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_reduce_fast_double_ring_for_910_93_executor.h"
12 :
13 : namespace hccl {
14 :
15 4 : CollAllReduceFastDoubleRingFor91093Executor::CollAllReduceFastDoubleRingFor91093Executor(
16 4 : const HcclDispatcher dispatcher, std::unique_ptr<TopoMatcher> &topoMatcher)
17 4 : : CollAllReduceRingFor91093Executor(dispatcher, topoMatcher)
18 : {
19 4 : if (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OP_BASE) {
20 4 : DMAReduceFlag_ = true;
21 : } else {
22 0 : DMAReduceFlag_ = false;
23 : }
24 4 : }
25 :
26 16 : HcclResult CollAllReduceFastDoubleRingFor91093Executor::DoubleRingReduceScatter(const std::string &tag,
27 : DeviceMem inputMem, DeviceMem outputMem, const u64 count, const HcclDataType dataType,
28 : const HcclReduceOp reductionOp, const std::vector<std::vector<Slice>> multRingsSliceZero, Stream stream,
29 : s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
30 : const std::vector<std::vector<Slice>> multRingsUserMemSlice, const bool disableDMAReduce)
31 : {
32 : (void)tag;
33 16 : HCCL_CONFIG_INFO(HCCL_ALG,
34 : "[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingReduceScatter] DoubleRingReduceScatter starts");
35 16 : HcclResult ret = HCCL_SUCCESS;
36 16 : u32 ringNum = multRingsSliceZero.size();
37 16 : CHK_RET(CheckCommSize(COMM_LEVEL0, ringNum));
38 : // 拿到ring环映射关系
39 16 : SubCommInfo level0ZeroCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
40 16 : auto nicList = topoAttr_.nicList;
41 : std::vector<std::vector<u32>> multiRingsOrder =
42 16 : GetRingsOrderByTopoType(level0ZeroCommInfo.localRankSize, topoType_, nicList);
43 16 : u64 reduceAttr = GetReduceAttr(inputMem, outputMem, dataType, reductionOp);
44 16 : SubCommInfo level0RingCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
45 : // 生成两个ring上的userMemIn_上对应的slices
46 16 : std::vector<std::vector<Slice>> userMemInputSlicesOfDoubleRing;
47 16 : CHK_RET(CollectMultiRingsUserMemSlices(ringNum, dataType, opInfo, multRingsSliceZero,
48 : multiRingsOrder, multRingsUserMemSlice, userMemInputSlicesOfDoubleRing));
49 : // 生成两个ring上的rankOrder
50 16 : std::vector<std::vector<u32>> rankOrders;
51 16 : CHK_RET(CollectMultiRingsRankOrder(ringNum, multiRingsOrder, rankOrders));
52 : // 初始化executor
53 16 : std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
54 16 : TemplateType::TEMPLATE_REDUCESCATTER_DB_RING_SLC, dispatcher_);
55 16 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_REDUCESCATTER_DB_RING_SLC in COMM_LEVEL0", __func__);
56 16 : CHK_SMART_PTR_NULL(tempAlg);
57 16 : ret = tempAlg->Prepare(inputMem, inputMem, outputMem, count, dataType, stream,
58 : multRingsSliceZero, reductionOp, LEVEL0_BRIDGE_RANK_ID, baseOffset, disableDMAReduce,
59 16 : reduceAttr, opInfo, topoAttr_.userRank, algResResp_->slaveStreams, algResResp_->notifiesMain,
60 16 : algResResp_->notifiesAux, rankOrders, userMemInputSlicesOfDoubleRing);
61 16 : CHK_PRT_RET(ret != HCCL_SUCCESS,
62 : HCCL_ERROR("[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingReduceScatter] Double ring "
63 : "ReduceScatter failed,return[%d]", ret), ret);
64 16 : u32 ringIndexOp = COMM_INDEX_0;
65 16 : u32 rankSize = level0RingCommInfo.localRankSize;
66 16 : ret = tempAlg->RegisterProfiler(((ringIndexOp + 1) << PROF_RINGINDEX_OFFSET_OF_PLANEID) +
67 16 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0RingCommInfo.localRank, profStage,
68 : HCCL_EXEC_STEP_NOT_SET, stream);
69 16 : CHK_PRT_RET(ret != HCCL_SUCCESS,
70 : HCCL_ERROR("[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingReduceScatter] Double ring "
71 : "ReduceScatter failed,return[%d]", ret), ret);
72 : // 空拷贝用于后续操作附着
73 16 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem, outputMem, stream, dispatcher_));
74 16 : ret = RunTemplate(tempAlg, level0RingCommInfo);
75 16 : CHK_PRT_RET(ret != HCCL_SUCCESS,
76 : HCCL_ERROR("[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingReduceScatter] Double ring "
77 : "ReduceScatter failed,return[%d]", ret), ret);
78 :
79 16 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem, outputMem, stream, dispatcher_));
80 16 : return HCCL_SUCCESS;
81 16 : }
82 :
83 16 : HcclResult CollAllReduceFastDoubleRingFor91093Executor::DoubleRingAllGather(
84 : const std::string &tag, DeviceMem inputMem, DeviceMem outputMem,
85 : const u64 count, const HcclDataType dataType, const std::vector<std::vector<Slice> > multRingsSliceZero,
86 : Stream stream, s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
87 : const std::vector<std::vector<Slice>> multRingsUserMemSlice)
88 : {
89 : (void)tag;
90 16 : HCCL_CONFIG_INFO(HCCL_ALG,
91 : "[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingAllGather] DoubleRingAllGather starts");
92 16 : HcclResult ret = HCCL_SUCCESS;
93 16 : u32 ringNum = multRingsSliceZero.size();
94 16 : CHK_RET(CheckCommSize(COMM_LEVEL0, ringNum));
95 : // 拿到ring环映射关系
96 16 : SubCommInfo level0ZeroCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
97 16 : auto nicList = topoAttr_.nicList;
98 : std::vector<std::vector<u32>> multiRingsOrder =
99 16 : GetRingsOrderByTopoType(level0ZeroCommInfo.localRankSize, topoType_, nicList);
100 : // 生成两个ring上的userMemOut_上对应的slices
101 16 : std::vector<std::vector<Slice>> userMemOutputSlicesOfDoubleRing;
102 16 : CHK_RET(CollectMultiRingsUserMemSlices(ringNum, dataType, opInfo, multRingsSliceZero,
103 : multiRingsOrder, multRingsUserMemSlice, userMemOutputSlicesOfDoubleRing));
104 : // 生成两个ring上的rankOrder
105 16 : std::vector<std::vector<u32>> rankOrders;
106 16 : CHK_RET(CollectMultiRingsRankOrder(ringNum, multiRingsOrder, rankOrders));
107 : // 初始化executor
108 16 : std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
109 16 : TemplateType::TEMPLATE_ALIGNED_ALL_GATHER_DOUBLE_RING, dispatcher_);
110 16 : HCCL_CONFIG_INFO(HCCL_ALG, "[%s] Run TEMPLATE_ALIGNED_ALL_GATHER_DOUBLE_RING in COMM_LEVEL0", __func__);
111 16 : CHK_SMART_PTR_NULL(tempAlg);
112 16 : CHK_RET(tempAlg->Prepare(const_cast<HcomCollOpInfo*>(opInfo), topoAttr_.userRank, algResResp_->slaveStreams,
113 : algResResp_->notifiesMain, algResResp_->notifiesAux, rankOrders, userMemOutputSlicesOfDoubleRing));
114 :
115 16 : ret = tempAlg->Prepare(outputMem, outputMem, inputMem, count, dataType, stream, multRingsSliceZero,
116 : HCCL_REDUCE_RESERVED, LEVEL0_BRIDGE_RANK_ID, baseOffset);
117 16 : CHK_PRT_RET(ret != HCCL_SUCCESS,
118 : HCCL_ERROR("[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingAllGather]Double ring "
119 : "AllGather failed, return[%d]", ret), ret);
120 16 : u32 ringIndexOp = COMM_INDEX_0;
121 16 : u32 rankSize = level0ZeroCommInfo.localRankSize;
122 16 : ret = tempAlg->RegisterProfiler(
123 16 : ((ringIndexOp + 1) << PROF_RINGINDEX_OFFSET_OF_PLANEID) +
124 16 : (rankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + level0ZeroCommInfo.localRank,
125 : profStage, HCCL_EXEC_STEP_NOT_SET, stream);
126 16 : CHK_PRT_RET(ret != HCCL_SUCCESS,
127 : HCCL_ERROR("[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingAllGather]Double ring "
128 : "AllGather failed, return[%d]", ret), ret);
129 :
130 : // 空拷贝用于后续操作附着
131 16 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem, outputMem, stream, dispatcher_));
132 16 : ret = RunTemplate(tempAlg, level0ZeroCommInfo);
133 16 : CHK_PRT_RET(ret != HCCL_SUCCESS,
134 : HCCL_ERROR("[CollAllReduceFastDoubleRingFor91093Executor][DoubleRingAllGather] Double ring "
135 : "AllGather failed,return[%d]", ret), ret);
136 : // 添加空task,保证执行时不乱序
137 16 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem, outputMem, stream, dispatcher_));
138 16 : return HCCL_SUCCESS;
139 16 : }
140 :
141 16 : HcclResult CollAllReduceFastDoubleRingFor91093Executor::RunIntraSeverReduceScatter(
142 : const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
143 : const u64 count, const HcclDataType &dataType, const HcclReduceOp &reductionOp,
144 : const std::vector<std::vector<Slice>> &multRingsSliceZero, const Stream &stream, s32 profStage,
145 : const u64 baseOffset, const HcomCollOpInfo *opInfo,
146 : const std::vector<std::vector<Slice>> &multRingsUserMemSlice, const bool disableDMAReduce)
147 : {
148 16 : CHK_RET(DoubleRingReduceScatter(tag, inputMem, outputMem, count, dataType, reductionOp,
149 : multRingsSliceZero, stream, profStage, baseOffset, opInfo, multRingsUserMemSlice, disableDMAReduce));
150 16 : return HCCL_SUCCESS;
151 : }
152 :
153 16 : HcclResult CollAllReduceFastDoubleRingFor91093Executor::RunIntraSeverAllGather(
154 : const std::string &tag, DeviceMem &inputMem, DeviceMem &outputMem,
155 : const u64 count, const HcclDataType &dataType, const std::vector<std::vector<Slice>> &multRingsSliceZero,
156 : const Stream &stream, s32 profStage, const u64 baseOffset, const HcomCollOpInfo *opInfo,
157 : const std::vector<std::vector<Slice>> &multRingsUserMemSlice)
158 : {
159 16 : CHK_RET(DoubleRingAllGather(tag, inputMem, outputMem, count, dataType,
160 : multRingsSliceZero, stream, profStage, baseOffset, opInfo, multRingsUserMemSlice));
161 16 : return HCCL_SUCCESS;
162 : }
163 :
164 : REGISTER_EXEC("AllReduceFastDoubleRingFor91093Executor", AllReduceFastDoubleRingFor91093,
165 : CollAllReduceFastDoubleRingFor91093Executor);
166 :
167 : } // namespace hccl
|
