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_reduce_scatter_v_mesh_executor.h"
12 : #include "alg_template_register.h"
13 :
14 : namespace hccl {
15 :
16 0 : CollReduceScatterVMeshExecutor::CollReduceScatterVMeshExecutor(
17 : const HcclDispatcher dispatcher,
18 0 : std::unique_ptr<TopoMatcher> &topoMatcher)
19 0 : : CollReduceScatterVExecutor(dispatcher, topoMatcher)
20 : {
21 0 : DMAReduceFlag_ = false;
22 0 : }
23 :
24 0 : void CollReduceScatterVMeshExecutor::ParseParam(const OpParam& param)
25 : {
26 : // 910B 图模式非确定计算,inlineReduce使能,MESH拓扑场景下,创建一个mesh平面
27 0 : bool isInlineReduce = IsSupportSDMAReduce(param.inputPtr, param.outputPtr, param.VDataDes.dataType,
28 0 : param.reduceType);
29 0 : meshSinglePlane_ = (topoAttr_.deviceType == DevType::DEV_TYPE_910B) &&
30 0 : topoMatcher_->GetExternalInputHcclDeterministic() == DETERMINISTIC_DISABLE &&
31 0 : isInlineReduce && (workflowMode_ == HcclWorkflowMode::HCCL_WORKFLOW_MODE_OPS_KERNEL_INFO_LIB);
32 :
33 : // 记录图模式总数据量
34 0 : u64 totalSize = 0;
35 0 : const u64* counts = static_cast<const u64*>(param.VDataDes.counts);
36 0 : for (u32 i = 0; i < topoAttr_.userRankSize; i++) {
37 0 : totalSize += counts[i] * SIZE_TABLE[param.VDataDes.dataType];
38 : }
39 0 : totalSize_ = totalSize;
40 0 : scratchMemFlag_ = false; // mesh算法不需要使用scrachMem
41 0 : aicpuUnfoldMode_ = param.aicpuUnfoldMode;
42 0 : }
43 :
44 0 : HcclResult CollReduceScatterVMeshExecutor::CalcStreamNum(u32& streamNum)
45 : {
46 0 : u32 totalStreamNum = topoAttr_.deviceNumPerAggregation > 1U ? topoAttr_.deviceNumPerAggregation - 1U : 1U;
47 0 : streamNum = totalStreamNum - 1U;
48 0 : HCCL_INFO("[CollReduceScatterVMeshExecutor][CalcStreamNum] tag[%s] streamNum[%u]",
49 : tag_.c_str(), streamNum);
50 0 : return HCCL_SUCCESS;
51 : }
52 :
53 0 : HcclResult CollReduceScatterVMeshExecutor::CalcCommInfo(
54 : std::vector<LevelNSubCommTransport>& opTransport)
55 : {
56 0 : TransportMemType inputType = TransportMemType::RESERVED;
57 0 : TransportMemType outputType = TransportMemType::RESERVED;
58 0 : CHK_RET(CalcTransportMemType(inputType, outputType));
59 0 : CHK_RET(CalcLevel0CommInfo(inputType, outputType, opTransport));
60 0 : return HCCL_SUCCESS;
61 : }
62 :
63 0 : HcclResult CollReduceScatterVMeshExecutor::CalcTransportMemType(TransportMemType &inputType,
64 : TransportMemType &outputType)
65 : {
66 0 : inputType = TransportMemType::PARAM_INPUT;
67 0 : outputType = TransportMemType::PARAM_OUTPUT;
68 :
69 0 : HCCL_INFO("[CollReduceScatterVMeshExecutor][CalcTransportMemType] tag[%s] inputType[%d],"
70 : " outputType[%d]", tag_.c_str(), inputType, outputType);
71 0 : return HCCL_SUCCESS;
72 : }
73 :
74 0 : HcclResult CollReduceScatterVMeshExecutor::CalcLevel0CommInfo(TransportMemType inputType,
75 : TransportMemType outputType,
76 : std::vector<LevelNSubCommTransport>& opTransport)
77 : {
78 0 : CommParaInfo commParaLevel0(COMM_LEVEL0, CommType::COMM_TAG_MESH);
79 0 : commParaLevel0.meshSinglePlane = meshSinglePlane_;
80 0 : CHK_RET(CalcCommPlaneInfo(tag_, commParaLevel0, opTransport[COMM_LEVEL0], inputType, outputType));
81 0 : return HCCL_SUCCESS;
82 0 : }
83 :
84 0 : HcclResult CollReduceScatterVMeshExecutor::KernelRun(const OpParam ¶m, ExecMem &execMem)
85 : {
86 0 : HCCL_CONFIG_INFO(HCCL_ALG, "[CollReduceScatterVMeshExecutor] ReduceScatterV mesh run");
87 0 : HcclDataType dataType = param.VDataDes.dataType;
88 0 : const u32 unitSize = SIZE_TABLE[dataType];
89 :
90 0 : CHK_RET(CheckCommSize(COMM_LEVEL0, COMM_INDEX_0 + 1));
91 0 : SubCommInfo subCommInfo = GetSubCommInfo(COMM_LEVEL0, COMM_INDEX_0);
92 0 : u32 rankSize = subCommInfo.localRankSize;
93 :
94 0 : CHK_RET(ActiveSlaveStreams(param.stream));
95 :
96 : /* *******************节点内reducescatter ******************************************/
97 : // reduce_scatter_v 计算slice,数据分成ranksize份,每份的起始偏移和大小
98 0 : std::vector<Slice> inputSlices;
99 0 : const auto counts = static_cast<u64*>(param.VDataDes.counts);
100 0 : const auto displs = static_cast<u64*>(param.VDataDes.displs);
101 0 : for (u32 rank = 0; rank < rankSize; ++rank) {
102 0 : Slice userslice;
103 0 : userslice.offset = displs[rank] * unitSize;
104 0 : userslice.size = counts[rank] * unitSize;
105 0 : inputSlices.emplace_back(std::move(userslice));
106 : }
107 :
108 0 : u64 reduceAttr = GetReduceAttr(execMem.inputMem, execMem.outputMem, dataType, param.reduceType);
109 0 : std::unique_ptr<AlgTemplateBase> tempAlg = AlgTemplateRegistry::Instance().GetAlgTemplate(
110 0 : TemplateType::TEMPLATE_REDUCESCATTER_MESH_ATOMIC, dispatcher_);
111 0 : CHK_SMART_PTR_NULL(tempAlg);
112 :
113 0 : CHK_RET(tempAlg->Prepare(execMem.inputMem, execMem.outputMem, execMem.scratchMem, execMem.count, dataType,
114 : param.stream, param.reduceType, LEVEL0_BRIDGE_RANK_ID, inputSlices, 0, reduceAttr,
115 : algResResp_->slaveStreams, algResResp_->notifiesMain, algResResp_->notifiesAux,
116 : topoAttr_.userRank));
117 :
118 0 : CHK_RET(tempAlg->RegisterProfiler(
119 : (subCommInfo.localRankSize << PROF_RANKSIZE_OFFSET_OF_PLANEID) + subCommInfo.localRank,
120 : PROF_STAGE_0, HCCL_EXEC_STEP_NOT_SET, param.stream));
121 :
122 0 : CHK_RET(RunTemplate(tempAlg, subCommInfo));
123 :
124 0 : return HCCL_SUCCESS;
125 0 : }
126 :
127 : REGISTER_EXEC("ReduceScatterVMeshExecutor",
128 : ReduceScatterVMesh, CollReduceScatterVMeshExecutor);
129 : }
|