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_COMMON_EXECUTOR_H
12 : #define COLL_COMMON_EXECUTOR_H
13 :
14 : #include "coll_native_executor_base.h"
15 : #include "coll_alg_exec_registry.h"
16 : #include "profiler_base_pub.h"
17 : #include "send_receive_pub.h"
18 : #include "alg_template_register.h"
19 : #include "alltoallv_staged_calculator_pub.h"
20 :
21 : namespace hccl {
22 : constexpr u32 NSLBDP_MIN_COUNT = 128;
23 : class CollCommExecutor : public CollNativeExecutorBase {
24 : public:
25 : CollCommExecutor(const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher);
26 179 : ~CollCommExecutor() override = default;
27 :
28 : // CCL Op Share
29 : HcclResult MultiRingAllReduce(const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
30 : const u64 count, const HcclDataType dataType,
31 : const HcclReduceOp reductionOp,
32 : const std::vector<std::vector<Slice>> &multRingsSliceZero, Stream stream,
33 : s32 profStage, const u64 baseOffset = 0);
34 : HcclResult CollectMultiRingsUserMemSlices(u32 ringNum, const HcclDataType dataType,
35 : const HcomCollOpInfo *opInfo, const std::vector<std::vector<Slice>> &multRingsSliceZero,
36 : const std::vector<std::vector<u32>> &multiRingsOrder,
37 : const std::vector<std::vector<Slice>> &multRingsUserMemSlice,
38 : std::vector<std::vector<Slice>> &userMemSlicesOfMultiRings);
39 : HcclResult CollectMultiRingsRankOrder(u32 ringNum,
40 : const std::vector<std::vector<u32>> &multiRingsOrder,
41 : std::vector<std::vector<u32>> &rankOrders);
42 : u32 CalcOptimalIntraRingsize(u64 count, HcclDataType dataType, HcclCMDType opType);
43 :
44 : HcclResult MultiRingReduceScatter(const std::string &tag, DeviceMem inputMem, DeviceMem outputMem, const u64 count,
45 : const HcclDataType dataType, const HcclReduceOp reductionOp,
46 : const std::vector<std::vector<Slice>> multRingsSliceZero, Stream stream,
47 : s32 profStage, const u64 baseOffset = 0, const HcomCollOpInfo *opInfo = nullptr,
48 : const std::vector<std::vector<Slice>> multRingsUserMemSlice = std::vector<std::vector<Slice>> (0),
49 : const CommPlane levelIndex = COMM_LEVEL0);
50 :
51 : HcclResult MultiRingReduceScatterConcurrent(const std::string &tag, DeviceMem inputMem,DeviceMem outputMem,
52 : const u64 count, const HcclDataType dataType, const HcclReduceOp reductionOp,
53 : const std::vector<std::pair<bool, std::vector<Slice>>> multRingsSliceZero, Stream stream,
54 : s32 profStage, const u64 baseOffset = 0, const HcomCollOpInfo *opInfo = nullptr,
55 : const std::vector<std::pair<bool, std::vector<Slice>>> multRingsUserMemSlice =
56 : std::vector<std::pair<bool, std::vector<Slice>>> (0));
57 :
58 : HcclResult Level1ReduceScatterConcurrent(DeviceMem inputMem, DeviceMem scratchMem,const u64 count,
59 : const HcclDataType dataType, const HcclReduceOp reductionOp, Stream stream, s32 profStage,
60 : std::vector<Slice> &level1DataSegsSlice, u32 syncTrans, u64 reduceAttr);
61 :
62 : HcclResult UpdateOffsetBasedOnStrideCount(const OpParam ¶m,
63 : std::vector<std::vector<Slice>> &multRingsUserMemSlice) const;
64 :
65 : HcclResult MultiRingAllGather(const std::string &tag, DeviceMem inputMem, DeviceMem outputMem, const u64 count,
66 : const HcclDataType dataType,
67 : const std::vector<std::vector<Slice> > multRingsSliceZero, Stream stream,
68 : s32 profStage, const u64 baseOffset = 0, const HcomCollOpInfo *opInfo = nullptr,
69 : const std::vector<std::vector<Slice>> multRingsUserMemSlice = std::vector<std::vector<Slice>> (0),
70 : const CommPlane leveIndex = COMM_LEVEL0);
71 :
72 : HcclResult MultiRingAllGatherConcurrent(const std::string &tag, DeviceMem inputMem, DeviceMem outputMem,
73 : const u64 count, const HcclDataType dataType,
74 : const std::vector<std::pair<bool, std::vector<Slice>>> multRingsSliceZero, Stream stream,
75 : s32 profStage, const u64 baseOffset = 0, const HcomCollOpInfo *opInfo = nullptr,
76 : const std::vector<std::pair<bool, std::vector<Slice>>> multRingsUserMemSlice =
77 : std::vector<std::pair<bool, std::vector<Slice>>> (0));
78 :
79 : HcclResult Level1AllGatherConcurrent(DeviceMem inputMem, DeviceMem outputMem, const u64 count,
80 : const HcclDataType dataType, Stream stream, s32 profStage,
81 : std::vector<Slice> &level1DataSegsSlice, u32 syncTrans);
82 :
83 : HcclResult MultiRingMultiRootScatter(const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
84 : const u64 count, const HcclDataType dataType, const std::vector<std::vector<Slice>> &multRingsSliceZero,
85 : u32 root, Stream stream, const u64 baseOffset);
86 :
87 : HcclResult MultiStreamReduceScatterMesh(const std::string &tag, DeviceMem inputMem, DeviceMem outputMem,
88 : const u64 count, const HcclDataType dataType,
89 : const HcclReduceOp reductionOp,
90 : const std::vector<std::vector<Slice>>& multStreamsSlice,
91 : Stream stream,
92 : const CommPlane commLevelIndex,
93 : const u64 baseOffset = 0);
94 :
95 : HcclResult MultiRingGather(const std::string &tag, DeviceMem inputMem, DeviceMem outputMem, const u64 count,
96 : const HcclDataType dataType, const std::vector<std::vector<Slice>> multRingsSliceZero,
97 : HcclReduceOp op, u32 root, Stream stream, s32 profStage);
98 :
99 : HcclResult MultiStreamReduceScatterMeshAtomic(const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
100 : const u64 count, const HcclDataType dataType,
101 : const HcclReduceOp reductionOp,
102 : const std::vector<Slice> &dataSliceVct,
103 : Stream &stream,
104 : const CommPlane commLevelIndex,
105 : const u64 baseOffset = 0, HcomCollOpInfo *opInfo = nullptr);
106 : HcclResult PrepareReduceScatterSliceData(u64 dataCount, u32 unitSize, u32 sliceNum, std::vector<Slice> &dataSlice);
107 :
108 : HcclResult MultiRingScatter(const std::string &tag, DeviceMem inputMem, DeviceMem outputMem, const u64 count,
109 : const HcclDataType dataType, const std::vector<std::vector<Slice> > multRingsSliceZero,
110 : u32 root, Stream stream, const HcomCollOpInfo *opInfo, const u64 baseOffset = 0);
111 : std::vector<std::vector<u32>> GetRingsOrderByTopoType(u32 ranksSize, TopoType topoType, std::vector<u32> &nicList);
112 : HcclResult MutliSegSlicePrepare(const std::vector<Slice> &dataSegsSlice,
113 : std::vector<std::vector<Slice> >& mutliSegsSlices, u32 ringCount);
114 : HcclResult MutliSegSlicePrepareAvoidCceRewrite(const std::vector<Slice> &dataSegsSlice,
115 : std::vector<std::vector<Slice> >& mutliSegsSlices, u32 ringCount) const;
116 : void NicSendSizeCal(const std::vector<std::vector<Slice>> &mutliSegsSlices, u32 ringCount, u32 chunkSize,
117 : const std::vector<u32> &nicList, const std::string &tag);
118 : std::vector<std::vector<Slice> > PrepareMultiRingSlice(const std::vector<Slice> &dataSegsSlice,
119 : const std::string &tag, bool avoidCceRewrite = false, std::vector<u32> nicList = {0, 1, 2, 3, 4, 5, 6, 7}, CommPlane commLevelIndex = COMM_LEVEL0);
120 : // AnyPath特性使用
121 : std::vector<std::vector<u32>> GetRingsOrderForAnyPath(u32 ranksSize, TopoType topoType, std::vector<u32> &nicList);
122 : std::vector<std::vector<Slice> > AnyPathPrepareMultiRingSlice(const std::vector<Slice> &dataSegsSlice,
123 : const std::string &tag, bool avoidCceRewrite = false, std::vector<u32> nicList = {0, 1, 2, 3, 4, 5, 6, 7});
124 :
125 : bool Is2U2PInfer();
126 : bool Is910BSingleMesh();
127 : bool NeedCreateSingleMeshPlane(const bool isInlineReduce);
128 : bool SingleMeshInlineReduce(void *inputPtr, void *outputPtr, HcclDataType dataType, HcclReduceOp op);
129 : bool IsMultiMeshInlineReduce(void *inputPtr, void *outputPtr, HcclDataType dataType, HcclReduceOp op);
130 :
131 : u64 GetReduceAttr(DeviceMem &inputMem, DeviceMem &outputMem, HcclDataType dataType, HcclReduceOp op);
132 : HcclResult PrepareLevel1CommInfo(u32 &segmentIdx, u32 &commIndex, u64 &hdSize,
133 : const SubCommInfo &commInfo,
134 : const std::vector<std::vector<Slice> > &multRingsSliceZero,
135 : const std::string &tag);
136 : HcclResult GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& adjInfo) override;
137 :
138 : protected:
139 : virtual HcclResult GetSubStreamInfoOnOneRing(const u32 ringIndex,
140 : std::vector<Stream> &subStreamsInOneRing,
141 : std::vector<std::shared_ptr<LocalNotify>> &mainSignalsInOneRing,
142 : std::vector<std::shared_ptr<LocalNotify>> &subSignalsInOneRing);
143 : virtual u32 GetLevel0RingNum() const;
144 : HcclResult CalUserMemSlices(const HcclDataType dataType, const HcomCollOpInfo *opInfo,
145 : const std::vector<Slice> &singleRingSliceZero, u32 ringIndex,
146 : const std::vector<std::vector<u32>> &multiRingsOrder,
147 : std::vector<Slice> &userMemSlices);
148 : HcclResult GetRankOrder(const std::vector<std::vector<u32>> &multiRingsOrder, u32 ringIndex,
149 : std::vector<u32> &rankOrder);
150 : HcclResult SetRingNics(const std::string &tag, const std::vector<std::vector<u32>> &ringNics);
151 : HcclResult GetRingNics(const std::string &tag, std::vector<std::vector<u32>> &ringNics);
152 : HcclResult SetNicSendSize(const std::string &tag, std::vector<u64> &sizeList);
153 :
154 : // 用于ZerocopyExecutor
155 : HcclResult CalcIntraServerDataSlicesDiscontinuous(const OpParam ¶m, const ExecMem &execMem,
156 : u32 level0RankSize, u32 level1RankSize, u32 level2RankSize, std::vector<Slice> &dataSegsSlice);
157 : HcclResult CalcIntraServerDataSlicesContinuous(const OpParam ¶m, const ExecMem &execMem,
158 : u32 level0RankSize, u32 level1RankSize, u32 level2RankSize, std::vector<Slice> &dataSegsSlice);
159 : void CalcLevel1DataSlices(u64 sliceSize, u32 level1RankSize, u32 level2RankSize, std::vector<Slice> &level1DataSegsSlice);
160 : HcclResult GetCommRankInfoNormal(u32 &level0Rank, u32 &level0RankSize,
161 : u32 &level1Rank, u32 &level1RankSize, u32 &level2Rank, u32 &level2RankSize, bool isAHCAlgo = false);
162 :
163 : // 用于ExchangeExecutor
164 : HcclResult CalExchangeRemoteRankForReduceScatter(u32 &remoteRankSend, u32 &remoteRankRecv);
165 : HcclResult GetTransportForExchange(u32 remoteUserRank, LINK &targetLink);
166 : bool IsLevel0Neighbor(u32 remoteRank, u32 level0RankSize);
167 :
168 : std::mutex ringNicListLock_;
169 : std::map<std::string, std::vector<std::vector<u32>>> ringNicList_;
170 : std::mutex nicSendSizeListLock_;
171 : std::map<std::string, std::vector<u64>> nicSendSizeList_;
172 : };
173 : } // namespace hccl
174 :
175 : #endif /** __COLL_COMMON_EXECUTOR_H__ */
|
