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_v_ring_for_910_93_executor.h"
12 : #include <numeric>
13 :
14 : namespace hccl {
15 0 : CollAllGatherVRingFor91093Executor::CollAllGatherVRingFor91093Executor(const HcclDispatcher dispatcher,
16 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
17 0 : : CollAllGatherRingFor91093Executor(dispatcher, topoMatcher)
18 : {
19 0 : isAllGatherV_ = true;
20 0 : desc_.level1SupportedAlgos = {
21 : AlgTypeLevel1::ALG_LEVEL1_NHR,
22 : AlgTypeLevel1::ALG_LEVEL1_NB,
23 : AlgTypeLevel1::ALG_LEVEL1_RING
24 0 : };
25 0 : }
26 :
27 0 : bool CollAllGatherVRingFor91093Executor::IsSmallData(const u64 size)
28 : {
29 : (void) size;
30 0 : return false;
31 : }
32 :
33 0 : u64 CollAllGatherVRingFor91093Executor::CalcDstMemOffset(const OpParam ¶m, u32 perDataSize, u64 inputMemSize) const
34 : {
35 : (void) inputMemSize;
36 0 : const auto *counts = static_cast<const u64 *>(param.VDataDes.counts);
37 0 : const u64 offset = std::accumulate(counts, counts + topoAttr_.userRank, 0ULL);
38 0 : return offset * perDataSize;
39 : }
40 :
41 0 : HcomCollOpInfo CollAllGatherVRingFor91093Executor::GetHcomCollOpInfo(const OpParam ¶m, const ExecMem &execMem) const
42 : {
43 0 : HcomCollOpInfo opInfo = {
44 0 : "", execMem.inputPtr, execMem.outputPtr, execMem.count, param.VDataDes.dataType, param.root,
45 0 : param.reduceType, 0 // 暂不支持MC2的strideCount特性
46 0 : };
47 0 : return opInfo;
48 : }
49 :
50 0 : std::vector<Slice> CollAllGatherVRingFor91093Executor::PrepareSlicesL2(const OpParam ¶m,
51 : const SubCommInfo &level2CommInfo, const SubCommInfo &level1CommInfo, const SubCommInfo &level0CommInfo,
52 : u32 perDataSize, u64 inputMemSize) const
53 : {
54 : (void) inputMemSize;
55 0 : std::vector<Slice> level2DataSegsSlice;
56 0 : const auto *counts = static_cast<u64 *>(param.VDataDes.counts);
57 0 : const u32 level0RankSize = level0CommInfo.localRankSize;
58 0 : const u32 level0ServerIndex = level0CommInfo.localRank;
59 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
60 0 : const u32 level1ServerIndex = level1CommInfo.localRank;
61 0 : const u32 level2RankSize = level2CommInfo.localRankSize;
62 0 : for (u32 i = 0; i < level2RankSize; i++) {
63 0 : const u32 rank = i * level1RankSize * level0RankSize + level1ServerIndex * level0RankSize + level0ServerIndex;
64 0 : Slice sliceTemp;
65 0 : sliceTemp.size = counts[rank] * perDataSize;
66 0 : const u64 offset = std::accumulate(counts, counts + rank, 0ULL);
67 0 : sliceTemp.offset = offset * perDataSize;
68 0 : level2DataSegsSlice.push_back(sliceTemp);
69 : }
70 0 : return level2DataSegsSlice;
71 0 : }
72 :
73 0 : std::vector<Slice> CollAllGatherVRingFor91093Executor::PrepareSlicesL1(const OpParam ¶m,
74 : const SubCommInfo &level2CommInfo, const SubCommInfo &level1CommInfo, const SubCommInfo &level0CommInfo,
75 : u32 perDataSize, u64 inputMemSize) const
76 : {
77 : (void) inputMemSize;
78 0 : const auto *counts = static_cast<u64 *>(param.VDataDes.counts);
79 0 : const u32 level0RankSize = level0CommInfo.localRankSize;
80 0 : const u32 level0ServerIndex = level0CommInfo.localRank;
81 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
82 0 : const u32 level2RankSize = level2CommInfo.localRankSize;
83 0 : std::vector<Slice> level1DataSegsSlice;
84 0 : for (u32 j = 0; j < level1RankSize; j++) {
85 0 : for (u32 i = 0; i < level2RankSize; i++) {
86 0 : Slice level1Slice;
87 0 : const u32 rank = i * level1RankSize * level0RankSize + j * level0RankSize + level0ServerIndex;
88 0 : level1Slice.size = counts[rank] * perDataSize;
89 0 : const u64 offset = std::accumulate(counts, counts + rank, 0ULL);
90 0 : level1Slice.offset = offset * perDataSize;
91 0 : level1DataSegsSlice.push_back(level1Slice);
92 : }
93 : }
94 0 : return level1DataSegsSlice;
95 0 : }
96 :
97 0 : HcclResult CollAllGatherVRingFor91093Executor::PrepareSlicesL0(std::vector<std::vector<Slice>> &multRingsSlice,
98 : const OpParam ¶m, const SubCommInfo &level2CommInfo, const SubCommInfo &level1CommInfo,
99 : const SubCommInfo &level0CommInfo, u32 perDataSize, u64 inputMemSize)
100 : {
101 : (void) inputMemSize;
102 0 : HCCL_CONFIG_INFO(HCCL_ALG,
103 : "[CollAllGatherVRingFor91093Executor][PrepareSlicesL0] userRank[%u] starts.", topoAttr_.userRank);
104 0 : const auto *counts = static_cast<u64 *>(param.VDataDes.counts);
105 0 : const u32 level0RankSize = level0CommInfo.localRankSize;
106 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
107 0 : const u32 level2RankSize = level2CommInfo.localRankSize;
108 0 : std::vector<Slice> dataSegsSlice;
109 0 : for (u32 k = 0; k < level0RankSize; k++) { // 根据数据量计算每个环上数据的偏移和大小
110 0 : for (u32 i = 0; i < level2RankSize; i++) {
111 0 : for (u32 j = 0; j < level1RankSize; j++) {
112 0 : Slice sliceTemp;
113 0 : const u32 rank = i * level1RankSize * level0RankSize + j * level0RankSize + k;
114 0 : sliceTemp.size = counts[rank] * perDataSize;
115 0 : const u64 offset = std::accumulate(counts, counts + rank, 0ULL);
116 0 : sliceTemp.offset = offset * perDataSize; // no displs
117 0 : dataSegsSlice.push_back(sliceTemp);
118 : }
119 : }
120 : }
121 0 : multRingsSlice.push_back(dataSegsSlice);
122 0 : return HCCL_SUCCESS;
123 0 : }
124 :
125 : // AGV不支持MC2的strideCount特性
126 0 : HcclResult CollAllGatherVRingFor91093Executor::PrepareUserMemSlices(std::vector<std::vector<Slice>> &userMemSlices,
127 : const std::vector<std::vector<Slice>> &multRingsSlice, const OpParam ¶m, const SubCommInfo &level2CommInfo,
128 : const SubCommInfo &level1CommInfo, const SubCommInfo &level0CommInfo, u32 perDataSize, u64 inputMemSize)
129 : {
130 : (void) inputMemSize;
131 : (void) multRingsSlice;
132 0 : const auto *counts = static_cast<u64 *>(param.VDataDes.counts);
133 0 : const auto *displs = static_cast<u64 *>(param.VDataDes.displs);
134 0 : std::vector<Slice> dataSegsSlice;
135 0 : const u32 level0RankSize = level0CommInfo.localRankSize;
136 0 : const u32 level1RankSize = level1CommInfo.localRankSize;
137 0 : const u32 level2RankSize = level2CommInfo.localRankSize;
138 0 : for (u32 k = 0; k < level0RankSize; k++) { // 根据数据量计算每个环上数据的偏移和大小
139 0 : for (u32 i = 0; i < level2RankSize; i++) {
140 0 : for (u32 j = 0; j < level1RankSize; j++) {
141 0 : Slice sliceTemp;
142 0 : const u32 rank = i * level1RankSize * level0RankSize + j * level0RankSize + k;
143 0 : sliceTemp.size = counts[rank] * perDataSize;
144 0 : sliceTemp.offset = displs[rank] * perDataSize; // with displs
145 0 : dataSegsSlice.push_back(sliceTemp);
146 : }
147 : }
148 : }
149 0 : userMemSlices.push_back(dataSegsSlice);
150 0 : return HCCL_SUCCESS;
151 0 : }
152 :
153 : REGISTER_EXEC("AllGatherVRingFor91093Executor", AllGatherVRingFor91093, CollAllGatherVRingFor91093Executor);
154 :
155 : } // namespace hccl
|
