgather comm: using nhr_v1 algo inter-server.");
140 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_NB) {
141 : level1TempAlg
142 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_NB, dispatcher_);
143 0 : HCCL_INFO("allgather comm: using nonuniform-bruck algo inter-server.");
144 0 : } else if (algType_.algoLevel1 == AlgTypeLevel1::ALG_LEVEL1_HD) {
145 0 : level1TempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
146 0 : TemplateType::TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING, dispatcher_);
147 0 : HCCL_INFO("allgather comm: using halving-doubling algo inter-server.");
148 : } else {
149 : level1TempAlg
150 0 : = AlgTemplateRegistry::Instance().GetAlgTemplate(TemplateType::TEMPLATE_ALL_GATHER_RING, dispatcher_);
151 0 : HCCL_INFO("allgather comm: ring algo inter-server.");
152 : }
153 0 : CHK_SMART_PTR_NULL(level1TempAlg);
154 0 : return HCCL_SUCCESS;
155 : }
156 0 : return HCCL_E_UNAVAIL;
157 : }
158 : REGISTER_EXEC("AllGatherComm", AllGatherComm, CollAllGatherCommExecutor);
159 :
160 : } // namespace hccl
|
