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 ALG_TEMPLATE_BASE_PUB_H
12 : #define ALG_TEMPLATE_BASE_PUB_H
13 :
14 : #include <cstring>
15 : #include <vector>
16 : #include <memory>
17 : #include <list>
18 : #include "hccl/base.h"
19 : #include "externalinput_pub.h"
20 : #include "mem_device_pub.h"
21 : #include "stream_pub.h"
22 : #include "transport_pub.h"
23 : #include "adapter_pub.h"
24 : #include "dispatcher.h"
25 : #include "local_notify.h"
26 : #include "template_v1_utils.h"
27 : #include "op_context.h"
28 : #include "comm_ahc_pub.h"
29 :
30 : namespace hccl {
31 : constexpr s32 HCCL_EXEC_STAGE_NOT_SET = -1;
32 : constexpr s32 HCCL_EXEC_STEP_NOT_SET = -1;
33 : constexpr s32 HCCL_EXEC_PLANE_NOT_SET = -1;
34 : constexpr u64 ZERO_SLICE = 0;
35 : constexpr u32 TWO_RANK_SIZE = 2;
36 : constexpr u32 DMA_REDUCE_TWO_OFFSET = 2;
37 : constexpr u32 DMA_REDUCE_THREE_OFFSET = 3;
38 : constexpr u64 HCCL_CHUNK_SIZE = 1024 * 1024 * 1024; // 1024*1024*1024的size
39 : constexpr u64 HCCL_MIN_PIPLINE_SLICE_ALIGN = 512;
40 : constexpr u64 HCCL_MIN_SLICE_ALIGN_910B = 16384;
41 : constexpr u64 HCCL_MIN_SLICE_ALIGN_910_93 = 16384;
42 : constexpr u64 HCCL_MIN_SLICE_ALIGN_ONCHIP = 512;
43 : constexpr u64 HCCL_MIN_SLICE_ALIGN = 128;
44 : constexpr u64 HCCL_NIC_MAX_NUM = 8;
45 : constexpr u64 DOUBLE_RING_NUM = 2;
46 : constexpr u64 DOUBLE_RING_STREAM_NUM = 3;
47 : constexpr u32 ALIGNED_SUB_RING_INDEX = 0;
48 : constexpr u32 ALIGNED_MAIN_RING_INDEX = 1;
49 :
50 : // AnyPath相关，SDMA数据量切分比例
51 : constexpr u32 MAX_SPLIT_VALUE = 100;
52 : constexpr u32 BEST_SPLIT_VALUE_SR = 87;
53 : constexpr u32 BEST_SPLIT_VALUE_DR = 90;
54 : constexpr u64 HCCL_SPLIT_SIZE_INTER_SERVER = 8388608; // 每卡通信量的切分边界
55 :
56 : enum TemplateType {
57 : // 内置template
58 : TEMPLATE_ALL_GATHER_HD_STAGE = 0, // AllGatherHDStage
59 : TEMPLATE_ALL_2_ALL_V_DIRECT_FULL_MESH = 1, // AlltoAllVDirectFullMesh
60 : TEMPLATE_ALL_REDUCE_REDUCE_BCAST = 2, // AllReduceReduceBcast
61 : TEMPLATE_BROADCAST_NHR_V1 = 3, // BroadcastNHRV1
62 : TEMPLATE_BROADCAST_NHR = 4, // BroadcastNHR
63 : TEMPLATE_BROADCAST_NHR_ONESHOT = 5, // BroadcastNHROneshot
64 : TEMPLATE_BROADCAST_NB = 6, // BroadcastNB
65 : TEMPLATE_BROADCAST_NB_BINARY = 7, // BroadcastNBBinary
66 : TEMPLATE_BROADCAST_HD = 8, // BroadcastHD
67 : TEMPLATE_BROADCAST_RECURSIVE_HD = 10, // BcastRecursiveHalvingDoubling
68 : TEMPLATE_BROADCAST_RING = 11, // BroadcastRing
69 : TEMPLATE_BROADCAST_STAR = 12, // BroadcastStar
70 : TEMPLATE_ALL_2_ALL_V_FOR310P = 13, // AlltoAllVFor310P
71 : TEMPLATE_ALL_2_ALL_V_PAIRWISE = 15, // AlltoAllVPairwise
72 : TEMPLATE_ALL_2_ALL_V_STAGED_MESH = 16, // AlltoAllVStagedMesh
73 : TEMPLATE_ALL_2_ALL_V_STAGED_PAIRWISE = 17, // AlltoAllVStagedPairwise
74 : TEMPLATE_REDUCESCATTER_HDSTAGE = 18, // ReduceScatterHDStage
75 : TEMPLATE_REDUCESCATTER_LOCAL_REDUCE = 19, // ReduceScatterLocalReduce
76 : TEMPLATE_REDUCESCATTER_NB = 20, // ReduceScatterNB
77 : TEMPLATE_REDUCESCATTER_NHR = 21, // ReduceScatterNHR
78 : TEMPLATE_REDUCESCATTER_NHR_V1 = 22, // ReduceScatterNHRV1
79 : TEMPLATE_REDUCESCATTER_PIPELINE = 23, // ReduceScatterPipeline
80 : TEMPLATE_REDUCESCATTER_UNIFIED_MARCH = 24, // ReduceScatterUnifiedMarch
81 : TEMPLATE_REDUCESCATTER_DB_RING_SLC = 25, // AlignedReduceScatterDoubleRingWithSerialLocalCopy
82 : TEMPLATE_REDUCESCATTER_DB_RING = 26, // AlignedReduceScatterDoubleRing
83 : TEMPLATE_REDUCESCATTER_HD = 27, // ReduceScatterHalvingDoubling
84 : TEMPLATE_REDUCESCATTER_MESH_DIRECT = 28, // ReduceScatterMeshDirect
85 : TEMPLATE_REDUCESCATTER_MESH_ATOMIC = 29, // ReduceScatterMeshAtomic
86 : TEMPLATE_REDUCESCATTER_MESH_MIX_SS = 30, // ReduceScatterMeshMixSingleStream
87 : TEMPLATE_REDUCESCATTER_MESH_MIX = 31, // ReduceScatterMeshMix
88 : TEMPLATE_REDUCESCATTER_MESH = 32, // ReduceScatterMesh
89 : TEMPLATE_REDUCESCATTER_RECURSIVE_HD = 33, // ReduceScatterRecursiveHalvingDoubling
90 : TEMPLATE_REDUCESCATTER_RING_DIRECT = 34, // ReduceScatterRingConcurrentDirect
91 : TEMPLATE_REDUCESCATTER_RING = 35, // ReduceScatterRing
92 : TEMPLATE_ALL_REDUCE_RECURSIVE_HALVING_DOUBLING = 36,
93 : TEMPLATE_ALL_REDUCE_RING = 37,
94 : TEMPLATE_REDUCE_RECURSIVE_HALVING_DOUBLING = 38,
95 : TEMPLATE_REDUCE_RING = 39,
96 : TEMPLATE_REDUCE_NHR_ONE_SHOT = 40,
97 : TEMPLATE_ALL_REDUCE_CHUNK_MESH = 41,
98 : TEMPLATE_ALL_REDUCE_DOUBLING_DIRECT = 42,
99 : TEMPLATE_ALL_REDUCE_DOUBLING = 43,
100 : TEMPLATE_ALL_REDUCE_HD_OPTIM = 44,
101 : TEMPLATE_ALL_REDUCE_LOCAL_REDUCE_BCAST = 45,
102 : TEMPLATE_ALL_REDUCE_LOCAL_REDUCE = 46,
103 : TEMPLATE_ALL_REDUCE_MESH_DIRECT_ONESHOT = 47,
104 : TEMPLATE_ALL_REDUCE_MESH_DIRECT = 48,
105 : TEMPLATE_ALL_REDUCE_NB = 49,
106 : TEMPLATE_ALL_REDUCE_NHR_ONESHOT = 50,
107 : TEMPLATE_ALL_REDUCE_NHR_V1 = 51,
108 : TEMPLATE_ALL_REDUCE_NHR = 52,
109 : TEMPLATE_ALL_REDUCE_OPBASE_PIPELINE = 53,
110 : TEMPLATE_ALIGNED_ALL_GATHER_DOUBLE_RING = 54, // AlignedAllGatherDoubleRing
111 : TEMPLATE_ALL_GATHER_HALVING_DOUBLING = 55, // AllGatherHalvingDoubling
112 : TEMPLATE_ALL_GATHER_MESH = 56, // AllGatherMesh
113 : TEMPLATE_ALL_GATHER_MESH_ATOMIC = 57, // AllGatherMeshAtomic
114 : TEMPLATE_ALL_GATHER_MESH_DIRECT = 58, // AllGatherMeshDirect
115 : TEMPLATE_ALL_GATHER_MESH_MIX = 59, // AllGatherMeshMix
116 : TEMPLATE_ALL_GATHER_RECURSIVE_HALVING_DOUBLING = 60, // AllGatherRecursiveHalvingDoubling
117 : TEMPLATE_ALL_GATHER_RING_CONCURRENT_DIRECT = 61, // AllGatherRingConcurrentDirect
118 : TEMPLATE_ALL_GATHER_RING = 62, // AllGatherRing
119 : TEMPLATE_ALL_GATHER_NB = 63, // AllGatherNB
120 : TEMPLATE_ALL_GATHER_NHRV1 = 64, // AllGatherNHRV1
121 : TEMPLATE_ALL_GATHER_NHR = 65, // AllGatherNHR
122 : TEMPLATE_ALL_GATHER_PIPELINE = 66, // AllGatherPipeline
123 : TEMPLATE_ALL_GATHER_UNIFIED_MARCH = 67, // AllGatherUnifiedMarch
124 :
125 : TEMPLATE_MULTI_ROOT_SCATTER_RING = 68, // MultiRootScatterRing
126 : TEMPLATE_SCATTER_DOUBLE_RING_DIRECT = 69, // ScatterDoubleRingDirect
127 : TEMPLATE_SCATTER_MESH = 70, // ScatterMesh
128 : TEMPLATE_SCATTER_RING_CONCURRENT_DIRECT = 71, // ScatterRingConcurrentDirect
129 : TEMPLATE_SCATTER_RING = 72, // ScatterRing
130 : TEMPLATE_SCATTER_NB = 73, // ScatterNB
131 : TEMPLATE_SCATTER_NHR = 74, // ScatterNHR
132 : TEMPLATE_SCATTER_RING_DIRECT = 75, // ScatterRingDirect
133 :
134 : TEMPLATE_GATHER_MESH = 76, // GatherMesh
135 : TEMPLATE_GATHER_RING = 77, // GatherRing
136 : TEMPLATE_GATHER_STAR = 78, // GatherStar
137 :
138 : TEMPLATE_ALL_2_ALL_PIPELINE_MESH_PAIRWISE_CCL_ENOUGH = 79, // AlltoallPipelineMeshPairwiseCCLEnough
139 : TEMPLATE_ALL_2_ALL_PIPELINE_MESH_PAIRWISE_PING_PONG = 80, // AlltoallPipelineMeshPairwisePingPong
140 :
141 : TEMPLATE_ALL_REDUCE_AHC = 81,
142 : TEMPLATE_ALL_REDUCE_AHC_BROKE = 82,
143 : TEMPLATE_ALL_GATHER_AHC = 83,
144 : TEMPLATE_ALL_GATHER_AHC_BROKE = 84,
145 : TEMPLATE_REDUCESCATTER_AHC = 85,
146 : TEMPLATE_REDUCESCATTER_AHC_BROKE = 86,
147 : TEMPLATE_ALL_GATHER_RING_DIRECT = 87, // AllGatherRingDirect
148 : TEMPLATE_ALL_GATHER_HCCS_SIO = 88,
149 : TEMPLATE_REDUCESCATTER_HCCS_SIO = 89,
150 :
151 : TEMPLATE_ALLREDUCE_GRAPH_PIPELINE = 90, // AllReduceGraphPipeline
152 :
153 : TEMPLATE_ALL_GATHER_GRAPH_PIPELINE = 91, // AllGatherGraphPipeline AG图模式pipeline
154 : TEMPLATE_REDUCESCATTER_GRAPH_PIPELINE = 92, // ReduceScatterGraphPipeline AG图模式pipeline
155 :
156 : TEMPLATE_ALL_GATHER_SLIM_RING = 93,
157 : TEMPLATE_REDUCESCATTER_SLIM_RING =94,
158 :
159 : TEMPLATE_REDUCESCATTER_PLANT_LOCAL_REDUCE = 95, // ReduceScatterPlantLocalReduce RS规约保序单机
160 : TEMPLATE_REDUCESCATTER_PLANT_LOCAL_REDUCE_COMBINE = 96, // ReduceScatterPlantLocalReduceCombine RS规约保序跨机
161 : TEMPLATE_REDUCESCATTER_V_PIPELINE = 97, //ReduceScatterVPipeline RSV多机Pipeline
162 :
163 : TEMPLATE_ALL_GATHER_V_PIPELINE = 98, // AllGatherV pipeline
164 :
165 : TEMPLATE_ALL_REDUCE_DOUBLING_LOCAL_REDUCE = 99, // AllReduceDoublingLocalReduce AR 910A单机小数据量tbe reduce优化
166 :
167 : TEMPLATE_ALL_2_ALL_V_CONTINUOUS_PIPELINE = 100, // AlltoallvContinuousPipeline
168 :
169 : TEMPLATE_ALL_GATHER_V_GRAPH_PIPELINE = 101, // AllGatherV Graph pipeline
170 : TEMPLATE_REDUCESCATTER_MULTI_DETERMINISTIC_PIPELINE = 102,
171 : TEMPLATE_ALL_REDUCE_MULTI_DETERMINISTIC_PIPELINE = 103,
172 : TEMPLATE_ALL_2_ALL_FULL_MESH_SYMMETRIC_MEMORY = 104,
173 :
174 : TEMPLATE_NATIVE_MAX_NUM, // 内置template最大值
175 :
176 : TEMPLATE_CUSTOM_BEGIN = 1000, // 用户自定义template起始值
177 : TEMPLATE_CUSTOM_MAX_NUM = 2000 // 用户自定义template最大值
178 : };
179 :
180 : enum class SliceType {
181 : SLICE_TYPE_TX,
182 : SLICE_TYPE_RX
183 : };
184 :
185 : using GroupSlicesInfo = std::vector<MemBlockInfo>;
186 :
187 : enum class HalvingDoublingType {
188 : BINARY_BLOCK_HALVING_DOUBLING,
189 : RECURSIVE_HALVING_DOUBLING,
190 : RESERVED_ALGORITHM_TYPE
191 : };
192 :
193 : using SliceType = enum SliceType;
194 :
195 : enum class RunStage {
196 : RUN_PREPARE,
197 : RUN_REDUCE_SCATTER,
198 : RUN_ALLGATHER,
199 : RUN_ALLREDUCE,
200 : RUN_DEFAULT
201 : };
202 :
203 : struct PrepareData {
204 : u32 root = INVALID_VALUE_RANKID;
205 : u32 userRank = INVALID_VALUE_RANKID;
206 : u32 userRankSize = 0;
207 : u32 interRank = INVALID_VALUE_RANKID;
208 : u32 interRankSize = 0;
209 :
210 : u64 count = 0;
211 : HcclDataType dataType = HCCL_DATA_TYPE_RESERVED;
212 : HcclReduceOp reductionOp = HCCL_REDUCE_RESERVED;
213 : u64 baseOffset = 0;
214 :
215 : DeviceMem inputMem;
216 : DeviceMem outputMem;
217 : DeviceMem scratchMem;
218 : DeviceMem cclInMem;
219 : DeviceMem cclOutMem;
220 :
221 : Stream stream;
222 : const std::vector<Stream>* subStreamsPtr = nullptr;
223 : const std::vector<std::shared_ptr<LocalNotify>>* signalPtr = nullptr;
224 : const std::vector<std::shared_ptr<LocalNotify>>* signalAuxPtr = nullptr;
225 :
226 : const std::vector<LINK>* linksPtr = nullptr;
227 : const std::vector<Slice>* slicesPtr = nullptr;
228 : const std::vector<std::vector<Slice>>* multRingsSlicesPtr = nullptr;
229 : const std::vector<u32>* nicRankListPtr = nullptr;
230 :
231 : HcclWorkflowMode workMode = HcclWorkflowMode::HCCL_WORKFLOW_MODE_RESERVED;
232 : HcomCollOpInfo *opInfo = nullptr;
233 : bool disableDMAReduce = false;
234 : bool isSuPodAsym = false;
235 : HcclCMDType opType = HcclCMDType::HCCL_CMD_INVALID;
236 :
237 : const SendRecvInfo *localSendRecvInfoPtr = nullptr;
238 : const ZCopySendRecvInfo *sendRecvInfoPtr = nullptr;
239 : u32 devNumInlocalPod = 0;
240 : u32 rankIdxInPod = 0;
241 : u64 reduceAttr = 0;
242 :
243 : AlgOpContext algOpContext;
244 :
245 : bool needAlltoallvCache = false; // 用于alltoallv类算子的aicpu cache
246 : };
247 :
248 : struct HcclTopoInfo;
249 : class TopoMatcher;
250 : class ExecutorBase {
251 : public:
252 : explicit ExecutorBase(const HcclDispatcher dispatcher);
253 : virtual ~ExecutorBase();
254 :
255 : virtual HcclResult RunAsync();
256 : virtual HcclResult RunAsync(const u32 rank, const u32 rankSize,
257 : const std::vector<std::shared_ptr<Transport> > &links);
258 : virtual HcclResult RunAsyncStaged(const u32 rank, const u32 rankSize,
259 : const std::vector<std::shared_ptr<Transport> > &links, RunStage stage);
260 : /* 12个参数 */
261 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
262 : const HcclDataType dataType, const Stream &stream,
263 : const HcclReduceOp reductionOp = HCCL_REDUCE_RESERVED,
264 : const u32 root = INVALID_VALUE_RANKID,
265 : const std::vector<Slice> &slices = std::vector<Slice>(ZERO_SLICE),
266 : const u64 baseOffset = 0, std::vector<u32> nicRankList = {0, 1, 2, 3, 4, 5, 6, 7},
267 : const bool disableDMAReduce = false);
268 :
269 : /* 11个参数 */
270 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &scratchMem, const u64 count,
271 : const HcclDataType dataType,
272 : const Stream &stream, const HcclReduceOp reductionOp = HCCL_REDUCE_RESERVED,
273 : const u32 root = INVALID_VALUE_RANKID,
274 : const std::vector<Slice> &slices = std::vector<Slice>(ZERO_SLICE),
275 : const u64 baseOffset = 0, std::vector<u32> nicRankList = {0, 1, 2, 3, 4, 5, 6, 7},
276 : const bool disableDMAReduce = false);
277 :
278 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
279 : const HcclDataType dataType, const Stream &stream,
280 : const std::vector<std::vector<Slice>> &multRingsSlices,
281 : const HcclReduceOp reductionOp = HCCL_REDUCE_RESERVED,
282 : const u32 root = INVALID_VALUE_RANKID,
283 : const u64 baseOffset = 0,
284 : const bool disableDMAReduce = false);
285 :
286 : virtual HcclResult Prepare(PrepareData ¶m);
287 :
288 : /* 1个参数 */
289 : // AllGatherNHR, ScatterNHR
290 : virtual HcclResult Prepare(bool needSaveRankMap);
291 :
292 : // AHC 扩展参数
293 : virtual HcclResult Prepare(AHCExtendPreparePara &extendParam);
294 :
295 : // GatherStar
296 : virtual HcclResult Prepare(u32 userRank);
297 :
298 : /* 2个参数 */
299 : // ReduceScatterNB, ReduceScatterNHRV1, ReduceScatterRing, ReduceScatterRecursiveHalvingDoubling
300 : virtual HcclResult Prepare(u64 reduceAttrBitMap, HcomCollOpInfo *opInfo = nullptr);
301 :
302 : // ReduceScatterNHR
303 : virtual HcclResult Prepare(u64 reduceAttrBitMap, bool needMerge);
304 :
305 : // ReduceScatterMeshMixSingleStream, ReduceScatterMesh
306 : virtual HcclResult Prepare(u64 reduceAttrBitMap, u32 streamIndex);
307 :
308 : // ScatterMesh
309 : virtual HcclResult Prepare(u32 interRank, u32 interRankSize);
310 :
311 : /* 3个参数 */
312 : // for AllGatherHalvingDoubling based on input_scratch_Mem_nicRankList Prepare
313 : // and should be called soon template AllGatherHalvingDoubling created
314 : virtual HcclResult Prepare(u32 blockSize, UserMemType hdInputMemType, UserMemType hdOutputMemType);
315 :
316 : /* 4个参数 */
317 : // ScatterRingDirect
318 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, const u32 userRank, const std::vector<u32> &ringsOrders,
319 : const std::vector<Slice> &userMemInputSlices);
320 :
321 : // AllGatherRingDirect
322 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, u32 userRank,
323 : const std::vector<Slice> &userMemOutputSlices, bool isSdma = true);
324 :
325 : /* 5个参数 */
326 : // AHC 5个参数，带扩展参数
327 : virtual HcclResult Prepare(u64 totalCount, const std::vector<std::vector<std::vector<u32>>> &subGroups,
328 : std::map<AHCConcOpType, TemplateType> &ahcAlgOption, bool extendFlag = false,
329 : AHCExtendPreparePara extendPara = AHCExtendPreparePara());
330 :
331 : /* 6个参数 */
332 : // AlltoAllVStagedPairwise
333 : virtual HcclResult Prepare(DeviceMem &sendMem, DeviceMem &recvMem, StageAlltoAllVAddrInfo &sendAddrInfo,
334 : StageAlltoAllVAddrInfo &recvAddrInfo, bool isAlltoAllZCopyMode, Stream &mainStream);
335 :
336 : /* 7个参数 */
337 : virtual HcclResult Prepare(u64 reduceAttrBitMap, std::vector<Stream> &meshStreams,
338 : std::vector<std::shared_ptr<LocalNotify>> &meshSignal, std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
339 : u32 userRank, HcomCollOpInfo *opInfo, bool aicpu);
340 :
341 : // AlltoAllVPairWise
342 : virtual HcclResult Prepare(AlltoAllVBufferInfo &sendBuffer, AlltoAllVBufferInfo &recvBuffer,
343 : bool isAlltoAllZCopyMode, const Stream &stream, HcclWorkflowMode workMode,
344 : std::map<u32, std::vector<u64>> &rankSendDisplsMap, std::map<u32, std::vector<u64>> &rankRecvDisplsMap);
345 :
346 : // AlignedAllGatherDoubleRing
347 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, const u32 userRank, std::vector<Stream> &subStreams,
348 : std::vector<std::shared_ptr<LocalNotify>> &mainSignals, std::vector<std::shared_ptr<LocalNotify>> &subSignals,
349 : const std::vector<std::vector<u32>> &ringsOrders,
350 : const std::vector<std::vector<Slice>> &userMemOutputSlicesOfDoubleRing);
351 :
352 : // AllGatherMeshAtomic, AllgatherMeshDirect, AllGatherMesh, AllGatherMeshMix, GatherMesh
353 : virtual HcclResult Prepare(std::vector<Stream> &meshStreams, std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
354 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, u32 userRank = INVALID_VALUE_RANKID,
355 : HcomCollOpInfo *opInfo = nullptr, u32 interRank = INVALID_VALUE_RANKID, u32 interRankSize = 0);
356 :
357 : /* 8个参数 */
358 : virtual HcclResult Prepare(u64 reduceAttrBitMap, std::vector<Stream> &meshStreams,
359 : std::vector<std::shared_ptr<LocalNotify>> &meshSignal, std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
360 : u32 interRank, u32 interRankSize, u32 userRank, HcomCollOpInfo *opInfo);
361 :
362 : // AlltoAllVStagedPairwise
363 : virtual HcclResult Prepare(DeviceMem &sendMem, DeviceMem &recvMem, DeviceMem &scratchInputMem,
364 : DeviceMem &scratchOutputMem, StageAlltoAllVAddrInfo &sendAddrInfo, StageAlltoAllVAddrInfo &recvAddrInfo,
365 : bool isAlltoAllZCopyMode, Stream &mainStream);
366 :
367 : // AllGatherRingConcurrentDirect ScatterRingConcurrentDirect
368 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, const u32 userRank, std::vector<Stream> &subStreams,
369 : const std::vector<std::shared_ptr<LocalNotify>> &mainSignals,
370 : const std::vector<std::shared_ptr<LocalNotify>> &subSignals, const std::vector<u32> &ringsOrder,
371 : const std::vector<Slice> &userMemSlices, bool isSdma = true);
372 :
373 : /* 9个参数 */
374 : // scatterDoubleRingDirect
375 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, const u32 userRank, const u32 subRingRank,
376 : std::vector<Stream> &subStreams, const std::vector<std::shared_ptr<LocalNotify>> &mainSignals,
377 : const std::vector<std::shared_ptr<LocalNotify>> &subSignals, const std::vector<std::vector<u32>> &ringsOrders,
378 : const std::vector<std::vector<Slice>> &multiRingSlices,
379 : const std::vector<std::vector<Slice>> &userMemInputSlices);
380 :
381 : // ReduceScatterRingConcurrentDirect
382 : virtual HcclResult Prepare(const u64 reduceAttrBitMap, const HcomCollOpInfo *opInfo, const u32 userRank,
383 : std::vector<Stream> &subStreams, const std::vector<std::shared_ptr<LocalNotify>> &mainSignals,
384 : const std::vector<std::shared_ptr<LocalNotify>> &subSignals, const std::vector<u32> &ringsOrder,
385 : const std::vector<Slice> &userMemInputSlices, bool isSdma = true);
386 :
387 : // AlltoAllVPairWise
388 : virtual HcclResult Prepare(AlltoAllVBufferInfo &sendBuffer, AlltoAllVBufferInfo &recvBuffer,
389 : DeviceMem &scratchInputMem, DeviceMem &scratchOutputMem, bool isAlltoAllZCopyMode, const Stream &stream,
390 : HcclWorkflowMode workMode, std::map<u32, std::vector<u64>> &rankSendDisplsMap,
391 : std::map<u32, std::vector<u64>> &rankRecvDisplsMap);
392 :
393 : /* 10个参数 */
394 : virtual HcclResult Prepare(const HcomCollOpInfo *opInfo, DeviceMem &cclBufferA, DeviceMem &cclBufferB,
395 : const u64 count, const SubCommInfo &level1CommInfo, const SubCommInfo &level0CommInfo, Stream &mainStream,
396 : std::vector<Stream> &subStream, std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain,
397 : std::vector<std::shared_ptr<LocalNotify>> ¬ifySub);
398 :
399 : // AlltoAllPipelineMeshPairwiseCCLEnough, AlltoAllPipelineMeshPairwisePingPong
400 : virtual HcclResult Prepare(u32 userRank, A2aPipelineMemory A2aPipelineMemory, const SubCommInfo &level0CommInfo,
401 : const SubCommInfo &level1CommInfo, Stream &mainStream, std::vector<Stream> &subStream,
402 : std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain, std::vector<std::shared_ptr<LocalNotify>> ¬ifySub,
403 : std::vector<SendRecvInfo> &allMeshAggregationSendRecvInfo,
404 : HcclWorkflowMode workMode = HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE);
405 :
406 : // AlltoAllVSatgedMesh
407 : virtual HcclResult Prepare(DeviceMem &sendMem, DeviceMem &recvMem, StageAlltoAllVAddrInfo &sendAddrInfo,
408 : StageAlltoAllVAddrInfo &recvAddrInfo, bool isAlltoAllZCopyMode, u32 userRank, Stream &mainStream,
409 : std::vector<Stream> &subStreams, std::vector<std::shared_ptr<LocalNotify>> &meshSignalMainToSub,
410 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalSubToMain);
411 :
412 : // ReduceScatterPlantLocalReduce
413 : virtual HcclResult Prepare(void *inputMemPtr, DeviceMem &cclInMem, DeviceMem &outputMem,
414 : const Stream &stream, std::vector<Stream> &subStreams,
415 : std::vector<std::shared_ptr<LocalNotify>> &meshSignal, std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
416 : GroupSlicesInfo &grouSlicesInfo, const HcclReduceOp reductionOp, u32 all2allOffset, const HcclDataType dataType,
417 : bool isNeedSpaceBorrow, bool reverseMemUsage = false, bool isA3CrossNode = false);
418 :
419 : // ReduceScatterPlantLocalReduceCombine
420 : virtual HcclResult Prepare(DeviceMem &cclInMem, DeviceMem &outputMem,
421 : const Stream &stream, std::vector<Stream> &subStreams, std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
422 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, MemBlockInfo &memBlockInfo,
423 : const HcclReduceOp reductionOp, const HcclDataType dataType, bool isUseCclIn, bool isLevel0LastRank, bool isNeedSpaceBorrow);
424 :
425 : /* 11个参数 */
426 : // Prepare for AllGatherPipeline
427 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, u32 userRank, u64 &count, DeviceMem &cclBufferPartOne,
428 : DeviceMem &cclBufferPartTwo, SubCommInfo &level0CommInfo, SubCommInfo &level1CommInfo, Stream &mainStream,
429 : std::vector<Stream> &subStream, std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain,
430 : std::vector<std::shared_ptr<LocalNotify>> ¬ifySub);
431 :
432 : // Prepare for AllGatherUnifiedMarch
433 : virtual HcclResult Prepare(const Stream &mainStream, SubCommInfo &level0CommInfo, DeviceMem &userInput,
434 : DeviceMem &userOutput, DeviceMem &usrInMem, DeviceMem &usrOutMem, u64 blockDataByte,
435 : std::vector<Stream> &subStreams, const std::vector<std::shared_ptr<LocalNotify>> &meshSignalMainToSub,
436 : const std::vector<std::shared_ptr<LocalNotify>> &meshSignalSubToMain,
437 : const std::vector<std::vector<Slice>> &multRingsUserMemSlice, const u64 baseOffset = 0);
438 :
439 : // Prepare for AllGatherHccsSio
440 : virtual HcclResult Prepare(SubCommInfo &outerCommInfoHccs, SubCommInfo &outerCommInfoSio, DeviceMem &usrInMem,
441 : DeviceMem &usrOutMem, u64 totalCount, const HcclDataType dataType, const Stream &mainStream,
442 : std::vector<Stream> &meshStreams, std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
443 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, u32 userRank, HcomCollOpInfo *opInfo);
444 :
445 : // Prepare for AlltoAllvContinuousPipeline
446 : virtual HcclResult Prepare(const u32 userRank, const A2aPipelineMemory &a2aPipelineMemory,
447 : const SubCommInfo &level0CommInfo, const SubCommInfo &level1CommInfo,
448 : const Stream &mainStream, std::vector<Stream> &subStream,
449 : std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain, std::vector<std::shared_ptr<LocalNotify>> ¬ifySub,
450 : std::vector<SendRecvInfo> &sendRecvInfoList, const HcclDataType dataType,
451 : const HcclWorkflowMode workMode);
452 :
453 : /* 12个参数 */
454 : // AlltoAllVFor310P
455 : virtual HcclResult Prepare(DeviceMem &userInput, DeviceMem &userOutput, DeviceMem &cclInMem, DeviceMem &cclOutMem,
456 : const std::vector<std::shared_ptr<LocalNotify>> &signalMainToSub,
457 : const std::vector<std::shared_ptr<LocalNotify>> &signalSubToMain, Stream &mainStream,
458 : std::vector<Stream> &subStreams, const std::vector<LINK> &links, u32 userRank, u32 userRankSize,
459 : std::vector<SendRecvInfo> &allMeshAggregationSendRecvInfo);
460 :
461 : // AlltoAllVStagedMesh
462 : virtual HcclResult Prepare(DeviceMem &sendMem, DeviceMem &recvMem, DeviceMem &scratchInputMem,
463 : DeviceMem &scratchOutputMem, StageAlltoAllVAddrInfo &sendAddrInfo, StageAlltoAllVAddrInfo &recvAddrInfo,
464 : bool isAlltoAllZCopyMode, u32 userRank, Stream &mainStream, std::vector<Stream> &subStreams,
465 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalMainToSub,
466 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalSubToMain);
467 :
468 : // ReduceScatterPipeline
469 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, DeviceMem &cclBuffer, const u64 count, const u64 bufferSize,
470 : const u64 offset, const SubCommInfo &level0CommInfo, const SubCommInfo &level1CommInfo, Stream &mainStream,
471 : std::vector<Stream> &subStream, std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain,
472 : std::vector<std::shared_ptr<LocalNotify>> ¬ifySub, u64 reduceAttrBitMap);
473 :
474 : // ReduceScatterVPipeline
475 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, DeviceMem &cclBuffer, const u64 bufferSize,
476 : const std::vector<Slice> &slices, const SubCommInfo &level0CommInfo, const SubCommInfo &level1CommInfo,
477 : Stream &mainStream, std::vector<Stream> &subStream,
478 : std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain,
479 : std::vector<std::shared_ptr<LocalNotify>> ¬ifySub,
480 : u64 reduceAttrBitMap);
481 :
482 : // BroadcastStar
483 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
484 : const HcclDataType dataType, const Stream &stream, const HcclReduceOp reductionOp, const u32 root,
485 : const std::vector<Slice> &slices, const u64 baseOffset, std::vector<u32> nicRankList, u32 userRank);
486 :
487 : // Prepare for AllGatherVPipeline
488 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, u32 userRank, u64 &count, DeviceMem &cclBufferPartOne,
489 : DeviceMem &cclBufferPartTwo, SubCommInfo &level0CommInfo, SubCommInfo &level1CommInfo, Stream &mainStream,
490 : std::vector<Stream> &subStream, std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain,
491 : std::vector<std::shared_ptr<LocalNotify>> ¬ifySub, std::vector<Slice>& userOutSlice);
492 :
493 : /* 13个参数 */
494 : // BroadcastHD
495 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
496 : const HcclDataType dataType, const Stream &stream, const HcclReduceOp reductionOp, const u32 root,
497 : std::vector<Stream> &meshStreams, const std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
498 : const std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, u32 interRank, const HcomCollOpInfo *opInfo);
499 :
500 : /* 14个参数 */
501 : // ReduceScatterUnifiedMarch
502 : virtual HcclResult Prepare(Stream &mainStream, SubCommInfo &level0CommInfo, DeviceMem &userInput,
503 : DeviceMem &userOutput, DeviceMem &usrInMem, DeviceMem &scratchMem, u64 totalCount,
504 : std::vector<Stream> &subStreams, const std::vector<std::shared_ptr<LocalNotify>> &meshSignalMainToSub,
505 : const std::vector<std::shared_ptr<LocalNotify>> &meshSignalSubToMain, const HcclDataType dataType,
506 : const HcclReduceOp reductionOp, const std::vector<std::vector<Slice>> &multRingsUserMemSlice,
507 : u64 reduceAttrBitMap);
508 :
509 : // ReduceScatterHalvingDoubling
510 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
511 : const HcclDataType dataType, const Stream &stream, const HcclReduceOp reductionOp, const u32 root,
512 : const std::vector<Slice> &slices, const u64 baseOffset, const u32 blockSize, const u64 reduceAttrBitMap,
513 : const UserMemType hdInputMemType, const UserMemType hdOutputMemType);
514 :
515 : /* 15个参数 */
516 : // AlltoAllVMeshReadOnly
517 : virtual HcclResult Prepare(DeviceMem &userInput, DeviceMem &userOutput, DeviceMem &scratchPingMem,
518 : DeviceMem &scratchPongMem, StageAlltoAllVAddrInfo &sendAddrInfo, StageAlltoAllVAddrInfo &recvAddrInfo,
519 : HcclWorkflowMode workMode, Stream &mainStream, std::vector<Stream> &subStreams,
520 : const std::vector<std::shared_ptr<LocalNotify>> &meshSignalMainToSub,
521 : const std::vector<std::shared_ptr<LocalNotify>> &meshSignalSubToMain, u32 userRank, u32 intraRankSize,
522 : const std::vector<LINK> &links, std::vector<SendRecvInfo> &allMeshAggregationSendRecvInfo);
523 :
524 : /* 16个参数 */
525 : // ReduceScatterHDStage, ReduceScatterLocalReduce, ReduceScatterMeshAtomic, ReduceScatterMeshDirect
526 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
527 : const HcclDataType dataType, const Stream &stream, const HcclReduceOp reductionOp, const u32 root,
528 : const std::vector<Slice> &slices, const u64 baseOffset, const u64 reduceAttrBitMap,
529 : std::vector<Stream> &meshStreams, std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
530 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, u32 userRank, const HcomCollOpInfo *opInfo = nullptr);
531 :
532 : //ReduceScatterHccsSio
533 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
534 : const HcclDataType dataType, const Stream &stream, const HcclReduceOp reductionOp,
535 : const u32 root, const u64 baseOffset,
536 : const u64 reduceAttrBitMap, std::vector<Stream> &meshStreams,
537 : std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
538 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
539 : u32 userRank, SubCommInfo subCommInfoHccs, SubCommInfo subCommInfoSio, HcomCollOpInfo *opInfo);
540 :
541 : /* 17个参数 */
542 : // ReduceScatterMeshMix
543 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
544 : const HcclDataType dataType, const Stream &stream, const HcclReduceOp reductionOp, const u32 root,
545 : const std::vector<Slice> &slices, const u64 baseOffset, const u64 reduceAttrBitMap,
546 : std::vector<Stream> &meshStreams, const std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
547 : const std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux, u32 interRank, u32 interRankSize,
548 : HcomCollOpInfo *opInfo);
549 :
550 : /* 19个参数 */
551 : // AlignedReduceScatterDoubleRing, AlignedReduceScatter, DoubleRingWithSerialLocalCopy
552 : virtual HcclResult Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem, const u64 count,
553 : const HcclDataType dataType, const Stream &stream, const std::vector<std::vector<Slice>> &multRingsSlices,
554 : const HcclReduceOp reductionOp, const u32 root, const u64 baseOffset, const bool disableDMAReduce,
555 : const u64 reduceAttrBitMap, const HcomCollOpInfo *opInfo, const u32 userRank, std::vector<Stream> &subStreams,
556 : const std::vector<std::shared_ptr<LocalNotify>> &mainSignals,
557 : const std::vector<std::shared_ptr<LocalNotify>> &subSignals, const std::vector<std::vector<u32>> &ringsOrders,
558 : const std::vector<std::vector<Slice>> &userMemInputSlicesOfDoubleRing);
559 :
560 : // ReduceScatterDeterPipeline
561 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, DeviceMem &buffer, const u64 count,
562 : const u64 offset, const std::vector<Slice> &slices, const SubCommInfo &level0CommInfo,
563 : const SubCommInfo &level1CommInfo, Stream &mainStream, std::vector<Stream> &subStream,
564 : std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain, std::vector<std::shared_ptr<LocalNotify>> ¬ifySub);
565 :
566 : // AllReduceDeterPipeline
567 : virtual HcclResult Prepare(HcomCollOpInfo *opInfo, DeviceMem &inBuffer, DeviceMem &outBuffer, const u64 count,
568 : const std::vector<Slice> &slices, const SubCommInfo &level0CommInfo,
569 : const SubCommInfo &level1CommInfo, Stream &mainStream, std::vector<Stream> &subStream,
570 : std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain, std::vector<std::shared_ptr<LocalNotify>> ¬ifySub);
571 :
572 : HcclResult Sum(const std::vector<Slice> &inputSlices, u32 start, u32 num, u64 &sizeOut);
573 : HcclResult RegisterProfiler(s32 planeId, s32 stage, s32 step, const Stream &stream);
574 : static HcclResult ExecEmptyTask(DeviceMem &inputMem, DeviceMem &outputMem, Stream &stream,
575 : const HcclDispatcher dispatcher);
576 : HcclResult CheckConcurrentDirectParameters(const u32 rank, const u32 rankSize, const std::vector<LINK> &links);
577 185 : u32 DataUnitSize(HcclDataType dataType) const
578 : {
579 185 : if (dataType >= HCCL_DATA_TYPE_RESERVED) {
580 0 : HCCL_ERROR("[AlgTemplateBase][DataUnitSize]data type[%s] out of range[%d, %d]",
581 : GetDataTypeEnumStr(dataType).c_str(), HCCL_DATA_TYPE_INT8, HCCL_DATA_TYPE_RESERVED - 1);
582 0 : return 0;
583 : }
584 :
585 185 : return SIZE_TABLE[dataType];
586 : }
587 :
588 : static std::vector<bool> CalcLinksRelation(const u32 rank, const u32 rankSize, const u32 rootRank = 0,
589 : HalvingDoublingType algorithmType = HalvingDoublingType::RECURSIVE_HALVING_DOUBLING);
590 :
591 : static HcclResult PrepareSliceData(u64 dataCount, u32 unitSize, u32 sliceNum, u64 piplineOffset,
592 : std::vector<Slice> &dataSlice);
593 : static HcclResult PrepareSliceMeshStreams(const std::vector<Slice> &rankSegsSlice, u32 streamCount,
594 : std::vector<std::vector<Slice>> &mutliStreamsSlices);
595 :
596 136 : static inline u64 RoundUpWithDivisor(u64 value, u64 divisor)
597 : {
598 136 : if ((value == 0) || (divisor == 0)) {
599 0 : return divisor;
600 : }
601 : // divisor必须大于等于1, 返回value向上取divisor的整数倍的值
602 136 : return ((value + (divisor - 1)) / divisor) * divisor;
603 : }
604 : inline u64 ByteOffset(u64 countOffset) const
605 : {
606 : return countOffset * DataUnitSize(dataType_);
607 : }
608 : inline u64 SliceOffset(u32 sliceIndex, u64 countPerSlice) const
609 : {
610 : return sliceIndex * countPerSlice * DataUnitSize(dataType_);
611 : }
612 1 : inline void CloseBarrier()
613 : {
614 1 : barrierSwitchOn_ = false;
615 1 : }
616 : virtual HcclResult GetNslbAdjInfo(const u32 rank, const u32 rankSize,
617 : const std::vector<LINK> &links, AdjInfo& nslbAdjInfo);
618 :
619 : // 只用于alltoallv类算子的aicpu cache
620 : virtual HcclResult GetHcclOffsetDstRanksMap(std::unordered_map<uint64_t, std::vector<uint32_t>>& hcclOffsetDstRanksMap) const;
621 :
622 : protected:
623 : HcclResult ExecuteBarrier(const std::shared_ptr<Transport> &preLink, const std::shared_ptr<Transport> &aftLink);
624 : HcclResult ExecuteBarrier(const std::shared_ptr<Transport> &preLink,
625 : const std::shared_ptr<Transport> &aftLink, Stream &stream);
626 : HcclResult ExecuteBarrier(const std::shared_ptr<Transport> &preLink,
627 : const std::shared_ptr<Transport> &aftLink, u32 notifyIdx);
628 : HcclResult ExecuteBarrier(const std::shared_ptr<Transport> &preLink,
629 : const std::shared_ptr<Transport> &aftLink, u32 notifyIdx, Stream &stream);
630 : HcclResult ExecuteBarrier(std::shared_ptr<Transport> link, Stream &stream);
631 : HcclResult ExecuteRxSync(std::shared_ptr<Transport> link, UserMemType srcMemType, u64 srcOffset,
632 : void *dst, u64 len, Stream &stream) const;
633 : HcclResult ExecuteTxSync(std::shared_ptr<Transport> link, UserMemType dstMemType, u64 dstOffset,
634 : void *src, u64 len, Stream &stream) const;
635 : virtual HcclResult PrepareRunAsync(const u32 rank, const u32 rankSize,
636 : const std::vector<std::shared_ptr<Transport> > &links);
637 : const HcclDispatcher dispatcher_;
638 : std::vector<Slice> slicesDummy_;
639 : std::vector<Slice> &slices_;
640 : DeviceMem inputMem_; /* * 输入memory */
641 : DeviceMem outputMem_; /* * 输出memory */
642 : DeviceMem scratchMem_; /* * 草稿memory */
643 :
644 : u64 count_; // 需处理的每块memory数据总个数
645 : u64 dataBytes_; // 数据所占的字节数
646 : HcclDataType dataType_;
647 : HcclReduceOp reductionOp_;
648 : u32 root_;
649 : bool disableDMAReduce_;
650 :
651 : // Added on Mar.24th, for profiling template
652 : StepData profilerInput_;
653 : u64 baseOffset_;
654 :
655 : Stream stream_;
656 :
657 : // 用于chunk算法
658 : std::vector<u32> nicRankList_;
659 : std::vector<std::vector<u32>> rankSliceLists_;
660 : bool barrierSwitchOn_;
661 : // 用于91093 aligend double ring算法
662 : std::vector<std::vector<Slice>> multRingsSlices_;
663 : AlgOpContext algOpContext_;
664 : private:
665 : static void CalcBinaryBlockParams(u32 rank, u32 rankSize, u32 &stepsInBlock, u32 &lowerBlockSize,
666 : u32 &myBlockSize, u32 &rankInMyBlock, u32 &myBlockOffset, u32 &higherBlockSize);
667 : static HcclResult CalcBinaryBlockHalvingDoubleLinkReleation(u32 rank, u32 rankSize,
668 : std::vector<bool> &linkRelation);
669 :
670 : static void CalcLinkInBlock(u32 blockSize, u32 rankInBlock, std::list<u32> &linkRankIndexInBlock);
671 : static void CalcLinkBetweenParts(u32 part1Size, std::list<u32> &linkRankIndexInBlock,
672 : std::list<u32> &linkRankIndex, bool oddRank);
673 : static void CalcRecursiveHalvingDobuleLinkReleation(u32 rank, u32 rankSize, u32 rootRank,
674 : std::vector<bool> &linkRelation);
675 : static void CalcRecursiveHdLinkRelationForFirstScene(u32 rank,
676 : u32 part1Size, u32 blockSize, std::vector<bool> &linkRelation);
677 : static void CalcRecursiveHdLinkRelationForSecondScene(u32 rank,
678 : u32 part1Size, u32 blockSize, std::vector<bool> &linkRelation);
679 : };
680 : using AlgTemplateBase = ExecutorBase;
681 : } // namespace hccl
682 :
683 : #endif /* EXECUTOR_BASE_PUB_H */
|
