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 <ios>
12 : #include <iostream>
13 :
14 : #include "log.h"
15 :
16 : #include "ccu_instruction_all_reduce_mesh1d.h"
17 : #include "ccu_rank_group.h"
18 : #include "ccu_ctx_creator_registry.h"
19 : #include "ccu_context_all_reduce_mesh1d.h"
20 : #include "ccu_temp_all_reduce_mesh_1D.h"
21 :
22 : namespace Hccl {
23 : static CcuInstRegister<CcuContextAllReduceMesh1D> g_registrarAllReduce(
24 : CcuInstType::CCU_ALL_REDUCE_MESH_1D_DIRECT);
25 :
26 0 : CcuTempAllReduceMesh1D::CcuTempAllReduceMesh1D(const RankId virtualRank, const u32 tempRankSize,
27 : const std::vector<std::vector<RankId>> &tempVTopo,
28 0 : const std::map<RankId, u32> &tempVirtRankMap)
29 0 : : CcuAlgTemplateBase(virtualRank, tempRankSize, tempVTopo, tempVirtRankMap)
30 : {
31 0 : }
32 :
33 0 : CcuTempAllReduceMesh1D::~CcuTempAllReduceMesh1D()
34 : {
35 0 : }
36 :
37 0 : void CcuTempAllReduceMesh1D::InitReduceInfo(const ReduceOp &reduceOp, const DataType &dataType) {
38 0 : reduceOp_ = reduceOp;
39 0 : dataType_ = dataType;
40 0 : }
41 :
42 0 : HcclResult CcuTempAllReduceMesh1D::CalcSliceInfo(const AllignInfo &allignInfo, const u64 dataSize,
43 : RankSliceInfo &sliceInfoVec)
44 : {
45 0 : CHK_RET(CalcSliceInfoAllReduce(allignInfo, tempRankSize_, dataSize, sliceInfoVec));
46 0 : return HcclResult::HCCL_SUCCESS;
47 : }
48 :
49 0 : HcclResult CcuTempAllReduceMesh1D::CalcRes(AlgTempResReq &tempResReq)
50 : {
51 0 : tempResReq.queNum = 1;
52 0 : tempResReq.streamNum = tempResReq.queNum;
53 0 : HCCL_INFO("[CalcRes] tempResReq.queNum[%u]", tempResReq.queNum);
54 0 : CHK_RET(CalcResLinksMesh(myRank_, tempRankSize_, tempVTopo_, linkNumBtwPeers_, tempResReq));
55 0 : return HcclResult::HCCL_SUCCESS;
56 : }
57 :
58 : /* CCU数据类型校验规则
59 : * Reduce算子：
60 : * 高精度模式，当dataType==outputDataType时，可选类型为FP32、FP16、BF16、UINT8、INT16、INT32；
61 : * 低精度模式，当dataType!=outputDataType时，dataType可选范围HIF8、E4M3、E5M2、INT8；outputDataType可选范围FP32、FP16、BF16；
62 : * 非Reduce算子：任意数据类型，dataType==outputDataType即可。
63 : */
64 0 : HcclResult CcuTempAllReduceMesh1D::CheckCcuDataType() const
65 : {
66 0 : if (op_.dataType == op_.outputDataType) {
67 : // allreduce算子高精度模式
68 0 : HCCL_INFO("HIGH PRECISION");
69 : set<DataType> highPrecisionSupportedInputDataType
70 0 : = {DataType::FP32, DataType::FP16, DataType::BFP16, DataType::UINT8, DataType::INT16, DataType::INT32};
71 0 : if (highPrecisionSupportedInputDataType.count(op_.dataType) == 0) {
72 0 : HCCL_ERROR("Unsupported DataType [%s] For OpType [%s].",
73 : op_.dataType.Describe().c_str(), op_.opType.Describe().c_str());
74 0 : return HcclResult::HCCL_E_PARA;
75 : }
76 0 : } else if (op_.outputDataType != DataType::INVALID) {
77 : // allreduce算子低精度模式
78 0 : HCCL_INFO("LOW PRECISION");
79 0 : HCCL_ERROR("Unsupported LOW PRECISION, Output DataType [%s] For OpType [%s].",
80 : op_.outputDataType.Describe().c_str(), op_.opType.Describe().c_str());
81 0 : return HcclResult::HCCL_E_PARA;
82 : }
83 0 : HCCL_INFO("CheckCcuDataType Success!");
84 0 : return HcclResult::HCCL_SUCCESS;
85 : }
86 :
87 0 : HcclResult CcuTempAllReduceMesh1D::Run(const TempFuncs &tempFuncs, const RankSliceInfo &sliceInfoVec,
88 : const BuffInfo &buffInfo, const ResLinks &tempLinks,
89 : std::vector<InsQuePtr> &tempInsQues)
90 : {
91 0 : CHK_PRT_RET(tempInsQues.empty(),
92 : HCCL_ERROR("[CcuTempAllReduceMesh1D] empty queue"), HcclResult::HCCL_E_INTERNAL);
93 0 : CHK_PTR_NULL(tempInsQues[0]);
94 0 : opMode_ = tempFuncs.opMode;
95 0 : buffInfo_ = buffInfo;
96 0 : CcuInstructionAllReduceMesh1D ccuInsAllReduceMesh1D;
97 0 : std::vector<uint64_t> dimSize;
98 0 : dimSize.push_back(tempRankSize_);
99 : uint64_t inputAddr;
100 : uint64_t outputAddr;
101 :
102 0 : if (op_.outputDataType == DataType::INVALID) {
103 0 : op_.outputDataType = op_.dataType;
104 : }
105 0 : CHK_RET(CheckCcuDataType());
106 0 : if (opMode_ == OpMode::OPBASE) {
107 0 : if (tempFuncs.isForepart) {
108 0 : inputAddr = BufferTypeToAddr(tempFuncs.usrData.usrInSlices[0].GetType()) + tempFuncs.usrData.usrInSlices[0].GetOffset();
109 : } else {
110 0 : inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff;
111 : }
112 0 : if (tempFuncs.isBottom) {
113 0 : outputAddr = BufferTypeToAddr(tempFuncs.usrData.usrOutSlices[0].GetType()) + tempFuncs.usrData.usrOutSlices[0].GetOffset();
114 : } else {
115 0 : outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff;
116 : }
117 : } else {
118 0 : inputAddr = BufferTypeToAddr(buffInfo_.inBuffType) + buffInfo_.inBuffBaseOff + tempFuncs.usrData.usrInSlices[0].GetOffset();
119 0 : outputAddr = BufferTypeToAddr(buffInfo_.outBuffType) + buffInfo_.outBuffBaseOff + tempFuncs.usrData.usrOutSlices[0].GetOffset();
120 : }
121 0 : HCCL_INFO("inputAddr[%llu], outputAddr[%llu]", inputAddr, outputAddr);
122 :
123 0 : uint64_t sliceSize = sliceInfoVec[myRank_][0].size; // 获取本rank需要处理的数据量
124 0 : uint64_t offset = sliceInfoVec[myRank_][0].offset; // 自己需要 reduce 的数据基于 inputAddr 的偏移
125 : uint64_t token;
126 0 : CHK_RET(GetToken(op_, token));
127 0 : ccuInsAllReduceMesh1D.Init(static_cast<uint32_t>(myRank_), inputAddr, outputAddr, sliceSize, offset, token, op_, tempVTopo_);
128 0 : HCCL_INFO("[CcuTempAllReduceMesh1D] Run Init: myRank_[%d], dimSize[%llu], inputAddr[%llu], outputAddr[%llu],"\
129 : "sliceSize[%llu], offset[%llu]", myRank_, dimSize[0], inputAddr, outputAddr, sliceSize, offset);
130 :
131 0 : std::vector<LinkData> links;
132 0 : for (auto &pair : tempLinks) {
133 0 : if (pair.second.empty()) {
134 0 : continue;
135 : }
136 0 : links.push_back(pair.second[0]);
137 : }
138 0 : HCCL_INFO("[CcuTempAllReduceMesh1D] links.size[%zu]", links.size());
139 0 : ccuInsAllReduceMesh1D.SetLinks(links);
140 :
141 0 : RankGroup rankGroup;
142 :
143 0 : for (auto &peer : tempVTopo_[0]) {
144 0 : rankGroup.AddRank(peer);
145 : }
146 0 : u32 cntCkeNum = 4;
147 0 : ccuInsAllReduceMesh1D.SetCntCkeNum(cntCkeNum);
148 0 : ccuInsAllReduceMesh1D.SetRankGroup(rankGroup);
149 0 : HCCL_INFO("CCUInsAllReducemesh1D is [%s]", ccuInsAllReduceMesh1D.Describe().c_str());
150 0 : tempInsQues[0]->Append(std::move(std::make_unique<CcuInstructionAllReduceMesh1D>(ccuInsAllReduceMesh1D)));
151 :
152 0 : return HcclResult::HCCL_SUCCESS;
153 0 : }
154 : } // namespace Hccl
|
