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 : #ifndef COLL_NATIVE_EXECUTOR_BASE_H
12 : #define COLL_NATIVE_EXECUTOR_BASE_H
13 :
14 : #include "coll_executor_base.h"
15 : #include "device_capacity.h"
16 : #include "dispatcher.h"
17 : #include "stream_active_manager.h"
18 : #include "comm_factory_pub.h"
19 : #include "rank_consistentcy_checker.h"
20 : #include "hccl_aiv.h"
21 : #include "config_log.h"
22 :
23 : namespace hccl {
24 : constexpr u64 HCCL_INPLACE_MEMCOPY_SIZE = 131072; // 128K数据量 = 131072B数据量
25 : constexpr u64 HCCL_POST_SYNC_MEMCOPY_SIZE = 131072; // 128K数据量 = 131072B数据量
26 : struct ExecMem {
27 : u64 count = 0;
28 : DeviceMem inputMem; /* 单算子模式时是InCCLMem, 图模式时是InUserMem */
29 : DeviceMem outputMem; /* 单算子模式时是OutCCLMem, 图模式时是OutUserMem */
30 : DeviceMem scratchMem;
31 : void *inputPtr = nullptr; /* InUserMem的地址，图模式时与inputMem的地址相同 */
32 : void *outputPtr = nullptr; /* OutUserMem的地址，图模式时与outputMem的地址相同 */
33 : };
34 :
35 : class CollNativeExecutorBase : public CollExecutorBase {
36 : public:
37 : CollNativeExecutorBase(const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher);
38 184 : ~CollNativeExecutorBase() override = default;
39 :
40 : HcclResult CalcResRequest(const OpParam& param, AlgResourceRequest &resourceRequest) override;
41 :
42 : protected:
43 : /* *************** 资源计算 *************** */
44 : virtual void ParseParam(const OpParam& param);
45 : virtual HcclResult CalcCommInfo(std::vector<LevelNSubCommTransport>& opTransport);
46 : virtual HcclResult CalcLevel0CommInfo(TransportMemType inputType,
47 : TransportMemType outputType,
48 : std::vector<LevelNSubCommTransport>& opTransport);
49 : virtual HcclResult CalcLevel1CommInfo(TransportMemType inputType,
50 : TransportMemType outputType,
51 : std::vector<LevelNSubCommTransport>& opTransport); // 默认情况下可根据algType_支持NHR、NHRV1、NB、HD、Ring算法。
52 : virtual HcclResult CalcLevel2CommInfo(TransportMemType inputType,
53 : TransportMemType outputType,
54 : std::vector<LevelNSubCommTransport>& opTransport);
55 : virtual HcclResult CalcStreamNum(u32& streamNum);
56 : virtual HcclResult CalcScratchMemSize(u64& scratchMemSize);
57 : virtual HcclResult CalcNotifyNum(u32 streamNum, u32 ¬ifyNum);
58 : virtual HcclResult CalcAivBufferRequest(u64 &aivBufferRequest);
59 :
60 : // 考虑新建一个资源计算类ResourceCalculator，将资源推导、资源解析的都放进去。
61 : // 推导通信域信息的公用函数，不同Executor的在计算Level0、Level1、Level2时使用。
62 : HcclResult CalcCommPlaneInfo(const std::string &tag, const CommParaInfo &commParaInfo,
63 : std::vector<SingleSubCommTransport> &commTransport, TransportMemType inPutMemType,
64 : TransportMemType outPutMemType);
65 : HcclResult BuildResourceRequest(u64 scratchMemSize, u32 streamNum, u32 notifyNum, u64 aivBufferRequest,
66 : std::vector<LevelNSubCommTransport>& opTransport, AlgResourceRequest& resourceRequest);
67 : HcclResult PrintTransportRequest(AlgResourceRequest& resourceRequest);
68 : virtual HcclResult CalcOptimalIntraRing(const OpParam& param);
69 : HcclResult SetCommInfoForARS(u32 ringSize);
70 : HcclResult SetCommInfoForIntraARS(u32 intraRingsize, std::vector<u32> commPlaneVector);
71 : HcclResult SetCommInfoForInterARS(u32 intraRingsize, std::vector<u32> commPlaneVector);
72 : /* *************** 算法编排 *************** */
73 : // 非零拷贝场景走KernelRun
74 : virtual HcclResult KernelRun(const OpParam ¶m, ExecMem &execMem);
75 : // 零拷贝场景走KernelRunIntraServerPre、KernelRunInterServer、KernelRunIntraServerPost
76 0 : virtual HcclResult KernelRunInterServer(const OpParam ¶m, ExecMem &execMem) {return HCCL_SUCCESS;}
77 0 : virtual HcclResult KernelRunIntraServerPre(const OpParam ¶m, ExecMem &execMem) {return HCCL_SUCCESS;}
78 0 : virtual HcclResult KernelRunIntraServerPost(const OpParam ¶m, ExecMem &execMem) {return HCCL_SUCCESS;}
79 : virtual HcclResult Getlevel1CommRank(SubCommInfo& level1CommInfo);
80 : virtual HcclResult SelectTempAlg(std::unique_ptr<AlgTemplateBase> &level1TempAlg, u32 level1RankSize);
81 : virtual HcclResult GetDevNumInlocalPod(u32& devNumInlocalPod);
82 :
83 : // 图模式下激活从流
84 : HcclResult ActiveSlaveStreams(const Stream &stream);
85 : // 将从流添加至Profiling
86 : HcclResult AddSubStreamToProfiling();
87 : // 检查通信域大小
88 : HcclResult CheckCommSize(const CommPlane levelIndex, const u32 subLevelIndex);
89 :
90 : // 获取不同类型通信域中的 transport 信息
91 : // 为了避免循环调用时反复校验Range引发性能问题，此处不做Range校验，建议调用该接口前先调用CheckCommSize避免OutOfRange问题
92 : SubCommInfo GetSubCommInfo(const CommPlane levelIndex, const u32 subLevelIndex);
93 :
94 : HcclResult GetRankByUserRank(CommPlane levelIndex, u32 subLevelIndex, u32 userRank, u32 &rank);
95 : HcclResult GetUserRankByRank(CommPlane levelIndex, u32 subLevelIndex, u32 rank, u32 &userRank);
96 : HcclResult GenerateStreams(PrepareData &prepareData, std::vector<Stream> &streams);
97 : HcclResult NotifySubStreamStart(
98 : Stream &stream,
99 : std::vector<Stream> &substreams,
100 : std::vector<std::shared_ptr<LocalNotify>> &signalsSubToMain,
101 : u32 substreamNum);
102 : HcclResult WaitSubStreamFinish(
103 : Stream &stream,
104 : std::vector<Stream> &substreams,
105 : std::vector<std::shared_ptr<LocalNotify>> &signalsMainToSub,
106 : u32 substreamNum);
107 : HcclResult GenerateRecordWaitStreams(
108 : std::vector<Stream> &streams,
109 : u32 recordStreamNum, u32 waitStreamNum,
110 : std::vector<Stream> &recordStreams, std::vector<Stream> &waitStreams);
111 : HcclResult HoldAllRanksOnCurrentOp(const OpParam ¶m, ExecMem &execMem, PrepareData &prepareData, std::vector<LINK> links);
112 : HcclResult HoldAllRanksOnCurrentOpWithSingleStream(OpParam ¶m, ExecMem &execMem, std::vector<LINK> links);
113 : HcclResult SendRecvSignalOnLinks(OpParam ¶m, ExecMem &execMem, std::vector<LINK> links);
114 : bool OpSyncCheckCommSize(const CommPlane levelIndex, const u32 expectedSize);
115 : HcclResult PostSyncWithSubstream(OpParam ¶m, ExecMem &execMem, PrepareData &prepareData);
116 : HcclResult PostSyncWithoutSubstream(OpParam ¶m, ExecMem &execMem);
117 : HcclResult InplaceOpSync(OpParam ¶m, ExecMem &execMem);
118 :
119 : virtual HcclResult CopyAivCommInfoToDevice(const CommPlane levelIndex, const u32 subLevelIndex,
120 : AlgResourceResponse& algResource);
121 :
122 : HcclResult SetOpCache(const AivOpArgs& opArgs, const AivTopoArgs& topoArgs, const AivResourceArgs& resourceArgs,
123 : const AivAlgArgs& algArgs, ExtraArgs& extraArgs, AivProfilingInfo& aivProfilingInfo, bool isA3CrossNode);
124 :
125 : void SetWorkflowMode(HcclWorkflowMode workflowMode) override;
126 : HcclWorkflowMode GetExecutorWorkflowMode() const override;
127 :
128 : /* ---------------以下为 protected 成员变量定义领域-------------------------- */
129 : std::string tag_;
130 : u32 root_ = INVALID_VALUE_RANKID;
131 : AlgResourceResponse *algResResp_ = nullptr;
132 : HcclCMDType opType_ = HcclCMDType::HCCL_CMD_INVALID;
133 :
134 : // Infos got from topoMatcher_
135 : const HcclTopoInfo topoAttr_;
136 : const HcclAlgoInfo algoAttr_;
137 : TopoType topoType_;
138 : bool is310P3Common_ = false;
139 : bool aicpuUnfoldMode_ = false;
140 : HcclWorkflowMode workflowMode_;
141 : };
142 : std::vector<std::vector<u32>> GetARSRingsOrder(u32 ranksSize, TopoType topoType, std::vector<u32> &RingList);
143 : }
144 : #endif
|
